text
stringlengths 1
2.56M
| id
stringlengths 40
40
| metadata
dict |
|---|---|---|
\section{Introduction}\label{Sec:Intro}
Experimental advances in studying the non-equilibrium dynamics of isolated quantum-many-body systems have in recent years led to the quest for a better theoretical understanding of such dynamics.
Systems of interest in this context include ultracold Bose gases \cite{Anderson2001a,Eiermann2004a,Sadler2006a,Weller2008a.PRL101.130401,Weiler2008a,Neely2010a},
semi-conductor exciton–polariton
superfluids \cite{Kasprzak2006a.etal,Lagoudakis2008a,Lagoudakis2009a,Amo2011a,roumpos2011single} and
topological condensed matter systems \cite{Qi2009a}.
The non-equilibrium dynamics of such systems can be observed after a quench into a far-from-equilibrium initial state,
as it can be realized experimentally by a sudden change of a parameter of the system, \textit{e.\,g.}\ temperature or the scattering length of the constituent particles.
During the relaxation process and before thermal equilibrium is reached the system's dynamics is generically strongly nonlinear and exhibits properties of quantum turbulence, the quantum analogue of classical turbulence.
This line of research adds new aspects to the general phenomena of quantum and classical turbulence which have attracted considerable attention since decades and include, among others, universal scaling of correlation functions \cite{Kolmogorov1941a,Kolmogorov1941b,Kolmogorov1941c,Obukhov1941} with characteristic energy cascades in the system, quasi-stationary flow \cite{Richardson1920}, and the observation of non-thermal fixed points \cite{Berges:2008wm,Berges:2008sr,Scheppach:2009wu,Berges:2010ez,Prufer:2018hto,Erne:2018gmz} (see \cite{Schmied:2018mte} for a review).
The strongly nonlinear nature of the dynamics poses a great challenge as it renders a theoretical description very intricate.
Even if the system's constituents are weakly coupled, commonly used methods such as perturbation theory break down
as a consequence of strong correlations that build up dynamically,
often related to the occurrence of highly nonlinear topological excitations in the system.
These include in particular quantized vortices, the quantum analogues of classical eddies.
In modern experiments with Bose--Einstein-condensed ultracold quantum gases
vortex ensembles can even be produced by stirring the condensate with a
laser \cite{Raman2001a.PRL.87.210402R,Abo-Shaeer2001a}.
Quantized vortices are known to play a crucial role also in superfluid helium \cite{Donnelly1991a} where strong
coupling prevails in the system and the theoretical description is even more challenging.
In the present work we will be concerned with the dynamics of a vortex--anti-vortex pair
which is the most basic configuration that allows one to study the interaction of topological defects.
We will use the theoretical framework of holography or gauge/gravity duality for our study.
It offers a promising pathway towards a better understanding
of non-equilibrium dynamics and is particularly well suited for the description of strongly coupled quantum systems.
In holography, a strongly coupled quantum field theory without gravity in $d$ dimensions is mapped
to a theory with classical gravity in $d+1$ dimensions.
The quantum behavior of the field theory is thus encoded in a classical gravitational theory in a higher-dimensional spacetime.
A variety of holographic dualities has been devised since such a duality was first proposed in the form
of the AdS/CFT correspondence \cite{Maldacena:1997re,Gubser:1998bc,Witten:1998qj}.
AdS/CFT posits that $\mathcal{N}=4$ super Yang--Mills (SYM) theory, a
supersymmetric and conformal gauge theory (CFT) in $3+1$ dimensions, is dual to type IIB string theory
on $\text{AdS}_5 \times S^5$, where AdS$_5$ stands for an Anti-de Sitter spacetime and $S^5$ denotes a five-sphere.
A key feature is that the AdS/CFT correspondence simplifies in the limit of strong coupling and large number of colors on the field theory side,
corresponding to which the string theory side simplifies to a weakly coupled theory of Einstein gravity.
Generalizations to dualities between other strongly coupled quantum theories and weakly coupled gravitational theories
have become a powerful tool to address many problems in strongly coupled and hence manifestly nonperturbative quantum field theories.
The quantum theory in $d$ dimensions is usually referred to as the `boundary theory' as one can imagine it to `live' on
the boundary of the $(d+1)$-dimensional AdS spacetime.
The AdS spacetime away from the boundary, on the other hand, is referred to as the `bulk'.
Holographic dualities can in particular be used to describe quantum field theories at finite temperature.
Here, typically, the dual gravitational representation includes a black hole in the bulk of the AdS space, with its
Hawking temperature being equal to the temperature in the boundary theory. Similarly, a chemical
potential in the field theory corresponds to a charged black hole in the bulk.
Holography has been applied to a variety of physical systems (see \textit{e.\,g.}\ \cite{Adams:2012th} for a review), ranging
from quark--gluon plasma physics, see \textit{e.\,g.}\ \cite{CasalderreySolana:2011us}, to condensed matter
systems, see for instance \cite{Hartnoll:2016apf}, including the pioneering work on
holographic superconductors and superfluids
in \cite{Gubser:2008px,Hartnoll:2008vx,Herzog:2008he} (see \cite{Herzog:2009xv} for a review).
Specifically, it has been found \cite{Gubser:2008px,Hartnoll:2008vx,Herzog:2008he} that an
Abelian Higgs model coupled to gravity on a $(3+1)$-dimensional Anti-de Sitter spacetime with a black hole
has a dual interpretation in terms of a superconductor or superfluid (for a review, see \cite{Hartnoll:2009sz}).
Superfluidity is associated with the breaking of a global U$(1)$ symmetry.
In the holographic model, this is incorporated by the breaking of a local U$(1)$ symmetry and the associated
condensation of the scalar field in the bulk.
Remarkably, the simple Abelian Higgs model with classical gravity encodes the quantum behavior
of the superfluid, even if strong correlations are present in the system.
It is therefore well suited for studying the real-time dynamics of the superfluid far from equilibrium.
Extensive analytical as well as numerical studies of the holographic superfluid have been performed.
Among them are studies of the details of the phase transition into the superfluid
phase\footnote{Some attention has also been paid to modified holographic superfluid models and how gravity modifications
affect the nature of the phase transition, in particular with regard to critical exponents that deviate
from the mean-field values they take for the original holographic superfluid,
\textit{e.\,g.}\ \cite{Franco:2009yz,Franco:2009if,Herzog:2010vz}.} \cite{Herzog:2008he,Maeda:2009wv}, computations of the
speed of sound and the investigation of sound waves,
see \textit{e.\,g.}\ \cite{Herzog:2008he,Basu:2008bh,Herzog:2009ci,Yarom:2009uq,Amado:2009ts}, and studies of topological defects in the
system, see \textit{e.\,g.}\ \cite{Keranen:2009vi,Keranen:2009ss,Keranen:2009re,Keranen:2010sx,Adams:2012pj,Dias:2013bwa,Chesler:2014gya,Ewerz:2014tua,Du:2014lwa,Lan:2016cgl,Lan:2017qxm,Guo:2018mip,Lan:2018llf,Li:2019swh,Xia:2019eje,Yang:2019ibe,Xu:2019msl}.
With regard to topological defects, much attention has been devoted to the study of vortices and in particular
ensembles of vortices and anti-vortices.
While quantized vortices constitute an intense field of research in general, they play a particular role in the holographic superfluid as
the main sites of energy and momentum dissipation in this system \cite{Adams:2012pj}.
To wit, they have a geometric interpretation in that tubes emanating from the vortex cores at the boundary punch holes
through the scalar charge cloud in the bulk so that energy can flow from the boundary to the black hole where it is absorbed,
see \cite{Adams:2012pj}.
In the present work, we will study the annihilation process of a single vortex--anti-vortex pair, or vortex dipole, in the holographic superfluid.
By considering just one vortex and one anti-vortex we have a clean environment for investigating their mutual interaction.
At the same time, vortex pairs can be viewed as basic building blocks for the interactions in larger ensembles of vortices
and are thus relevant in a more general context. Our method builds on previous work in \cite{Ewerz:2014tua} where the vortex dynamics of
an ensemble of vortices was studied for certain classes of quench-like initial conditions, and we use the same numerical
implementation.
While the annihilation process of a vortex dipole in a holographic superfluid has been studied before \cite{Lan:2018llf},
we are able to study new aspects of this system by using new methods. Specifically, by introducing
an advanced tracking technique, we drastically increase the precision of the vortex trajectories
that we extract from the numerical simulations. This gain in precision allows us to compute directly
the vortex velocities and accelerations. We can then infer universal properties of the motion of the vortex dipole
and can cleanly identify a power law for the behavior of the vortex--anti-vortex separation.
We also compare the vortex trajectories to a general point-vortex model from the literature.
Furthermore, we study the temperature dependence of the vortex dynamics.
Finally, we discuss some potential further applications of the methods for tracking vortices
developed here.
Our paper is organized as follows.
In section \ref{Sec:HoloFluid} we review the gravitational model for the holographic superfluid. We discuss
its equations of motion and their numerical implementation used for the real-time simulation of the system.
Section \ref{Sec:VortAntiVort} is concerned with the study of the vortex--anti-vortex pair in the holographic superfluid.
We describe the numerical initialization of vortices and introduce a new method to determine their positions during
the evolution of the system with high precision. Subsequently, we present our results for the vortex trajectories
from the initial separation to the annihilation. Here we also discuss universal properties of the motion of the vortex dipole
and compute velocities and accelerations of the vortices.
In section \ref{Sec:MuDependence} we study the dependence of the vortex
dynamics on the temperature of the superfluid. We close with a summary and outlook in section \ref{Sec:Summary}.
Some technical details are presented in the appendices.
\section{Holographic Superfluid}
\label{Sec:HoloFluid}
In this section we review the $3+1$-dimensional gravitational model dual to a superfluid in $2+1$ dimensions.
\subsection{Holographic Model}
\label{Sec:Model}
In his seminal work \cite{Gubser:2008px}, Gubser showed that, if coupled to gravity with a negative cosmological constant,
even a simple Abelian Higgs model without self-interaction exhibits spontaneous symmetry breaking.
This laid the ground for a holographic description of superconductivity and superfluidity as subsequently developed
in \cite{Hartnoll:2008vx,Herzog:2008he}.\footnote{The
holographic superfluid was devised in \cite{Herzog:2008he} after the dual description of a superconductor
was introduced and discussed in \cite{Hartnoll:2008vx}. From the field-theory perspective,
the main difference between a superfluid and a superconductor is that for a superfluid a global symmetry is
spontaneously broken whereas for a superconductor a local symmetry is spontaneously broken.
The differences between the holographic superfluid and the holographic superconductor with regard
to topological vortex defects are discussed in \cite{Dias:2013bwa}, for example.}
The spontaneous breaking of the local $U(1)$ symmetry in the AdS bulk occurs due to the negative contribution of the gauge
field to the effective scalar mass in the vicinity of the black-hole horizon which drives the system unstable and leads to the
formation of a scalar charge cloud. In the dual field theory this corresponds to the breaking of a global $U(1)$ symmetry and
a phase transition from the normal fluid to a superfluid state.
The gravity model holographically describing a $(2+1)$-dimensional superfluid is given by the action \cite{Gubser:2008px}
\begin{align}\label{action}
S=\frac{1}{2\kappa}\int\,\text{d}^4x\,\sqrt{-g}\left( \mathcal{R}-2\Lambda+\frac{1}{q^2}\mathcal{L_\text{matter}}\right)\,,
\end{align}
where $g$ is the determinant of the metric $g_{MN}$ ($M,N=0,\dots,3$), $\mathcal{R}$ its associated Ricci scalar,
$\kappa$ Newton's constant in $4$ spacetime dimensions,
and $\Lambda=-3/L^2_\text{AdS}$ the cosmological constant of the $(3+1)$-dimensional
Anti-de Sitter spacetime with curvature radius $L_\text{AdS}$.
The matter part $\mathcal{L_\text{matter}}$ of the Lagrangian is given by
\begin{align}
\mathcal{L_\text{matter}}=-\frac{1}{4}F_{MN}F^{MN}-|(\nabla_M- \text{i} A_M)\Phi|^2 -m^2|\Phi|^2\,,
\end{align}
where the charge $q$ of the scalar field has been pulled out of the Lagrangian $\mathcal{L_\text{matter}}$ by an appropriate rescaling of the fields.
Using this convention makes evident that $q$ controls the effect of the matter part of the model on the dynamical gravity, \textit{cf.}\ \eqref{action}.
The mass of the scalar field $\Phi$ is $m$.
We use $\nabla_M$ to denote the Levi-Civita connection of the metric.
The gravity model is thus given by a simple Abelian Higgs model formulated on a $(3+1)$-dimensional
gravity background with minimal coupling between the complex scalar field $\Phi$ and the gauge
connection $A_M$. The kinetic term for the Abelian gauge field is formulated in terms of its field
strength tensor $F_{MN}=\nabla_MA_N-\nabla_NA_M$. Note that no terms of higher order in the scalar field than the quadratic
mass term are needed. This is different from the usual Higgs mechanism in the Standard Model
where the $\Phi^4$ interaction term is essential for spontaneous symmetry breaking to occur.
In contrast to flat Minkowski spacetime, the AdS-type spacetime allows for
an instability to occur even if there is no self-interaction of the scalar field.
The squared scalar mass $m^2$ may even be negative without the symmetry being broken as
long as it is above the Breitenlohner--Freedman bound \cite{Breitenlohner:1982bm,Breitenlohner:1982jf,Mezincescu:1984ev}.
At low temperatures, the effective scalar mass can be lowered below this bound by the negative
contribution of the bulk gauge field, thus giving rise to an instability and symmetry breaking.
In this work we choose $m^2=-2/L^2_\text{AdS}$ which is above the Breitenlohner--Freedman bound,
and we set $L_\text{AdS}=1$ for convenience.
The interpretation of this model according to the holographic dictionary is the following.
The Abelian gauge field in the bulk $A_M$ is dual to a global conserved $U(1)$ current operator $j^\mu$ in
the boundary field theory ($\mu=0,1,2$). Similarly, the complex scalar field $\Phi$ with mass $m$ and charge $q$, to
which the gauge field $A_M$ is dynamically coupled, is dual to a complex scalar field operator $\Psi$.
This operator plays a crucial role as its quantum expectation value $\psi=\langle\Psi\rangle$ is the order
parameter for the phase transition from a normal fluid to the superfluid.
In a usual field-theory context, superfluidity is described by a complex scalar field whose vacuum expectation value
assumes non-zero values in the symmetry-broken phase, corresponding to the existence of a superfluid condensate.
Instead of considering the full set of equations of motion corresponding to the action \eqref{action}
we simplify the model by taking the so-called probe approximation. This means that the gravity part of the action
is solved independently of the matter fields, \textit{i.\,e.}\ the back-reaction of the matter fields on the metric is neglected.
This approximation is justified if the scalar charge $q$ is large.
The matter part of the model is then solved in the fixed gravity background.
For a detailed discussion of the probe limit and its validity see \cite{Hartnoll:2008vx} and, in the context of
vortex defects, \cite{Albash:2009iq}.
The pure gravity part which comprises only the first two terms in the action \eqref{action}, \textit{i.\,e.}\ the
Ricci scalar and the cosmological constant, can be solved straightforwardly.
The vacuum Einstein equations are given by
\begin{align}
\mathcal{R}_{MN}-\frac{1}{2}\mathcal{R}\,g_{MN}+\Lambda\,g_{MN}=0\,.
\end{align}
They are solved by a $(3+1)$-dimensional Schwarzschild Anti-de Sitter spacetime that
has a planar black hole with horizon situated at the bulk position $z=z_\text{h}$.
In this work we choose to work with infalling Eddington--Finkelstein coordinates in which the line element is given by
\begin{align}\label{metric}
\text{d}s^2=\frac{L_\text{AdS}^2}{z^2}\left(-h(z)\text{d}t^2+\text{d}\bm{x}^2-2\text{d}t\,\text{d}z\right)\,,
\end{align}
with the horizon function
\begin{align}
h(z)=1-\left(\frac{z}{z_\text{h}}\right)^3\,.
\end{align}
The choice of infalling Eddington--Finkelstein coordinates is convenient for two reasons.
Firstly, unlike the usual Schwarzschild coordinates they do not exhibit a coordinate
singularity at the black-hole horizon. Secondly, all null geodesics falling into the black hole are of radial nature.
These two features are particularly useful for numerical
simulations in holography \cite {Chesler:2013lia} due to the ensuing natural implementation of boundary conditions.
In these coordinates, $z$ is the holographic (radial) coordinate and $(t, \bm{x})=(t, x, y)$ are
the time and remaining spatial directions, which also make up the coordinates of the dual field
theory that can be thought of as living on the boundary at $z=0$ of the Anti-de Sitter spacetime.
As a radial coordinate $z$ is taken to be positive and one refers to the spacetime for $z>0$ as the bulk.
The Hawking temperature associated with the metric \eqref{metric} is
\begin{align}\label{temp}
T=\frac{3}{4\pi z_\text{h}}\,,
\end{align}
and it coincides with the temperature of the boundary theory.
Taking the background metric to be fixed one can vary the action \eqref{action} with
respect to the matter fields to obtain their equations of motion,
\begin{align}
\nabla_MF^{MN}&=J^N\,,\label{Maxwell}\\
\left(-D^2+m^2\right)\Phi&=0\,, \label{KGE}
\end{align}
with the current
\begin{align}\label{current}
J^N&=\text{i}\left(\Phi^\ast D^N\Phi-\Phi\left(D^N\Phi\right)^\ast\right)\,,
\end{align}
and the general covariant derivative $ D_M=\nabla_M-\text{i}A_M $.
This is a classical system of coupled partial differential equations on the curved $4$-dimensional
Anti-de Sitter spacetime \eqref{metric}. The Maxwell and Klein-Gordon
equations, \eqref{Maxwell} and \eqref{KGE} respectively, are coupled through the electromagnetic current \eqref{current}.
This classical system of equations of motion together with the appropriate boundary conditions
encode the quantum behavior of the dual field theory.
We are interested in computing
the full quantum evolution of the order parameter field $\psi=\langle \Psi \rangle$, where $\Psi$ is the
operator dual to the complex scalar field $\Phi$ in the bulk.
Finding the time evolution of the
expectation value $\psi$ thus reduces to solving the classical equations of motion for $\Phi$.
To solve the equations of motion, we first need to impose boundary conditions for the gauge field $A_M$ as well as for
the scalar field $\Phi$.
The boundary condition for the temporal component $A_t$ of the gauge field at the conformal boundary $z=0$ fixes the chemical potential $\mu$ in the dual quantum theory, \textit{i.\,e.},
\begin{align}
A_t(t, \bm{x}, z)=\mu+\mathcal{O}(z)\,,
\end{align}
where $\mu$ is independent of the superfluid's coordinates $(t, \bm{x})$.
The only parameter of the holographic superfluid, uniquely characterizing it, is the dimensionless ratio $\mu/T\sim (\mu z_\text{h})$ (\textit{cf.}\ \eqref{temp}).
One finds that the scalar field $\Phi$ forms a charge condensate above a critical value $(\mu z_\text{h})_\text{c}=4.06371$ (see \textit{e.\,g.}\ \cite{Herzog:2008he}) which sets the system in the superfluid phase.
It is then straightforward to relate $\mu/T\sim (\mu z_\text{h})$ to the temperature ratio $T/T_\text{c}$, where $T_\text{c}$ denotes the critical temperature,
\begin{equation}\label{TvsMu}
\frac{T}{T_\text{c}}=\frac{(\mu z_\text{h})_\text{c}}{\mu z_\text{h}}\,.
\end{equation}
We fix our units by setting $z_\text{h}=1$.
This also fixes the superfluid's temperature to $T=\frac{3}{4\pi}$ (see \eqref{temp}).
Therefore, $\mu$ is the only free parameter that controls the phase transition and hence the temperature ratio $T/T_\text{c}$ according to \eqref{TvsMu}.
Henceforth, we will characterize and discuss the superfluid in terms of its temperature ratio $T/T_\text{c}$ as this appears to be the more intuitive parameter. But we keep in mind that what we actually vary is the chemical potential.
According to the holographic dictionary, the scalar expectation value $\psi = \langle\Psi\rangle$ is
identified as the next-to-leading order term in the near-boundary expansion of the scalar field $\Phi$,
\begin{align}\label{bcPhi}
\Phi(t, \bm{x}, z)=\eta(t, \bm{x})z + \psi(t, \bm{x}) z^2 +\mathcal{O}(z^3)\,,
\end{align}
where $\eta(t, \bm{x})$ constitutes a source for the scalar field which has to be set
to zero, \textit{i.\,e.}\ $\eta\equiv 0$, to not break the $U(1)$-symmetry explicitly.
This condition can be expressed as
\begin{align}
\partial_z\Phi(t, \bm{x},z)|_{z=0}=0\,.
\end{align}
Similarly, all sources for the spatial gauge field components, which in the holographic interpretation
correspond to an external superfluid flow, are switched off, \textit{i.\,e.}\ $A_x(t, \bm{x},z=0)=A_y(t, \bm{x},z=0)=0$.
At the black-hole horizon, $z=z_\text{h}$, the gauge field components must satisfy regularity
conditions, given by $A_t(t, \bm{x},z_\text{h})=A_x(t, \bm{x},z_\text{h})=A_y(t, \bm{x},z_\text{h})=0$.
Similarly, the scalar field $\Phi$ must behave regularly at the horizon which is
intrinsically implemented by the choice of infalling Eddington--Finkelstein coordinates.
More details on the equations of motion and the boundary conditions are given in appendix \ref{eom}.
The scalar order parameter $\psi$ is a complex field whose
absolute value squared gives the density of the superfluid,
\begin{equation}
n(t, \bm{x})=|\psi(t, \bm{x})|^2.
\end{equation}
We denote the homogeneous background density of the superfluid in thermal equilibrium, where no vortices are present, as $n_0$.
The phase $\varphi(t, \bm{x})=\text{arg}(\psi(t, \bm{x}))$ of the order parameter encodes the superfluid velocity
via $\bm{v}^\text{s}(\bm{x},t) = \vec{\nabla}\,\varphi(\bm{x},t)$, where $\vec{\nabla} = (\partial_x,\partial_y)$.
In the probe approximation the black hole constitutes an infinitely sized static heat bath for the superfluid.
Due to its presence, the superfluid is equipped with an energy and momentum dissipation mechanism caused by modes falling into the black hole.
In this sense, the black hole can loosely be interpreted as a normal component connected to the superfluid condensate in the sense of Tisza's two-fluid model \cite{Tisza1938TPiHII}.
In the case with back-reaction it has been shown that to non-dissipative
order\footnote{This refers to the hydrodynamic expansion at the boundary with only non-dissipative terms included.
We stress that the near-boundary expansion \eqref{bcPhi} is not a hydrodynamic expansion in any sense.
Instead, it yields the order parameter field of the superfluid condensate that captures the physics at all scales.}
the holographic superfluid is indeed governed by a relativistic version of Tisza's two-fluid model \cite{Sonner:2010yx}.
It appears likely that a similar identification holds, at least approximately, also at dissipative order, \textit{cf.}\ \cite{Sonner:2010yx}.
Also for the non-back-reacted system an analogy to the two-fluid model could be expected, but a rigorous
identification has not been established so far. Keeping this caveat in mind, it can be useful to view the results of
the following sections in light of the two-fluid picture.
\subsection{Remarks on the Numerical Implementation}
\label{Subsec:NumImp}
We are interested in performing a real-time simulation of the dynamics of the holographic superfluid.
The coupled system of equations of motion (see appendix \ref{eom}) cannot be solved analytically and hence must be dealt with numerically.
The superfluid is described by the order parameter field $\psi$
which can be extracted from the solution of the equations of motion according to the near-boundary expansion \eqref{bcPhi}.
For $T/T_\text{c}<1$ the system is in the superfluid phase, \textit{i.\,e.}\ the global $U(1)$ symmetry is spontaneously broken.
In section \ref{Sec:VortAntiVort} we will keep $T/T_\text{c}$ fixed at $T/T_\text{c}=0.68$ (corresponding to $\mu=6$)
and vary only the initial condition of our physical setup, that is the initial size of the vortex dipole.
In section \ref{Sec:MuDependence}, on the other hand, we will keep the initial condition fixed and only vary the temperature, $T/T_\text{c}$.
All our numerical simulations are performed on a grid of $512\times 512$ points in the $(x, y)$-plane, with periodic boundary conditions imposed.
For a given $T/T_\text{c}$ we choose the grid constant, $a$, such that an isolated vortex is resolved by $13$ grid points in diameter at $95\%$ of the background density $n_0$.
For $T/T_\text{c}=0.68$ this yields $a=1/3.5$. Identical grid constants are used in $x$- and $y$-direction
such that an isolated vortex is resolved in a rotationally symmetric fashion.
In section \ref{Sec:MuDependence}, where we discuss the dependence of our physical system on the temperature $T/T_\text{c}$, we give a table with all temperatures that we consider and the corresponding grid constants.
For the holographic $z$-direction we use $32$ collocation points and accordingly a basis of $32$ Chebyshev polynomials.
The system is propagated in time by an explicit and adaptive Runge-Kutta time-stepping scheme.
In the following sections we often show vortex positions etc.\ at unit timesteps.
One unit timestep (also: unit of time) is composed of an adaptive number of much smaller timesteps $\tau\ll 1$
that we use in the solver for the differential equations.
Further details on the numerical implementation are given in appendix \ref{App:Numerics}.
\section{Dynamics of Holographic Vortex Dipoles}
\label{Sec:VortAntiVort}
In this section we discuss the dynamics and the annihilation process of a single vortex--anti-vortex pair, or vortex dipole.
Our main interest concerns the vortex trajectories and the vortices' local velocities and accelerations.
We introduce a high-precision tracking algorithm for the position of the vortices that allows us to extract the trajectories in a quasi-continuous manner.
It is this high precision that then enables us to accurately determine the velocity and acceleration fields of the vortices.
Knowing the acceleration of the vortices we can in turn infer their (effective) interaction if we assume their (unknown) mass to be constant.
Throughout this section the temperature of the superfluid will be fixed to $T/T_\text{c}=0.68$.
\subsection{Initial Conditions}
\label{Sec:InitialCond}
We study a system composed of only one vortex and one anti-vortex, separated by an initial distance of $u_0$ grid points.
This is the most basic non-trivial configuration and thus provides a clean environment for studying vortex dynamics.
One possible source of numerical uncertainty originates from the periodicity of the grid and thus of the initial configuration.
To minimize it, we restrict ourselves to sufficiently small initial dipole sizes. In other words, we initialize our vortex dipole such
that the `inner' distance, \textit{i.\,e.}\ the $u_0$ grid points in the initial condition, is much shorter than the `outer'
distance, \textit{i.\,e.}\ the distance from vortex to anti-vortex obtained by crossing the boundary to the next periodic cell
of the lattice. (We always assume the dipole to be placed close to the center of a cell of the periodic lattice.)
With regard to this problem, we can access somewhat larger initial distances $u_0$ by placing the two vortices
on the diagonal of the quadratic grid.
We perform all numerical simulations presented in this work on a grid of $512\times512$ points in the $(x,y)$-directions.
We have checked that our results are indeed independent of the grid size by performing one simulation on a $1024\times1024$
grid which gave identical results. We conclude that for all initial dipole sizes $u_0$ considered here the observed dynamics is not affected
by periodicity effects. When we align the vortices along the diagonal it is in general not possible to choose $u_0$ to be an integer.
For simplicity, we then quote the closest integer as the value of $u_0$.
A vortex is a topological defect of the superfluid and has an intrinsic structure solely determined by the superfluid itself.
The main property of a vortex is its phase winding $w$ around the vortex position at which the superfluid density vanishes.
The quantized phase winding is given by a contour integral encircling the vortex position,
$w=\frac{1}{2\pi}\oint\,\text{d}\varphi$ for $\varphi=\text{arg}(\psi)$,
and is by definition positive for vortices and negative for anti-vortices.
Vortices can be built into the system by multiplying the complex scalar expectation value $\psi$ by a localized
phase, $\psi \to \psi \cdot e^{\text{i} w\phi_\text{v}}$, where $\phi_\text{v}$ is
the appropriate phase angle in the $(x, y)$-plane. In our holographic system
we multiply the complex scalar bulk field $\Phi$ by a localized
phase $\Phi(t, \bm{x}, z) \to \Phi(t, \bm{x}, z)\cdot e^{\text{i} w\phi_\text{v}}$ at every $z$-slice along the holographic direction.
At the position of the vortex we in addition set the scalar field $\Phi$ to zero, again for all $z$-values.
For the boundary superfluid this corresponds to a zero in the condensate density at the vortex position.
As the system is evolved in time, the topological defect imprinted in this way fully builds up within a few unit timesteps
and becomes a vortex of characteristic size and shape.\footnote{The flow field for a vortex--anti-vortex pair
takes significantly longer to become fully developed, although a good approximation is reached within a few unit timesteps.
This will be discussed in section \ref{phasehealing} below.}
In figure \ref{Dens_Phase} we show one of the initial phase and density configurations (after
it has fully built up at $t=10$) used in this work.
\begin{figure}[t]
\begin{center}
\includegraphics{Dens_Phase.pdf}
\caption{Initial condition for a typical vortex-dipole configuration used in this work. Left: Superfluid density with one
vortex and one anti-vortex aligned along the diagonal of the quadratic grid with initial separation of $u_0=80$ grid points.
The upper-right zero in the density is the vortex, the lower-left the anti-vortex.
The superfluid density is scaled by $n_0$, the background condensate density without any defects.
Right: The corresponding phase configuration of the system.
The color coding is chosen to be periodic to capture the nature of the periodic phase.
The snapshots of the density and
the phase are taken after 10 units of time, when the imprinted vortices are fully developed.
\label{Dens_Phase}}
\end{center}
\end{figure}
The vortex--anti-vortex pair is aligned along the diagonal
of the grid, initially separated by a distance $u_0$ of 80 grid points. The vortices actually start
moving immediately after they have been imprinted due to the phase configuration of the respective other vortex.
When the vortex shape has fully built up the position of the vortices has therefore already changed and does no longer
exactly correspond to the initial distance.
The plot on the left shows the superfluid density $n=|\psi|^2$ scaled to the background
condensate density $n_0$ (the equilibrium density of the superfluid without any topological defects). The plot on the right shows the
phase of the order parameter field, $\varphi=\text{arg}(\psi)$. Both plots comprise the full grid.
The initial conditions we study in this section are all similar to the one shown in figure \ref{Dens_Phase} and
only differ in the chosen initial distance $u_0$ and the orientation of the dipole.
\subsection{Vortex Dynamics}
\label{Sec:Dynamics}
As the system evolves in time the vortices move on characteristic trajectories, approaching each other while simultaneously
performing a motion in the perpendicular direction.
We denote the motion along the dipole axis as longitudinal motion, and the one perpendicular to it as transverse motion.
In figure \ref{Snapshots_4} we show four snapshots of the superfluid density during the dipole evolution
for an exemplary initial separation of $u_0=80$ grid points
at different times, from upper left to lower right with proceeding time.
It takes $t=1551$ unit timesteps for the vortices to annihilate for the chosen initial condition.
\begin{figure}[t]
\centering
\includegraphics[trim={0 0 0 1.1cm}]{Vortex_Gauss_4times.pdf}
\caption{Snapshots of the superfluid density during the evolution of a dipole of initial size of $u_0=80$ grid points,
taken at four different times (a)\;$t=400$, (b)\;$t=1530$, (c)\;$t=1546$, and (d)\;$t=1558$.
The initially circular vortices deform as they approach and finally annihilate each other.
At time $t=1558$ in plot (d) the vortices have already annihilated, producing a soliton-like excitation in the system.
This excitation is quickly damped to sound waves and eventually dissipated into the thermal bath of the system.
\label{Snapshots_4} }
\end{figure}
The first snapshot (a) is taken at time $t=400$, \textit{i.\,e.}\ early in the evolution when the vortices are far apart and still perfectly spherical.
At this time the width of the vortices is much smaller than the vortex separation and the vortices can thus be interpreted as point-like defects in the superfluid density.
We will come back to this in section \ref{hvisection} where we compare the trajectories to equations of motion in the point-vortex approximation.
Snapshot (b) is taken at time $t=1530$, already shortly before the annihilation.
At this stage, the superfluid depletions of the vortices start to deform non-elliptically and subsequently merge.
This merging process begins only at very late time, here $\sim 25$ timesteps before the final annihilation, and proceeds very quickly.
The transition from non-overlapping to overlapping density depletions is of course not sharp, but one can
still reasonably well estimate the time when it starts.
The transition will be relevant for the precise tracking of the vortices, see section \ref{Sec:Tracking} below.
In the following, we will refer to the last $25$ timesteps as late times.
In snapshot (c), four unit timesteps before the annihilation, the vortices have merged into a peanut-shaped
system with strongly overlapping density depletions.
At this time, the dipole accelerates most strongly.
After the annihilation, we observe a soliton-like excitation propagating
along the transverse direction of the vortex-dipole. This is shown in snapshot (d). The soliton-like excitation
transitions into sound waves.
Each annihilation event of a vortex--anti-vortex system therefore releases energy
into the system in the form of sound waves.
The sound waves eventually dissipate completely, leaving the system in a homogeneous
equilibrium state.
\subsection{Tracking Algorithm}
\label{Sec:Tracking}
So far, in all real-time simulations of holographic superfluids vortices were tracked by locating their
phase windings on the numerical grid, see \textit{e.\,g.}\ \cite{Adams:2012pj,Ewerz:2014tua,Lan:2018llf,Xia:2019eje,Du:2014lwa,Chesler:2014gya,Li:2019swh,Yang:2019ibe,Lan:2016cgl}.
The same holds for many studies of superfluids using the Gross--Pitaevskii equation (GPE)\footnote{The Gross--Pitaevskii equation \cite{Gross1961a,Pitaevskii1961a} or nonlinear Schr\"odinger equation is a mean-field description for the classical order parameter field $\psi$ of dilute ultra-cold Bose gases in the ground state, \textit{i.\,e.}\ Bose-Einstein condensates.}, see for instance \cite{Numasato2010,Karl:2016wko,Shanquan2020}.
This procedure yields the vortex positions with plaquette resolution which is sufficient for many applications.
We find, however, that to study and resolve the full dynamics of a vortex dipole a sub-grid-spacing resolution is required.
Only then can velocities and accelerations be accurately computed.
There exists a number of tracking methods in the literature, see \textit{e.\,g.}\ \cite{Nazarenko_2006,PhysRevE.78.026601,PhysRevE.86.055301,Zuccher_2012,Taylor_2014,rorai2014approach} for applications of different methods, each with different advantages and drawbacks.
For the purpose of this work we propose a new tracking algorithm for vortices on a two-dimensional grid
and combine it with an existing Newton--Raphson (NR) tracking method, see for example \cite{GALANTAI200025}.
This ensures that at each stage of the evolution the vortices are most accurately and numerically efficiently tracked.
\subsubsection{Gaussian Fitting Routine}
In the following, we introduce and discuss a new tracking routine that has the advantage of being
numerically very efficient and applicable for most of the evolution of the vortex dipole.
At a given time $t$ of the evolution we consider the condensate density $n(t, \bm{x})=|\psi(t, \bm{x})|^2$ in a subregion\footnote{The results are independent of the chosen subregion as long as it is large enough to fully enclose both, the vortex and the anti-vortex. In practice we choose a rectangle spanned by the vortex dipole plus $40$ grid points in the positive and negative $x$- and $y$-directions.} that encloses both vortices.
We then use a fitting routine that locates the minima of the density depletions, \textit{i.\,e.}\ the vortex positions, based on an interpolation
between grid points. This permits a determination of vortex positions with sub-plaquette precision.
As fitting function we use a linear combination of two two-dimensional upside-down Gaussians,
\begin{align}
G(x, y)=A&-B\exp\left\{-\frac{(x-x_{1})^2}{2\sigma_{x,1}^2}-\frac{(y-y_{1})^2}{2\sigma_{y,1}^2}\right\}\\
&-C\exp\left\{-\frac{(x-x_{2})^2}{2\sigma_{x,2}^2}-\frac{(y-y_{2})^2}{2\sigma_{y,2}^2}\right\}\,,
\end{align}
where $(x_{1},\,y_{1})$ and $(x_{2},\,y_{2})$ are the $(x,\,y)$-positions of the vortex and anti-vortex, respectively.
The positions are the fitting parameters we are mainly interested in.
The widths $\sigma_{x/y,i/j}$ of the Gaussians in $x$- and $y$-direction are also separate fitting parameters to allow for non-spherical deformations of the vortices.
Finally, $A, B$ and $C$ are additional parameters in the fit, accounting for the background density and a scaling of the Gaussians.
For the fitting routine we use a Levenberg--Marquardt least-squares fitting algorithm.
The reason that this fitting routine accurately tracks the vortex positions is the following.
Essentially, the method relies on the fact that an extremum of a function of simple and approximately known shape can be found with high precision
even when its values are known only on a much coarser discrete lattice.
Vortices are characterized by their quantized phase windings which implies that the superfluid density vanishes at their exact location.
Consequently, to track the vortices one has to find these minima.
For spherical or elliptically deformed vortices the position of the minimum of the density depletions is fully encoded in the flanks.
Thus, one only has to accurately describe the flanks (in our method by Gaussians) and interpolate to the minimum.
The minimum of the Gaussian then coincides with the vortex location.
A Gaussian turns out to be a very good functional approximation in the region of the flanks and can thus be
used to determine the vortex positions.
We point out that in order to most accurately describe the flanks, the Gaussian fit should not be
forced to reach zero density since it does not capture the full vortex shape.
Notably, any other function that describes the flanks of the vortices similarly well could also be used.
We have checked this explicitly and found very good agreement between the results.
For instance, another choice could be the approximate shape \cite{Schakel2008,Pethick2006} of a vortex from
the Gross--Pitaevskii equation, given by
\begin{align}\label{GPEVortex}
n(r) = \frac{n_0 r^2}{2 \xi^2 + r^2}\,,
\end{align}
where $n(r)$ is the
density profile with $r$ denoting the radial distance from the vortex center,
$n_0$ the background density, and $\xi$ a free parameter determining the width of the vortex.
Using this function in the fitting routine yields a discrepancy with the Gaussian fit of approximately $\pm 0.02$ grid points in the vortex positions.
We estimate that the error of the vortex position resulting from the Gaussian fits is of the same order of magnitude.
This can be reasoned by testing how well a GPE vortex (\textit{cf.}\ \eqref{GPEVortex}) captures the position of a grid-discretized Gaussian (of typical parameters
found in the fits to our vortices) and vice versa.
A simple analysis show that the discrepancies are also of the order of a $10^{-2}$ grid points.
Furthermore, a comparison with known tracking methods in the literature yields the same estimate.
In particular, we compared the results with the NR method discussed below and found excellent agreement.
In all plots shown in this work the error in the vortex positions is much smaller than the size of the plot markers.
Figure \ref{GaussFits} shows the cross section of a two-dimensional Gaussian fit to a vortex--anti-vortex system.
\begin{figure}[t]
\centering
\includegraphics{Vortex_Gauss_1D_2D.pdf}
\caption{Graphical illustration of the Gaussian fitting routine for the localization of the vortices.
The plot on the left shows a cross section of the superfluid density (blue dots) along the $x$-grid for fixed $y$ and a
Gaussian fit of the data (solid red line).
The plot on the right shows the superfluid density in the vicinity of a vortex with contour
lines of the two-dimensional Gaussian fit in grayscale.
\label{GaussFits}}
\end{figure}
Instead of showing the full dipole system we representatively focus on just a single vortex but stress that we in fact did a combined fit to both vortex and anti-vortex as described above.
The left plot in figure \ref{GaussFits} shows the cross section of the two-dimensional Gaussian fit to the
superfluid density $n(x,y)$ along the $x$-direction for fixed $y$, with the $y$-position taken as the closest integer $y$-grid point to the position from the fit.
Crucially, the Gaussian fit does not reach zero but captures the vortex flanks to high precision which is sufficient to find the position of the minimum.
On the right in figure \ref{GaussFits} a contour plot of the same fitted vortex profile is shown.
The contour lines of the fit are displayed in grayscale while for the condensate density $n$ we use the same color coding as in figure \ref{Dens_Phase}.
This Gaussian fitting algorithm can be used to track the vortices as soon as their shapes have fully developed after initialization, that is after $5$ unit timesteps.
Only for the last $25$ unit timesteps, when the vortices merge and deform non-elliptically, this method breaks down and thus fails to track the vortices as the Gaussians can non longer accurately capture the vortex flanks.
For this regime we instead rely on a well-known Newton--Raphson tracking algorithm which we discuss next.
\subsubsection{Newton--Raphson Tracking Algorithm}
A reliable way to track the vortices also in the late-time regime when the density depletions of the vortices overlap is given by the NR method.
As a standard algorithm for numerical root finding, the NR method is optimally suited for vortex tracking on two-dimensional grids, see for example \cite{PhysRevE.86.055301,PhysRevE.78.026601,Villois2016}.
In the following we give a brief overview of the method and discuss its advantages and drawbacks as compared to the Gaussian fitting routine.
As discussed above, vortices are characterized by $\psi(\bm{x}_i)=0$, with $\bm{x}_i$ the position of the $i$th vortex, which is a direct consequence of the phase winding.
The NR method can therefore be used to solve $\psi(\bm{x}_i)=0$ directly, and no assumptions regarding the shape of the density depletion are needed.
To ensure that a zero at $\bm{x}_i$ indeed signals a vortex one subsequently has to check that there is a corresponding phase winding at this position.
To solve $\psi(\bm{x}_i)=0$ for $\bm{x}_i$ using the NR method, one performs a Taylor expansion of $\psi(\bm{x})$ around an initial guess $\bm{x}_i^g$ for the vortex position,
\begin{align}\label{NRTaylor}
0=\psi(\bm{x}_i)=\psi(\bm{x}_i^g) + J(\bm{x}_i^g)(\bm{x}_i - \bm{x}_i^g) + \mathcal{O}\left[(\bm{x}_i - \bm{x}_i^g)^2\right]\,,
\end{align}
with the Jacobian matrix
\begin{align}
J(\bm{x})=\begin{pmatrix}
\operatorname{Re}[ \partial_x \psi(\bm{x})] & \operatorname{Re}[ \partial_y \psi(\bm{x})] \\
\operatorname{Im}[ \partial_x \psi(\bm{x})] &\operatorname{Im}[ \partial_y \psi(\bm{x})] \\
\end{pmatrix} \,.
\end{align}
If $J(\bm{x})$ is invertible, \textit{i.\,e.}\ has full rank, \eqref{NRTaylor} can be solved for $\bm{x}_i$,
\begin{align}\label{NRiteration}
\bm{x}_i = \bm{x}_i^g - J^{-1}(\bm{x}_i^g)\,\psi(\bm{x}_i^g) + \mathcal{O}\left[(\bm{x}_i - \bm{x}_i^g)^2\right]\,.
\end{align}
This procedure can then be iterated such that $\bm{x}_i$ approaches the exact vortex position to sub-grid-spacing precision.
The precision is mainly limited by the evaluation of the Jacobian at intermesh points.
This is done by Fourier interpolating $\psi(\bm{x})$ which also guarantees that the periodicity of the order parameter field is retained throughout the iteration procedure.
As initial guess for the algorithm we locate the phase winding of the vortices on the numerical grid. This plaquette precision is sufficient to ensure convergence of the NR algorithm.
We stop the iteration when the condition $\psi(\bm{x_i}) < 10^{-10}n_0$ is satisfied.
The advantage of this algorithm as compared to the Gaussian fitting routine is that it directly tracks the zeros in the superfluid density.
The Gaussian fitting routine, on the other hand, determines the minima of the density depletions only via their flanks.
That is very precise as long as the vortices deform only mildly and elliptically but fails as soon as this is no longer the case.
For the vortex dynamics under investigation this happens as part of the final annihilation process where the vortices deform strongly.
For all earlier times, when the vortices are spherically symmetric or only slightly elliptically deformed, we checked that the two methods agree to a very high precision with a maximum absolute deviation of $\pm 0.02$ grid points.
This error agrees with the estimate made above for the Gaussian fitting routine.
The key advantages of the Gaussian fitting routine are its simplicity and its numerical efficiency which make it much faster than the NR method.
Therefore, we use the Gaussian fitting algorithm to track the vortices for all vortex--anti-vortex separations $d(t)$ larger than $25$ grid points and the NR method for all smaller separations.
As stated above, the two methods yield the same results for the first stage, \textit{i.\,e.}\ for $d(t)>25$.
We show a comparison of the trajectories obtained with the two methods in figure \ref{GaussVsNR}.
\begin{figure}[t]
\begin{center}
\includegraphics{Trajectories_Gauss_NR.pdf}
\caption{Comparison of the vortex trajectories obtained with the two different tracking algorithms,
for an initial separation of $u_0=80$ grid points.
For most of the evolution, both methods give the same results with a discrepancy smaller than the marker size.
At late times the Newton--Raphson method is superior to the Gaussian fitting routine as it relies only on the zero in the superfluid density and not on the vortex flanks. Due to its numerical efficiency, the Gaussian fitting routine is superior for early and intermediate times, and hence for the larger part of the evolution.
\label{GaussVsNR}}
\end{center}
\end{figure}
The difference in the vortex position at large separations is smaller than the marker size.
For small separations, on the other hand, one clearly sees that the Gaussian fitting routine can no longer accurately track the vortices.
The trajectories obtained with this method cannot resolve the final annihilation process and have a systematic error,
as is evident from the almost parallel final pieces of the trajectories which would, if real, not lead to an annihilation of the vortices.
The NR method, in contrast, provides the accurate vortex positions until immediately before the annihilation.
\subsubsection{Sub-plaquette Vortex Tracking}
Having discussed the tracking algorithms, we now focus on the gain of these methods as compared to the standard plaquette resolution obtained from locating the phase windings on the grid.
The position found by locating the phase winding is inherently plagued with an uncertainty of one grid point
as the vortex position is always assigned to integer grid points.
Consequently the trajectories determined by this method take a stair-like shape, with
the vortex position being allocated to one grid point for a certain number of timesteps before jumping to the next one.
The tracking algorithms described above, on the other hand, allow to extract quasi-continuous trajectories (only limited by finite time-stepping resolution).
In figure \ref{Trajectories} we compare the vortex trajectories determined by our (combined) tracking method to the trajectories
determined by using only the phase windings to locate the vortices.
\begin{figure}[t]
\centering
\includegraphics{Trajectories.pdf}
\caption{Trajectories of the vortex-dipole with an initial separation of $u_0=80$ grid points
computed by only localizing the phase windings of the superfluid phase (red dots)
and with sub-plaquette resolution using our localization method (blue dots).
The phase-winding localization is only as precise as one grid point and the trajectory is therefore discrete or stair-like,
while our localization method results in quasi-continuous trajectories.
\label{Trajectories} }
\end{figure}
Clearly, the gain from the combined use of Gaussian fitting routine and NR method has a strong impact on the extracted trajectories.
We note that there is a systematic contribution to the deviation between the two methods coming from the fact that, when locating the vortices via their phase windings on a plaquette, we always attribute the position to the upper left corner of that plaquette. Any other choice would imply a similar systematic deviation that would be stronger in other parts of a given trajectory.
Let us give an empirical estimate of the difference in precision between the two methods for our specific vortex-dipole system.
At early and intermediate times (to be discussed in more detail below) it takes roughly $50$ unit timesteps for the vortex to move as far as one grid point.
Locating the vortex solely by its phase winding would thus give the same grid point for all of these $50$ unit timesteps.
By using our fitting routine, on the other hand, we can clearly locate the vortex at $50$ distinct positions.
Therefore, our method yields quasi-continuous trajectories for the vortices.
As compared to the locations obtained from the phase windings this is a strong increase in precision.
Obviously, the gain in precision depends on the velocity of the vortices and, in relative terms, is smaller for fast vortex motion.
We will see that the increase in precision due to sub-plaquette tracking
is essential for obtaining some of the main results in our analysis of the vortex dipole.
For obtaining the data shown in figure \ref{Trajectories}, as well as for all other results presented in this work,
we choose $t=5$ as the starting time for our fitting procedure.\footnote{Note, however, that in figure \ref{Trajectories} the
discrete trajectories of plaquette-resolution are shown starting at $t=0$.}
The trajectories are shown up to the point where the vortices annihilate.
Since the vortices accelerate strongly just before annihilation, we increase the frequency of tracking during the last 8 unit timesteps
and determine the position 100 times per unit timestep here.
\subsection{Trajectories of the Dipole System}
\label{DynamicsDipole}
Let us now investigate in detail the dynamics of a vortex--anti-vortex pair.
We first present and discuss the trajectories of the vortex
dipole, here for initial separations along the horizontal of the quadratic grid.
The results for five different initial separations are shown in figure \ref{VortPosBeide}.
\begin{figure}[t]
\begin{center}
\includegraphics{VortPosBeide.pdf}
\caption{Trajectories of the vortex--anti-vortex dipole system with five different initial separations.
The vortices are imprinted with initial separation $u=u_0$ along the horizontal of the quadratic grid.
(For details of the initialization see section \ref{Sec:InitialCond}.)
Here we show only the part of the grid containing the trajectories.
We observe that, after an initial phase healing process (see section \ref{phasehealing}),
the trajectories become universal for all different initial separations.
\label{VortPosBeide}}
\end{center}
\end{figure}
Note that for conciseness the plot in figure \ref{VortPosBeide} does not
show the full numerical grid but only the part containing the actual vortex trajectories.
We observe that the trajectories of the vortices at intermediate and late times are independent of the initial separation and follow universal curves.
That is, the short-distance behavior is independent of the earlier evolution of the system.
We checked this explicitly by shifting the trajectories such that the respective annihilations take place at the same point on the grid.
Figure \ref{VortPosBeide} shows the un-shifted vortex trajectories.
We refer to all times from which on the trajectories are universal, but prior to the last $25$ unit timesteps, as intermediate times.
We stress that the distinction between early, intermediate and late times in this sense is only approximate
and does not split the evolution into three strictly distinguishable phases.
The universality of the trajectories at intermediate and late times is a remarkable
feature of the system as it shows that the physics of a vortex--anti-vortex pair in the superfluid is
entirely governed by the size of the dipole and independent of its previous evolution.
In subsection \ref{Velocities} we will discuss this point in more detail in the context of the vortex velocities and accelerations.
\subsubsection{Phase Healing}
\label{phasehealing}
In figure \ref{VortPosBeide} we observe a feature of the trajectories at early times that is, at first sight, very surprising
and gives rise to the non-universality of the trajectories in this regime. Both the vortex and the anti-vortex
show a slight outward bending in their motion before they approach each other. The effect is discernible only when
the vortices are tracked with sub-plaquette resolution. This is obvious in figure \ref{Trajectories} where
the quasi-continuous trajectories (blue dots) exhibit a clear outward bending while the discrete trajectories (red dots)
limited to plaquette resolution do not show any sign of outward bending.
In figure \ref{VortPosBeide}, the outward bending is seen in the early-time evolution for all initial separations.
It turns out that the effect is not physical but is an artifact of the method used for imprinting the initial configuration
of vortices in the superfluid order parameter field.
As discussed above, we initialize two single vortices by simply imprinting and linearly superimposing their individual phase
configurations on the entire grid.
But this does not take into account that a proper vortex--anti-vortex pair is actually one coupled system.
Imprinting such a pair would require us to imprint the coupled phase configuration instead of the linear superposition of
two separate vortices. However, an analytic formula for the correct phase configuration of a proper vortex dipole is not
known. Imprinting the superposition of the phase fields of two separate vortices is therefore the best practical procedure.
Fortunately, during the numerical evolution of the system the phase configuration `heals itself' and transitions into that of a proper vortex dipole.
This `phase healing' process is much slower than the build-up process of the vortex shapes, as can be expected because it
is necessarily a non-local process. After the phase healing the dynamics of the dipole becomes independent
of its initial configuration and is then solely determined by the distance between vortex and anti-vortex.
We refer to `intermediate times' of the evolution for times after the phase healing is completed, while we call
times during the phase healing process `early times'. Equivalently, intermediate times refer to the universal regime of the
evolution which is independent of the initial separation of the vortex--anti-vortex pair. We find that
the phase healing process is mildly dependent on the initial condition, \textit{i.\,e.}\ the initial separation of the vortices, \textit{cf.}\ figure \ref{VortPosBeide}.
However, the effect of the phase healing on the trajectories is rather subtle and fades out slowly.
It is therefore difficult to determine the exact time of the transition from the phase-healing regime to the universal regime.
The analysis in section \ref{Velocities} will show that the vortex velocities permit a reasonably accurate estimate.
For an initial separation of $u_0=80$ grid points we find that it takes about $\Delta t \sim 350$ unit timesteps
for the phase configuration to approach that of a proper vortex dipole.
We can understand how the phase healing leads to a slightly outward pointing motion
by comparing the phase configurations of two dipoles of identical size, with one of them being in its initial phase configuration
and the other one being in a phase configuration after the healing process. The latter can be obtained by evolving a larger dipole
through the phase healing.
The difference between the two corresponding phase configurations gives the initially `missing configuration'
(as compared to the fully coupled dipole configuration), and we can derive the superfluid flow resulting from it.
To visualize this effect, we initialize a vortex pair as a superposition of the individual phase fields at large distance
and propagate it in time for long enough so that its phase field has fully healed. From the corresponding
phase configuration we subtract the phase configuration we would have imprinted for a new pair of vortices
at this very separation. This gives us the missing phase configuration of the newly initialized dipole
as compared to a fully coupled dipole. Its negative gradient
is the superfluid flow that acts on the newly imprinted vortices at this separation due to their missing phase configuration.
We show the streamlines of this flow field superimposed on the superfluid density of the dipole
configuration in figure \ref{DenAndStream_Remnant}.
\begin{figure}[t]
\begin{center}
\includegraphics{DenAndStream_Remnant.pdf}
\caption{Density plot of a vortex configuration after the phase healing with remnant streamlines.
The streamlines show the negative gradient of the difference of the phase field of a fully coupled (healed) dipole
and of the phase configuration used for initializing a vortex--anti-vortex system of the same separation.
The phase of the evolved dipole, originally imprinted at a separation of $u_0=80$ grid points, is taken at $t=350$.
The outward pointing streamlines at the position of the vortices explain the initial
slight outward bending of the vortex trajectories (see figure \ref{VortPosBeide}).
\label{DenAndStream_Remnant}}
\end{center}
\end{figure}
Here we have chosen the dipole orientation along the diagonal of the grid so that
the vortex dipole motion is pointed towards the lower-right corner of the plot.
We can clearly see the outward pointing streamlines that lead to the observed outward bending of the trajectories.
Furthermore, we observe streamlines pointing in the direction opposite to the vortex motion.
In subsection \ref{Velocities} we will see that this leads to an initial phase of deceleration in the vortex motion.
Understanding the artifacts of the initialization procedure allows us to separate them from actual dynamics
of the vortex dipole by concentrating on the universal regime of the evolution.
It is worth pointing out that related effects due to coupled phase fields have been observed in
the context of kinks in scalar field theory \cite{Christov:2018ecz,Christov:2018wsa}.
\subsubsection{Comparison to Hall--Vinen--Iordanskii Equations}
\label{hvisection}
There exist well-known equations of motion, first introduced by Hall, Vinen and Iordanskii (HVI) \cite{Hall1956a.PRSLA.238.204,Iordansky1964a.AnnPhys.29.335}, for the dynamics of a number of vortices in the point-vortex approximation.
They can be derived on very general grounds in a hydrodynamical setup, taking into account the interactions of the vortices
with the normal and the condensate components of the fluid.
The forces that drive the vortex dynamics are the Magnus force, orthogonal to the vortex motion, and drag forces.
The equations of motion are derived by setting the sum of all forces in the closed system to zero and solving for the vortex velocities \cite{Ambegaokar1980a,Sonin1997a.PhysRevB.55.485}.
Applied to our setup, there is no velocity of the normal component of the fluid,
and each vortex only feels the velocity field created by the respective other vortex.
The coupled differential equations for the two-dimensional trajectories $\bm{x}_i$, where $i \in \{ 1,2 \}$ labels the vortices, are then given by
\begin{equation}\label{eomVortices}
\frac{\text{d}\bm{x}_i}{\text{d}t}=(1-C)\bm{v}^\text{s}_i -w _i C' \bm{e}_\perp\times \bm{v}^\text{s}_i \,.
\end{equation}
Here, $C$ and $C'$ are phenomenological friction coefficients characterizing the mutual friction between the vortices and the fluid,
$w_i$ denotes the winding number of the $i$th vortex, and $\bm{v}^\text{s}_i$ is the superfluid velocity induced by the respective other vortex.
We further denote by $\bm{e}_\perp$ a unit vector perpendicular to the $(x,y)$-plane in a right-handed coordinate system,
pointing in an auxiliary direction used only for a convenient notation of this equation and not to be confused with the holographic direction.
For a given set of initial conditions $\bm{x}_i(0)=\bm{x}_{0,i}$, equation \eqref{eomVortices} can straightforwardly be solved.
For one vortex $(w_i=1)$ and one anti-vortex $(w_i=-1)$ we give a derivation of these solutions in appendix \ref{VortexTrajCGLE}.
It turns out that the vortices move along straight lines and annihilate under a fixed angle, depending on the parameters of the theory.
As we have seen, vortices in the holographic superfluid have a finite extension and in general cannot be considered as point-like defects.
However, at large vortex--anti-vortex separation as compared to the vortex width, a point-vortex approximation is possible as far as
the effective interaction between the vortices is concerned.
We can therefore compare the analytic results from the HVI equations to the vortex trajectories discussed above.
Figure \ref{VortPosBeide} clearly shows that the vortices do not follow straight lines for most of the evolution.
We note in particular that at early times, when the vortices are furthest apart, the motion is affected by the phase healing
and thus not expected to obey the equations of motion for a proper dipole.
Furthermore, we find that after the phase healing the vortices quickly accelerate
and approach each other, again rendering the point particle approximation invalid.
This is clearly visible in the vortex trajectories as the final annihilation process with overlapping density depletions is
preceded by a strong bending of the trajectories which sets in not long after the phase healing.
Only in between these two regimes a possible straight segment of the trajectories may be found.
It turns out that for most of the initial separations we study here, this segment, if it exists, is too short to be properly identified.
For an initial separation of $u_0=120$, however, we indeed find that the vortices follow straight lines
for a short period of time after the phase healing and prior to the strong bending and ensuing annihilation.
We present more details and a plot illustrating this behavior in appendix \ref{VortexTrajCGLE}.
The phase healing ultimately is an artifact of our method of preparing the initial state.
Thus we find that for sufficiently large separation the fully developed holographic dipole system indeed obeys
the HVI equations of motion.
Comparing in detail the evolution of a vortex dipole in the holographic superfluid,
in the Gross--Pitaevskii framework and in the HVI equations,
using the tracking methods developed in this work, offers even the possibility to determine
the characteristic HVI friction coefficients of the holographic superfluid, see \cite{Wittmer:2020mnm}.
\subsection{Size of the Dipole System}
\label{Dipole}
Next, we study how the vortex--anti-vortex separation behaves as a function of time.
The separation, or size of the vortex dipole, in units of grid points is defined as
\begin{equation}\label{distance}
u(t)=|\bm{x}_1(t) - \bm{x}_2(t)|\,,
\end{equation}
where $\bm{x}_i(t)$ denotes the position of vortex $i \in \{1, 2\}$ on the numerical grid.
In physical units, the distance is given by $d(t)=u(t)a$, where $a$ denotes the grid constant as discussed above.
(We recall that in this section we consider the temperature $T/T_\text{c}=0.68$ for which a grid constant of $a=1/3.5$ is chosen.)
In figure \ref{dist} we show the physical distance as a function of time, again for an initial separation of $u_0=80$ grid points.
\begin{figure}[t]
\begin{center}
\includegraphics{distance.pdf}
\caption{Size of the vortex dipole defined as distance between
the vortex and the anti-vortex as function of time, starting from
an initial separation of $u_0=80$ grid points along the diagonal of the grid.
The size in physical units $d(t)$ is obtained as $d(t)=u(t) a$, where in this example
the grid constant is $a=1/3.5$.
\label{dist}}
\end{center}
\end{figure}
For all other initial separations discussed in this work $d(t)$ has the same functional behavior, as we discuss in
more detail in appendix \ref{DipoleSize}.
We expect the dipole size at early times to be affected by the phase healing process and thus not to reflect
the correct behavior of a proper vortex dipole. Therefore we focus on intermediate and late times in the following.
There we find that $d(t)$ follows a power law for all initial separations.
Specifically, we perform a fit to our data with the ansatz
\begin{align}\label{PowerLaw}
d(t)=A(t_0 - t)^b\,,
\end{align}
where $t_0, A$ and $b$ are free fit parameters, and find that excellent agreement is achieved.
According to \eqref{PowerLaw} one would expect $t_0$
to be the time at which the vortices annihilate. However, during the final stages of the annihilation process, the
behavior of the vortices is dominated by the deforming and merging density depletions and is not adequately
described by the notion of moving vortices in the sense of a formula like \eqref{PowerLaw}.
We therefore exclude the final 5 unit timesteps from the analysis and leave $t_0$ as a free parameter of the fit.
For all initial separations we find for the exponent $b= 0.53 \pm 0.02$.
We present more details of the fit and further discussion of the results in appendix \ref{DipoleSize}.
We compare this result to the behavior expected from the HVI equations \eqref{eomVortices} for point vortices.
As discussed in appendix \ref{VortexTrajCGLE}, the HVI equations predict
the vortex--anti-vortex separation $d(t)$ to follow a power law of the form \eqref{PowerLaw}
with $b=0.5$, see equation \eqref{dPowerLaw}.
This is very close to what we find numerically for the holographic vortex dipole.
However, we note that the holographic vortices in general cannot be considered as point-like,
and deviations from the square root behavior are therefore expected.
Nevertheless, our results show that also for the case of finite-width vortices their spatial separation
obeys a universal power law after the initial phase healing.
The dynamics of a vortex dipole and its annihilation in the holographic superfluid were also
addressed in the recent study \cite{Lan:2018llf}, including in particular a calculation of the
time dependence of the dipole size. In the following we comment on how our results compare to
those of \cite{Lan:2018llf}. We first note that we study exactly the same system at
the same temperature. The methods for the numerical simulation are similar, with Fourier decomposition
in the $(x, y)$-plane and a pseudo-spectral method with an expansion in Chebyshev polynomials
in the holographic $z$-direction. A grid of $241\times 241\times 25$ points in the $(x, y, z)$-directions is
used in \cite{Lan:2018llf} while we use a larger domain of $512 \times 512 \times 32$ points.
A key difference between our work and \cite{Lan:2018llf} is that we use a high-resolution method for tracking
vortices while the analysis of \cite{Lan:2018llf} relies on tracking by phase winding, thus providing only
plaquette resolution.
The authors of \cite{Lan:2018llf} observe that $d(t)$ splits into two regimes each of which is described by a power law,
with a transition at a separation of twice the vortex radius (defined as the radius where $99\%$ of the background density are reached).
At large separations the square root behavior predicted by the HVI equations is found,
while an exponent of $2/5$ is found at small separations. Velocities, accelerations and effective interaction strengths
of the vortices are then derived from these power laws, in contrast to our calculation in which velocities and accelerations
are computed directly.
With improved precision in vortex tracking, we find different results for $d(t)$. For large distances, we find a similar power law,
although with a slightly different exponent. For small distances, we do not observe a second, distinct power law.
We further observe that the point-vortex approximation breaks down at distances much larger than twice the vortex radius.
The likely reason for the different results is the higher precision of our study, but also the larger numerical grid might play a role.
\subsection{Velocities and Accelerations}
\label{Velocities}
The high-precision tracking of the vortices allows us to directly compute their velocities and accelerations.
We can use the quasi-continuous trajectories determined above to compute derivatives via finite differences,
as we describe in appendix \ref{fdm}. Furthermore, we can then determine the dependence of
velocities and accelerations on the dipole size which is interesting with regard to the observed universality
of the dipole motion.
The results of this section again rely heavily on the improved tracking procedure, in particular for the slow motion
of vortices at early and intermediate times. Here, a vortex needs many unit timesteps to traverse a single grid spacing.
Accordingly, the velocity calculated from tracking with only plaquette resolution would be zero for most of the time,
except for occasional jumps when the next grid point is reached. From quasi-continuous trajectories, on the other
hand, we can compute quasi-continuous velocities and accelerations.
\subsubsection{Vortex Velocities}
In this section we consider physical distances (and velocities), \textit{i.\,e.}\ we use $d(t)=a u(t)$,
with $a$ the grid constant and $u(t)$ the dipole size in grid point.
Consequently, velocities are expressed in units of the speed of light.
To ease comparison with the trajectories discussed in the previous sections, we quote the
initial separation of the dipole $u_0$ in units of grid points.
We recall that we can split the dipole motion into longitudinal and transverse parts.
Analogously, we denote the velocity component along the dipole axis as longitudinal velocity $v_\parallel$,
and the component perpendicular to the dipole axis as transverse velocity $v_\perp$.
In figure \ref{vels} we show the two velocity components as function of the
dipole size for all five different initial separations we considered before.
\begin{figure}[t]
\begin{center}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics{LongitudinalVelOverDist.pdf}
\end{subfigure}\vspace{0cm}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics{TransversalVelOverDist.pdf}
\end{subfigure}
\caption{
Longitudinal (upper plot) and transverse (lower plot)
components of the vortex velocities for five different initial separations on double-logarithmic scale.
Shown are the velocities for the anti-vortex. For the vortex the curves
are analogous except for an additional sign in the longitudinal component.
The velocities are given in physical units, \textit{i.\,e.}\ all lengths in grid points are multiplied by the grid constant $a=1/3.5$.
The curves show universality of the velocities at intermediate and late times.
The behavior of the velocities at early times can be attributed to the phase-healing process.
The straight line in the lower plot denotes an inverse-distance behavior for comparison.
Due to strong dissipative effects in the holographic system, deviations
from this behavior are expected.
\label{vels}}
\end{center}
\end{figure}
To wit, in these plots time proceeds from right to left as the dipole shrinks with increasing time.
Figure \ref{vels} shows that after some time the velocities for all different initial separations approach a universal curve.
This holds for both the longitudinal and the transverse component.
The deviations at early times are caused by the initial phase healing discussed above.
At intermediate and late times, the velocities are universal and depend only on the size of the system,
confirming the universality observed in the trajectories in \ref{Trajectories}.
We also find that for most of the evolution the total velocity of the dipole system stays well below
the speed of light, such that the vortex motion can be considered non-relativistic.
This confirms the finding of \cite{Adams:2012pj} that the boundary dynamics of the holographic superfluid is
non-relativistic although the holographic model has an inherent Poincar\'{e} invariance.
So far we have only considered the evolution of vortex dipoles of different initial separations.
To test the universality of the vortex motion at intermediate and late times, we also performed simulations
with a different type of initial condition. For that we imprint the vortex and anti-vortex as before, but in addition
perturb the system with random noise\footnote{We perturb the system with noise by adding a random (real) phase $\theta (\bm{x})$ to the initial phase configuration. $\theta (\bm{x})$ is obtained by populating low-energy Fourier modes with (complex) random Gaussian noise and then transforming back to real space. For more details see \textit{e.\,g.}\ \cite{Ewerz:2014tua}.} at time $t=0$. Due to that, the vortices acquire a different initial velocity both
in absolute value and in direction. As a result, the trajectories at early times are very different from what we observed in figure \ref{VortPosBeide}.
At intermediate and late times, however, they can be shown to agree again after an appropriate rotation of the coordinate system (with an
angle depending on the initial motion due to the respective noise).
Analogously, we find excellent agreement of the velocities and accelerations after the phase healing and after
the initial noise has dissipated. This again confirms the universal dynamics of the vortex-dipole
in which the system loses information about its initial condition.
In the following we will consider again the usual initial dipole configurations without noise.
The phase healing at early times has a marked effect on the velocities.
The transverse component of the velocity even decreases initially before it increases again.
We note that at this stage the velocity is exclusively induced by the initialized phase configuration
which, however, does not match the correct phase field of a coupled dipole. Therefore, the
velocities receive a contribution according to the streamlines shown in figure \ref{DenAndStream_Remnant}.
There, we indeed found a component pointing in the direction opposite to the center-of-mass motion of
the dipole. Once in the universal regime, on the other hand, the velocities exhibit the expected behavior
of a vortex dipole with a purely attractive force due to the strongly dissipative nature of the holographic superfluid
in the presence of vortices.
In a superfluid condensate without dissipation, in contrast, vortices are known to exhibit so-called Helmholtz propagation
in which vortex and anti-vortex move at fixed distance on straight parallel lines.
The velocity of the dipole is then purely dictated by the vortex--anti-vortex distance and follows
a simple inverse-distance law due to angular-momentum conservation in a closed system.
As our system is dissipative, Helmholtz propagation is suppressed and we expect deviations from the inverse-distance law.
To illustrate this, we plot a $1/d$ curve in figure \ref{vels} for the transverse component of the velocity.
\subsubsection{Vortex Accelerations}
Although the velocity for given initial positions of the vortices
already contains the full information about the dipole motion, we still find it
instructive to show the corresponding acceleration, which we again
split it into its longitudinal and transverse components.
Figure \ref{accs} shows the longitudinal component $a_\parallel$ and the modulus of the transverse component $|a_\perp|$ of the acceleration
in physical units.
\begin{figure}[t]
\begin{center}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics{LongitudinalAccOverDist.pdf}
\end{subfigure}\vspace{0cm}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics{TransversalAccOverDist.pdf}
\end{subfigure}
\caption{Modulus of the longitudinal (upper plot) and
transverse (lower plot) components of the vortex and anti-vortex accelerations for five different initial conditions
on double-logarithmic scale.
All curves are displayed in physical units with a numerical grid constant of $a=1/3.5$.
The kink in the transverse component and the initial behavior of the longitudinal
component can be attributed to the phase healing of the system.
\label{accs}}
\end{center}
\end{figure}
The accelerations exhibit the same universal behavior at intermediate and late times after the phase healing.
We note that the dips in $|a_\perp|$ are due to a change in sign of $a_\perp$.
A comment is in order concerning the interpretation of vortex accelerations in terms of force fields.
Force fields usually require an interpretation in terms of test particles to be placed at a given position.
However, such an interpretation is questionable for a quantized vortex as it always has a strong back-reaction on the fluid.
Formally, one could of course compute acceleration vectors and interpret them in terms of a mutual force field acting
on the vortex and the anti-vortex. This would furthermore require the knowledge of an effective mass of a vortex.
The latter problem has been addressed in the literature, mainly in the context of Gross--Pitaevskii theory,
suggesting a definition for the mass of an isolated single vortex, see for instance \cite{Simula_2018}.
Unfortunately, none of these models can be directly applied to our system and the effective vortex mass remains unknown.
Nevertheless, we can infer from the accelerations in figure \ref{accs} that such an effective force depends only on the spatial separation
and is not a central force.
\section{Vortex Dipole at Varying Superfluid Temperature}
\label{Sec:MuDependence}
So far, we have considered vortex dipole dynamics only at a fixed temperature $T/T_\text{c}=0.68$ of the holographic superfluid.
In the following we compare the vortex velocities for a given initial separation while
varying $T/T_\text{c}$. In general, we do not expect the probe approximation of the holographic superfluid to fully capture
all effects of the temperature. However, given that a fully back-reacted computation has not yet been performed, we
find it interesting to see how the dynamics changes with temperature. We keep in mind that the results might
be different in a fully back-reacted setting.
\subsection{General Considerations and Initial Conditions}
We recall from section \ref{Sec:Model} that the holographic superfluid has only one dimensionless parameter $\mu/T$.
Setting units by $z_\text{h}=1$ leaves only $\mu$ as free parameter which fixes the temperature $T/T_\text{c}$.
In figure \ref{cond_2d_T} we show the modulus of the order parameter field in thermal equilibrium, which serves as our background
solution, as function of the temperature $T/T_\text{c}$ or the chemical potential $\mu$.
\begin{figure}[t]
\begin{center}
\includegraphics{cond_2d_T.pdf}
\caption{Order parameter as a function of temperature (lower abscissa) or
chemical potential (upper abscissa). As the temperature $T$
falls below the critical temperature $T_\text{c}$ a condensate
forms, corresponding to a non-vanishing order parameter that
increases with decreasing temperature.
On the chemical potential axis we mark the values of $\mu$ used in this section.
\label{cond_2d_T}}
\end{center}
\end{figure}
On the $\mu$-axis we mark those values that we use in this section.
Changing the temperature of the holographic superfluid implies a change in
the superfluid condensate density (see figure \ref{cond_2d_T}), a change in the time scales of the system,
and a change in the typical size of a vortex. In order to always resolve a vortex by the same number
of grid points we account for the changing vortex size by adjusting the grid constant for each fixed temperature.
This allows us to avoid any potential artifacts due to velocities and accelerations at different temperatures
being determined at different numerical resolution of the vortex structure.
Again, all our numerical simulations are performed on a $512 \times 512$ grid in the $(x, y)$-direction.
We note that choosing different grid constants at different temperatures implies
different physical sizes of the total system.
We choose five different temperatures covering a broad range of the values in which the system is numerically accessible.
Above and below our chosen temperatures the numerical solution turns out to be plagued by instabilities.
One of the temperatures is the value chosen above, $T/T_\text{c}=0.68$ in the center of the accessible temperature range.
In order to compare the dynamics of the system at different temperatures, we use dimensionless units for the quantities
in this section.
To account for the dependence of the vortex size on the temperature $T/T_\text{c}$ we study the vortex
velocities for a given $T/T_\text{c}$ as a function of the dimensionless distance $d(t)/\xi$, where $\xi$
is the vortex width as determined by fitting \eqref{GPEVortex} to the vortex profile for the given temperature.
Choosing instead the same initial separation in grid points would imply that the vortices are well separated for
some temperature while they would already start to overlap for another.
We recall that for given temperature we choose the grid constant $a$ such that an isolated vortex is resolved
by $13$ grid points in diameter at $95\%$ of the background density $n_0$.
For the initial separation we adapt $u_0$ such that $d_0/\xi$ (with $d_0=au_0$) it the same for the different temperatures.
In table \ref{mu_params} we summarize our choices for the temperature, the corresponding chemical potentials,
and the resulting background densities $n_0$ and vortex width $\xi$. The initial separation $d_0$ is chosen
as $39.2 \,\xi$. In addition, we show the corresponding size $L$ of the computational domain in units of $\xi$.
\begin{table}[t]
\begin{center}
\caption{Parameters used for different choices of temperature $T/T_\text{c}$ or chemical potential $\mu$.
The vortex width $\xi$ and the initial dipole size $d_0$ have physical units.
Also given are the background condensate density $n_0$ and the size $L$ of the computational domain, here in
units of $\xi$.
\label{mu_params}}
\begin{tabular}{cccccc}
\toprule
$T/T_\text{c}$&$\mu$ &$n_0$ & $\xi$ & $d_0/\xi$ & $L/\xi$ \\
\midrule
$0.90 $&$4.5$ & $5.5 $ & $1.16$ & $39.2$&$126.5$ \\
$0.80$ &$5.1 $& $16.1$ & $0.85$ & $39.2$ & $171.1$\\
$0.68 $& $6$ & $41.7$ & $0.44$ &$39.2$ & $334.4$\\
$0.54 $& $7.5$ & $120.2$ & $0.32$ &$39.2$ &$457.1$ \\
$0.45 $& $9 $ &$ 265.2$ & $0.24$ & $39.2$ & $606.4$ \\
\bottomrule
\end{tabular}
\label{tab:MuDynamics2D}
\end{center}
\end{table}
\subsection{Velocities of the Vortex Dipole}
With these parameters we evolve the systems in time and track the vortices until they annihilate.
We find that the initial time it takes the vortices to fully build up their shapes after we imprint them
is slightly dependent on the temperature but is always shorter or equal to $\Delta t=12$ unit timesteps
which is well within the regime of the phase healing. As the velocities in this regime are anyway
affected by artifacts of the vortex initialization we can disregard any small differences in the
build-up process. We start the tracking for all chosen temperatures after the vortices have fully developed.
We point out that the tracking based on Gaussian fitting of the vortices works equally well for all temperatures
since the vortices are always resolved by the same number of grid points.
The longitudinal and transverse velocities, defined as before and again computed via finite differences from the tracked trajectories,
are shown in figure \ref{vels_mu}.
\begin{figure}[t]
\begin{center}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics{LongitudinalVelOverDist_mu.pdf}
\end{subfigure}\vspace{0cm}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics{TransversalVelOverDist_mu.pdf}
\end{subfigure}
\caption{Longitudinal (upper plot) and transverse (lower plot)
components of the vortex velocities for five different temperatures.
The initial vortex separations $d_0$ are given in table \ref{mu_params}.
Shown are the velocities for the anti-vortex.
For the vortex the curves are analogous except for an additional sign in the longitudinal component.
The effects due to the phase healing at early times (or large separations) are
stronger for higher temperatures.
\label{vels_mu}}
\end{center}
\end{figure}
The figure shows that the effect of the phase healing is present for all temperatures, but more pronounced at higher temperatures.
We further see that after the phase healing process the velocities exhibit a clear ordering according with temperature.
Both transverse and longitudinal velocities decrease with increasing temperature.
An interpretation of the ordering of the velocities with temperature in terms of dissipation, for example
along the lines of Tisza's two-fluid model, would appear speculative as
the results are obtained in the probe-approximation and might be affected when back-reaction is included.
\section{Summary and Outlook}
\label{Sec:Summary}
We have studied the dynamics of a single vortex--anti-vortex pair in a holographic superfluid in numerical real-time simulations.
A new tracking method has been introduced which allows us to locate the vortices with strongly improved precision
as compared to the conventional plaquette-based technique of locating the phase windings in the condensate density.
The new tracking method relies on locating the minimum of the density depletion of the vortex rather than only its phase winding.
Using a fit with Gaussians, supplemented by a Newton-Raphson method at small vortex separations, we can determine the vortex
positions with sub-plaquette resolution in a numerically efficient manner. Fitting both vortices simultaneously with
a linear combination of two two-dimensional Gaussians even allows us to capture the mutual deformation at small distances,
thus eliminating a further source of error in the vortex positions. Our tracking method can be applied throughout the evolution
of the vortex dipole. With this method we have, for the first time for the holographic superfluid, computed quasi-continuous
trajectories of the vortices during their approach and eventual annihilation. In addition, we use large numerical grids of
$512 \times 512$ points in the $(x, y)$-directions such that finite-size (or periodicity) effects are efficiently suppressed, as
we have checked on a $1024 \times 1024$ grid.
The increased precision makes it possible to study many aspects of the dynamics of the quantized vortices
which are not accessible with plaquette-resolution.
In particular, we are able to derive directly the velocities and accelerations of the vortices during the whole evolution of the dipole.
This allows us to improve earlier results of the pioneering study \cite{Lan:2018llf} in which the velocities were derived from
plaquette-resolution trajectories. We partially confirm the results found there for the power law for the velocity during the
dipole evolution. We find a very similar though not exactly equal exponent for large vortex separations, with our result indicating a
slight deviation from the behavior expected from the Hall--Vinen--Iordanskii equations for point-like vortices.
However, we do not confirm the finding of a second power law for small distances reported in \cite{Lan:2018llf}.
A key result of our study is the observation of a universal behavior of the vortex dipole in the holographic superfluid.
Specifically, we find that its
trajectory and velocities are universal independent of the initial separation. Hence their interaction in the superfluid
is only determined by their separation and does not depend on the previous motion. This finding holds for all fixed
temperatures of the superfluid condensate. We point out that we observe in our simulations some
deviations from this universal behavior at early times, visible as a slight initial outward bending of the trajectories.
But we have shown that these can be fully attributed to artifacts of the numerical initialization of the phase field of
the vortices. Such artifacts occur not only in the holographic superfluid but also in other numerical simulations of superfluids like for
example in Gross--Pitaevskii dynamics. It might be interesting to study this problem in more detail so that one could
potentially identify better methods for initializing ensembles of vortices in superfluid simulations.
We note that our finding of universality in the vortex dynamics could again only be obtained due to the improved
precision of the vortex tracking.
We have investigated how the dynamics of the vortex dipole depends on the temperature of the superfluid.
Our results indicate an overall decrease in the vortex velocities with increasing temperature if
we compare the system for suitably rescaled dimensionless quantities. However, the probe approximation
for the holographic superfluid might not be well suited to obtain a full picture of the temperature dependence.
It would therefore be interesting to study this question in a fully back-reacted system.
The vortex--anti-vortex system studied here is one of the most basic non-trivial examples of topological
defects in a superfluid, and in some sense a basic building block for understanding more complicated
ensembles of vortices. Such systems of vortices and anti-vortices, as well as solitons, have
been widely studied for the holographic superfluid. We expect that the improved method for vortex tracking
can be very advantageous also for improving our understanding of such systems. It would for example be
interesting to see whether the universality of the dynamics of a vortex dipole has an equivalent for systems of
many vortices.
The tracking method presented here can also be used for simulations of vortices in Gross--Pitaevskii dynamics.
An interesting result of such a study is the quantitative determination of the strong dissipation and of the friction
coefficients of vortices in the holographic superfluid \cite{Wittmer:2020mnm}.
This result has been obtained by a high-precision comparison of vortex trajectories
in the holographic superfluid with those of Gross--Pitaevskii dynamics and of the Hall--Vinen--Iordanskii equations.
Finally, we expect that our tracking method will also be useful in the study of vortex dynamics in
three-dimensional superfluids, including holographic superfluids.
\acknowledgments
We would like to thank Thomas Gasenzer, Markus Karl, Christian-Marcel Schmied and Ulrich Uwer for helpful discussions, and Panayotis Kevrekidis for drawing our attention to \cite{Christov:2018ecz,Christov:2018wsa}.
The work of P.\,W.\ was supported by the Studien\-stiftung des Deutschen Volkes e.\,V.
| 2c22d3427354bcc232180bb49c16101ac8c95a52 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Finite dimensional Alexandrov spaces with lower curvature bound
can be viewed as a generalization of Riemannian manifolds with lower sectional curvature bound
(\cite{BGP}). Compared to Riemannian manifolds, Alexandrov spaces might have some singularities,
so that some important subsets appear as some kinds of analogues of totally geodesic submanifolds,
such as convex subsets, extremal subsets and quasi-geodesics ([PP1-2]). Recently,
such a kind of subsets named quasi-convex subsets has been introduced (\cite{SSW}).
They include not only all convex subsets without boundary and extremal subsets but also more other subsets,
such as fixed point sets of isometries on Alexandrov spaces.
Moreover, all shortest pathes in a quasi-convex subset are quasi-geodesics.
To show the definition of quasi-convex subsets, we first make some conventions on notations.
\noindent$\bullet$\ \ $\text{Alex}(k)$: the set of complete and finite dimensional Alexandrov spaces with
curvature $\geqslant k$.
\noindent$\bullet$\ \ $\Bbb S^2_k$: the complete and simply connected space form of dimension 2 and curvature $k$.
\noindent$\bullet$\ \ $|pq|$, $[pq]$: the distance, a minimal geodesic (i.e. shortest path) between $p$ and $q$.
\noindent$\bullet$\ \ $\uparrow_p^q$: the direction from $p$ to $q$ for a given $[pq]$.
\noindent$\bullet$\ \ $\Uparrow_p^q$: the set of all directions from $p$ to $q$.
\noindent For a point $p\in X\in \text{Alex}(k)$, we denote by $\Sigma_pX$ the space of directions of $X$ at $p$ which
belongs to $\text{Alex}(1)$ ([BGP]). In $\Sigma_pX$, $\Uparrow_p^q$ is a closed subset, and
$|\Uparrow_p^q\Uparrow_p^r|$ is the distance between $\Uparrow_p^q$ and $\Uparrow_p^r$. And to $p,q,r\in X$, we associate $\tilde p, \tilde q, \tilde r\in \mathbb S^2_k$
with $|\tilde p\tilde q|=|pq|, |\tilde p\tilde r|=|pr|$ and $|\tilde q\tilde r|=|qr|$ \footnote{If $k>0$ and
$|pq|+|pr|+|qr|=\frac{2\pi}{\sqrt k}$, it is necessary to add a condition that $\tilde p, \tilde q$ and $\tilde r$ lie in a geodesic of length $\frac{\pi}{\sqrt k}$.},
and then we denote by $\tilde\angle_k qpr$ the angle at $\tilde p$ of the triangle
$\triangle \tilde p\tilde q\tilde r$.
\begin{defn}\label{def0.1}{\rm In an $X\in \text{Alex}(k)$, a closed subset $F$ is called to be {\it quasi-convex}
if the following condition is satisfied: if the distance function to $q\not\in F$ restricted to $F$, ${\operatorname{dist}}_q|_F$, attains a minimum at $p\in F$, then for all $r\in F\setminus\{p\}$
\begin{equation} \label{eqn0.1}
|\Uparrow_p^q\Uparrow_p^r|\leqslant\frac{\pi}{2} \text{\ \ (or equivalently, $\tilde\angle_k qpr\leqslant\frac{\pi}{2}$)}.
\end{equation}
Here, we make a convention that both the empty set and a single point are quasi-convex in $X$.}
\end{defn}
By Toponogov's Theorem\footnote{For the theorem, one can refer to Section 3 in [BGP] (or Theorem 1.1 in \cite{SSW}).}, it is obvious that `$|\Uparrow_p^q\Uparrow_p^r|\leqslant\frac{\pi}{2}$' implies `$\tilde\angle_k qpr\leqslant\frac\pi2$', but not vice versa; however, they are equivalent to each other in the situation of Definition \ref{def0.1} (\cite{SSW}). Moreover,
due to the arbitrariness of $q$, (\ref{eqn0.1}) is in fact equivalent to $|\uparrow_p^q\Uparrow_p^r|\leqslant\frac{\pi}{2}$ for any $[pq]$.
\begin{remark}\label{rem0.2}{\rm In this paper, we say that $F$ is {\it extremal} in $X$ if (\ref{eqn0.1}) holds for all $r\in X\setminus\{p\}$ in Definition \ref{def0.1}. This coincides with the concept `extremal' in \cite{PP1} if $F$ contains at least two points (\cite{SSW}). If $F$ is the empty set or a single point and if $k=1$,
some extra conditions are added in \cite{PP1} for some kind of completeness (cf. \cite{SSW}).}
\end{remark}
\begin{remark}\label{rem0.3}{\rm In \cite{SSW}, locally quasi-convex subsets are also defined. In detail,
a subset $F$ in an $X\in \text{Alex}(k)$ is {\it locally quasi-convex}
if for any $x\in F$ there is a neighborhood $U_x$ of $x$ such that $U_x\cap F$ is closed and if
${\operatorname{dist}}_q|_F$ with $q\not\in F$ attains a minimum at $p\in F\cap U_x$, then the corresponding
(\ref{eqn0.1}) holds for all $r\in F\cap U_x\setminus\{p\}$. In a complete Riemannian manifold, a closed and locally quasi-convex subset must be a totally geodesic submanifold.}
\end{remark}
It is obvious that the quasi-convexity (as well as the extremality) of $F$ is determined by the geometry at points realizing minimums of ${\operatorname{dist}}_q|_F$ with $q\notin F$. A natural question is what is the essential geometry to a general point of $F$.
The first result of this paper gives an answer to it.
\vskip2mm
\noindent {\bf Theorem A.}{\it \label{thmA}
Let $F$ be a closed subset in an $X\in \text{\rm Alex}(k)$. Then $F$ is quasi-convex in $X$ if and only if, for any two distinct points $p, r\in F$ and $\eta\in\Sigma_pX$, there is $\zeta\in\Uparrow_p^r$ and $\xi\in\Sigma_pF$ with $|\eta\xi|\leqslant\frac\pi2$ such that
\begin{equation} \label{eqn0.2}\cos|\eta\zeta|\geqslant\cos|\eta\xi|\cos|\zeta\xi|;
\end{equation}
where if $\Sigma_pF=\emptyset$ (i.e. $p$ is an isolated point of $F$), then (\ref{eqn0.2}) means that $\cos|\eta\zeta|\geqslant 0$. }
\vskip2mm
Note that, in Theorem A, if $\eta$ lies in $\Sigma_pF$, then we can let $\xi=\eta$.
And an alternative formulation of (\ref{eqn0.2}) is $\tilde\angle_1 \eta\xi\zeta\leqslant\frac{\pi}{2}$.
\begin{remark}\label{rem0.4}{\rm In Theorem A, according to the proof of Theorem A,
if $F$ is quasi-convex in $X$, we can in fact select $\xi$
with $|\eta\xi|\leqslant\frac\pi2$ to satisfy a bit stronger version of (\ref{eqn0.2}):
$$\cos|\eta\Uparrow_p^r|\geqslant\cos|\eta\xi|\cos|\Uparrow_p^r\xi|.$$}
\end{remark}
Restricted to `extremal' case, Theorem A can be formulated as follows, which can be seen almost obviously from the proof of Theorem A.
\vskip2mm
\noindent {\bf Corollary B.}{\it \label{croB}
Let $F$ be a closed subset in an $X\in \text{\rm Alex}(k)$. Then
$F$ is extremal in $X$ if and only if, for any $p\in F$ and any $\eta,\zeta\in\Sigma_pX$, there is $\xi\in\Sigma_pF$ with $|\eta\xi|\leqslant\frac\pi2$ such that $\cos|\eta\zeta|\geqslant\cos|\eta\xi|\cos|\zeta\xi|$ (which means that $\cos|\eta\zeta|\geqslant 0$ if $\Sigma_pF=\emptyset$).}
\vskip2mm
From Theorem A, we can derive another equivalent condition of quasi-convexity, which will be
very helpful to see some nice properties of quasi-convex subsets.
\vskip2mm
\noindent {\bf Corollary C.}{\it \label{corC}
Let $F$ be a closed subset in an $X\in \text{\rm Alex}(k)$. Then $F$ is quasi-convex in $X$ if and only if, for any two distinct points $p, r\in F$, if there is $\eta\in\Sigma_pX$ satisfying $|\eta\Uparrow_p^r|>\frac\pi2$, then the farthest direction to $\Uparrow_p^r$ in $\Sigma_pX$ belongs to $\Sigma_pF$. }
\vskip2mm
Note that the farthest direction $\xi$ to $\Uparrow_p^r$ in $\Sigma_pX$,
i.e. $|\xi\Uparrow_p^r|=\max\{|\nu\Uparrow_p^r| : \nu\in \Sigma_pX\}$, is unique if the maximum is bigger than $\frac\pi2$
(by Toponogov's Theorem).
In Corollary C, if $\Sigma_pF=\emptyset$, then ``if there is $\eta\in\Sigma_pX$ satisfying $|\eta\Uparrow_p^r|>\frac\pi2$, then ...'' means that $|\eta\Uparrow_p^r|\leqslant\frac\pi2$ for all $\eta\in\Sigma_pX$.
In Sections 2 and 4, we will present two equivalent versions of Corollary C (see Propositions \ref{prop2.1} and \ref{prop4.1}).
Similarly, for `extremal' case, Corollary C can be formulated as follows:
{\it In an $X\in \text{\rm Alex}(k)$, a closed subset $F$ is extremal if and only if,
for any $p\in F$ and $\zeta\in\Sigma_pX$, if there is $\eta\in\Sigma_pX$ satisfying $|\eta\zeta|>\frac\pi2$,
then the farthest direction to $\zeta$ in $\Sigma_pX$ belongs to $\Sigma_pF$.}
The necessity of the extremality of $F$ here can be seen from \cite{PP1},
and the sufficiency has been given in \cite{Pet}.
\begin{remark}\label{rem0.5}{\rm Based on Remark \ref{rem0.3}, one can give the corresponding versions
of Theorem A and Corollary C (and Propositions \ref{prop2.1} and \ref{prop4.1}) for local quasi-convexity. }
\end{remark}
\begin{remark}\label{rem0.6}{\rm
Let $F$ be a quasi-convex subset in $X\in \text{\rm Alex}(k)$. \cite{SSW} has shown two important properties of $F$.
One is that $\Sigma_pF$ is also quasi-convex in $\Sigma_pX$ for any $p\in F$,
and the other is that a shortest path in $F$ is a quasi-geodesic in $X$. They can be proven
just from Definition \ref{def0.1} without involving the new viewpoints in Theorem A and Corollary C.}
\end{remark}
In the rest of the paper, we first give proofs of Theorem A and Corollary C in Sections 1 and 2.
Then, as applications of them, we will show several properties of quasi-convex subsets.
For instance, two points in a quasi-convex subset can be jointed with a curve in the subset if they are close sufficiently
(see Section 3), and the intersection of two quasi-convex subsets is also quasi-convex (see Section 4.2).
Moreover, based on Corollary C, we can illustrate quasi-convex subsets by gradient curves of distance functions (see Section 4.1). This will make it easy to judge some kinds of subsets to be quasi-convex (see Sections 4.2-4), such as the fixed point set of an isometry.
\section{Proofs of Theorem A and Corollary B}
The main goal of this section is to give a proof for Theorem A.
In the proof, we will use the following facts in Alexandrov geometry.
\begin{lemma}[\cite{BGP}]\label{lem1.1}
Let $X\in \text{\rm Alex}(k)$ and $p\in X$. Then for any small $\epsilon>0$
there is a neighborhood $U_\epsilon$ of $p$ such that, for any $[pq],[pr], [qr]\subset U_\epsilon$,
$$0\leqslant|\uparrow_p^q\uparrow_p^r|-\tilde\angle_k qpr<\epsilon,\ \
0\leqslant|\uparrow_q^p\uparrow_q^r|-\tilde\angle_k pqr<\epsilon\ \text{ and }\
0\leqslant|\uparrow_r^p\uparrow_r^q|-\tilde\angle_k prq<\epsilon.$$
\end{lemma}
\begin{lemma}[\cite{BGP}]\label{lem1.2}
Let $X\in\text{\rm Alex}(k)$, and let $p, r, q_i\in X$ with $q_i\to p$ as $i\to\infty$.
If there is $[pq_i]$ such that $\uparrow_p^{q_i}$ converges to $\eta\ (\in\Sigma_pX)$ as $i\to\infty$,
then
$$\lim_{i\to\infty}\frac{|q_ir|-|pr|}{|pq_i|}=-\cos|\Uparrow_p^{r}\eta|.$$
As a result, $\lim\limits_{i\to\infty}\tilde\angle_krpq_i=|\Uparrow_p^{r}\eta|$.
\end{lemma}
\begin{proof}[Proof of Theorem A]\
We first show the necessity of the quasi-convexity in the theorem.
Note that we can let $\xi=\eta$ if $\eta\in \Sigma_{p} F$,
so we can assume that $\eta\notin \Sigma_{p} F$. Then
there is $q_n\in X\setminus F$ such that $q_n \to p $ and $\uparrow_{p}^{q_n}\to \eta$ as $n\to \infty$.
Let $p_n\in F$ satisfy $|q_np_n|=|q_nF|$. Note that $p_n\to p$ as $n\to \infty$,
and we can complete the proof according to the following two cases.
Case 1: there is at least a subsequence of $\{p_n\}$ which belongs to $F\setminus\{p\}$.
In this case, there must be a subsequence
$\{p_{n_i}\}$ with $p_{n_i}\neq p$ such that $\uparrow_{p}^{p_{n_i}}$ converges to some $\xi\in\Sigma_pF$ as $i\to\infty$.
We claim that $|\eta\xi|\leqslant\frac\pi2$ and
\begin{equation} \label{eqn1.1}\cos|\eta\Uparrow_p^r|\geqslant\cos|\eta\xi|\cos|\Uparrow_p^r\xi|.
\end{equation}
This implies that there is $\zeta\in\Uparrow_p^r$ such that (\ref{eqn0.2}) holds. So, we just need to verify the claim in this case. For simpleness, we still denote by $\{p_n\}$ the subsequence $\{p_{n_i}\}$.
Since $|q_np_n|<|q_np|$ and $q_n, p_n\to p$ as $n\to\infty$,
by Lemma \ref{lem1.1} it is easy to see that $|\eta\xi|\leqslant\frac\pi2$. For (\ref{eqn1.1}),
due to the similarity, we only give a proof for the case where $k=0$. By the quasi-convexity of $F$, we have that
\begin{equation} \label{eqn1.2}
\tilde \angle_k q_np_nr\leqslant \frac{\pi}{2}.
\end{equation}
Consequently, by the Law of Cosine, we have that
$$|rp_n|^2+|q_np_n|^2\geqslant |q_nr|^2,$$
where
\begin{align*}
|rp_n|^2 & =|rp|^2+|pp_n|^2-2|rp|\cdot |pp_n|\cdot \cos \tilde \angle_0 rpp_n, \\
|q_np_n|^2 & =|q_np|^2+|pp_n|^2-2|q_np|\cdot |pp_n|\cdot \cos \tilde \angle_0 q_npp_n,\\
|q_nr|^2 & =|q_np|^2+|pr|^2-2|q_np|\cdot |pr|\cdot \cos \tilde \angle_0 q_npr.
\end{align*}
It then follows that
\begin{align}
\cos \tilde\angle_0 rpq_n & \geqslant -\frac{|pp_n|^2}{|pq_n|\cdot |pr|} +\frac{|pp_n|}{|pq_n|}\cos\tilde \angle_0 rpp_n+\frac{|pp_n|}{|pr|}\cos\tilde \angle_0 q_npp_n. \label{eqn1.3}
\end{align}
Since $q_n, p_n\to p$, $\uparrow_{p}^{q_n}\to \eta$ and $\uparrow_{p}^{p_n}\to \xi$ as $n\to\infty$, by Lemma \ref{lem1.2} we have that
\begin{equation} \label{eqn1.4}
\tilde \angle_0 rpq_n\to |\Uparrow_{p}^r\eta |\ \text{ and }\
\tilde \angle_0 rpp_n \to |\Uparrow_{p}^r\xi|\ \text{ as }\ n\to \infty.
\end{equation}
Moreover, we make a subclaim:
\begin{equation} \label{eqn1.5}
\lim_{n\to\infty}\frac{|pp_n|}{|pq_n|}=\cos|\eta \xi|.
\end{equation}
Note that (\ref{eqn1.3})-(\ref{eqn1.5}) together with $\frac{|pp_n|}{|pr|}\to 0$ as $n\to \infty$ implies (\ref{eqn1.1}).
In order to see the subclaim, by Lemma \ref{lem1.1} it suffices to show that
there is $[p_nq_n]$ and $[p_np]$ such that $|\uparrow_{p_n}^{q_n}\uparrow_{p_n}^p|\to\frac\pi2$ as $n\to\infty$. By the quasi-convexity of $F$, there is $[p_nq_n]$, $[p_np]$ and $[p_np_m]$ with $m\neq n$
such that
$$|\uparrow_{p_n}^{q_n}\uparrow_{p_n}^p|\leqslant\frac\pi2 \text{ and }
|\uparrow_{p_n}^{q_n}\uparrow_{p_n}^{p_{m}}|\leqslant\frac\pi2.$$
On the other hand, since $p_n\to p$ and $\uparrow_{p}^{p_n}\to \xi$ as $n\to\infty$,
it is not hard to see that $|\uparrow_{p_n}^p\uparrow_{p_n}^{p_{m}}|\to \pi$
where $n\gg m$ and $m\to\infty$ (by Toponogov's Theorem), which implies
$$|\uparrow_{p_n}^{q_n}\uparrow_{p_n}^p|+|\uparrow_{p_n}^{q_n}\uparrow_{p_n}^{p_{m}}|\to \pi.$$ It therefore follows that
$|\uparrow_{p_n}^{q_n}\uparrow_{p_n}^p|\to\frac\pi2$ as $n\to\infty$ (i.e. the subclaim is verified, so is the claim).
Case 2: $p_n=p$ for all large $n$. In this case, the quasi-convexity of $F$ guarantees that $|\Uparrow_p^{q_n} \Uparrow_{p}^{r}|\leqslant\frac{\pi}{2}$, and thus $|\eta \Uparrow_{p}^{r}|\leqslant\frac{\pi}{2}$
(note that `$q_n\to p$ and $\uparrow_{p}^{q_n}\to \eta$' implies $\Uparrow_{p}^{q_n}\to \eta$, cf. \cite{BGP}).
I.e., there is $\zeta\in\Uparrow_p^r$ such that $\cos|\eta\zeta|\geqslant 0$, so the proof of the necessity is done if $\Sigma_p F=\emptyset$.
If $\Sigma_p F\neq\emptyset$, we claim that
\begin{align*}
& |\uparrow_p^{q_n}\xi|=\frac{\pi}{2} \ \text{ for any } \ \xi\in \Sigma_p F,
\end{align*}
which implies $|\eta\xi|=\frac{\pi}{2}$.
In fact, we have that $|\uparrow_p^{q_n}\xi|\geqslant\frac\pi2$ because $|q_np|=|q_nF|$ (by Lemma \ref{lem1.2}),
and meanwhile $|\uparrow_p^{q_n}\xi|\leqslant\frac\pi2$ (by the quasi-convexity of $F$).
Note that the claim right above implies that any $\xi\in \Sigma_p F$ satisfies (\ref{eqn0.2}), and thus the proof of the necessity is done.
\vskip2mm
Next, we will verify the sufficiency of the quasi-convexity in Theorem A. We argue by contradiction.
Suppose that $F$ is not quasi-convex in $X$. Then there exists $q\not\in F$ and $p,r\in F$ such that $|qp|=|qF|$, but for some $[qp]$ we have that
\begin{equation} \label{eqn1.6}
|\uparrow_{p}^q\Uparrow_{p}^r|>\frac{\pi}{2}.
\end{equation}
We let $\eta\triangleq \uparrow_{p}^q$. By the assumption, there is $\zeta\in\Uparrow_{p}^r$ and $\xi\in\Sigma_pF$ with $|\eta\xi|\leqslant\frac\pi2$ such that
$$\cos |\eta\zeta| \geqslant \cos|\eta\xi| \cdot \cos|\zeta\xi|.$$
Note that `$|\eta\zeta |>\frac\pi2$' (see (\ref{eqn1.6})) together with `$|\eta\xi|\leqslant\frac\pi2$' implies that $|\eta\xi|$ must be less than $\frac\pi2$. However, since $|qp|=|qF|$, by Lemma \ref{lem1.2} we have that $|\eta\xi|\geqslant\frac\pi2$,
a contradiction.
\end{proof}
By restricting the above proof to `extremal' case, we can easily derive Corollary B.
\begin{proof}[Proof of Corollary B]\
If $\zeta$ is the direction of some $[pr]$ at $p$ (here $r$ is not needed to lie in $F$), one just need to replace all
`quasi-convexity' with `extremality' in the proof of Theorem A.
Then for the case where $\zeta$ cannot be realized by a minimal geodesic, we can draw the conclusion by a limiting argument (note that there is $[pr_i]$ with $r_i\to p$ as $i\to\infty$ such that $\uparrow_p^{r_i}\to\zeta$).
\end{proof}
\begin{remark}\label{rem1.3}{\rm Similar to Corollary B, if $F$ is a nonempty quasi-convex subset in
$X\in \text{\rm Alex}(k)$, and if $\Sigma_pF\neq\emptyset$ with $p\in F$, then for any $\eta\in\Sigma_pX$ and $\zeta\in\Sigma_pF$ there is $\xi\in\Sigma_pF$ with $|\eta\xi|\leqslant\frac\pi2$ such that $\cos|\eta\zeta|\geqslant\cos|\eta\xi|\cos|\zeta\xi|$ (by Theorem A). However, the converse might not be true (as an example,
one can consider a submanifold in a Riemannian manifold which is not totally geodesic). }
\end{remark}
\section{Proof of Corollary C}
In this section, we will first prove Corollary C, and then show an alternative version of it.
\begin{proof} [Proof of Corollary C]\
We first verify the necessity of the quasi-convexity in the corollary.
Let $p,r\in F$. Since there is $\eta\in\Sigma_pX$ such that $|\eta\Uparrow_p^r|>\frac\pi2$,
the farthest direction $\xi_0$ to $\Uparrow_p^r$ in $\Sigma_pX$ is unique and $|\xi_0\Uparrow_p^r|>\frac\pi2$.
By Theorem A (and Remark \ref{rem0.4}, see (\ref{eqn1.1})), the quasi-convexity of $F$ implies that there is $\xi\in\Sigma_pF$ with $|\xi_0\xi|\leqslant\frac\pi2$ such that
$$\cos|\xi_0\Uparrow_p^r|\geqslant\cos|\xi_0\xi|\cos|\Uparrow_p^r\xi|.$$
Note that `$|\xi_0\Uparrow_p^r|>\frac\pi2$', `$|\xi_0\Uparrow_p^r|\geqslant|\xi\Uparrow_p^r|$' and `$|\xi_0\xi|\leqslant\frac\pi2$' imply that $\xi_0$ has to be equal to $\xi$, so it follows that $\xi_0\in\Sigma_pF$.
Next, we will verify the sufficiency, and argue by contradiction.
Suppose that $F$ is not quasi-convex in $X$. Then there exists $q\not\in F$ and $p,r\in F$ such that $|qp|=|qF|$, but for some $[qp]$ we have that
$$|\Uparrow_{p}^r\uparrow_{p}^q|>\frac{\pi}{2}. $$
Hence, by the assumption, the farthest direction $\xi$ to $\Uparrow_p^r$ in $\Sigma_pX$
belongs to $\Sigma_pF$. We claim that $|\uparrow_{p}^q\xi|<\frac\pi2$, which
contradicts `$|qp|=|qF|$' (by Lemma \ref{lem1.2}).
In fact, if $|\uparrow_{p}^q\xi|\geqslant\frac\pi2$, by Toponogov's Theorem it is not hard to see that
$|\Uparrow_p^r\xi'|>|\Uparrow_p^r\xi|$ for $\xi'$ near $\xi$ in any $[\uparrow_{p}^q\xi]$
because $|\Uparrow_{p}^r\uparrow_{p}^q|>\frac{\pi}{2}$ and $|\Uparrow_p^r\xi|\geqslant|\Uparrow_{p}^r\uparrow_{p}^q|>\frac{\pi}{2}$; a contradiction
(because $\xi$ is the farthest direction to $\Uparrow_p^r$).
\end{proof}
We now provide an alternative version of Corollary C, which will be easy to be used
in next section. We formulate it in the following proposition.
\begin{prop}\label{prop2.1}
Let $F$ be a closed subset in an $X\in \text{\rm Alex}(k)$. Then $F$ is quasi-convex in $X$ if and only if, for any two distinct points $p, r\in F$, if $\eta \in \Sigma_p X$ satisfies $-\cos|\Uparrow_{p}^r \eta|>0$, then there exists $\xi\in \Sigma_{p}F$ such that
\begin{equation} \label{eqn2.1}
-\cos |\Uparrow_{p}^r\xi | \geqslant -\cos|\Uparrow_{p}^r \eta|.
\end{equation}
Moreover, when $F$ is quasi-convex and $-\cos|\Uparrow_{p}^r \eta|>0$, the $\xi$ in (\ref{eqn2.1}) can be chosen to satisfy $\cos |\eta\xi|\geqslant-\cos|\Uparrow_{p}^r \eta|$; as a result, there is $\xi'\in \Sigma_{p}F$ such that
\begin{equation} \label{eqn2.2}
\cos |\Uparrow_{p}^r\xi' | \geqslant -\cos|\Uparrow_{p}^r \eta|.
\end{equation}
\end{prop}
\begin{proof} Note that the first statement is an alternative formulation of Corollary C. So, we just need to verify the second one. Assume that $F$ is quasi-convex and $\eta \in \Sigma_p X$ satisfies $-\cos|\Uparrow_{p}^r \eta|>0$.
By Theorem A, there is $\zeta\in\Uparrow_p^r$ and $\xi\in\Sigma_pF$ with $|\eta\xi|\leqslant\frac\pi2$ such that
$$\cos|\eta\zeta|\geqslant\cos|\eta\xi|\cos|\zeta\xi|.$$
This implies that $\cos |\eta\xi|\geqslant -\cos|\Uparrow_{p}^r \eta|$ because $-\cos|\eta\zeta|\geqslant -\cos|\Uparrow_{p}^r \eta|>0$ and $\cos|\eta\xi|\geqslant0$.
We next show that there is $\xi'\in \Sigma_{p}F$ such that (\ref{eqn2.2}) holds. Note that there
is a sequence of $r_i\in F$ such that $r_i\to p$ and $\Uparrow_p^{r_i}\to\xi$ as $i\to\infty$.
And for any $\zeta\in \Uparrow_p^r$, by replacing $\Uparrow_p^r$ and $\eta$ with $\Uparrow_p^{r_i}$ and $\zeta$ respectively,
we can conclude that there is $\xi_i\in \Sigma_{p}F$ such that $\cos |\zeta\xi_i|\geqslant -\cos|\Uparrow_p^{r_i}\zeta|$,
which implies $\cos |\Uparrow_p^{r}\xi_i|\geqslant -\cos|\Uparrow_p^{r}\Uparrow_p^{r_i}|$.
Thereby, for the limit $\xi'$ of any converging subsequence of $\{\xi_i\}$ (note that $\Sigma_pF$ is a
closed subset in $\Sigma_pX$), we have that
$\cos |\Uparrow_p^{r}\xi'|\geqslant -\cos|\Uparrow_p^{r}\xi|\geqslant -\cos|\Uparrow_{p}^r \eta|$.
\end{proof}
\begin{remark}\label{rem2.2}{\rm By replacing $\Uparrow_{p}^r$ with any $\zeta\in\Sigma_pX$ in Proposition \ref{prop2.1},
we can get the corresponding version of the proposition for `extremal' case. Here, the existences of $\xi$ and $\xi'$ by the extremality of $F$ can be seen from Proposition 1.6 in \cite{PP1}. }
\end{remark}
\vskip2mm
In the rest of the paper, we will present some properties of quasi-convex subsets, as applications of
the idea of Theorem A (and its equivalent versions---Corollary C and Proposition 2.1).
\section{Connectedness of quasi-convex subsets}
It is known that the number of extremal subsets in a compact space of ${\operatorname{Alex}}(k)$ is finite (Proposition 3.6 in \cite{PP1}). Unfortunately, there is no such strong conclusion on quasi-convex subsets. For example, in a standard sphere, there are infinitely many great circles each of which is quasi-convex. Nevertheless, we have
a weaker conclusion for quasi-convex subsets.
\begin{prop} \label{prop3.1}
In a compact space $X\in {\operatorname{Alex}}(k)$, any quasi-convex subset has a finite number of connected components.
\end{prop}
In fact, we have the following stronger conclusion than Proposition \ref{prop3.1} (which
corresponds to (2) of Corollary 3.2 in \cite{PP1} for `extremal' case).
\begin{prop} \label{prop3.2}
Let $X$ be a compact space in ${\operatorname{Alex}}(k)$, and let $F$ be a quasi-convex subset in $X$. Then there is $\epsilon>0$ (depending on $X$) such that, for any two distinct points $p,q\in F$ with $|pq|<\epsilon^2$, there exists a curve in $F$ jointing $p$ and $q$ with length bounded from above by $\frac{|pq|}{\epsilon}$.
\end{prop}
Moreover, we can see the following property which corresponds to Proposition 3.3 in \cite{PP1} for `extremal' case.
\begin{prop}\label{prop3.3}
Let $F$ be a quasi-convex subset in an $X\in{\operatorname{Alex}}(k)$, and let $p\in F$ and $\xi\in \Sigma_{p} F$. Then there exists a curve in $F$ starting from $p$ and tangent to the direction $\xi$.
\end{prop}
In proving Propositions 3.1 and 3.2, the following lemma (Lemma 3.1 in \cite{PP1}) is needed, which is some kind of essential geometry of Alexandrov spaces with lower curvature bound.
\begin{lemma}\label{lem3.4}
Let $X$ be a compact space in ${\operatorname{Alex}}(k)$. Then there is $\epsilon>0$ (depending on $X$) such that,
for any two distinct points $p, q \in X$ with $|p q|<\epsilon^{2}$, at least one of the following holds:
$$\max _{\eta \in \Sigma_{q} X}\operatorname{dist}_p'|_q(\eta)>\epsilon
\text{\ \ and }\ \max _{\eta \in \Sigma_{p} X}\operatorname{dist}_q'|_p(\eta)>\epsilon,$$
where $\operatorname{dist}_p'|_q(\eta)$ denotes the derivative of $\operatorname{dist}_p$ (the distance function to $p$) at $q$ along the direction $\eta$.
\end{lemma}
Note that ${\operatorname{dist}}_p'|_q(\eta)=-\cos|\Uparrow_{q}^p\eta|$ (by Lemma \ref{lem1.2}). Hence, by Proposition \ref{prop2.1} (see (\ref{eqn2.1}) and (\ref{eqn2.2})) we can easily see the following property.
\begin{lemma} \label{lem3.5}
In Lemma \ref{lem3.4}, if $p, q$ lie in a quasi-convex subset $F$ of $X$ additionally, then
$$\max _{\xi \in \Sigma_{q} F}{\operatorname{dist}}_p'|_q(\xi)>\epsilon \quad \text{and}\quad \min _{\xi \in \Sigma_{q} F}{\operatorname{dist}}_p'|_q(\xi)<-\epsilon,$$
or
$$\max _{\xi \in \Sigma_{p} F}{\operatorname{dist}}_q'|_p(\xi)>\epsilon \quad \text{and}\quad \min _{\xi \in \Sigma_{p} F }{\operatorname{dist}}_q'|_p(\xi)<-\epsilon.$$
\end{lemma}
Lemma \ref{lem3.5} corresponds to (1) of Corollary 3.2 in \cite{PP1} for `extremal' case, where $p$ and $q$ can lie in two distinct extremal subsets.
Since our proofs for Propositions 3.1 and 3.2 are imitations of their corresponding versions for `extremal' case in \cite{PP1}, we just provide rough proofs for them.
\begin{proof} [Proof of Proposition \ref{prop3.1}]\
Let $F$ be a quasi-convex subset in $X$, and let $F_1$ and $F_2$ be two connected components of $F$. It suffices to show that the distance between $F_1$ and $F_2$ is bigger than $\epsilon^2$,
where $\epsilon$ is the number associated to $X$ satisfying Lemma \ref{lem3.4}. If this is not true, then there is $p\in F_1$ and $q\in F_2$ such that $|pq|=|F_1F_2|<\epsilon^2$ (note that
$X$ is compact and $F$ is closed in $X$). It then follows that
$$\min_{\xi\in \Sigma_p F_1}{\operatorname{dist}}_q'|_p(\xi)\geqslant 0 \quad \text{and}\quad \min_{\xi\in \Sigma_q F_2}{\operatorname{dist}}_p'|_q(\xi)\geqslant 0,$$
which contradicts Lemma \ref{lem3.5}.
\end{proof}
\begin{proof} [Proof of Proposition \ref{prop3.2}]\
Let $p,q$ be two distinct points in $F$. By Lemma \ref{lem3.5}, there is an $\epsilon$ such that if $|pq|<\epsilon^2$, then
$$
\min _{\xi \in \Sigma_{q} F} \operatorname{dist}_{p}^{\prime} |_{q}(\xi)<-\epsilon\ \text{ or } \
\min _{\xi \in \Sigma_{p} F} \operatorname{dist}_{q}^{\prime} |_{p}(\xi)<-\epsilon.
$$
Then, for each $n\in \Bbb N^+$,
it is not hard to see that there are two sequences of points $\{p_i\}_{i=1}^\infty$ and $\{q_i\}_{i=1}^\infty$ (depending on $n$)
in $F$ with
\begin{equation}\label{eqn3.1}
\text{either }\ \ p_i=p_{i-1} \text{ and } 0<|q_{i}q_{i-1}|<\frac1n\ \ \text{ or } \ \
q_{i}=q_{i-1} \text{ and } 0<|p_ip_{i-1}|<\frac1n
\end{equation}
(where $p_0=p$ and $q_0=q$) such that
\begin{equation}\label{eqn3.2}
|p_{i-1}q_{i-1}|-|p_iq_i|>\epsilon (|p_{i-1}p_i|+|q_iq_{i-1}|)
\end{equation}
and
\begin{equation}\label{eqn3.3}
|p_iq_i|\to 0 \text{ as } i\to\infty.
\end{equation}
Let $\{p_{i,n}, q_{i,n}\}_{i=0}^\infty$ denote the above two point sequences corresponding to each $n$.
Note that $$\sum_{i=1}^\infty(|p_{i-1,n}p_{i,n}|+|q_{i,n}q_{i-1,n}|)<\frac{|pq|}{\epsilon}$$ for all $n$.
Hence, as $n\to \infty$ and passing to a subsequence of $\{n\}$,
$\{p_{i,n}, q_{i,n}\}_{i=0}^\infty$ converges to a curve in $F$ jointing $p$ and $q$ with length $\leqslant\frac{|pq|}{\epsilon}$.
\end{proof}
\begin{remark}\label{rem3.6}{\rm In the proof of Proposition \ref{prop3.2}, for general $\{p_i\}_{i=1}^\infty$ and $\{q_i\}_{i=1}^\infty$ satisfying (\ref{eqn3.1}) and (\ref{eqn3.2}), it might occur that
$p_i\to\bar p$ and $q_i\to\bar q$ as $i\to\infty$ with $\bar p\neq\bar q$. Note that $|\bar p\bar q|<|pq|<\epsilon^2$, so
similarly there is $p'$ and $q'$ in $F$ with either $p'=\bar p$ and $|q'\bar q|<\frac1n$ or
$q'=\bar q$ and $|p'\bar p|<\frac1n$ such that
$|\bar p\bar q|-|p'q'|>\epsilon (|\bar pp'|+|q'\bar q|)$. Hence, for sufficiently large $i$, we can reset
$p_i=p_{i-1}$ and $q_i=q'$ or $q_i=q_{i-1}$ and $p_i=p'$ so that the new $p_i$ and $q_i$ still satisfy (\ref{eqn3.1}) and (\ref{eqn3.2}). Such an idea enables us to find $\{p_i\}_{i=1}^\infty$ and $\{q_i\}_{i=1}^\infty$ satisfying (\ref{eqn3.1})-(\ref{eqn3.3}). }
\end{remark}
As for Proposition \ref{prop3.3}, we can almost directly copy the proof of Proposition 3.3 in \cite{PP1} (here the basis is Lemma \ref{lem3.5}
instead of Corollary 3.2 in \cite{PP1}). However, we would like to provide a proof for it via Corollary C without involving Lemmas \ref{lem3.5} and \ref{lem3.4}.
\begin{proof} [Proof of Proposition \ref{prop3.3}]\
Since $\xi\in\Sigma_pF$, there is $\{p_i\}_{i=1}^\infty\subset F$ such that $p_i\to p$ and $\Uparrow_p^{p_i}\to\xi$ as $i\to\infty$. Note that, for sufficiently small $\delta>0$, there is $i_0$ such that
\begin{equation} \label{eqn3.4}
|\Uparrow_p^{p_i}\Uparrow_p^{p_{i_0}}|<\frac{\delta}3 \text{ for all } i>i_0.
\end{equation}
Moreover, by Lemma \ref{lem1.1}, we can assume that
\begin{equation} \label{eqn3.5}
|\uparrow_p^x\uparrow_p^y|-\tilde{\angle }_k x p y<\frac{\delta}{3} \text{ for any } [px], [py]\subset\overline{B_{p}(|pp_{i_0}|)}.
\end{equation}
In the rest of the proof, for any $i>i_0$ with $|pp_i|\ll|pp_{i_0}|$, we will first construct an arc-length parameterized curve $\alpha_i(t)|_{t\in[0,\frac{|p_ip_{i_0}|}{2}]}$ $\subset F$ with $\alpha_i(0)=p_i$ such that
$$|\Uparrow_p^{\alpha_i(t)}\xi|<\delta \text{ for all } t\in[0, \frac{|p_ip_{i_0}|}{2}].$$
And then we will show that, as $i\to\infty$, $\alpha_i(t)$ converges to a curve we want.
In order to construct $\alpha_i(t)$, we claim that, for any $\epsilon\ll |pp_i|$, there is $\{z_j\}_{j=1}^{N(\epsilon)}\subset F$ with $z_1=p_i$ and a constant $C$ such that
$$|z_jz_{j+1}|<\epsilon,\ \left|\sum_{j=1}^{N(\epsilon)-1}|z_jz_{j+1}|-\frac{|p_ip_{i_0}|}{2}\right|<\epsilon \text{ and } \tilde\angle_kz_{j+1}pp_{i_0}<\tilde\angle_kz_jpp_{i_0}+\frac{C\epsilon}{|pp_i|}|z_jz_{j+1}|.$$
Note that (\ref{eqn3.4}) implies that $\tilde\angle_kz_1pp_{i_0}<\frac\delta3$ (by Toponogov's Theorem) and
$|\Uparrow_p^{p_{i_0}}\xi|\leqslant\frac{\delta}3$. Then taking into account (\ref{eqn3.5}), we can see that
$\{z_j\}_{j=1}^{N(\epsilon)}$ converges to the desired $\alpha_i(t)$ as $\epsilon\to 0$.
We now verify the above claim. Since $\tilde\angle_kp_ipp_{i_0}<\frac{\delta}3$
and $\tilde\angle_kpp_{i_0}p_i\ll \frac{\delta}3$ (note that $|pp_i|\ll |p_ip_{i_0}|$), we can assume that $\tilde\angle_kpp_ip_{i_0}>\pi-\delta$, and thus by Toponogov's Theorem we have that
$$|\Uparrow_{p_i}^{p}\Uparrow_{p_i}^{p_{i_0}}|>\pi-\delta.$$
By Corollary C, the farthest direction $\xi_i$ to $\Uparrow_{p_i}^p$ in $\Sigma_{p_i}X$ belongs to $\Sigma_{p_i}F$.
By Lemma \ref{lem1.2} and Toponogov's Theorem, the `farthest' property of $\xi_i$ implies that for any $\eta\in\Uparrow_{p_i}^{p_{i_0}}$ there is $\zeta\in\Uparrow_{p_i}^{p}$
such that $\tilde\angle_1\zeta\xi_i\eta\leqslant\frac\pi2$. It then follows that
$$
\cos |\Uparrow_{p_i}^p\Uparrow_{p_i}^{p_{i_0}}|\geqslant\cos |\zeta\eta|\geqslant \cos|\xi_i\zeta|\cdot\cos|\xi_i\eta|\geqslant \cos|\xi_i\zeta|\cdot\cos|\xi_i\Uparrow_{p_i}^{p_{i_0}}|
$$
(note that $|\xi_i\zeta|\geqslant |\xi_i\Uparrow_{p_i}^{p}|\geqslant |\Uparrow_{p_i}^{p}\Uparrow_{p_i}^{p_{i_0}}|>\pi-\delta$, and thus it holds that $|\xi_i\eta|<\frac\pi2$),
which implies that
\begin{equation}\label{eqn3.6}
\cos |\Uparrow_{p_i}^p\Uparrow_{p_i}^{p_{i_0}}|\geqslant -\cos|\xi_i\Uparrow_{p_i}^{p_{i_0}}|,
\text{ or equivalently, } |\Uparrow_{p_i}^p\Uparrow_{p_i}^{p_{i_0}}|+|\xi_i\Uparrow_{p_i}^{p_{i_0}}|\leqslant\pi.
\end{equation}
Denote by $z_1$ the point $p_i$, and note that $\xi_i\in\Sigma_{z_1}F$. Then for any $\epsilon\ll |pz_1|$, (\ref{eqn3.6}) together with Lemma \ref{lem1.2} implies that there is $z_2\in F$ such that $|z_1z_2|<\epsilon$ and
\begin{equation}\label{eqn3.7}
\tilde\angle_kpz_1p_{i_0}+\tilde\angle_kp_{i_0}z_1z_2<\pi+\epsilon.
\end{equation}
Let $\bar z_2,\bar p,\bar p_{i_0}\in \Bbb S_k^2$ satisfy $|\bar p\bar p_{i_0}|=|pp_{i_0}|$, $|\bar z_2\bar p_{i_0}|=|z_2p_{i_0}|$ and $|\bar p\bar z_2|=|pz_1|+|z_1z_2|$.
For a special case of (\ref{eqn3.7}) where $\tilde\angle_kpz_1p_{i_0}+\tilde\angle_kp_{i_0}z_1z_2\leqslant\pi$,
we notice that $\angle \bar z_2\bar p\bar p_{i_0}\leqslant\tilde\angle_kz_1pp_{i_0}\ (<\frac\delta3)$
by Alexandrov's lemma (Lemma 2.5 in \cite{BGP}),
which implies $\angle \bar p\bar z_2\bar p_{i_0}>\pi-\delta\ (>\frac\pi2)$. From `$\angle \bar p\bar z_2\bar p_{i_0}>\frac\pi2$',
it follows that $\tilde\angle_kz_2pp_{i_0}\leqslant\angle \bar z_2\bar p\bar p_{i_0}$ (note that $|pz_2|\leqslant|\bar p\bar z_2|$), so
\begin{equation}\label{eqn3.8}
\tilde\angle_kz_2pp_{i_0}<\tilde\angle_kz_1pp_{i_0} \text{ and } \tilde\angle_kpz_2p_{i_0}>\pi-\delta.
\end{equation}
In general, it is not so hard to conclude that
$\angle \bar z_2\bar p\bar p_{i_0}<\tilde\angle_kz_1pp_{i_0}+\frac{C\epsilon}{|pp_i|}|z_1z_2|$, where
$C$ is a constant depending only on $|pp_{i_0}|$.
Then we can similarly see that
\begin{equation}\label{eqn3.9}
\tilde\angle_kz_2pp_{i_0}\leqslant\angle \bar z_2\bar p\bar p_{i_0}<\tilde\angle_kz_1pp_{i_0}+\frac{C\epsilon}{|pp_i|}|z_1z_2| \text{ and } \tilde\angle_kpz_2p_{i_0}>\pi-\delta.
\end{equation}
Note that $|\Uparrow_{z_1}^p\xi_i|>\pi-\delta$, which implies that $|pz_2|>|pz_1|$ (by Lemma \ref{lem1.2}).
Then based on (\ref{eqn3.8}) and (\ref{eqn3.9}), we can similarly locate $z_j$ with $j\geqslant 3$ one by one.
Moreover, for a similar reason in Remark \ref{rem3.6}, there is an $N(\epsilon)$ such that
$\left|\sum_{j=1}^{N(\epsilon)-1}|z_jz_{j+1}|-\frac{|p_ip_{i_0}|}{2}\right|<\epsilon$. So far, the claim has been verified.
Note that $\alpha_i(t)$ converges to a curve $\alpha(t)|_{t\in[0,\frac{|pp_{i_0}|}{2}]}$ with $\alpha(0)=p$ as $i\to\infty$ (here, there might be a need of passing to a subsequence, but in fact not; cf. Remark \ref{rem3.7} below). We just need to show that $\alpha(t)$ is tangent to $\xi$ at $p$.
In fact, for any integer $n>1$, there is $i_n\gg i_0$ such that
$|\Uparrow_p^{p_i}\Uparrow_p^{p_{i_n}}|<\frac{\delta^n}3 \text{ for all } i>i_n$ (cf. (\ref{eqn3.4})).
Then we can similarly construct another curve
$\bar\alpha_i(t)|_{t\in[0,\frac{|p_ip_{i_n}|}{2}]}\subset F$ with $\bar\alpha_i(0)=p_i$ such that
$|\Uparrow_p^{\bar\alpha_i(t)}\xi|<\delta^n \text{ for all } t\in[0, \frac{|p_ip_{i_n}|}{2}]$.
Note that $z_j$ can be chosen to ensure that $\bar\alpha_i(t)|_{t\in[0,\frac{|p_ip_{i_n}|}{2}]}=\alpha_i(t)|_{t\in[0,\frac{|p_ip_{i_n}|}{2}]}$;
namely, for sufficiently large $i$ we have that
$$|\Uparrow_p^{\alpha_i(t)}\xi|<\delta^n \text{ for all } t\in[0, \frac{|p_ip_{i_n}|}{2}].$$
This implies that $|\Uparrow_p^{\alpha(t)}\xi|<\delta^n \text{ for all } t\in(0, \frac{|pp_{i_n}|}{2}]$,
i.e. $\alpha(t)$ is tangent to $\xi$ at $p$.
\end{proof}
\begin{remark}\label{rem3.7}{\rm
We would like to point out that (\ref{eqn3.8}) and (\ref{eqn3.9}) are inspired by the proof of Proposition 3.3 in \cite{PP1}.
Moreover, in the proof right above, $\alpha_i(t)$ is in fact the beginning part of the gradient curve of ${\operatorname{dist}}_p$ starting from $p_i$,
and $\alpha(t)$ is just the beginning part of the radial curve starting from $p$ with direction $\xi$ (cf. Corollary \ref{coro4.2} below).}
\end{remark}
\section{Gradient curve and its applications}
\subsection{Gradient and radial curves}
In \cite{Pet}, gradient curve has been introduced for semi-concave functions,
distance functions in particular, on an $X\in {\operatorname{Alex}}(k)$. Let $p$ and $q$ be two distinct points in $X$, and let $\alpha(t)|_{t\in[0,+\infty)}$ be a continuous curve in $X$ with $\alpha(0)=q$.
Then $\alpha(t)|_{t\in[0,+\infty)}$ is the gradient curve of ${\operatorname{dist}}_p$ starting from $q$
if and only if, for each $t_0\in [0,+\infty)$, either $\Uparrow_{\alpha(t_0)}^{\alpha(t)}$ converges to $\eta\in\Sigma_{\alpha(t_0)}X$
as $t\to t_0^+$ with
$$|\Uparrow_{\alpha(t_0)}^{p}\eta|=\max\{|\Uparrow_{\alpha(t_0)}^{p}\nu| : \nu\in \Sigma_{\alpha(t_0)}X\}>\frac\pi2$$
(more precisely, it is required that ${\operatorname{dist}}_p'(\alpha(t))|_{t=t_0^+}=-\cos|\Uparrow_{\alpha(t_0)}^{p}\eta|$), or $\alpha(t)\equiv\alpha(t_0)$ for all $t>t_0$ and there is no $\nu\in \Sigma_{\alpha(t_0)}X$ such that $|\Uparrow_{\alpha(t_0)}^{p}\nu|>\frac\pi2$.
Moreover, given $\xi\in\Sigma_pX$ and $\{p_i\}_{i=1}^\infty$ with $p_i\to p$ and $\Uparrow_p^{p_i}\to\xi$
as $i\to\infty$, we can define the radial curve starting from $p$ with direction $\xi$ by
the limit of gradient curves of ${\operatorname{dist}}_p$ starting from $p_i$ (\cite{Pet}).
Consequently, similar to `extremal' case (cf. Sections 3.1 and 4.1 in \cite{Pet}), we can see
the following two properties from Corollary C.
\begin{prop}\label{prop4.1}
Let $F$ be a subset in an $X\in {\operatorname{Alex}}(k)$. Then $F$ is quasi-convex if and only if, for any two distinct points $p,r\in F$, the gradient curve of ${\operatorname{dist}}_p$ starting from $r$ lies in $F$.
\end{prop}
\begin{coro}\label{coro4.2}
Let $F$ be a quasi-convex subset in an $X\in{\operatorname{Alex}}(k)$. Then, for any $p \in F$ and any $\xi \in \Sigma_{p} F$, the radial curve starting from $p$ with direction $\xi$ lies in $F$.
\end{coro}
In Proposition \ref{prop4.1}, there is no need to assume that $F$ is closed in $X$; and for `extremal' case, $p$ can be an arbitrary point in $X$.
As another corollary of Proposition \ref{prop4.1}, we can see that the limit of quasi-convex subsets is also quasi-convex
(refer to Lemma 4.1.3 in \cite{Pet} for `extremal' case).
\begin{coro}\label{coro4.3}
Let $\{X_{n}\}_{n=1}^\infty$ be $m$-dimensional spaces in ${\operatorname{Alex}}(\kappa)$, and let $F_{n}$ be quasi-convex in $X_{n}$. If $X_{n} \stackrel{\text { GH }}{\longrightarrow} X$ with $F_{n} \rightarrow F \subset X$ as $n\to \infty$, then $F$ is also quasi-convex in $X$.
\end{coro}
\begin{proof} By Proposition \ref{prop4.1}, we just need to verify that, for any two distinct points $p,r\in F$, the gradient curve of ${\operatorname{dist}}_p$ starting from $r$ lies in $F$. Let $p_n, r_n\in F_n$ with $p_{n}\to p$ and $r_n\to r$ as $n\to \infty$. It is clear that ${\operatorname{dist}}_{p_n}$ converges to ${\operatorname{dist}}_p$ as $n\to\infty$. A fundamental fact is that the gradient curve
of ${\operatorname{dist}}_{p_n}$ starting from $r_n$ converges to the gradient curve
of ${\operatorname{dist}}_{p}$ starting from $r$ (Lemma 2.1.5 in \cite{Pet}). Due to the quasi-convexity of $F_n$, the gradient curve
of ${\operatorname{dist}}_{p_n}$ starting from $r_n$ lies in $F_n$, so the gradient curve of ${\operatorname{dist}}_p$ starting from $r$ lies in $F$.
\end{proof}
\begin{remark}\label{rem4.4}{\rm If $X_n$ converges to $X$ without collapse (in the Gromov-Hausdorff sense) in Corollary \ref{coro4.3}, then similar to `extremal' case (cf. Section 4.1 in \cite{Pet}) one can
show that $F_{n}$ also converges to $F$ with respect to induced intrinsic metrics from $X_n$ and $X$.}
\end{remark}
In the rest of this section, we will provide three applications of Proposition \ref{prop4.1}.
\subsection{Intersection of two quasi-convex subsets}
Via Proposition \ref{prop4.1}, we can see that the intersection of two quasi-convex subsets is also quasi-convex.
\begin{prop}\label{prop4.5}
Let $F$ and $G$ be two quasi-convex subsets in an $X\in{\operatorname{Alex}}(k)$. Then $F\cap G$ is also quasi-convex in $X$;
moreover, $\Sigma_{p} (F\cap G)=\Sigma_{p} F\cap \Sigma_{p}G$ for any $p\in F\cap G$.
\end{prop}
\begin{proof} Since $F$ and $G$ are quasi-convex in $X$, for any $p,r\in F\cap G$ with $p\neq r$,
the gradient curve of ${\operatorname{dist}}_p$ starting from $r$ belongs to both $F$ and $G$ (by Proposition \ref{prop4.1}), which
implies that $F\cap G$ is also quasi-convex (by Proposition \ref{prop4.1} again).
Next, for any $p\in F\cap G$, we show that $\Sigma_{p} (F\cap G)=\Sigma_{p} F\cap \Sigma_{p}G$.
It is obvious that $\Sigma_{p} (F\cap G)\subseteq \Sigma_{p} F\cap \Sigma_{p}G$. On the other hand,
for any $\xi\in \Sigma_{p} F\cap \Sigma_{p} G$, the radial curve starting from $p$ with direction $\xi$ lies in $F\cap G$
(by Corollary \ref{coro4.2}), which implies that $\Sigma_{p} (F\cap G)\supseteq \Sigma_{p} F\cap \Sigma_{p}G$.
\end{proof}
\begin{remark}\label{rem4.6}{\rm It is true that the union of two extremal subsets is also extremal (\cite{PP1}).
However, in general, the union of two quasi-convex subsets might not be quasi-convex
(e.g., the union of two lines in a plane is not quasi-convex).}
\end{remark}
\begin{remark}\label{rem4.7}{\rm
Let $F$ and $G$ be two extremal subsets in an $X\in{\operatorname{Alex}}(k)$.
Without involving the concept of gradient curve,
\cite{PP1} has proven that both $F\cap G$ and $\overline{F\setminus G}$ are also
extremal in $X$ by showing $\Sigma_{p} (F\cap G)=\Sigma_{p} F\cap \Sigma_{p}G$
and $\Sigma_{p} \overline{F\setminus G}=\overline{\Sigma_{p}F\setminus \Sigma_{p}G}$ firstly. However, we cannot give a proof for Proposition \ref{prop4.5} in such a way
(because the condition of `quasi-convex' is much weaker than `extremal').
Moreover, so far we cannot either show $\Sigma_{p} \overline{F\setminus G}=\overline{\Sigma_{p}F\setminus \Sigma_{p}G}$
or prove that $\overline{F\setminus G}$ is still quasi-convex if
$F$ and $G$ are quasi-convex in $X$ with $\dim(X)\geqslant 3$.}
\end{remark}
\subsection{Quasi-convex subsets in spherical suspensions}
Let $Z\triangleq\{z_1,z_2\}*Y$ with $|z_1z_2|=\pi$ and $Y\in \text{\rm Alex}(1)$ be a spherical suspension
(for details about such suspension structure refer to \cite{BGP}).
As examples of quasi-convex subsets, \cite{SSW} has shown that
if a quasi-convex subset $F$ in $Z$ contains at least two points including $z_1$, then $F=\{z_1,z_2\}*(F\cap Y)$.
In this paper, we provide a short proof and a stronger version of it via Proposition \ref{prop4.1}.
\begin{prop}\label{prop4.8}
Let $Z=\{z_1,z_2\}*Y$ with $|z_1z_2|=\pi$ and $Y\in \text{\rm Alex}(1)$,
and let $F$ be a quasi-convex subset in $Z$ containing at least two points. Then either $F\subseteq Y$, or there is $\bar z_1$, $\bar z_2$ and $\bar Y\in \text{\rm Alex}(1)$ with $|\bar z_1\bar z_2|=\pi$ such that $Z=\{\bar z_1,\bar z_2\}*\bar Y$ and $F=\{\bar z_1,\bar z_2\}*(F\cap \bar Y)$.
\end{prop}
\begin{proof} We first consider a special case where $z_1$ belongs to $F$. As mentioned above,
$F=\{z_1,z_2\}*(F\cap Y)$ in this case. In fact,
for any point $r\in F\setminus\{z_1\}$, $[rz_2]$ is the gradient curve of ${\operatorname{dist}}_{z_1}$ starting from $r$
by the spherical suspension structure of $Z$, and thus has to lie in $F$ by Proposition \ref{prop4.1}. In particular, $z_2\in F$. Similarly, if $r\neq z_2$, then $[rz_1]$ also belongs to $F$;
namely, the minimal geodesic $[z_1z_2]$ passing $r$ belongs to $F$. This implies that $F=\{z_1,z_2\}*(F\cap Y)$.
We now can assume that $z_1\not\in F$, $z_2\not\in F$ and $F\cap(Z\setminus Y)\neq\emptyset$.
Then we can assume that there is $\bar z_1\in F$ and $\bar z_2\in F$ such that $\frac\pi2>|z_1F|=|z_1\bar z_1|\leqslant|z_2\bar z_2|=|z_2F|$. So, if $|\bar z_1\bar z_2|<\pi$, then the spherical suspension structure of $Z$ guarantees that, for any $[\bar z_1\bar z_2]$,
$$|\uparrow_{\bar z_2}^{\bar z_1}\uparrow_{\bar z_2}^{z_1}|<\frac\pi2
\text{ and } |\uparrow_{\bar z_2}^{\bar z_1}\uparrow_{\bar z_2}^{z_2}|=\pi-|\uparrow_{\bar z_2}^{\bar z_1}\uparrow_{\bar z_2}^{z_1}|.$$
However, the quasi-convexity of $F$ implies that $|\uparrow_{\bar z_2}^{\bar z_1}\uparrow_{\bar z_2}^{z_2}|\leqslant\frac\pi2$, a contradiction. Namely, it has to hold that $|\bar z_1\bar z_2|=\pi$ (so $z_i$ and $\bar z_i$ lie in a (great) circle of perimeter $2\pi$). Hence, there is $\bar Y\in \text{\rm Alex}(1)$ such that $Z=\{\bar z_1,\bar z_2\}*\bar Y$; and thus, similar to the special case above, $F=\{\bar z_1,\bar z_2\}*(F\cap \bar Y)$.
\end{proof}
\subsection{Fixed point set of an isometry}
\begin{prop}\label{prop4.9}
Let $X\in \text{\rm Alex}(k)$, and let $F$ be the fixed point set of an isometry on $X$. Then $F$ is quasi-convex in $X$.
\end{prop}
Recall that the fixed point set of an isometry on a complete Riemannian manifold is totally geodesic, while a quasi-convex subset in a complete Riemannian manifold must be totally geodesic (\cite{SSW}).
\begin{proof} We need only to consider the case where $F$ contains at least two points.
Let $p$ and $r$ be arbitrary two distinct points in $F$. By the uniqueness of the gradient curve of ${\operatorname{dist}}_p$ starting from $r$, it must be fixed by the isometry. I.e., the gradient curve of ${\operatorname{dist}}_p$ starting from $r$
belongs to $F$, so $F$ is quasi-convex by Proposition \ref{prop4.1}.
\end{proof}
\begin{remark}\label{rem4.10}{\rm
Let $\gamma$ be the isometry fixing $F$ in Proposition \ref{prop4.9}, and let $p\in F$. Note that there is a naturally induced isometry $\bar \gamma$ on $\Sigma_pX$, and $\Sigma_pF$ belongs to the fixed point set $\bar F$ of $\bar\gamma$.
On the other hand, by the uniqueness of radial curve starting from a point with a fixed direction,
$\gamma$ must fix the radial curve starting from $p$ with any direction $\xi\in\bar F$. Namely, $\Sigma_pF=\bar F$.}
\end{remark}
\begin{remark}\label{rem4.11}{\rm
Let $X\in \text{\rm Alex}(k)$, and let $\Gamma$ be a compact group which acts on $X$ by isometries
with nonempty fixed point set $F$. In \cite{PP1}, it has been shown that $F$
is extremal as a subset of the orbit space $X/\Gamma$ (where a key tool is `strictly convex hull'). Based on this, \cite{SSW} has proven that $F$ is quasi-convex in $X$. Note that we can prove it using the same arguments as in the proof of Proposition \ref{prop4.9}. (We would like to point out that, using the technique of strictly convex hull, one can also see that $\Sigma_pF=\bar F$ in Remark \ref{rem4.10}.)}
\end{remark}
| 79a8c6f2b7657034eec00680f2c2cb01b9994f06 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{Introduction and statement of the main results}
We consider the nonlinear Korteweg-de Vries (KdV) equation in a bounded interval $(0, L)$ equipped with the Dirichlet boundary condition and the Neumann boundary condition on the right:
\begin{equation}\label{KdV-NL}\left\{
\begin{array}{cl}
u_t (t, x) + u_x (t, x) + u_{xxx} (t, x) + u (t,x) u_x (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
u(t, x=0) = u(t, x=L) = u_x(t , x= L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
u(t = 0, \cdot) = u_0 & \mbox{ in } (0, L),
\end{array}\right.
\end{equation}
where $u_0 \in L^2(0, L)$ is the initial data. The KdV equation has been introduced by Boussinesq \cite{1877-Boussinesq} and Korteweg
and de Vries \cite{KdV} as a model for propagation of surface water waves along a
channel.
This equation also furnishes a very useful nonlinear approximation model
including a balance between a weak nonlinearity and weak dispersive effects and has been studied extensively, see e.g.~\cite{Whitham74, Miura76}.
Regarding \eqref{KdV-NL}, Rosier \cite{Rosier97} introduced a set of critical lengths ${\mathcal N}$ defined by
\begin{equation}\label{def-cN}
{\mathcal N} : = \left\{ 2 \pi \sqrt{\frac{k^2 + kl + l^2}{3}}; \, k, l \in \mathbb{N}_*\right\}.
\end{equation}
This set plays an important role in both the decay property of the solution $u$ of \eqref{KdV-NL} and the controllability property of the system associated with \eqref{KdV-NL} where $u_x(t, L)$ is a control instead of $0$.
Let us briefly review the known results on the controllability of \eqref{KdV-NL} where $u_x(t, L)$ is a control:
\begin{equation}\label{KdV-NLC}\left\{
\begin{array}{cl}
u_t (t, x) + u_x (t, x) + u_{xxx} (t, x) + u (t, x) u_x (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
u(t, x=0) = u(t, x=L) = 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
u_x(\cdot , x= L) : \mbox{ is a control}, \\[6pt]
u(t = 0, \cdot) = u_0 & \mbox{ in } (0, L).
\end{array}\right.
\end{equation}
For initial and final datum in $L^2(0, L)$ and controls in $L^2(0, T)$,
Rosier~\cite{Rosier97} proved that system \eqref{KdV-NLC}
is small-time locally controllable around 0 provided that the length $L$ is not critical,
i.e., $L \not \in {\mathcal N}$. To this end, he studied the controllability of the linearized system using the Hilbert
Uniqueness Method and compactness-uniqueness arguments. He also established
that when the length $L$ is critical, i.e., $L \in {\mathcal N}$, the linearized system is not controllable. More precisely, he showed
that there exists a non-trivial finite-dimensional subspace ${\mathcal M}$ ($= {\mathcal M}_L$) of $L^2(0, L)$ such that its orthogonal space in $L^2(0, L) $ is reachable from $0$ whereas ${\mathcal M}$ is not. To tackle the control problem for the critical length $L \in {\mathcal N}$, Coron and Cr\'epeau
introduced the power series expansion method \cite{CC04}. The idea is to take into account the effect of
the nonlinear term $u u_x$ absent in the linearized system. Using this
method, they showed \cite{CC04} (see also \cite[section 8.2]{Coron07}) that system \eqref{KdV-NLC} is
small-time locally controllable if $L = m 2 \pi$ for $m \in \mathbb{N}_*$ satisfying
\begin{equation}
\nexists (k, l) \in \mathbb{N}_* \times \mathbb{N}_* \mbox{ with } k^2 + kl + l^2 = 3 m^2 \mbox{ and }
k \neq l,
\end{equation}
with initial and final datum in $L^2(0, L)$ and controls in $L^2(0, T)$.
In this case, $\dim {\mathcal M} = 1$ and ${\mathcal M}$ is spanned by $1 - \cos x$. Cerpa \cite{Cerpa07}
developed the analysis in \cite{CC04} to prove that \eqref{KdV-NLC} is locally
controllable at \emph{a finite time} in the case $\dim {\mathcal M} = 2$. This corresponds to
the
case where
\[
L = 2 \pi \sqrt{\frac{k^2 + kl + l^2}{3}}
\]
for some $k, \, l \in \mathbb{N}_*$ with $k>l$, and there is no $(m, n) \in \mathbb{N}_* \times \mathbb{N}_*$ with $m>n$
and $m^2 + mn + n^2 = k^2 + kl + l^2$. Later, Cr\'epeau and Cerpa \cite{CC09}
succeeded to extend the ideas in \cite{Cerpa07} to obtain the local
controllability for all other critical lengths at {\it a finite time}. Recently, Coron, Koenig, and Nguyen \cite{CKN-20} prove that when
$(2k + l) / 3 \not \in \mathbb{N}_*$, one cannot achieve the small time local controllability for initial datum in $H^3(0, L)$ and controls in $H^1$ (in time). We also establish the local controllability for finite time of \eqref{KdV-NLC} for some subclass of these pairs $(k, l)$ with initial datum in $H^3(0, L)$ and the controls in $H^1(0, T)$. This is surprising when compared with known results on internal controls for system \eqref{KdV-NL}. It is known, see \cite{CPR15, PVZ02, Pazoto05}, that system \eqref{KdV-NL} is locally controllable using internal controls {\it whenever} the control region contains an {\it arbitrary} open subset of $(0, L)$.
We next discuss the decay property of \eqref{KdV-NL}. Multiplying the equation of $u$ (real) with $u$ and integrating by parts, one obtains
\begin{equation}\label{key-identity}
\int_{0}^L |u(t, x)|^2 \, dx + \int_0^t |u_x(s, 0)|^2 \, ds = \int_{0}^L |u(0, x)|^2 \, dx \mbox{ for all } t > 0.
\end{equation}
As a consequence of \eqref{key-identity}, one has
\begin{equation}\label{key-identity-0}
\int_{0}^L |u(t, x)|^2 \, dx \le \int_{0}^L |u(0, x)|^2 \, dx \mbox{ for all } t > 0.
\end{equation}
In the case $L \not \in {\mathcal N}$, Menzala, Vasconcellos, and Zuazua \cite{PVZ02} proved that the solutions of \eqref{KdV-NL} with small initial datum in $L^2(0, L)$ decay exponentially to 0. Their analysis is based on the exponential decay of the linearized system for which it holds, see \cite[Proposition 3.3]{Rosier97},
\begin{equation}\label{key-identity-1}
\int_0^t |u_x(s, 0)|^2 \, ds \ge c_t \int_{0}^L |u(0, x)|^2 \mbox{ for all } t > 0.
\end{equation}
When a local damping was added, they also obtained the global exponential stability using the multiplier technique, compactness arguments, and the unique continuation for the KdV equations.
Related results on modified nonlinear KdV equations can be found in \cite{RZ06,LP07}. It is known from the work of Rosier \cite{Rosier97} that for $u_0 \in {\mathcal M}$, the solution $u$ of the linearized system satisfies
\begin{equation}\label{key-identity-2}
\int_0^t |u_x(s, 0)|^2 \, ds = 0 \mbox{ for all } t > 0,
\end{equation}
which implies in particular that \eqref{key-identity-1} does not hold for any $t > 0$. The work of Menzala, Vasconcellos, and Zuazua naturally raises the question whether or not the solutions of \eqref{KdV-NL} go to 0 as the time goes to infinity (see \cite[Section 4]{PVZ02} and also \cite[Section 5]{Pazoto05}). Quite recently, progress has been made for this problem. Concerning the decay property of \eqref{KdV-NL} for critical lengths, when $\dim {\mathcal M} = 1$, Chu, Coron, and Shang \cite{CCS15} showed that the solution $u(t, \cdot)$ goes to 0 as $t \to+ \infty$ for all small initial data in $L^2(0, L)$. Moreover, they showed that there exists a constant $C$ depending only on $L$ such that
\begin{equation}\label{CCS15}
\| u(t, \cdot) \|_{L^2(0, L)} \le \frac{C}{\sqrt{t}} \mbox{ for } t > 0.
\end{equation}
It is worth mentioning that the set of $L \in {\mathcal N}$ such that $\dim {\mathcal M} = 1$ is infinite \cite{CC04}. When $k = 2$ and $l = 2$ (the smallest length for which $\dim {\mathcal M} = 2$), Tang, Chu, Sang, and Coron \cite{TCSC18} also established the decay to 0 of the solutions
by establishing an estimate equivalent to \eqref{CCS15} (see \cite[(1.20) in Theorem 1.1]{TCSC18}). The analysis in \cite{CCS15,TCSC18} is based on the center manifold theory in infinite dimensions, see e.g. \cite{HI11}, in particular the work \cite{VMW04}. To this end, the authors showed the existence and smoothness of a center manifold associated with \eqref{KdV-NL}, which have their own interests.
\medskip
In this paper, we show that all solutions of \eqref{KdV-NL} decay to 0 at least with a rate $1/t^{1/2}$ provided their initial data in $L^2(0, L)$ is small enough when $\dim {\mathcal M} = 1$ or when condition \eqref{main-assumption} below holds (this requires in particular that $\dim {\mathcal M}$ is even). Given
a critical length $L$, condition~\eqref{main-assumption} can be checked numerically, a scilab program is given in the appendix (see \Cref{cor1} for a range of validation). Our approach is inspired by the spirit of the power series expansion due to Coron and Cr\'epeau \cite{CC04} and involves the theory of quasi-periodic functions.
Before stating our results, let us introduce some notations associated with the structure of ${\mathcal M}$, see e.g. \cite{Rosier97,CC04,Cerpa14}. Recall that, for each $L \in {\mathcal N}$, there exists exactly $n_L \in \mathbb{N}_*$ pairs $(k_m, l_m) \in \mathbb{N}_* \times \mathbb{N}_*$ ($1 \le m \le n_L$) such that $k_m \ge l_m$, and
\begin{equation}\label{def-L}
L = 2 \pi \sqrt{\frac{k_m^2 + k_m l_m + l_m^2}{3}}.
\end{equation}
For $1 \le m \le n_L$, set
\begin{equation}\label{def-pm}
p_m = p(k_m, l_m) = \frac{(2k_m + l_m)(k_m - l_m)(2 l_m + k_m)}{3 \sqrt{3}(k_m^2 + k_m l_m + l_m^2)^{3/2}},
\end{equation}
and denote
\begin{equation}
{\mathcal P}_L = \Big\{p_m \mbox{ given by } \eqref{def-pm}; 1 \le m \le n_L \Big\}.
\end{equation}
For $L \in {\mathcal N}$ and $1 \le m \le n_L$ with $p_m \neq 0$, let $\sigma_{j, m}$ ($1 \le j \le 3$) be the solutions of
$$
\sigma^3 - 3 (k_m^2 + k_m l_m + l_m^2) \sigma + 2(2 k_m + l_m)(2 l_m + k_m) (k_m - l_m) = 0,
$$
and set, with the convention $\sigma_{j+3, m} = \sigma_{j, m}$ for $j \ge 1$,
\begin{equation}\label{def-sm}
s_m = s(k_m, l_m) : = \sum_{j=1}^3 \sigma_{j, m} (\sigma_{j+2, m} - \sigma_{j+1, m} ) \left( e^{\frac{4 \pi i (k_m - l_m)}{ 3} } e^{2 \pi i \sigma_{j, m}} + e^{- 2 \pi i \sigma_{j, m}}\right).
\end{equation}
We are ready to state the main result of the paper:
\begin{theorem}\label{thm1} Let $L \in {\mathcal N}$. Assume that either $\dim {\mathcal M} = 1$ or
\begin{equation}\label{main-assumption}
p_m \neq 0 \quad \mbox{ and } \quad s_m \neq 0 \quad \mbox{ for all } 1 \le m \le n_L.
\end{equation}
There exists $\varepsilon_0 > 0$ depending only on $L$ such that for all (real) $u_0 \in L^2(0, L)$ with $\| u_0 \|_{L^2(0, L)} \le \varepsilon_0$, the unique solution $u \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$
of \eqref{KdV-NL} satisfies
\begin{equation}\label{thm1-cl1}
\lim_{t \to 0} \| u(t, \cdot) \|_{L^2(0, L)} = 0.
\end{equation}
More precisely, there exists a constant $C$ depending only on $L$ such that, for $t \ge C/ \| u_0 \|_{L^2(0, L)}^2$ and $\| u_0 \|_{L^2(0, L)} \le \varepsilon_0$, it holds
\begin{equation}\label{thm1-cl2}
\|u(t, \cdot) \|_{L^2(0, L) } \le \frac{1}{2} \| u(0, \cdot) \|_{L^2(0, L)}.
\end{equation}
As a consequence, we have
\begin{equation}\label{thm1-cl3}
\| u(t, \cdot) \|_{L^2(0, L)} \le c /t^{1/2} \mbox{ for } t \ge 0,
\end{equation}
for some positive constant $c$ depending only on $L$.
\end{theorem}
\begin{remark} \rm Let $L \in \mathbb{N}$. Condition $p_m \neq 0$ for all $1 \le m \le n_L$ is equivalent to the fact that $\dim {\mathcal M}$ is even, see e.g. \cite{Cerpa14}.
\begin{remark}\rm Note that $s_m$ is a antisymmetric function of $(\sigma_{1, m}, \sigma_{2, m}, \sigma_{3, m})$ and hence the condition \eqref{main-assumption} does not depend on the order of $(\sigma_{1, m}, \sigma_{2, m}, \sigma_{3, m})$.
\end{remark}
\begin{remark} \rm Assume \eqref{main-assumption}. Applying \Cref{thm1}, one derives from \eqref{key-identity-0} that $0$ is (locally) asymptotically stable with respect to $L^2(0, L)$-norm for system \eqref{KdV-NL}.
\end{remark}
\begin{remark} \rm Assume that \eqref{thm1-cl2} holds. By the regularity properties of the KdV equations, one derives that the same rate of decay holds for $t>1$ when $\| \cdot \|_{L^2(0, L) }$ is replaced by $\| \cdot \|_{H^m(0, L) }$ for $m \ge 1$.
\end{remark}
\end{remark}
Condition \eqref{main-assumption} can be checked numerically. For example, using scilab (the program is given in the appendix), we can check $s_m \neq 0$ for all $(k_m, l_m) \in \mathbb{N}_*$ with $1 \le l_m < k_m < 2000$. As a consequence, we have
\begin{corollary}\label{cor1} Let $L \in {\mathcal N}$. Assume that either $\dim {\mathcal M} = 1$ or $1 \le k_m, l_m \le 1000$ for some $1 \le m \le n_L$. Then \eqref{thm1-cl3} holds if $p_m \neq 0$ for all $1 \le m \le n_L$.
\end{corollary}
We thus rediscover the decay results in \cite{CCS15,TCSC18} by a different approach and obtain new results.
\begin{remark} \rm
Concerning \eqref{main-assumption}, we expect that $s_m \neq 0$ holds for all $L \in {\mathcal N}$ but we are not able to show it.
\end{remark}
The optimality of the decay rate $1/ t^{1/2}$ given in \eqref{thm1-cl3} is open. However, we can establish the following result for all critical lengths.
\begin{proposition}\label{pro-opt} Let $L \in {\mathcal N}$. There exists $c > 0$ such that for all $\varepsilon > 0$, there exists $u_0 \in L^2(0, L)$ such that
$$
\| u_0\|_{L^2(0, L)} \le \varepsilon \quad \mbox{ and } \quad \| u(t, \cdot) \|_{L^2(0, L)} \ge c \ln (t+2) /t \mbox{ for some } t > 0.
$$
\end{proposition}
It is natural to ask if the decay holds globally, i.e., without the assumption on the smallness of the initial data. In fact, this cannot hold even for non-critical lengths. More precisely, Doronin and Natali \cite{DN14} showed that there exist (infinite) stationary states of \eqref{KdV-NL} for any length $L$, which is critical or not.
\subsection{Ideas of the analysis and structure of the paper} The key of the analysis of \Cref{thm1} is to (observe and) establish the following fact (see \Cref{lemK}): Let $L \in {\mathcal N}$. Under condition \eqref{main-assumption} or $\dim {\mathcal M} =1$, there exist two constants $T_0>0$ and $C>0$ depending only on $L$ such that for $T \ge T_0$, one has, for all $u_0 \in L^2(0, L)$ with $\| u_0\|_{L^2(0, L)}$ sufficiently small,
\begin{equation}\label{decayK-I}
\| u(T, \cdot) \|_{L^2(0, L)} \le \| u_0 \|_{L^2(0, L)} \Big(1 - C \| u_0 \|_{L^2(0, L)}^2 \Big) \mbox{ for } T \ge T_0,
\end{equation}
where $u$ is the unique solution of \eqref{KdV-NL}.
To get an idea of how to prove \eqref{decayK-I}, let us consider the case $u_0 \in {\mathcal M} \setminus \{0 \}$, which is somehow the worst case. The analysis is inspired by the spirit of the power expansion method \cite{CC04}. Let $\widetilde u_1$ be the unique solution of
\begin{equation} \label{hu1-Int}\left\{
\begin{array}{cl}
\widetilde u_{1, t} (t, x) + \widetilde u_{1, x} (t, x) + \widetilde u_{1, xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\widetilde u_1(t, x=0) = \widetilde u_1(t, x=L) = \widetilde u_{1, x} (t , x= L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
\widetilde u_1(t = 0, \cdot) = u_0/ \varepsilon & \mbox{ in } (0, L),
\end{array}\right.
\end{equation}
with $\varepsilon = \| u_0\|_{L^2(0, L)} > 0$, and let $\widetilde u_2$ be the unique solution of
\begin{equation}\label{hu2-Int}\left\{
\begin{array}{cl}
\widetilde u_{2, t} (t, x) + \widetilde u_{2, x} (t, x) + \widetilde u_{2, xxx} (t, x) + \widetilde u_{1, x} (t, x) \widetilde u_1 (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\widetilde u_2(t, x=0) = \widetilde u_2(t, x=L) = \widetilde u_{2, x} (t , x= L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
\widetilde u_2(t = 0, \cdot) = 0 & \mbox{ in } (0, L).
\end{array}\right.
\end{equation}
By considering the system of $\varepsilon \widetilde u_1 + \varepsilon^2 \widetilde u_2 - u$, we can prove that, for arbitrary $\tau > 0$,
\begin{equation}\label{diff-I}
\| (\varepsilon \widetilde u_1 + \varepsilon^2 \widetilde u_2 - u)_x (\cdot, 0) \|_{L^2(0, \tau)} \le c_\tau \varepsilon^3,
\end{equation}
for some $c_\tau > 0$ depending only on $\tau$ and $L$, provided that $\varepsilon$ is sufficiently small. Since $\widetilde u_{1}(t, \cdot) \in {\mathcal M}$ for all $t > 0$, one can then derive that
$$
\widetilde u_{1, x} (t, 0) = 0 \mbox{ for } t \ge 0.
$$
Thus, if one can show that, for some $\tau_0 > 0$ and for some $c_0> 0$
\begin{equation}\label{cond-hu2-I}
\| \widetilde u_{2, x}(\cdot, 0) \|_{L^2(0, \tau_0)} \ge c_0,
\end{equation}
then from \eqref{diff-I} one has, for $\varepsilon$ small enough,
$$
\| u_{x}(\cdot, 0) \|_{L^2(0, \tau_0)} \ge c_0 \varepsilon^2.
$$
This implies \eqref{decayK-I} with $T_0 = \tau_0$ by \eqref{key-identity}.
To establish \eqref{cond-hu2-I}, we first construct a special solution $W$ of the system
\begin{equation}\label{W-Int}\left\{
\begin{array}{cl}
W_t (t, x) + W_{x} (t, x) + W_{xxx} (t, x) + \widetilde u_{1, x} (t, x) \widetilde u_1 (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
W(t, x=0) = W(t, x=L) = W (t , x= L)= 0 & \mbox{ for } t \in (0, +\infty),
\end{array}\right.
\end{equation}
via a separation-of-variable process. Moreover, we can prove for such a solution $W$ that
\begin{multline}\label{W_x-quasi-I}
\mbox{$W$ is bounded by $\| \widetilde u_1(0, \cdot) \|_{L^2(0, L)}$ up to a positive constant,} \\
\mbox{and $W_{x}(\cdot, 0)$ is a non-trivial quasi-periodic function.}
\end{multline}
The proof of this property is based on some useful observations on $p_m$ and the boundary conditions considered in \eqref{KdV-NL}, and involves some arithmetic arguments. It is in the proof of the existence of $W$ and the second fact of \eqref{W_x-quasi-I} that assumption \eqref{main-assumption} or $\dim {\mathcal M} = 1$ is required. Note that, for all $\delta > 0$, there exists $T_\delta > 0$ such that it holds, for $\tau \ge T_\delta$,
\begin{equation}\label{decay-I}
\| y_x(\cdot, 0) \|_{L^2(\tau, 2 \tau)} \le \delta \| y_0 \|_{L^2(0, L)},
\end{equation}
for all solution $y \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$ of the system
\begin{equation*}\left\{
\begin{array}{cl}
y_t (t, x) + y_x (t, x) + y_{xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
y(t, x=0) = y(t, x=L) = y_x(t , x= L)= 0 & \mbox{ for } t \in (0, +\infty).
\end{array}\right.
\end{equation*}
Combining \eqref{W_x-quasi-I} and \eqref{decay-I}, we can derive \eqref{cond-hu2-I} after applying the theory of quasi-periodic functions, see e.g. \cite{Bohr47}.
\medskip The proof of \Cref{pro-opt} is inspired by the approach which is used to prove \Cref{thm1} and is mentioned above.
\medskip
The paper is organized as follows. The elements for the construction of $W$ are given in \Cref{sect-construction} and the elements for the
proof of \eqref{W_x-quasi-I} are given in \Cref{sect-quasi}. The proof of \Cref{thm1} is given in \Cref{sect-thm1} where
\eqref{decayK-I} is formulated in \Cref{lemK}. The proof of \Cref{pro-opt} is given in \Cref{sect-opt}.
In the appendix, we reproduce a proof of a technical result, which is obtained in \cite{CKN-20}, and provide the scilab code.
\section{Construction of auxiliary functions} \label{sect-construction}
Let us begin with recalling and introducing some useful notations motivated by the structure of ${\mathcal M}$, see e.g. \cite{Rosier97,CC04,Cerpa14}. For $L \in {\mathcal N}$ and for $1 \le m \le n_L$, denote
\begin{equation}\label{def-etam}
\left\{\begin{array}{c}
\displaystyle \eta_{1, m} = - \frac{2 \pi i (2 k_m + l_m) }{3 L },\\[6pt]
\displaystyle \eta_{2, m} = \eta_{1, m} + \frac{2 \pi i }{L} k_m = \frac{2 \pi i (k_m - l_m) }{3 L }, \\[6pt]
\displaystyle \eta_{3, m} = \eta_{2, m} + \frac{2 \pi i }{L} l_m = \frac{2 \pi i (k_m + 2 l_m) }{3 L }.
\end{array}\right.
\end{equation}
Set
\begin{equation}\label{def-psi}
\left\{ \begin{array}{cl}
\psi_m(x) = \sum_{j=1}^3 (\eta_{j+1, m} - \eta_{j, m}) e^{\eta_{j+2, m} x} & \mbox{ for } x \in [0, L], \\[6pt] \Psi_m(t, x) = e^{- i t p_m} \psi_m(x) & \mbox{ for } (t, x) \in \mathbb{R} \times [0, L],
\end{array} \right.
\end{equation}
(recall that $p_m$ is defined in \eqref{def-pm}). It is clear from the definition of $\eta_{j, m}$ in \eqref{def-etam} that
\begin{equation}\label{pro-etam}
e^{\eta_{1, m} L} = e^{\eta_{2, m} L} = e^{\eta_{3, m} L}.
\end{equation}
This property of $\eta_{j,m}$ associated with $L$ is used several times in our analysis.
\begin{remark} \rm One can check that $\eta_{j, m}$ are the solutions of the equation
$$
\lambda^3 + \lambda - i p_m \lambda = 0.
$$
This implies in particular that $p_{m_1} \neq p_{m_2}$ if $(k_{m_1}, l_{m_1}) \neq (k_{m_2}, l_{m_2})$ as observed in \cite{Cerpa07}.
\end{remark}
It is known that $\Psi_m$ is a solution of the linearized KdV system; moreover,
$$
\Psi_{m, x}(\cdot, 0) = 0,
$$
i.e.,
\begin{equation}\label{KdV-Psi}\left\{
\begin{array}{cl}
\Psi_{m, t} (t, x) + \Psi_{m, x} (t, x) + \Psi_{m, xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\Psi_m(t, 0) = \Psi_m(t, L) = \Psi_{m, x} (t , 0) =\Psi_{m, x}(t , L) = 0 & \mbox{ for } t \in (0, +\infty).
\end{array}\right.
\end{equation}
These properties of $\Psi_m$ can be easily checked. It is known that, see e.g. \cite{Cerpa14},
\begin{equation}\label{span-M}
{\mathcal M} = \mbox{span} \Big\{ \big\{ \Re( \psi_m(x)); 1 \le m \le n_L \Big\} \cup \Big\{ \Im( \psi_m(x)); 1 \le m \le n_L \big\} \Big\}.
\end{equation}
Here and in what follows, for a complex number $z$, we denote $\Re z$, $\Im z$, and $\bar z$ its real part, its imaginary part, and its conjugate, respectively.
In this section, we prepare elements to construct the function $W$ mentioned in the introduction. Assume that $u_0 \in {\mathcal M} \setminus \{0\}$ and let $\varepsilon = \| u_0 \|_{L^2(0, L)}$. By \eqref{span-M}, there exists $(\alpha_m)_{m=1}^{n_L} \subset \mathbb{C}$ such that
\begin{equation}\label{tu1-***}
\frac{1}{\varepsilon} u_{0} = \Re \left\{ \sum_{m=1}^{n_L} \alpha_m \Psi_m (0, x) \right\}.
\end{equation}
The function $\widetilde u_1$ defined by \eqref{hu1-Int} is then given by
$$
\widetilde u_1(t, x) = \Re \left\{ \sum_{m=1}^{n_L} \alpha_m \Psi_m (t, x) \right\} = \Re \left\{ \sum_{m=1}^{n_L} \alpha_m e^{- i p_m t} \psi_m (x) \right\}.
$$
Using the fact, for an appropriate complex function $f$,
$$
\Re f(t, x) \Re f_x(t, x) = \frac{1}{2} \Big( (\Re f(t, x) )^2 \Big)_x= \frac{1}{8} \Big( \big( f(t, x)^2)_x + \big( \bar f(t, x) ^2\big)_x + 2 ( |f(t, x)|^2)_x \Big),
$$
we derive from \eqref{def-psi} and \eqref{tu1-***} that
\begin{align}\label{motivation-sect-construction}
\widetilde u_{1, x} (t, x) \widetilde u_1(t, x) = & \frac{1}{8} \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \Big( \alpha_{m_1} \alpha_{m_2} e^{-i (p_{m_1} + p_{m_2}) t} \psi_{m_1} (x) \psi_{m_2} (x) \Big)_x \\[6pt]
& + \frac{1}{8} \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \Big( \overline{ \alpha_{m_1} \alpha_{m_2} e^{-i (p_{m_1} + p_{m_2}) t} \psi_{m_1} (x) \psi_{m_2} (x)} \Big)_x \nonumber \\[6pt]
& + \frac{1}{4} \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \Big( \alpha_{m_1} \bar \alpha_{m_2} e^{-i (p_{m_1} - p_{m_2}) t} \psi_{m_1} (x) \bar \psi_{m_2} (x) \Big)_x. \nonumber
\end{align}
Motivated by \eqref{motivation-sect-construction}, in this section, we construct solutions of system \eqref{pro1-sys-1}-\eqref{pro1-sys-2} and system \eqref{pro1-sys-1-Co}-\eqref{pro1-sys-2-Co} below.
\medskip
We begin with the following simple result whose proof is omitted.
\begin{lemma}\label{lem-der} Let $L \in {\mathcal N}$ and $1 \le m_1, m_2 \le n_L$. We have, in $[0, L]$,
\begin{multline}\label{lem-der-cl1}
\Big(\psi_{m_1} \psi_{m_2} \Big)'(x) \\[6pt]
= \sum_{j=1}^3 \sum_{k=1}^3 (\eta_{j+1, m_1} - \eta_{j, m_1})(\eta_{k+1, m_2} - \eta_{k, m_2}) (\eta_{j+2, m_1} + \eta_{k+2, m_2}) e^{(\eta_{j+2, m_1} + \eta_{k+2, m_2}) x},
\end{multline}
and
\begin{multline}\label{lem-der-cl2}
\Big(\psi_{m_1} \bar \psi_{m_2} \Big)'(x) \\[6pt]
= \sum_{j=1}^3 \sum_{k=1}^3 (\eta_{j+1, m_1} - \eta_{j, m_1})(\bar \eta_{k+1, m_2} - \bar \eta_{k, m_2}) (\eta_{j+2, m_1} + \bar \eta_{k+2, m_2}) e^{(\eta_{j+2, m_1} + \bar \eta_{k+2, m_2}) x} .
\end{multline}
\end{lemma}
We next introduce
\begin{definition}\label{def1}
For $z \in \mathbb{C}$, let $\lambda_j = \lambda_{j} (z)$ $(1 \le j \le 3)$ be the roots of the equation
\begin{equation}\label{def-lambda}
\lambda^3 + \lambda - i z = 0,
\end{equation}
and set
\begin{equation}\label{def-Q}
Q (z) = \left(\begin{array}{ccc}
1 & 1 & 1 \\[6pt]
e^{\lambda_1 L } & e^{\lambda_2 L} & e^{\lambda_3 L} \\[6pt]
\lambda_1 e^{\lambda_1 L} & \lambda_2 e^{\lambda_2 L} & \lambda_3 e^{\lambda_3 L}
\end{array}\right).
\end{equation}
\end{definition}
\begin{remark}\rm
Some comments on the definition of $Q$ are in order.
The matrix $Q$ is antisymmetric with respect
to $\lambda_j$ ($j=1, 2, 3$), and its definitions depend on a choice of the order of $(\lambda_1, \lambda_2, \lambda_3)$. Nevertheless, we later consider either the equation $\det Q = 0$ or a quantity depending on $Q$ in such a way that the order of $(\lambda_1, \lambda_2, \lambda_3)$ does not matter. The definition of $Q$ is only considered in these contexts.
\end{remark}
\begin{remark} \rm The definition of $\lambda_j(z)$ in \Cref{def1} is slightly different from the one given in \cite{CKN-20} where $i z$ is used instead of $-iz$ in \eqref{def-lambda}.
\end{remark}
\begin{remark} \rm \label{rem-pm-lambda} It is known that if $z \in {\mathcal P}_L$ for some $L \in {\mathcal N}$, then
$$
\lambda_j = \eta_{j, m} \mbox{ for some } 1 \le m \le n_L.
$$
Hence, by \eqref{pro-etam},
$$
e^{\lambda_1 L } = e^{\lambda_2 L } = e^{\lambda_3 L }.
$$
\end{remark}
\begin{remark}\label{rem-lambda} \rm Note that \eqref{def-lambda} has simple roots for $z \neq \pm 2/(3 \sqrt{3})$. Thus, a general solution of the equation
$$
y'''(x) + y'(x) - i z y (x) = 0 \mbox{ in } [0, L],
$$
is of the form $\sum_{j=1}^3 a_j e^{\lambda_j(z) x}$ when $z \neq \pm 2/ (3 \sqrt{3})$.
For $z = \pm 2/(3 \sqrt{3})$, equation \eqref{def-lambda} has three roots
$$
\lambda_1 = \mp 2i / \sqrt{3} \quad \mbox{ and } \quad \lambda_2 = \lambda_3 = \pm i/ \sqrt{3}.
$$
\end{remark}
We now recall a useful property of solutions of the equation $\det Q = 0$ which is established in \cite{CKN-20}
(a consequence of \cite[Remark 2.7]{CKN-20}).
\begin{lemma}\label{lem-Q} Let $z \in \mathbb{R}$. Then $\det Q(z) = 0$ if and only if either $z = \pm 2/ \sqrt{3}$ or ($L \in {\mathcal N}$ and $z \in {\mathcal P}_L$). Moreover,
$$
\big\{ \pm 2/ \sqrt{3} \big\} \cap {\mathcal P}_{L} = \emptyset \mbox{ for all } L \in {\mathcal N}.
$$
\end{lemma}
The proof of \Cref{lem-Q} is reproduced in the appendix for the convenience of the reader.
\medskip
Let $L \in {\mathcal N}$ and $1 \le m_1, m_2 \le n_L$. As mentioned above, we are interested in constructing a solution of the system
\begin{equation}\label{pro1-sys-1}
- i (p_{m_1} + p_{m_2}) \varphi_{m_1, m_2}(x) + \varphi_{m_1, m_2}' (x) + \varphi_{m_1, m_2}''' (x) + \Big(\psi_{m_1} \psi_{m_2} \Big)'(x)= 0 \mbox{ in } (0, L),
\end{equation}
and
\begin{equation}\label{pro1-sys-2}
\varphi_{m_1, m_2}(0) = \varphi_{m_1, m_2}(L) = \varphi_{m_1, m_2}'(L)= 0.
\end{equation}
\medskip
We have
\begin{proposition}\label{pro1} Let $L \in {\mathcal N}$ and $1 \le m_1, m_2 \le n_L$. Let $\lambda_j = \lambda_j(p_{m_1} + p_{m_2})$ and $\mathcal Q = Q(i p_{m_1} + i p_{m_2})$ where $\lambda_j$ and $Q$ are defined by \eqref{def-lambda} and \eqref{def-Q}. When $p_{m_1} \neq 0$ and $p_{m_2} \neq 0$, set
\begin{equation}\label{pro1-D}
D = D_{m_1, m_2}= \sum_{j=1}^3 \sum_{k=1}^3 \frac{(\eta_{j+1, m_1} - \eta_{j, m_1})(\eta_{k+1, m_2} - \eta_{k, m_2}) }{3 \eta_{j+2, m_1} \eta_{k+2, m_2}},
\end{equation}
and
\begin{equation}\label{pro1-fm1m2}
\chi_{m_1, m_2}(x) = - \sum_{j=1}^3 \sum_{k=1}^3 \frac{(\eta_{j+1, m_1} - \eta_{j, m_1})(\eta_{k+1, m_2} - \eta_{k, m_2}) }{3 \eta_{j+2, m_1} \eta_{k+2, m_2}} e^{(\eta_{j+2, m_1} + \eta_{k+2, m_2}) x} \mbox{ in } [0, L].
\end{equation}
We have
\begin{enumerate}
\item[1)] Assume that $p_{m_1} \neq 0$, $p_{m_2} \neq 0$, and $p_{m_1} + p_{m_2} \not \in {\mathcal P}_L \cup \big\{ 2 / (3 \sqrt{3}) \big\}$.
The unique solution of system \eqref{pro1-sys-1}-\eqref{pro1-sys-2} is given by
\begin{equation}\label{pro1-wm1m2}
\varphi_{m_1, m_2} (x) = \chi_{m_1, m_2}(x) + \sum_{j=1}^3 a_j e^{\lambda_j x},
\end{equation}
where $(a_1, a_2, a_3) $ is uniquely determined via \eqref{pro1-sys-2}, i.e.,
\begin{equation}\label{pro1-aj}
\mathcal Q (a_1, a_2, a_3)^\mathsf{T} = D(1, e^{(\eta_{1, m_1} + \eta_{1, m_2}) L }, 0)^\mathsf{T}.
\end{equation}
\item[2)] Assume that $p_{m_1} \neq 0$, $p_{m_2} \neq 0$, and $p_{m_1} + p_{m_2} \in {\mathcal P}_L$. A solution of system \eqref{pro1-sys-1}-\eqref{pro1-sys-2}
is given by \eqref{pro1-wm1m2} where $(a_1, a_2, a_3)$ satisfies
\begin{equation}\label{pro1-aj-2}
a_1 + a_2 + a_3 = D \quad \mbox{ and } \quad \lambda_1 a_1 + \lambda_2 a_2 + \lambda_3 a_3 = 0.
\end{equation}
\item[3)] Assume that $p_{m_1} \neq 0$, $p_{m_2} \neq 0$, and $p_{m_1} + p_{m_2} = 2/ (3 \sqrt{3})$. Consider the convention
\begin{equation}\label{lem2-lambda}
\lambda_1 = - 2 i / \sqrt{3} \quad \mbox{ and } \quad \lambda_2 = \lambda_3 = i / \sqrt{3}.
\end{equation}
System \eqref{pro1-sys-1}-\eqref{pro1-sys-2} has a unique solution
given by
\begin{equation}\label{def-wmm-*}
\varphi_{m_1, m_2} (x) = \chi_{m_1, m_2}(x) +
a_1 e^{\lambda_1 x} + (a_2 + a_3 x) e^{\lambda_2 x},
\end{equation}
where $(a_1, a_2, a_3)$ is uniquely determined via \eqref{pro1-sys-2}, i.e.,
\begin{equation}\label{pro1-aj-3}
\mathcal Q_1 (a_1, a_2, a_3)^\mathsf{T} = D(1, e^{(\eta_{1, m_1} + \eta_{1, m_2}) L }, 0)^\mathsf{T},
\end{equation}
where
\begin{equation}\label{def-Qm}
\mathcal Q_1 = \left(\begin{array}{ccc}
1 & 1 & 0 \\[6pt]
e^{\lambda_1 L } & e^{\lambda_2 L} & L e^{\lambda_2 L} \\[6pt]
\lambda_1 e^{\lambda_1 L} & \lambda_2 e^{\lambda_2 L} &(\lambda_2 L + 1) e^{\lambda_2 L}
\end{array}\right).
\end{equation}
\item[4)] Assume that $p_{m_1} = p_{m_2} = 0$ and thus $m_1 = m_2 = m$. A solution of system \eqref{pro1-sys-1}-\eqref{pro1-sys-2} is
\begin{equation}\label{def-wmm-0}
\varphi_{m, m} (x) = 4 \left( L \sin x + \frac{1}{6} - x \sin x - \frac{1}{6} \cos (2x) \right).
\end{equation}
\end{enumerate}
\end{proposition}
\begin{proof} We proceed with the proof of 1), 2), 3), and 4) in Steps 1, 2, 3, and 4 below, respectively.
\medskip
\noindent{\it Step 1}: Proof of 1). Since $\eta = \eta_{j, m}$ $(1 \le j \le 3)$ is a root of the equation
\begin{equation*}
\eta^3 + \eta - ip_m = 0,
\end{equation*}
it follows that
\begin{equation*}
\eta_{j, m_1} \neq - \eta_{k, m_2}
\end{equation*}
(since otherwise $p_{m_1} = - p_{m_2}$ which is impossible), and
\begin{equation*}
(\eta_{j, m_1} + \eta_{k, m_2})^3 + (\eta_{j, m_1} + \eta_{k, m_2}) - i (p_{m_1} + p_{m_2}) = 3 \eta_{j, m_1} \eta_{k, m_2} (\eta_{j,m_1} + \eta_{k, m_2}).
\end{equation*}
Since $p_{m_1} \neq 0$ and $p_{m_2} \neq 0$, we derive from \Cref{lem-der} that $\chi_{m_1, m_2}$ is a solution of \eqref{pro1-sys-1}. Since a general solution of the equation $\xi''' + \xi' = i (p_{m_1} + p_{m_2}) \xi$ is of the form $\sum_{j=1}^3 a_j e^{\lambda_j x}$ by \Cref{rem-lambda},
it follows that
\begin{equation}\label{pro1-gen-sol}
\mbox{ a general solution of \eqref{pro1-sys-1} is of the form $\chi_{m_1, m_2} (x) + \sum_{j=1}^3 a_j e^{\lambda_j x}$}.
\end{equation}
We have
\begin{equation}\label{pro1-pro-chi}
- \chi_{m_1, m_2} (0) = D, \quad - \chi_{m_1, m_2} (L) \mathop{=}^{\eqref{pro-etam}} D e^{(\eta_{1, m_1} + \eta_{1, m_2}) L}, \quad \mbox{ and } \quad - \chi_{m_1, m_2, x} (L) \mathop{=}^{\eqref{pro-etam}} 0.
\end{equation}
It follows that a function of the form $\chi_{m_1, m_2} (x) + \sum_{j=1}^3 a_j e^{\lambda_j x}$ satisfies \eqref{pro1-sys-2} if and only if
\begin{equation*}
\sum_{j=1}^3 a_j = D, \quad \sum_{j=1}^3 a_j e^{\lambda_j L} = D e^{(\eta_{1, m_1} + \eta_{1, m_2}) L}, \quad \sum_{j=1}^3 a_j \lambda_j e^{\lambda_j L} = 0,
\end{equation*}
which is equivalent to \eqref{pro1-aj}.
Since $p_{m_1} + p_{m_2} \not \in {\mathcal P}_{L} \cup \big\{2 /( 3 \sqrt{3}) \big\}$ and $p_{m_1} + p_{m_2} > 0$, it follows from \Cref{lem-Q} that $\det \mathcal Q \neq 0$. Therefore, one obtains 1).
\medskip
\noindent{\it Step 2:} Proof of 2). A solution of \eqref{pro1-sys-1} is of the form $\chi_{m_1, m_2} (x) + \sum_{j=1}^3 a_j e^{\lambda_j x}$.
This function satisfies \eqref{pro1-sys-2} if and only if, by \Cref{rem-pm-lambda} (recall that $p_{m_1} + p_{m_2} \in {\mathcal P}_L$),
\begin{equation*}
\sum_{j=1}^3 a_j = D, \quad e^{\lambda_1 L} \sum_{j=1}^3 a_j \mathop{=}^{\eqref{pro-etam}} D e^{(\eta_{1, m_1} + \eta_{1, m_2}) L}, \quad \sum_{j=1}^3 a_j \lambda_j \mathop{=}^{\eqref{pro-etam}} 0.
\end{equation*}
This system has a solution if
\begin{equation}\label{Step2-key}
e^{\lambda_1 L} = e^{(\eta_{1, m_1} + \eta_{1, m_2}) L},
\end{equation}
and a solution is given
by \eqref{pro1-wm1m2} where $(a_1, a_2, a_3)$ satisfies \eqref{pro1-aj-2}.
It remains to prove \eqref{Step2-key}. Assume, for some $p_{m_3} \in {\mathcal P}_L$, that
\begin{equation}\label{pro1-S2-p}
p_{m_1} + p_{m_2} = p_{m_3}.
\end{equation}
To establish \eqref{Step2-key}, it suffices to prove that, by \eqref{pro-etam} and \Cref{rem-pm-lambda},
$$
e^{( \eta_{2, m_1} + \eta_{2, m_2} ) L } = e^{\eta_{2, m_3} L }
$$
which is equivalent to the fact, by \eqref{def-etam},
\begin{equation}\label{pro1-S2-mod}
\frac{k_{m_3} - l_{m_3}}{3} - \frac{k_{m_1} - l_{m_1}}{3} - \frac{k_{m_2} - l_{m_2}}{3} \in \mathbb{Z}.
\end{equation}
From \eqref{pro1-S2-p} and the definition of $p_m$ in \eqref{def-pm}, we have
\begin{multline}\label{pro1-S2-p1}
(k_{m_3} - l_{m_3}) (2k_{m_3} + l_{m_3}) (2 l_{m_3} + k_{m_3}) \\[6pt]
= (k_{m_1} - l_{m_1}) (2k_{m_1} + l_{m_1}) (2 l_{m_1} + k_{m_1}) + (k_{m_2} - l_{m_2}) (2k_{m_2} + l_{m_2}) (2 l_{m_2} + k_{m_2}).
\end{multline}
Since
$$
(k_{m_j} - l_{m_j}) (2k_{m_j} + l_{m_j}) (2 l_{m_j} + k_{m_j}) = l_{m_j} - k_{m_j} \mod 3,
$$
It follows from \eqref{pro1-S2-p1} that
$$
k_{m_3} - l_{m_3} = k_{m_1} - l_{m_1} + \big( k_{m_2} - l_{m_2} \big) \mod 3,
$$
which yields \eqref{pro1-S2-mod}. The proof of Step 2 is complete.
\medskip
\noindent{\it Step 3:} Proof of 3). A solution of \eqref{pro1-sys-1} is of the form $\chi_{m_1, m_2}(x) + a_1 e^{\lambda_1 x} + (a_2 + a_3 x) e^{\lambda_2 x}$. This function satisfies \eqref{pro1-sys-2} if and only if, by \eqref{pro1-pro-chi},
$$
a_1 + a_2 = D, \quad a_1 e^{\lambda_1 L } + a_2 e^{\lambda_2 L } + a_3 L e^{\lambda_2 L} = De^{(\eta_{1, m_1} + \eta_{1, m_2} ) L},
$$
and
$$
a_1 \lambda_1 e^{\lambda_1 L } + a_2 \lambda_2 e^{\lambda_2 L } + a_3 (\lambda_2 L + 1) e^{\lambda_2 L} = 0,
$$
which is equivalent to \eqref{pro1-aj-3}.
Hence, it suffices to prove that $\mathcal Q_1$ is invertible. Replacing the third row of $\mathcal Q_1$ by itself minus $\lambda_2$ times the second row, we obtain
\begin{equation}
\mathcal Q_2 = \left(\begin{array}{ccc}
1 & 1 & 0 \\[6pt]
e^{\lambda_1 L } & e^{\lambda_2 L} & L e^{\lambda_2 L} \\[6pt]
(\lambda_1 - \lambda_2) e^{\lambda_1 L} & 0 &e^{\lambda_2 L}
\end{array}\right).
\end{equation}
We have
$$
\det \mathcal Q_2 = e^{2 \lambda_2 L} - \big(1 - L (\lambda_1 - \lambda_2) \big) e^{(\lambda_1 + \lambda_2 ) L }.
$$
Using \eqref{lem2-lambda}, we derive that $\det \mathcal Q_2 = 0$ if and only if
$$
e^{3 \lambda_2 L } = 1 + 3 \lambda_2 L.
$$
Since the equation $e^{ix } = 1 + i x$ has only one solution $x = 0$ in the real line, one derives that $\det \mathcal Q_2 \neq 0$. Therefore, $\mathcal Q_1$ is invertible. The proof of Step 3) is complete.
\medskip
\noindent{\it Step 4:} Proof of 4). Since $ p_{m} =0$, it follows that $k_m = l_m$, and $L = 2 \pi k_m$. One then has
$$
\eta_{1, m} = -i, \quad \eta_{2, m} = 0, \quad \eta_{3, m} = i.
$$
It follows from the definition of $\psi_m$ in \eqref{def-psi} that
\begin{equation}\label{psi-complex}
\psi_m(x) = 2 i (\cos x - 1).
\end{equation}
This implies
$$
\big( \psi_m^2(x) \big)_x = 8 (\cos x - 1) \sin x.
$$
A straightforward computation gives the conclusion.
\medskip
The proof of \Cref{pro1} is complete.
\end{proof}
\begin{remark} \rm In the case, $p_{m_1} = 0$ and $p_{m_2} \neq 0$, one cannot construct a solution of \eqref{pro1-sys-1}-\eqref{pro1-sys-2} in general. In fact, one can check that
\begin{multline}
\chi_{m_1, m_2} (x) = - \sum_{j=1, 2} \sum_{k=1}^3 \frac{(\eta_{j+1, m_1} - \eta_{j, m_1})(\eta_{k+1, m_2} - \eta_{k, m_2}) }{3 \eta_{j+2, m_1} \eta_{k+2, m_2}} e^{(\eta_{j+2, m_1} + \eta_{k+2, m_2}) x} \\[6pt]
- \sum_{k=1}^3 \frac{(\eta_{1, m_1} - \eta_{3, m_1})(\eta_{k+1, m_2} - \eta_{k, m_2}) \eta_{k+2, m_2} }{3 {\eta_{k+2, m_2}}^2 + 1} x e^{ \eta_{k+2, m_2} x}
\end{multline}
is a solution of \eqref{pro1-sys-1}. However,
$$
\chi_{m_1, m_2} (0) \neq e^{-\eta_{1, m_2} L} \chi_{m_1, m_2} (L)
$$
since, in general,
$$
\sum_{k=1}^3 \frac{(\eta_{k+1, m_2} - \eta_{k, m_2}) \eta_{k+2, m_2} }{3 {\eta_{k+2, m_2}}^2 + 1} \neq 0.
$$
Hence one cannot find $(a_1, a_2, a_3) \in \mathbb{C}^3$ such that the function $
\chi_{m_1, m_2}(x) + \sum_{j=1}^3 a_j e^{\lambda_j x}$, with $\lambda_j = \lambda_j (p_{m_2})$,
verifies \eqref{pro1-sys-2}.
\end{remark}
Let $L \in {\mathcal N}$ and $1 \le m_1, m_2 \le n_L$. We are next interested in constructing a solution of the system
\begin{equation}\label{pro1-sys-1-Co}
- i (p_{m_1} - p_{m_2}) \phi_{m_1, m_2}(x) + \phi_{m_1, m_2}'(x) + \phi_{m_1, m_2}''' (x) + \Big(\psi_{m_1} \bar \psi_{m_2} \Big)'(x)= 0 \mbox{ in } (0, L),
\end{equation}
and
\begin{equation}\label{pro1-sys-2-Co}
\phi_{m_1, m_2}(0) = \phi_{m_1, m_2}(L) = \phi_{m_1, m_2}'(L)= 0.
\end{equation}
We have
\begin{proposition}\label{pro1-Co} Let $L \in {\mathcal N}$ and $1 \le m_1, m_2 \le n_L$. Let $\widetilde \lambda_j = \lambda_j(p_{m_1} - p_{m_2})$
and $\widetilde \mathcal Q = Q(i p_{m_1} - i p_{m_2})$ where $\lambda_j$ and $Q$ are defined by \eqref{def-lambda} and \eqref{def-Q}. When $p_{m_1} \neq 0$ and $p_{m_2} \neq 0$, set
\begin{equation}\label{pro1-D-Co}
\widetilde D = \widetilde D_{m_1, m_2} = \sum_{j=1}^3 \sum_{k=1}^3 \frac{(\eta_{j+1, m_1} - \eta_{j, m_1})(\bar \eta_{k+1, m_2} - \bar \eta_{k, m_2}) }{3 \eta_{j+2, m_1} \bar \eta_{k+2, m_2}}
\end{equation}
and
\begin{equation}\label{pro1-chi-Co}
\widetilde \chi_{m_1, m_2}(x) = - \sum_{j=1}^3 \sum_{k=1}^3 \frac{(\eta_{j+1, m_1} - \eta_{j, m_1})(\bar \eta_{k+1, m_2} - \bar \eta_{k, m_2}) }{3 \eta_{j+2, m_1} \bar \eta_{k+2, m_2}} e^{(\eta_{j+2, m_1} + \bar \eta_{k+2, m_2}) x} \mbox{ in } [0, L].
\end{equation}
We have
\begin{enumerate}
\item[1)] Assume that $p_{m_1} \neq 0$, $p_{m_2} \neq 0$, $p_{m_1} \neq p_{m_2}$, and $p_{m_1} - p_{m_2} \not \in {\mathcal P}_L $.
The unique solution of system \eqref{pro1-sys-1-Co}-\eqref{pro1-sys-2-Co} is given by
\begin{equation}\label{pro1-wm1m2-Co}
\phi_{m_1, m_2} (x) = \widetilde \chi_{m_1, m_2}(x)
+ \sum_{j=1}^3 a_j e^{\widetilde \lambda_j x},
\end{equation}
where $(a_1, a_2, a_3)$ is uniquely determined via \eqref{pro1-sys-2-Co}, i.e.,
\begin{equation}\label{def-aj-Co}
\widetilde \mathcal Q (a_1, a_2, a_3)^\mathsf{T} = \widetilde D(1, e^{(\eta_{1, m_1} + \bar \eta_{1, m_2}) L }, 0)^\mathsf{T}.
\end{equation}
\item[2)] Assume that $p_{m_1} \neq 0$, $p_{m_2} \neq 0$, $p_{m_1} \neq p_{m_2}$, and $p_{m_1} - p_{m_2} \in {\mathcal P}_L$. A solution of system \eqref{pro1-sys-1-Co}-\eqref{pro1-sys-2-Co}
is given by \eqref{pro1-wm1m2-Co} where $(a_1, a_2, a_3)$ satisfies
\begin{equation}
a_1 + a_2 + a_3 = \widetilde D \quad \mbox{ and } \widetilde \lambda_1 a_1 + \widetilde \lambda_2 a_2 + \widetilde \lambda_3 a_3 = 0.
\end{equation}
\item[3)] Assume that $p_{m_1} = p_{m_2} \neq 0$ and thus $m_1 = m_2 = m$. System \eqref{pro1-sys-1-Co}-\eqref{pro1-sys-2-Co} has a unique solution
\begin{multline*}
\phi_{m, m} (x) = - \sum_{j=1}^3 \sum_{k=1}^3 \frac{(\eta_{j+1, m} - \eta_{j, m})(\bar \eta_{k+1, m} - \bar \eta_{k, m}) }{3 \eta_{j+2, m} \bar \eta_{k+2, m}} e^{(\eta_{j+2, m} + \bar \eta_{k+2, m}) x} \\[6pt]
+ \sum_{j=1}^3 \sum_{k=1}^3 \frac{(\eta_{j+1, m} - \eta_{j, m})(\bar \eta_{k+1, m} - \bar \eta_{k, m}) }{3 \eta_{j+2, m} \bar \eta_{k+2, m}}.
\end{multline*}
\end{enumerate}
\item[4)] Assume that $p_{m_1} = p_{m_2} = 0$ and thus $m_1 = m_2 = m$. A solution of system \eqref{pro1-sys-1-Co}-\eqref{pro1-sys-2-Co} is
\begin{equation}\label{def-wmm-0-Co}
\phi_{m, m} (x) = - 4 \left( L \sin x + \frac{1}{6} - x \sin x - \frac{1}{6} \cos (2x) \right).
\end{equation}
\end{proposition}
\begin{proof} We proceed with the proof of 1), 2), 3), and 4) in Steps 1, 2, 3, and 4 below, respectively.
\medskip
\noindent{\it Step 1}: Proof of 1). The proof is similar tos Step 1 in the proof of \Cref{pro1}. One just notes that
\begin{equation*}
(\eta_{j, m_1} + \bar \eta_{k, m_2})^3 + (\eta_{j, m_1} + \bar \eta_{k, m_2}) - i (p_{m_1} - p_{m_2}) = 3 \eta_{j, m_1} \bar \eta_{k, m_2} (\eta_{j, m_1} + \bar \eta_{k, m_2}),
\end{equation*}
and
$$
\eta_{j, m_1} + \bar \eta_{k, m_2} \neq 0
$$
since $p_{m_1} \neq p_{m_2}$.
\medskip
\noindent{\it Step 2:} Proof of 2). The proof is almost the same as Step 2 in the proof of \Cref{pro1}. The details are omitted.
\medskip
\noindent{\it Step 3:} Proof of 3). One can check that $\phi_{m, m}$ is a solution of \eqref{pro1-sys-1-Co}-\eqref{pro1-sys-2-Co}. The uniqueness follows from the fact that equation \eqref{def-lambda} has simple roots for $z=0$.
\medskip
\noindent{\it Step 4:} Proof of 4). The conclusion is from 4) of \Cref{pro1} by noting that
$$
|\psi_m(x)|^2 \mathop{=}^{\eqref{psi-complex}} - \psi_m(x)^2 \mbox{ if } p_m =0.
$$
\medskip
The proof is complete.
\end{proof}
\section{Properties of auxiliary functions} \label{sect-quasi}
The main goal of this section is to establish, for $L \in {\mathcal N}$ and $1 \le m \le n_L$ with $p_m \neq 0$, that
\begin{equation}\label{der-varphi-0}
\varphi_{m, m}'(0) \neq 0
\end{equation}
provided \eqref{main-assumption} holds (see \Cref{pro-chi}) where $\varphi_{m, m}$ is determined in \Cref{pro1}. We begin with
\begin{lemma} \label{lemE} Let $L \in {\mathcal N}$ and $1 \le m \le n_L$ with $p_m \neq 0$.
Set
\begin{equation}
E_m : = \sum_{j=1}^3 \frac{\eta_{j+1, m} - \eta_{j, m}}{\eta_{j+2, m}}.
\end{equation}
We have
\begin{equation}\label{lemE-cl1}
D_{m, m} = - \chi_{m, m}(0) = \frac{1}{3} E_m^2,
\end{equation}
and
\begin{equation}\label{lemE-cl2}
E_m
= - \frac{2 7 k_m l_m (k_m+l_m)}{(k_m+2l_m) (2 k_m + l_m) (k_m-l_m)} \neq 0.
\end{equation}
\end{lemma}
\begin{proof} It is clear to see from \eqref{pro1-fm1m2} that
$$
D_{m, m} = - \chi_{m, m}(0) = \frac{1}{3} E_m^2.
$$
With the notation $\gamma_{j, m} = L \eta_{j, m}/ (2 \pi i)$, we have
\begin{equation}\label{lem-E-gamma}
\gamma_{1, m} = - \frac{2 k_m + l_m}{3}, \quad \gamma_{2, m} = \frac{k_m -l_m}{3}, \quad \gamma_{3, m} = \frac{k_m+ 2 l_m}{3}.
\end{equation}
It follows that
\begin{equation*}
E_m = \sum_{j=1}^3 \frac{\gamma_{j+1, m} - \gamma_{j, m}}{\gamma_{j+2, m}}
= \frac{3k_m}{k_m + 2 l_m } - \frac{3 l_m }{2 k_m + l_m} - \frac{3 (k_m+l_m)}{k_m -l_m}.
\end{equation*}
Since
\begin{multline*}
k_m(2k_m + l_m) (k_m - l_m) - l_m(k_m + 2 l_m) (k_m - l_m) - (k_m + l_m) (k_m + 2 l_m) (2 k_m + l_m) \\[6pt]
= 2 (k_m^2 - l_m^2) (k_m - l_m) - (k_m + l_m) (k_m + 2 l_m) (2 k_m + l_m) \\[6pt]
= (k_m + l_m) \Big( 2 k_m^2 - 4 k_m l_m + 2 k_m^2 - 2 k_m^2 - 2 l_m^2 - 5 k_m l_m \Big) = - 9 k_m l_m (k_m + l_m),
\end{multline*}
we derive that
\begin{equation*}
E_m
= - \frac{2 7 k_m l_m (k_m+l_m)}{(k_m+2l_m) (2 k_m + l_m) (k_m-l_m)} \neq 0.
\end{equation*}
The proof is complete.
\end{proof}
We next show in \Cref{lem-2/3,lem-2pm} below that for $L \in {\mathcal N}$ and for $1 \le m \le n_L$ with $p_m \neq 0$, it holds
$$
2 p_m \neq 2 / (3 \sqrt{3}) \quad \mbox{ and } \quad 2 p_m \not \in {\mathcal P}_L.
$$
As a consequence $\varphi_{m, m}$ is constructed via 1) and 4) in \Cref{pro1}. We begin with
\begin{lemma}\label{lem-2/3} Let $L \in {\mathcal N}$ and $1 \le m \le n_L$. Then
$$
2 p_m \neq 2 / (3 \sqrt{3}).
$$
\end{lemma}
\begin{proof} We first claim that there is no $k, l \in \mathbb{N}_*$ with $k \ge l $ such that
\begin{equation}\label{lem-2/3-p1}
(2k + l) (2 l + k) (k-l) = (k^2 + l^2 + kl)^{3/2}.
\end{equation}
We prove this by contradiction. Assume that there exists such a pair $(k, l)$. Set
$$
H = \Big\{(k, l) \in \mathbb{N}_*\times \mathbb{N}_*, \; k \ge l, \mbox{ and \eqref{lem-2/3-p1} holds}\Big\}.
$$
Set
$$
h = \min \Big\{ k + l; (k, l) \in H \Big\} > 0.
$$
Fix $(k, l) \in H$ such that $k + l = h$. Since
$$
(2k + l) (2 l + k) (k-l) \mbox{ is even},
$$
it follows from \eqref{lem-2/3-p1} that $k^2 + l^2 + kl $ is even. Hence both $k$ and $l$ are even. We write $k = 2 k_1$ and $l = 2 l_1$ for some $k_1, l_1 \in \mathbb{N}_*$. It is clear that
$$
k_1 \ge l_1,
$$
and
\begin{equation*}
(2k_1 + l_1) (2 l_1 + k_1) (k_1-l_1) = (k_1^2 + l_1^2 + k_1l_1)^{3/2}.
\end{equation*}
This implies
$$
(k_1, l_1) \in H.
$$
We have
$$
k_1 + l_1 = (k+ l)/ 2 = h/2 \quad \mbox{ and } \quad h > 0.
$$
This contradicts the definition of $h$. The claim is proved.
\medskip
We are ready to derive the conclusion of \Cref{lem-2/3}. Since $2 p_m = 2 / (3 \sqrt{3})$ for some $1 \le m \le n_L$ and for some $L \in {\mathcal N}$ if and only if, by the definition of $p_m$ in \eqref{def-pm},
$$
(2k_m + l_m)(k_m - l_m)(2 l_m + k_m) = (k_m^2 + l_m^2 + k_m l_m)^{3/2},
$$
the conclusion follows from the claim.
\end{proof}
We next prove
\begin{lemma}\label{lem-2pm} There is no quadruple $(k_1, l_1, k_2, l_2) \in \mathbb{N}_*^4$ satisfying the system
\begin{equation}\label{sys-kl}
\left\{\begin{array}{c}
k_1 > l_1, \quad k_2 > l_2, \\[6pt]
k_1^2 + k_1 l_1 + l_1^2 = k_2^2 + k_2 l_2 + l_2^2, \\[6pt]
(2k_2 + l_2) (2 l_2 + k_2) (k_2 - l_2) = 2 (2k_1 + l_1) (2 l_1 + k_1) (k_1 - l_1).
\end{array}\right.
\end{equation}
Consequently, for $L \in {\mathcal N}$ and $1 \le m \le n_L$, we have
\begin{equation}\label{lem-2pm-cl2}
2 p_m \not \in {\mathcal P}_L \mbox{ if } p_m \neq 0.
\end{equation}
\end{lemma}
\begin{proof} We prove the non-existence by contradiction. Assume that there exists a quadruple $(k_1, l_1, k_2, l_2) \in \mathbb{N}_*^4$ satisfying \eqref{sys-kl}.
Set
\begin{equation}
G = \Big\{ (k_1, l_1, k_2, l_2) \in \mathbb{N}_*^4; \eqref{sys-kl} \mbox{ holds} \Big\},
\end{equation}
and let
\begin{equation}
g = \min \Big\{ k_1 + l_1 + k_2 + l_2; (k_1, l_1, k_2, l_2) \in G \Big\} > 0.
\end{equation}
Fix $(k_1, l_1, k_2, l_2) \in G$ such that $k_1 + l_1 + k_2 + l_2 = g$. Set
\begin{equation}\label{lem-2pm-A}
A := k_1^2 + k_1 l_1 + l_1^2 = k_2^2 + k_2 l_2 + l_2^2 \quad (\mbox{by the second line of \eqref{sys-kl}}).
\end{equation}
Since, for $(k, l) \in \mathbb{R}$,
$$
(2k + l)(2l + k) = 2 (k^2 + kl + l^2) + 3 kl \quad \mbox{ and } \quad (k-l)^2 = (k^2 + k l + l^2) - 3 k l,
$$
it follows from the square of the last line of \eqref{sys-kl}, with
\begin{equation}\label{def-x1x2}
x_1 = 3 k_1 l_1 \quad \mbox{ and } \quad x_2 = 3 k_2 l_2,
\end{equation}
that
$$
(2 A + x_2)^2 (A - x_2) = 4 (2 A + x_1)^2 (A - x_1).
$$
This implies
\begin{equation}
(4 A^3 - 3A x_2^2 - x_2^3) = 4(4 A^3 - 3A x_1^2 - x_1^3),
\end{equation}
or equivalently
\begin{equation}\label{lem-2pm-p1}
12 A^3 = 3 A (4 x_1^2 - x_2^2) + 4 x_1^3 - x_2^3.
\end{equation}
Using \eqref{def-x1x2}, we derive that $A^3 = 0 \mod 3$, which yields
$$
A = 0 \mod 3.
$$
Putting this information into \eqref{lem-2pm-p1} and using again \eqref{def-x1x2}, we obtain
$$
x_1^3 - x_2^3 = 0 \mod 3^4.
$$
We deduce from \eqref{def-x1x2} that
\begin{equation}\label{lem-2pm-p2}
(k_1 l_1)^3 - (k_2 l_2)^3 = 0 \mod 3.
\end{equation}
By writing $k_1 l_1$ under the form $k_2 l_2 + 3 q + r$ with $q \in \mathbb{Z}$ and $r \in \mathbb{N}$ with $0 \le r \le 2$, we have
\begin{equation}\label{lem-2pm-p3}
(k_1 l_1)^3 - (k_2 l_2)^3 = 3 k_2^2 l_2^2 (3 q + r) + 3 k_2 l_2 (3 q + r)^2 + (3 q + r)^3.
\end{equation}
Combining \eqref{lem-2pm-p2} and \eqref{lem-2pm-p3} yields that $r = 0$. Putting this information into \eqref{lem-2pm-p1}, we obtain
$$
A^3 = 0 \mod 3^4.
$$
This implies
$$
A = 0 \mod 9.
$$
We deduce from \eqref{lem-2pm-A} that
$$
k_1 = 0 \mod 3, \quad l_1 = 0 \mod 3, \quad k_2 = 0 \mod 3, \quad l_2 = 0 \mod 3.
$$
Let $\hat k_1, \hat l_1, \hat k_2, \hat l_2 \in \mathbb{N}_*$ be such that
$$
k_1 = 3 \hat k_1, \quad l_1 = 3 \hat l_1, \quad k_2 = 3 \hat k_2, \quad l_2 = 3 \hat l_2.
$$
One can easily check that $(\hat k_1, \hat l_1, \hat k_2, \hat l_2) \in G$ and
$$
\hat k_1 + \hat l_1 + \hat k_2 + \hat l_2 = g/3 < g.
$$
We obtain a contradiction. The non-existence associated with \eqref{sys-kl} is proved.
\medskip
It is clear that \eqref{lem-2pm-cl2} is just a consequence of the non-existence by the definition of $L$ and $p_m$ as a function of $k_m$ and $l_m$ in \eqref{def-L} and \eqref{def-pm}. The proof is complete.
\end{proof}
We are ready to state and prove the main result of this section:
\begin{proposition}\label{pro-chi} Let $L \in {\mathcal N}$ and $1 \le m \le n_L$. Then
\begin{equation}\label{pro-chi-cl1}
\varphi_{m, m}' (0) = 4 \pi L = - \phi_{m, m}'(0) \mbox{ if } p_m = 0,
\end{equation}
and, if $p_m \neq 0$ and $s_m \neq 0$ then
\begin{equation}\label{pro-chi-cl2}
\varphi_{m, m}' (0) \neq 0.
\end{equation}
\end{proposition}
\begin{proof} Assertion \eqref{pro-chi-cl1} follows immediately from 4) of \Cref{pro1,pro1-Co}. We next consider the case $p_m \neq 0$. By \Cref{lem-2/3,lem-2pm},
we have
$$
\varphi_{m, m}'(0) = 0
$$
only if, with $\alpha = e^{2 \eta_{2, m} L}$ and $\lambda_j = \lambda_j (2 p_m)$,
\begin{equation}
\left\{\begin{array}{c}
\sum_{j=1}^3 \lambda_j a_j = 0 \quad (= \varphi_{m, m}' (0) \mbox{ since $\chi_{m, m}'(0) = 0$}), \\[6pt]
\sum_{j=1}^3 \lambda_j e^{\lambda_j L} a_j = 0 \quad (= \varphi_{m, m}'(L) \mbox{ since $\chi_{m, m}'(L) = 0$} ), \\[6pt]
\sum_{j=1}^3 (e^{\lambda_j L} - \alpha ) a_j = 0 \quad (= - \chi_{m, m}(L) + \alpha \chi_{m, m}(0) \mbox{ since $\chi_{m, m}(L) = \alpha \chi_{m, m } (0)$} ).
\end{array}\right.
\end{equation}
Since $E_m \neq 0$ by \Cref{lemE}, one has a non-trivial solution $(a_1, a_2, a_3)$ of this system. This implies
\begin{equation}\label{pro-chi-p1}
\det K_1 = 0 \quad \mbox{ where } K_1 : = \quad \left(\begin{array}{ccc}
\lambda_1 & \lambda_2 & \lambda_3 \\[6pt]
\lambda_1 e^{\lambda_1 L} & \lambda_2 e^{\lambda_2 L} &
\lambda_3 e^{\lambda_3 L} \\[6pt]
e^{\lambda_1 L} - \alpha & e^{\lambda_2 L} - \alpha &
e^{\lambda_3 L} - \alpha
\end{array}\right).
\end{equation}
Set
$$
\hat \lambda_j = \lambda_j L.
$$
Condition \eqref{pro-chi-p1} is equivalent to
\begin{equation}\label{pro-chi-p2}
\det K_2 = 0 \quad \mbox{ where } \quad K_2 : = \left(\begin{array}{ccc}
\hat \lambda_1 & \hat \lambda_2 & \hat \lambda_3 \\[6pt]
\hat \lambda_1 e^{\hat \lambda_1} & \hat \lambda_2 e^{\hat \lambda_2} &
\hat \lambda_3 e^{\hat \lambda_3} \\[6pt]
e^{\hat \lambda_1} - \alpha & e^{\hat \lambda_2} - \alpha &
e^{\hat \lambda_3} - \alpha
\end{array}\right).
\end{equation}
A computation yields
\begin{equation*}
\det K_2 = \sum_{j=1}^3 \hat \lambda_j \Big( ( \hat \lambda_{j+1} - \hat \lambda_{j+2}) e^{\hat \lambda_{j+1} + \hat \lambda_{j+2}} - \alpha (\hat \lambda_{j+1}e^{\hat \lambda_{j+1}} - \hat \lambda_{j+2}e^{\hat \lambda_{j+2}} ) \Big),
\end{equation*}
which implies
\begin{equation}\label{det-K2}
\det K_2 = \sum_{j=1}^3 \hat \lambda_j ( \hat \lambda_{j+1} - \hat \lambda_{j+2}) \Big( e^{- \hat \lambda_{j}} + \alpha e^{\hat \lambda_{j}} \Big).
\end{equation}
Here we used the fact $\sum_{j=1}^3 \hat \lambda_j = L \sum_{j=1}^3 \lambda_j =0 $. From the definition of $\lambda_j = \lambda_{j} (2 p_m)$ given in \Cref{def1}, we have
\begin{equation*}
\left\{\begin{array}{c}
\hat \lambda_1 + \hat \lambda_2 + \hat \lambda_3 = 0, \\[6pt]
\hat \lambda_1 \hat \lambda_2+ \hat \lambda_1 \hat \lambda_3 + \hat \lambda_2 \hat \lambda_3 = L^2, \\[6pt]
\hat \lambda_1 \hat \lambda_2 \hat \lambda_3 = 2 i p_m L^3.
\end{array}\right.
\end{equation*}
Define $\sigma_{j, m}$ by
$$
\hat \lambda_{j} = \frac{2 \pi i \sigma_{j, m}}{3}.
$$
We then have
\begin{equation*}
\left\{\begin{array}{c}
\sigma_{1, m} + \sigma_{2, m} + \sigma_{3, m} = 0, \\[6pt]
\sigma_{1, m} \sigma_{2, m} + \sigma_{1, m} \sigma_{3, m} + \sigma_{2, m} \sigma_{3, m} = - 3 (k_m^2 + l_m^2 + k_m l_m), \\[6pt]
\sigma_{1, m} \sigma_{2, m} \sigma_{3, m} = - 2 (2k_m + l_m) (2 l_m + k_m) (k_m - l_m),
\end{array}\right.
\end{equation*}
where in the last identity, we used the fact
$$
p_m L^3 = \frac{1}{27} (2 \pi)^3 (2k_m + l_m) (2 l_m + k_m) (k_m - l_m).
$$
It is clear that $\det K_2 = 0$ if and only if \eqref{main-assumption} holds. The proof is complete.
\end{proof}
\section{Useful properties related to quasi-periodic functions}
In this section, we derive some properties for $W_x(\cdot, 0)$ given in the introduction using the quasi-periodic-function theory.
The main result of this section is \Cref{pro-quasi}.
We begin with its weaker version.
\begin{lemma}\label{lem-quasi} Let $\ell \in \mathbb{N}_*$, $a_j \in \mathbb{C}$, $q_j \ge 0$ for $1 \le j \le \ell$, and
$M_{j_1, j_2}, N_{j_1, j_2} \in \mathbb{C}$ with $1 \le j_1, j_2 \le \ell$.
Assume that
\begin{equation}\label{lem-quasi-A1}
\left\{\begin{array}{c}
q_{j_1} \neq q_{j_2} \mbox{ for } 1 \le j_1 \neq j_2 \le \ell, \\[6pt]
M_{j, j} \neq 0 \mbox{ for } 1 \le j \le \ell, \\[6pt]
(M_{j, j} \mbox{ is real and } N_{j, j} \neq 0) \mbox{ if } q_j = 0, \\[6pt]
a_{j} \in i \mathbb{R} \mbox{ if } q_j =0,
\end{array}\right.
\end{equation}
and
\begin{equation}\label{lem-quasi-A2}
\sum_{j=1}^\ell |a_{j}|^2 > 0.
\end{equation}
Set, for $t \in \mathbb{R}$,
\begin{multline}\label{lem-quasi-g}
g(t) \\[6pt]:= \sum_{j_1 = 1}^\ell \sum_{j_2 = 1}^\ell \Big( a_{j_1} a_{j_2} M_{j_1, j_2} e^{- i (q_{j_1} + q_{j_2}) t} + \bar a_{j_1} \bar a_{j_2} \bar M_{j_1, j_2} e^{ i (q_{j_1} + q_{j_2}) t} + 2 a_{j_1} \bar a_{j_2} N_{j_1, j_2} e^{-i (q_{j_1} - q_{j_2})} \Big).
\end{multline}
There exists $t \in \mathbb{R}_+$ such that
\begin{equation}\label{lem-quasi-cl}
g(t) \neq 0.
\end{equation}
\end{lemma}
\begin{proof} We prove \eqref{lem-quasi-cl} by recurrence in $\ell$. It is clear that the conclusion holds for $\ell=1$. Indeed, if $q_1 \neq 0$ then since $e^{2 q_1 t}$, $0$, and $e^{-2 q_1 t}$ are independent, the conclusion follows. Otherwise, $q_1 = 0$.
Since $M_{1, 1}$ is real and $a_{1} \in i \mathbb{R}$, we have
$$
g(t) = 2 |a_1|^2 N_{1, 1}.
$$
The conclusion in the case $\ell = 1$ follows since $N_{1, 1} \neq 0$.
Assume that the conclusion holds for $\ell \ge 1$, we prove that the conclusion holds for $\ell +1$.
Without loss of generality, one might assume that
\begin{equation}\label{lem-quasi-q}
0 \le q_1 < q_2 < \dots < q_{\ell} < q_{\ell +1}.
\end{equation}
We will prove \eqref{lem-quasi-cl} for $\ell+1$ by contradiction. Assume that there exist $a_j$ and $q_j \ge 0$ with $1 \le j \le \ell +1$, $M_{j_1, j_2}, \, N_{j_1, j_2} \in \mathbb{C}$ with $1 \le j_1, j_2 \le \ell+1$ such that \eqref{lem-quasi-A1}, \eqref{lem-quasi-A2}, and \eqref{lem-quasi-q} hold, and, for all $t \in \mathbb{R}_+$,
\begin{equation}\label{lem-quasi-p1}
\sum_{j_1 = 1}^{\ell+1} \sum_{j_2 = 1}^{\ell + 1} \Big( a_{j_1} a_{j_2} M_{j_1, j_2} e^{- i (q_{j_1} + q_{j_2}) t} + \bar a_{j_1} \bar a_{j_2} \bar M_{j_1, j_2} e^{ i (q_{j_1} + q_{j_2}) t} + 2 a_{j_1} \bar a_{j_2} N_{j_1, j_2} e^{-i (q_{j_1} - q_{j_2})} \Big) = 0.
\end{equation}
Since the function $e^{- 2 i q_{\ell+1} t}$ defined in $\mathbb{R}_+$ does not belong to the space
\begin{multline*}
\mbox{span} \left( \Big \{e^{-it (q_{j_1} + q_{j_2})}; 1 \le j_1 \le \ell +1; 1 \le j_2 \le \ell \Big\}, \right. \\[6pt]
\left. \Big \{e^{it (q_{j_1} + q_{j_2})}; 1 \le j_1 \le \ell +1; 1 \le j_2 \le \ell + 1 \Big\}, \right. \\[6pt]
\left. \Big \{e^{-it (q_{j_1} - q_{j_2})}; 1 \le j_1 \le \ell +1; 1 \le j_2 \le \ell +1 \Big\}
\right),
\end{multline*}
for $t \in \mathbb{R}_+$ by \eqref{lem-quasi-q}, we have
$$
a_{\ell+1}^2 M_{\ell +1, \ell +1} = 0.
$$
This yields, since $M_{\ell+1, \ell+1} \neq 0$,
$$
a_{\ell+1} = 0.
$$
It follows from \eqref{lem-quasi-p1} that
\begin{equation}
\sum_{j_1 = 1}^{\ell} \sum_{j_2 = 1}^{\ell} \Big( a_{j_1} a_{j_2} M_{j_1, j_2} e^{- i (q_{j_1} + q_{j_2}) t} + \bar a_{j_1} \bar a_{j_2} \bar M_{j_1, j_2} e^{ i (q_{j_1} + q_{j_2}) t} + 2 a_{j_1} \bar a_{j_2} N_{j_1, j_2} e^{-i (q_{j_1} - q_{j_2})} \Big) = 0.
\end{equation}
We now can use the assumption on the recurrence to obtain a contradiction. The proof of \eqref{lem-quasi-cl} is complete.
\end{proof}
Using \Cref{lem-quasi} and the theory of quasi-periodic functions, see e.g. \cite{Bohr47}, we can derive the following useful result for the proof of \Cref{thm1}.
\begin{proposition}\label{pro-quasi}
Let $\ell \in \mathbb{N}_*$, $a_j \in \mathbb{C}$, $q_j \ge 0$ for $1 \le j \le \ell$, and
$M_{j_1, j_2}, N_{j_1, j_2} \in \mathbb{C}$ with $1 \le j_1, j_2 \le \ell$.
Assume that \eqref{lem-quasi-A1} holds and denote $g$ by \eqref{lem-quasi-g}. For all $0 < \gamma_1 < \gamma_2 $ there exist $\gamma_0>0$ and $\tau_0 > 0$ depending only on $\gamma_1$, $\gamma_2$, $\ell$, $q_j$, $M_{j_1, j_2}$, and $N_{j_1, j_2}$ such that if
\begin{equation}
\gamma_1 \le \sum_{j=1}^\ell |a_j|^2 \le \gamma_2,
\end{equation}
then
\begin{equation}\label{pro-cl}
\| g \|_{L^2(\tau, 2 \tau)} \ge \gamma_0 \mbox{ for all } \tau \ge \tau_0.
\end{equation}
\end{proposition}
\begin{proof} Instead of \eqref{pro-cl}, it suffices to prove
\begin{equation}\label{pro-cl1}
\| g \|_{L^\infty(\tau, 2 \tau)} \ge \gamma_0 \mbox{ for } \tau \ge \tau_0
\end{equation}
by contradiction since $|g'(t)| \le C$ in $\mathbb{R}$.
Assume that for all $n \in \mathbb{N}_*$ there exist $(a_{j, n})_{j=1}^\ell \subset \mathbb{C}$ and $(t_n) \subset \mathbb{R}$ such that $\gamma_1 \le \sum_{j=1}^\ell |a_{j, n}|^2 \le \gamma_2$, $t_n \ge n$, and
\begin{equation}\label{pro-p1}
\|g_n\|_{L^\infty(t_n, 2t_n)} \le 1/n,
\end{equation}
where $g_n$ is defined in \eqref{lem-quasi-g} where $a_{j_1}$ and $a_{j_2}$ are replaced by $a_{j_1, n}$ and $a_{j_2, n}$. Without loss of generality, one might assume that
$$
\lim_{n \to + \infty} a_{j, n} = a_j \in \mathbb{C}
$$
and $\gamma_1 \le \sum_{j=1}^N |a_j|^2 \le \gamma_2$. Consider $g$ defined by \eqref{lem-quasi-g} with these $a_j$. We have
\begin{equation}\label{pro-p2}
\lim_{n \to + \infty} \| g_n - g\|_{L^\infty(\mathbb{R})} = 0.
\end{equation}
Since $g$ is an almost-periodic function with respect to $t$ (see e.g. \cite[Corollary on page 38]{Bohr47}), it follows from the definition of almost-periodic functions, see e.g. \cite[Section 44 on pages 32 and 33]{Bohr47}, that for every $\varepsilon > 0$, there exists ${\mathcal L}_\varepsilon > 0$
such that every interval $(\alpha, \alpha + {\mathcal L}_\varepsilon)$ containing a number $\tau(\varepsilon, \alpha)$ for which it holds
\begin{equation}\label{pro-ap}
|g(t + \tau(\varepsilon, \alpha) ) - g(t)| \le \varepsilon \mbox{ for all } t \in \mathbb{R}.
\end{equation}
The proof is now divided into two cases.
\medskip
\noindent{\it Case 1:} $\liminf_{\varepsilon \to 0 } {\mathcal L}_\varepsilon < + \infty$. Denote ${\mathcal L}_0= \liminf_{\varepsilon \to 0} {\mathcal L}_\varepsilon$.
We claim that $g$ is $T$-periodic for some period $T \le {\mathcal L}_0 + 1$. Indeed, by \eqref{pro-ap} applied with $\alpha = 1/2$, there exists a sequence $(\tau_n) \subset (1/2, {\mathcal L}_0 +1)$ such that, for large $n$,
$$
|g(t+ \tau_n) - g(t)| \le 1/n \mbox{ for all } t \in \mathbb{R}.
$$
By choosing $T = \liminf_{n \to + \infty} \tau_n$, we have
$$
g(t + T) = g(t) \mbox{ for all } t \in \mathbb{R}.
$$
The claim is proved.
Since $g$ is $T$-periodic, we have
$$
\|g \|_{L^\infty(t_n, t_n+ T + 1)} = \|g \|_{L^\infty(0, T + 1)} \mbox{ for } n \in \mathbb{N}_*,
$$
and since $g$ is analytic and $g \neq 0$ by \Cref{lem-quasi}, we obtain
$$
\|g \|_{L^\infty(0, T + 1)} > 0.
$$
This contradicts \eqref{pro-p1} and \eqref{pro-p2}. The proof of Case 1 is complete.
\medskip
\noindent{\it Case 2:} $\lim_{\varepsilon \to 0}L_\varepsilon = + \infty$. Set
\begin{equation}\label{pro-quasi-rho}
\rho = \|g \|_{L^\infty(0, 1)}.
\end{equation}
It follows from \Cref{lem-quasi} that $g$ is not identically equal to 0. Since $g$ is analytic, we derive that
\begin{equation}\label{pro-quasi-rho>0}
\rho > 0.
\end{equation}
Let $n_0 \ge 2$ be such that
$$
\|g_n - g \|_{L^\infty(\mathbb{R})} < \rho/4, \quad \| g_n\|_{L^\infty(t_n, 2 t_n)} < \rho / 4 \mbox{ for } n \ge n_0.
$$
Such an $n_0$ exists by \eqref{pro-p1}, \eqref{pro-p2}, and \eqref{pro-quasi-rho>0}. We have, for $n \ge n_0$,
\begin{equation}\label{pro-p3}
\|g \|_{L^\infty(t_n, 2 t_n)} \le \|g_n - g\|_{L^\infty(t_n, 2 t_n)} + \|g_n\|_{L^\infty(t_n, 2 t_n)} \le \rho/4 + \rho/4 = \rho/2.
\end{equation}
Fix $0< \varepsilon < \rho/4$ and fix $n \ge n_0$ such that $1 \le {\mathcal L}_\varepsilon \le t_n/2$. Such a number $n$ exists since $t_n \ge n$.
It follows from the definition of $\tau( \varepsilon, t_n)$ that
\begin{equation}\label{pro-tau1}
\tau(\varepsilon, t_n) \in (t_n, t_n + {\mathcal L}_\varepsilon) \subset (t_n, 3 t_n/2),
\end{equation}
and
\begin{equation}\label{pro-tau2}
\big|g\big(t + \tau(\varepsilon, t_n)\big) - g(t)\big| \le \varepsilon \mbox{ for all } t \in \mathbb{R}.
\end{equation}
This yields
\begin{equation}\label{pro-p4}
\|g \|_{L^\infty(t_n, 2 t_n)} \mathop{\ge}^{\eqref{pro-tau1}}
\|g \|_{L^\infty(\tau(\varepsilon, t_n), \tau(\varepsilon, t_n) + 1)} \mathop{\ge}^{\eqref{pro-tau2}}
\|g \|_{L^\infty(0, 1)} - \varepsilon \ge \rho - \rho/4 = 3 \rho/4.
\end{equation}
Combining \eqref{pro-p3} and \eqref{pro-p4} yields a contradiction since $\rho > 0$ by \eqref{pro-quasi-rho>0}. The proof of Case 2 is complete.
\end{proof}
\section{An upper bound for the decay rate - Proof of \Cref{thm1}} \label{sect-thm1}
This section containing two subsections is devoted to the proof of \Cref{thm1}. The main ingredient is given in the first section and the proof is presented in the second one.
\subsection{A key lemma} In this section, we prove
\begin{lemma}\label{lemK} Let $L \in {\mathcal N}$. Assume that $\dim {\mathcal M} = 1$ or \eqref{main-assumption} holds. There exist $\varepsilon_0 > 0$, $C>0$, and $T_0 > 0$ depending only on $L$ such that for all (real) $u_0 \in L^2(0, L)$ with $\| u_0 \|_{L^2(0, L)} \le \varepsilon_0$, the unique solution $u \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$
of system \eqref{KdV-NL} satisfies
\begin{equation}\label{decayK}
\| u(T, \cdot) \|_{L^2(0, L)} \le \| u_0 \|_{L^2(0, L)} \Big(1 - C \| u_0 \|_{L^2(0, L)}^2 \Big) \mbox{ for } T \ge T_0.
\end{equation}
\end{lemma}
\begin{proof} We first collect several known facts. Let $T_1 > 0$ be such that
\begin{equation}\label{lemK-T1}
\| v_x(\cdot, 0) \|_{L^2(0, t)} \ge \frac{1}{2} \| v(0, \cdot) \|_{L^2(0, L)} \mbox{ for } t \ge T_1,
\end{equation}
for all solutions $v \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$ of the system
\begin{equation}\left\{
\begin{array}{cl}
v_t (t, x) + v_x (t, x) + v_{xxx} (t, x) = 0 & \mbox{ in } (0, +\infty) \times (0, L), \\[6pt]
v(t, x=0) = v(t, x=L) = v_x(t , x= L)= 0 & \mbox{ in } (0, +\infty),
\end{array}\right.
\end{equation}
with $v(0, \cdot) \in L^2(0, L)$ satisfying the condition
$$
v(0, \cdot) \perp {\mathcal M}
$$
(the orthogonality is considered with respect to $L^2(0, L)$-scalar product). The existence of such a constant $T_1$ follows from \cite{Rosier97}.
There exist two positive constants $\varepsilon_0$ and $c_1$ such that if $\| u_0 \|_{L^2(0, L)} \le \varepsilon_0$, then
\begin{equation}\label{lemK-c1}
\| u \|_{C\big([0, T_1]; L^2(0, L) \big)} + \| u \|_{L^2\big( (0, T_1); H^1(0, L) \big)} \le c_1 \| u_0\|_{L^2(0, L)}
\end{equation}
(see e.g., \cite[Proposition 14]{CC04}).
There is a positive constant $c_2$ such that if $\widetilde u_0 \in L^2(0, L)$, $\widetilde f \in L^1\big((0, T_1); L^2(0, L) \big)$, and $\widetilde y \in C \big([0, T_1); L^2(0, L) \big) \cap L^2 \big([0, T_1); H^1(0, L) \big) $ is the unique solution of the system
\begin{equation}\left\{
\begin{array}{cl}
\widetilde u_t (t, x) + \widetilde u_x (t, x) + \widetilde u_{xxx} (t, x) = \widetilde f & \mbox{ in } (0, T_1) \times (0, L), \\[6pt]
\widetilde u(t, x=0) = \widetilde u(t, x=L) = \widetilde u_x(t , x= L)= 0 & \mbox{ in } (0, T_1), \\[6pt]
\widetilde u(t= 0, \cdot) = \widetilde u_0 & \mbox{ in } (0, L),
\end{array}\right.
\end{equation}
then
\begin{multline}\label{lemK-c2}
\| \widetilde u_x(\cdot, 0) \|_{L^2(0, T_1)} + \| \widetilde u \|_{C\big([0, T_1]; L^2(0, L) \big)} + \| \widetilde u \|_{L^2\big( (0, T_1); H^1(0, L) \big)} \\[6pt]
\le c_2 \Big( \| \widetilde u_0\|_{L^2(0, L)} + \|\widetilde f\|_{L^1\big((0, T_1); L^2(0, L) \big)} \Big).
\end{multline}
There exists a positive constant $c_3$ depending only on $L$ such that, for {\it all} $T>0$,
\begin{equation}\label{lemK-c3}
\| \xi \xi_x \|_{L^1\big( (0, T); L^2(0, L) \big)} \le c_3 \| \xi \|_{L^2\big( (0, T); H^1(0, L) \big)}^2
\end{equation}
(the constant $c_3$ is independent of $T$).
\medskip
We now decompose $u_0$ into two parts:
\begin{equation}\label{thm-u0-decomp}
u_0 = u_{0, 1} + u_{0, 2} \mbox{ in } (0, L),
\end{equation}
where
\begin{equation*}
u_{0, 1} = \mbox{Projection}_{{\mathcal M}} u_0
\end{equation*}
with respect to $L^2(0, L)$-scalar product.
\medskip
The proof is now divided into two cases, with $0< \varepsilon = \| u_{0} \|_{L^2(0, L)} < \varepsilon_0$ (the conclusion is clear if $\varepsilon =0$),
\begin{itemize}
\item Case 1: $\| u_{0, 2} \|_{L^2(0, L)} \ge \beta \varepsilon^2 = \beta \| u_0\|_{L^2(0, L)}^2$,
\item Case 2: $\| u_{0, 2} \|_{L^2(0, L)} < \beta \varepsilon^2 = \beta \| u_0\|_{L^2(0, L)}^2$,
\end{itemize}
where
\begin{equation}\label{def-beta}
\beta = 4c_1^2 c_2 c_3.
\end{equation}
\medskip
\noindent{\it Case 1}: Assume that
\begin{equation}\label{Case1}
\| u_{0, 2} \|_{L^2(0, L)} \ge \beta \varepsilon^2 = \beta \| u_0\|_{L^2(0, L)}^2.
\end{equation}
Let $\hat u \in C \big([0, T_1); L^2(0, L) \big) \cap L^2 \big([0, T_1); H^1(0, L) \big)$ be the unique solution of
\begin{equation}\label{lemK-hu}\left\{
\begin{array}{cl}
\hat u_t (t, x) + \hat u_x (t, x) + \hat u_{xxx} (t, x) = 0 & \mbox{ in } (0, T_1) \times (0, L), \\[6pt]
\hat u(t, 0) = \hat u(t, L) = \hat u_x(t , L)= 0 & \mbox{ in } (0, T_1), \\[6pt]
\hat u(0, \cdot) = u_0 & \mbox{ in } (0, L).
\end{array}\right.
\end{equation}
Then
\begin{equation}\label{lemK-c1c2}
\|(\hat u - u)_x(\cdot, 0) \|_{L^2(0, T_1)} \mathop{\le}^{\eqref{lemK-c2}} c_2 \| u u_x \|_{L^1 \big((0, T_1) ; L^2(0, L) \big)} \mathop{\le}^{\eqref{lemK-c1}, \eqref{lemK-c3}} c_1^2 c_2 c_3 \varepsilon^2.
\end{equation}
Let $\hat u_j \in C \big([0, T_1); L^2(0, L) \big) \cap L^2\big([0, T_1); H^1(0, L) \big)$ with $j=1, \, 2$ be the unique solution of
\begin{equation}\left\{
\begin{array}{cl}
\hat u_{j, t} (t, x) + \hat u_{j, x} (t, x) + \hat u_{j, xxx} (t, x) = 0 & \mbox{ for } t \in (0, T), \, x \in (0, L), \\[6pt]
\hat u_j(t, 0) = \hat u_j(t, L) = \hat u_{j, x} (t , L)= 0 & \mbox{ for } t \in (0, T), \\[6pt]
\hat u_j(0, \cdot) = u_{0, j} & \mbox{ in } (0, L).
\end{array}\right.
\end{equation}
Then
$$
\hat u = \hat u_1 + \hat u_2 \mbox{ in } [0, T_1] \times [0, L].
$$
We have
\begin{equation}\label{Case1-p1}
\hat u_{1, x} (\cdot, 0) =0 \mbox{ in } [0, T_1],
\end{equation}
and, by the choice of $T_1$ via \eqref{lemK-T1},
\begin{equation}\label{Case1-p2}
\| \hat u_{2, x} (\cdot, 0)\|_{L^2(0, T_1)} \ge \frac{1}{2} \| \hat u_2(0, \cdot)\|_{L^2(0, L)} = \frac{1}{2} \| u_{0,2}\|_{L^2(0, L)}.
\end{equation}
It follows from \eqref{Case1} that
\begin{equation}\label{Case1-p3}
\| \hat u_{x} (\cdot, 0)\|_{L^2(0, T_1)} \ge \frac{1}{2} \beta \varepsilon^2.
\end{equation}
From \eqref{lemK-c1c2} and \eqref{Case1-p3}, we obtain
\begin{multline*}
\| u_{x} (\cdot, 0)\|_{L^2(0, T_1)} \ge \| \hat u_{x} (\cdot, 0)\|_{L^2(0, T_1)} - \| (u - \hat u)_{x} (\cdot, 0)\|_{L^2(0, T_1)} \ge \left( \frac{1}{2} \beta - c_1^2 c_2 c_3 \right) \varepsilon^2 \mathop{\ge}^{\eqref{def-beta}} c_1^2 c_2 c_3 \varepsilon^2.
\end{multline*}
In other words,
\begin{equation}\label{Case1-*}
\| u_{x} (\cdot, 0)\|_{L^2(0, T_1)}
\ge c_1^2 c_2 c_3 \| u_0\|_{L^2(0, L)}^2.
\end{equation}
\medskip
\noindent{\it Case 2}: Assume that
\begin{equation}\label{Case2}
\| u_{0, 2} \|_{L^2(0, L)} < \beta \varepsilon^2 = \beta \| u_0 \|_{L^2(0, L)}^2.
\end{equation}
Since
$$
\| u_{0, 1} \|_{L^2(0, L)}^2 + \| u_{0, 2} \|_{L^2(0, L)}^2 = \| u_{0} \|_{L^2(0, L)}^2 = \varepsilon^2,
$$
by considering $\varepsilon$ sufficiently small, one can assume that
$$
\| u_{0, 1} \|_{L^2(0, L)} \ge \varepsilon/2.
$$
Let $\alpha_m \in \mathbb{C}$ ($1 \le m \le n_L$) be such that
\begin{equation}\label{thm-def-y1}
\frac{1}{\varepsilon} u_{0, 1} = \Re \left\{ \sum_{m=1}^{n_L} \alpha_m \Psi_m (0, x) \right\}.
\end{equation}
Since $u_{0, 1} \in {\mathcal M}$, such a family of $(\alpha_m)_{m=1}^{n_L}$ exists. Since
$$
1/2 \le \| \frac{1}{\varepsilon} u_{0, 1} \|_{L^2(0, L)} \le 1
$$
and $\Big( \Psi_m (0, \cdot) \Big)$ is orthogonal in $L^2(0, L)$ (with respect to the complex field), one can assume in addition that
$$
0< \gamma_1 \le \sum_{m=1}^{n_L} |\alpha_m|^2 \le \gamma_2,
$$
for some constants $\gamma_1$, $\gamma_2$ depending only on $L$. Moreover, since $\Psi_m (0, x) \in i \mathbb{R}$ for $x \in [0, L]$ (by \eqref{psi-complex}) if $p_m = 0$ (see e.g. \eqref{psi-complex}), one can also assume that $a_m \in i \mathbb{R}$ if $p_m =0$.
Let $\gamma_0 > 0$ and $\tau_0> 0$ be the constants given in \Cref{pro-quasi} with $\ell= n_L$, $\gamma_1$ and $\gamma_2$ determined above, $q_m= p_m$ given by \eqref{def-pm},
\begin{equation}\label{Case2-M}
M_{m_1, m_2} = \frac{1}{8} \varphi_{m_1, m_2}' (0) \quad \mbox{ and } \quad N_{m_1, m_2} = \frac{1}{8} \phi_{m_1, m_2}' (0),
\end{equation}
where $\varphi_{m_1, m_2}$ and $\phi_{m_1, m_2}$ are defined in \Cref{pro1} and \Cref{pro1-Co}, respectively; in the case the definition of $\varphi_{m_1, m_2}$ and $\phi_{m_1, m_2}$ in \Cref{pro1} and \Cref{pro1-Co} are not unique, we fix a choice of $\varphi_{m_1, m_2}$ and $\phi_{m_1, m_2}$.
By \Cref{pro-chi}, we have
$$
M_{m,m} \neq 0,
$$
and
$$
(M_{m, m} \mbox{ is real and $N_{m, m} \neq 0$) if } p_m = 0.
$$
Then, by \Cref{pro-quasi}, for all
$a_j \in \mathbb{C}$ $(1 \le j \le N)$ satisfying
$\gamma_1 \le \sum_{j=1}^N |a_j|^2 \le \gamma_2$, it holds
\begin{equation}\label{Case2-pro-g}
\| g \|_{L^2(\tau, 2 \tau)} \ge \gamma_0 \mbox{ for all } \tau \ge \tau_0,
\end{equation}
where
\begin{multline}\label{Case2-def-g}
g(t) = \sum_{m_1 = 1}^{n_L} \sum_{m_2 = 1}^{n_L} \Big( a_{m_1} a_{m_2} M_{m_1, m_2} e^{- i (p_{m_1} + p_{m_2}) t} \\[6pt] + \bar a_{m_1} \bar a_{m_2} \bar M_{m_1, m_2} e^{ i (p_{m_1} + p_{m_2}) t}
+ 2 a_{m_1} \bar a_{m_2} N_{m_1, m_2} e^{-i (p_{m_1} - p_{m_2})} \Big).
\end{multline}
Define
\begin{equation}\label{Case2-defA}
A = \beta + 2 \sum_{m_1 = 1}^{n_L} \sum_{m_2 = 1}^{n_L} \|\varphi_{m_1, m_2}\|_{L^2(0, L)} + 2 \sum_{m_1 = 1}^{n_L} \sum_{m_2 = 1}^{n_L} \|\phi_{m_1, m_2}\|_{L^2(0, L)},
\end{equation}
and set
\begin{equation}\label{Case2-def-c4}
c_4 = 1/ (2 A).
\end{equation}
Let $T_2 \ge 2 \tau_0$ be such that
\begin{equation}\label{Case2-T2}
\| y_x(\cdot, 0) \|_{L^2(T_2/2, T_2)} \le c_4 \gamma_0 \| y(0, \cdot) \|_{L^2(0, L)},
\end{equation}
for all solutions $y \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$ of
\begin{equation}\label{thm-sys1}\left\{
\begin{array}{cl}
y_t (t, x) + y_x (t, x) + y_{xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
y(t, 0) = y(t, L) = y_x(t , L)= 0 & \mbox{ for } t \in (0, +\infty),
\end{array}\right.
\end{equation}
with $y(0, \cdot) \in L^2(0, L)$. Note that $T_2$ is independent of $y(0, \cdot)$. The existence of $T_2$ can be proved by decomposing $y(0, \cdot) = y_1(0, \cdot) + y_2(0, \cdot)$ with $y_1(0, \cdot) \in {\mathcal M}$, and noting that \eqref{Case2-T2} holds for the solution with initial data being $y_2(0, \cdot)$
since the solution is exponential decay, and the contribution for $y_x(\cdot, 0)$ from the solution with initial data is $y_1(0, \cdot)$ is 0.
Let $\widetilde u_1, \; \widetilde u_2 \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$ be the unique solution of
\begin{equation}\label{sys-NL}\left\{
\begin{array}{cl}
\widetilde u_{1, t} (t, x) + \widetilde u_{1, x} (t, x) + \widetilde u_{1, xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\widetilde u_1(t, 0) = \widetilde u_1(t, L) = \widetilde u_{1, x}(t , L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
\displaystyle \widetilde u_1(0, \cdot ) = \frac{1}{\varepsilon} u_{0, 1} & \mbox{ in } [0, L],
\end{array}\right.
\end{equation}
and
\begin{equation}\label{Case2-sys}\left\{
\begin{array}{cl}
\widetilde u_{2, t} (t, x) + \widetilde u_{2, x} (t, x) + \widetilde u_{2, xxx} (t, x) + \widetilde u_1 \widetilde u_{1, x} = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\widetilde u_2(t, 0) = \widetilde u_2(t, L) = \widetilde u_{2, x}(t , L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
\displaystyle \widetilde u_2(0, \cdot) = \frac{1}{\varepsilon^2} u_{0, 2} & \mbox{ in } [0, L].
\end{array}\right.
\end{equation}
Set
$$
V(t, x) = \sum_{m=1}^{n_L} \alpha_m \Psi_m(t, x) \quad \mbox{ and } \quad U(t, x) = \Re V(t, x).
$$
We have
\begin{equation}\label{sys-U}\left\{
\begin{array}{cl}
U(t, x) + U_x (t, x) + U_{xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
U(t, x=0) = U(t, x=L) = U_x(t , x= L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
\displaystyle U(t=0, \cdot) = \frac{1}{\varepsilon} u_{0, 1} & \mbox{ in } [0, L].
\end{array}\right.
\end{equation}
This implies
$$
\widetilde u_1 = U \mbox{ in } (0, + \infty) \times (0, L).
$$
Define
\begin{equation}\label{def-V1}
V_1(t, x) = \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \alpha_{m_1} \alpha_{m_2} \varphi_{m_1, m_2}(x) e^{- i (p_{m_1} + p_{m_2} ) t},
\end{equation}
and
\begin{equation}\label{def-V2}
V_2(t, x) = \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \alpha_{m_1} \bar \alpha_{m_2} \phi_{m_1, m_2}(x) e^{- i (p_{m_1} - p_{m_2} ) t}.
\end{equation}
Then, by the construction of $\varphi_{m_1, m_2}$,
\begin{equation}\left\{
\begin{array}{cl}
V_{1, t} (t, x) + V_{1, x} (t, x) + V_{1, xxx} (t, x) + \big( V(t, x) V(t, x) \big)_x = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
V_1(t, 0) = V_1(t, L) = V_{1, x}(t , L)= 0 & \mbox{ for } t \in (0, +\infty),
\end{array}\right.
\end{equation}
and, by the construction of $\phi_{m_1, m_2}$,
\begin{equation}\left\{
\begin{array}{cl}
V_{2, t} (t, x) + V_{2, x} (t, x) + V_{2, xxx} (t, x) + \big( |V(t, x)|^2 \big)_x = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
V_2(t, 0) = V_2(t, L) = V_{2, x}(t , L)= 0 & \mbox{ for } t \in (0, +\infty).
\end{array}\right.
\end{equation}
Set
\begin{equation}\label{def-W}
W = \frac{1}{8} \Big( V_1 + \bar V_1 + 2 V_2 \Big) \mbox{ in } (0, + \infty) \times (0, L).
\end{equation}
It follows from \eqref{Case2-def-g} that $W_x(t, 0) = g(t)$ in $\mathbb{R}_+$ and hence, by \eqref{Case2-pro-g},
\begin{equation}\label{Case2-pro-W}
\| W_x(t, 0) \|_{L^2(\tau, 2 \tau)} \ge \gamma_0 \mbox{ for all } \tau \ge \tau_0.
\end{equation}
Since
$$
\big( V(t, x) V(t, x) \big)_x + \overline{\big( V(t, x) V(t, x) \big)_x} + 2 \big( |V(t, x)|^2 \big)_x = 8 U(t, x) U_x(t, x),
$$
we derive from \eqref{def-W} that
\begin{equation}\left\{
\begin{array}{cl}
W_{t} (t, x) + W_{x} (t, x) + W_{xxx} (t, x) + U(t, x) U_{x}(t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
W(t, 0) = W(t, L) = W_{x}(t , L)= 0 & \mbox{ for } t \in (0, +\infty). \end{array}\right.
\end{equation}
Let $\widetilde W \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$ be the unique solution of
\begin{equation}\left\{
\begin{array}{cl}
\widetilde W_{t} (t, x) + \widetilde W_{x} (t, x) + \widetilde W_{xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\widetilde W(t, 0) = \widetilde W(t, L) = \widetilde W_{x}(t , L) = 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
\displaystyle \widetilde W(0, \cdot) = \widetilde u_{2} (0, \cdot) - W(0, \cdot).
\end{array}\right.
\end{equation}
Then
\begin{equation}\label{Case2-hu2WW}
\widetilde u_2 = \widetilde W + W \mbox{ in } (0, + \infty) \times (0, L).
\end{equation}
We have
\begin{equation}\label{Case2-tt1}
\| \widetilde u_{2, x} (\cdot, 0) \|_{L^2(T_2/2, T_2)} \mathop{\ge}^{\eqref{Case2-hu2WW}} \| W_{x} (\cdot, 0) \|_{L^2(T_2/2, T_2)} - \| \widetilde W_{x} (\cdot, 0) \|_{L^2(T_2/2, T_2)},
\end{equation}
\begin{equation}\label{Case2-tt2}
\| \widetilde W_{x} (\cdot, 0) \|_{L^2(T_2/2, T_2)} \mathop{\le}^{\eqref{Case2-T2}} c_4 \gamma_0 \|\widetilde W(0, \cdot) \|_{L^2(0, L)},
\end{equation}
and, since $T_2 \ge \tau_0$,
\begin{equation}\label{Case2-tt3}
\| W_{x} (\cdot, 0) \|_{L^2(T_2/2, T_2)} \mathop{\ge}^{\eqref{Case2-pro-W}} \gamma_0.
\end{equation}
Since, by \eqref{def-V1}, \eqref{def-V2}, and \eqref{def-W}
\begin{multline*}
8W(0, x) = \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \alpha_{m_1} \alpha_{m_2} \varphi_{m_1, m_2}(x) \\[6pt]
+ \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \bar \alpha_{m_1} \bar \alpha_{m_2} \bar \varphi_{m_1, m_2}(x)
+ 2 \sum_{m_1=1}^{n_L} \sum_{m_2=1}^{n_L} \alpha_{m_1} \bar \alpha_{m_2} \phi_{m_1, m_2}(x),
\end{multline*}
it follows that
\begin{equation}\label{Case2-W0}
\|W(0, \cdot) \|_{L^2(0, L)} \le 2 \sum_{m_1 = 1}^{n_L} \sum_{m_2 = 1}^{n_L} \|\varphi_{m_1, m_2}\|_{L^2(0, L)} + 2 \sum_{m_1 = 1}^{n_L} \sum_{m_2 = 1}^{n_L} \|\phi_{m_1, m_2}\|_{L^2(0, L)}.
\end{equation}
By the definition of $A$ in \eqref{Case2-defA}, we obtain from \eqref{Case2} and \eqref{Case2-W0} that
\begin{equation}\label{Case2-A2}
A \ge
\| \widetilde u_2 (0, \cdot) \|_{L^2(0, L)} + \|W(0, \cdot) \|_{L^2(0, L)} \mathop{\ge}^{\eqref{Case2-hu2WW}} \| \widetilde W(0, \cdot) \|_{L^2(0, L)}.
\end{equation}
Combining \eqref{Case2-tt1}, \eqref{Case2-tt2}, \eqref{Case2-tt3}, and \eqref{Case2-A2} yields
\begin{equation*}
\|\widetilde u_{2, x} (\cdot, 0) \|_{L^2(T_2/2, T_2)} \ge \gamma_0 - c_4 \gamma_0 A.
\end{equation*}
Since $c_4 = 1 / (2 A)$ by \eqref{Case2-def-c4}, we obtain
\begin{equation}\label{lemK-p1}
\|\widetilde u_{2, x} (\cdot, 0) \|_{L^2(T_2/2, T_2)} \ge \gamma_0/2.
\end{equation}
Set
\begin{equation*}
u_d = \varepsilon \widetilde u_1 + \varepsilon^2 \widetilde u_2 - u \mbox{ in } (0, + \infty) \times (0, L),
\end{equation*}
and
\begin{equation}
f_d = u u_x - \varepsilon^2 \widetilde u_1 \widetilde u_{1, x} \mbox{ in } (0, + \infty) \times (0, L).
\end{equation}
We have, by \eqref{sys-NL} and \eqref{Case2-sys},
\begin{equation}\left\{
\begin{array}{cl}
u_{d, t} (t, x) + u_{d, x} (t, x) + u_{d, xxx} (t, x) = f_d (t, x) & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
u_d(t, x=0) = u_d(t, x=L) = u_d(t , x= L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
u_d(t = 0, \cdot) = 0 & \mbox{ in } (0, L).
\end{array}\right.
\end{equation}
It is clear that
\begin{equation*}
\| f_d\|_{L^1\big((0, T_2); L^2(0, L) \big)} \le C \varepsilon^2,
\end{equation*}
where $C$ is a positive constant depending only on $T_2$ and $L$. It follows that
\begin{equation*}
\| u_d \|_{C\big([0, T_2]; L^2(0, L) \big)} + \| u_d \|_{L^2\big( (0, T_2); H^1(0, L) \big)} \le C \varepsilon^2.
\end{equation*}
This in turn implies that
\begin{equation*}
\| f_d\|_{L^1\big((0, T_2); L^2(0, L) \big)} \le C \varepsilon^3.
\end{equation*}
and therefore,
\begin{equation}\label{lemK-p2}
\| u_{d, x} (\cdot, 0) \|_{L^2( 0, T_2) } \le C \varepsilon^3.
\end{equation}
Combining \eqref{lemK-p1} and \eqref{lemK-p2}, and noting that $\widetilde u_{1, x} (t, 0) = 0$ yield
\begin{equation}\label{est-case2}
\|u_x (\cdot, 0) \|_{L^2(T_2/2, T_2) } \ge C \varepsilon^2.
\end{equation}
The analysis of Step 2 is complete.
\medskip
The conclusion now follows from Case 1 where one obtains \eqref{Case1-*} and Case 2 where one obtains \eqref{est-case2} by choosing $T_0 = \max\{T_1, T_2\}$ and using \eqref{key-identity}.
The proof is complete.
\end{proof}
We are ready to give
\subsection{Proof of \Cref{thm1}} By \Cref{lemK}, we have
$$
\|u(T_2, \cdot) \|_{L^2(0, L) } \le \| u(0, \cdot) \|_{L^2(0, L)} \Big(1 - C \| u(0, \cdot) \|_{L^2(0, L)}^2 \Big).
$$
This yields, with $\| u(0, \cdot) \|_{L^2(0, L)} = \varepsilon>0$ and $p$ being the largest integer less than $1/ (2 C \varepsilon^2)$,
$$
\|u(p T_2, \cdot) \|_{L^2(0, L) } \le \frac{1}{2} \| u(0, \cdot) \|_{L^2(0, L)}.
$$
Here we also used \eqref{key-identity}. Using \eqref{key-identity} again, it follows that, for $T \ge C/ \| u(0, \cdot) \|_{L^2(0, L)}^2$,
$$
\|u(T, \cdot) \|_{L^2(0, L) } \le \frac{1}{2} \| u(0, \cdot) \|_{L^2(0, L)}.
$$
This implies, by recurrence, that
$$
\| u(T, \cdot) \|_{L^2(0, L)} \le 2^{-n} \| u_0 \|_{L^2(0, L)}\mbox{ for } T \ge C \sum_{p=0}^{n-1} 2^{2p} /\| u(0, \cdot) \|_{L^2(0, L)}^2
$$
since $\|u(t, \cdot) \|_{L^2(0, L)}$ is a non-increasing function with respect to $t$. In particular, we obtain, since $\|u(t, \cdot) \|_{L^2(0, L)}$ is a non-increasing function with respect to $t$,
\begin{equation}
\|u(t, \cdot) \|_{L^2(0, L)} \le C/ t^{1/2}.
\end{equation}
The proof is complete. \qed
\section{A lower bound for the decay rate - Proof of \Cref{pro-opt}} \label{sect-opt}
Fix $1 \le m \le n_L$ and $\alpha_m \in \mathbb{C}$ with $|\alpha_m| = 1$ such that
$$
\Re ( \alpha_m \varphi_{m, m}(x)) \mbox{ is not identically equal to 0 in } [0, L].
$$
Let $\widetilde u_1 \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$ be the unique solution of
\begin{equation}\label{sys-NL-L}\left\{
\begin{array}{cl}
\widetilde u_{1, t} (t, x) + \widetilde u_{1, x} (t, x) + \widetilde u_{1, xxx} (t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\widetilde u_1(t, x=0) = \widetilde u_1(t, x=L) = \widetilde u_{1, x}(t , x= L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
\displaystyle \widetilde u_1(0, \cdot ) = \Re ( \alpha_m \varphi_{m, m}).
\end{array}\right.
\end{equation}
Set
\begin{equation}\label{def-V1-L}
V_1(t, x) = \alpha_{m}^2 \varphi_{m, m}(x) e^{- 2 i p_{m} t},
\end{equation}
\begin{equation}\label{def-V2-L}
V_2(t, x) = |\alpha_{m}|^2 \phi_{m, m}(x),
\end{equation}
and denote
\begin{equation}\label{def-W-L}
\widetilde u_2 = \frac{1}{8} \Big( V_1 + \bar V_1 + 2 V_2 \Big) \mbox{ in } (0, + \infty) \times (0, L).
\end{equation}
Since $\phi_{m, m}$ is real by 3) of \Cref{pro1-Co}, it follows that $V_2$ is real and hence so is $\widetilde u_2$.
As in the proof of \Cref{lemK}, we have
$$
\widetilde u_1(t, x) = \Re \Big(\alpha_m \varphi_{m, m} (x)e^{- i p_m t} \Big),
$$
and
\begin{equation*}\left\{
\begin{array}{cl}
\widetilde u_{2, t} (t, x) + \widetilde u_{2, x} (t, x) + \widetilde u_{2, xxx} (t, x) + \widetilde u_1(t, x) \widetilde u_{1, x}(t, x) = 0 & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
\widetilde u_2(t, x=0) = \widetilde u_2(t, x=L) = \widetilde u_{2, x}(t , x= L)= 0 & \mbox{ for } t \in (0, +\infty). \end{array}\right.
\end{equation*}
Let $u \in C \big([0, + \infty); L^2(0, L) \big) \cap L^2_{\operatorname{loc}} \big([0, + \infty); H^1(0, L) \big)$ be a (real) solution of \eqref{KdV-NL} with
$$
\| u(0, \cdot) \|_{L^2(0, L)} \le \Gamma \varepsilon,
$$
where
$$
\Gamma : = \sup_{t} \| \widetilde u_{1} (t, \cdot) \|_{L^2(0, L)} + 1.
$$
Set
$$
\widetilde u_2 (t, x) = W(t, x) \mbox{ in } (0, + \infty) \times (0, L),
$$
\begin{equation*}
u_d = \varepsilon \widetilde u_1 + \varepsilon^2 \widetilde u_2 - u \mbox{ in } (0, + \infty) \times (0, L),
\end{equation*}
\begin{equation*}
f_d = u u_x - \varepsilon^2 \widetilde u_1 \widetilde u_{1, x} \mbox{ in } (0, + \infty) \times (0, L).
\end{equation*}
We have
\begin{equation}\left\{
\begin{array}{cl}
u_{d, t} (t, x) + u_{d, x} (t, x) + u_{d, xxx} (t, x) = f_d (t, x) & \mbox{ for } t \in (0, +\infty), \, x \in (0, L), \\[6pt]
u_d(t, x=0) = u_d(t, x=L) = u_d(t , x= L)= 0 & \mbox{ for } t \in (0, +\infty), \\[6pt]
u_d(t = 0, \cdot) = 0 & \mbox{ in } (0, L).
\end{array}\right.
\end{equation}
Denote
\begin{equation}\label{pro-opt-def-g}
g_d = \varepsilon^3 ( \widetilde u_1 \widetilde u_{2, x} + \widetilde u_2 \widetilde u_{1, x} ) + \varepsilon^4 \widetilde u_2 \widetilde u_{2, x}.
\end{equation}
We write $f_d$ under the form
\begin{align*}
f_d = & (u- \varepsilon \widetilde u_1 - \varepsilon^2 \widetilde u_2) u_x + (\varepsilon \widetilde u_1 + \varepsilon^2 \widetilde u_2) (u- \varepsilon \widetilde u_1 - \varepsilon^2 \widetilde u_2)_x + g_d \\[6pt]
= & - u_d u_x - (\varepsilon \widetilde u_1 + \varepsilon^2 \widetilde u_2) u_{d, x} + g_d.
\end{align*}
Multiplying the equation of $u_d$ with $u_d$ (which is real), integrating by parts in $(1, t) \times (0, L)$, and using the form of $f_d$ just above give
\begin{multline}\label{pro-opt-p1}
\int_0^L |u_d(t, x)|^2 \, dx \le \int_0^L |u_d(1, x)|^2 \, dx + 2 \int_1^t \int_0^L |u_d|^2 |u_x| \, dx \, ds \\[6pt]
+ \int_1^t \int_{0}^L ( \varepsilon |\widetilde u_1| + \varepsilon^2 |\widetilde u_2| )_x |u_d|^2 \,dx \, ds + 2 \int_1^t \int_0^L |g_d| |u_d|.
\end{multline}
Since
$$
\int_0^L |u(t, x)|^2 \, dx \le \int_0^L |u(0, x)|^2 \le C \varepsilon \mbox{ for } t \ge 0,
$$
and the effect of the regularity, one has
\begin{equation}\label{pro-opt-p2}
|u(t, x)| + |u_x(t, x)| \le C \varepsilon \mbox{ for } t \ge 1, \, x \in [0, L].
\end{equation}
Let $a$ be a (small) positive constant defined later (the smallness of $a$ depending only on $L$). Let $t_0 \in [1, a/\varepsilon]$ be such that
$$
\int_0^L |u_d(t_0, x)|^2 \, dx = \max_{t \in [1, a/\varepsilon]} \int_0^L |u_d(t, x)|^2 \, dx.
$$
Combining \eqref{pro-opt-p1} with $t = t_0$ and \eqref{pro-opt-p2} yields
$$
\int_0^L |u_d(t_0, x)|^2 \, dx \le \int_0^L |u_d(1, x)|^2 \, dx + C a \int_0^L |u_d(t_0, x)|^2 \, dx + \int_1^{a/\varepsilon} \int_0^L \varepsilon^{-1} |g_d|^2 \, d x.
$$
This implies, if $a$ is sufficiently small,
$$
\int_0^L |u_d(t_0, \cdot)|^2 \, dx \, dx \le C \int_0^L |u_d(1, x)|^2 \, dx + C \varepsilon^4
$$
by \eqref{pro-opt-def-g}.
On the other hand, one has
$$
\int_0^L |u_d(t, \cdot)|^2 \, dx \, dx \le C \int_0^L |u_d(0, x)|^2 \, dx + C \varepsilon^4 \mbox{ for } t \in [0, 1].
$$
We have just proved that, for $a$ sufficiently small,
$$
\sup_{t \in [0, a/\varepsilon]} \| u_d(t, \cdot) \|_{L^2(0, L)} \le C \Big( \| u_d(0, \cdot) \|_{L^2(0, L)} + \varepsilon^2 \Big).
$$
Continuing this process, we obtain
\begin{equation}\label{pro-opt-p3}
\sup_{t \in [0, a n /\varepsilon]} \| u_d(t, \cdot) \|_{L^2(0, L)} \le C^n \| u_d(0, \cdot) \|_{L^2(0, L)} + \sum_{k=1}^n C^k \varepsilon^2.
\end{equation}
We now consider $u$ with
$$
u(0, \cdot) = \varepsilon \widetilde u_1(0, \cdot) + \varepsilon_2 \widetilde u_2 (0, \cdot).
$$
Thus
\begin{equation}\label{pro-opt-p4}
u_d(0, \cdot) = 0.
\end{equation}
Fix $ \gamma > 0$ such that
\begin{equation}
\inf_{t \in \mathbb{R}} \int_{0}^L |\widetilde u_1(t, x)|^2 \,dx \ge 4 \gamma.
\end{equation}
With $n $ being the largest integer number such that $C^{n+1} \le \gamma \varepsilon^{-1}$ (we assume now and later on that $C \ge 2$), we derive from \eqref{pro-opt-p3} and \eqref{pro-opt-p4} that
\begin{equation*}
\sup_{t \in [0, a n /\varepsilon]} \| u_d(t, \cdot) \|_{L^2(0, L)} \le \gamma \varepsilon.
\end{equation*}
Since
$$
u_d = \varepsilon \widetilde u_1 + \varepsilon^2 \widetilde u_2 - u,
$$
by the choice of $\gamma$, we have, for $\varepsilon $ sufficiently small,
\begin{equation*}
\| u(a n/ \varepsilon, \cdot) \|_{L^2(0, L)} \ge \gamma \varepsilon.
\end{equation*}
We deduce that, with $\tau = a n/ \varepsilon \sim \varepsilon^{-1} \ln \varepsilon^{-1}$ (hence $\varepsilon^{-1} \sim \tau / \ln \tau$),
\begin{equation*}
\| u(\tau, \cdot) \|_{L^2(0, L)} \ge \gamma \varepsilon \ge C \gamma \ln \tau/ \tau.
\end{equation*}
The proof is complete.
\qed
| 7af09c1371c5d1f01acc3ad73a4c6318527c415b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction and notations}
As this article deals with log-concave random variables and probability density functions, we first recall basic definitions and properties:
\begin{defi}
A density $f : \mathbb{R} \rightarrow \mathbb{R}_+$ is said to be log-concave if it satisfies the following inequality:
\begin{equation*}
\forall x,y \in \mathbb{R}, \forall \theta \in [0,1], f(\theta x + (1 - \theta) y) \geq f(x)^{\theta} f(y)^{1 - \theta}.
\end{equation*}
A real-valued random variable $X$ is said to be log-concave if it admits a density $f$ which is log-concave.
\end{defi}
Basic properties of log-concave distributions are surveyed in a recent paper by Saumard and Wellner, see \cite{review}.
\medskip
Log-concavity for integer-valued probability distributions, i.e. for families ${(f(k))}_{k \in \mathbb{Z}}$ of non-negative numbers with $\underset{k \in \mathbb{Z}}{\sum} f(k) = 1$, can be defined as follows:
\begin{defi}
We say that ${(f(k))}_{k \in \mathbb{Z}}$ is log-concave if we have the following inequality:
\begin{equation*}
\forall k \in \mathbb{Z} \ , \ f(k)^2 \geq f(k-1) f(k+1).
\end{equation*}
\end{defi}
\medskip
\begin{defi}
Given an integer $n \ge 2$, function $f : \mathbb{R} \rightarrow \mathbb{R}$ is said to be a P\'olya frequency function in the class $PF_n$ if for every $n$-uples $a_1 \leq \dots \leq a_n$ and $b_1 \leq \dots \leq b_n$ we have:
\begin{equation*}
\det {(f(a_i - b_j))}_{1 \leq i,j \leq n} \geq 0.
\end{equation*}
\end{defi}
For example, the density $x \mapsto \dfrac{1}{\sqrt{2 \pi}} \exp \left( - \dfrac{x^2}{2}\right)$ is in the class $PF_n$ for any $n \geq 2$\cite{schoenberg}. \\
We can find a characterization of $PF_3$ class and some examples in \cite{PF3}.
The notion of $PF_n$ class can be seen as a natural generalization of log-concavity due to the following:
\begin{prop}
\label{prop1}
The class $PF_2$ is the class of log-concave functions.
\end{prop}
The seminal paper \cite{schoenberg} by Schoenberg surveys basic properties of Pólya frequency functions.
\medskip
In \cite{efron}, Efron studied a non-trivial monotonicity property for log-concave distributions:
\begin{defi}
Let $X,Y$ be two real-valued random variables.
\begin{itemize}
\item The pair $(X,Y)$ is said to satisfy the restricted monotonicity property if for any non-decreasing function $\phi : \mathbb{R} \rightarrow \mathbb{R}$, the functions ${\Phi}_X,{\Phi}_Y: \mathbb{R} \rightarrow \mathbb{R}$, defined by
\begin{equation*}
{\Phi}_X(s) := \mathbb{E}[\phi(X)|X+Y = s] \ , \ {\Phi}_Y(s) := \mathbb{E}[\phi(Y)|X+Y = s].
\end{equation*}
are also non-decreasing.
\item The pair (X,Y) is said to satisfy the strong monotonicity property if for any function $\phi : \mathbb{R}^2 \rightarrow \mathbb{R}$ which is non-decreasing in each variable, the function $\Phi : \mathbb{R} \rightarrow \mathbb{R}$, defined by
\begin{equation*}
\Phi(s) := \mathbb{E}[\phi(X,Y)|X+Y =s].
\end{equation*}
is also non-decreasing.
\end{itemize}
\end{defi}
Efron's theorem is stated as follows:
\begin{theo}\label{thm_efron}
Let $(X,Y)$ be a pair of real-valued random variables.
\begin{itemize}
\item The pair $(X,Y)$ satisfies the strong monotonicity property if and only if satisfies the restricted monotonicity property.
\item If $X$ and $Y$ are log-concave, then they satisfy the restricted and strong monotonicity properties.
\end{itemize}
\end{theo}
By using Efron's theorem one can prove the stability of log-concavity by convolution. In the article \cite{Wellner}, there is also a proof of the stability of strong log-concavity by convolution.
\medskip
A recent work on Efron's monotonicity property is the article \cite{efron_2} by Saumard and Wellner. In this paper, the authors describe conditions on a pair $(X,Y)$, other than log-concavity, for the monotonicity property to hold.
\medskip
The purpose of this paper is to prove two others generalizations of Efron's theorem. These generalizations are actually suggested by the title of Efron's original paper, \textit{"Increasing properties of P\'olya frequency functions"}, which is a bit misleading since the paper only deals with log-concave distributions.
\medskip
In Section 2, we generalize a restricted version of Efron's Theorem, see Theorem \ref{theorem1}: a generalized restricted monotonicity property holds for pairs of $PF_n$ variables. A consequence of this theorem is another proof of the stability of the class $PF_n$ by convolution. The proof of Theorem \ref{theorem1} is very simple and only uses elementary properties of determinants. \\
In section 3, we reinforce the strong version of Efron's Theorem: we prove that for a stronger monotonicity assumption on the function $\phi$, we obtain a more powerful result. We can see Theorem \ref{thm_efron} as a particular case of Theorem \ref{theorem2}. \\
In section 4, we give a more general result of the section 3, where Theorem \ref{theorem2} and Theorem \ref{theorem2_2} are a particular case. We apply in particular this result to the function $\Phi : s \mapsto \mathbb{E}[XY|X+Y=s]$.
\section{A restricted Efron's theorem for $PF_n$ variables}
In this paragraph, we prove the following generalization of restricted Efron's theorem.
\begin{defi}
A $n$-uple of functions ${\phi}_1, \ldots ,{\phi}_n$ is in the class $GM_n$ if for every $x_1 \leq \ldots \leq x_n$, we have :
\begin{equation*}
\det {({\phi}_i(x_j))}_{1 \leq i,j \leq n} \geq 0.
\end{equation*}
\end{defi}
We can remark that, for a pair $(\phi_1,\phi_2)$ of positive functions, the condition $({\phi}_1,{\phi}_2) \in GM_2$ is equivalent to say that the function $\dfrac{{\phi}_2}{{\phi}_1}$ is non-decreasing.
\begin{theo}
\label{theorem1}
We fix some $n \geq 2$. Let $X,Y$ be independent random variables with $PF_n$ densities $f,g$. Let $({\phi}_1, \dots ,{\phi}_n)$ be a n-uple of functions in the class $GM_n$. For $1 \leq k \leq n$ we define the function ${\Phi}_k : \mathbb{R} \rightarrow \mathbb{R}$ by
\begin{equation*}
{\Phi}_k(s) := \mathbb{E}[{\phi}_k(X) | X+Y = s].
\end{equation*}
We assume that ${\Phi}_k$ is well-defined. Then the n-uple $({\Phi}_1, \ldots ,{\Phi}_n)$ is also in the class $GM_n$.
\end{theo}
Restricted Efron's theorem is the particular case where $n=2$, ${\phi}_1 = 1$ and ${\phi}_2$ is a non-decreasing function. In order to prove Theorem \ref{theorem1}, we will use two elementary propositions. \\
The first one is a well-known identity which can be traced back to Andreev, see \cite{Andreief} and \cite{forrester}.
\begin{prop}
\label{proposition1}
Let ${(f_{i,j}(x))}_{1 \leq i,j \leq n}$ be a $n \times n$ matrix of integrable functions. Then
\begin{equation*}
\det \left( \int f_{i,j}(x) dx \right) = \int_{\mathbb{R}^n} \det(f_{i,j}(x_i)) dx_1 \dots dx_n = \int_{\mathbb{R}^n} \det(f_{i,j}(x_j)) dx_1 \dots dx_n .
\end{equation*}
\end{prop}
The second elementary proposition is proven by noticing that the set of $(x_1, \dots ,x_n) \in \mathbb{R}^n$ such that $x_i = x_j$ for some $i \neq j$ has its Lebesgue measure equal to zero. We denote by $\mathfrak{S}_n$ the set of permutations on $\{ 1, \dots , n\}$.
\begin{prop}
\label{proposition2}
Let $h : \mathbb{R}^n \rightarrow \mathbb{R}$ be an integrable function. We have:
\begin{equation*}
\int_{\mathbb{R}^n} h(x_1, \dots ,x_n) dx_1 \dots dx_n = \int_{x_1 < \ldots < x_n} \underset{\sigma \in \mathfrak{S}_n}{\sum} h(x_{\sigma(1)}, \ldots ,x_{\sigma(n)}) dx_1 \ldots dx_n.
\end{equation*}
\end{prop}
\begin{proof}[Proof of Theorem \ref{theorem1}]
Let us fix an ordered n-uple $s_1 < \dots < s_n$. For $1 \leq i,j \leq n$, we have:
\begin{equation*}
{\Phi}_i(s_j) = \mathbb{E}[{\phi}_i(X)|X+Y=s_j] = \dfrac{\int_{\mathbb{R}} f(x_i) {\phi}_i(x_i) g(s_j- x_i ) dx_i}{\int_{\mathbb{R}} f(x) g(s_j - x) dx}.
\end{equation*}
We thus have
\begin{equation}
\label{equation1}
\det({\Phi}_i(s_j)) = \dfrac{1}{\overset{n}{\underset{j=1}{\prod}} \int_{\mathbb{R}} f(x) g(s_j - x)} \det \left( \int_{\mathbb{R}} f(x_i) {\phi}_i(x_i) g(s_j -x_i) dx_i \right).
\end{equation}
Theorem \ref{theorem1} will thus be proven if we show that the determinant of the right-hand of \eqref{equation1} is non-negative. We will denote this determinant by $D$. \\
By Proposition \ref{proposition1}, we have:
\begin{equation*}
D = \int_{\mathbb{R}^n} \left( \overset{n}{\underset{i=1}{\prod}} f(x_i) \right) \left( \overset{n}{\underset{i=1}{\prod}} {\phi}_i(x_i) \right) \det(g(s_j - x_i))_{1 \leq i,j \leq n} dx_1 \dots dx_n.
\end{equation*}
We now use Proposition \ref{proposition2} to write:
\begin{equation*}
D = \int_{x_1 < \ldots < x_n} \underset{\sigma \in \mathfrak{S}_n}{\sum} \left( \overset{n}{\underset{i=1}{\prod}} f(x_{\sigma(i)}) \right) \left( \overset{n}{\underset{i=1}{\prod}} {\phi}_i (x_{\sigma(i)}) \right) \det (g(s_j - x_{\sigma (i) })) dx_1 \ldots dx_n.
\end{equation*}
For any permutation $\sigma \in \mathfrak{S}_n$, we have
$$\overset{n}{\underset{i=1}{\prod}} f(x_{\sigma(i)}) = \overset{n}{\underset{i=1}{\prod}} f(x_i).$$
Moreover, the function $g$ being in $PF_n$, we know that the function $d$ defined by
\begin{equation*}
d(x_1, \ldots ,x_n) := \det (g(s_j - x_i))_{1 \leq i,j \leq n}.
\end{equation*}
is non-negative when $x_1 < \ldots < x_n$. Moreover, for any permutation $\sigma \in \mathfrak{S}_n$, we have
\begin{equation*}
d(x_{\sigma(1)}, \ldots ,x_{\sigma(n)}) := \varepsilon(\sigma) \det (x_1, \ldots ,x_n).
\end{equation*}
We thus have:
\begin{align*}
D & = \int_{x_1 < \ldots < x_n} \left( \overset{n}{\underset{i=1}{\prod}} f(x_i) \right) d(x_1, \ldots ,x_n) \left( \underset{\sigma \in \mathfrak{S}_n}{\sum} \varepsilon (\sigma) \overset{n}{\underset{i=1}{\prod}} {\phi}_i(x_{\sigma(i)}) \right) dx_1 \ldots dx_n \\
&= \int_{x_1 < \ldots < x_n} \left( \overset{n}{\underset{i=1}{\prod}} f(x_i) \right) d(x_1, \ldots ,x_n) \det({\phi}_i(x_j))_{1 \leq i,j \leq n} dx_1 \ldots dx_n \\
&\geq 0.
\end{align*}
the inequality coming from $({\phi}_1, \ldots ,{\phi}_n)$ being in $GM_n$.
\end{proof}
An interesting consequence of Theorem \ref{theorem1} is the stability of the $PF_n$ class by convolution:
\begin{prop}
Let $X,Y$ be two independent random variables in the class $PF_n$. Then the variable $X+Y$ is also in $PF_n$.
\end{prop}
This result is already known, a proof can be found in \cite{karlin} or in \cite{schoenberg}. In the following proof, we suppose that at least one density is positive.
\begin{proof}
Let $X,Y$ be two independent random variables in the class $PF_n$. We denote by $f$ and $g$ their densities and by $r = f\star g$ the density of the random variable $X+Y$. We assume that $f >0$. \\
We want to prove that for any $a_1 \leq \cdots \leq a_n$ and $b_1 \leq \cdots \leq b_n, ~ \det(r(a_i-b_j))_{1 \leq i,j \leq n} \geq 0.$ \\
Let us fix an ordered n-uple $b_1 \leq \cdots \leq b_n$.
Now, we denote the function ${\phi}_i$ by
\begin{equation*}
{\phi}_i(x) := \dfrac{f(x-b_i)}{f(x)} .
\end{equation*}
We have that $f \in PF_n$, then by definition of the class $PF_n$ for any $a_1 \leq \ldots \leq a_n$:
$$\det({\phi}_i(a_j))_{1 \leq i,j \leq n} = \frac{1}{\overset{n}{\underset{j=1}{\prod}} f(a_j)} \det(f(a_j-b_i))_{1 \leq i,j \leq n} \geq 0.$$
In other words, $({\phi}_i)_i$ is in the class $GM_n$. So we can apply Theorem \ref{theorem1} to the $n$-uple of functions $({\phi}_i)_i$. We can also remark that:
$${\Phi}_i(a_j) = \dfrac{\int f(a_j-x) {\phi}_i(a_j-x) g(x) dx}{r(a_j)} = \dfrac{\int f(a_j - b_i -x) g(x) dx}{r(a_j)} = \dfrac{r(a_j-b_i)}{r(a_j)}.$$
Then by Theorem \ref{theorem1}, we have, for any $a_1 \leq \ldots \leq a_n$ :
$$ \det({\Phi}_i(a_j))_{1 \leq i,j \leq n} = \underset{j}{\prod} \dfrac{1}{r(a_j)} \det(r(a_j-b_i))_{1 \leq i,j \leq n} \geq 0.$$
Finally, we have proved that for any
$a_1 \leq \ldots \leq a_n$ and $b_1 \leq \cdots \leq b_n$:
$$\det(r(a_i-b_j))_{1 \leq i,j \leq n} \geq 0.$$
\end{proof}
\section{Efron's Theorem with a stronger monotonicity assumption}
We turn now to a refined version of Efron's theorem: under a stronger assumption on the monotonicity of the function $\phi$ on each variable we have a stronger result. We state it in two different frameworks, one for real random variables and a second for integer-valued random variables.
\subsection{Statement and examples}
The continuous version of the generalized Efron's theorem is stated as follows:
\begin{theo}
\label{theorem2}
Let X,Y be two independent log-concave real-valued random variables having densities. Let $a \ge 0$ be a real parameter and $\phi : \mathbb{R}^2 \rightarrow \mathbb{R}$ be a measurable function such that:
\begin{enumerate}
\item For any $ y \in \mathbb{R}$, the function $x \mapsto \exp(-ax) \phi(x,y)$ is non-decreasing.
\item For any $ x \in \mathbb{R}$, the function $y \mapsto \exp(-ay) \phi(x,y)$ is non-decreasing.
\end{enumerate}
We define the function $\Phi : \mathbb{R} \rightarrow \mathbb{R}$ by
\begin{equation} \label{eq:defPhi}
\Phi(s) := \mathbb{E}[\phi(X,Y)|X+Y=s].
\end{equation}
We assume that $\Phi$ is well-defined. Then the function $s \mapsto \exp(-as) \Phi(s) $ is non-decreasing.
\end{theo}
If we take $a=0$ then conditions of Theorem \ref{theorem2} are equivalent to $\phi$ being a non-decreasing function in each variable and Theorem \ref{theorem2} is equivalent to Efron's theorem.
\medskip
Under additional regularity assumptions on $\phi$, Theorem~\ref{theorem2} can be restated as follows:
\begin{cor}
\label{corollary}
Let X,Y be two independent log-concave real random variables with densities. \\
Let $a \geq 0$ be a real parameter and $\phi : \mathbb{R}^2 \rightarrow \mathbb{R}$ be a function such that:
\begin{enumerate}
\item $\forall (x,y) \in \mathbb{R}^2 \ , \ {\partial}_1 \phi(x,y) \geq a \phi(x,y)$.
\item $\forall (x,y) \in \mathbb{R}^2 \ , \ {\partial}_2 \phi(x,y) \geq a \phi(x,y)$.
\end{enumerate}
If the function $\Phi : \mathbb{R} \rightarrow \mathbb{R}$ defined as in equation~\eqref{eq:defPhi} is differentiable, then it satisfies:
\begin{equation*}\forall s \in \mathbb{R}, \Phi'(s) \geq a \Phi(s).
\end{equation*}
\end{cor}
We give another equivalent statement for Theorem~\ref{theorem2}: if ${\phi}_1 : \mathbb{R}^2 \rightarrow \mathbb{R}$ is defined by
$${\phi}_1(x,y) := \exp(a(x+y))$$
and if we set $\phi_2 = \phi$, then the conclusion of Theorem \ref{theorem2} is equivalent to saying that $({\Phi}_1,{\Phi}_2) \in GM_2$ where $\forall s \in \mathbb{R}, {\Phi}_i(s) = \mathbb{E}[{\phi}_i(X,Y)|X+Y=s]$.
\bigskip
We turn now to the statement of the generalized Efron's Theorem in a discrete setting:
\begin{theo}
\label{theorem2_2} Let $X,Y$ be two independent log-concave random variables on $\mathbb{Z}$. Let $a \geq 0$ be a real parameter and $\phi : \mathbb{Z}^2 \rightarrow \mathbb{\mathbb{R}}$ be a function such that:
\begin{enumerate}
\item For any $ k_2 \in \mathbb{Z}$, the function $k_1 \mapsto \exp(-ak_1) \phi(k_1,k_2)$ is non-decreasing.
\item For any $ k_1 \in \mathbb{Z}$, the function $k_2 \mapsto \exp(-a k_2) \phi(k_1,k_2)$ is non-decreasing.
\end{enumerate}
We define the function $\Phi : \mathbb{Z} \rightarrow \mathbb{R} $ by
\begin{equation*}
\Phi(s) := \mathbb{E}[\phi(X,Y)|X+Y=s].
\end{equation*}
We assume that $\Phi$ is well-defined. Then the function $s \mapsto \exp(-as) \Phi(s)$ is non-decreasing.
\end{theo}
We might notice that conditions on $\phi$ are equivalent to:
\begin{enumerate}
\item $\forall (x,y) \in \mathbb{Z}^2, ~ \phi(x+1,y) \geq e^a \phi(x,y)$,
\item $\forall (x,y) \in \mathbb{Z}^2, ~ \phi(x,y+1) \geq e^a \phi(x,y)$.
\end{enumerate}
Furthermore, if we denote ${\Delta}_1 \phi(x,y) = \phi(x+1,y) - \phi(x,y)$ and ${\Delta}_2 \phi(x,y) = \phi(x,y+1) - \phi(x,y)$ the discrete derivative operators, then the same conditions are also equivalent to:
\begin{enumerate}
\item $\forall (x,y) \in \mathbb{Z}^2, ~ {\Delta}_1 \phi(x,y) \geq (e^a-1) \phi(x,y)$,
\item $\forall (x,y) \in \mathbb{Z}^2, ~ {\Delta}_2 \phi(x,y) \geq (e^a-1) \phi(x,y)$.
\end{enumerate}
We can remark that, under this form, these conditions can be seen as a discrete version of conditions of Corollary \ref{corollary}.
\subsubsection{Examples}
We will consider, in this paragraph, examples for which variables are separated, i.e. functions $\phi : \mathbb{R}^2 \rightarrow \mathbb{R}$ that can be writtten as a product of two non-negative real functions:
\begin{equation*}
\forall (x,y) \in \mathbb{R}^2, \phi(x,y) = f(x) g(y).
\end{equation*}
We assume that:
\begin{enumerate}
\item For any $y \in \mathbb{R}$, the function $x \mapsto \exp(-ax) f(x)$ is non-decreasing.
\item For any $x \in \mathbb{R}$, the function $y \mapsto \exp(-ay) g(y)$ is non-decreasing.
\end{enumerate}
We denote the function $\Phi$ by
$$\Phi(s) := \mathbb{E}[\phi(X,Y) | X+Y=s].$$
We assume that the function $\Phi$ is well-defined. Then by Theorem \ref{theorem2}, we have that the function $s \mapsto \exp(-as) \Phi(s)$ is non-decreasing.
\medskip
If $f$ and $g$ can be written under the form $f(x) = \exp(u(x))$ and $g(y) = \exp(v(y))$ where $u$ and $v$ are two real functions, then \begin{equation*}
\forall (x,y) \in \mathbb{R}^2 \ , \ \phi(x,y)= \exp(u(x)) \exp(v(y)) = \exp(u(x) + v(y)).
\end{equation*} Then, conditions of Theorem \ref{theorem2} are equivalent to:
\begin{enumerate}
\item For any $y \in \mathbb{R}$, the function $x \mapsto u(x) - ax$ is non-decreasing.
\item For any $x \in \mathbb{R}$, the function $y \mapsto v(y) - ay$ is non-decreasing.
\end{enumerate}
For a non-trivial example, we can consider
\begin{equation*} \phi(x,y) = \exp[x(x^2+ \alpha) + y(y^2 +\beta)].
\end{equation*}
By Theorem \ref{theorem2}, we have that the function $s \mapsto \exp(-\gamma s) \Phi(s)$ is non-decreasing with the choice of parameter $\gamma = \min(\alpha,\beta)$.
\subsection{Proof of Theorem 3 and 4}
Theorem \ref{theorem2} and Theorem \ref{theorem2_2} is a direct consequence of Efron's theorem. \\
We have assumed that $\Phi$ is well-defined. Then, $\mathbb{E}[\exp(-a(X+Y))\phi(X,Y)|X+Y=s]$ is also well-defined because we have that for any $s \in \mathbb{R}$:
\begin{equation*}
\mathbb{E}[\exp(-a(X+Y))\phi(X,Y)|X+Y=s] = \mathbb{E}[\exp(-as)\phi(X,Y)|X+Y=s] = \exp(-as) \Phi(s).
\end{equation*}
We can remark that for any $(x,y) \in \mathbb{R}^2$ :
\begin{equation*}
\exp(-a(x+y)) \phi(x,y) = \exp(-ax) \exp(-ay) \phi(x,y).
\end{equation*}
Then, the function $(x,y) \mapsto \exp(-a(x+y)) \phi(x,y)$ is non-decreasing in each variable because $x \mapsto \exp(-ax)\phi(x,y)$ and $y \mapsto \exp(-ay)\phi(x,y)$ are non-decreasing. \\
By the Efron's theorem, apply to the function $(x,y) \mapsto \exp(-a(x+y))\phi(x,y)$, we have that the function $s \mapsto \mathbb{E}[\exp(-a(X+Y))\phi(X,Y)|X+Y=s]$ is non-decreasing. But we have that :
\begin{equation*}
\mathbb{E}[\exp(-a(X+Y))\phi(X,Y)|X+Y=s] = \alpha(s) \Phi(s).
\end{equation*}
By definition of $\Phi$, we obtain $\mathbb{E}[\exp(-a(X+Y))\phi(X,Y)|X+Y=s] = \exp(-as) \Phi(s)$. \\
Finally, the function $s \mapsto \exp(-as) \Phi(s)$ is a non-decreasing function.
\section{Some generalizations of Theorem \ref{theorem2} and Theorem \ref{theorem2_2}}
In this part, we will show a more general result. We will adapt the result of the previous part for a more general real function $\alpha$. After, we will apply this result to some example.
\begin{prop}
\label{efron_general}
Let $X,Y$ be two independent log-concave real-valued random variables. Let $\phi : \mathbb{R}^2 \rightarrow \mathbb{R}$ a measurable function and $\alpha : \mathbb{R} \rightarrow \mathbb{R}$ a measurable function such as :
\begin{enumerate}
\item For any $y \in \mathbb{R}$, the function $x \mapsto \alpha(x+y) \phi(x,y)$ is non-decreasing.
\item For any $x \in \mathbb{R}$, the function $y \mapsto \alpha(x+y) \phi(x,y)$ is non-decreasing.
\end{enumerate}
We define the function $\Phi : \mathbb{R} \rightarrow \mathbb{R}$ by :
\begin{equation*}
\Phi(s):= \mathbb{E}[\phi(X,Y)|X+Y=s].
\end{equation*}
We assume that $\Phi$ is well-defined. Then, we have that the function $s \mapsto \alpha(s) \Phi(s)$ is non-decreasing. \\
\end{prop}
\begin{proof}
By assumption, we have that $\Phi$ is well-defined. Then, we have that $\mathbb{E}[\alpha(X+Y) \phi(X,Y) | X+Y =s]$ is well-defined because :
$$\mathbb{E}[\alpha(X+Y) \phi(X,Y) | X+Y =s] = \mathbb{E}[\alpha(s) \phi(X,Y) | X+Y =s] = \alpha(s) \Phi(s).$$
We apply the Efron's theorem to the function $(x,y) \mapsto \alpha(x+y) \phi(x,y)$. Then we have that $s \mapsto \mathbb{E}[\alpha(X+Y) \phi(X,Y)|X+Y=s]$ is non-decreasing. But, we can remark that :
\begin{equation*}
\forall s \in \mathbb{R}, \mathbb{E}[\alpha(X+Y) \phi(X,Y)|X+Y=s]= \mathbb{E}[\alpha(s) \phi(X,Y)|X+Y=s].
\end{equation*}
Finally, the function $s \mapsto \alpha(s) \Phi(s)$ is non-decreasing.
\end{proof}
\begin{cor}
Let $X,Y$ be two independent log-concave real-valued random variables. Let $f: \mathbb{R} \rightarrow \mathbb{R}$ and $g: \mathbb{R} \rightarrow \mathbb{R}$ be two real functions. Let $\alpha : \mathbb{R} \rightarrow \mathbb{R}$ a measurable function such as :
\begin{enumerate}
\item For any $\lambda \in \mathbb{R}$, the function $x \mapsto \alpha(x+\lambda) f(x)$ is non-decreasing.
\item For any $\lambda \in \mathbb{R}$, the function $y \mapsto \alpha(y + \lambda) g(y)$ is non-decreasing.
\end{enumerate}
We define the function $\Phi : \mathbb{R} \rightarrow \mathbb{R}$ by :
\begin{equation*}
\Phi(s):= \mathbb{E}[f(X)g(Y)|X+Y=s].
\end{equation*}
We assume that $\Phi$ is well-defined. Then, we have that the function $s \mapsto \alpha(s) \Phi(s)$ is non-decreasing.
\end{cor}
\begin{proof}
To prove the corollary, we have only to apply the Proposition \ref{efron_general} to the function $\phi : (x,y) \mapsto f(x) g(y)$.
\end{proof}
We can remark that the Theorem 3 and 4 are a particular case of the proposition \ref{efron_general}. In the case of Theorem 3 and 4, we have that $\alpha(s) := \exp(-as)$. \\
The Efron's Theorem is the particular case where $\alpha(s):= 1$. \\
\medskip
Now, we will show an example of application of the Proposition \ref{efron_general}. \\
In this proposition, we will apply the proposition \ref{efron_general} to the function $\phi(x,y) := xy$ and $\alpha(s) := \dfrac{1}{s}$.
\begin{prop}
Let $X,Y$ be two independent log-concave real-valued random variables. We assume that $X,Y$ are positive.
We define the function $\Phi : \mathbb{R}_{+}^{*} \rightarrow \mathbb{R}$ by :
\begin{equation*}
\Phi(s):= \mathbb{E}[XY|X+Y=s].
\end{equation*}
We assume that $\Phi$ is well-defined. Then, we have that the function $s \mapsto \dfrac{\Phi(s)}{s}$ is non-decreasing on $\mathbb{R}_{+}^{*}$.
\end{prop}
\begin{proof}
We define the function $\Psi : (\mathbb{R}_{+}^{*})^2 \rightarrow \mathbb{R}$ by :
\begin{equation*}
\Psi(x,y) := \dfrac{xy}{x+y}.
\end{equation*}
The function $\Psi$ is differentiable and we have that :
\begin{equation*}
\forall (x,y) \in (\mathbb{R}_{+}^{*})^2 , \dfrac{\partial \Psi}{\partial x}(x,y) = \dfrac{y^2}{(x+y)^2} \geq 0.
\end{equation*}
\begin{equation*}
\forall (x,y) \in (\mathbb{R}_{+}^{*})^2, \dfrac{\partial \Psi}{\partial y}(x,y) = \dfrac{x^2}{(x+y)^2} \geq 0.
\end{equation*}
Then, the function $\Psi$ is non-decreasing in each variable. By the proposition \ref{efron_general}, we obtain that the function $s \mapsto \dfrac{\Phi(s)}{s}$ is a non-decreasing function.
\end{proof}
| 0589c8408903b7f27e26cdab8e5d2dc56029a3f5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Robots are predicted to increasingly participate in everyday life of humans over the next decade, but so far most research applications deal with interactions between a robot and a single human. However, if robots are to perform general tasks such as pick-up and delivery in public spaces, they will need to be able to cope and interact with large crowds. This poses a specific challenge for navigation control, where robots are supposed to move fast and efficient, while not endangering the safety of humans walking about in the same location.
Most work so far on robot navigation in environments with moving obstacles has been focused on a combination of pathfinding algorithms and human behaviour prediction \cite{bai2015intention}. However, more recently, reinforcement learning has been proposed for robot navigation through crowds \cite{chen2017decentralized}, but there has been no conclusive empirical comparison so far to pathfinding with prediction. In this paper, we present an evaluation of various pedestrian prediction techniques combined with the state-of-the-art pathfinding algorithm D* Lite \cite{d-star-lite}, and a recent reinforcement learning approach SA-CADRL \cite{chen2017socially}.
We evaluate the different approaches on a dataset extracted from real crowd movements recorded by CCTV at Grand Central Station in New York \cite{yi2015understanding}. The evaluation criteria are speed of movement from start to end point, and number of collisions of the virtual robot with humans. Our empirical results demonstrate that the reinforcement learning approach results in one of the fastest robots, while significantly reducing the number of collisions with humans, as measured in simulation based on real human movement data.
\section{Related work}
Research in the area of socially aware robot motion follows three main directions: behavior prediction, reinforcement learning, and behavior cloning.
Most of the work so far has focused on predicting humans' trajectories and plan the robot movement accordingly \cite{ferrer2014proactive}. For the prediction of humans' trajectories, clustering and typical behavior extraction was shown to be effective (e.g. \cite{ferrer2014behavior}). More recently, Deep Learning techniques have been successfully applied to the prediction of individual and crowd motion \cite{yi2016pedestrian}, which is currently the best performing technique. For robot route planning with dynamic obstacles the approaches vary. The simplest approach is to use a pre-defined set of rules to move the robot \cite{sisbot2007human}. Another approach is to compute attracting and repelling steering forces based on current and (where available) predicted future positions of obstacles \cite{reynolds1999steering}. This steering technique has been extended with Markov Decision Processes to deal with uncertainty in the predictions \cite{bai2015intention}. Svenstrup et al. \cite{svenstrup2010trajectory} iteratively improve the current best path using a Rapidly-exploring Random Tree with obstacles represented as potential fields.
In a recent reinforcement learning approach \cite{chen2017decentralized}, the robot was trained in simulation to move optimally through a crowd.
The third main approach to socially aware robot motion, behavior cloning, relies on the fact that humans can navigate in human crowds quite successfully and tries to imitate human behavior. This movement behaviour imitation problem has recently been tackled with inverse reinforcement learning, computing a reward function for crowd navigation from human crowd movement demonstrations, that was subsequently used to train a reinforcement learning agent \cite{kretzschmar2016socially}.
In this paper, we are focusing on the first two approaches, and perform a comparative evaluation in a simulation based on real-world pedestrian crowd recordings.
\section{Crowd movement dataset}
In order to simulate realistic crowd movements, we used a dataset obtained from surveillance video from Grand Central Station (New York, USA). The dataset consists of 6001 frames of camera images taken from the top, and 12684 labeled pedestrian trajectories as obtained by Yi et al. \cite{yi2015understanding}. An example frame with a single pedestrian trajectory is shown in Figure~\ref{fig:big_path}. The frames were sampled from a video of 4000 secs (i.e., a frequency is $1.5$ frames per second), and each frame has a resolution of $1920 \times 1080$. Trajectories were filtered by minimum length, where only those with at least 10 frames were retained. This left 12244 valid trajectories in the dataset with an average length of $37$ frames.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/big_path.jpg}
\caption{Example frame with sample trajectory for one person.}
\label{fig:big_path}
\end{figure}
In a data pre-processing step, we overlaid the frames with a $64 \times 36$ grid, as shown in Figure~\ref{fig:grid}. This allows discretization of the pedestrian positions and reduces the size of the state space for the path planning. The grid resolution was chosen based on average human speed from frame to frame which are $35$ pixels for the horizontal coordinate and $30$ pixels for the vertical one.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/frame.jpg}
\caption{Overlaying a $64 \times 36$ grid.}
\label{fig:grid}
\end{figure}
\section{Path planning methods} \label{sec:prediction}
We employed two general approaches to the path planning problem. In the first approach pedestrians are considered as moving obstacles and the goal is to predict their positions over the next five time steps. These positions are then used as dynamic obstacles for the path planning algorithms such as D* Lite \cite{d-star-lite}, which will be described later. The second approach is to learn a robot movement policy via reinforcement learning in simulation based on the recorded pedestrian movements. We start with a discussion of prediction methods for pedestrian positions.
\subsection{Pedestrian position prediction approach}
We compared several methods for position prediction in our study to demonstrate their impact on the robot policy.
\subsubsection{Baseline prediction}
A very simple method to predict the future position of a pedestrian is to assume that the pedestrian can move into any neighboring grid square in a time step. All these squares are considered to be occupied by the pedestrian in the next time step when performing the path planning.
\subsubsection{Random forest regression prediction}
A more sophisticated approach is to employ machine learning techniques for the prediction task. We chose Random Forest Regressor \cite{liaw2002classification} for this as a standard technique which has demonstrated robust performance in the literature. In this method, the prediction outputs are the five expected coordinates of the pedestrian over the next five time steps, relative to their original position.
The state for which a prediction is made was represented using features based on the five previous grid positions of the pedestrian as follows:
\begin{itemize}
\item Absolute coordinates (10 features).
\item Coordinates relative to the previous point ($x$ and $y$ speed, 8 features).
\item Speed (distance between neighboring points, 4 features) and average speed (1 feature).
\item Acceleration (change in speed from one step to the next one, 3 features) and average acceleration (1 feature).
\item Angle of movement for each step (4 features).
\end{itemize}
The features are illustrated on Fig. \ref{fig:features}
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{images/features_white_corrected.jpg}
\caption{Features for Random Forest Regressor. The top left shows the x and y coordinates of the last five trajectory points. The top right picture shows the distance between neighbouring trajectory points. The bottom left picture shows the coordinates relative to the previous trajectory point. The bottom right picture shows the changes in movement angle between two time steps.}
\label{fig:features}
\end{figure}
\subsubsection{Behavior-CNN prediction}
Behavior-CNN \cite{yi2016pedestrian} is a Neural Network architecture which was designed especially for movement prediction. It takes as input the embedded positions of the pedestrian over the last five time steps, and outputs the predicted positions over the next five time steps.
These embedded positions are called a displacement volume and correspond to a three-dimensional matrix of size $X \times Y \times 10$, where $X$ and $Y$ denote the area size (in our case $X = 1920$ and $Y = 1080$). The third dimension corresponds to the distances between the pedestrian positions over the last 5 time steps, along the x and y dimension. This results in 10 values. The distance values are first normalized to be between 0 and 1 according to $X$ and $Y$ respectively, and then 1 is added to is to avoid too small numbers. More formally, the 10-dimensional vector is computed as given in equation~\ref{eq:bcnn}, where ($x_t$,$y_t$) is the $i$-th position of the pedestrian with $t=1$ being the time five steps ago, and $t=5$ the current time step:
\begin{equation} \label{eq:bcnn}
[1 + \frac{x_5 - x_1}{X}, 1 + \frac{y_5 - y_1}{Y}, ..., 1 + \frac{x_5 - x_5}{X}, 1 + \frac{y_5 - y_5}{Y}]
\end{equation}
All other cells in the matrix are set to zero. The output the CNN has the same shape as the input and represents the predicted embedded positions over the next five time steps.
For other technical details we refer the reader to the original paper.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{images/beh_framework.png}
\caption{Behaviour-CNN framework \cite{yi2016pedestrian}.}
\label{fig:beh-cnn}
\end{figure*}
\subsubsection{Path planning with the D* Lite algorithm} \label{sec:d-star-lite}
D* Lite \cite{d-star-lite} is a heuristic path planning algorithm that takes as input a start and goal position, and recalculates the current path at each step taking into account the predicted coordinates of obstacles over a number of time steps. The method uses the freespace assumption, i.e. it assumes that unknown space is empty unless the predictions tell otherwise. We refer the reader to the original paper for more details.
Realization in Python 3 was based on the java implementation GitHub \footnote{https://github.com/daniel-beard/DStarLiteJava}.
\subsection{Reinforcement learning approach}
\subsubsection{SA-CADRL}
A state-of-the-art RL approach to socially aware robot movement is SA-CADRL (Socially Aware Collision Avoidance with Deep Reinforcement Learning) \cite{chen2017socially} which aims to learn an optimal robot policy based on the observed state of the robot and nearby agents (i.e. pedestrians). During training, the robot receives a positive reward for reaching the destination and a negative reward for delays and collisions. The policy is represented as a neural network whose input is formed from state information on the robot and nearby agents (pedestrians) and that outputs the robot velocity. This state information includes the current position, size, and velocity of the agent/robot.
Formally speaking, by using SA-CADRL architecture authors are trying to restore robot's policy $\pi$ such as expected value to goal would be minimal:
$$\arg \min_{\pi(s)} \mathbf{E}[t_g | \pi]$$
With the following restrictions:
\begin{itemize}
\item $||p_t - \hat{p}_t||_2 \geq r + \hat{r}$ -- for any frame distance between robot and other agents should be greater than their radii sum.
\item $p_{t_g} = p_g$ -- at the last frame robot should be in the goal coordinates.
\item $p_t = p_{t-1} + \Delta t \cdot \pi(s)$ -- current robot's position is defined by previous position and chosen action (moving direction).
\end{itemize}
To evaluate this approach on the same trajectories and with the same metrics as the other approaches, we used the pre-trained network provided by the Chen et al. which can be found in the SA-CADRL project GitHub page \footnote{https://github.com/mfe7/cadrl\_ros}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{images/sacadrl.png}
\caption{Original SA-CADRL network architecture \cite{chen2017socially}.}
\label{fig:sa-cadrl}
\end{figure}
\subsubsection{Grid-SA-CADRL}
A limitation of the SA-CADRL approach is that the robot can't move backwards. We extended the SA-CADRL approach to provide the robot with these abilities. Furthermore, we modified SA-CADRL to work on a grid-environment rather than a coordinate space as the original approach. This means that rather than having a continuous movement action, the actions are made discrete corresponding to a move into one of the 8 neighboring grid squares or standing still.
The resulting network architecture is shown in Figure~\ref{fig:network}. It takes the robots and three nearby pedestrians' states as inputs into an LSTM layer (size 64), followed by 3 fully connected layers with the last one outputting probabilities of actions through a softmax transformation. The robot then choose the action with the highest probability.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/jbrsacadrl.png}
\caption{Grid-SA-CADRL network architecture.}
\label{fig:network}
\end{figure}
We trained the Grid-SA-CADRL network for 1000 episodes, taking the pre-trained SA-CADRL network as a starting point. In each training episode we randomly picked a start and goal position for the robot, and a random starting time from the recordings. The pedestrians in the simulation walk according to the recording data, and an episode ends as soon as the robot has reached the goal position, or the recording has reached its end.
\section{Evaluation}
We first present the evaluation of the prediction method accuracy, followed by the evaluation of the path planning approaches.
\subsection{Evaluation of prediction methods}
We chose the Normalized Mean Square Error (NMSE, see equation below) to evaluate the pedestrian position prediction methods presented in Section~\ref{sec:prediction}, and optimize the hyper-parameters.
\begin{equation}
NMSE = \frac{1}{n}\sum^{n}_{i=1}{\sqrt{(\frac{x_i - \hat{x_i}}{X})^2 + (\frac{y_i - \hat{y_i}}{Y})^2}}
\end{equation}
Here $x_i$ and $y_i$ are ground truth values, $x_i$ and $y_i$ are predicted. $X$ and $Y$ are upper limits for $x_i$ and $y_i$ respectively, and in our case set to $X = 1920$ and $Y = 1080$.
We randomly split the data set of pedestrian trajectories into 80\% training, 10\% validation, and 10\% test data. For Random Forest Regression, the NMSE was 0.27, and Behavioral CNN significantly lowered this to 0.02. To show how this accuracy difference impacts on the performance of the robot, we used both prediction methods in the D* lite path planning algorithm.
\subsection{Evaluation of path planning}
All path planning algorithms were evaluated in terms of pedestrian safety and time taken to move from the start to the goal position. These metrics were measured as follows:
\begin{itemize}
\item Number of collisions (i.e. states in which the robot ends up in the same cell as a pedestrian) which were split into 3 groups as shown in Figure~\ref{fig:collision_types}. These represent a robot colliding a stationary pedestrian, a pedestrian colliding with a stationary robot, and a moving robot and a moving pedestrian colliding with each other.
\item Relative delay compared to the time of the optimal route with no pedestrians present, as computed by $D = (\frac{\hat{t}}{t} - 1) \times 100\%$, where $t$ is the time taken to move along the optimal route from start to goal, and $\hat{t}$ is the actual time taken.
\end{itemize}
In order to evaluate the path planning approaches, we generated 1000 episodes with random start and goal positions for the robot, and random starting times of the recording data.
While the SA-CADRL agent moves in a continuous coordinate space, this movement needed to be transformed into the grid-space model of the environment. Figure~\ref{fig:route} shows an example of this mapping. In the first time step the SA-CADRL agent moves one step to the north, and then moves diagonally to the north east. In the third step, however, the agent moves, but does not leave the current grid cell. This means that the agent was stationary for one time step in the grid-space representation. Note that the preferred speed of the SA-CADRL agent was set to match the grid cell size.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/route.jpg}
\caption{Mapping of vector space movement of SA-CADRL agent to the grid-space representation.}
\label{fig:route}
\end{figure}
Note that the pedestrian movement was recorded without a robot being present, which limits the realism of the simulation to some extent since it ignores the effect that robot movements could influence pedestrian decisions. However, the definition of the three different collision types has been specifically designed to take this limitation into account. Specifically, a moving robot colliding with a moving pedestrian is always an indicator for sub-optimal robot behaviour, no matter whether the pedestrian movement was influenced by the robot or not.
\section{Results}
The results of our evaluation are shown in table~\ref{tab:results}. While the baseline prediction approach may appear to be the most cautious one by assuming that a pedestrian is going to occupy all neighboring squares in the next time step, it caused the most collisions. There were due to some pedestrians moving faster than one square per time step. Increasing the range of squares that are assumed to be occupied to two does not improve the solution, since this causes the robot to stop moving at all. This confirms that a simple over-cautious solution is not feasible in a crowded space. In terms of movement speed, all other techniques demonstrate comparable performance.
The D* Lite method does not cause any SR collisions because it never is stationary. While the RL approaches do cause a relatively large number of these collisions, these are not indicative of poor robot behaviour, since pedestrians in a real-world situation are unlikely to collide with a stationary robot, and if they do it would not be the robot's fault. SP collisions do not occur for any of the approaches.
Most importantly, MRP collisions are relatively frequent for the D* Lite methods, corresponding to the accuracy of the chosen prediction method. The RL methods clearly outperform D* Lite on this measure, and our adaptation of SA-CADRL is more than halving the count of MRP collisions to a relatively small number of 3 in 1000. This clearly shows the superiority of Grid-SA-CADRL in this environment.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/collision_types.jpg}
\caption{The three types of collisions between the robot (red square) and a pedestrian (blue square).}
\label{fig:collision_types}
\end{figure}
\begin{table}[ht]
\caption{Evaluation of different approaches to robots navigation in crowded environment (taken over 1000 paths)}
\begin{tabular}{|l|p{1cm}|p{1cm}|p{1cm}|p{1cm}|p{1cm}|}
\hline
\multirow{3}{*}{Approach} &
\multirow{3}{*}{delay} &
\multicolumn{3}{c|}{collisions} \\ \cline{3-5}
& &
SR &
SP &
MRP \\ \hline \hline
\multicolumn{5}{|c|}{D* Lite based approach} \\ \hline \hline
Perfect prediction (0.0) & $1.44\%$ & 0 & 0 & 0 \\ \hline
Baseline & $7.57\%$ & 0 & 0 & 143 \\ \hline \hline
RFR (0.27) & $1.39\%$ & 0 & 0 & 67 \\ \hline
Behavior-CNN (0.02) & $1.42\%$ & 0 & 0 & 25 \\ \hline \hline
\multicolumn{5}{|c|}{Deep Reinforcement Learning approach} \\ \hline \hline
SA-CADRL (checkpoint) & $1.57\%$ & 183 & 0 & 8 \\ \hline
Grid-SA-CADRL & $1.46\%$ & 51 & 0 & 3 \\ \hline
\end{tabular}
\label{tab:results}
\end{table}
\section{Conclusions}
In this paper, we have presented a comparative evaluation of various path planning methods for robot navigation through crowds. The evaluation was carried out in a simulation based on real-world recorded pedestrian data in a crowded space. The results show the clear superiority of reinforcement learning in terms of number of collisions with moving pedestrians, while also never colliding with stationary pedestrians.
\section{Introduction}
Robots are predicted to increasingly participate in everyday life of humans over the next decade, but so far most research applications deal with interactions between a robot and a single human. However, if robots are to perform general tasks such as pick-up and delivery in public spaces, they will need to be able to cope and interact with large crowds. This poses a specific challenge for navigation control, where robots are supposed to move fast and efficient, while not endangering the safety of humans walking about in the same location.
Most work so far on robot navigation in environments with moving obstacles has been focused on a combination of pathfinding algorithms and human behaviour prediction \cite{bai2015intention}. However, more recently, reinforcement learning has been proposed for robot navigation through crowds \cite{chen2017decentralized}, but there has been no conclusive empirical comparison so far to pathfinding with prediction. In this paper, we present an evaluation of various pedestrian prediction techniques combined with the state-of-the-art pathfinding algorithm D* Lite \cite{d-star-lite}, and a recent reinforcement learning approach SA-CADRL \cite{chen2017socially}.
We evaluate the different approaches on a dataset extracted from real crowd movements recorded by CCTV at Grand Central Station in New York \cite{yi2015understanding}. The evaluation criteria are speed of movement from start to end point, and number of collisions of the virtual robot with humans. Our empirical results demonstrate that the reinforcement learning approach results in one of the fastest robots, while significantly reducing the number of collisions with humans, as measured in simulation based on real human movement data.
\section{Related work}
Research in the area of socially aware robot motion follows three main directions: behavior prediction, reinforcement learning, and behavior cloning.
Most of the work so far has focused on predicting humans' trajectories and plan the robot movement accordingly \cite{ferrer2014proactive}. For the prediction of humans' trajectories, clustering and typical behavior extraction was shown to be effective (e.g. \cite{ferrer2014behavior}). More recently, Deep Learning techniques have been successfully applied to the prediction of individual and crowd motion \cite{yi2016pedestrian}, which is currently the best performing technique. For robot route planning with dynamic obstacles the approaches vary. The simplest approach is to use a pre-defined set of rules to move the robot \cite{sisbot2007human}. Another approach is to compute attracting and repelling steering forces based on current and (where available) predicted future positions of obstacles \cite{reynolds1999steering}. This steering technique has been extended with Markov Decision Processes to deal with uncertainty in the predictions \cite{bai2015intention}. Svenstrup et al. \cite{svenstrup2010trajectory} iteratively improve the current best path using a Rapidly-exploring Random Tree with obstacles represented as potential fields.
In a recent reinforcement learning approach \cite{chen2017decentralized}, the robot was trained in simulation to move optimally through a crowd.
The third main approach to socially aware robot motion, behavior cloning, relies on the fact that humans can navigate in human crowds quite successfully and tries to imitate human behavior. This movement behaviour imitation problem has recently been tackled with inverse reinforcement learning, computing a reward function for crowd navigation from human crowd movement demonstrations, that was subsequently used to train a reinforcement learning agent \cite{kretzschmar2016socially}.
In this paper, we are focusing on the first two approaches, and perform a comparative evaluation in a simulation based on real-world pedestrian crowd recordings.
\section{Crowd movement dataset}
In order to simulate realistic crowd movements, we used a dataset obtained from surveillance video from Grand Central Station (New York, USA). The dataset consists of 6001 frames of camera images taken from the top, and 12684 labeled pedestrian trajectories as obtained by Yi et al. \cite{yi2015understanding}. An example frame with a single pedestrian trajectory is shown in Figure~\ref{fig:big_path}. The frames were sampled from a video of 4000 secs (i.e., a frequency is $1.5$ frames per second), and each frame has a resolution of $1920 \times 1080$. Trajectories were filtered by minimum length, where only those with at least 10 frames were retained. This left 12244 valid trajectories in the dataset with an average length of $37$ frames.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/big_path.jpg}
\caption{Example frame with sample trajectory for one person.}
\label{fig:big_path}
\end{figure}
In a data pre-processing step, we overlaid the frames with a $64 \times 36$ grid, as shown in Figure~\ref{fig:grid}. This allows discretization of the pedestrian positions and reduces the size of the state space for the path planning. The grid resolution was chosen based on average human speed from frame to frame which are $35$ pixels for the horizontal coordinate and $30$ pixels for the vertical one.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/frame.jpg}
\caption{Overlaying a $64 \times 36$ grid.}
\label{fig:grid}
\end{figure}
\section{Path planning methods} \label{sec:prediction}
We employed two general approaches to the path planning problem. In the first approach pedestrians are considered as moving obstacles and the goal is to predict their positions over the next five time steps. These positions are then used as dynamic obstacles for the path planning algorithms such as D* Lite \cite{d-star-lite}, which will be described later. The second approach is to learn a robot movement policy via reinforcement learning in simulation based on the recorded pedestrian movements. We start with a discussion of prediction methods for pedestrian positions.
\subsection{Pedestrian position prediction approach}
We compared several methods for position prediction in our study to demonstrate their impact on the robot policy.
\subsubsection{Baseline prediction}
A very simple method to predict the future position of a pedestrian is to assume that the pedestrian can move into any neighboring grid square in a time step. All these squares are considered to be occupied by the pedestrian in the next time step when performing the path planning.
\subsubsection{Random forest regression prediction}
A more sophisticated approach is to employ machine learning techniques for the prediction task. We chose Random Forest Regressor \cite{liaw2002classification} for this as a standard technique which has demonstrated robust performance in the literature. In this method, the prediction outputs are the five expected coordinates of the pedestrian over the next five time steps, relative to their original position.
The state for which a prediction is made was represented using features based on the five previous grid positions of the pedestrian as follows:
\begin{itemize}
\item Absolute coordinates (10 features).
\item Coordinates relative to the previous point ($x$ and $y$ speed, 8 features).
\item Speed (distance between neighboring points, 4 features) and average speed (1 feature).
\item Acceleration (change in speed from one step to the next one, 3 features) and average acceleration (1 feature).
\item Angle of movement for each step (4 features).
\end{itemize}
The features are illustrated on Fig. \ref{fig:features}
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{images/features_white_corrected.jpg}
\caption{Features for Random Forest Regressor. The top left shows the x and y coordinates of the last five trajectory points. The top right picture shows the distance between neighbouring trajectory points. The bottom left picture shows the coordinates relative to the previous trajectory point. The bottom right picture shows the changes in movement angle between two time steps.}
\label{fig:features}
\end{figure}
\subsubsection{Behavior-CNN prediction}
Behavior-CNN \cite{yi2016pedestrian} is a Neural Network architecture which was designed especially for movement prediction. It takes as input the embedded positions of the pedestrian over the last five time steps, and outputs the predicted positions over the next five time steps.
These embedded positions are called a displacement volume and correspond to a three-dimensional matrix of size $X \times Y \times 10$, where $X$ and $Y$ denote the area size (in our case $X = 1920$ and $Y = 1080$). The third dimension corresponds to the distances between the pedestrian positions over the last 5 time steps, along the x and y dimension. This results in 10 values. The distance values are first normalized to be between 0 and 1 according to $X$ and $Y$ respectively, and then 1 is added to is to avoid too small numbers. More formally, the 10-dimensional vector is computed as given in equation~\ref{eq:bcnn}, where ($x_t$,$y_t$) is the $i$-th position of the pedestrian with $t=1$ being the time five steps ago, and $t=5$ the current time step:
\begin{equation} \label{eq:bcnn}
[1 + \frac{x_5 - x_1}{X}, 1 + \frac{y_5 - y_1}{Y}, ..., 1 + \frac{x_5 - x_5}{X}, 1 + \frac{y_5 - y_5}{Y}]
\end{equation}
All other cells in the matrix are set to zero. The output the CNN has the same shape as the input and represents the predicted embedded positions over the next five time steps.
For other technical details we refer the reader to the original paper.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{images/beh_framework.png}
\caption{Behaviour-CNN framework \cite{yi2016pedestrian}.}
\label{fig:beh-cnn}
\end{figure*}
\subsubsection{Path planning with the D* Lite algorithm} \label{sec:d-star-lite}
D* Lite \cite{d-star-lite} is a heuristic path planning algorithm that takes as input a start and goal position, and recalculates the current path at each step taking into account the predicted coordinates of obstacles over a number of time steps. The method uses the freespace assumption, i.e. it assumes that unknown space is empty unless the predictions tell otherwise. We refer the reader to the original paper for more details.
Realization in Python 3 was based on the java implementation GitHub \footnote{https://github.com/daniel-beard/DStarLiteJava}.
\subsection{Reinforcement learning approach}
\subsubsection{SA-CADRL}
A state-of-the-art RL approach to socially aware robot movement is SA-CADRL (Socially Aware Collision Avoidance with Deep Reinforcement Learning) \cite{chen2017socially} which aims to learn an optimal robot policy based on the observed state of the robot and nearby agents (i.e. pedestrians). During training, the robot receives a positive reward for reaching the destination and a negative reward for delays and collisions. The policy is represented as a neural network whose input is formed from state information on the robot and nearby agents (pedestrians) and that outputs the robot velocity. This state information includes the current position, size, and velocity of the agent/robot.
Formally speaking, by using SA-CADRL architecture authors are trying to restore robot's policy $\pi$ such as expected value to goal would be minimal:
$$\arg \min_{\pi(s)} \mathbf{E}[t_g | \pi]$$
With the following restrictions:
\begin{itemize}
\item $||p_t - \hat{p}_t||_2 \geq r + \hat{r}$ -- for any frame distance between robot and other agents should be greater than their radii sum.
\item $p_{t_g} = p_g$ -- at the last frame robot should be in the goal coordinates.
\item $p_t = p_{t-1} + \Delta t \cdot \pi(s)$ -- current robot's position is defined by previous position and chosen action (moving direction).
\end{itemize}
To evaluate this approach on the same trajectories and with the same metrics as the other approaches, we used the pre-trained network provided by the Chen et al. which can be found in the SA-CADRL project GitHub page \footnote{https://github.com/mfe7/cadrl\_ros}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{images/sacadrl.png}
\caption{Original SA-CADRL network architecture \cite{chen2017socially}.}
\label{fig:sa-cadrl}
\end{figure}
\subsubsection{Grid-SA-CADRL}
A limitation of the SA-CADRL approach is that the robot can't move backwards. We extended the SA-CADRL approach to provide the robot with these abilities. Furthermore, we modified SA-CADRL to work on a grid-environment rather than a coordinate space as the original approach. This means that rather than having a continuous movement action, the actions are made discrete corresponding to a move into one of the 8 neighboring grid squares or standing still.
The resulting network architecture is shown in Figure~\ref{fig:network}. It takes the robots and three nearby pedestrians' states as inputs into an LSTM layer (size 64), followed by 3 fully connected layers with the last one outputting probabilities of actions through a softmax transformation. The robot then choose the action with the highest probability.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/jbrsacadrl.png}
\caption{Grid-SA-CADRL network architecture.}
\label{fig:network}
\end{figure}
We trained the Grid-SA-CADRL network for 1000 episodes, taking the pre-trained SA-CADRL network as a starting point. In each training episode we randomly picked a start and goal position for the robot, and a random starting time from the recordings. The pedestrians in the simulation walk according to the recording data, and an episode ends as soon as the robot has reached the goal position, or the recording has reached its end.
\section{Evaluation}
We first present the evaluation of the prediction method accuracy, followed by the evaluation of the path planning approaches.
\subsection{Evaluation of prediction methods}
We chose the Normalized Mean Square Error (NMSE, see equation below) to evaluate the pedestrian position prediction methods presented in Section~\ref{sec:prediction}, and optimize the hyper-parameters.
\begin{equation}
NMSE = \frac{1}{n}\sum^{n}_{i=1}{\sqrt{(\frac{x_i - \hat{x_i}}{X})^2 + (\frac{y_i - \hat{y_i}}{Y})^2}}
\end{equation}
Here $x_i$ and $y_i$ are ground truth values, $x_i$ and $y_i$ are predicted. $X$ and $Y$ are upper limits for $x_i$ and $y_i$ respectively, and in our case set to $X = 1920$ and $Y = 1080$.
We randomly split the data set of pedestrian trajectories into 80\% training, 10\% validation, and 10\% test data. For Random Forest Regression, the NMSE was 0.27, and Behavioral CNN significantly lowered this to 0.02. To show how this accuracy difference impacts on the performance of the robot, we used both prediction methods in the D* lite path planning algorithm.
\subsection{Evaluation of path planning}
All path planning algorithms were evaluated in terms of pedestrian safety and time taken to move from the start to the goal position. These metrics were measured as follows:
\begin{itemize}
\item Number of collisions (i.e. states in which the robot ends up in the same cell as a pedestrian) which were split into 3 groups as shown in Figure~\ref{fig:collision_types}. These represent a robot colliding a stationary pedestrian, a pedestrian colliding with a stationary robot, and a moving robot and a moving pedestrian colliding with each other.
\item Relative delay compared to the time of the optimal route with no pedestrians present, as computed by $D = (\frac{\hat{t}}{t} - 1) \times 100\%$, where $t$ is the time taken to move along the optimal route from start to goal, and $\hat{t}$ is the actual time taken.
\end{itemize}
In order to evaluate the path planning approaches, we generated 1000 episodes with random start and goal positions for the robot, and random starting times of the recording data.
While the SA-CADRL agent moves in a continuous coordinate space, this movement needed to be transformed into the grid-space model of the environment. Figure~\ref{fig:route} shows an example of this mapping. In the first time step the SA-CADRL agent moves one step to the north, and then moves diagonally to the north east. In the third step, however, the agent moves, but does not leave the current grid cell. This means that the agent was stationary for one time step in the grid-space representation. Note that the preferred speed of the SA-CADRL agent was set to match the grid cell size.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/route.jpg}
\caption{Mapping of vector space movement of SA-CADRL agent to the grid-space representation.}
\label{fig:route}
\end{figure}
Note that the pedestrian movement was recorded without a robot being present, which limits the realism of the simulation to some extent since it ignores the effect that robot movements could influence pedestrian decisions. However, the definition of the three different collision types has been specifically designed to take this limitation into account. Specifically, a moving robot colliding with a moving pedestrian is always an indicator for sub-optimal robot behaviour, no matter whether the pedestrian movement was influenced by the robot or not.
\section{Results}
The results of our evaluation are shown in table~\ref{tab:results}. While the baseline prediction approach may appear to be the most cautious one by assuming that a pedestrian is going to occupy all neighboring squares in the next time step, it caused the most collisions. There were due to some pedestrians moving faster than one square per time step. Increasing the range of squares that are assumed to be occupied to two does not improve the solution, since this causes the robot to stop moving at all. This confirms that a simple over-cautious solution is not feasible in a crowded space. In terms of movement speed, all other techniques demonstrate comparable performance.
The D* Lite method does not cause any SR collisions because it never is stationary. While the RL approaches do cause a relatively large number of these collisions, these are not indicative of poor robot behaviour, since pedestrians in a real-world situation are unlikely to collide with a stationary robot, and if they do it would not be the robot's fault. SP collisions do not occur for any of the approaches.
Most importantly, MRP collisions are relatively frequent for the D* Lite methods, corresponding to the accuracy of the chosen prediction method. The RL methods clearly outperform D* Lite on this measure, and our adaptation of SA-CADRL is more than halving the count of MRP collisions to a relatively small number of 3 in 1000. This clearly shows the superiority of Grid-SA-CADRL in this environment.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.45\textwidth]{images/collision_types.jpg}
\caption{The three types of collisions between the robot (red square) and a pedestrian (blue square).}
\label{fig:collision_types}
\end{figure}
\begin{table}[ht]
\caption{Evaluation of different approaches to robots navigation in crowded environment (taken over 1000 paths)}
\begin{tabular}{|l|p{1cm}|p{1cm}|p{1cm}|p{1cm}|p{1cm}|}
\hline
\multirow{3}{*}{Approach} &
\multirow{3}{*}{delay} &
\multicolumn{3}{c|}{collisions} \\ \cline{3-5}
& &
SR &
SP &
MRP \\ \hline \hline
\multicolumn{5}{|c|}{D* Lite based approach} \\ \hline \hline
Perfect prediction (0.0) & $1.44\%$ & 0 & 0 & 0 \\ \hline
Baseline & $7.57\%$ & 0 & 0 & 143 \\ \hline \hline
RFR (0.27) & $1.39\%$ & 0 & 0 & 67 \\ \hline
Behavior-CNN (0.02) & $1.42\%$ & 0 & 0 & 25 \\ \hline \hline
\multicolumn{5}{|c|}{Deep Reinforcement Learning approach} \\ \hline \hline
SA-CADRL (checkpoint) & $1.57\%$ & 183 & 0 & 8 \\ \hline
Grid-SA-CADRL & $1.46\%$ & 51 & 0 & 3 \\ \hline
\end{tabular}
\label{tab:results}
\end{table}
\section{Conclusions}
In this paper, we have presented a comparative evaluation of various path planning methods for robot navigation through crowds. The evaluation was carried out in a simulation based on real-world recorded pedestrian data in a crowded space. The results show the clear superiority of reinforcement learning in terms of number of collisions with moving pedestrians, while also never colliding with stationary pedestrians.
| 7acafb082c69ce4d4984ca2a3e3bd8ffcac6abb0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Person re-identification (ReID) is an important task that matches person images across times/spaces/cameras, which has many applications such as people tracking in smart retail, image retrieval for finding lost children. Existing approaches achieve remarkable performance when the training and testing data are from the same dataset/domain. But they usually fail to generalize well to other datasets where there are domain gaps~\cite{ge2020mutual}. To address this practical problem, unsupervised domain adaptive (UDA) person ReID attracts much attention for both the academic and industrial communities, where labeled source domain and unlabeled target domain data are exploited for training.
Typical UDA person ReID approaches~\cite{ge2020mutual,zhai2020adcluster,zhong2019invariance,2020Hierarchical,song2020unsupervised} include three steps: feature pre-training with labeled source domain data, clustering-based pseudo-label prediction for the target domain data, and feature representation learning/fine-tuning with the pseudo-labels. The last two steps are usually iteratively conducted to promote each other. However, the pseudo-labels obtained/assigned through clustering usually contain noisy (wrong) labels due to the divergence/domain gap between the source and target data, and the imperfect results of the clustering algorithm. Such noisy labels would mislead the feature learning and harm the domain adaptation performance. Thus, \emph{alleviating the negative effects of those samples with unreliable/noisy pseudo labels is important for the success of domain adaptation}.
The challenge lies in 1) how to identify samples that are prone to have noisy pseudo labels; 2) how to alleviate their negative effects during the optimization. In this paper, to answer the first question, we have observed abundant samples and analyzed the relationship between the characteristics of the samples and the correctness of pseudo labels. Based on the theory on uncertainty~\cite{kendall2017uncertainties}, a model has uncertainty on its prediction of an input sample. Here, we measure the inconsistency level of the output features of two models (the student model and the teacher model based on the mean teacher method~\cite{tarvainen2017mean}) and take it as the estimated uncertainty of a target domain sample. \textcolor{black}{As shown in Fig.~\ref{fig:intro}, we observe the distribution of the uncertainty (inconsistency levels) for correct/clean pseudo labels and wrong pseudo labels. We found that the uncertainty values for the samples with wrong pseudo labels are usually larger} than those with correct pseudo labels. This motivates us to estimate and exploit the uncertainty of samples to alleviate the negative effects of noisy pseudo labels, enabling effective domain adaptation. We answer the second question by carefully incorporating the uncertainty of samples into classification loss, triplet loss, and contrastive loss, respectively.
We summarize our main contributions as follows:
\begin{itemize}[leftmargin=*,noitemsep,nolistsep]
\item We propose a network named Uncertainty-guided Noise Resilient Network (UNRN) to explore the credibility of the predicted pseudo labels of target domain samples for effective domain adaptive person ReID.
\item We develop an uncertainty estimation strategy by calculating the inconsistency of two models in terms of their predicted soft multilabels.
\item We incorporate the uncertainty of samples to the ID classification loss, triplet loss, and contrastive loss through re-weighting to alleviate the negative influence of noisy pseudo labels.
\end{itemize}
Extensive experiments demonstrate the effectiveness of our framework and the designed components on unsupervised person ReID benchmark datasets. Our scheme achieves the state-of-the-art performance on all the benchmark datasets.
\section{Related Work}
\subsection{Unsupervised Domain Adaptive Person ReID}
Existing unsupervised domain adaptive person ReID approaches can be grouped into three main categories.
\noindent\textbf{Clustering-based methods} in general generate hard or soft pseudo labels based on clustering results and then fine-tune/train the models based on the pseudo labels ~\cite{fan2018unsupervised,zhang2019self,yang2019selfsimilarity,ge2020mutual,yu2019unsupervised,zhong2019invariance, song2020unsupervised, jin2020global}. They are widely used due to their superior performance. PUL~\cite{fan2018unsupervised} iteratively obtains hard clustering labels and trains the model in self-training manner. SSG~\cite{yang2019selfsimilarity} exploits the potential similarity for the global body and local body parts, respectively, to build multiple independent clusters. These clusters are then assigned with labels to supervise the training. PAST~\cite{zhang2019self} optimizes the network with triplet-based loss function (to capture the local structure of target-domain data points) and classification loss by appending a classification layer (to use global information about the data distribution) based on clustering results.
Pseudo label noise caused by unsupervised clustering is always an obstacle to the self-training. Such noisy labels would mislead the feature learning and impede the achievement of high performance. Recently, some methods introduce mutual learning among two/three collaborative networks to mutually exploit the refined soft pseudo labels of the peer networks as supervision \cite{ge2020mutual,zhai2020multiple}. To suppress the noises in the pseudo labels, NRMT \cite{zhao2020unsupervised} maintains two networks during training to perform collaborative clustering and mutual instance selection, which reduces the fitting to noisy instances by using the mutual supervision and the reliable instance selection. These approaches need mutual learning of two or more networks and are somewhat complicated. Besides, the selection of reliable instances in NRMT~\cite{zhao2020unsupervised} is a hard rather than a soft selection, which requires a careful determination of threshold parameters and may lose the chance of exploiting useful information of the abandoned samples.
Different from the above works, in this paper, we propose a simple yet effective framework which estimates the reliability of the pseudo labels through uncertainty estimation and softly exploit them in the ReID losses to alleviate the negative effects of noise-prone samples. Note that we do not need two networks for mutual learning. We build our framework based on the mean teacher method, which maintains a temporally averaged model of the base network during the training, to facilitate the estimation of the uncertainty of each target domain sample.
\noindent\textbf{Domain translation-based methods} \cite{deng2018image,huang2020real,wei2018person,ge2020structured} transfer source domain labeled images to the style of the target domain images and use these transferred images and the inherited ground-truth labels to fine-tune the model. However, the quality of translated images is still not very satisfactory which hinders the advancement of these approaches.
\noindent\textbf{Memory bank based methods} have been widely used for unsupervised representation learning which facilitates the introduction of contrastive loss~\cite{he2020momentum} for the general tasks. He \textit{et al}.~ leverage the memory bank to better train the model to exploit the similarity between a sample and the instances in the global memory bank~\cite{he2020momentum}. Wang et al. propose to use memory bank to facilitate hard negative instance mining across batches \cite{wang2020cross}. ECN~\cite{zhong2019invariance} leverages memory bank to enforce exemplar-invariance, camera-invariance and neighborhood-invariance over global training data (instead of local batch) for UDA person ReID. We introduce contrastive loss to the target instance memory bank to enable the joint optimization of positive pairs and negative pairs for a query/anchor sample over all the samples in the memory bank, which serves as our strong baseline.
\begin{figure*}[t]
\centering
\includegraphics[width=0.90\linewidth]{figure2_1.jpg}
\caption{Overview of the proposed Uncertainty-guided Noise Resilient Network (UNRN) for UDA person ReID. We build our baseline framework with the mean teacher method (where the mean teacher model is a temporally moving average of weights of the student network) together with contrastive loss (supported by a memory bank). Our method belongs to clustering-based methods. In the model pre-training stage, we pre-train the network using source domain labeled data. In the clustering stage, we do clustering on the unlabeled target domain data using the more accurate features from the mean teacher model and assign pseudo labels based on the clustering results. Because of domain gap, some of the pseudo-labels are noisy/incorrect. In the joint fine-tuning stage, we propose to exploit the estimated uncertainty to evaluate the reliability of the pseudo-labels to alleviate the negative influence of the samples with error-prone pseudo-labels, by carefully incorporating the uncertainty to re-weight the contribution of samples in ID classification loss, triplet loss, and contrastive loss, respectively. Stage 2 and Stage 3 are performed alternatively.
\label{fig:fm}
\end{figure*}
\subsection{Uncertainty Estimation}
Uncertainty modeling helps us to understand what a model does not know (or is not confident). In order to suppress the noise during the training, existing works~\cite{kendall2017uncertainties, chang2020data, zheng2020rectifying,zheng2019abstract} have explored uncertainty estimation from different aspects, such as the data-dependent uncertainty, model uncertainty, and the uncertainty on annotation. For fully supervised learning tasks with clean groudtruth labels, some works learn the data-dependent uncertainty in an end-to-end manner to alleviate the influence of observation noise of an input sample for better network optimization~\cite{kendall2017uncertainties, chang2020data}.
Zheng \textit{et al}.~~\cite{zheng2020rectifying} use an extra auxiliary classifier to help estimate the uncertainty of the predicted pseudo labels for semantic segmentation. For clustering-based UDA person ReID, some of the pseudo labels generated by clustering are incorrect/noisy. However, the investigation on how to softly evaluate the reliability of clustering-generated pseudo labels is still underexplored. In this work, we explore the reliability estimation of the clustering-generated pseudo labels and alleviate the influence of noise-prone samples by re-weighting their contributions in ReID losses.
\section{\!\!\!\!Uncertainty-guided Noise Resilient Network}
Unsupervised domain adaptive person ReID aims at adapting the model trained on a labeled source domain dataset $\mathbb{D}_{s}\!\!=\!\!\left\{\!\!\left.\left(\boldsymbol{x}_{i}^{s}, {y}_{i}^{s}\right)\right|_{i=1} ^{N_{s}}\!\right\}$ \textcolor{black}{of $C_{s}$ identities} to an annotation-free target domain dataset $\mathbb{D}_{t}\!\!=\!\!\left\{\!\left.\boldsymbol{x}_{i}^{t}\right|_{i=1} ^{N_{t}}\right\}$. $N_{s}$ and $N_{t}$ denote the number of samples. $\boldsymbol{x}_{i}$ denotes a sample and ${y}_{i}$ denotes its groudtruth label.
Fig.~\ref{fig:fm} shows an overview of our proposed Uncertainty-guided Noise Resilient Network (UNRN) for UDA person ReID. We aim to address the negative influence of noise-prone pseudo labels during adaptation/fine-tuning under the clustering-based framework. We build a clustering-based strong baseline scheme \emph{SBase.} for UDA person ReID. On top of the strong baseline, we introduce an uncertainty calculation module to estimate the reliability of psedudo labels and incorporate the uncertainty into the ReID losses (ID classification loss (ID loss), triplet loss, and contrastive loss, respectively) to alleviate the influence of noise-prone pseudo labels. Our uncertainty-guided optimization design brings significant improvement on top of this strong baseline. We introduce the strong baseline in Section \ref{subsec:baseline} and elaborate on our proposed uncertainty-guided noise-resilient optimization designs in Section \ref{subsec:uncertainty}.
\subsection{Clustering-based Strong Baseline}
\label{subsec:baseline}
We build a clustering-based strong baseline scheme \emph{SBase.} for UDA person ReID. We follow the general pipeline of clustering-based UDA methods \cite{fan2018unsupervised, song2020unsupervised, jin2020global} which consists of three main stages, \textit{i}.\textit{e}., model pre-training, clustering, and fine-tuning. As shown in Fig.~\ref{fig:fm}, we exploit the labeled source domain data for fine-tuning, incorporate contrastive loss, and leverage the simple yet effective mean teacher method to have a strong baseline. We will present the ablation study of each component in the experimental section.
In the model pre-training stage, we pre-train the network using source domain labeled data. In the clustering stage, we do clustering on the unlabeled target domain data using the more accurate features from the mean teacher model and generate pseudo labels based on the clustering results.
\noindent\textbf{Joint fine-tuning using source data.} In the fine-tuning stage~\cite{zhang2019self}, \textcolor{black}{most works} fine-tune the networks only using target domain pseudo labels. Here, we also re-use the valuable source domain data with reliable groudtruth labels. \textcolor{black}{For a source sample, we add ID classification loss, where maintained $C_{s}$ class centers are used as the classification weight vectors for classification.}
\noindent\textbf{Contrastive loss across memory bank.} For image retrieval task, to enable the pair similarity optimization over informative negative instances, Wang \textit{et al}.~ propose a cross-batch memory mechanism that memorizes the feature embeddings of the past iterations to allow collecting sufficient hard negative pairs across the memory bank for network optimization~\cite{wang2020cross}. Motivated by this, for a \textcolor{black}{target domain} query sample $a$, we add contrastive loss to maximize the within-class similarity and minimize the between-class similarity across \textcolor{black}{the memory bank. Particularly, the memory bank consists of $N$ target domain instances (which is maintained similar to \cite{wang2020cross}) and $C_{s}$ source class center features (as the negative samples). Based on the pseudo labels of the target domain samples and the additional source class centers, we have $N_a^+$ positive samples and $N_a^- = N-N_a^+ + C_{s}$ negative samples. We optimize their similarities with respect to the query sample.} Following circle loss \cite{sun2020circle}, we use self-paced weighting to softly emphasize harder sample pairs by giving larger weights to them to get effective update.
\noindent\textbf{Mean teacher method.} Mean teacher is a method that temporally averages model weights over training steps, which tends to produce a more accurate model than using the final weights directly \cite{tarvainen2017mean}. As illustrated in Fig.~\ref{fig:fm}, there is no gradient back-propagation over the teacher model which just maintains a temporal moving average of the student model. This method is simple yet effective. We use the features from the teacher model to perform clustering and the final ReID inference.
\subsection{Uncertainty-guided Optimization}
\label{subsec:uncertainty}
\begin{figure}[!t]
\centering
\includegraphics[width=1.0\linewidth]{figure3.jpg}
\caption{Uncertainty estimation module. For a sample $\boldsymbol{x}_i$ and its feature $\boldsymbol{\tilde{f}}_i$ from the mean teacher model and feature $\boldsymbol{f}_i$ from the student model, we estimate the uncertainty of its pseudo label by calculating the KL divergence of their soft multilabels.}
\label{fig:fm2}
\end{figure}
The noisy pseudo labels would mislead the feature learning in the fine-tuning stage and hurt the adaptation performance. We aim to reduce the negative influence of noisy pseudo labels by evaluating the credibility of the pseudo label of each sample and suppress the contributions of samples with error-prone pseudo labels in the ReID losses.
\noindent{\textbf{Uncertainty estimation.}} Intuitively, the higher of the uncertainty the model has on its output of a sample, the lower of the reliability (larger of observation noise) of the output and it is more likely to have incorrect pseudo labels through clustering. We have observed in Fig.~\ref{fig:intro} that the samples with wrong pseudo labels in general have higher uncertainty values than those with correct pseudo labels. We leverage the uncertainty to softly assess the credibility of samples.
We estimate the uncertainty based on the output consistency between the mean teacher model and the student model. For a sample $\boldsymbol{x}_i$ of the target domain, we denote the extracted feature from the student model as $\boldsymbol{f}_i \in \mathbb{R}^{D}$ of $D$ dimensions, and the feature from the teacher model as $\boldsymbol{\tilde{f}}_i \in \mathbb{R}^{D}$. One straightforward solution to calculate the consistency between the mean teacher model and the student model is to calculate the distance (\egno, cosine distance) between the two features. However, this is clustering ignorance which does not capture/explore the global distribution of the target domain samples.
In MAR \cite{yu2019unsupervised}, they measure the class likelihood vector (\textit{i}.\textit{e}., soft multilabel) of a person image with a set of reference persons (from an auxiliary domain). The inconsistency with the soft multilabel vector of an anchor sample is used to mine the hard negative sample that is visually similar to the anchor image but is of a different identity. Inspired by this, we propose to evaluate the uncertainty based on the inconsistency of the two features $\boldsymbol{f}_i$ and $\boldsymbol{\tilde{f}}_i$, by calculating the distance (\egno, KL distance, L1 distance) of their soft multilabels with respect to the same set of cluster centers. Particularly, as illustrated in Fig.~\ref{fig:fm2}, we use the class centers of source dataset and the cluster centers of the target domain data together to form the set of ``reference persons''. The soft multilabel agreement is analog to the voting by the ``reference persons'' to evaluate the relative consistency. Besides, through the comparison with the cluster centers of the target domain data, the soft multilabel captures some global information of the clustering centers which is related to pseudo label assignment. As an auxiliary domain, the source centers provide additional references.
Let ${R} =[R_t, R_s] \in \mathbb{R}^{K_r \times D}$ denote a matrix which stores the features of the $K_r$ ``reference persons'', with each column denoting the feature of a ``reference person''. $R_t \in \mathbb{R}^{K_t \times D}$ denotes the $K_t$ cluster centers of the target domain data, and $R_s \in \mathbb{R}^{K_s \times D}$ denotes the $K_s$ class centers of the source domain data (which are obtained from the weights of the fully-connected layer of the ID classifier). $K_r=K_t+K_s$.
For a feature $\boldsymbol f_i \in \mathbb{R}^{D}$ from the student model, we calculate the similarity to the $K_r$ ``reference persons'' and obtain the soft multilabel (likelihood vector) as:
\begin{equation}
\boldsymbol p_i = {\rm{Softmax}} \left( R \cdot \boldsymbol f_i \right),
\end{equation}
where Softmax($\cdot$) denotes softmax function which normalizes the similarity scores. Similarly, for the feature $\boldsymbol{\tilde{f}}_i$ from the mean teacher model, we obtain its soft multilabel as $\boldsymbol{\tilde{p}}_i$. We use KL divergence to measure the difference between the two probability distributions from the two models as the uncertainty $u_i$ of the sample $\boldsymbol x_i$:
\begin{equation}
u_i = D_{KL}(\boldsymbol{\tilde{p}}_i || \boldsymbol p_i)= \sum_{k=1}^{K_r} \tilde{p}_{i,k} \log \frac{\tilde{p}_{i,k}}{p_{i,n}}.\\
\end{equation}
\noindent{\textbf{Optimization.}} We have observed that a sample with a wrong pseudo-label (through clustering), in general, has a higher uncertainty. Based on this observation, we propose to exploit the uncertainty to estimate the unreliability of the pseudo-label of a sample and use it to re-weight the contributions of samples in various ReID losses.
For a sample $\boldsymbol{x}_i$ with high uncertainty, we will reduce its contribution to the losses. Therefore, we could assign $\omega_i= 1/u_{i}$ as the credibility weight. To enable more stable training, we adopt the policy in~\cite{kendall2017uncertainties} and define $\omega_i = {\rm{exp}}({-u_i})$.
We incorporate the uncertainty-guided optimization in the classification loss, triplet loss, and contrastive loss, respectively.
For \emph{ID Classification loss}, we define the uncertainty-guided ID classification loss in a min-batch of $n_t$ target domain samples as
\begin{equation}
\mathcal{L}_{UID} =-\frac{1}{n_{t}} \sum_{i=1}^{n_{t}} \mathcal \omega_{i} \log p\left(\tilde{y}_{i} | \boldsymbol{x}_{i}\right),
\label{eq:9}
\end{equation}
where $p\left(\tilde{y}^t_{i} | \boldsymbol{x}^t_{i}\right)$ denotes the probability of being class $\tilde{y}^t_{i}$, where $\tilde{y}^t_{i}$ denotes the pseudo groudtruth class (based on the pseudo label assigned after clustering). For a sample with high uncertainty, a smaller weight is used to reduce its contribution to the overall loss to reduce its negative effect.
As a typical sample-pair similarity optimization, triplet loss is widely used in ReID to make the similarity between an anchor sample and a positive sample to be much larger than that between this anchor sample and negative sample. For the $j^{th}$ triplet of an anchor sample, a positive sample and a negative sample that correspond to three pseudo labels, we approximate the reliability of a sample pair, \egno, the positive sample pair, by a function of the two uncertainties as
\begin{equation}
\mathcal{\omega}_{ap}^j = \varphi(u_{a}^j, u_{p}^j),
\label{eq:omega}
\end{equation}
where $u_a^j$ and $u_p^j$ denote the estimated uncertainty for the anchor sample and positive sample in the $j^{th}$ triplet, respectively. For simplicity, we define the pair credibility as the average of two credibility as $\varphi(u_{a}^j, u_{p}^j) = \omega_a^j + \omega_p^j = {\rm{exp}}({-u_a^j}) + {\rm{exp}}({-u_p^j})$. Similarly, we get $\mathcal{\omega}_{an}^j$ for the negative sample pair.
For \emph{triplet loss}, we define the uncertainty-guided triplet loss in a min-batch of $n_{tr}$ triplets as
\small
\begin{equation}
\mathcal{L}_{UTRI}\!=-\frac{1}{n_{tr}} \sum\limits_{j=1}^{n_{tr}} \log \frac{\omega_{ap}^j \exp(s_{ap}^j)}{ \omega_{ap}^j \exp(s_{ap}^j)\!+\! \omega_{an}^j \exp(s_{an}^j)},
\end{equation}
\normalsize
where $s_{an}^j$ denotes the similarity for the $j^{th}$ positive sample pair. Mathematically, the lower credibility (higher uncertainty) of a sample pair, the smaller of a weight on the similarity and thus a smaller gradient in optimization, \textit{i}.\textit{e}., contributing smaller to the optimization.
For \emph{contrastive loss}, given a query/anchor sample $\boldsymbol{x}_k$, we have $N_k^+$ positive samples and $N_k^-$ negative samples in the memory bank. For a batch of $n_t$ samples, by introducing the sample pair credibility weights, the uncertainty-guided contrastive loss is as
\small
\begin{equation}
\mathcal{L}_{\text {UCT}}\!=\!\frac{1}{n_t} {\sum\limits_{k=1}^{n_t}}\log\!\!\left[\!\! 1+{\sum\limits_{j=1}^{N_k^-} \omega_{kj}^- {\rm{exp}}(s_{kj}^-) \sum\limits_{i=1}^{N_k^+} \omega_{ki}^+ {\rm{exp}}(-s_{ki}^+) }\right],
\label{eq:UCT}
\end{equation}
\normalsize
where $s_{kj}^-$ denotes the similarity between the query sample $\boldsymbol{x}_k$ and the $j^{th}$ negative sample, and $\omega_{kj}^-$ denotes the approximated reliability of the sample pair (see (\ref{eq:omega})). The lower credibility of a sample pair, the smaller the gradient and the contribution of this pair to the optimization. \textcolor{black}{Note that similar to our strong baseline, we also use self-paced weighting ~\cite{sun2020circle} to softly emphasize harder sample pairs (whose similarity score deviates far from the optimum) by giving larger weights to them to get effective update. For simplicity, we do not present it in (\ref{eq:UCT}), where $s_{kj}^-$, $s_{ki}^+$ denote the similarities that already re-weighted. }
The total loss for the target domain data in the fine-tuning stage could be formulated as:
\begin{equation}
\mathcal L_{target} = \mathcal L_{UID} + \lambda_{tri} \mathcal L_{UTRI} + \lambda_{ct} \mathcal L_{UCT} + \lambda_{reg} \mathcal L_{reg},
\end{equation}
where $L_{reg} = \frac{1}{n_t} {\sum_{i=1}^{n_t}} u_i$ is a regularization loss which prevents large uncertainty (\textit{i}.\textit{e}., small credibility which could reduce the first three losses) all the time. $\lambda_{tri}$, $\lambda_{ct}$, and $\lambda_{reg}$ are weighting factors.
\begin{comment}
==================================================
\begin{equation}
\mathcal{L}_{u i d} =-\frac{1}{m_{t}} \sum_{i=1}^{m_{t}} \mathcal U_{i} \log p\left(\tilde{y}^t_{i} | \boldsymbol{x}^t_{i}\right),
\label{eq:9}
\end{equation}
For the pair similarity optimization losses, triplet loss
When we obtain a pair uncertainty about a similarity score, it should get a weighting factor so as to get effective update with different gradient. Large uncertainty should bring the small gradient and vice versa. To this end, we define the triplet loss with uncertainty re-weighting:
\begin{equation}
\mathcal{L}_{utri} =\frac{1}{m_{t}} \sum\limits_{i=1}^{m_{t}} \max \left(0,s^n_i (\mathcal U_{n}+\mathcal U_{i})/2 - s^p_i (\mathcal U_{p}+\mathcal U_{i})/2 \right),
\end{equation}
where $U_{i}$ indicates the uncertainty of anchor, $U_{n}$and $U_{p}$ indicate the uncertainty of negative and positive samples, respectively.
In a similar spirit, weighted contrastive loss and identification loss also can be rectified by the uncertainty:
\begin{equation}
\mathcal{L}_{\text {ucl}}=\frac{1}{m_t} \sum_{i=1}^{m_t}\log\left[ 1+{\sum\limits_{j=1}^{L} S_{i,j}^p (\mathcal U_{i}+\mathcal U_{j})\sum\limits_{k=1}^{K} S_{i,k}^n (\mathcal U_{i}+\mathcal U_{k})} \right],\\
\end{equation}
\begin{equation}
\mathcal{L}_{u i d} =-\frac{1}{m_{t}} \sum_{i=1}^{m_{t}} \mathcal U_{i} \log p\left(\tilde{y}^t_{i} | \boldsymbol{x}^t_{i}\right),
\label{eq:9}
\end{equation}
Since most pseudo labels are correct, the model still could learn from the noisy labels. In Section 4.4, we show the uncertainty with pseudo labels further boosts the performance on the target domain despite the noise in pseudo labels.
We could obtain two distributions by the student model and the mean-teacher model, respectively.
To rectify the learning from noisy labels, we calculate the KL divergence between predicted likelihood of classification of two models as the uncertainty of samples:
\begin{equation}
\mathcal L_{KL}(\boldsymbol p_i,\boldsymbol{\tilde{p}}_i)= \sum_{n=1}^{N_t+N_s} p_{i,n} \log{p_{i,n}/\tilde{p}}_{i,n}\\
\end{equation}
It is worthy to note that the larger the KL divergence, the less the sample uncertainty. Therefore, we can use the $1/L_{kl}$ to re-weight the ReID loss.
\begin{equation}
[\boldsymbol V, \boldsymbol W]=[\boldsymbol V_1, \boldsymbol V_2, ..., \boldsymbol V_{N_t} , \boldsymbol W_1, \boldsymbol W_2, ..., \boldsymbol W_{N_s}],
\end{equation}
===========
Particularly, we calculate the similarity distribution of features between samples and the target cluster centers. Then we adopt the Kullback-Leibler divergence between the two similarity distribution as the uncertainty of the sample. The consistency levels for the samples with wrong pseudo labels are much lower than those with correct pseudo labels.
Let $\boldsymbol{V} \in \mathbb{R}^{N_t \times D}$ denotes a target prototype which stores all target clustering center and $\boldsymbol W \in \mathbb{R}^{N_s \times D}$ denotes a source prototype which stores all source center. These prototype weight vectors indicates the fully-connected weights or the identification center of features. $D$ is the feature dimension in the last linear layer:
\begin{equation}
[\boldsymbol V, \boldsymbol W]=[\boldsymbol V_1, \boldsymbol V_2, ..., \boldsymbol V_{N_t} , \boldsymbol W_1, \boldsymbol W_2, ..., \boldsymbol W_{N_s}],
\end{equation}
As shown in Fig.~\ref{fig:fm2}, given a feature $\boldsymbol f_i \in \mathbb{R}^{1 \times D}$ from student model and a feature $\boldsymbol{\tilde{f}}_i \in \mathbb{R}^{1 \times D}$ from mean teacher model, we calculate the similarity to target clustering center and source center for each mini-batch of target features:
\begin{equation}
\boldsymbol p_i = [\boldsymbol V, \boldsymbol W] \cdot {\boldsymbol f_i}^{T} \in \mathbb{R}^{(N_t+N_s) \times 1},
\end{equation}
where $\boldsymbol p_i$ indicates the similarity distribution from the student model, the $\boldsymbol{\tilde{p}}_i$ is from the mean-teacher model. We could obtain two distributions by the student model and the mean-teacher model, respectively.
To rectify the learning from noisy labels, we calculate the KL divergence between predicted likelihood of classification of two models as the uncertainty of samples:
\begin{equation}
\mathcal L_{KL}(\boldsymbol p_i,\boldsymbol{\tilde{p}}_i)= \sum_{n=1}^{N_t+N_s} p_{i,n} \log{p_{i,n}/\tilde{p}}_{i,n}\\
\end{equation}
It is worthy to note that the larger the KL divergence, the less the sample uncertainty. Therefore, we can use the $1/L_{kl}$ to re-weight the ReID loss.
To stabilize the training, we adopt the policy in~\cite{kendall2017uncertainties} that replace $1/L_{kl}$ as $\exp^{-L_{kl}}$:
\begin{equation}
\mathcal U_i= \exp(-L_{KL}(\boldsymbol p_i,\boldsymbol{\tilde{p}}_i))
\label{eq:uncer}
\end{equation}
where $\mathcal U_i$ indicates the credibility of the $i$-th samples.
We have observed that a sample with a wrong pseudo-label through clustering, in general, has a weaker consistency between the output of the mean teacher model and the student model. Based on this observation, we propose to exploit the uncertainty (measured by consistency levels) to evaluate the reliability of the pseudo-label of a sample and incorporate the uncertainty to re-weight its contributions in ReID losses. We incorporate the uncertainty-guided optimization in the classification loss, triplet loss, and contrastive loss for ReID.
When we obtain a pair uncertainty about a similarity score, it should get a weighting factor so as to get effective update with different gradient. Large uncertainty should bring the small gradient and vice versa. To this end, we define the triplet loss with uncertainty re-weighting:
\begin{equation}
\mathcal{L}_{utri} =\frac{1}{m_{t}} \sum\limits_{i=1}^{m_{t}} \max \left(0,s^n_i (\mathcal U_{n}+\mathcal U_{i})/2 - s^p_i (\mathcal U_{p}+\mathcal U_{i})/2 \right),
\end{equation}
where $U_{i}$ indicates the uncertainty of anchor, $U_{n}$and $U_{p}$ indicate the uncertainty of negative and positive samples, respectively.
In a similar spirit, weighted contrastive loss and identification loss also can be rectified by the uncertainty:
\begin{equation}
\mathcal{L}_{\text {ucl}}=\frac{1}{m_t} \sum_{i=1}^{m_t}\log\left[ 1+{\sum\limits_{j=1}^{L} S_{i,j}^p (\mathcal U_{i}+\mathcal U_{j})\sum\limits_{k=1}^{K} S_{i,k}^n (\mathcal U_{i}+\mathcal U_{k})} \right],\\
\end{equation}
\begin{equation}
\mathcal{L}_{u i d} =-\frac{1}{m_{t}} \sum_{i=1}^{m_{t}} \mathcal U_{i} \log p\left(\tilde{y}^t_{i} | \boldsymbol{x}^t_{i}\right),
\label{eq:9}
\end{equation}
Since most pseudo labels are correct, the model still could learn from the noisy labels. In Section 4.4, we show the uncertainty with pseudo labels further boosts the performance on the target domain despite the noise in pseudo labels.
\subsection{Optimization}
We integrate the above-mentioned losses. The total loss of the training could be formulated as:
\begin{equation}
\mathcal L_{target} = \mathcal L_{utri} + \mathcal L_{uid} + \mathcal L_{ucl} + \lambda_{kl} \mathcal L_{kl},
\end{equation}
where $lambda_{kl}$ is the weight for the consistency regularization. We do not intend to minimize the ReID loss under all conditions. If the uncertainty has received one large value, we will not punish the ReID loss. Meanwhile, to prevent that the model predicts the credibility all the time, as a trade-off, we introduce the regularization term via adding $L_{kl}$.
......
\paragraph{Overview}
Deep-cluster framework is the general pipeline due to the conciseness and effectiveness of this framework. To be specific, a ReID model $F(\cdot | \boldsymbol{\theta})$ is generally pre-trained via a supervised task on source domain data, where $\theta$ denotes the parameters of network. Then we adopt this pre-trained network to extract the features of target domain images which denote as $\left\{\boldsymbol{f_i^t}=F\left(\boldsymbol{x}_{i}^{t} | \boldsymbol{\theta}\right)\right\}|_{i=1} ^{N_{t}} \in \mathbb{R}^{D}$.
After that, the target domain images are grouped into $K$ classes by clustering these embedding features. Let $\{\tilde{\boldsymbol{y}}_{i}^{t}\}|^{N_t}_i\in\{1, \ldots, K\}$ denotes the pseudo labels generated for target domain images $\{x_{i}^{t}\}|_{i}^{N_t}$, which contain the noisy label.
Then, to calculate a cross-entropy loss, a classification head $FC_k: \mathbb{R}^{D} \rightarrow \mathbb{R}^{K}$ is adopted to convert the embedding feature to a score vector, in the form of $p\left(\tilde{y}^{t}_{i} | \boldsymbol{x}^{t}_{i}\right)=\operatorname{softmax}\left(FC( F(\boldsymbol{x}_{i}^{t} | \boldsymbol{\theta}))\right)$.
The network parameters $\theta$ and a learnable target-domain classifier $FC_k$ are then optimized with respect to a cross-entropy loss $\mathcal{L}_{i d}^{t}(\boldsymbol{\theta})$ and a triplet loss $\mathcal{L}_{t r i}^{t}(\boldsymbol{\theta})$ in the form of,
\begin{equation}
\begin{aligned}
\mathcal{L}_{i d}^{t}(\boldsymbol{\theta}) =-\frac{1}{m_{t}} \sum_{i=1}^{m_{t}} \log p\left(\tilde{y}^t_{i} | \boldsymbol{x}^t_{i}\right)
\\ \mathcal{L}_{t r i}^{t}(\boldsymbol{\theta}) =\frac{1}{m_{t}} \sum_{i=1}^{m_{t}} \max \left(0,s^n_i-s^p_i+m\right),
\end{aligned}
\label{eq:1}
\end{equation}
where $m_t$ is the mini-batch size of target data, $s^p_i=\boldsymbol{f}_{i}^{t} \boldsymbol{f}_{i, p}^{t}/ \| {\boldsymbol{f}_{i}^{t}}^T \| \cdot \| \boldsymbol{f}_{i, p}^{t}\|$ denotes the hardest positive cosine similarity, $s^n_i=\boldsymbol{f}_{i}^{t} \boldsymbol{f}_{i, n}^{t}/ \| {\boldsymbol{f}_{i}^{t}}^T \| \cdot \| \boldsymbol{f}_{i, n}^{t}\|$ denotes the hardest negative cosine similarity, scripts $i,p$ and $j,n$ indicate the indices of hardest positive and negative feature in each mini-batch for the sample $x^t_i$, and $m = 0$ denotes the triplet distance margin.
In the stage of clustering, some noisy labels may be generated. This can be partly alleviated by using the mean teacher model~\cite{tarvainen2017mean} at inference time, and consequently a mean teacher can yield more accurate targets. Therefore, we use the mean teacher model to enhance our baseline, which maintains a exponential moving average (EMA) parameters from student model. The parameters of the temporally average model at current
iteration $T$ are denoted $\boldsymbol\theta_T$ respectively, which can be calculated as:
\begin{equation}
\boldsymbol\theta^T_{t} = \alpha \boldsymbol\theta^{T-1}_{t} + (1-\alpha) \boldsymbol\theta^{T}_{s},
\end{equation}
where $\boldsymbol\theta^T_{t}$ indicates the temporal average parameters in the $T$-th iteration and $\boldsymbol\theta^{T}_{s}$ indicates the parameters of student model. The initial temporal average parameter of teacher model is same as the student model.
Existing methods for person ReID mostly use the triplet loss and classification loss to train the model. However, the hard-mining strategy of these works is intrinsically limited by mini-batch training. It means that only the local negative and positive pairs are accessible. Mining informative negative and positive instances are of central importance to this task. Thus, we adopt a memory bank to memorize the embeddings of past iterations, which allows the model to collect sufficient negative and positive pairs from multiple mini-batches to over the whole dataset.
Such two steps, pseudo label generation by clustering and feature learning with pseudo labels, are alternated until the training converges. However, because of domain gap, the pseudo-labels are not always reliable and there are noisy labels. This would mislead the feature representation learning towards learning noisy information. So, we need to exploit the credibility of the assigned pseudo-label of each sample to alleviate the influence of noisy labels, by suppressing the contribution of noisy samples and encouraging a higher contribution of good samples in the optimization.
As illustrated in Fig.~\ref{fig:fm}, the framework is alternated to gradually train the ReID model on target domain with the guidance of pseudo labels.
\subsection{Target Instance Memory Bank}
We first describe our target instance memory bank module, with updating mechanism. Then we show that a weighting scheme focusing on informative pairs is beneficial for contrastive loss with multiple positive samples.
\paragraph{Updating mechanism.} We use the memory bank~\cite{he2019momentum} to memory the features of whole dataset. When the mini-batch of samples are coming, we maintain the memory bank as a queue of data samples. This allows us to reuse the embeddings of features from the immediate preceding mini-batches. The introduction of a queue decouples the memory bank size from the mini-batch size. Our memory bank size can be much larger than a typical mini-batch size, and can be flexibly and independently set as a hyper-parameter. The samples in the dictionary are progressively replaced. The current mini-batch is enqueued to the dictionary, and the oldest mini-batch in the queue is removed. The dictionary always represents a sampled subset of all data, while the extra computation of maintaining this dictionary is manageable. Moreover, removing the oldest mini-batch can be beneficial, because its encoded keys are the most outdated and thus the least consistent with the newest ones.
\paragraph{Weighted contrastive loss} Then we use the cosine distance to calculate the similarity $s$ of features between mini-batch and memory bank. Moreover, according to the pseudo labels, we split the samples to the $L$ negative pairs and $K$ positive pairs, where weighted contrastive loss~\cite{sun2020circle} is suitable for this issue. Thus, weighted contrastive loss is proposed to better calculate the loss in terms of multiple positive samples, which is beneficial for the optimization of model:
\begin{equation}
\begin{aligned}
\mathcal{L}_{\text {cl}}=-\frac{1}{m_t} \sum_{i=1}^{m_t}\log[ \frac{\sum_{j=1}^{L} S_{i,j}^p}{\sum_{j=1}^{L} S_{i,j}^p+\sum_{k=1}^{K} S_{i,k}^n} ],
\end{aligned}
\label{eq:9}
\end{equation}
where $S_{i,j}^n= \exp \left( \alpha^{n}_{j}\left(s^{n}_{j}-m\right)\right)$ refers to the multiple positive paired similarity, $S_{i,k}^p=
\exp \left( \alpha^{p}_{k}\left(s^{p}_{k}-1+m\right)\right)$ indicates the multiple negative paired similarity, $\alpha^{n}_{j}=[m+s^n_j]_{+}$ and $\alpha^{p}_{k}=[1+m-s^p_k]_{+}$ are non-negative weighting factors respectively, $[\cdot]_{+}$ is the ``cut-off at zero'' operation, $m$ refers to a margin in order to better separate similarity, and $L$ and $K$ denote the number of negative pairs and positive pairs in the memory bank for the sample $x^t_i$ of mini-batch.
==========================================
\end{comment}
\section{Experiments}
\begin{table*}[t!]
\footnotesize
\centering
\caption{Performance (\%) comparison with the state-of-the-art methods for UDA person ReID on the datasets of DukeMTMC-reID, Market-1501 and MSMT17. We mark the results of the second best by \underline{underline} and the best results by \textbf{bold} text.
}
\vspace{-3mm}
\begin{center}
\begin{tabular}{|P{6.0cm}|C{1.0cm}C{1.0cm}C{1.0cm}C{1.0cm}|C{1.0cm}C{1.0cm}C{1.0cm}C{1.0cm}|}
\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{Methods}} & \multicolumn{4}{c|}{DukeMTMC$\to$Market1501} & \multicolumn{4}{c|}{Market1501$\to$DukeMTMC} \\
\cline{2-9}
\multicolumn{1}{|c|}{} & mAP & R1 & R5 & R10 & mAP & R1 & R5 & R10 \\
\hline
ATNet~\cite{Liu2019cvpr}(CVPR'19) & 25.6& 55.7& 73.2& 79.4& 24.9 &45.1 &59.5& 64.2\\
SPGAN+LMP~\cite{DengWeijian2018cvpr}(CVPR'18) &26.7 &57.7 &75.8 &82.4& 26.2&46.4 &62.3 &68.0 \\
CFSM~\cite{chang2018disjoint} (AAAI'19) & 28.3 & 61.2 & - & - & 27.3 & 49.8 &- & - \\
BUC~\cite{lin2019aBottom} (AAAI'19) & 38.3 & 66.2 & 79.6 & 84.5 & 27.5 & 47.4 & 62.6 & 68.4 \\
ECN~\cite{zhong2019invariance} (CVPR'19) & 43.0 & 75.1 & 87.6 & 91.6 & 40.4 & 63.3 & 75.8 & 80.4 \\
UCDA~\cite{qi2019novel} (ICCV'19) & 30.9 & 60.4 & - & - & 31.0 & 47.7 & - & - \\
PDA-Net~\cite{li2019cross} (ICCV'19) & 47.6 & 75.2 & 86.3 & 90.2 & 45.1 & 63.2 & 77.0 & 82.5 \\
PCB-PAST~\cite{zhang2019self} (ICCV'19) & 54.6 & 78.4 & - & - & 54.3 & 72.4 & - & - \\
SSG~\cite{yang2019selfsimilarity} (ICCV'19) & 58.3 & 80.0 & 90.0 & 92.4 & 53.4 & 73.0 & 80.6 & 83.2 \\
ACT~\cite{Yang2019Asymmetric} (AAAI'20) & 60.6 & 80.5 & - & - & 54.5 & 72.4 & - & - \\
MPLP~\cite{WANG2020cvpr1} (CVPR'20) & 60.4 &84.4 &92.8& 95.0 & 51.4&72.4 &82.9& 85.0 \\
DAAM~\cite{Huang2020aaai} (AAAI'20)& {67.8} & {86.4} & {-} & {-} & {63.9} & {77.6} & {-} & {-} \\
AD-Cluster~\cite{zhai2020adcluster} (CVPR'20)& {68.3} & {86.7} & {94.4} & {96.5} & {54.1} & {72.6} & {82.5} & {85.5} \\
MMT~\cite{ge2020mutual} (ICLR'20) & {71.2} & {87.7} & {94.9} & {96.9} & {65.1} & {78.0} & \underline{88.8} & \underline{92.5} \\
NRMT~\cite{zhao2020unsupervised}(ECCV'20) & 71.7& 87.8& 94.6& 96.5& 62.2& 77.8& 86.9& 89.5 \\
B-SNR+GDS-H~\cite{jin2020global}(ECCV'20) & 72.5 & 89.3 & - & - & 59.7 & 76.7 &- &-\\
MEB-Net~\cite{zhai2020multiple}(ECCV'20) & \underline{76.0} &\underline{89.9}& \underline{96.0}& \underline{97.5}& \underline{66.1}& \underline{79.6} & 88.3& 92.2 \\
\hline
UNRN (Ours) & \textbf{78.1} & \textbf{91.9} & \textbf{96.1} & \textbf{97.8} & \textbf{69.1} & \textbf{82.0} & \textbf{90.7} & \textbf{93.5} \\
\hline
\end{tabular}\\
\begin{tabular}{|P{6.0cm}|C{1.0cm}C{1.0cm}C{1.0cm}C{1.0cm}|C{1.0cm}C{1.0cm}C{1.0cm}C{1.0cm}|}
\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{Methods}} & \multicolumn{4}{c|}{Marke1501$\to$MSMT17} & \multicolumn{4}{c|}{DukeMTMC$\to$MSMT17} \\
\cline{2-9}
\multicolumn{1}{|c|}{} & mAP & R1 & R5 & R10 & mAP & R1 & R5 & R10 \\
\hline
ECN~\cite{zhong2019invariance} (CVPR'19) & 8.5 & 25.3 & 36.3 & 42.1 & 10.2 & 30.2 & 41.5 & 46.8 \\
SSG~\cite{yang2019selfsimilarity} (ICCV'19) & 13.2 & 31.6 &- & 49.6 & 13.3 & 32.2 & - & 51.2 \\
DAAM~\cite{Huang2020aaai} (AAAI'20)& {20.8} & { 44.5} & {-} & {-} & { 21.6} & { 46.7} & {-} & {-} \\
NRMT~\cite{zhao2020unsupervised}(ECCV'20) & 19.8& 43.7& 56.5 &62.2& 20.6 &45.2& 57.8& 63.3 \\
MMT~\cite{ge2020mutual} (ICLR'20) & \underline{22.9} & \underline{49.2} & \underline{63.1} & \underline{68.8} & \underline{23.3} & \underline{50.1} & \underline{63.9} & \underline{69.8} \\
\hline
UNRN (Ours) & \textbf{25.3} & \textbf{52.4} & \textbf{64.7} & \textbf{69.7} & \textbf{26.2} & \textbf{54.9} & \textbf{67.3} & \textbf{70.6} \\
\hline
\hline
\end{tabular}
\end{center}
\label{tab:sota}
\end{table*}
\subsection{Datasets and Evaluation Metrics}
We evaluate our methods using three person ReID datasets, including DukeMTMC-reID (Duke) ~\cite{dukemtmc}, Market-1501 (Market) ~\cite{market} and MSMT17~\cite{wei2018person}.
DukeMTMC-reID~\cite{dukemtmc} has 36,411 images, where 702 identities are used for training and 702 identities for testing.
Market-1501~\cite{market} contains 12,936 images of 751 identities for training and 19,281 images of 750 identities for testing.
MSMT17~\cite{wei2018person} contains 126,441 images of 4,101 identities, where 1,041 identities and 3060 identities are used for training and testing respectively.
We adopt mean average precision (mAP) and CMC Rank-1/5/10 (R1/R5/R10) accuracy for evaluation.
\subsection{Implementation Details}
\label{sec:imp}
We use ResNet50 pretrained on ImageNet as our backbone networks. As ~\cite{luo2019bag}, we perform data agumentation of randomly erasing, cropping, and flipping. For source pre-training, each mini-batch contains 64 images of 4 identities. For our fine-tuning stage, when the source data is also used, each mini-batch contains 64 source-domain images of 4 identities and 64 target-domain images of 4 pseudo identities, where there are 16 images for each identity. All images are resized to 256$\times$128. Similar to \cite{ge2020mutual,yang2019selfsimilarity}, we use the clustering algorithm of DBSCAN. For DBSCAN, the maximum distance between neighbors is set to $eps=0.6$ and the minimal number of neighbors for a dense point is set to 4. ADAM optimizer is adopted. The initial learning rate is set to 0.00035. We set the weighting factors $\lambda_{tri}=1$, $\lambda_{ct}=0.05$, and $\lambda_{reg}=1$, where we determine them simply by making the several losses on the same order of magnitude.
\subsection{Comparison with the State-of-the-arts}
We compare our proposed UNRN with the state-of-the-art methods on four domain adaptation settings in Tab.~\ref{tab:sota}. Our UNRN significantly outperforms the second best UDA methods by $2.1\%$, $3.0\%$, $2.4\%$ and $2.9\%$, in mAP accuracy, for Duke$\to$Market, Market$\to$Duke, Market$\to$MSMT, and Duke$\to$MSMT, respectively.
SSG~\cite{yang2019selfsimilarity} performs multiple clustering on both global body and local body parts. DAAM~\cite{Huang2020aaai} introduces an attention module and incorporates domain alignment constraints. MMT~\cite{ge2020mutual} use two networks (four models) and MEB-Net~\cite{Zhai2020} use three networks (six models) to perform mutual mean teacher training, which have high computation complexity in training. Our UNRN uses only one network (two models) in training but still significantly outperforms the best-performing MEB-Net ~\cite{Zhai2020}.
%
\subsection{Ablation Studies}
\begin{table}[h!]
\caption{Ablation studies on the effectiveness of components in our proposed UNRN on Market and Duke. \textbf{Source}: source data is also used in fine-tuning stage with ID classification loss. \textbf{Contrastive loss (CT)}: pair similarity optimization across the memory bank. \textbf{Mean teacher}: use mean teacher method where a temporally moving averaged model is taken as the mean teacher. \textbf{ID}: target domain identity classification loss. \textbf{UID}: ID loss with \emph{uncertainty}. \textbf{TRI}: target domain triplet loss. \textbf{UTRI}: target domain triplet loss with \emph{uncertainty}. \textbf{UCT}: contrastive loss with \emph{uncertainty}.}
\centering
\footnotesize
\vspace{-5pt}
\label{tab:ablation1}
\begin{center}
\begin{tabular}{P{3.4cm}|C{0.7cm}C{0.7cm}|C{0.7cm}C{0.85cm}}
\hline
\multicolumn{1}{c|}{\multirow{2}{*}{Methods}} & \multicolumn{2}{c|}{Duke$\to$Market}&\multicolumn{2}{c}{Market$\to$Duke} \\
\cline{2-5}
\multicolumn{1}{c|}{}& mAP & R1 & mAP & R1 \\
\hline \hline
Supervised learning &85.7&94.1&75.8&86.2\\
\hline
\hline
Model pretraining &32.9&62.6&35.2&53.3\\
\hline
Baseline &68.2&87.9&60.4&75.9\\
+Source &70.4&88.3&61.3&76.4\\
+Contrastive loss &72.1&88.7&62.1&77.6\\
+Mean teacher (SBase.) &75.4&89.8&64.8&79.7\\
\hline
SBase.~(ID+TRI+CT) &75.4&89.8&64.8&79.7\\
SBase.~w/~UID &77.3&91.2&68.2&81.3\\
SBase.~w/~UTRI &76.1&90.9&66.3&81.0\\
SBase.~w/~UID+UTRI &77.8&91.5&68.9&81.7\\
SBase.~w/~UID+UTRI+UCT &78.1&91.9&69.1&82.0\\
\hline
\end{tabular}
\end{center}
\vspace{-15pt}
\end{table}
\noindent\textbf{Effectiveness of components in our strong baseline.} We build our basic baseline \emph{Baseline} following the commonly used baselines in UDA methods~\cite{ge2020mutual,song2020unsupervised,jin2020global}, where in the fine-tuning stage, the identity classification loss and triplet loss are used to fine-tune the network based on the pseudo labels for the target domain data. On top of \emph{Baseline}, we add three components (as described in Section \ref{subsec:baseline}. Tab.~\ref{tab:ablation1} shows that each component brings additional significant gain and finally we have a strong baseline \emph{SBase.}.
\noindent\textbf{Effectiveness of our uncertainty-guided optimization.} We validate the effectiveness of our proposed design on top of a strong baseline \emph{SBase.}. In general, the stronger of a baseline, the harder one can achieve gains since the cases easy to address are mostly handled by the strong baseline. Once the new design is complementary to the strong baseline, it is valuable to advance the development of techniques.
In the strong baseline \emph{SBase.}, for the target domain samples, identity classification loss (ID), triplet loss (TRI), and contrastive loss (CT) are used for supervision based on pseudo labels. To alleviate the negative influence of noisy/wrong pseudo labels, we exploit uncertainty to re-weight the contributions of samples.
Tab.~\ref{tab:ablation1} shows the comparisons. When we replace ID loss by UID loss, which is the ID loss with uncertainty, the mAP accuracy is significantly improved by 1.9\% and 3.4\% for Duke$\to$Market and Market$\to$Duke, respectively. When we replace triplet (TRI) loss by UTRI loss, similar improvements are observed. We have our final scheme \emph{UNRN} when the uncertainty-guided optimization is applied to all the three losses and it achieves \textbf{2.7\%} and \textbf{4.3\%} improvement in mAP accuracy over \emph{SBase.} for Duke$\to$Market and Market$\to$Duke, respectively.
\subsection{Design Choices}
\noindent\textbf{Influence of different designs in uncertainty estimation.} We have discussed the estimation of uncertainty in Section \ref{subsec:uncertainty}. There are some design choices and Tab.~\ref{tab:ablation2} shows the comparisons. \textbf{1)} \emph{UNRN-feat. consist.} denotes that we estimate the uncertainty based on the distance of the features $\boldsymbol{\tilde{f}_i}$ and $\boldsymbol{{f}_i}$ instead of the distance of derived soft multilabels. We can see that the gain over \emph{SBase.} is negligible due to the poor estimation of uncertainty. In contrast, using the consistency between soft multilabels (\textit{i}.\textit{e}. \emph{UNRN-$R$ (Ours)}) captures global information of the target cluster centers and brings significant improvement. \textbf{2)} When we estimate the uncertainty, we leverage a set of ``reference persons'' to get the soft multilabels. \emph{UNRN-$R_t$} denotes the scheme when only the target cluster centers (see Fig.~\ref{fig:fm2}) are used as ``reference persons''. \emph{UNRN-$R_s$} denotes the scheme when only the source centers are used as ``reference persons''. \emph{UNRN-$R$} denotes both are used as ``reference persons''. We can see that the performance of \emph{UNRN-$R_s$} is very similar to that of \emph{SBase.}, where the source centers only cannot provide the clustering information of target domain data and is helpless to estimate the reliability of pseudo labels. \emph{UNRN-$R_t$} outperforms \emph{SBase.} significantly, which captures the target domain global clustering information that is helpful to estimate the reliability of pseudo labels. Interestingly, \emph{UNRN-$R$} which jointly considers the target cluster centers and source centers provides the best performance. That may
be because the source centers provide more references which enables the soft multilabels more informative.
\noindent\textbf{Influence of regularization loss $\mathcal{L}_{reg}$.} The regularization loss $\mathcal{L}_{reg}$ prevents larger uncertainty all the time. As shown in Tab.~\ref{tab:ablation2}, our final scheme \emph{UNRN-$R$} outperforms \emph{UNRN w/o $\mathcal{L}_{reg}$} by 0.6\% and 1.1\% in mAP accuracy for Duke$\to$Market and Market$\to$Duke, respectively.
\begin{table}[t
\caption{Influence of different designs in uncertainties estimation, and the influence of regularization loss $\mathcal{L}_{reg}$ on performance under our framework.}
\centering
\small
\vspace{-2mm}
\label{tab:ablation2}
\begin{center}
\begin{tabular}{P{3.6cm}|C{0.7cm}C{0.7cm}|C{0.7cm}C{0.85cm}}
\hline
\multicolumn{1}{c|}{\multirow{2}{*}{Methods}} & \multicolumn{2}{c}{Duke$\to$Market}&\multicolumn{2}{c}{Market$\to$Duke} \\
\cline{2-5}
\multicolumn{1}{c|}{}& mAP & R1 & mAP& R1 \\
\hline \hline
SBase. &75.4&89.8&64.8&79.7\\
\hline
UNRN w/o $\mathcal{L}_{reg}$ &77.5&91.4&68.0&81.4\\
\hline
UNRN-feat. consist. &76.5&91.0&66.7&80.9 \\
UNRN-$R_s$ &76.0&91.3&66.6&81.0 \\
UNRN-$R_t$ &77.8&91.7&68.3&81.7 \\
UNRN-$R$ (Ours) &78.1&91.9&69.1&82.0 \\
\hline
\end{tabular}
\end{center}
\vspace{-10pt}
\end{table}
\noindent\textbf{Influence of size of memory bank.} We use a queue to maintain \textcolor{black}{$N$ target domain instances in the} memory bank. Tab.~\ref{tab:memory-size} shows that as the queue length $N$ increases, the performance increases but saturates when the size is around 8192.
\begin{table}[t]
\caption{Influence of the number ($N$) of target instances in the memory bank. We study this on top of baseline scheme \emph{Baseline+Source+Contrastive loss} (see Tab.~\ref{tab:ablation1}). ``All'' denotes the size is equal to the size of target training dataset.}
\label{tab:memory-size}
\vspace{-8pt}
\begin{center}
\small
\begin{tabular}{c|c|c|c|c|c}
\hline
\multicolumn{1}{c|}{\multirow{2}{*}{Size of memory bank}} & \multicolumn{5}{c}{Market$\to$Duke}\\ \cline{2-6}
\multicolumn{1}{c|}{}& 0 & 1024 & 4096 & 8192 & All \\
\hline
\hline
mAP & 62.8 & 63.4 & 63.9 & 64.8 & 64.5 \\ \hline
R1 & 77.9 & 78.9 & 79.3 & 79.7 & 79.3 \\ \hline
\end{tabular}
\end{center}
\end{table}
\vspace{-8pt}
\section{Conclusion}
In this paper, for clustering-based UDA person ReID, we aim to alleviate the negative influence of wrong/noisy labels during the adaptation. We have observed that a sample with a wrong pseudo-label through clustering in general has a weaker consistency between the output of the mean teacher model and the student model. Based on this finding, we propose to exploit the uncertainty (measured by consistency levels) to evaluate the reliability of the pseudo-label of a sample and incorporate the uncertainty to re-weight its contribution within various ReID losses, including the ID classification loss per sample, the triplet loss, and the contrastive loss. Our uncertainty-guided optimization brings significant improvement over our strong baseline and our scheme achieves the state-of-the-art performance on benchmark datasets.
\section{Acknowledgments}
This work was in part supported by the National Key R$\&$D Program of China under Grand 2020AAA0105702, National Natural Science Foundation of China (NSFC) under Grants U19B2038 and 61620106009.
\section{Ethical Impact}
Our method is proposed to help match/identify different persons across images, which can facilitate the development of smart retail systems in the future. When the person re-ID system is used to identify the pedestrian, it may cause a violation of human privacy. Therefore, governments and officials need to carefully formulate strict regulations and laws to ensure the legal use of ReID technology and strictly protect the data.
\bibliographystyle{aaai}
| 1f3c2da1acfb35f0f9b2620665b3bfea1d4d5e6f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
One of the prominent fields in robotics is the development of artificial or prosthetic limbs. The most advanced devices are controlled via electromagnetic sensors that measure sEMG signals either from muscles \cite{fajardo2015} or the central nervous system \cite{Li2017}.
Artificial limbs are normally controlled via gesture recognition from the sEMG sensor data. This gesture recognition can be either discrete \cite{Du2017} or continuous \cite{Quivira2018}. In the discrete case, there is a limited number of available gestures that the limb can perform, and the goal is to map the sensor data to one of these gestures. Therefore, the recognition task is a classification problem. In our work, we focus instead on continuous gesture recognition, which derives the angles of the limb joints from the data and allows for more life-like limb movement.
In this paper, we perform experiments on the task of controlling a robotic hand via an arm band sensor. Previous work (e.g. \cite{fajardo2015, Du2017, Quivira2018}) assumed robotic hands with immobile wrists, which significantly restricts the functionality of the robotic limb. We present, for the first time, experiments with both immobile and mobile wrist data, and show that our proposed techniques outperform the state-of-the-art in both cases.
The most successful gesture recognition from the sEMG sensor data methods to date employ recurrent neural networks (RNNs) \cite{Quivira2018}. However, RNNs take a very long time to train to reach a desirable accuracy \cite{Lei2017}. In our work, we propose to use RNNs with simple recurrent units (SRU) \cite{Lei2017}, instead, for continuous gesture recognition. This method has the advantage of reaching a higher gesture recognition accuracy than regular RNNs in the same training time.
In addition, we show how domain adaptation techniques \cite{Ganin2014} can be used for continuous gesture recognition, and for the first time empirically demonstrate that these techniques improve the transferability between subjects, where a limb controller trained on data from one person is used for gesture recognition of another person.
\section{Background and Related Work}
\subsection{sEMG sensors}
Surface electromyography is the method of obtaining electric signals coming from neurons that are responsible from muscle contractions. As the name suggests, it is a non-invasive method and electrodes are attached to the skin surface. These signals can be then used in a control system for powered upper-limb prostheses because motions of wrist and fingers are mostly controlled by muscles in the forearm \cite{phinyomark2012}.
sEMG sensors range from non-portable units with high measurement accuracy to light and wearable units, albeit with a lower accuracy.
In order to retrieve sEMG data, we utilized the Myo armband \cite{Sathiyanarayanan2016}, as shown in Figure~\ref{img:myo}. This device has the advantages of being wearable and having an affordable price\footnote{At the time of writing the price is in the range of \$200}. The Myo armband consists of 8 sEMG sensors capable of recording sEMG data at frequency of 200 Hz. The raw signal is digitized with 8-bit analog-to-digital converter to yield a $[-128,+128]$ range. For wireless connection the Myo armband uses Bluetooth Low Energy technology.
\begin{figure}[thtb]
\centering
\includegraphics[width=0.45\textwidth]{images/band_on_my_arm}
\caption{Myo Armband}
\label{img:myo}
\end{figure}
\subsection{Motion capture system}
In order to tackle the task of continuous gesture recognition, a motion capture system to extract the limb angles is required for the generation of the data set. Existing motion capture systems are based on a wide range of technologies, including video recording \cite{kim2008}, ultrasound \cite{hettiarachchi2015}, and tracking gloves \cite{kim2009}.
In our work, the Leap Motion Controller \cite{leapmotion} was used, which provided a cheap and efficient way to track and record the limb angles of a user's hand. This device is specifically designed for hand movement detection. The controller itself consists of two monochromatic IR cameras and three IR LEDs. The overall accuracy of the controller was shown to be 0.7 millimeters. Motion capture data is transmitted with a frequency from 50 Hz up to 120 Hz.
\subsection{Recurrent Neural Networks}
Recurrent neural networks (RNNs) \cite{Jain:1999:RNN} are a special type of neural networks designed to work with (numerical) sequence data. The input of an RNN is the sequence and the output is a numerical vector. In our case, the input sequence are the sEMG measurements, and the output is a vector of limb angles. Every neuron of the network on each element of the input sequence updates its state and outputs a value as defined by:
\begin{equation}
h_t = \sigma_h(W_h x_t + U_h h_{t - 1} + b_h) \\
\end{equation}
\begin{equation}
y_t = \sigma_y(W_y h_t + b_y)
\end{equation}
where $x_t$ is the sequence element, $h_t$ is the hidden state, $y_t$ is the output vector, $W, U, b$ are cell parameters, and $\sigma_h$ and $\sigma_y$ are activation functions.
\subsection{Hand Gesture Recognition}
Initial work on controlling artificial limbs using EMG signals focused on creating a control scheme, i.e. mapping specific signals to specific actions of the limb (e.g. \cite{fajardo2015}). These approaches are not able to tackle a large variety of gestures and the resulting controllers are specific to an individual and can not be transferred to other persons. Gesture recognition methods have been introduced to overcome these limitations.
The gesture recognition task can be split into:
\begin{enumerate}
\item Classification with a number of discrete gestures or poses.
\item Regression with respect to recognizing continuous values such as joint angles.
\end{enumerate}
In most recent and advanced work, Du et al.\ \cite{Du2017} propose a hand gesture classification method employing domain adaption (specifically adaptive batch normalization, AdaBN). The authors generated several datasets with the number of gestures varying from 8 to 12, none of which involved wrist movement. The results showed that AdaBN can improve generalization and tranferability of a convolutional neural network with respect to accuracy on the discrete gesture recognition tasks.
Following up on this work, \cite{Quivira2018} focused on continuous gesture recognition (joint angles) using Gaussian processes. Again, unlike our work, this research did not involve any wrist movement.
\section{Method}
In this section, we present our approach for continuous gesture recognition. Specifically, we propose the use of RNNs with simple recurrent units (SRUs) to achieve a higher accuracy in the same time of training than regular RNNs using gated recurrent units (GRUs). Furthermore, we introduce a novel technique to apply domain adaptation to continuous gesture recognition.
\subsection{Simple Recurrent Units}
Due to vanishing gradient and exploding gradient problems vanilla RNN is not used nowadays \cite{chung2014}. The most common approach to date is to employ RNNs with GRU cells.
\subsubsection{GRU-cell}
The GRU-cell features special gates to reduce vanishing and exploding gradient problems. It is defined as:
\begin{equation} \label{eq:gru_z_t}
z_t = \sigma_g (W_z x_t + U_z h_{t - 1} + b_z)
\end{equation}
\begin{equation} \label{eq:gru_r_t}
r_t = \sigma_g (W_r x_t + U_r h_{t - 1} + b_r)
\end{equation}
\begin{equation}
h_t = (1 - z_t) \odot h_{t - 1} + z_t \odot \sigma_h (W_h x_t + U_h (r_t \odot h_{t - 1}) + b_h)
\end{equation}
where $x_t$ is the sequence element, $h_t$ is the hidden state and output, $z_t$ is the update gate, $r_t$ is the reset gate, and $W, U, b$ are cell parameters.
Figure~\ref{img:rec_scheme_num} illustrates an RNN architecture with GRU cells.
\begin{figure}[thtb]
\centering
\includegraphics[width=0.40\textwidth]{images/rec_scheme_num}
\caption{Recurrent neural network architecture \newline 1 - Input data, \newline 2 - The first GRU/SRU layer, \newline 3 - The second GRU/SRU layer, \newline 4 - Fully-connected layer of Predictor, \newline 5 - Fully-connected layer without activation-function - result vector.}
\label{img:rec_scheme_num}
\end{figure}
\subsubsection{SRU-cell}
The GRU-cell, due to its internal dependencies, is difficult to parallelize efficiently. In order to reduce computation cost for RNN training and inference, a simple recurrent unit was proposed in \cite{Lei2017} as follows:
\begin{equation}
\hat{x_t} = W x_t
\end{equation}
\begin{equation}
f_t = \sigma(W_f x_t + b_f)
\end{equation}
\begin{equation}
r_t = \sigma(W_r x_t + b_r)
\end{equation}
\begin{equation}
c_t = f_t \odot c_{t - 1} + (1 - f_t) \odot \hat{x_t}
\end{equation}
\begin{equation}
h_t = r_t \odot \sigma(c_t) + (1 - r_t) \odot x_t
\end{equation}
where $x_t$ is a sequence element, $h_t$ is the output vector, $c_t$ is the hidden state, $f_t$ is the forget gate, $r_t$ is the reset gate, and $W, U, b$ are cell parameters.
The RNN architecture with SRU cells is analogous to the RNN architecture shown in Figure~\ref{img:rec_scheme_num}.
\subsection{Domain adaptation}
When transferring a gesture recognition network trained on one person to another person, specialized transfer learning techniques are required. Adversarial domain adaptation (ADA) has been shown in recent work \cite{ganin2014unsupervised} to perform well for transfer learning tasks, and we have thus selected it for our research. ADA employs an additional neural network that predicts the current domain of input data from a feature vector. There is a gradient reversal layer between the feature vector and the discriminator network. This layer multiplies gradients by some value between -1 and 0. To train a neural network with ADA, we sum the loss function of the predictor and the loss function of the discriminator and perform back-propagation with a reversal of the discriminator gradient before the feature layer. The goal of this process is to inhibit domain-specific features and thus prevent over-fitting.
Figure~\ref{img:rec_scheme_adapt} shows the RNN architecture with domain adaptation.
We have also tried adaptive batch normalization, as used for discrete hand gesture recognition in \cite{Du2017}, but this did not show good performance on our continuous gesture recognition tasks.
\begin{figure}[thtb]
\centering
\includegraphics[width=0.40\textwidth]{images/rec_scheme_num_adapt}
\caption{Recurrent neural network with domain adaptation architecture \newline 1 - Input data, \newline 2 - The first GRU/SRU layer, \newline 3 - The second GRU/SRU layer, \newline 4 - Fully-connected layer of Predictor, \newline 5 - Fully-connected layer without activation-function - result vector, \newline 6 - Fully-connected layer of Discriminator, \newline 7 - Fully-connected layer without activation function - vector of classes (domains) probabilities.}
\label{img:rec_scheme_adapt}
\end{figure}
\section{Data acquisition}
sEMG data was obtained by means of the Myo armband device and finger and wrist angles were captured with the Leap Motion motion capture system.
We recorded data recorded from 5 healthy volunteers ranging in age from 20 to 28 years. Each participant was asked to sit on a chair and establish a relaxed position with the wrist slightly above the Leap Motion sensor to ensure accurate motion capture. The participant was then asked to perform a given sequence of hand movements such as bending or stretching the index finger.
Two datasets were collected: the first dataset only included movement of fingers with a static wrist (similar to \cite{Quivira2018}, and the second one was aimed to add wrist movements and some casual gestures. Data was collected in sessions with a duration of 4 minutes each. The specifics of the datasets are as follows.
\hfill
\noindent
Immobile wrist dataset:
\hfill
\begin{enumerate}
\item Movements of distinct fingers, 150 s
\item Simultaneous movement of all fingers, 60 s
\item Free finger movements, 30 s
\end{enumerate}
\hfill
\noindent
Mobile wrist dataset:
\hfill
\begin{enumerate}
\item Movements of distinct fingers, 60 s
\item Simultaneous movement of all fingers, 30 s
\item Pinch, 30 s
\item Open palm movements, 30 s
\item Thumb movements, 30 s
\item Free finger movements, 30 s
\end{enumerate}
\hfill
For each volunteer 8 sessions per dataset were recorded. For the mobile wrist dataset, the volunteers were asked to do the finger movements while keeping the wrist angle at one of 6 specified positions, including four angles on the up-down axis of the wrist, bending the wrist to the right, and keeping it straight.
Due to different data acquisition rates of the Myo armband and Leap Motion there is a need for data synchronization and alignment. Global alignment was achieved by matching the dataframes from the two devices, minimizing the difference between the time stamps of the individual dataframes, with a maximum difference of 10 ms.
After alignment, the sEMG and limb angle data was filtered with a low pass filter with a frequency of 10 Hz and 4Hz respectively to achieve noise reduction. The sEMG data was split into windows of size 128, which formed the input vector for the recurrent neural network.
\section{Experiment Setup}
In this section we provide the design parameters of the two RNN approaches used in our experiments. The first approach represents standard RNN technology, while the second approach incorporates optimizations selected by us for the task of continuous gesture recognition.
Adversarial domain adaptation was implemented with a gradient reversal layer after the feature generation block. The discriminator block consists of 2 fully connected layers with 256 neurons in the first layer and 15 or 18 neurons in the second (output) layer, since there are 15 angles to predict in the case of an immobile wrist and 18 in case of a mobile wrist.
The hyper-parameters of the recurrent neural network with GRU used in our experiments are:
\begin{itemize}
\item 2 GRU recurrent layers with 256 neurons each. Only the last output was fed to the next layer.
\item 2 fully connected layers with 256 and 15 (18) neurons respectively.
\item Optimizer: Adam \cite{kingma2014adam} with a learning rate of 0.001.
\item Trained for 30 epochs or until there was no validation gain during 8 epochs.
\end{itemize}
The hyper-parameters for the recurrent neural network with SRU are:
\begin{itemize}
\item 2 SRU recurrent layers with 256 neurons each.
\item Global Average Pooling layer.
\item 2 fully connected layers with 256 and 15 (18) neurons respectively.
\item Optimizer: Adam with learning rate equals to 0.001.
\item Net was trained for 30 epochs or until there was no validation gain during 8 epochs.
\end{itemize}
\section{Evaluation Metrics}
\makeatletter
\DeclareRobustCommand{\vardivision}{%
\mathbin{\mathpalette\@vardivision\relax
}
\newcommand{\@vardivision}[2]{%
\reflectbox{$\m@th\smallsetminus$}%
}
\makeatother
In our evaluation, we measure the root mean square error (RMSE) and normalized root mean square error (NRMSE). RMSE is used as a standard way to measure regression error. In addition, we used NRMSE to normalize the errors to account for different limb movement ranges. RMSE and NRMSE are given in equations~\ref{rmse} and~\ref{nrmse} respectively.
$y_i$ denote the real values and $\hat{y_i}$ the predicted values. $y_{range}$ is the range of real values. $\Theta_i = y_i \vardivision y_{range}$ and $\hat{\Theta_i} = \hat{y_i} \vardivision y_{range}$.
where $n$ is the number of angles predicted (15 for an immobile wrist and 18 for a mobile wrist).
\begin{equation}
\label{rmse}
RMSE = \sqrt[]{\frac{1}{n}\sum^{n}_{i=0}{(y_i - \hat{y_i})^2}}
\end{equation}
\begin{equation}
\label{nrmse}
NRMSE = \sqrt[]{\frac{1}{n}\sum^{n}_{i=0}{(\Theta_i - \hat{\Theta_i})^2}}
\end{equation}
\newcolumntype{C}[1]{>{\centering\let\newline\\\arraybackslash\hspace{0pt}}m{#1}}
\begin{table*}[tb]
\centering
\caption{Results for immobile wrist}
\label{all_table_static}
\begin{tabular}{|l|p{32mm}|c||C{20mm}|C{20mm}||C{20mm}|C{20mm}|}
\hline
\multirow{2}{10mm}{Metric} & \multirow{2}{12mm}{Model} & \multirow{2}{15mm}{Intra session} & \multicolumn{2}{p{16mm}||}{Inter session} & \multicolumn{2}{p{16mm}|}{Inter subjects} \\ \cline{4-7}
& & & no ADA & with ADA & no ADA & with ADA \\
\hline
\multirow{3}{*}{RMSE}
& Gaussian Process & 23.29$\pm$0.15 & 23.44$\pm$0.05 & - & 23.91$\pm$0.01 & - \\ \cline{2-7}
& GRU Recurrent neural net & 18.26$\pm$0.04 & 19.80$\pm$0.14 & 20.50$\pm$0.07 & 23.45$\pm$0.30 & \textbf{22.29$\pm$0.24} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{17.99$\pm$0.04} & \textbf{19.24$\pm$0.04} & 19.70$\pm$0.03 & 23.14$\pm$0.06 & \textbf{22.07$\pm$0.18} \\
\cline{2-7}
\hline \hline
\multirow{3}{*}{NRMSE}
& Gaussian Process & 0.2456$\pm$0.0016 & 0.1924$\pm$0.0000 & - & 0.2034$\pm$0.0001 & - \\ \cline{2-7}
& GRU Recurrent neural net & 0.1552$\pm$0.0004 & 0.1604$\pm$0.0008 & 0.1684$\pm$0.0005 & 0.1984$\pm$0.0023 & \textbf{0.1890$\pm$0.0021} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{0.1512$\pm$0.0004} & \textbf{0.1558$\pm$0.0004} & 0.1613$\pm$0.0004 & 0.1967$\pm$0.0002 & \textbf{0.1879$\pm$0.0019} \\
\cline{2-7}
\hline
\end{tabular}
\end{table*}
\begin{table*}[tb]
\centering
\caption{Results for mobile wrist}
\label{all_table_wrist}
\begin{tabular}{|l|p{32mm}|c||C{20mm}|C{20mm}||C{20mm}|C{20mm}|}
\hline
\multirow{2}{10mm}{Metric} & \multirow{2}{12mm}{Model} & \multirow{2}{15mm}{Intra session} & \multicolumn{2}{p{16mm}||}{Inter session} & \multicolumn{2}{p{16mm}|}{Inter subjects} \\ \cline{4-7}
& & & no ADA & with ADA & no ADA & with ADA \\
\hline
\multirow{3}{*}{RMSE}
& Gaussian Process & 22.33$\pm$0.24 & 22.23$\pm$0.00 & - & 22.39$\pm$0.01 & - \\ \cline{2-7}
& GRU Recurrent neural net & 18.60$\pm$0.05 & 19.19$\pm$0.05 & 19.93$\pm$0.06 & 21.74$\pm$0.20 & \textbf{20.94$\pm$0.17} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{18.26$\pm$0.11} & \textbf{18.83$\pm$0.04} & 19.55$\pm$0.03 & 21.50$\pm$0.11 & \textbf{20.89$\pm$0.09} \\ \cline{2-7}
\hline \hline
\multirow{3}{*}{NRMSE}
& Gaussian Process & 0.2360$\pm$0.0033 & 0.1770$\pm$0.0000 & - & 0.1824$\pm$0.0001 & - \\ \cline{2-7}
& GRU Recurrent neural net & 0.1558$\pm$0.0004 & 0.1518$\pm$0.0004 & 0.1580$\pm$0.0006 & 0.1757$\pm$0.0016 & \textbf{0.1698$\pm$0.0012} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{0.1516$\pm$0.0013} & \textbf{0.1490$\pm$0.0000} & 0.1550$\pm$0.0000 & 0.1742$\pm$0.0008 & \textbf{0.1695$\pm$0.0007} \\
\cline{2-7}
\hline
\end{tabular}
\end{table*}
\section{Evaluation} \label{sec:evaluation}
In order to show the superiority of the improved RNN approach (RNN with SRU) over other recent approaches from the literature, namely a standard RNN with GRU approach (similar to \cite{Du2017}) and Gaussian Process (as in \cite{Quivira2018}), we applied these three methods to our datasets with and without mobile wrists.
In addition, we applied adversarial domain adaption (ADA) to the RNN approaches to evaluate the transfer ability of each of these methods. Note, that ADA is not applicable to Gaussian Processes.
We split the data into training, validation and test sets as follows:
\begin{enumerate}
\item Each session was split into blocks of 12 seconds each.
\item A period of 3 seconds was randomly sampled from each block.
\item Half of these periods formed a validation set, while the other half formed the test set
\item The training set was created from the sessions after removing all periods that were used for the validation and test set.
\end{enumerate}
In other words, to get intra-session results all training, validation and test sets from all session were combined together. Half of this data was used for training, 25\% was used for validation and the remaining 25\% for testing.
In order to obtain inter-session results, we split the data set into five partitions, where the partitions do not share data from the same session. We used four of the partitions for training and validation, and the remaining partition for testing. This was repeated five times where each partition is used as a test set exactly once.
A similar approach was used to evaluate inter-subject performance, where each of the five partitions contained data from a single person.
\section{Results}
The results of the evaluation with immobile wrist and mobile wrist data are shown in tables~\ref{all_table_static} and~\ref{all_table_wrist} respectively. The evaluation results for the three machine learning algorithms are presented in terms of RMSE and NRMSE for the three types of experiments: intra-session, inter-session, and inter-subject (as described in Section~\ref{sec:evaluation}). For inter-session and inter-subject we also compare results with and without the application of adversarial domain adaptation (ADA).
First of all, recurrent neural networks with SRUs show better performance in most cases, and comparable performance in the case of inter-subject experiments. These results hold for both mobile and immobile wrists. The apparent gesture recognition accuracy improvement when comparing mobile wrist with immobile wrist results is due to an increased number of predicted angles in the mobile wrist case (18 vs. 15). Therefore, the two cases are not really comparable to each other.
As can be seen, adversarial domain adaptation (ADA) improves inter-subject accuracy, while actually showing worse results for inter-session experiments. This demonstrates that inter-session differences are not significant enough for the inhibition of session-specific features (and the resulting prevention of over-fitting) offered by ADA to have a positive impact on accuracy. However the case is different for inter-subject testing, where the difference in recorded data can be quite substantial.
\section{Conclusion and Outlook}
In this paper, we presented experiments with three machine learning approaches applied to continuous gesture recognition based on sEMG measurements. Our main contributions are:
\begin{enumerate}
\item We presented for the first time an evaluation of machine learning techniques with mobile wrists, which significantly enhance the functionality of artificial limbs.
\item We demonstrated that RNNs with SRUs are superior to other state-of-the-art approaches to gesture recognition proposed in the literature, specifically Gaussian Processes and RNNs with GRUs.
\item We presented for the first time inter-subject gesture recognition experiments and showed the advantages of adversarial domain adaption for this task.
\end{enumerate}
In future work, we intend to further investigate the case of inter-subject gesture recognition, including the development of improved and customized domain adaptation techniques. Also, we plan to look into the application of Bayesian methods to further improve continuous gesture recognition accuracy.
\section{Introduction}
One of the prominent fields in robotics is the development of artificial or prosthetic limbs. The most advanced devices are controlled via electromagnetic sensors that measure sEMG signals either from muscles \cite{fajardo2015} or the central nervous system \cite{Li2017}.
Artificial limbs are normally controlled via gesture recognition from the sEMG sensor data. This gesture recognition can be either discrete \cite{Du2017} or continuous \cite{Quivira2018}. In the discrete case, there is a limited number of available gestures that the limb can perform, and the goal is to map the sensor data to one of these gestures. Therefore, the recognition task is a classification problem. In our work, we focus instead on continuous gesture recognition, which derives the angles of the limb joints from the data and allows for more life-like limb movement.
In this paper, we perform experiments on the task of controlling a robotic hand via an arm band sensor. Previous work (e.g. \cite{fajardo2015, Du2017, Quivira2018}) assumed robotic hands with immobile wrists, which significantly restricts the functionality of the robotic limb. We present, for the first time, experiments with both immobile and mobile wrist data, and show that our proposed techniques outperform the state-of-the-art in both cases.
The most successful gesture recognition from the sEMG sensor data methods to date employ recurrent neural networks (RNNs) \cite{Quivira2018}. However, RNNs take a very long time to train to reach a desirable accuracy \cite{Lei2017}. In our work, we propose to use RNNs with simple recurrent units (SRU) \cite{Lei2017}, instead, for continuous gesture recognition. This method has the advantage of reaching a higher gesture recognition accuracy than regular RNNs in the same training time.
In addition, we show how domain adaptation techniques \cite{Ganin2014} can be used for continuous gesture recognition, and for the first time empirically demonstrate that these techniques improve the transferability between subjects, where a limb controller trained on data from one person is used for gesture recognition of another person.
\section{Background and Related Work}
\subsection{sEMG sensors}
Surface electromyography is the method of obtaining electric signals coming from neurons that are responsible from muscle contractions. As the name suggests, it is a non-invasive method and electrodes are attached to the skin surface. These signals can be then used in a control system for powered upper-limb prostheses because motions of wrist and fingers are mostly controlled by muscles in the forearm \cite{phinyomark2012}.
sEMG sensors range from non-portable units with high measurement accuracy to light and wearable units, albeit with a lower accuracy.
In order to retrieve sEMG data, we utilized the Myo armband \cite{Sathiyanarayanan2016}, as shown in Figure~\ref{img:myo}. This device has the advantages of being wearable and having an affordable price\footnote{At the time of writing the price is in the range of \$200}. The Myo armband consists of 8 sEMG sensors capable of recording sEMG data at frequency of 200 Hz. The raw signal is digitized with 8-bit analog-to-digital converter to yield a $[-128,+128]$ range. For wireless connection the Myo armband uses Bluetooth Low Energy technology.
\begin{figure}[thtb]
\centering
\includegraphics[width=0.45\textwidth]{images/band_on_my_arm}
\caption{Myo Armband}
\label{img:myo}
\end{figure}
\subsection{Motion capture system}
In order to tackle the task of continuous gesture recognition, a motion capture system to extract the limb angles is required for the generation of the data set. Existing motion capture systems are based on a wide range of technologies, including video recording \cite{kim2008}, ultrasound \cite{hettiarachchi2015}, and tracking gloves \cite{kim2009}.
In our work, the Leap Motion Controller \cite{leapmotion} was used, which provided a cheap and efficient way to track and record the limb angles of a user's hand. This device is specifically designed for hand movement detection. The controller itself consists of two monochromatic IR cameras and three IR LEDs. The overall accuracy of the controller was shown to be 0.7 millimeters. Motion capture data is transmitted with a frequency from 50 Hz up to 120 Hz.
\subsection{Recurrent Neural Networks}
Recurrent neural networks (RNNs) \cite{Jain:1999:RNN} are a special type of neural networks designed to work with (numerical) sequence data. The input of an RNN is the sequence and the output is a numerical vector. In our case, the input sequence are the sEMG measurements, and the output is a vector of limb angles. Every neuron of the network on each element of the input sequence updates its state and outputs a value as defined by:
\begin{equation}
h_t = \sigma_h(W_h x_t + U_h h_{t - 1} + b_h) \\
\end{equation}
\begin{equation}
y_t = \sigma_y(W_y h_t + b_y)
\end{equation}
where $x_t$ is the sequence element, $h_t$ is the hidden state, $y_t$ is the output vector, $W, U, b$ are cell parameters, and $\sigma_h$ and $\sigma_y$ are activation functions.
\subsection{Hand Gesture Recognition}
Initial work on controlling artificial limbs using EMG signals focused on creating a control scheme, i.e. mapping specific signals to specific actions of the limb (e.g. \cite{fajardo2015}). These approaches are not able to tackle a large variety of gestures and the resulting controllers are specific to an individual and can not be transferred to other persons. Gesture recognition methods have been introduced to overcome these limitations.
The gesture recognition task can be split into:
\begin{enumerate}
\item Classification with a number of discrete gestures or poses.
\item Regression with respect to recognizing continuous values such as joint angles.
\end{enumerate}
In most recent and advanced work, Du et al.\ \cite{Du2017} propose a hand gesture classification method employing domain adaption (specifically adaptive batch normalization, AdaBN). The authors generated several datasets with the number of gestures varying from 8 to 12, none of which involved wrist movement. The results showed that AdaBN can improve generalization and tranferability of a convolutional neural network with respect to accuracy on the discrete gesture recognition tasks.
Following up on this work, \cite{Quivira2018} focused on continuous gesture recognition (joint angles) using Gaussian processes. Again, unlike our work, this research did not involve any wrist movement.
\section{Method}
In this section, we present our approach for continuous gesture recognition. Specifically, we propose the use of RNNs with simple recurrent units (SRUs) to achieve a higher accuracy in the same time of training than regular RNNs using gated recurrent units (GRUs). Furthermore, we introduce a novel technique to apply domain adaptation to continuous gesture recognition.
\subsection{Simple Recurrent Units}
Due to vanishing gradient and exploding gradient problems vanilla RNN is not used nowadays \cite{chung2014}. The most common approach to date is to employ RNNs with GRU cells.
\subsubsection{GRU-cell}
The GRU-cell features special gates to reduce vanishing and exploding gradient problems. It is defined as:
\begin{equation} \label{eq:gru_z_t}
z_t = \sigma_g (W_z x_t + U_z h_{t - 1} + b_z)
\end{equation}
\begin{equation} \label{eq:gru_r_t}
r_t = \sigma_g (W_r x_t + U_r h_{t - 1} + b_r)
\end{equation}
\begin{equation}
h_t = (1 - z_t) \odot h_{t - 1} + z_t \odot \sigma_h (W_h x_t + U_h (r_t \odot h_{t - 1}) + b_h)
\end{equation}
where $x_t$ is the sequence element, $h_t$ is the hidden state and output, $z_t$ is the update gate, $r_t$ is the reset gate, and $W, U, b$ are cell parameters.
Figure~\ref{img:rec_scheme_num} illustrates an RNN architecture with GRU cells.
\begin{figure}[thtb]
\centering
\includegraphics[width=0.40\textwidth]{images/rec_scheme_num}
\caption{Recurrent neural network architecture \newline 1 - Input data, \newline 2 - The first GRU/SRU layer, \newline 3 - The second GRU/SRU layer, \newline 4 - Fully-connected layer of Predictor, \newline 5 - Fully-connected layer without activation-function - result vector.}
\label{img:rec_scheme_num}
\end{figure}
\subsubsection{SRU-cell}
The GRU-cell, due to its internal dependencies, is difficult to parallelize efficiently. In order to reduce computation cost for RNN training and inference, a simple recurrent unit was proposed in \cite{Lei2017} as follows:
\begin{equation}
\hat{x_t} = W x_t
\end{equation}
\begin{equation}
f_t = \sigma(W_f x_t + b_f)
\end{equation}
\begin{equation}
r_t = \sigma(W_r x_t + b_r)
\end{equation}
\begin{equation}
c_t = f_t \odot c_{t - 1} + (1 - f_t) \odot \hat{x_t}
\end{equation}
\begin{equation}
h_t = r_t \odot \sigma(c_t) + (1 - r_t) \odot x_t
\end{equation}
where $x_t$ is a sequence element, $h_t$ is the output vector, $c_t$ is the hidden state, $f_t$ is the forget gate, $r_t$ is the reset gate, and $W, U, b$ are cell parameters.
The RNN architecture with SRU cells is analogous to the RNN architecture shown in Figure~\ref{img:rec_scheme_num}.
\subsection{Domain adaptation}
When transferring a gesture recognition network trained on one person to another person, specialized transfer learning techniques are required. Adversarial domain adaptation (ADA) has been shown in recent work \cite{ganin2014unsupervised} to perform well for transfer learning tasks, and we have thus selected it for our research. ADA employs an additional neural network that predicts the current domain of input data from a feature vector. There is a gradient reversal layer between the feature vector and the discriminator network. This layer multiplies gradients by some value between -1 and 0. To train a neural network with ADA, we sum the loss function of the predictor and the loss function of the discriminator and perform back-propagation with a reversal of the discriminator gradient before the feature layer. The goal of this process is to inhibit domain-specific features and thus prevent over-fitting.
Figure~\ref{img:rec_scheme_adapt} shows the RNN architecture with domain adaptation.
We have also tried adaptive batch normalization, as used for discrete hand gesture recognition in \cite{Du2017}, but this did not show good performance on our continuous gesture recognition tasks.
\begin{figure}[thtb]
\centering
\includegraphics[width=0.40\textwidth]{images/rec_scheme_num_adapt}
\caption{Recurrent neural network with domain adaptation architecture \newline 1 - Input data, \newline 2 - The first GRU/SRU layer, \newline 3 - The second GRU/SRU layer, \newline 4 - Fully-connected layer of Predictor, \newline 5 - Fully-connected layer without activation-function - result vector, \newline 6 - Fully-connected layer of Discriminator, \newline 7 - Fully-connected layer without activation function - vector of classes (domains) probabilities.}
\label{img:rec_scheme_adapt}
\end{figure}
\section{Data acquisition}
sEMG data was obtained by means of the Myo armband device and finger and wrist angles were captured with the Leap Motion motion capture system.
We recorded data recorded from 5 healthy volunteers ranging in age from 20 to 28 years. Each participant was asked to sit on a chair and establish a relaxed position with the wrist slightly above the Leap Motion sensor to ensure accurate motion capture. The participant was then asked to perform a given sequence of hand movements such as bending or stretching the index finger.
Two datasets were collected: the first dataset only included movement of fingers with a static wrist (similar to \cite{Quivira2018}, and the second one was aimed to add wrist movements and some casual gestures. Data was collected in sessions with a duration of 4 minutes each. The specifics of the datasets are as follows.
\hfill
\noindent
Immobile wrist dataset:
\hfill
\begin{enumerate}
\item Movements of distinct fingers, 150 s
\item Simultaneous movement of all fingers, 60 s
\item Free finger movements, 30 s
\end{enumerate}
\hfill
\noindent
Mobile wrist dataset:
\hfill
\begin{enumerate}
\item Movements of distinct fingers, 60 s
\item Simultaneous movement of all fingers, 30 s
\item Pinch, 30 s
\item Open palm movements, 30 s
\item Thumb movements, 30 s
\item Free finger movements, 30 s
\end{enumerate}
\hfill
For each volunteer 8 sessions per dataset were recorded. For the mobile wrist dataset, the volunteers were asked to do the finger movements while keeping the wrist angle at one of 6 specified positions, including four angles on the up-down axis of the wrist, bending the wrist to the right, and keeping it straight.
Due to different data acquisition rates of the Myo armband and Leap Motion there is a need for data synchronization and alignment. Global alignment was achieved by matching the dataframes from the two devices, minimizing the difference between the time stamps of the individual dataframes, with a maximum difference of 10 ms.
After alignment, the sEMG and limb angle data was filtered with a low pass filter with a frequency of 10 Hz and 4Hz respectively to achieve noise reduction. The sEMG data was split into windows of size 128, which formed the input vector for the recurrent neural network.
\section{Experiment Setup}
In this section we provide the design parameters of the two RNN approaches used in our experiments. The first approach represents standard RNN technology, while the second approach incorporates optimizations selected by us for the task of continuous gesture recognition.
Adversarial domain adaptation was implemented with a gradient reversal layer after the feature generation block. The discriminator block consists of 2 fully connected layers with 256 neurons in the first layer and 15 or 18 neurons in the second (output) layer, since there are 15 angles to predict in the case of an immobile wrist and 18 in case of a mobile wrist.
The hyper-parameters of the recurrent neural network with GRU used in our experiments are:
\begin{itemize}
\item 2 GRU recurrent layers with 256 neurons each. Only the last output was fed to the next layer.
\item 2 fully connected layers with 256 and 15 (18) neurons respectively.
\item Optimizer: Adam \cite{kingma2014adam} with a learning rate of 0.001.
\item Trained for 30 epochs or until there was no validation gain during 8 epochs.
\end{itemize}
The hyper-parameters for the recurrent neural network with SRU are:
\begin{itemize}
\item 2 SRU recurrent layers with 256 neurons each.
\item Global Average Pooling layer.
\item 2 fully connected layers with 256 and 15 (18) neurons respectively.
\item Optimizer: Adam with learning rate equals to 0.001.
\item Net was trained for 30 epochs or until there was no validation gain during 8 epochs.
\end{itemize}
\section{Evaluation Metrics}
\makeatletter
\DeclareRobustCommand{\vardivision}{%
\mathbin{\mathpalette\@vardivision\relax
}
\newcommand{\@vardivision}[2]{%
\reflectbox{$\m@th\smallsetminus$}%
}
\makeatother
In our evaluation, we measure the root mean square error (RMSE) and normalized root mean square error (NRMSE). RMSE is used as a standard way to measure regression error. In addition, we used NRMSE to normalize the errors to account for different limb movement ranges. RMSE and NRMSE are given in equations~\ref{rmse} and~\ref{nrmse} respectively.
$y_i$ denote the real values and $\hat{y_i}$ the predicted values. $y_{range}$ is the range of real values. $\Theta_i = y_i \vardivision y_{range}$ and $\hat{\Theta_i} = \hat{y_i} \vardivision y_{range}$.
where $n$ is the number of angles predicted (15 for an immobile wrist and 18 for a mobile wrist).
\begin{equation}
\label{rmse}
RMSE = \sqrt[]{\frac{1}{n}\sum^{n}_{i=0}{(y_i - \hat{y_i})^2}}
\end{equation}
\begin{equation}
\label{nrmse}
NRMSE = \sqrt[]{\frac{1}{n}\sum^{n}_{i=0}{(\Theta_i - \hat{\Theta_i})^2}}
\end{equation}
\newcolumntype{C}[1]{>{\centering\let\newline\\\arraybackslash\hspace{0pt}}m{#1}}
\begin{table*}[tb]
\centering
\caption{Results for immobile wrist}
\label{all_table_static}
\begin{tabular}{|l|p{32mm}|c||C{20mm}|C{20mm}||C{20mm}|C{20mm}|}
\hline
\multirow{2}{10mm}{Metric} & \multirow{2}{12mm}{Model} & \multirow{2}{15mm}{Intra session} & \multicolumn{2}{p{16mm}||}{Inter session} & \multicolumn{2}{p{16mm}|}{Inter subjects} \\ \cline{4-7}
& & & no ADA & with ADA & no ADA & with ADA \\
\hline
\multirow{3}{*}{RMSE}
& Gaussian Process & 23.29$\pm$0.15 & 23.44$\pm$0.05 & - & 23.91$\pm$0.01 & - \\ \cline{2-7}
& GRU Recurrent neural net & 18.26$\pm$0.04 & 19.80$\pm$0.14 & 20.50$\pm$0.07 & 23.45$\pm$0.30 & \textbf{22.29$\pm$0.24} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{17.99$\pm$0.04} & \textbf{19.24$\pm$0.04} & 19.70$\pm$0.03 & 23.14$\pm$0.06 & \textbf{22.07$\pm$0.18} \\
\cline{2-7}
\hline \hline
\multirow{3}{*}{NRMSE}
& Gaussian Process & 0.2456$\pm$0.0016 & 0.1924$\pm$0.0000 & - & 0.2034$\pm$0.0001 & - \\ \cline{2-7}
& GRU Recurrent neural net & 0.1552$\pm$0.0004 & 0.1604$\pm$0.0008 & 0.1684$\pm$0.0005 & 0.1984$\pm$0.0023 & \textbf{0.1890$\pm$0.0021} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{0.1512$\pm$0.0004} & \textbf{0.1558$\pm$0.0004} & 0.1613$\pm$0.0004 & 0.1967$\pm$0.0002 & \textbf{0.1879$\pm$0.0019} \\
\cline{2-7}
\hline
\end{tabular}
\end{table*}
\begin{table*}[tb]
\centering
\caption{Results for mobile wrist}
\label{all_table_wrist}
\begin{tabular}{|l|p{32mm}|c||C{20mm}|C{20mm}||C{20mm}|C{20mm}|}
\hline
\multirow{2}{10mm}{Metric} & \multirow{2}{12mm}{Model} & \multirow{2}{15mm}{Intra session} & \multicolumn{2}{p{16mm}||}{Inter session} & \multicolumn{2}{p{16mm}|}{Inter subjects} \\ \cline{4-7}
& & & no ADA & with ADA & no ADA & with ADA \\
\hline
\multirow{3}{*}{RMSE}
& Gaussian Process & 22.33$\pm$0.24 & 22.23$\pm$0.00 & - & 22.39$\pm$0.01 & - \\ \cline{2-7}
& GRU Recurrent neural net & 18.60$\pm$0.05 & 19.19$\pm$0.05 & 19.93$\pm$0.06 & 21.74$\pm$0.20 & \textbf{20.94$\pm$0.17} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{18.26$\pm$0.11} & \textbf{18.83$\pm$0.04} & 19.55$\pm$0.03 & 21.50$\pm$0.11 & \textbf{20.89$\pm$0.09} \\ \cline{2-7}
\hline \hline
\multirow{3}{*}{NRMSE}
& Gaussian Process & 0.2360$\pm$0.0033 & 0.1770$\pm$0.0000 & - & 0.1824$\pm$0.0001 & - \\ \cline{2-7}
& GRU Recurrent neural net & 0.1558$\pm$0.0004 & 0.1518$\pm$0.0004 & 0.1580$\pm$0.0006 & 0.1757$\pm$0.0016 & \textbf{0.1698$\pm$0.0012} \\ \cline{2-7}
& SRU Recurrent neural net & \textbf{0.1516$\pm$0.0013} & \textbf{0.1490$\pm$0.0000} & 0.1550$\pm$0.0000 & 0.1742$\pm$0.0008 & \textbf{0.1695$\pm$0.0007} \\
\cline{2-7}
\hline
\end{tabular}
\end{table*}
\section{Evaluation} \label{sec:evaluation}
In order to show the superiority of the improved RNN approach (RNN with SRU) over other recent approaches from the literature, namely a standard RNN with GRU approach (similar to \cite{Du2017}) and Gaussian Process (as in \cite{Quivira2018}), we applied these three methods to our datasets with and without mobile wrists.
In addition, we applied adversarial domain adaption (ADA) to the RNN approaches to evaluate the transfer ability of each of these methods. Note, that ADA is not applicable to Gaussian Processes.
We split the data into training, validation and test sets as follows:
\begin{enumerate}
\item Each session was split into blocks of 12 seconds each.
\item A period of 3 seconds was randomly sampled from each block.
\item Half of these periods formed a validation set, while the other half formed the test set
\item The training set was created from the sessions after removing all periods that were used for the validation and test set.
\end{enumerate}
In other words, to get intra-session results all training, validation and test sets from all session were combined together. Half of this data was used for training, 25\% was used for validation and the remaining 25\% for testing.
In order to obtain inter-session results, we split the data set into five partitions, where the partitions do not share data from the same session. We used four of the partitions for training and validation, and the remaining partition for testing. This was repeated five times where each partition is used as a test set exactly once.
A similar approach was used to evaluate inter-subject performance, where each of the five partitions contained data from a single person.
\section{Results}
The results of the evaluation with immobile wrist and mobile wrist data are shown in tables~\ref{all_table_static} and~\ref{all_table_wrist} respectively. The evaluation results for the three machine learning algorithms are presented in terms of RMSE and NRMSE for the three types of experiments: intra-session, inter-session, and inter-subject (as described in Section~\ref{sec:evaluation}). For inter-session and inter-subject we also compare results with and without the application of adversarial domain adaptation (ADA).
First of all, recurrent neural networks with SRUs show better performance in most cases, and comparable performance in the case of inter-subject experiments. These results hold for both mobile and immobile wrists. The apparent gesture recognition accuracy improvement when comparing mobile wrist with immobile wrist results is due to an increased number of predicted angles in the mobile wrist case (18 vs. 15). Therefore, the two cases are not really comparable to each other.
As can be seen, adversarial domain adaptation (ADA) improves inter-subject accuracy, while actually showing worse results for inter-session experiments. This demonstrates that inter-session differences are not significant enough for the inhibition of session-specific features (and the resulting prevention of over-fitting) offered by ADA to have a positive impact on accuracy. However the case is different for inter-subject testing, where the difference in recorded data can be quite substantial.
\section{Conclusion and Outlook}
In this paper, we presented experiments with three machine learning approaches applied to continuous gesture recognition based on sEMG measurements. Our main contributions are:
\begin{enumerate}
\item We presented for the first time an evaluation of machine learning techniques with mobile wrists, which significantly enhance the functionality of artificial limbs.
\item We demonstrated that RNNs with SRUs are superior to other state-of-the-art approaches to gesture recognition proposed in the literature, specifically Gaussian Processes and RNNs with GRUs.
\item We presented for the first time inter-subject gesture recognition experiments and showed the advantages of adversarial domain adaption for this task.
\end{enumerate}
In future work, we intend to further investigate the case of inter-subject gesture recognition, including the development of improved and customized domain adaptation techniques. Also, we plan to look into the application of Bayesian methods to further improve continuous gesture recognition accuracy.
| 7f4051036f2d93e1282a4cebec8521680fcbde38 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
In the rapidly growing field of quantum science and technologies, the quest for new reliable and effective techniques to manipulate systems at the quantum scale is of paramount importance, and control engineers are actively contributing to this effort (see e.g. \cite{altafini-tutorial} and references therein). Due to the intrinsic nature of quantum systems, one of the central resources of classical control design - {\em feedback} control - is particularly difficult to harness. In the last two decades, great progress has been made in this sense, building on the foundation of quantum probability and filtering theory \cite{belavkin,quantum-filtering,ticozzi2012stabilization}, and arriving to remarkable experimental implementations \cite{haroche-feedback}.
In this paper, we combine the advantages offered by measurement-based feedback with switching control strategies, which allow us to include {\em dissipative} control resources in a systematic way, as opposed to the more typical Hamiltonian control (see e.g. \cite{VanHandel2005, Mirrahimi2007,} and \cite{altafini-tutorial} for a review). { The main contribution is a switching control strategy for the stabilization of a target pure state or subspace, where the current dynamics is selected based on the estimation of the state at the switching times, and maintained for a given dwell time. Such strategy guarantees practical stability of the target in mean under minimal assumptions (existence of a switching control Lyapunov function), while under typical control assumptions the stability is guaranteed both in mean and almost surely.}
The paper is structured as follows: Section \ref{sec:prob_def} defines the problem of interest and present some theoretical results that are used in the rest of the paper; { Section \ref{sec:control_strategies} discusses the assumptions under which effective strategies can be derived, recalls open-loop switching strategies previously introduced in \cite{Scaramuzza-Ticozzi}, presents the novel measurement-based, closed-loop control laws and proves its stability. The analysis of convergence uses techniques that depart from those based on linear systems used in \cite{Scaramuzza-Ticozzi} and build on specific properties of the stochastic filtering dynamics \cite{Ticozzi2017}.} Section \ref{sec:case_study} provides some insights on the performances of the novel control law by presenting a relevant case study and simulation results for the stabilization of entangled states on networks of two-level systems.
\section{PROBLEM DEFINITION}
\label{sec:prob_def}
In this article we consider a system described by a finite-dimensional Hilbert space $\mathcal{H}$. Following the standard in physics and quantum information science, we shall employ Dirac's notation for vectors $\ket{\psi}\in\cal{H},$ and their duals $\bra{\psi}\in\cal{ H}^\dag.$ Let $\mathcal{B}(\mathcal{H})$ denote the set of linear operators on $\mathcal{H}$.
The state of a quantum system is completely described by a density operator $\rho\in\mathfrak{D}(\mathcal{H}) = \{\rho\in \mathcal{B}(\mathcal{H}):\rho=\rho^\dag\geq0, \ \text{tr}(\rho)=1\}$.
We will suppose that the system is controlled with a series of driving dynamics and that it is subjected to an homodyne detection measurement. The resulting dynamics is thus described by processes $(\rho_t)_{t\in\mathbb{R}_+}$ of states associated to the stochastic master equation (SME) \cite{quantum-filtering,altafini-tutorial}
\begin{equation}
d\rho_t=\mathcal{L}_j(\rho_t)dt+\mathcal{G}_C(\rho_t)dW_t,
\label{eqn:model}
\end{equation}
where:
\begin{align*}
\mathcal{L}_j(\rho_t) &= -i[H_j,\rho_t] + \mathcal{D}_{L_j}(\rho_t) +\mathcal{D}_C(\rho_t),\\
\mathcal{D}_A(\rho_t) &= A\rho_t A^\dag-\frac{1}{2}\{A^\dag A, \rho_t\},\\
\mathcal{G}_C(\rho_t) &= C\rho_t +\rho_t C^\dag -\text{tr}\left((C+C^\dag)\rho_t\right)\rho_t
\end{align*}
In equation \eqref{eqn:model}, the term $-i[H_j,\rho_t] + \mathcal{D}_{L_j}(\rho_t)$ represents the driving dynamics associated to the Hamiltonian and noise operators $(H_j, L_j)$. We suppose that a list of possible driving dynamics is provided and the role of the controller will be to choose which dynamics to activate. This concept will be further formalized in the following.
The term $\mathcal{D}_C(\rho_t)dt+\mathcal{G}_C(\rho_t)dW_t$ accounts for the homodyne detection measurement process associated to the fixed noise operator $C$. For simplicity we here consider unit detection efficiency, but the control strategy works with imperfect detection as well. The process $dW_t$ is a Wiener process, adapted to the filtration ${\cal F}_t$ \cite{quantum-filtering}, and it can be seen as the innovation for the
homodyne measurement output $ dY_t =\text{tr}\left((C^\dag+C)\rho_t\right)dt+dW_t.$
If the measurement record is not accessible, the best description of the state evolution can be obtained as the expectation of \eqref{eqn:model} over the outcomes of the measurement process. Namely, defining $\hat{\rho}_t = \mathbb{E}[\rho_t|{\cal F}_t]$, we have that the time evolution of $\hat{\rho}_t$ is described by the Markovian master equation (MME) \cite{alicki-lendi}:
\begin{equation}
\frac{d}{dt}\hat{\rho}_t = \mathcal{L}_j(\hat{\rho}_t).
\label{eqn:MME_model}
\end{equation}
Throughout this work, we will consider the stabilization of linear subspaces of $\mathcal{H}$, and the relevant particular case of pure states.
Let $\mathcal{H_S}$ be the target subspace of $\mathcal{H}$. Denoting $P_S$ the orthogonal projector on $\mathcal{H_S}$, we can describe the set of states whose support is $\mathcal{H_S}$ or a subspace of $\mathcal{H_S}$ as
\begin{equation}
\mathcal{I_S(H)} = \{\rho\in\mathfrak{D}(\mathcal{H}): \text{tr}(P_S\rho)=1\}.
\end{equation}
With a slight abuse of terminology, we say that $\mathcal{H_S}$ is invariant, stable, or attractive for the dynamics if such is the supported state-set $\mathcal{I_S(H)}.$
The subspace $\mathcal{H_S}$ is said globally asymptotically stable (GAS) for \eqref{eqn:model}:\vspace{1mm}\\ - in mean if $\lim\limits_{t\to\infty}\norm{\hat{\rho}_t - P_S \hat{\rho}_t P_S }_1=0$, $\forall\rho_0\in\mathfrak{D}(\mathcal{H})$;\\
- almost surely ({\em a.s.}) if $\mathbb{P}\left(\lim\limits_{t\to\infty}\norm{\rho_t - P_S \rho_t P_S }_1=0\right)=1$, $\forall\rho_0\in\mathfrak{D}(\mathcal{H})$.
We next summarize some results of \cite{Ticozzi2017} that will be of use to our aims in the following theorem. { A Lyapunov function is a functional $V:\mathfrak{D}(\mathcal{H})\to\mathbb{R}_+$ such that: $V(\rho)\geq0$, with $V(\rho)=0$ if and only if $\rho\in\mathcal{I_S(H)}$ and $\frac{d}{dt}V(\rho)<0$ for all $\rho\notin\mathcal{I_S(H)}$.
\begin{theorem}
Consider system \eqref{eqn:model}, with a fixed $\mathcal{L_j$. A subspace $\mathcal{H_S}$ of $\mathcal{H}$ is:
\begin{itemize}
\item GAS in mean if and only if it is GAS almost surely;
\item if GAS in mean if and only if there exists an operator $K\geq0$ such that $V(\rho)=\text{tr}(K\rho)$ is a Lyapunov function.
\end{itemize}
\label{thm:mean_as_equ}
\end{theorem}
{\em Remark:} These results are the starting point for our analysis and deserve some explanations: (i) The linear $V$ is associated with a $K$ such that $K_R:=P_SKP_S>0,$ and the latter matrix is derived from a (perturbed) Perron-Frobenius eigen-operator for the dual dynamics, reduced to the complement of the target. (ii) The equivalence of asymptotic stability in mean and a.s. is {\em crucially dependent on the fact the target is a subspace}, and not true otherwise. In our main theorem we will use the same proof idea, applied to switching evolutions.}
We shall be interested in sequences of switching times that are unbounded countable set of times $0,t_1,t_2,\dots$ such that $t_k-t_{k-1}>\epsilon$ for some $\epsilon>0.$ The assumption $t_k-t_{k-1}>\epsilon$ for $\epsilon>0$ is introduced to prevent chattering: for this reason, we shall refer to a sequence with the properties above as a {\em non-chattering}. { Non-chattering time sequences are essential for practical implementations, as well as ensuring well-behaved solution of the SME. This point will not be explicitly discussed in this work but the proof follows the discussion of \cite{Mirrahimi2007}.} Finally, we state the control problem of interest.
\begin{problem}
Given a target subspace $\mathcal{H_S}$ of $\mathcal{H}$ and a finite set of generators $\{\mathcal{L}_j\}_{j=1,\dots,m}$ for model \eqref{eqn:model}, find a piece-wise constant switching control law $j(t):[0,+\infty)\to\{1,\dots,m\}$ that admits a set of non-chattering switching times, so that $\mathcal{H_S}$ is made GAS (in mean and/or almost surely) by selecting $\mathcal{L}_{j(t_k)}$ on $[t_k,t_{k+1})$.
\end{problem}
\section{CONTROL STRATEGIES}
\label{sec:control_strategies}
In this section, we will first recall two previously proposed switching techniques, based on {\em open-loop} switching, and next present our {\em closed-loop} proposal.
\subsection{Control Assumptions}
The following assumption is typically required to prove convergence of a switching law.
\begin{assumption}
\label{assumption}
{Each Lindblad generator has the target subspace $\mathcal{H_S}$ as invariant , and}
there exists $\alpha\in[0,1]^m$, $\norm{\alpha}_1=1$ such that $\mathcal{H_S}$ is GAS in mean for
\begin{equation}
\label{eq:assump1} \frac{d}{dt}\hat{\rho}_t = \mathcal{L}_{c}(\hat{\rho}_t) = \sum_{j=1}^m\alpha_j \mathcal{L}_{j}(\hat{\rho}_t).
\end{equation}
\end{assumption}
We also introduce a second working assumption, which relaxes Assumption 1 above. Essentially, it requires a linear control Lyapunov function.
\begin{assumption}
\label{assumption_2}
There exists a linear Lyapunov function $V(\rho)$ such that $\forall \rho\notin\mathcal{I_S(H)}$, $\exists j$: $V(\mathcal{L}_j(\rho))<0$ and $\forall \rho\in\mathcal{I_S(H)}$, $\exists j$: $V(\mathcal{L}_j(\rho))=0$.
\end{assumption}
One of the key differences between the assumptions is that the second does not require invariance of the target for each generator, and is thus weaker, as we argue in the following.
\begin{proposition}\label{prop:1implies2}
Assumption \ref{assumption} implies Assumption \ref{assumption_2}.
\end{proposition}
\begin{proof}
If Assumption 1 holds, from Theorem \ref{thm:mean_as_equ} we have that there exists a linear Lyapunov function $V(\rho)$ such that $V(\rho)>0$ for all $\rho\notin\mathcal{I_S}(\mathcal{H})$, $V(\rho)=0$ for all $\rho\in\mathcal{I_S}(\mathcal{H})$ and $V(\mathcal{L}_c(\rho))<0$ for all $\rho\notin\mathcal{I_S}(\mathcal{H})$. This means that we have
$ V(\mathcal{L}_c(\rho))=\sum_{j=1}^m\alpha_j V(\mathcal{L}_{j}(\rho))<0
$
for all $\rho\notin\mathcal{I_S}(\mathcal{H})$, thanks to the linearity of $V(\rho)$. Then, since $\alpha_i>0$ and $\norm{\alpha}_1=1$, we have that there exist an index $
j^* = \arg\min_{j=1,\dots,m} V(\mathcal{L}_{j}(\rho))
$
such that
$ V(\mathcal{L}_{j^*}(\rho)) \leq \sum_{j=1}^m\alpha_j V(\mathcal{L}_{j}(\rho)) = V(\mathcal{L}_c(\rho))< 0
$
for any $\rho\notin\mathcal{I_S}(\mathcal{H})$.
{Finally, since from Assumption \ref{assumption} we have that every Lindblad generator leaves $\mathcal{H_S}$ invariant, we have that $\forall j$ $V(\mathcal{L}_j(\rho))=0$, $\forall\rho\in\mathcal{I_S(H)}$.}
\end{proof}
Proposition \ref{prop:1implies2} will be instrumental to the proof of the main theorem. It is possible to prove that the converse implication is not true. For example, consider the following three-level system in ${\cal H}=\textrm{span}\{\ket{0},\ket{1},\ket{2}\}$. Let $\ket{0}\bra{0}$ be the target state, and consider the two following generators:
\begin{align*}
\mathcal{L}_1(\rho):( &H_1 = 0, L_1=\Outer{0}{2}+\Outer{1}{1}+\Outer{2}{2}),\\
\mathcal{L}_2(\rho):( &H_2=\Outer{0}{2}+\Outer{2}{0},\\
&\quad L_2=\Outer{0}{1} +\Outer{1}{1}+\Outer{2}{2} ).
\end{align*}
Notice that the generator $\mathcal{L}_1(\rho)$ stabilizes the space generated by $\textrm{span}\{\ket{0},\ket{1}\},$ being the only invariant space for $L_1$ and has trivial dynamics on it \cite{ticozzi-QDS}, but does not make the target GAS. For the generator $\mathcal{L}_2(\rho)$, instead, $\ket{0}$ is not even invariant.
It is then quite clear that there exists no convex combination of $\mathcal{L}_1(\rho)$ and $\mathcal{L}_2(\rho)$ that makes $\ket{0}\bra{0}$ GAS in mean.
However, by considering $V(\rho) = \text{tr}(K\rho)$ with $K$ the projector on the subspace orthogonal to the target,
it is possible to prove by direct computation that $\forall\rho\neq\ket{0}\bra{0}$, $V(\mathcal{L}_j(\rho))<0$ for some $j$ and $V(\mathcal{L}_j(\ket{0}\bra{0}))=0$ for $j=1,2$.
This shows that Assumption \ref{assumption_2} is genuinely more general than Assumption \ref{assumption}. However, if one requires non-chattering control strategies, we will argue it only allows for a weaker stability notion, and convergence in mean.
\subsection{Open-loop strategies: cyclic and state-based switching}
The first control strategy we consider is a {\em time-based} solution \cite{Scaramuzza-Ticozzi}, based on Assumption \ref{assumption}.
\begin{definition}[Cyclic switching control law]
Given the vector $\alpha$ that satisfies Assumption \ref{assumption} for the set of Lindblad dynamics $\{\mathcal{L}_j\}$, the cyclic switching control law selects each index $j$ for a fraction $\alpha_j$ of the total cycle period $\varepsilon>0$.
\end{definition}
Note that this control law depends only on the vector $\alpha$ from Assumption \ref{assumption}, and no information on the initial state of the system is required. Essentially, for $\varepsilon\rightarrow 0$ it mimics the evolution generated by the convex combination \eqref{eq:assump1}, and makes $\mathcal{H_S}$ GAS in mean. A full proof can be found in \cite{Scaramuzza-Ticozzi}.
The second control law we recall is also proposed in \cite{Scaramuzza-Ticozzi}, and here specialized to the case of a linear Lyapunov function.
\begin{definition}[State-based switching control law]
Given a set of Lindblad operators $\{\mathcal{L}_j\}$, an estimate of the initial state $\rho_0$, a linear Lyapunov function $V(\cdot)$ that guarantees that $\mathcal{H_S}$ is GAS for the system $\mathcal{L}_c$ as in \eqref{eq:assump1} and a {non-chattering} switching time sequence $\{t_k\}$, the state-based switching control law is defined as:
\begin{equation}
j(t) = \arg\min_{j=1,\dots,m} V(\mathcal{L}_j(\hat{\rho}_{t_k})) , \quad \forall t\in[t_k,t_{k+1})
\end{equation}
where $\hat{\rho}_{t_k}$ is the average of the system state at time $t_k$, solution of the MME.
\end{definition}
The original proposal in \cite{Scaramuzza-Ticozzi} employed a quadratic function to show the existence of a non-chattering sequence; in this case, it follows from Theorem \ref{thm:main} in the following section.
\subsection{Measurement-based switching control}
The approach we propose exploits continuous measurements to have a {\em closed-loop} estimate of the current state of the system, and then apply the state-based switching control law based on this estimate - as opposed to the MME average state.
An important step in setting up the feedback loop is the choice of the measurement operator $C$, so that the measurement process does not destabilize the target - namely, the target should be an eigenstate or, more generally, an eigenspace, of $C$.
A natural choice is to resort again to Theorem \ref{thm:mean_as_equ} and choose $C=K$, where $K$ is the matrix of the Lyapunov function $V(\rho)=\text{tr}(K\rho)$ that satisfies Assumption \ref{assumption_2} for the set of MME generators $\{\mathcal{L}_j\}$. In this way we have that $\mathcal{H_S}\subseteq \textrm{ker}(C)$ is naturally satisfied by the construction of $K$, and also it holds that $V(\mathcal{D}_K(\rho))=0$ for all $\rho$. From now on we will only consider this choice of $C$ for sake of simplicity.
The switching control law we consider is:
\begin{definition}[Measurement-based switching control law]
\label{def:measurement_based}
Given a set of Lindblad operators $\{\mathcal{L}_j\}$, an estimate of the initial state $\rho_0$, a $V(\rho) = \text{tr}(K\rho)$ that satisfies Assumption \ref{assumption_2} for the set $\{\mathcal{L}_j\}$
and a {non-chattering} switching time sequence $\{t_k\}$, the measurement-based switching control law is defined as:
\begin{equation}
j(t) = \arg\min_{j=1,\dots,m} V(\mathcal{L}_j(\rho_{t_k})), \quad \forall t\in[t_k,t_{k+1})
\end{equation}
where $\rho_{t_k}$ is the estimate of the state at $t_k$ obtained as the solution of equation \eqref{eqn:model}.
\end{definition}
We will now prove stability of the proposed switching law.
\begin{theorem}\label{thm:main}
{ If Assumption \ref{assumption} holds, then there exists a non-chattering time sequence such that} the measurement-based switching control law makes $\mathcal{I_S}(\mathcal{H})$ GAS in mean and almost surely.
\end{theorem}
\begin{proof}
{
We start by proving the existence of a non-chattering time sequence, then we prove stability in mean, and finally, we argue that a.s. stability follows from stability in mean and Theorem \ref{thm:mean_as_equ}.
If Assumption \ref{assumption} holds, then by Proposition \ref{prop:1implies2} we can construct a linear Lyapunov $V(\rho)=\text{tr}(K\rho)$ function as in Assumption \ref{assumption_2}.
Assume $\dot{V}(\rho_{t_k}) := V(\mathcal{L}_j(\rho_{t_k}))<0$ and define $\hat{t} = \min\{t \text{ s.t. }V(\rho_t) \leq r V(\rho_{t_{k}})\}$ for some $r\in[0,1)$ fixed. We want to prove that $\exists\varepsilon>0$ such that $\hat{t}-t_k>\varepsilon$, and that it does not depend on the state at time $t_k$.
We can then apply the mean value theorem: there exists $s\in[t_k,\hat{t}]$ such that $\ddot{V}(\rho_s)(\hat t-t_k) = \dot{V}(\rho_{\hat t}) - \dot{V}(\rho_{t_k}) = (1-r)|\dot{V}(\rho_{t_k})|$.
Since by Assumption \ref{assumption} all evolutions leave $\mathcal{I_S}(\mathcal{H})$ invariant we have that:
\begin{align}
| \dot V(\rho_{t_k}) | & = |\text{tr}( \mathcal{L_{j}^\dag(K) \rho_{t_k})| \geq |\text{tr}( \mathcal{L_C^\dag(K) \rho_{t_k})|\nonumber \\
&= |\text{tr}( \mathcal{L_{C,R}^\dag(K_R) \rho_{R,t_k})|\geq k_{C,min} \text{tr}(\rho_{R,t_k}), \label{eq:bound}
\end{align}
where the $R$ subscript denotes the restriction of the operators to the complement of the target space and $k_{C,min}$ is the minimum absolute value of the eigenvalues of $\mathcal{L_{C,R}^\dag(K_R),$ and in the second line we used Proposition 2.5 of \cite{Ticozzi2017}.
On the left, we then have
$| \ddot V(\rho_s) | = | \text{tr}( K \mathcal{L^2_j(\rho_S)|= |\text{tr}( K_R \mathcal{L^2_{R,j}(\rho_{R,s})| \leq \bar k_{2,max} \text{tr}(\rho_{R,s}),$
where $\bar k_{2,max}$ is the maximum absolute value of the eigenvalues of $(\mathcal{L_{R,j}^{\dag})^2(K_R),$ for all $j$.
Then, since $e^{(t-t_j)\mathcal{L_{R,j}}$ is a trace-non-increasing map for any finite interval and $\rho_{R,t}=e^{(t-t_k)\mathcal{L_{R,j}}\rho_{R,t_k}$, we have $\text{tr}(\rho_{R,t_k})\geq\text{tr}(\rho_{R,t})$, $\forall t \in[t_k,\hat{t}]$.
Combining all the above inequalities we obtain:
$ \hat t-t_k\geq {(1-r)k_{C,min}}/{\bar k_{2,max}} $
then taking $t_{k+1}\in[t_k,\hat{t}]$ we obtain a non-chattering time sequence such that $\dot{V}(\rho_t)<0$ $\forall t \in[t_k,t_{k+1}]$.}
We will now prove stability in mean. We shall use a generalized version of Barbalat's Lemma and in particular Corollary 1 in \cite{Huang2012} to prove stability.
We have that $V(\rho)$ is lower bounded since $V(\rho)\geq0.$ By construction, $\dot{V}(\rho) =V(\mathcal{L}(\rho))$ is piece-wise continuous, and
non-positive in the time interval for a non-chattering sequence constructed as above. Moreover, being $V(\mathcal{L}(\rho))$ linear in $\rho$ for any time interval $t\in[t_i,t_{i+1})$, we have that $V(\rho)$ is twice differentiable in any time interval. Computing the second derivative of $V(\cdot)$ with respect to time, we get $\ddot{V}(\rho)=\text{tr}(\mathcal{L}_j^\dag(K)\mathcal{L}_j(\rho))$ which is linear in $\rho$, hence bounded for all $t\in[t_i,t_{i+1})$. Thus all hypothesis of Corollary 1 in \cite{Huang2012} are satisfied, and $\dot{V}(\rho)\to0$ in mean for $t\to\infty$.
{ This proves stability in mean, and that $V(\rho_t)$ is a (continuous) positive supermartingale for SME. Convergence in mean implies $L^1$ convergence to 0; on the other hand, by bounded supermartingale convergence theorem $V(\rho)$ converges both in $L^1$ and a.s. to some $V_\infty$ (see also proof of Theorem 1.1 in \cite{Ticozzi2017}.) Since we know that the $L^1$-limit is $0,$ convergence a.s. is also guaranteed. Being $V$ positive on the complement of the target, $V(\rho)=0$ a.s. implies $\rho\in{\cal I}_S({\cal H})$ a.s. as well.}
\end{proof}
{
{\em Remark:} For $\Delta t=t_{k+1}-t_k\rightarrow 0$ the proposed strategy tends to select {\em each time} $\mathcal{L_{j(t)}$ such that
$j(t)=\textrm{argmin}_jV(\mathcal{L_j\rho_t)$. While granting the {\em optimal} convergence rate at each time, at least if the estimated state is correct, this continuous strategy is not practically viable, and could lead to chattering. Assumption \ref{assumption}, however, allows us to derive non-chattering switching sequences (and a worst-case exponential bound using the dominant eigenvalue of $\mathcal{L_{C,R}$). Hence it is key to assess how well the proposed strategy fares with respect to the optimal one with faulty initializations. In the next sections we shall focus on this, leaving the estimate of the optimal convergence rate for future work.
We next show that the weaker Assumption \ref{assumption_2} is still sufficient to prove {\em practical} stability in mean.
\begin{theorem}
Assume Assumption 2 to hold for $V(\rho)=\text{tr}(K\rho)$. Then for every $\varepsilon>0$ there exists a non-chattering sequence such that the measurement-based strategy stabilizes the $\varepsilon$-``neighborhood" of the target ${\cal I}_{S,\varepsilon}=\{\ \rho\in \mathfrak{D}(\mathcal{H})|\min_{\sigma\in{\cal I}_S} \|\rho - \sigma\|_1<\varepsilon\}$ in mean, and enters it in finite time.
\end{theorem}
\begin{proof}
First notice that
$\min_{\sigma\in{\cal I}_S} \|\rho - \sigma\|_1=\min_{\sigma\in{\cal I}_S} \text{tr}(|\Pi_S(\rho - \sigma)|)+\text{tr}(\rho_R)\leq 2\text{tr}(\rho_R),$ with $\Pi_S$ the projector on the support of the target.
So $\text{tr}(\rho_R)\leq \varepsilon/2$ implies $\rho\in{\cal I}_{S,\varepsilon},$ and, since $V(\rho)<k_{max}\text{tr}(\rho_R),$ we have $\Omega_{\varepsilon/2k_{max}}\subseteq{\cal I}_{S,\varepsilon},$ where we define the sub-level set $\Omega_\ell=\{\rho|V(\rho)\leq \ell\}.$ So it is sufficient to prove that $\Omega_\beta=\Omega_{\varepsilon/2k_{max}}$ is stabilized by the average dynamics.
Consider $\alpha<\beta=\varepsilon/2k_{max}.$
By Assumption 2, and since $ \mathfrak{D}(\mathcal{H})$ is compact and $\dot V$ and $\ddot V$ are linear in $\rho$, there exists $\delta>0$ such that $\min_j V(\mathcal{L_j\rho)<-\delta$ for all $\rho\notin \Omega_\alpha,$ as well as two constants such that $\max_{\rho,j}|V(\mathcal{L_j\rho)|<M_1$ and $\max_{\rho,j}| V(\mathcal{L_j^2\rho)|<M_2.$ Then, if we take $t_{k+1}-t_k=\Delta t< \delta/M_2$ we are guaranteed that $\dot V$ remains negative for the whole switching interval. Since $\dot V(\rho)<-\delta+M_2\Delta t<0,\;\forall\rho\notin \Omega_\alpha,$ we have $V(t)\leq V(t_k)+(-\delta+M_2\Delta t)(t-t_k),$ and $\Omega_\alpha$ is reached in finite time during some interval $[t_\ell,t_{\ell+1})$, and so is $\Omega_\beta,$ and remain in there until at least $t_{\ell+1}.$ We next prove that if we start from $\Omega_\alpha$ we do not exit $\Omega_\beta,$ making it invariant. To this aim, it is sufficient to note that if $\rho_{t_{\ell+1}}\in\Omega_\alpha$, then for $t>t_{\ell+1}$:
$V(\rho_t)\leq \alpha + M_1\Delta t$. The right-hand side can be made arbitrarily small by reducing $\alpha$ and $\Delta t$, so it is always possible to ensure $\alpha + M_1\Delta t\leq \beta.$
\end{proof}
{\em Remark:} In this case, the lack of invariance of the target prevents from guaranteeing non-zero dwell times close to it; as a consequence, stability in mean is proved towards a set that does not have limited support, hence convergence a.s. cannot be established as before.}
\subsection{On robustness with respect to initialization errors}
\label{sec:initialization}
The knowledge of the initial condition plays a central role in determining the effectiveness of the control law, as already highlighted above.
Let suppose that the system has a true initial state $\rho_0$ but our best estimate on the initial state is $\rho_{0,e}$.
If we aim to stabilize a pure state, then by a simple majorization argument (see \cite{Scaramuzza-Ticozzi}) it is easy to show that convergence for the state-based strategy is guaranteed for any initial condition such that $supp(\rho_0)\subseteq supp(\rho_{0,e}),$ even if the switching generators are selected using the projected evolution of the wrong state. However, it is not straightforward to estimate how fast convergence is attained, or what would be the worst-case scenario.
When considering the SME, in the case of a non-exact estimate of the initial condition, the evolution of the true state is still described by model \eqref{eqn:model}. Then, by using the fact that the output signal $Y_t$ is independent of how we model the state, we obtain the evolution of the estimated state:
\begin{align}
d\rho_{t,e} &= \mathcal{L}_j(\rho_{t,e})dt + \mathcal{G}_C(\rho_{t,e})d\widetilde{W}_t,\\
d\widetilde{W}_t &=dW_t + \text{tr}[(C+C^\dag)(\rho_t - \rho_{t,e})]dt.
\label{eqn:filter_est}
\end{align}
It is important to highlight that, while $dW_t$ is a Wiener process, $d\widetilde{W}_t$ is not, as it presents a drift term. This complicates the study of the stability and convergence of the filter: some useful results are provided in \cite{VanHandel2009, Amini2011}.
{ In particular, for homodyne-detection SMEs as in our case, it is known that the filter is stable, and the measurement outcomes estimate converge, namely $\text{tr}[(C+C^\dag)(\rho_t - \rho_{t,e})]\rightarrow 0$ in mean. For extremal eigenvalues of $C$, as in our case, this directly implies that {\em attaining stabilization for the estimate also implies stabilization of the actual state, and vice-versa.} If the measurement effect is small with respect to the drift (e.g. we consider $\lambda C$ for a positive, sufficiently small $\lambda$) then the SME behavior is well approximated by the MME, and robustness is inherited whenever $supp(\rho_0)\subseteq supp(\rho_{0,e})$.
General convergence of the density operator is harder to obtain, and may not always be granted. When Assumption \ref{assumption} holds, the target is invariant and $\text{tr}(\hat\rho_{R,t})=\text{tr}(P_R\hat\rho_t )$ is decreasing, for both the actual and the estimated state. This implies that the SME dynamics are stable and the two evolutions do not diverge in mean.
It goes beyond the scope of this work to analyze what are the conditions under which the switching law stabilizes both. we tested initialization robustness numerically finding good results in all simulations. What is affected by faulty initialization, in general, is the convergence speed and we will exhibit some evidence in the next section.
Being the measurement-based control law a true {\em closed-loop} solution, we expect it to have a faster convergence than its open-loop counterpart, with the state estimate getting more accurate using the measurements and converging to the true state faster. }
\section{Application to multipartite entanglement generation: graph states}
\label{sec:case_study}
%
\begin{figure*}[h]\centering
\begin{subfigure}{.49\textwidth}\centering
\includegraphics[scale=0.55]{imgs/case1_1.pdf}
\caption{Results of simulation 1.}
\label{fig:results_1}
\end{subfigure}
\begin{subfigure}{.49\textwidth}\centering
\includegraphics[scale=0.55]{imgs/case2_1.pdf}
\caption{Results of simulation 2.}
\label{fig:results_2}
\end{subfigure}
\caption{Results of the two simulations scenarios described in paragraphs \ref{par:sim_1} and \ref{par:sim_2}. Only the trajectories of the true state of the system are here reported (not the estimated one).}
\label{fig:results}
\end{figure*}
%
In this section, we test the performance of the proposed control law in stabilizing an entangled state of a network of qubits (two-dimensional systems). We focus on the case of graph states: these are of practical interest \cite{graph} and offer a highly symmetric structure on which we can rely upon to design our control, which we briefly recall in the following.
Let us consider a graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ with $|\mathcal{V}|=n$. Each node of the graph $j\in\mathcal{V}$ is associated to a qubit:
$
\mathcal{H} = \bigotimes_{a=1}^n\mathcal{H}_a, \quad \textrm{dim}(\mathcal{H}_a)=2\quad\forall a.
$
Each edge $(i,j)\in\mathcal{E}$ describes the interaction between two adjacent qubits, this interaction is described by the operator $U_{(j,k)}$. We will assume that $U_{(j,k)}$ acts as the generalized controlled-Z gate acting on the qubits $j$ and $k$. Thus we have $U_{(j,k)}=C^Z_{(j,k)}\otimes I_{\overline{(j,k)}}$ with $C^Z = \text{diag}(1,1,1,-1)$.
The product of all such unitary matrices, $ U_G = \prod_{(j,k)\in\mathcal{E}}U_{(j,k)}$
can be seen as a global unitary, or in quantum computation terms a quantum circuit, which is used to map a factorized states with respect to the original qubit subsystems into entangled states, called {\em graph states}. A more comprehensive description of these states and their open-loop stabilization can be found in \cite{Ticozzi2012,Johnson2016}.
We consider a particular pure state as our target state but the procedure that follows can be adapted to any pure target state. Assume that we want to prepare the state $\rho_G = U_G \ket{+}\bra{+}^{\otimes n} U_G^\dag$.
To construct (locality constrained) MME generators that stabilize $\rho_G$ we simply construct the local generators which prepare each factor into $\rho_j=\ket{+}\bra{+}$ in the un-rotated basis, extend each generator to the whole system by tensor product with the identity and then transform these operators by $U_G$. The operators needed for local stabilization are thus of the form
$L_{G,j} = U_G^\dagger (\ket{+}\bra{-}_j\otimes I_{\bar{j}})U_G.$
We can then create a $\mathcal{D}_{L_j}(\rho)$ for each qubit, using these noise operators.
In \cite{Johnson2016} it is possible to find the proof of the fact that
$\mathcal{D}_c(\rho) = \sum_{j=1}^n\frac{1}{n}\mathcal{D}_{L_j}(\rho)$
makes $\rho_G$ GAS for the model $\frac{d}{dt}\rho_t = \mathcal{D}_c(\rho_t)$.
In order to find a Lyapunov function suitable to our aims we introduce the graph Hamiltonian:
$ H_G =-U_G^\dagger\Big(\sum_jX_j\Big)U_G$
where $X_j=\sigma_{x,j}\otimes I_{\bar{j}}$ and $\sigma_{x,j}$ represents the Pauli-$x$ matrix acting on the $j$-th qubit.
The target state $\rho_G$ may be seen as the unique ground state of the graph Hamiltonian $H_G$, i.e. $\rho_G$ is the only state such that $\text{tr}(H_G\rho_G)=\min\lambda_i$, where $\lambda_i$ are the eigenvalues of $H_G$.
This implies that $V(\rho) = \text{tr}(H_G\rho)$ is a valid linear Lyapunov function.
We can thus consider models \eqref{eqn:model} and \eqref{eqn:MME_model} with $\mathcal{L}_j(\rho) = \mathcal{D}_{L_j}(\rho)+\mathcal{D}_{H_G}(\rho)$ and $\mathcal{G}_{H_G}(\rho)$ where $L_j$ and $H_G$ are as defined above. The results of \cite{Johnson2016} imply that any convex combinations $\sum_j\alpha_j\mathcal{L}_j(\rho)$ makes GAS in mean, and thus satisfies Assumption \ref{assumption}.
\subsection{Numerical simulations}
We will now present two simulations realized on a 5-qubit system that show the advantages of the proposed method.
Both simulations will have the same graph (depicted in Figure \ref{fig:sys_graph}), the same target state $\rho_G = U_G \ket{+}\bra{+}^{\otimes n}U_G^\dag$, and the same true initial condition $\rho_0 = U_G \ket{--++-}\bra{--++-}U_G^\dag$. The only difference between the two simulations will be the estimated initial condition $\rho_{0,e}$.
\input{graph_figure}
In order to numerically compute the solution of the SME, we used the method proposed in \cite{Amini2011}, which we found the most reliable. Regarding the simulations of the cyclic and the state-based control laws, we simply used the Euler integration method.
The two simulations have been run with
step length $d_t=0.005,$ number of steps $N=1000,$ steps between switching $\epsilon=10,$ each with $1000$ realizations.
The results of the two simulations will be shown as graphs of the trace norm distance of the true state from the target state $\frac{1}{2}\norm{\rho_t-\rho_G}_1$ against time.
\subsubsection{Simulation 1}
\label{par:sim_1}
The first estimated initial state we consider is $\rho_{0,e} = 0.5[\rho_0^t+I_d/d]$. This introduces an additional ``uniform'' uncertainty on the correct initialization of the filter. From Figure \ref{fig:results_1} we can observe that the measurement-based trajectory converges on average with a speed that is similar to the optimal one when the true initial state is available to both strategies. Both do improve convergence with respect to cyclic switching, which does not depend of course on the initial condition estimate.
For the wrong initial condition case, the measurement-based trajectory improves the convergence with respect to the state-based one, yet mostly in the central phase of the evolution. The measurement-based trajectories shown in Figure \ref{fig:results} are the average obtained from 1000 realizations. The true trajectories would resemble the dashed ones in Figure \ref{fig:case_detail}. These results already highlight how the performance depends on the estimated state. This is further illustrated by the next set of simulations.
\begin{figure}[h]
\centering
\includegraphics[scale=0.55]{imgs/case1_2.pdf}
\caption{Evolution of the measurement-based trajectory in Simulation 1. In solid blue the average trajectory over 1000 realizations is shown, the light-blue area shows the average plus or minus one standard deviation while the dashed lines represent five typical realizations of the quantum trajectories.}
\label{fig:case_detail}
\end{figure}
\subsubsection{Simulation 2}
\label{par:sim_2}
The second estimated initial state we consider is $\rho_{0,e} = 0.5[\rho_0^t +U_G \ket{++--+}\bra{++--+}U_G^\dag]$. This second case has been designed to highlight the strength of the proposed strategy. In fact, the state is a mixture of the true state and a state which is not only orthogonal to the true one but is obtained by flipping single-qubit states, so the marginal states are also orthogonal. From the results shown in Figure \ref{fig:results_2} we can observe that this case is very different from the previous simulation. In particular, we can notice that correct initialization still provides the best convergence for both control laws. However, for the wrong initial condition case, the state-based trajectory has the same convergence rate as the cyclic trajectory, while the measurement based-trajectory provides a noticeable improvement, closer to the optimal performance. These results reinforce the intuition that closed-loop control, which can take advantage of the measurement outcome to obtain a better estimation of the state, can adapt in real-time and better avoid switching sequences that are not effective.
\printbibliography
\end{document} | af0da53928040b80a4b384ac7f9d3382f9935be7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\IEEEPARstart{W}{ith} the rapid development of high-definition displays, the demand for high resolution image/video content has been increasing rapidly. To improve the user-end visual experience, image Super-Resolution (SR) technique is developed to interpolate High-Resolution (HR) images from their Low-Resolution (LR) references. Examples of these interpolated HR images can be seen in Fig. \ref{fig:SR}. The past two decades have witnessed a booming of image SR algorithms with widespread applications including medical image processing, video surveillance, remote sensing, and face recognition, among many others. In the SR process, an Image Quality Assessment (IQA) metric is critical as both a performance indicator and a guidance for further improvement. However, the mainstream IQA methods, such as Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM) index \cite{Wang2004}, do not have high correlation with subjective opinions in image SR \cite{Wang2020}.
\begin{figure}[t]
\centering
\subfigure[BICUBIC\_2]{
\includegraphics[width=1.8cm]{fig/animal9/B2.jpg}
}
\subfigure[RLLR\_2]{
\includegraphics[width=1.8cm]{fig/animal9/R2.jpg}
}
\subfigure[SRCNN\_2]{
\includegraphics[width=1.8cm]{fig/animal9/S2.jpg}
}
\subfigure[VDSR\_2]{
\includegraphics[width=1.8cm]{fig/animal9/V2.jpg}
}
\subfigure[BICUBIC\_3]{
\includegraphics[width=1.8cm]{fig/animal9/B3.jpg}
}
\subfigure[RLLR\_3]{
\includegraphics[width=1.8cm]{fig/animal9/R3.jpg}
}
\subfigure[SRCNN\_3]{
\includegraphics[width=1.8cm]{fig/animal9/S3.jpg}
}
\subfigure[VDSR\_3]{
\includegraphics[width=1.8cm]{fig/animal9/V3.jpg}
}
\subfigure[BICUBIC\_4]{
\includegraphics[width=1.8cm]{fig/animal9/B4.jpg}
}
\subfigure[RLLR\_4]{
\includegraphics[width=1.8cm]{fig/animal9/R4.jpg}
}
\subfigure[SRCNN\_4]{
\includegraphics[width=1.8cm]{fig/animal9/S4.jpg}
}
\subfigure[VDSR\_4]{
\includegraphics[width=1.8cm]{fig/animal9/V4.jpg}
}
\caption{Examples of HR images generated by SR. (a-d), (e-h) and (i-l) are reconstructed HR images via BICUBIC \cite{Hou1979}, RLLR \cite{Liu2012}, SRCNN \cite{Dong2016} and VDSR \cite{Kim2016} algorithms with scaling factors of 2, 3 and 4 respectively. Quality variations are observed across both scaling factor and SR methods.}
\label{fig:SR}
\end{figure}
Existing IQA methods can be divided into two categories: subjective tests by human observers and objective models by automatic algorithms. Between them, the more reliable way is the subjective test, since human users are the ultimate viewers of images. Several SR methods assess their performances using small-scale subjective tests \cite{Mehdi2017, WangRecovering2018}. However subjective test is time-consuming and hard to be embedded into practical applications. Instead, it is often utilized to construct databases, which serve as standard testing sets to compare objective models.
Numerous objective IQA models have also been proposed. According to the availability of reference, objective models can be classified into full-reference, reduced-reference and no-reference IQA models. In a typical image SR problem, the perfect quality HR image is unavailable as a reference, therefore, full-reference models are generally inapplicable. In addition, SR algorithms often introduce mixed impairments, including blurring, ringing and aliasing artifacts, which are not well measured by the existing methods \cite{Ma2016,Tang2019}. Therefore, conventional general-purpose reduced-reference and no-reference IQA models should be redesigned specifically for image SR problem.
The key challenge in SR-IQA is how to effectively learn distinguishing feature representations of various LR and HR images and then map the features to image quality prediction. Convolutional Neural Network (CNN) has shown its advantage in IQA with promising successes in recent years \cite{Kang2014, Yang2019,Yan2019}. Compared with other data-driven models, CNN is capable of learning features and making quality predictors jointly in an end-to-end manner. Despite its superiority, CNN has not yet been well exploited in SR-IQA, for which most methods are not based on deep learning \cite{Chen2018, Yeganeh2015, Fang2016, Tang2019, Ma2016}. In this work, we propose a CNN-based SR-IQA algorithm. Considering the importance of reference information \cite{Bosse2018}, we employ CNNs to extract the features from both images before and after the SR process. A two-stream CNN architecture is thereby designed to simultaneously take the test HR and the corresponding LR reference images, followed by feature fusion and quality prediction of the HR image.
The performance of deep learning based models relies heavily on the quality and quantity of training images. Although several benchmark databases of image SR qualities have been constructed \cite{Yang2014, Yeganeh2015, Ma2016, Wang2017, Zhou2019}, they are limited in their sample sizes. The largest database contains only 1,620 HR images. A larger image SR database is thus imperative, for which the biggest challenge is in the enormous workload of human labeling. In this work, we observe a negatively exponential decay behavior of subjective scores after iterative downsample-and-SR processing of natural images. This helps us develop a large-scale database with reduced human labeling workload. Experiments on randomly selected samples demonstrate the high accuracy of this database. Based on this database, we are able to develop and train the two-stream deep network that predicts the quality of SR images with a high correlation to human scores.
Our major contributions are as follows:
\begin{enumerate}[1)]
\item Developed so far the largest IQA database for image SR with a novel semi-automatic labeling method, which greatly reduced the workload of human labeling.
\vspace{0.5mm}
\item Proposed a two-stream deep network that incorporates the available LR image into the quality evaluation of HR image. It results in an end-to-end model that jointly learns perceptually consistent features from the two images and a quality predictor.
\vspace{0.5mm}
\item Designed a feature fusion method that combines the two-stream deep CNN, leading to superior performance against state-of-the-art quality prediction.
\end{enumerate}
The remaining of this paper is organized as follows. Section \uppercase\expandafter{\romannumeral2} introduces the related work. Section \uppercase\expandafter{\romannumeral3} explains the methodology and process to construct the proposed large-scale database. Section \uppercase\expandafter{\romannumeral4} elaborates the proposed deep network, including network architecture, feature fusion method, and model training. Section \uppercase\expandafter{\romannumeral5} provides the experimental results. Finally, the paper is concluded in Section \uppercase\expandafter{\romannumeral6}.
\begin{figure*}[t]
\centering
\subfigure[ANIMALS: 0.9882]{
\begin{minipage}{0.19\linewidth}
\includegraphics[width=2.8cm]{fig/curve/B15_animal8_1.jpg}\vspace{4pt}
\includegraphics[width=2.85cm]{fig/curve/animal2.pdf}
\end{minipage}}\hspace{-7.5mm}
\subfigure[BUILDINGS: 0.9875]{
\begin{minipage}{0.19\linewidth}
\includegraphics[width=2.8cm]{fig/curve/S27_building2_1.jpg}\vspace{4pt}
\includegraphics[width=2.85cm]{fig/curve/building2.pdf}
\end{minipage}}\hspace{-7.5mm}
\subfigure[HUMANS: 0.9923]{
\begin{minipage}{0.19\linewidth}
\includegraphics[width=2.8cm]{fig/curve/R27_human16.jpg}\vspace{4pt}
\includegraphics[width=2.85cm]{fig/curve/human2.pdf}
\end{minipage}}\hspace{-7.5mm}
\subfigure[SPORTS: 0.9932]{
\begin{minipage}{0.19\linewidth}
\includegraphics[width=2.8cm]{fig/curve/V15_sports16_1.jpg}\vspace{4pt}
\includegraphics[width=2.85cm]{fig/curve/sports2.pdf}
\end{minipage}}\hspace{-7.5mm}
\subfigure[PLANTS: 0.9836]{
\begin{minipage}{0.19\linewidth}
\includegraphics[width=2.8cm]{fig/curve/S15_plant3_1.jpg}\vspace{4pt}
\includegraphics[width=2.85cm]{fig/curve/plant2.pdf}
\end{minipage}}\hspace{-7.5mm}
\subfigure[SCENERY: 0.9804]{
\begin{minipage}{0.19\linewidth}
\includegraphics[width=2.8cm]{fig/curve/V27_scenery4_1.jpg}\vspace{4pt}
\includegraphics[width=2.85cm]{fig/curve/scenery2.pdf}
\end{minipage}}\hspace{-7.5mm}
\centering
\caption{Sample images and the corresponding MOS versus DS-SR iteration testing points with exponential decaying fitting curves.}
\label{fig:curves}
\end{figure*}
\section{RELATED WORK}
Image SR has been an active research topic in recent years. Existing single image SR algorithms may be divided into three categories: interpolation-based \cite{Hou1979,Liu2012,Romano2014}, reconstruction-based \cite{Zhang2012,Vardan2016,Ren2017} and learning-based \cite{Dong2016,Kim2016,Lai2019,Ahn2018,Tian2020,Zhang2020}. Interpolation-based methods are simple, efficient and of low computational cost but has limited restoration performance, because such methods often introduce severe artifacts. Built upon models of prior domain knowledge, reconstruction-based methods suppress artifacts better, but often have much higher computational complexity. Learning-based methods learn LR-to-HR image mapping from training data.
During the past decades, numerous perceptual IQA metrics have been proposed to predict the visual quality of images with full-reference, reduced-reference and no-reference. Among them, the scope of application of full-reference metric is limited due to its requirement of unimpaired source. Recently, researchers have paid more attention to no-reference IQA. Early no-reference methods were based on Natural Scene Statistic (NSS) \cite{NSS2011}. Common NSS features include wavelet coefficients \cite{Moorthy2012}, locally normalized illumination coefficients \cite{Mittal2012}. Learning-based methods have been growing steadily \cite{Xu2016,Shao2016,Gu2018}, which automatically learn the mapping between image features and the perceptual quality.
Meanwhile, Image Sharpness Assessment (ISA) technique has emerged as an effective method for IQA \cite{HVS-MaxPool2019}. Hassen \textit{et al.} identified the Local Phase Coherence (LPC) of images computed in wavelet transform domain to assess image quality \cite{LPC-SI2013}. Bahrami \textit{et al.} proposed a fast no-reference ISA method based on the standard deviation of weighted Maximum Local Variation (MLV) distribution of images \cite{MLV2014}. Blanchet \textit{et al.} introduced an indicator of Global Phase Coherence (GPC), which decreases with blur, noise, and ringing \cite{GPC2015}. Li \textit{et al.} proposed a no-reference SPArse Representation-based Image SHarpness index (SPARISH), which used an overcomplete dictionary learned from natural images to measure the extent of blur \cite{SPARISH2016}. Hosseini \textit{et al.} designed two no-reference ISA metrics, called Synthetic-MaxPol \cite{Synthetic-MaxPool2018} and HVS-MaxPol \cite{HVS-MaxPool2019}, which were based on MaxPol convolution kernals and MaxPol filter library, respectively.
The use of deep learning has been a strong trend in recent IQA algorithms. Kang \textit{et al.} designed a shallow CNN to learn features from contrast normalized image patches \cite{Kang2014}. Ma \textit{et al.} proposed a multi-task end-to-end learning framework for no-reference IQA \cite{Ma2018}. Yang \textit{et al.} proposed an end-to-end SGDNet for no-reference IQA, which introduced saliency information to facilitate quality prediction \cite{Yang2019}. Yan \textit{et al.} integrated the NSS features prediction to the deep learning-based no-reference IQA to improve the representation and generalization ability \cite{Yan2019}.
However, most existing IQA methods are not suitable for SR images, since they are designed for images degraded by common distortions such as compression, white noise and blur. The distortions produced by SR are often mixed and more sophisticated. The study \cite{Wang2017} has shown that the popular no-reference IQA metrics are difficult to predict the perceptual quality of SR image based on an SR Image Database (SRID).
In a lab testing environment, the pristine reference HR image may be available when evaluating SR algorithms, for which objective full-reference metrics are directly applicable. It was shown that the Mean Squared Error (MSE) is a poor criterion in many cases \cite{Yang20191}. In \cite{Thapa2016}, Thapa \textit{et al.} adopted PSNR and SSIM to compare the performance of several SR algorithms. In SR benchmark study \cite{Yang2014}, Yang \textit{et al.} analyzed the effectiveness of six common full-reference IQA indicators, and also demonstrated that existing full-reference metrics fail to match the perceptual qualities of HR images well. Small-scale subjective tests have also been carried out. Reibman \textit{et al.} evaluated SR enhanced image quality through subjective tests, and pointed out that the full-reference IQA metrics cannot always capture visual quality of HR image \cite{Reibman2006}.
In most practical scenarios, the reference HR images are not available, thus reduced-reference and no-reference IQA algorithms are preferred. Yeganeh \textit{et al.} extracted three sets of NSS to assess the image SR performance with the available LR image as a reference, but the interpolation factors were strictly limited to integers \cite{Yeganeh2015}. Zhou \textit{et al.} proposed a Quality Assessment Database for Super-resolved images (QADS) and an IQA method considering the structural and textural components of images \cite{Zhou2019}. Chen \textit{et al.} presented a hybrid quality metric for non-integer image interpolation that combined both reduced-reference and no-reference philosophies \cite{Chen2018}. Fang \textit{et al.} introduced a reduced-reference quality assessment method for image SR by predicting the energy and texture similarity between LR and HR images \cite{Fang2016}. In \cite{Tang2019}, Tang \textit{et al.} proposed another reduced-reference IQA algorithm for SR reconstructed images with information gain and texture similarity combining saliency detection. Ma \textit{et al.} proposed a no-reference metric by supervised learning on a large set of reference-free HR images \cite{Ma2016}. In addtion, Fang \textit{et al.} designed a Blind model based on CNN for SR-IQA (BSRIQA) \cite{Fang2018}.
Despite the significant effort, existing SR-IQA methods are limited in three aspects. First, most of them are constrained to integer scaling factors. Second, the public image SR databases are limited in size, making it difficult to train SR-IQA models without encountering serious overfitting problems. Third, there is no deep learning based model for reduced-reference SR-IQA.
Focusing on the issues above, we firstly establish a large-scale image SR database, in which the scaling factor of HR images can be arbitrary. A reduced-reference CNN-based SR-IQA method is then proposed, which combines with LR images as references and is trained on the constructed large-scale database.
\section{SR Image Database Construction}
Learning-based IQA methods desire large-scale databases for training. The main challenge in building such databases is how to label a large number of images with quality ratings. Human subjective testing is desirable, but is time-consuming and expensive. Moreover, the fatigue effect when labeling large-scale datasets often affects the consistency and reliability of subjective ratings. Here we opt to a semi-automatic approach, aiming for largely reducing of workload of subjective testing.
\subsection{The Exponential Law in Image SR}
Presumably, when an image is going through a downsampling (DS) followed by SR interpolation process iteratively, the quality of the SR interpolated images decreases with iteration. To verify this effect, we apply the DS-SR operation iteratively on natural images of diverse content. We then conduct subjective tests to obtain Mean Opinion Score (MOS) values. Interestingly, we find that the MOS value decreases with iteration, approaximately following an exponentially decaying curve (with very few outliers), regardless of the image content or interpolation method, as shown in Fig. \ref{fig:curves}. Let $Q$ and $t$ denote the quality score and the number of SR iterations respectively, we have
\begin{equation}
Q(t) = e^{-bt}, t\geq0,
\label{eq:1}
\end{equation}
where $b$ is a positive constant depending on image content and SR method. We normalize the quality of pristine HR images to 1, thus all curves pass through a fixed point: $Q$ (0) = 1. Given $b$, the quality scores of a set of HR images can be quickly acquired. Therefore, this observation offers promises to significantly reduce the workload of image labeling.
Inspired by the exponential decaying relationship described above, we carry out a dedicated experimemt to investigate the feasibility of using the relationship for semi-automatical labeling, which may greatly reduce the workload of subjective ratings. Firstly, we randomly select 300 images of 30 natural scenes and perform iterative DS-SR on these images. Secondly, we utilize two approaches to obtain the subjective scores. In the first approach, 21 subjects are asked to score a subset of HR images. Using the exponential decay law of Eq. (\ref{eq:1}), we interpolate the MOS values of the remaining images. Detailed information of this approach is illustrated in the following Section \uppercase\expandafter{\romannumeral3}. \textit{C}. In the second approach, all images are scored by all subjects to obtain all MOS values. Thirdly, we compare the MOS values obtained by the above two approaches: semi-automatic rating and full subjective test.
The results are evaluated using Pearson Linear Correlation Coefficient (PLCC), Spearman Rank-order Correlation Coefficient (SRCC) and Kendall Rank Correlation Coefficient (KRCC), as shown in Table \ref{tab:curve_corr}, which indicates an encouragingly high correlation between the two approaches. To provide a reference point, we also compute the correlations between the scores given by each individual subject and MOS. The average across all subjects, as reported in Table \ref{tab:curve_corr}, is a meaningful indicator of the performance of an average subject. Clearly, the correlation between the semi-automatic rating and MOS is higher than an average subject, another strong evidence that helps justify the semi-aotumatic rating approach. Given that the semi-aotumatic approach can greatly reduce the workload of human subject labelings, it offers great promise to generate quality labels with the absence of excessive subjective testing. Furthermore, the proposed DISQ model (details given later) trained on the SISAR database created from the semi-automatic rating approach well generalizes to produce superior performance on other databases, which again validates the usefulness of the proposed semi-automatic rating method.
\begin{table}[h]
\centering
\caption{PLCC evaluation between Semi-Automatic Rating and Full Subjective Test}
\label{tab:curve_corr}
\setlength{\tabcolsep}{2mm}
\begin{tabular}{cccccc}
\toprule
Image Types &LR Images &HR Images &PLCC &SRCC &KRCC \\ \midrule
Animals &6 &60 &0.9776 &0.9716 &0.8667 \\
Buildings &6 &60 &0.9925 &0.9731 &0.8747 \\
Humans &6 &60 &0.9656 &0.9609 &0.8396 \\
Sports &6 &60 &0.9842 &0.9725 &0.8776 \\
Plants &3 &30 &0.9797 &0.9817 &0.9040 \\
Scenery &3 &30 &0.9731 &0.9791 &0.8915 \\ \midrule
\multicolumn{3}{c}{Average} &0.9825 &0.9719 &0.8619 \\ \midrule
\multicolumn{3}{c}{Average Subject Performance} &0.9510 &0.9314 &0.8184 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Source Images and SR Algorithms}
The first step to construct the database is to collect a batch of HR images generated by SR algorithms. In the semi-automatic rating approach, it is achieved by iterative DS-SR on source images. The source images are selected to cover diversified scenes including animals, buildings, humans, sports, plants and scenery. In total, 100 natural images of $1024\times 768$ resolution are selected. Then, four typical SR algorithms are employed to generate HR images, including two interpolation-based methods (BICUBIC \cite{Hou1979}, RLLR \cite{Liu2012}) and two learning-based methods (SRCNN \cite{Dong2016}, VDSR \cite{Kim2016}). In particular, SRCNN and VDSR are constrained to image SR with integer scaling factors. We integrate these two methods and the BICUBIC algorithm to achieve non-integer image SR. In order to obtain HR images with varying qualities, three scaling factors, 1.5, 2 and 2.7, are utilized in the iterative DS-SR process. Hence, the scaling factor 2 is supported by BICUBIC, RLLR, SRCNN and VDSR while the other two factors are supported by BICUBIC, RLLR, SRCNN+BICUBIC and VDSR+BICUBIC, as shown in the Table \ref{tab:database}.
\begin{table}[t]
\centering
\caption{Composition of SISAR Database}
\label{tab:database}
\setlength{\tabcolsep}{1mm}
\begin{tabular}{ccccc}
\toprule
Source Images & Algorithms & Factors & \begin{tabular}[c]{@{}c@{}}HR Images\\in a Batch\end{tabular} & \begin{tabular}[c]{@{}c@{}}Total of\\HR Images\end{tabular} \\ \midrule
100 & \begin{tabular}[c]{@{}c@{}}BICUBIC\\RLLR\end{tabular} & 1.5, 2, 2.7 & 8, 7, 6 & 4200 \\ \midrule
100 & \begin{tabular}[c]{@{}c@{}}SRCNN\\ VDSR\end{tabular} & 2 & 7 & 1400 \\ \midrule
100 & \begin{tabular}[c]{@{}c@{}}SRCNN+BICUBIC\\ VDSR+BICUBIC\end{tabular} & 1.5, 2.7 & 8, 6 & 2800\\ \bottomrule
\end{tabular}
\end{table}
Our experiment as demonstrated in Fig. \ref{fig:curves} suggests that the image quality drops to an unacceptable level after certain number of DS-SR iterations. To avoid this, we set the maximum DS-SR iterations to 8, 7, 6 for the scaling factors 1.5, 2, 2.7, respectively. As a result, there are 8,400 HR images generated by image SR. An example of HR images of the same source is presented in Fig. \ref{fig:example}, where the SR algorithm BICUBIC is used with a scaling factor of 2.
\begin{figure}[t]
\centering
\subfigure[MOS: 0.7769]
\includegraphics[width=2.5cm]{fig/human6/7769.jpg
}\vspace{-0.3mm}
\subfigure[MOS: 0.6036]
\includegraphics[width=2.5cm]{fig/human6/6036.jpg
}\vspace{-0.3mm}
\subfigure[MOS: 0.4290]
\includegraphics[width=2.5cm]{fig/human6/4290.jpg
}\vspace{-0.3mm}
\subfigure[MOS: 0.2831]{
\includegraphics[width=2.5cm]{fig/human6/2831.jpg
}\hspace{-1mm}
\subfigure[MOS: 0.1709]{
\includegraphics[width=2.5cm]{fig/human6/1709.jpg
}\hspace{-1mm}
\subfigure[MOS: 0.0958]{
\includegraphics[width=2.5cm]{fig/human6/1328.jpg}
}
\caption{Sample HR Images created by SR algorithms.}
\label{fig:example}
\end{figure}
\subsection{Image Quality Labeling}
We use the semi-automatic rating approach to label the quality scores of images in the SISAR database. According to Eq. (\ref{eq:1}), these scores can be calculated for given parameter $b$, which is dependent on image content and SR algorithm, and can be estimated through subjective test.
For each (source image, SR algorithm, scaling factor) combination, we extract one HR test image for subjective testing, 25 subjects, aged between 20 and 30, are asked to give scores between 0 and 10 to assess the image quality. The HR images for test are randomly shown on the same screen together with their corresponding original images, which are considered as pristine references with quality score of 10. The experiment procedure follows the Double Stimulus Continous Quality Scale (DSCQS) method defined in ITU R BT. 500-13 \cite{Recommendation2012}. After statistical analysis, 2 participants are identified to be outliers and the remaining 23 subjective scores are averaged to an interpreted Mean Opinion Score (iMOS) for each test image. We normalize the iMOS values to [0,1] to obtain a point ($k$, iMOS) on the corresponding exponential decaying curve, where $k$ is the number of SR iterations. Substituting the point into Eq. (\ref{eq:1}), we obtain the parameter $b$.
With the parameter $b$ and Eq. (\ref{eq:1}), we compute the iMOS values of all other HR images in the group. Eventually, this semi-automatic labeling process saves the workload of human subjective testing by a factor of approximately 7.
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{fig/histogram.eps}
\caption{Histogram of the iMOS Values of the SISAR database.}
\label{fig:histogram}
\end{figure}
\subsection{Summary of SISAR Database}
In total, the SISAR database contains 8,400 HR images, which are generated by three scaling factors with six types of SR algorithms that includes four SR algorithms (BICUBIC, RLLR, SRCNN and VDSR) and two combined SR operations (SRCCN+BICUBIC and VDSR+BICUBIC). The proposed database covers diverse image contents as shown in Table \ref{tab:img_num}, and the range of interpolated image resolution is from 864$\times$608 to 1054$\times$811. The histogram of the final iMOS values obtained by the proposed semi-automatic rating approach are shown in Fig. \ref{fig:histogram}, where it appears that the test HR images well cover the full range of quality levels.
\begin{table}[t]
\centering
\caption{Composition of SISAR database}
\label{tab:img_num}
\setlength{\tabcolsep}{1.2mm}
\begin{tabular}{@{}ccccccc@{}}
\toprule
& Animals & Buildings & Humans & Sports & Plants & Scenery \\ \midrule
Source Images & 20&20&20&20 &11 &9 \\
HR Images & 1680&1680&1680&1680 &924 &756 \\ \bottomrule
\end{tabular}
\end{table}
\section{SR Image Quality Assessment Model}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{fig/framework2.eps
\caption{Architecture of DISQ Framework.}
\label{fig:framework}
\end{figure*}
We propose an end-to-end reduced-reference method for SR-IQA. The proposed model, namely Deep Image Super-resolution Quality (DISQ) index, is a two-stream CNN to evaluate the perceptual quality of HR images with LR images as references. The network architecture is illustrated in Fig. \ref{fig:framework}. Firstly, an HR image and its corresponding LR image are fed into each stream CNN to extracte image features. Then, a feature fusion step is conducted in order to combine discriminative features for the quality regression. Finally, the fused features are regressed onto the quality score of HR image by fully connected layers. In the following subsections, the network structure, feature fusion and model training will be discussed in detail.
\subsection{Network Architecture}
The proposed DISQ is built on an end-to-end learning framework that estimates the quality score of a test HR image with the HR image and LR image as inputs. The role of LR image is to provide reference information that may help the quality assessment of HR image. Considering the different scaling factors, we first split them into small non-overlapping image patches and input to the feature extractors to generate feature maps representing the whole images. Inspired by the architecture of VGG-16 network \cite{Simonyan2014}, we design our CNN for SR-IQA with a similar philosophy.
The proposed network contains two convolutional modules, a feature fusion module and a fully connected module. The convolutional module is utilized as feature extractor to learn various features from inputs, and the fully connected module is used to derive global features to regress quality scores. The feature fusion step is the bridge to connect these modules.
In the convolutional modules, all convolutions have a kernel size of 3$\times$3 with altered kernels. Combining with max-pooling layers with a stride of 2, the network extracts local features while reducing the size of the feature maps. We employ Rectified Linear Unit (ReLU) as the nonlinear activation function. The first convolutional module, named as ${\rm CNN}_{\rm LR}$, is designed to learn features of LR images. The inputs are the patches of an LR image with the size of 32$\times$32. Another convolutional module, denoted as ${\rm CNN}_{\rm HR}$, is a feature extractor of HR images. We divide each HR image into 128$\times$128 patches as inputs to generate feature maps of a whole HR image. To facilitate feature fusion, the two CNNs contain 2 and 4 max-pooling layers correspondingly, and several convolutional layers to obtain features with similar shapes.
After feature fusion, the fused features are fed into the fully connected module to be regressed onto the perceptual scores. There are four fully connected layers with 2048, 1024, 256 and 1 neurons, respectively. We apply dropout into the first three fully connected layers with a probability of 0.5. By randomly masking out the neurons, dropout helps prevent overfitting. The last layer is a simple linear regression with a scalar output that predicts the quality score.
\subsection{Feature Fusion}
We adopt \begin{math} {\rm CNN}_{\rm LR}\end{math} and \begin{math} {\rm CNN}_{\rm HR}\end{math} as the feature extractors to produce the feature maps \begin{math} F_{\rm H} \end{math} and \begin{math} F_{\rm L} \end{math} directly from the input HR image patches \begin{math} I_{\rm p} \end{math} and LR image patches \begin{math} I_{\rm rp} \end{math}, respectively. According to the internal structure of two convolutional modules, we can calculate the shape of output feature maps.
\begin{equation}
\begin{split}
\textbf{Feature } &\textbf{Extraction:} \\
F_{\rm L} &= {\rm CNN}_{\rm LR}(I_{\rm rp}; \theta_1), {\rm shape} = (N_{\rm L}, 8, 8, 512), \\
F_{\rm H} &= {\rm CNN}_{\rm HR}(I_{\rm p}; \theta_2), {\rm shape} = (N_{\rm H}, 8, 8, 512).
\end{split}
\end{equation}
Here $\theta_1$ and $\theta_2$ indicate the parameters of CNNs. $N_{\rm H}$ and $N_{\rm L}$ denote the numbers of HR and LR image patches. The $F_{\rm H}$ and $F_{\rm L}$ are collections of image patch features.
In order to transform the patch-level features into the image-level features, we first perform a pooling operation before image feature fusion. Specifically, the feature map is pooled into one mean tensor, max tensor and min tensor in the first dimension. Then the three tensors are concatenated into a new feature set without any further modifications.
\begin{equation}
\begin{split}
&\textbf{Feature } \textbf{Pooling:} \\
&F_{\rm mean} = {\rm mean}(F; {\rm axis}=1), {\rm shape} = (1, 8, 8, 512), \\
&F_{\rm max} = {\rm max}(F; {\rm axis}=1), {\rm shape} = (1, 8, 8, 512),\\
&F_{\rm min} = {\rm min}(F; {\rm axis}=1), {\rm shape} = (1, 8, 8, 512),\\
&F_{\rm pool} = (F_{\rm mean}, F_{\rm max}, F_{\rm min}), {\rm shape} = (3, 8, 8, 512).
\end{split}
\end{equation}
After feature pooling, there are two pooled features $F_{\rm Hpool}$ and $F_{\rm Lpool}$ with identical structure created by $F_{\rm H}$ and $F_{\rm L}$, respectively. We fuse the extracted image feature maps $F_{\rm Hpool}$ and $F_{\rm Lpool}$ before inputting to the regression part of the network. In image SR, the HR images are reconstructed based on the global information of LR images, which renders the difference between HR and LR image features to a meaningful representation in the feature space. In order to calculate their distance, the feature fusion step follows \cite{Bosse2018}:
\begin{equation}
\textbf{Feature Fusion: } \quad F_{\rm fuse} = F_{\rm Hpool} - F_{\rm Lpool}.
\label{eq:4}
\end{equation}
Idealy, $F_{\rm Lpool}$ in Eq. (\ref{eq:4}) should be replaced by $F_{\rm Hpool}^0$, where $F_{\rm Hpool}^0$ denotes the pooled feature set of the perfect HR image. However, this perfect image is unfortunately unavailable. To investigate the impact of replacing $F_{\rm Hpool}^0$ with $F_{\rm Lpool}$, we carry out two empirical studies.
First, we perform test on a reference image and several corresponding LR images to calculate the difference of $F_{\rm Hpool}^0$ and $F_{\rm Lpool}$. The results are given in Table \ref{tab:distance}, which shows that the MSE of the two features are negligible relative to $F_{\rm Hpool}-F_{\rm Lpool}$.
Second, we compute the correlation between the norm of $ F_{\rm Hpool}-F_{\rm Lpool}$ and subjective scores of images. The results shown in Fig. \ref{fig:norm} suggest that the correlation is strong. These empirical studies lead us to adopt Eq. (\ref{eq:4}). Furthermore, the effectiveness
of this fusion method is also validated by experiments in the following Section \uppercase\expandafter{\romannumeral5}. \textit{C}.
\begin{table}[h]
\centering
\caption{MSE between $F_{\rm Hpool}^0$ and $F_{\rm Lpool}$, which is relatively small compare with the MSE between $F_{\rm Hpool}$ and $F_{\rm Lpool}$ (~$10^{-2}$)}
\label{tab:distance}
\setlength{\tabcolsep}{2.2mm}
\begin{tabular}{@{}ccccc@{}}
\toprule
& \begin{tabular}[c]{@{}c@{}}Downsample\\by 2\end{tabular} & \begin{tabular}[c]{@{}c@{}}Downsample\\by 2.5\end{tabular} & \begin{tabular}[c]{@{}c@{}}Downsample\\by 2.7\end{tabular} & \begin{tabular}[c]{@{}c@{}}Downsample\\by 3\end{tabular} \\ \midrule
MSE& 3.357E-03 & 3.363E-03 & 3.401E-03 & 3.380E-03 \\ \bottomrule
\end{tabular}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=9.2cm]{fig/norm2.eps
\caption{The Correlation between MOS and the Norm of $ F_{\rm Hpool}-F_{\rm Lpool}$.}
\label{fig:norm}
\end{figure}
\begin{figure*}[h]
\small
\centering
\includegraphics[width=18cm]{fig/Databases-7/SISAR7.jpg}\vspace{0.15cm}
\includegraphics[width=18cm]{fig/Databases-7/Waterloo7.jpg}\vspace{0.15cm}
\includegraphics[width=18cm]{fig/Databases-7/CVIU7.jpg}\vspace{0.15cm}
\includegraphics[width=18cm]{fig/Databases-7/QADS7.jpg}\vspace{0.15cm}
\centering
\caption{\small Typical HR images of SISAR (line 1), Waterloo-15 (line 2), CVIU-17 (line 3) and QADS (line 4).}
\label{fig:ref}
\vspace{0.15cm}
\end{figure*}
\subsection{Model Learning}
For an input HR image $I$ and a reference LR image $I_{\rm r}$, the proposed SR-IQA network $M$ is used to predict the perceptual quality of HR image $Q_{\rm pred}$:
\begin{equation}
Q_{\rm pred} = M((I_{\rm r}, I); \theta),
\end{equation}
where $\theta$ indicates all parameters of this network.
Denote the ground truth quality of the input as $Q_{\rm gt}$. The training goal of network $M$ is to find the optimal parameter setting, so as to minimize the overall quality prediction loss between $Q_{\rm pred}$ and $Q_{\rm gt}$ of all HR images in the training dataset. We apply the MSE as loss function in the training process, which is widely used in various regression tasks. Moreover, in order to avoid overfitting, $l$2-norm is added to the loss as a penalty term. Thus, the loss function is:
\begin{equation}
{\rm Loss} = \frac{1}{N} \sum_{i=0}^{N} \left\| Q_{\rm pred,i} - Q_{\rm gt,i} \right\| ^2 + \lambda \left\| \theta \right\| ^2,
\end{equation}
where the subscript $i$ of $Q_{\rm pred, i}$ and $Q_{\rm gt, i}$ represent the predicted quality and group truth of the $i$-th image, respectively. $\lambda$ is the regularization coefficient set to 0.0005 in our work.
The Adam optimizer \cite{2014Adam} is adopted to minimize the loss function. The main advantage of Adam is that after bias correction, the learning rate of each epoch has a definite range, which makes the parameters more stable. The learning rate is $\eta$=0.0001, and other parameters of Adam optimizer are of default settings.
\section{EXPERIMENTAL RESULTS}
In this section, we evaluate the performance of the proposed DISQ model and compare it with 16 other IQA metrics on four datasets. The cross-database test is also performed to assess the generalizibility of the algorithms.
\subsection{Experimental Setups}
\textbf{Databases}: We train the DISQ model on the proposed SISAR database.
To verify the generalizibility of the proposed method, we also employ three publicly available databases, Waterloo-15 \cite{Yeganeh2015}, CVIU-17 \cite{Ma2016} and QADS \cite{Zhou2019}, for cross-database validations. Typical images of these databases are listed in Fig. \ref{fig:ref}.
\textbf{SISAR} contains 8,400 HR images. We divide it into training and test sets with a 80/20 split and no content overlapped.
\textbf{Waterloo-15} is an interpolation image database containing 312 interpolated HR images and the corresponding LR images from 13 source images \cite{Yeganeh2015}. The HR images were created by eight interpolation algorithms combined with scaling factors of 2, 4 and 8.
\textbf{CVIU-17} is a collection of 180 LR images and 1,620 HR images, which were generated by nine SR methods and six integer scaling factors. There is another available SR image database, ECCV-14 \cite{Yang2014}, which contains 540 HR images. It should be noted that ECCV-14 is a subset of CVIU-17. Therefore, we only provide the experimental result on CVIU-17.
\textbf{QADS} is a super-resolved image database, which created 980 HR images using 21 image SR methods from 20 reference images.
The score ranges and types are not unified in these database, we choose the settings of SISAR as our standard in these experiments. Subjective scores on the other three databases are linearly scaled to the range of [0,1].
\textbf{Evaluation}: We use three common measurements to evaluate the performance of our algorithm by calculating the correlation between the subjective and objective quality scores: PLCC, SRCC and KRCC. PLCC is used to measure the accuracy of IQA algorithms. SRCC and KRCC are used to evaluate the monotonicity of quality predictions. For these metrics, a higher value up to 1 indicates better performance of a specific IQA method. The results in this work are the average values obtained by calculating the correlation coefficients based on different image contents.
\subsection{Performance Comparison}
\begin{table*}[t]
\centering
\caption{Performance Comparison of IQA Methods: The datasets used for training the test algorithms are denoted by ‘—’.}
\label{tab:comparison}
\begin{tabular}{@{}lccc|ccc|ccc|ccc@{}}
\toprule
& \multicolumn{3}{c}{Waterloo-15} & \multicolumn{3}{c}{CVIU-17} &\multicolumn{3}{c}{QADS} &\multicolumn{3}{c}{SISAR} \\ \cmidrule(l){2-13}
\multirow{-2}{*}{Methods} & PLCC & SRCC & KRCC & PLCC & SRCC & KRCC & PLCC & SRCC & KRCC & PLCC & SRCC & KRCC \\ \midrule
DIIVIVE & 0.6202 & 0.5827 & 0.4225 & 0.4602 & 0.4580 & 0.3310 &0.4404 &0.4654 &0.3184 &0.3349 &0.3201 &0.2576 \\
BRISQUE & 0.7527 & 0.7626 & 0.5682 & 0.5025 & 0.4644 & 0.3557 &0.5242 &0.5461 &0.3830 &0.3408 &0.3487 &0.2790 \\
CNN-IQA &{0.6659} &{0.6579} &{0.4856} & {0.2933} & {0.2894} & {0.2141} &{0.4037} &{0.3734} &{0.2694} &{0.2345} &{0.2752}&{0.1910} \\
HOSA & 0.8287 & {\color[HTML]{CB0000} \textbf{0.8236}} & {\color[HTML]{CB0000} \textbf{0.6375}} & {\color[HTML]{3166FF} \textbf{0.5927}} & 0.5450 & 0.4169 &0.6343 &0.6409 &0.4598 &0.3299 &0.3180 &0.2629 \\
LPC-SI & 0.8017 & 0.4689 & 0.4188 & 0.4547 & 0.4164 & 0.3294 &0.5027 &0.4902 &0.3358 &0.8562 &0.8076 &0.6267 \\
MLV & 0.7313 & 0.3508 & 0.2903 & 0.5655 & 0.4628 & 0.3675 &0.2471 &0.2456 &0.1644 &0.7085 &0.6399 &0.5631 \\
GPC & {\color[HTML]{3166FF} \textbf{0.8310}} & 0.5058 & 0.4289 &0.5878 &{\color[HTML]{3166FF} \textbf{0.5576}} &{\color[HTML]{3166FF} \textbf{0.4549}} &0.4002 &0.4470 &0.3116 &0.7057 &0.8755 &0.7221 \\
SPARISH & 0.6988 & 0.6584 & 0.4924 & 0.4711 & 0.4390 & 0.3430 &0.5849 &0.6530 &0.4713 &0.8750 &0.8933 &{\color[HTML]{3166FF} \textbf{0.7373}} \\
Synthetic-MaxPol &0.5909 &0.2938 &0.2249 &0.4920 &0.4537 &0.3536 &0.4057 &0.3845 &0.2645 &0.4607 &0.3305 &0.2867 \\
HVS-MaxPol &0.7905 &0.7272 &0.5401 &0.5448 &0.5557 &0.4343 &0.5918 &0.5876 &0.4098 &0.3880 &0.3151 &0.2788 \\
FQPath &0.8038 &0.7544 &0.5672 &0.5122 &0.5111 &0.3884 &0.4879 &0.4750 &0.3232 &0.3200 &0.3429 &0.3127 \\
FocusLiteNN &{\color[HTML]{CB0000} \textbf{0.8368}} &0.5359 &0.4441 &0.5338 &0.4464 &0.3361 &0.3775 &0.3721 &0.2533 &{\color[HTML]{3166FF} \textbf{0.9123}} &{\color[HTML]{3166FF} \textbf{0.9016}} &0.7361 \\
NSS-SR & {—} & — & {—} & {0.5504} & {0.4993} & {0.3616} & {0.3670} & {0.2160} & {0.1424} & {0.3706} & {0.3280} & {0.2635} \\
HYQM & 0.4108 & 0.2096 & 0.2004 &0.2225 & 0.1125 & 0.0941 &0.4506 & 0.4443 & 0.2973 & 0.3660 & 0.2732 & 0.2112 \\
LNQM & 0.7488 &0.7022 & 0.4974 & — & — & — &{\color[HTML]{3166FF} \textbf{0.7220}} & {\color[HTML]{3166FF} \textbf{0.7274}} & {\color[HTML]{3166FF} \textbf{0.5394}} & 0.5192 &0.4944 &0.3449 \\
BSRIQA &0.6991 &0.6736 &0.4689 &— &— &— &0.5863 &0.5841 &0.4173 & 0.6906 &0.4341 &0.3099 \\ \midrule
DISQ & 0.8232 & {\color[HTML]{3166FF} \textbf{0.7897}} & {\color[HTML]{3166FF} \textbf{0.5770}} & {\color[HTML]{CB0000} \textbf{0.6440}} & {\color[HTML]{CB0000} \textbf{0.7071}} & {\color[HTML]{CB0000} \textbf{0.5280}} &{\color[HTML]{CB0000} \textbf{0.8173}} &{\color[HTML]{CB0000} \textbf{0.7960}} &{\color[HTML]{CB0000} \textbf{0.6302}} & {\color[HTML]{CB0000} \textbf{0.9143}} & {\color[HTML]{CB0000} \textbf{0.9124}} & {\color[HTML]{CB0000} \textbf{0.7742}} \\ \bottomrule
\end{tabular}
\end{table*}
The performance evaluation results of the proposed DISQ model are listed in Table \ref{tab:comparison} and compared with other IQA methods, where the best and second-best results are shown in {\color[HTML]{CB0000} \textbf{red}} and {\color[HTML]{3166FF} \textbf{blue}}, respectively. The compared algorithms include four no-reference IQA methods (DIIVIVE \cite{Moorthy2012}, BRISQUE \cite{Mittal2012}, HOSA \cite{Xu2016}, CNN-IQA \cite{Kang2014}) designed for general distorted images, six no-reference ISA indicators (LPC-SI \cite{LPC-SI2013}, MLV \cite{MLV2014}, GPC \cite{GPC2015}, SPARISH \cite{SPARISH2016}, Synthetic-MaxPol \cite{Synthetic-MaxPool2018}, HVS-MaxPol \cite{HVS-MaxPool2019}), two focus quality assessment algorithms (FQPath \cite{FQPath2020}, FocusLiteNN \cite{wang2020focuslitenn}), and four related SR-IQA works (NSS-SR \cite{Yeganeh2015}, HYQM \cite{Chen2018}, LNQM \cite{Ma2016}, BSRIQA \cite{Fang2018}).
The source codes of these metrics are obtained from the authors’ public websites. There are two other SR-IQA metrics \cite{ Fang2016, Tang2019} without publicly available codes and parameters, thus they are not compared in the table.
From Table \ref{tab:comparison}, we have the following observations:
First, FocusLiteNN obtains the highest PLCC on the Waterloo-15 database, and the HOSA achieves the best performance on SRCC and KRCC. In this database, all images were interpolated with integers, which implies that the pixels at LR image will be reused in HR image. An SR-IQA algorithm can utilize this feature to enhance their evaluation performance. However, the proposed DISQ algorithm is designed for general purpose SR without prior knowledge of scaling factors. Nevertheless, its performance is still close to the best. In particular, DISQ ranks the second in terms of SRCC and KRCC.
Second, DISQ achieves a good performance on the CVIU-17 database, while other metrics present moderate performances. CVIU-17 was created by several popular image SR algorithms, in which HR images are not limited to the common distortions. Therefore, the features modeled in common IQA methods cannot cover diverse visual contents. Although NSS-SR and NYQM are designed for image SR, they do not work well on this database, because the hand-crafted features of these methods are built more specifically for image interpolation.
Third, most non-SR-IQA methods are less effective on the QADS database. DISQ and LNQM, both learning-based, show excellent performance in comparison with other methods. Between them, DISQ utilizes the large-scale database and deep learning to effectively extract intrinsic features, and shows much better performance. The moderate performance of the CNN-based BSRIQA may be attributed to the small training database.
Fourth, it can be clearly observed that DISQ outperforms other metrics on the SISAR database. FocusLiteNN achieves excellent performance, and most no-reference ISA algorithms show a good correlation with SISAR database. The reason is that the methods perform better on the local focused images in SISAR database than other metrics. However, their performance drops significantly on the other three databases. Among the SR-IQA methods, although LNQM and BSRIQA obtains better performance than NSS-SR and NYQM, they are not among the best performers. Both of them are trained on CVIU-17, in which the HR images were generated only by integer scaling factors. By contrast, SISAR contains HR images produced by non-integer scaling factors, which benefits the DISQ model that uses SISAR for training.
To further investigate model generalization ability, Table \ref{tab:average} reports the average performances of all algorithms under comparison on different databases. The performance of the algorithms above the dividing line are the average of four databases, and the rest are average of three, except for the database used for training the corresponding algorithm. As can be observed, the DISQ achieves the best average performance in cross-database experiments, and the PLCC and SRCC values are nearly 10\% higher than the second-best model. The results demonstrate the robustness of the proposed DISQ model.
\begin{table}[h]
\centering
\caption{Average Performance in Cross-Database Comparison}
\label{tab:average}
\setlength{\tabcolsep}{5mm}
\begin{tabular}{lccc}
\toprule
Methods & PLCC &SRCC &KRCC \\ \midrule
DIIVIVE & 0.4639 & 0.4566 &0.3324 \\
BRISQUE & 0.5301 & 0.5305 &0.3965 \\
CNN-IQA & 0.3994 & 0.4040 &0.2900 \\
HOSA & 0.5964 & 0.5819 &0.4443 \\
LPC-SI & 0.6538 & 0.5458 &0.4277 \\
MLV & 0.5631 & 0.4248 &0.3326 \\
GPC & 0.6312 & 0.5965 &0.4794 \\
SPARISH & 0.6575 & {\color[HTML]{3166FF} \textbf{0.6609}} &{\color[HTML]{3166FF} \textbf{0.5110}} \\
Synthetic-MaxPol & 0.4873 & 0.3656 &0.2824 \\
HVS-MaxPol & 0.5788 & 0.5464 &0.4157 \\
FQPath & 0.5310 & 0.5209 &0.3979 \\
FocusLiteNN & {\color[HTML]{3166FF} \textbf{0.6651}} & 0.5640 &0.4424 \\
HYQM & 0.3625 & 0.2599 &0.2008 \\ \midrule
NSS-SR & 0.4293 & 0.3478 &0.2558 \\
LNQM & 0.6633 & 0.6413 &0.4606 \\
BSRIQA & 0.6587 & 0.5639 &0.3987 \\
DISQ &{\color[HTML]{CB0000} \textbf{0.7615}} & {\color[HTML]{CB0000} \textbf{0.7643}} &{\color[HTML]{CB0000} \textbf{0.5784}}\\ \bottomrule
\end{tabular}
\end{table}
\subsection{Ablation Study}
To evaluate the contribution of each component in the proposed DISQ method, we conduct a series of ablation experiments. The ablation results are presented in Table \ref{tab:reference} to \ref{tab:fusion}. In each column, the best results are highlighted in {\color[HTML]{CB0000} \textbf{red}}.
\subsubsection{Influence of the references LR images} In this experiment, we exclude the reference LR image in our model, resulting in a network with ${\rm CNN}_{\rm HR}$ and fully connected modules only. This no-reference model is trained on SISAR under the same parameter settings and training steps as the proposed reduced-reference DISQ model. Its testing performance on different databases is listed in Table \ref{tab:reference}. Although the performance of the no-reference model is slightly better on the QADS database, in all other three databases, the reduced-reference model performs much better, suggesting significant benefits of including the LR image as reference.
\begin{table}[h]
\centering
\caption{PLCC Performance Comparison with or without LR image as Reference}
\label{tab:reference}
\setlength{\tabcolsep}{3.5mm}
\begin{tabular}{lcccc}
\toprule
Methods & SISAR & Waterloo-15 & CVIU-17 &QADS\\ \midrule
w/o &0.8610 &0.6207 &0.5024 &{\color[HTML]{CB0000} \textbf{0.8403}} \\
w/ &{\color[HTML]{CB0000} \textbf{0.9143}} &{\color[HTML]{CB0000} \textbf{0.8232}} &{\color[HTML]{CB0000} \textbf{0.6440}} &0.8173 \\ \bottomrule
\end{tabular}
\end{table}
It is also interesting to observe that the performance of this no-reference network is superior to many other algorithms shown in Table \ref{tab:comparison}. Experiments also reveal that further training of this no-reference model may increase its PLCC performance on SISAR but with a lower generalizibility to other databases. This suggests that including the LR image reference may help reduce overfitting.
\subsubsection{Selection of feature map pooling method} In the proposed network, the feature map is pooled into a concatenated tensor, which is described in Section \uppercase\expandafter{\romannumeral4}. \textit{B}. Previous studies \cite{Kang2014,Fang2018} have proved that the concatenated pooling feature has certain advantages compared with common mean pooling or max pooling. To intuitively illustrate the effectiveness of the joint features, we report the performance comparison of different pooling methods under identical training epochs in Table \ref{tab:pooling}, where the joint pooling method significantly improves the accuracy of quality prediction, especially on the CVIU-17 database. The merged feature map possesses rich and robust image features, which contributes to the mapping from image features to quality.
\begin{table}[h]
\centering
\caption{PLCC Comparison of different feature pooling methods}
\label{tab:pooling}
\setlength{\tabcolsep}{3mm}
\begin{tabular}{lcccc}
\toprule
Methods & SISAR & Waterloo-15 & CVIU-17 & QADS\\ \midrule
Max Pooling &0.8803 &0.6593 &0.4512 &0.6714 \\
Mean Pooling &0.8838 &0.7583 &0.4590 &0.7712\\
Joint Pooling &{\color[HTML]{CB0000} \textbf{0.9143}} &{\color[HTML]{CB0000} \textbf{0.8232}} &{\color[HTML]{CB0000} \textbf{0.6440}} &{\color[HTML]{CB0000} \textbf{0.8173}}\\ \bottomrule
\end{tabular}
\end{table}
\subsubsection{Effectiveness of feature fusion methods} We incorporate the LR and HR image features in two other methods, which were also discussed in \cite{Bosse2018}. In summary, the image features $F_{\rm Hpool}$ and $F_{\rm Lpool}$ are merged in the following three methods
\begin{equation*}
\begin{split}
&\textbf{Method 1: } F_{\rm fuse} = F_{\rm Hpool} - F_{\rm Lpool}, \\
&\textbf{Method 2: } F_{\rm fuse} = (F_{\rm Hpool}, F_{\rm Lpool}), \\
&\textbf{Method 3: } F_{\rm fuse} = (F_{\rm Hpool}, F_{\rm Lpool}, F_{\rm Hpool} - F_{\rm Lpool}). \\
\end{split}
\end{equation*}
Under the same training steps, the performance of our model combined with different feature fusion methods are provided in Table \ref{tab:fusion}. Method 1 exhibits the optimal performance with the fewest parameters on most databases. The network that merges $F_{\rm Hpool}$ and $F_{\rm Lpool}$ in Method 2 is theoretically capable of learning \begin{math} F_{\rm Hpool}-F_{\rm Lpool} \end{math}
in the regression part. However, the performance is worse than Method 1, except for the QADS database. Method 3 unites the first two but fails to further improve the quality prediction accuracy even with an increasing number of paramaters. With better or similar PLCC performance, Method 1 uses only 1/2 and 1/3 of the numbers of parameters when compared with Methods 2 and 3, respectively.
\begin{table}[h]
\centering
\caption{PLCC of Models Using Different Feature Fusion Methods}
\label{tab:fusion}
\setlength{\tabcolsep}{2mm}
\begin{tabular}{ccccc}
\toprule
\begin{tabular}[c]{@{}c@{}}Fusion Methods\\\& Shape of $F_{\rm fuse}$\end{tabular} & SISAR & Waterloo-15 & CVIU-17 &QADS\\ \midrule
Method 1: (3, 8, 8, 512) &{\color[HTML]{CB0000} \textbf{0.9143}} &{\color[HTML]{CB0000} \textbf{0.8232}} &{\color[HTML]{CB0000} \textbf{0.6440}} &0.8173\\
Method 2: (6, 8, 8, 512) & 0.8938 & 0.7678 & 0.5626 &{\color[HTML]{CB0000} \textbf{0.8474}}\\
Method 3: (9, 8, 8, 512) & 0.8973 & 0.7306 & 0.5214 &0.6897\\ \bottomrule
\end{tabular}
\end{table}
\section{CONCLUSIONS}
We build the largest of its kind SISAR database for SR-IQA, where a novel semi-automatic rating approach is proposed, which greatly reduces the workload for human labeling. We train an end-to-end DISQ model for SR-IQA which uses a two-stream DNN for feature extraction, followed by a feature fusion network for quality prediction. We train the DISQ model using the SISAR database, and the experimental results demonstrate the superior performance of the DISQ model against state-of-the-art. Cross-database validation also shows better generalizibility of the proposed DISQ model. The SISAR database and the DISQ model will be made public available to facilitate reproducible research.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\balance
\bibliographystyle{IEEEtran}
| b4592263fbd4709bf30fd1faa7aad2a423711ef3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Replicating human movements and behaviour in humanoid robots is a formidable challenge with many exciting applications, ranging from health care (e.g. artificial limbs \cite{alshamsi2016}) to space exploration \cite{nasa2015}. Since the manual engineering of controllers for such tasks is extremely difficult, machine learning and specifically reinforcement learning has received much attention in this area. A recent NIPS competition \cite{kidzinski2018learningtorun} focused on the creation of a simulated humanoid running robot in a continuous and high-dimensional environment. The top entries to the competition employed state-of-the-art deep reinforcement learning techniques \cite{jaskowski2018rltorunfast,kidzinski2018l2rsolutions}, which resulted in strong, but not optimal, performance.
In this paper, we demonstrate that videos showing human running behaviour can be used to significantly improve the learning performance. In our approach, we use the coordinates of specific body parts (e.g. the foot) to define a potential function that is fed into potential-based reward shaping \cite{ng1999policy}.
To create a strong baseline for the evaluation of our approach, we combined selected RL techniques of the top ten competition entries with further optimizations to create a running agent that displays a significantly faster learning rate than the top entry. We then add the reward shaping from video data sampled from various YouTube videos, which resulted in a running agent that reached twice the running speed as our baseline in 12 hours of training.
Since potential-based reward shaping has the nice theoretical property of not changing the optimal policy \cite{ng1999policy}, data taken from sub-optimal running behaviour does not prevent the RL agent from overcoming the sub-optimalities and produce humanoid running that outperforms the data source. We demonstrate this theoretical property empirically by sampling limb positions from a slower-running agent and show how our approach generates a running robot, that after a relatively short time of training, starts to run faster than the original agent.
Overall, the main contribution of our work is to demonstrate how data extracted from videos of human movements can be used to significantly speed up the reinforcement learning of humanoid robots. While our work focuses on the training of humanoid running behaviour, the proposed techniques can easily be applied to any other form of humanoid movements.
\section{Background}
\subsection{Reinforcement Learning}
Reinforcement learning is a paradigm which allows agents to learn by reward and punishment from interactions with the environment \cite{sutton1984temporal}. The numeric feedback received from the environment is used to improve the agent's actions. The majority of work in the area of reinforcement learning applies a Markov Decision Process (MDP) as a mathematical model \cite{puterman2014markov}.
An MDP is a tuple $\big(S, A, T, R)$, where $S$ is the state space, A is the action space, $T(s,a,s') = Pr(s'|s,a)$ is the probability that action a in state s will lead to state $s'$, and $R(s, a, s')$ is the immediate reward $r$ received when action $a$ taken in state $s$ results in a transition to state $s'$. The problem of solving an MDP is to find a policy (i.e., mapping from states to actions) which maximises the accumulated reward. When the environment dynamics (transition probabilities and reward function) are available, this task can be solved using policy iteration \cite{bertsekas1995dynamic}.
When the environment dynamics are not available, as with most real problem domains, policy iteration cannot be used. However, the concept of an iterative approach remains the backbone of the majority of reinforcement learning algorithms. These algorithms apply so called temporal-difference updates to propagate information about values of states and/or state-action pairs, $Q(s, a)$ [20]. These updates are based on the difference of the two temporally different estimates of a particular state or state-action value. The Q-learning algorithm is such a method [21]. After each transition, $(s, a) \rightarrow (s', r)$, in the environment, it updates state-action values by the formula:
\begin{equation}
\label{eq:qlearn}
Q(s,a) \leftarrow Q(s,a) + \alpha[r + \gamma\max Q(s',a') - Q(s,a)]
\end{equation}
where $\alpha$ is the rate of learning and $\gamma$ is the discount factor. It modifies the value of taking action $a$ in state $s$, when after executing this action the environment returned reward $r$, and moved to a new state $s'$.
\subsection{Potential Based reward shaping}
The idea of reward shaping is to provide an additional reward representative of prior knowledge to reduce the number of suboptimal actions made and so reduce the time needed to learn \cite{ng1999policy, randlov1998learning}. This concept can be represented by the following formula for the Q-learning algorithm:
\begin{equation} \label{eq:shaping}
Q(s,a) \leftarrow Q(s,a)+\alpha[r+F(s,s')+\gamma\max Q(s',a') - Q(s,a)]
\end{equation}
where $F(s,s')$ is the general form of any state-based shaping reward.
Even though reward shaping has been powerful in many experiments it quickly became apparent that, when used improperly, it can change the optimal policy \cite{randlov1998learning}. To deal with such problems, potential-based reward shaping was proposed \cite{ng1999policy} as the difference of some potential function $\Phi$ defined over a source s and a destination state $s':F(s,s')=\gamma \Phi(s') - \Phi(s)$ where $\gamma$ must be the same discount factor as used in the agent's update rule (see Equation \ref{eq:qlearn}).
Ng et al. \cite{ng1999policy} proved that potential-based reward shaping, defined according to Equation \ref{eq:shaping}, guarantees learning a policy which is equivalent to the one learned without reward shaping in both infinite and finite horizon MDPs.
Wiewiora \cite{wiewiora2003potential} later proved that an agent learning with potential-based reward shaping and no knowledge-based Q-table initialization will behave identically to an agent without reward shaping when the latter agent's value function is initialized with the same potential function.
These proofs, and all subsequent proofs regarding potential-based reward shaping including those presented in this paper, require actions to be selected by an advantage-based policy \cite{wiewiora2003potential}. Advantage-based policies select actions based on their relative differences in value and not their exact value. Common examples include greedy, $\epsilon$-greedy and Boltzmann softmax.
\subsection{Deep-RL}
In Deep RL, the Q value function is represented as a multi-layer neural network \cite{Goodfellow-et-al-2016}. Deep RL algorithms have been shown to perform strongly on RL tasks which have been infeasible to tackle before. Over recent years, a number of algorithms and optimizations have been proposed, and we have chosen to apply
the Deep Deterministic Policy Gradient (DDPG) algorithm for our application domain \cite{lillicrap2015continuous}. DDPG has been shown to be effective in continuous action domains where classic reinforcement learning methods struggled.
Specifically, in the DDPG algorithm two neural networks are used:
$\mu(S)$ is a network (the {\it actor}) that returns the action vector whose components are the values of the corresponding control signals.
$Q^w(s, a)$ is a second neural network (the {\it critic}, that returns the $Q$ value, i.e. the value estimate of the action of $a$ in state $s$.
\begin{equation} \label{eq:policy5}
\begin{aligned}
\nabla_\theta J(\pi_\theta) & =\int_{S}\rho^\pi(s)\int_{A}\nabla_\theta\pi_\theta(a|s)Q^w(s,a) da ds \\
&=\mathbb{E}_{s\sim \rho^\pi,a\sim\pi_\theta}[\nabla_\theta \log \pi_\theta (a|s)Q^w(s,a)]
\end{aligned}
\end{equation}
where $\theta$ is the parameter vector of the probabilistic policy and $\rho^\pi(s)$ is the probability of reaching state $s$ with policy $\pi$.
For a more complete description of DDPG, see \cite{lillicrap2015continuous}.
\subsection{Reinforcement Learning from Demonstration}
Human expert demonstrations have been demonstrated to improve the learning speed and accuracy of RL agent on a wide range of tasks. Most work in this area (e.g. \cite{suay2016learning, brys2015reinforcement}) focused on the use of state-action recordings as demonstration. This is infeasible in the case of video data, where only state information is available and the demonstration actions are not explicitly provided and often can not be derived either (as is the case with running). More recently, various methods for state-only demonstrations have been proposed (e.g. \cite{peng2018deepmimic, liu2017imitation}). However, all of these methods target the imitation of demonstrations. Our work is employing potential-based reward shaping which uses the demonstrations to speed up the learning, but is also able to overcome any sub-optimalities in the demonstration rather than purely imitating them.
\section{Simulation Environment}
\label{sec:domain}
\begin{figure}
\centering
\includegraphics[width=0.40\textwidth]{img/opensim3.jpg}
\caption{A screenshot of the "Learning to Run" simulation environment}
\label{img:environment}
\end{figure}
The simulation environment has been provided by the "Learning to Run" competition and is based on the OpenSim environment employing the Simbody physics engine. The environment simulates a three-dimensional race course with small obstacles, along which a humanoid robot with 6 joints (ankle, knee, and hip on two legs) and corresponding muscles is running (see Figure \ref{img:environment}). The actions of the running humanoid robot are excitation values applied to the muscles implemented in the robot model. The next state of the environment is computed by the physics engine based on the resulting muscle activations, forces, velocities and positions of the joints.
The OpenSim \cite{kidzinski2018learningtorun} model environment represents the robot state using a vector of 41 features:
\begin{itemize}
\item position of the pelvis (rotation, x, y)
\item velocity of the pelvis (rotation, x, y)
\item rotation of each ankle, knee and hip (6 values)
\item angular velocity of each ankle, knee and hip (6 values)
\item position of the center of mass (2 values)
\item velocity of the center of mass (2 values)
\item positions of head, pelvis, torso, left and right toes, left and right talus (14 values)
\item strength of left and right psoas (a muscle at the lower spine)
\item next obstacle: x distance from the pelvis, y position of the center relative to the the ground, radius.
\end{itemize}
The reward of an agent is provided at each simulation step and is the distance covered in the run minus the muscle strain as computed by the simulation environment.
\section{Baseline Agent}
When designing the baseline agent, we combined selected techniques from the top 10 competition entries \cite{kidzinski2018l2rsolutions} with further optimizations. In this section, we summarize the most beneficial techniques used, all of which are taken from various contributions published in \cite{kidzinski2018l2rsolutions}.
In all exerimental results presented in the remainder of this paper, the experiments have been repeated 5 times, and the graphs show the standard error from the mean. The RL parameter choice was $\alpha = 0.08$ and $\gamma = 0.9$, which have been determined experimentally.
\subsection{State representation}
The original state representation provided by the competition software contained 41 features, described in Section \ref{sec:domain}. In our state representation we added 71 features, including:
\begin{itemize}
\item Two-dimensional coordinates of key body positions relative to the pelvis at the center point (0,0).
\item Two-dimensional velocity and acceleration vectors for key body points.
\end{itemize}
The new state representation allowed us to significantly speed up the learning process as seen on Figure \ref{img:centerandfeatures}.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/modified.jpg}
\caption{This Figure shows significant learning speed increase after adding velocity and acceleration features and centering the coordinates system at the pelvis position}
\label{img:centerandfeatures}
\end{figure}
\subsection{Additional training experience}
After running a simulation episode, we trained the RL agent with additional mirrored data, which represented the agents experience during the episode and reflecting it along the $xy$ plane. This adds valuable training for the value estimator (i.e. the critic), since the task is symmetrical. Figure \ref{img:mirrored} shows the resulting performance improvement.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/mirrored.jpg}
\caption{Speeding up learning process through adding mirrored data}
\label{img:mirrored}
\end{figure}
\subsection{Repeating the chosen action}
Each time the running agent chooses an action, this is repeated three times. Because we employed an actor-critic method, this reduced the number of computations needed to generate the next action during an episode by a factor of three. The resulting performance gain can be seen in Figure \ref{img:flipaction}.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/flip.jpg}
\caption{Performance gains when repeating each action three times}
\label{img:flipaction}
\end{figure}
\subsection{Reducing state resolution}
In this optimization step, all the state representation data was changed from {\it double} to {\it float}. This resulted in a speed-up of the computations and a somewhat smaller state space, while also reducing the precision of the state representation. Figure \ref{img:doublefloat} shows the resulting performance increase.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/Speedup.jpg}
\caption{Switching from double to float}
\label{img:doublefloat}
\end{figure}
\subsection{Neural network topology}
After applying all of the techniques above, we compared 5 different network architectures by arbitrarily varying the number of layers and neurons per layer. The results are presented in Figure \ref{img:architectures}. For our baseline agent we chose the best performing layer, using 5 layers with 128 neurons each.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/6.jpg}
\caption{Various neural network topologies: LxN denotes L layers with N neurons per layer}
\label{img:architectures}
\end{figure}
\section{Reward Shaping from Video Data}
After designing our baseline, we added potential-based reward shaping from video data taken from arbitrary YouTube videos depicting running of humans and human-like characters. In this section we describe how the potential function was generated.
\subsection{Potential function}
The overall potential function is defined as the sum of potential functions for every body part: pelvis, two knees and two feet. Following the potential-based reward shaping approach, an additional reward is given to an agent on each simulation step corresponding to the change in potentials of the source and target state.
We considered the following three different potential functions for each body part (knee and foot) in our research, all of them based on the inverse of the distance between the respective body part coordinate in the video-generated data and the humanoid robot. The three potential functions represent three different inverse distance functions:
\begin{itemize}
\item PF1: $\frac{1}{dx + dy}$
\item PF2: $\frac{1}{\sqrt{dx^2 + dy^2}} $
\item PF3: $\frac{1}{dx^2 + dy^2}$
\end{itemize}
where $dx$ ($dy$) is the absolute difference between the x (y) coordinate of the respective body part taken from the video data and the x (y) coordinate of the body part of the humanoid robot.
\subsection{Data collection}
For our potential function we have used the following three sources of video data:
\begin{itemize}
\item A video of a cartoon character running (see Figure \ref{img:cartoon} for a screenshot)
\item A video of a running character in a computer game (see Figure \ref{img:videogame} for a screenshot)
\item A video of a running human (see Figure \ref{img:runninghuman} for a screenshot)
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{img/cartoon.jpg}
\caption{Screenshot from a video depicting a cartoon character running (taken from http://y2u.be/2y6aVz0Acx0)}
\label{img:cartoon}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{img/videogame.jpg}
\caption{Screenshot from a video depicting a computer game character running (taken from http://y2u.be/YbYOsE7JyXs)}
\label{img:videogame}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{img/runninghuman.jpg}
\caption{Screenshot from a video depicting a human running (taken from http://y2u.be/5mVgThl-yMU)}
\label{img:runninghuman}
\end{figure}
Each of the sources was used to define a potential function. The performance of the resulting potential functions in RL are compared in Figure \ref{img:var_potential}. Note that the learning curves were not trained until final convergence, which would require much more time and based on the theoretical properties of potential-based reward shaping would ultimately reach the same performance. In each source we recorded the positions of the two knees and the two feet relative to the pelvis as a two-dimensional coordinate. The recording frequency was four positions per half step. The resulting coordinates were normalized according to the OpenSim simulation. While in our work the extraction of the coordinates was done manually, algorithms to accurately extract body part positions in images with a clear view of the body do exist (e.g. \cite{toshev2014deeppose,guler2018densepose}), and we intend to use these in future work.
\subsection{Selecting the data source and the potential function}
We first compared the performance of the potential functions based on three videos and the inverse distance measure PF2. The results for this experiment is shown in Figure \ref{img:differentvideo}, and demonstrates that the human video is the best data source for the reward shaping.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/differentvideo.jpg}
\caption{Comparing the three videos as a data source for the potential function based on PF2}
\label{img:differentvideo}
\end{figure}
After selecting the running human video as the data source, we compared the three different potential functions as depicted in Figure \ref{img:var_potential}. The results show that PF3 performs best.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/pot_functions.JPG}
\caption{Comparing three potential functions}
\label{img:var_potential}
\end{figure}
\section{Evaluation of Video-based Reward Shaping}
Figure \ref{img:baseline_vs_shaping} shows the comparison of our chosen reward shaping approach (PF3) to the RL baseline. The results show that the reward shaping speeds up the learning significantly, reaching double the running speed at 12 hours of training. The end result after 24 hours of training still shows a significant advantage of the reward shaping approach. It is also worth noting that the demonstration video is of a running human who is using his arms, while the simulation model does not include these.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/rewardshaping.jpg}
\caption{Performance comparison between the baseline and the reward shaping approach}
\label{img:baseline_vs_shaping}
\end{figure}
An important advantage of potential-based reward shaping is the theoretical guarantee that the shaping will not change the optimal policy. In order to demonstrate this advantage in our context, we used a weak running robot generated by the baseline RL agent after 12 hours of training as a sub-optimal data source for the potential function. Clearly, the resulting agent is not running optimally, and the positions of the feet and knees will not be in optimal positions most of the time. We then train our RL agent with the reward shaping generated from these sub-optimal coordinates (using PF3), and compared the performance to the weak runner. The results are shown in Figure \ref{img:suboptimal}, and demonstrate that the RL agent is able to overcome the suboptimal performance of the data source. In fact, after 20 hours of training, the performance is more than double that of the suboptimal running agent. Also, note that the suboptimal shaping did not hurt the learning performance significantly. After 12 hours of training the shaped agent performs comparable to the baseline agent with 12 hours of training.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/sim.jpg}
\caption{Performance of the reward shaping approach with suboptimal data. The dotted vertical line represents 12 hours of training (the training time of the shaping source).}
\label{img:suboptimal}
\end{figure}
\section{Conclusions}
In this paper, we presented a method to use videos of human and human-like running to shape the reward of an RL agent learning to run. Our results demonstrate that a significant improvement in learning speed can be achieved by our proposed method, as compared to a strong baseline which we designed combining selected techniques of the top ten entries to the "Learning to Run" competition at NIPS 2017.
In future work, we intend to employ automated body pose extraction methods such as the one presented in \cite{guler2018densepose} and widen our investigation to other humanoid movement apart from running, e.g. jumping.
\section{Introduction}
The \textit{proceedings} are the records of a conference.\footnote{This
is a footnote} ACM seeks
to give these conference by-products a uniform, high-quality
appearance. To do this, ACM has some rigid requirements for the
format of the proceedings documents: there is a specified format
(balanced double columns), a specified set of fonts (Arial or
Helvetica and Times Roman) in certain specified sizes, a specified
live area, centered on the page, specified size of margins, specified
column width and gutter size.
\section{The Body of The Paper}
Typically, the body of a paper is organized into a hierarchical
structure, with numbered or unnumbered headings for sections,
subsections, sub-subsections, and even smaller sections. The command
\texttt{{\char'134}section} that precedes this paragraph is part of
such a hierarchy.\footnote{This is a footnote.} \LaTeX\ handles the
numbering and placement of these headings for you, when you use the
appropriate heading commands around the titles of the headings. If
you want a sub-subsection or smaller part to be unnumbered in your
output, simply append an asterisk to the command name. Examples of
both numbered and unnumbered headings will appear throughout the
balance of this sample document.
Because the entire article is contained in the \textbf{document}
environment, you can indicate the start of a new paragraph with a
blank line in your input file; that is why this sentence forms a
separate paragraph.
\subsection{Type Changes and {\itshape Special} Characters}
We have already seen several typeface changes in this sample. You can
indicate italicized words or phrases in your text with the command
\texttt{{\char'134}textit}; emboldening with the command
\texttt{{\char'134}textbf} and typewriter-style (for instance, for
computer code) with \texttt{{\char'134}texttt}. But remember, you do
not have to indicate typestyle changes when such changes are part of
the \textit{structural} elements of your article; for instance, the
heading of this subsection will be in a sans serif\footnote{Another
footnote here. Let's make this a rather long one to see how it
looks.} typeface, but that is handled by the document class file.
Take care with the use of\footnote{Another footnote.} the
curly braces in typeface changes; they mark the beginning and end of
the text that is to be in the different typeface.
You can use whatever symbols, accented characters, or non-English
characters you need anywhere in your document; you can find a complete
list of what is available in the \textit{\LaTeX\ User's Guide}
\cite{Lamport:LaTeX}.
\subsection{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of
the three are discussed in the next sections.
\subsubsection{Inline (In-text) Equations}
A formula that appears in the running text is called an
inline or in-text formula. It is produced by the
\textbf{math} environment, which can be
invoked with the usual \texttt{{\char'134}begin\,\ldots{\char'134}end}
construction or with the short form \texttt{\$\,\ldots\$}. You
can use any of the symbols and structures,
from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a
few examples of in-text equations in context. Notice how
this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsubsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols
and structures available in \LaTeX\@; this section will just
give a couple of examples of display equations in context.
First, consider the equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\subsection{Citations}
Citations to articles~\cite{bowman:reasoning,
clark:pct, braams:babel, herlihy:methodology},
conference proceedings~\cite{clark:pct} or maybe
books \cite{Lamport:LaTeX, salas:calculus} listed
in the Bibliography section of your
article will occur throughout the text of your article.
You should use BibTeX to automatically produce this bibliography;
you simply need to insert one of several citation commands with
a key of the item cited in the proper location in
the \texttt{.tex} file~\cite{Lamport:LaTeX}.
The key is a short reference you invent to uniquely
identify each work; in this sample document, the key is
the first author's surname and a
word from the title. This identifying key is included
with each item in the \texttt{.bib} file for your article.
The details of the construction of the \texttt{.bib} file
are beyond the scope of this sample document, but more
information can be found in the \textit{Author's Guide},
and exhaustive details in the \textit{\LaTeX\ User's
Guide} by Lamport~\shortcite{Lamport:LaTeX}.
This article shows only the plainest form
of the citation command, using \texttt{{\char'134}cite}.
Some examples. A paginated journal article \cite{Abril07}, an enumerated
journal article \cite{Cohen07}, a reference to an entire issue \cite{JCohen96},
a monograph (whole book) \cite{Kosiur01}, a monograph/whole book in a series (see 2a in spec. document)
\cite{Harel79}, a divisible-book such as an anthology or compilation \cite{Editor00}
followed by the same example, however we only output the series if the volume number is given
\cite{Editor00a} (so Editor00a's series should NOT be present since it has no vol. no.),
a chapter in a divisible book \cite{Spector90}, a chapter in a divisible book
in a series \cite{Douglass98}, a multi-volume work as book \cite{Knuth97},
an article in a proceedings (of a conference, symposium, workshop for example)
(paginated proceedings article) \cite{Andler79}, a proceedings article
with all possible elements \cite{Smith10}, an example of an enumerated
proceedings article \cite{VanGundy07},
an informally published work \cite{Harel78}, a doctoral dissertation \cite{Clarkson85},
a master's thesis: \cite{anisi03}, an online document / world wide web
resource \cite{Thornburg01, Ablamowicz07, Poker06}, a video game (Case 1) \cite{Obama08} and (Case 2) \cite{Novak03}
and \cite{Lee05} and (Case 3) a patent \cite{JoeScientist001},
work accepted for publication \cite{rous08}, 'YYYYb'-test for prolific author
\cite{SaeediMEJ10} and \cite{SaeediJETC10}. Other cites might contain
'duplicate' DOI and URLs (some SIAM articles) \cite{Kirschmer:2010:AEI:1958016.1958018}.
Boris / Barbara Beeton: multi-volume works as books
\cite{MR781536} and \cite{MR781537}.
A couple of citations with DOIs: \cite{2004:ITE:1009386.1010128,
Kirschmer:2010:AEI:1958016.1958018}.
Online citations: \cite{TUGInstmem, Thornburg01, CTANacmart}.
\subsection{Tables}
Because tables cannot be split across pages, the best
placement for them is typically the top of the page
nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and
the table caption. The contents of the table itself must go
in the \textbf{tabular} environment, to
be aligned properly in rows and columns, with the desired
horizontal and vertical rules. Again, detailed instructions
on \textbf{tabular} material
are found in the \textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed
output of this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more desirable.
Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
It is strongly recommended to use the package booktabs~\cite{Fear05}
and follow its main principles of typography with respect to tables:
\begin{enumerate}
\item Never, ever use vertical rules.
\item Never use double rules.
\end{enumerate}
It is also a good idea not to overuse horizontal rules.
\subsection{Figures}
Like tables, figures cannot be split across pages; the best placement
for them is typically the top or the bottom of the page nearest their
initial cite. To ensure this proper ``floating'' placement of
figures, use the environment \textbf{figure} to enclose the figure and
its caption.
This sample document contains examples of \texttt{.eps} files to be
displayable with \LaTeX. If you work with pdf\LaTeX, use files in the
\texttt{.pdf} format. Note that most modern \TeX\ systems will convert
\texttt{.eps} to \texttt{.pdf} for you on the fly. More details on
each of these are found in the \textit{Author's Guide}.
\begin{figure}
\includegraphics{fly}
\caption{A sample black and white graphic.}
\end{figure}
\begin{figure}
\includegraphics[height=1in, width=1in]{fly}
\caption{A sample black and white graphic
that has been resized with the \texttt{includegraphics} command.}
\end{figure}
As was the case with tables, you may want a figure that spans two
columns. To do this, and still to ensure proper ``floating''
placement of tables, use the environment \textbf{figure*} to enclose
the figure and its caption. And don't forget to end the environment
with \textbf{figure*}, not \textbf{figure}!
\begin{figure*}
\includegraphics{flies}
\caption{A sample black and white graphic
that needs to span two columns of text.}
\end{figure*}
\begin{figure}
\includegraphics[height=1in, width=1in]{rosette}
\caption{A sample black and white graphic that has
been resized with the \texttt{includegraphics} command.}
\end{figure}
\subsection{Theorem-like Constructs}
Other common constructs that may occur in your article are the forms
for logical constructs like theorems, axioms, corollaries and proofs.
ACM uses two types of these constructs: theorem-like and
definition-like.
Here is a theorem:
\begin{theorem}
Let $f$ be continuous on $[a,b]$. If $G$ is
an antiderivative for $f$ on $[a,b]$, then
\begin{displaymath}
\int^b_af(t)\,dt = G(b) - G(a).
\end{displaymath}
\end{theorem}
Here is a definition:
\begin{definition}
If $z$ is irrational, then by $e^z$ we mean the
unique number that has
logarithm $z$:
\begin{displaymath}
\log e^z = z.
\end{displaymath}
\end{definition}
The pre-defined theorem-like constructs are \textbf{theorem},
\textbf{conjecture}, \textbf{proposition}, \textbf{lemma} and
\textbf{corollary}. The pre-defined de\-fi\-ni\-ti\-on-like constructs are
\textbf{example} and \textbf{definition}. You can add your own
constructs using the \textsl{amsthm} interface~\cite{Amsthm15}. The
styles used in the \verb|\theoremstyle| command are \textbf{acmplain}
and \textbf{acmdefinition}.
Another construct is \textbf{proof}, for example,
\begin{proof}
Suppose on the contrary there exists a real number $L$ such that
\begin{displaymath}
\lim_{x\rightarrow\infty} \frac{f(x)}{g(x)} = L.
\end{displaymath}
Then
\begin{displaymath}
l=\lim_{x\rightarrow c} f(x)
= \lim_{x\rightarrow c}
\left[ g{x} \cdot \frac{f(x)}{g(x)} \right ]
= \lim_{x\rightarrow c} g(x) \cdot \lim_{x\rightarrow c}
\frac{f(x)}{g(x)} = 0\cdot L = 0,
\end{displaymath}
which contradicts our assumption that $l\neq 0$.
\end{proof}
\section{Conclusions}
This paragraph will end the body of this sample document.
Remember that you might still have Acknowledgments or
Appendices; brief samples of these
follow. There is still the Bibliography to deal with; and
we will make a disclaimer about that here: with the exception
of the reference to the \LaTeX\ book, the citations in
this paper are to articles which have nothing to
do with the present subject and are used as
examples only.
\section{Introduction}
Replicating human movements and behaviour in humanoid robots is a formidable challenge with many exciting applications, ranging from health care (e.g. artificial limbs \cite{alshamsi2016}) to space exploration \cite{nasa2015}. Since the manual engineering of controllers for such tasks is extremely difficult, machine learning and specifically reinforcement learning has received much attention in this area. A recent NIPS competition \cite{kidzinski2018learningtorun} focused on the creation of a simulated humanoid running robot in a continuous and high-dimensional environment. The top entries to the competition employed state-of-the-art deep reinforcement learning techniques \cite{jaskowski2018rltorunfast,kidzinski2018l2rsolutions}, which resulted in strong, but not optimal, performance.
In this paper, we demonstrate that videos showing human running behaviour can be used to significantly improve the learning performance. In our approach, we use the coordinates of specific body parts (e.g. the foot) to define a potential function that is fed into potential-based reward shaping \cite{ng1999policy}.
To create a strong baseline for the evaluation of our approach, we combined selected RL techniques of the top ten competition entries with further optimizations to create a running agent that displays a significantly faster learning rate than the top entry. We then add the reward shaping from video data sampled from various YouTube videos, which resulted in a running agent that reached twice the running speed as our baseline in 12 hours of training.
Since potential-based reward shaping has the nice theoretical property of not changing the optimal policy \cite{ng1999policy}, data taken from sub-optimal running behaviour does not prevent the RL agent from overcoming the sub-optimalities and produce humanoid running that outperforms the data source. We demonstrate this theoretical property empirically by sampling limb positions from a slower-running agent and show how our approach generates a running robot, that after a relatively short time of training, starts to run faster than the original agent.
Overall, the main contribution of our work is to demonstrate how data extracted from videos of human movements can be used to significantly speed up the reinforcement learning of humanoid robots. While our work focuses on the training of humanoid running behaviour, the proposed techniques can easily be applied to any other form of humanoid movements.
\section{Background}
\subsection{Reinforcement Learning}
Reinforcement learning is a paradigm which allows agents to learn by reward and punishment from interactions with the environment \cite{sutton1984temporal}. The numeric feedback received from the environment is used to improve the agent's actions. The majority of work in the area of reinforcement learning applies a Markov Decision Process (MDP) as a mathematical model \cite{puterman2014markov}.
An MDP is a tuple $\big(S, A, T, R)$, where $S$ is the state space, A is the action space, $T(s,a,s') = Pr(s'|s,a)$ is the probability that action a in state s will lead to state $s'$, and $R(s, a, s')$ is the immediate reward $r$ received when action $a$ taken in state $s$ results in a transition to state $s'$. The problem of solving an MDP is to find a policy (i.e., mapping from states to actions) which maximises the accumulated reward. When the environment dynamics (transition probabilities and reward function) are available, this task can be solved using policy iteration \cite{bertsekas1995dynamic}.
When the environment dynamics are not available, as with most real problem domains, policy iteration cannot be used. However, the concept of an iterative approach remains the backbone of the majority of reinforcement learning algorithms. These algorithms apply so called temporal-difference updates to propagate information about values of states and/or state-action pairs, $Q(s, a)$ [20]. These updates are based on the difference of the two temporally different estimates of a particular state or state-action value. The Q-learning algorithm is such a method [21]. After each transition, $(s, a) \rightarrow (s', r)$, in the environment, it updates state-action values by the formula:
\begin{equation}
\label{eq:qlearn}
Q(s,a) \leftarrow Q(s,a) + \alpha[r + \gamma\max Q(s',a') - Q(s,a)]
\end{equation}
where $\alpha$ is the rate of learning and $\gamma$ is the discount factor. It modifies the value of taking action $a$ in state $s$, when after executing this action the environment returned reward $r$, and moved to a new state $s'$.
\subsection{Potential Based reward shaping}
The idea of reward shaping is to provide an additional reward representative of prior knowledge to reduce the number of suboptimal actions made and so reduce the time needed to learn \cite{ng1999policy, randlov1998learning}. This concept can be represented by the following formula for the Q-learning algorithm:
\begin{equation} \label{eq:shaping}
Q(s,a) \leftarrow Q(s,a)+\alpha[r+F(s,s')+\gamma\max Q(s',a') - Q(s,a)]
\end{equation}
where $F(s,s')$ is the general form of any state-based shaping reward.
Even though reward shaping has been powerful in many experiments it quickly became apparent that, when used improperly, it can change the optimal policy \cite{randlov1998learning}. To deal with such problems, potential-based reward shaping was proposed \cite{ng1999policy} as the difference of some potential function $\Phi$ defined over a source s and a destination state $s':F(s,s')=\gamma \Phi(s') - \Phi(s)$ where $\gamma$ must be the same discount factor as used in the agent's update rule (see Equation \ref{eq:qlearn}).
Ng et al. \cite{ng1999policy} proved that potential-based reward shaping, defined according to Equation \ref{eq:shaping}, guarantees learning a policy which is equivalent to the one learned without reward shaping in both infinite and finite horizon MDPs.
Wiewiora \cite{wiewiora2003potential} later proved that an agent learning with potential-based reward shaping and no knowledge-based Q-table initialization will behave identically to an agent without reward shaping when the latter agent's value function is initialized with the same potential function.
These proofs, and all subsequent proofs regarding potential-based reward shaping including those presented in this paper, require actions to be selected by an advantage-based policy \cite{wiewiora2003potential}. Advantage-based policies select actions based on their relative differences in value and not their exact value. Common examples include greedy, $\epsilon$-greedy and Boltzmann softmax.
\subsection{Deep-RL}
In Deep RL, the Q value function is represented as a multi-layer neural network \cite{Goodfellow-et-al-2016}. Deep RL algorithms have been shown to perform strongly on RL tasks which have been infeasible to tackle before. Over recent years, a number of algorithms and optimizations have been proposed, and we have chosen to apply
the Deep Deterministic Policy Gradient (DDPG) algorithm for our application domain \cite{lillicrap2015continuous}. DDPG has been shown to be effective in continuous action domains where classic reinforcement learning methods struggled.
Specifically, in the DDPG algorithm two neural networks are used:
$\mu(S)$ is a network (the {\it actor}) that returns the action vector whose components are the values of the corresponding control signals.
$Q^w(s, a)$ is a second neural network (the {\it critic}, that returns the $Q$ value, i.e. the value estimate of the action of $a$ in state $s$.
\begin{equation} \label{eq:policy5}
\begin{aligned}
\nabla_\theta J(\pi_\theta) & =\int_{S}\rho^\pi(s)\int_{A}\nabla_\theta\pi_\theta(a|s)Q^w(s,a) da ds \\
&=\mathbb{E}_{s\sim \rho^\pi,a\sim\pi_\theta}[\nabla_\theta \log \pi_\theta (a|s)Q^w(s,a)]
\end{aligned}
\end{equation}
where $\theta$ is the parameter vector of the probabilistic policy and $\rho^\pi(s)$ is the probability of reaching state $s$ with policy $\pi$.
For a more complete description of DDPG, see \cite{lillicrap2015continuous}.
\subsection{Reinforcement Learning from Demonstration}
Human expert demonstrations have been demonstrated to improve the learning speed and accuracy of RL agent on a wide range of tasks. Most work in this area (e.g. \cite{suay2016learning, brys2015reinforcement}) focused on the use of state-action recordings as demonstration. This is infeasible in the case of video data, where only state information is available and the demonstration actions are not explicitly provided and often can not be derived either (as is the case with running). More recently, various methods for state-only demonstrations have been proposed (e.g. \cite{peng2018deepmimic, liu2017imitation}). However, all of these methods target the imitation of demonstrations. Our work is employing potential-based reward shaping which uses the demonstrations to speed up the learning, but is also able to overcome any sub-optimalities in the demonstration rather than purely imitating them.
\section{Simulation Environment}
\label{sec:domain}
\begin{figure}
\centering
\includegraphics[width=0.40\textwidth]{img/opensim3.jpg}
\caption{A screenshot of the "Learning to Run" simulation environment}
\label{img:environment}
\end{figure}
The simulation environment has been provided by the "Learning to Run" competition and is based on the OpenSim environment employing the Simbody physics engine. The environment simulates a three-dimensional race course with small obstacles, along which a humanoid robot with 6 joints (ankle, knee, and hip on two legs) and corresponding muscles is running (see Figure \ref{img:environment}). The actions of the running humanoid robot are excitation values applied to the muscles implemented in the robot model. The next state of the environment is computed by the physics engine based on the resulting muscle activations, forces, velocities and positions of the joints.
The OpenSim \cite{kidzinski2018learningtorun} model environment represents the robot state using a vector of 41 features:
\begin{itemize}
\item position of the pelvis (rotation, x, y)
\item velocity of the pelvis (rotation, x, y)
\item rotation of each ankle, knee and hip (6 values)
\item angular velocity of each ankle, knee and hip (6 values)
\item position of the center of mass (2 values)
\item velocity of the center of mass (2 values)
\item positions of head, pelvis, torso, left and right toes, left and right talus (14 values)
\item strength of left and right psoas (a muscle at the lower spine)
\item next obstacle: x distance from the pelvis, y position of the center relative to the the ground, radius.
\end{itemize}
The reward of an agent is provided at each simulation step and is the distance covered in the run minus the muscle strain as computed by the simulation environment.
\section{Baseline Agent}
When designing the baseline agent, we combined selected techniques from the top 10 competition entries \cite{kidzinski2018l2rsolutions} with further optimizations. In this section, we summarize the most beneficial techniques used, all of which are taken from various contributions published in \cite{kidzinski2018l2rsolutions}.
In all exerimental results presented in the remainder of this paper, the experiments have been repeated 5 times, and the graphs show the standard error from the mean. The RL parameter choice was $\alpha = 0.08$ and $\gamma = 0.9$, which have been determined experimentally.
\subsection{State representation}
The original state representation provided by the competition software contained 41 features, described in Section \ref{sec:domain}. In our state representation we added 71 features, including:
\begin{itemize}
\item Two-dimensional coordinates of key body positions relative to the pelvis at the center point (0,0).
\item Two-dimensional velocity and acceleration vectors for key body points.
\end{itemize}
The new state representation allowed us to significantly speed up the learning process as seen on Figure \ref{img:centerandfeatures}.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/modified.jpg}
\caption{This Figure shows significant learning speed increase after adding velocity and acceleration features and centering the coordinates system at the pelvis position}
\label{img:centerandfeatures}
\end{figure}
\subsection{Additional training experience}
After running a simulation episode, we trained the RL agent with additional mirrored data, which represented the agents experience during the episode and reflecting it along the $xy$ plane. This adds valuable training for the value estimator (i.e. the critic), since the task is symmetrical. Figure \ref{img:mirrored} shows the resulting performance improvement.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/mirrored.jpg}
\caption{Speeding up learning process through adding mirrored data}
\label{img:mirrored}
\end{figure}
\subsection{Repeating the chosen action}
Each time the running agent chooses an action, this is repeated three times. Because we employed an actor-critic method, this reduced the number of computations needed to generate the next action during an episode by a factor of three. The resulting performance gain can be seen in Figure \ref{img:flipaction}.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/flip.jpg}
\caption{Performance gains when repeating each action three times}
\label{img:flipaction}
\end{figure}
\subsection{Reducing state resolution}
In this optimization step, all the state representation data was changed from {\it double} to {\it float}. This resulted in a speed-up of the computations and a somewhat smaller state space, while also reducing the precision of the state representation. Figure \ref{img:doublefloat} shows the resulting performance increase.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/Speedup.jpg}
\caption{Switching from double to float}
\label{img:doublefloat}
\end{figure}
\subsection{Neural network topology}
After applying all of the techniques above, we compared 5 different network architectures by arbitrarily varying the number of layers and neurons per layer. The results are presented in Figure \ref{img:architectures}. For our baseline agent we chose the best performing layer, using 5 layers with 128 neurons each.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/6.jpg}
\caption{Various neural network topologies: LxN denotes L layers with N neurons per layer}
\label{img:architectures}
\end{figure}
\section{Reward Shaping from Video Data}
After designing our baseline, we added potential-based reward shaping from video data taken from arbitrary YouTube videos depicting running of humans and human-like characters. In this section we describe how the potential function was generated.
\subsection{Potential function}
The overall potential function is defined as the sum of potential functions for every body part: pelvis, two knees and two feet. Following the potential-based reward shaping approach, an additional reward is given to an agent on each simulation step corresponding to the change in potentials of the source and target state.
We considered the following three different potential functions for each body part (knee and foot) in our research, all of them based on the inverse of the distance between the respective body part coordinate in the video-generated data and the humanoid robot. The three potential functions represent three different inverse distance functions:
\begin{itemize}
\item PF1: $\frac{1}{dx + dy}$
\item PF2: $\frac{1}{\sqrt{dx^2 + dy^2}} $
\item PF3: $\frac{1}{dx^2 + dy^2}$
\end{itemize}
where $dx$ ($dy$) is the absolute difference between the x (y) coordinate of the respective body part taken from the video data and the x (y) coordinate of the body part of the humanoid robot.
\subsection{Data collection}
For our potential function we have used the following three sources of video data:
\begin{itemize}
\item A video of a cartoon character running (see Figure \ref{img:cartoon} for a screenshot)
\item A video of a running character in a computer game (see Figure \ref{img:videogame} for a screenshot)
\item A video of a running human (see Figure \ref{img:runninghuman} for a screenshot)
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{img/cartoon.jpg}
\caption{Screenshot from a video depicting a cartoon character running (taken from http://y2u.be/2y6aVz0Acx0)}
\label{img:cartoon}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{img/videogame.jpg}
\caption{Screenshot from a video depicting a computer game character running (taken from http://y2u.be/YbYOsE7JyXs)}
\label{img:videogame}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{img/runninghuman.jpg}
\caption{Screenshot from a video depicting a human running (taken from http://y2u.be/5mVgThl-yMU)}
\label{img:runninghuman}
\end{figure}
Each of the sources was used to define a potential function. The performance of the resulting potential functions in RL are compared in Figure \ref{img:var_potential}. Note that the learning curves were not trained until final convergence, which would require much more time and based on the theoretical properties of potential-based reward shaping would ultimately reach the same performance. In each source we recorded the positions of the two knees and the two feet relative to the pelvis as a two-dimensional coordinate. The recording frequency was four positions per half step. The resulting coordinates were normalized according to the OpenSim simulation. While in our work the extraction of the coordinates was done manually, algorithms to accurately extract body part positions in images with a clear view of the body do exist (e.g. \cite{toshev2014deeppose,guler2018densepose}), and we intend to use these in future work.
\subsection{Selecting the data source and the potential function}
We first compared the performance of the potential functions based on three videos and the inverse distance measure PF2. The results for this experiment is shown in Figure \ref{img:differentvideo}, and demonstrates that the human video is the best data source for the reward shaping.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/differentvideo.jpg}
\caption{Comparing the three videos as a data source for the potential function based on PF2}
\label{img:differentvideo}
\end{figure}
After selecting the running human video as the data source, we compared the three different potential functions as depicted in Figure \ref{img:var_potential}. The results show that PF3 performs best.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/pot_functions.JPG}
\caption{Comparing three potential functions}
\label{img:var_potential}
\end{figure}
\section{Evaluation of Video-based Reward Shaping}
Figure \ref{img:baseline_vs_shaping} shows the comparison of our chosen reward shaping approach (PF3) to the RL baseline. The results show that the reward shaping speeds up the learning significantly, reaching double the running speed at 12 hours of training. The end result after 24 hours of training still shows a significant advantage of the reward shaping approach. It is also worth noting that the demonstration video is of a running human who is using his arms, while the simulation model does not include these.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/rewardshaping.jpg}
\caption{Performance comparison between the baseline and the reward shaping approach}
\label{img:baseline_vs_shaping}
\end{figure}
An important advantage of potential-based reward shaping is the theoretical guarantee that the shaping will not change the optimal policy. In order to demonstrate this advantage in our context, we used a weak running robot generated by the baseline RL agent after 12 hours of training as a sub-optimal data source for the potential function. Clearly, the resulting agent is not running optimally, and the positions of the feet and knees will not be in optimal positions most of the time. We then train our RL agent with the reward shaping generated from these sub-optimal coordinates (using PF3), and compared the performance to the weak runner. The results are shown in Figure \ref{img:suboptimal}, and demonstrate that the RL agent is able to overcome the suboptimal performance of the data source. In fact, after 20 hours of training, the performance is more than double that of the suboptimal running agent. Also, note that the suboptimal shaping did not hurt the learning performance significantly. After 12 hours of training the shaped agent performs comparable to the baseline agent with 12 hours of training.
\begin{figure}
\includegraphics[width=0.47\textwidth]{img/sim.jpg}
\caption{Performance of the reward shaping approach with suboptimal data. The dotted vertical line represents 12 hours of training (the training time of the shaping source).}
\label{img:suboptimal}
\end{figure}
\section{Conclusions}
In this paper, we presented a method to use videos of human and human-like running to shape the reward of an RL agent learning to run. Our results demonstrate that a significant improvement in learning speed can be achieved by our proposed method, as compared to a strong baseline which we designed combining selected techniques of the top ten entries to the "Learning to Run" competition at NIPS 2017.
In future work, we intend to employ automated body pose extraction methods such as the one presented in \cite{guler2018densepose} and widen our investigation to other humanoid movement apart from running, e.g. jumping.
\section{Introduction}
The \textit{proceedings} are the records of a conference.\footnote{This
is a footnote} ACM seeks
to give these conference by-products a uniform, high-quality
appearance. To do this, ACM has some rigid requirements for the
format of the proceedings documents: there is a specified format
(balanced double columns), a specified set of fonts (Arial or
Helvetica and Times Roman) in certain specified sizes, a specified
live area, centered on the page, specified size of margins, specified
column width and gutter size.
\section{The Body of The Paper}
Typically, the body of a paper is organized into a hierarchical
structure, with numbered or unnumbered headings for sections,
subsections, sub-subsections, and even smaller sections. The command
\texttt{{\char'134}section} that precedes this paragraph is part of
such a hierarchy.\footnote{This is a footnote.} \LaTeX\ handles the
numbering and placement of these headings for you, when you use the
appropriate heading commands around the titles of the headings. If
you want a sub-subsection or smaller part to be unnumbered in your
output, simply append an asterisk to the command name. Examples of
both numbered and unnumbered headings will appear throughout the
balance of this sample document.
Because the entire article is contained in the \textbf{document}
environment, you can indicate the start of a new paragraph with a
blank line in your input file; that is why this sentence forms a
separate paragraph.
\subsection{Type Changes and {\itshape Special} Characters}
We have already seen several typeface changes in this sample. You can
indicate italicized words or phrases in your text with the command
\texttt{{\char'134}textit}; emboldening with the command
\texttt{{\char'134}textbf} and typewriter-style (for instance, for
computer code) with \texttt{{\char'134}texttt}. But remember, you do
not have to indicate typestyle changes when such changes are part of
the \textit{structural} elements of your article; for instance, the
heading of this subsection will be in a sans serif\footnote{Another
footnote here. Let's make this a rather long one to see how it
looks.} typeface, but that is handled by the document class file.
Take care with the use of\footnote{Another footnote.} the
curly braces in typeface changes; they mark the beginning and end of
the text that is to be in the different typeface.
You can use whatever symbols, accented characters, or non-English
characters you need anywhere in your document; you can find a complete
list of what is available in the \textit{\LaTeX\ User's Guide}
\cite{Lamport:LaTeX}.
\subsection{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of
the three are discussed in the next sections.
\subsubsection{Inline (In-text) Equations}
A formula that appears in the running text is called an
inline or in-text formula. It is produced by the
\textbf{math} environment, which can be
invoked with the usual \texttt{{\char'134}begin\,\ldots{\char'134}end}
construction or with the short form \texttt{\$\,\ldots\$}. You
can use any of the symbols and structures,
from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a
few examples of in-text equations in context. Notice how
this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsubsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols
and structures available in \LaTeX\@; this section will just
give a couple of examples of display equations in context.
First, consider the equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\subsection{Citations}
Citations to articles~\cite{bowman:reasoning,
clark:pct, braams:babel, herlihy:methodology},
conference proceedings~\cite{clark:pct} or maybe
books \cite{Lamport:LaTeX, salas:calculus} listed
in the Bibliography section of your
article will occur throughout the text of your article.
You should use BibTeX to automatically produce this bibliography;
you simply need to insert one of several citation commands with
a key of the item cited in the proper location in
the \texttt{.tex} file~\cite{Lamport:LaTeX}.
The key is a short reference you invent to uniquely
identify each work; in this sample document, the key is
the first author's surname and a
word from the title. This identifying key is included
with each item in the \texttt{.bib} file for your article.
The details of the construction of the \texttt{.bib} file
are beyond the scope of this sample document, but more
information can be found in the \textit{Author's Guide},
and exhaustive details in the \textit{\LaTeX\ User's
Guide} by Lamport~\shortcite{Lamport:LaTeX}.
This article shows only the plainest form
of the citation command, using \texttt{{\char'134}cite}.
Some examples. A paginated journal article \cite{Abril07}, an enumerated
journal article \cite{Cohen07}, a reference to an entire issue \cite{JCohen96},
a monograph (whole book) \cite{Kosiur01}, a monograph/whole book in a series (see 2a in spec. document)
\cite{Harel79}, a divisible-book such as an anthology or compilation \cite{Editor00}
followed by the same example, however we only output the series if the volume number is given
\cite{Editor00a} (so Editor00a's series should NOT be present since it has no vol. no.),
a chapter in a divisible book \cite{Spector90}, a chapter in a divisible book
in a series \cite{Douglass98}, a multi-volume work as book \cite{Knuth97},
an article in a proceedings (of a conference, symposium, workshop for example)
(paginated proceedings article) \cite{Andler79}, a proceedings article
with all possible elements \cite{Smith10}, an example of an enumerated
proceedings article \cite{VanGundy07},
an informally published work \cite{Harel78}, a doctoral dissertation \cite{Clarkson85},
a master's thesis: \cite{anisi03}, an online document / world wide web
resource \cite{Thornburg01, Ablamowicz07, Poker06}, a video game (Case 1) \cite{Obama08} and (Case 2) \cite{Novak03}
and \cite{Lee05} and (Case 3) a patent \cite{JoeScientist001},
work accepted for publication \cite{rous08}, 'YYYYb'-test for prolific author
\cite{SaeediMEJ10} and \cite{SaeediJETC10}. Other cites might contain
'duplicate' DOI and URLs (some SIAM articles) \cite{Kirschmer:2010:AEI:1958016.1958018}.
Boris / Barbara Beeton: multi-volume works as books
\cite{MR781536} and \cite{MR781537}.
A couple of citations with DOIs: \cite{2004:ITE:1009386.1010128,
Kirschmer:2010:AEI:1958016.1958018}.
Online citations: \cite{TUGInstmem, Thornburg01, CTANacmart}.
\subsection{Tables}
Because tables cannot be split across pages, the best
placement for them is typically the top of the page
nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and
the table caption. The contents of the table itself must go
in the \textbf{tabular} environment, to
be aligned properly in rows and columns, with the desired
horizontal and vertical rules. Again, detailed instructions
on \textbf{tabular} material
are found in the \textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed
output of this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more desirable.
Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
It is strongly recommended to use the package booktabs~\cite{Fear05}
and follow its main principles of typography with respect to tables:
\begin{enumerate}
\item Never, ever use vertical rules.
\item Never use double rules.
\end{enumerate}
It is also a good idea not to overuse horizontal rules.
\subsection{Figures}
Like tables, figures cannot be split across pages; the best placement
for them is typically the top or the bottom of the page nearest their
initial cite. To ensure this proper ``floating'' placement of
figures, use the environment \textbf{figure} to enclose the figure and
its caption.
This sample document contains examples of \texttt{.eps} files to be
displayable with \LaTeX. If you work with pdf\LaTeX, use files in the
\texttt{.pdf} format. Note that most modern \TeX\ systems will convert
\texttt{.eps} to \texttt{.pdf} for you on the fly. More details on
each of these are found in the \textit{Author's Guide}.
\begin{figure}
\includegraphics{fly}
\caption{A sample black and white graphic.}
\end{figure}
\begin{figure}
\includegraphics[height=1in, width=1in]{fly}
\caption{A sample black and white graphic
that has been resized with the \texttt{includegraphics} command.}
\end{figure}
As was the case with tables, you may want a figure that spans two
columns. To do this, and still to ensure proper ``floating''
placement of tables, use the environment \textbf{figure*} to enclose
the figure and its caption. And don't forget to end the environment
with \textbf{figure*}, not \textbf{figure}!
\begin{figure*}
\includegraphics{flies}
\caption{A sample black and white graphic
that needs to span two columns of text.}
\end{figure*}
\begin{figure}
\includegraphics[height=1in, width=1in]{rosette}
\caption{A sample black and white graphic that has
been resized with the \texttt{includegraphics} command.}
\end{figure}
\subsection{Theorem-like Constructs}
Other common constructs that may occur in your article are the forms
for logical constructs like theorems, axioms, corollaries and proofs.
ACM uses two types of these constructs: theorem-like and
definition-like.
Here is a theorem:
\begin{theorem}
Let $f$ be continuous on $[a,b]$. If $G$ is
an antiderivative for $f$ on $[a,b]$, then
\begin{displaymath}
\int^b_af(t)\,dt = G(b) - G(a).
\end{displaymath}
\end{theorem}
Here is a definition:
\begin{definition}
If $z$ is irrational, then by $e^z$ we mean the
unique number that has
logarithm $z$:
\begin{displaymath}
\log e^z = z.
\end{displaymath}
\end{definition}
The pre-defined theorem-like constructs are \textbf{theorem},
\textbf{conjecture}, \textbf{proposition}, \textbf{lemma} and
\textbf{corollary}. The pre-defined de\-fi\-ni\-ti\-on-like constructs are
\textbf{example} and \textbf{definition}. You can add your own
constructs using the \textsl{amsthm} interface~\cite{Amsthm15}. The
styles used in the \verb|\theoremstyle| command are \textbf{acmplain}
and \textbf{acmdefinition}.
Another construct is \textbf{proof}, for example,
\begin{proof}
Suppose on the contrary there exists a real number $L$ such that
\begin{displaymath}
\lim_{x\rightarrow\infty} \frac{f(x)}{g(x)} = L.
\end{displaymath}
Then
\begin{displaymath}
l=\lim_{x\rightarrow c} f(x)
= \lim_{x\rightarrow c}
\left[ g{x} \cdot \frac{f(x)}{g(x)} \right ]
= \lim_{x\rightarrow c} g(x) \cdot \lim_{x\rightarrow c}
\frac{f(x)}{g(x)} = 0\cdot L = 0,
\end{displaymath}
which contradicts our assumption that $l\neq 0$.
\end{proof}
\section{Conclusions}
This paragraph will end the body of this sample document.
Remember that you might still have Acknowledgments or
Appendices; brief samples of these
follow. There is still the Bibliography to deal with; and
we will make a disclaimer about that here: with the exception
of the reference to the \LaTeX\ book, the citations in
this paper are to articles which have nothing to
do with the present subject and are used as
examples only.
| e8ec9d76bc3ac42a3d8b950b7352da6828031d53 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Emotion recognition in conversation (ERC) is an emerging task in natural language processing (NLP) that aims to identify the emotion of each utterance in a conversation. It can be regarded as an extension of traditional emotion detection from text, or an arising problem in dialogue systems that helps generate emotion-aware dialogues \cite{zhou2017emotional}. Empirical evidence shows that the conversational context of an utterance plays an indispensable role in this task \cite{poria2019emotion}. Moreover, the emotion also tends to stay unchanged within a short context of the conversation. It is thus very critical to effectively model the alternate utterances by different parties.
To solve this problem, many recent works focus on deep neural networks with hierarchical structures to model the conversational data \cite{majumder2019dialoguernn, ghosal2019dialoguegcn,jiao2019higru,zhong2019knowledge}. In these works, each utterance is firstly encoded separately into an utterance representation, which is then modeled sequentially and hierarchically. Although the structures seem to comply with the organization of utterances, they ignore the direct dependencies between words in different utterances. In addition, they are not conducive to the application of pre-trained language models such as BERT \cite{devlin2018bert} and XLNet \cite{yang2019xlnet}, which have achieved superior performance in many dialogue system tasks other than ERC \cite{madotto2018mem2seq,bao2020plato,henderson2019convert}.
There are two main challenges to directly apply these pre-trained language models to ERC. First, conversations in ERC are usually multi-party and there can be intra- and inter-speaker dependencies \cite{ghosal2019dialoguegcn}. Existing pre-trained language models are not readily feasible to encode these dependencies. Second, almost all language models are constrained by the input length. When the input sequence exceeds the limit, it has to be truncated, which may lead to loss of information in distant historical utterances \cite{majumder2019dialoguernn,ghosal2019dialoguegcn}.
To cope with the above challenges, we introduce an all-in-one XLNet model, namely DialogXL, for emotion recognition in multi-turn multi-party conversation. DialogXL intends to apply a strong pre-trained language model to ERC without constructing a complicated, hierarchical model in processing the conversational data. Specifically, it first replaces XLNet's \emph{segment recurrence} by a more flexible and memory-saving \emph{utterance recurrence} to utilize historical utterances. Utterance recurrence stores the hidden states of historical utterances in a memory bank and reuses them while identifying a query utterance. Next, the \emph{self-attention} in XLNet's Transformer layers is substituted for \emph{dialog-aware self-attention}, which consists of four different types of attention, namely local self-attention, global self-attention, speaker self-attention, and listener self-attention. Dialog-aware self-attention allows DialogXL to model the inter- and intra-speaker dependencies under different reception fields in the historical context. We conduct extensive experiments on four ERC benchmarks and the results show that the proposed model, DialogXL, outperforms all the baselines on the datasets. Furthermore, several studies are conducted to verify the modules of DialogXL, and an error analysis is used to delve into the reasons behind the errors.
To conclude, our contributions are as follows:
\begin{itemize}
\item DialogXL is the first effort of pre-trained language models designed for emotion recognition in conversation (ERC).
\item We propose a memory-saving utterance recurrence to replace XLNet's segment recurrence. The new approach allows DialogXL to cache up to 1000 historical words of a conversation, which is more powerful than the vanilla XLNet model.
\item Unlike the original self-attention that merely computes attention weights between words, our dialog-aware self-attention computes them by different reception fields and party roles, allowing us to capture useful intra- and inter-speaker dependencies.
\end{itemize}
\section{Related Work}
\subsection{Emotion Recognition in Conversation}
Emotion recognition in conversation (ERC) has emerged as an important problem in recent years and has attracted numerous interests from the NLP community. The availability of large conversational datasets \cite{busso2008iemocap, schuller2012avec, li2017dailydialog, chen2018emotionlines, poria2019meld} account partly for this phenomenon, and the increasing interests in dialogue systems may also explain it.
Recent works on ERC generally resort to deep learning models. For example, CMN \cite{hazarika2018conversational} and ICON \cite{hazarika2018icon} both utilize gated recurrent unit (GRU) and memory networks. \citeauthor{majumder2019dialoguernn} \shortcite{majumder2019dialoguernn} propose a recurrent-based model to model the party state, global state and emotional dynamics. \citeauthor{jiao2019higru} \shortcite{jiao2019higru} propose a hierarchical GRU structure that trains utterance-level and conversation-level encoders jointly. \citeauthor{ghosal2019dialoguegcn} \shortcite{ghosal2019dialoguegcn} propose a graph neural network based model to encode speaker dependencies and temporal information. \citeauthor{zhong2019knowledge} \shortcite{zhong2019knowledge} incorporate external knowledge bases to support the identification. \citeauthor{hazarika2019emotion} \shortcite{hazarika2019emotion} introduce transfer learning from utterance generation to ERC.
The modalities of data used in the above works are not the same. Specifically, \cite{hazarika2018icon,hazarika2018conversational,majumder2019dialoguernn} utilize textual, audio and video modalities, while the latest research \cite{jiao2019higru,ghosal2019dialoguegcn,zhong2019knowledge,hazarika2019emotion} tends to use only the textual modality.
\subsection{Pre-trained Language Models}
The effectiveness of large pre-trained language models \cite{devlin2018bert, yang2019xlnet, liu2019roberta, conneau2020unsupervised} has been well exhibited in many NLP tasks such as machine reading comprehension, text classification, machine translation. Among the language models, BERT \cite{devlin2018bert} utilizes bi-directional Transformer encoders as well as pre-training schemes of masked language modeling and next sentence prediction. XLNet \cite{yang2019xlnet} is another powerful pre-trained language model, which excels at processing long documents with the segment recurrence mechanism. In addition, it combines the strengths of both auto-encoding and auto-regressive language modeling. There have been some recent works that apply pre-trained language models to dialog-related tasks \cite{bao2020plato,ham2020end, henderson2019convert}, but they have yet to be applied to emotion recognition in conversation.
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{model_overview.pdf}
\caption{The architecture of our DialogXL.}
\label{fig:model_overview}
\end{figure*}
\section{Methodology}
There are two challenges to overcome in order to apply pre-trained language models to emotion recognition in conversation (ERC). The first challenge is how to encode a long historical context with hundreds of words. The second is how to model the intra- and inter-speaker dependencies of different parties. Instead of building a hierarchical network as previous work, we propose DialogXL\footnote{The implementation is available at https://github.com/shen-wzh3/DialogXL.} to address these two challenges on the basis of XLNet with two improvements.
The overview architecture of DialogXL is shown in Figure \ref{fig:model_overview}. It consists of an embedding layer, 12 Transformer layers, and a feed-forward neural network. The model identifies the emotion for each utterance in turn when a conversation comes in. Compared with XLNet, DialogXL has a more effective memory bank equipped during the training and testing phases, storing hidden states of historical utterances for future reuses. The memory bank is updated by a new \emph{utterance recurrence} mechanism. And the hidden states at each Transformer layer are derived by \emph{dialog-aware self-attention}.\footnote{XLNet's permutation language modeling and two-stream self-attention, which are designed for language modeling tasks, are not included in DialogXL.}
\subsection{Problem Definition}
In ERC, a conversation is defined as a list of utterances $\{u_1, u_2, ..., u_N\}$, where $N$ is the number of utterances. Each utterance $u_i$ consists of $n_i$ tokens, namely $u_i = \{w_{i1}, w_{i2},...,w_{in_i}\}$. A discrete value $y_i\in \mathcal{S}$ is used to denote the emotion label of $u_i$, where $\mathcal{S}$ is the set of emotion labels. The speaker is denoted by a function $p(\cdot)$. For example, $p(u_i)\in \mathcal{P}$ denotes the speaker of $u_i$ and $\mathcal{P}$ is the collection of all speaker roles in an ERC dataset. The objective of this task is to output the emotion label $y_t$ for a given query utterance $u_t$ based on its historical context $\{u_1, u_2, ..., u_{t-1}\}$ and the corresponding speaker information.
\subsection{Model Input}
At each time step $t$, the query utterance $u_t$ is prepended with the special token ``\texttt{[CLS]}'':
\begin{equation}
x_t = \{\text{[CLS]}, w_{t1}, w_{t2}, ..., w_{tn_t}\}.
\end{equation}
The utterance is then passed to the embedding layer. In DialogXL, this layer consists of only word embedding. The output of the embedding layer is treated as input hidden states to the first Transformer layer:
\begin{equation}
\mathbf{h}_t^0 = \text{Embedding}(x_t)
\end{equation}
\subsection{Utterance Recurrence}
XLNet \cite{yang2019xlnet} and Transformer-XL \cite{dai2019transformer} address the limitation of input size by a mechanism named \emph{segment recurrence}, which caches previous hidden states in a memory bank and revisits them in future computations. However, this mechanism is ineffective when directly applied to conversational emotion recognition for two reasons. First, the ``segment'' in XLNet refers to a fixed-length sequence rather than a linguistic unit such as a sentence. The conversation in ERC is defined in terms of utterances, which are typically full sentences or paragraphs. Therefore, it is essential to keep the utterances complete rather than segmented into pieces. Second, segment recurrence constrains segments in the same training batch to have the same length, which results in too many paddings stored in memory. By contrast, the proposed \emph{utterance recurrence} stores the historical context in memory without paddings, allowing the memory to store a longer historical context.
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{utterance_recurrence}
\caption{Illustration of the memory update strategies by utterance recurrence and segment recurrence. The batch size is 4, with each row corresponding to a conversation.}
\label{fig:utterance_recurrence}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth]{multi_perspective_attention}
\caption{Demonstration of dialog-aware self-attention: (a) global self-attention, (b) local self-attention, (c) speaker self-attention, and (d) listener self-attention. The utterance to be identified is $u_t$, while the hidden states of $u_{t-3}$, $u_{t-2}$, and $u_{t-1}$ are cached in memory. The speaker identities are: $p(u_t)=p(u_{t-2}) $ and $p(u_{t-1})=p(u_{t-3})$. The window size of local self-attention is 2. The upper part of each subgraph is the attention mask for $u_t$ and other utterances, with masked attention weights colored grey. The lower part of each subgraph is the attention flows between two consecutive Transformer layers, where solid blue lines represent self-attention within $u_t$ and dashed orange lines represent attention between $u_t$ and the memory $\mathbf{m}$.}
\label{fig:multi_perspective_attention}
\end{figure*}
The memory, denoted by $\mathbf{m}$, works like a stack. Every time a new set of hidden states are generated for a query utterance, they are concatenated with the current memory. To prevent from introducing noises into the memory, only the hidden states of the utterance tokens are stored, with the hidden states of the ``\texttt{[CLS]}'' and padding positions ignored. Formally, for the $t$-th utterance, at each Transformer layer $l$ the new memory $\mathbf{m}^{l'}$ is updated as:
\begin{equation}
\mathbf{m}^{l'} = \mathbf{m}^{l} \parallel \mathbf{h}_{t,1:1+n_t}^{l}
\end{equation}
where $\parallel$ denotes the concatenation operation. This update strategy is useful especially during the batching operation. As illustrated in Figure \ref{fig:utterance_recurrence}, updating memory with only the hidden states of utterance tokens makes the memory more compact, for the noises introduced by padding are mostly eliminated and more space is freed to cache a longer context.
\subsection{Dialog-Aware Self-Attention}
Utterances occur alternately by different parties in a conversation, and the vanilla self-attention in XLNet cannot be directly applied to the multi-party setting. To this end, we replace the self-attention by \emph{dialog-aware self-attention}, which enables our model to encode conversation contexts in a multi-turn multi-party setting. The new self-attention consists of four types of self-attention: \emph{global self-attention} and \emph{local self-attention} for different sizes of receptive fields, and \emph{speaker self-attention} and \emph{listener self-attention} for intra- and inter-speaker dependencies. We implement the dialog-aware self-attention by skillfully changing the masking strategies of self-attention, without the need to add any extra embeddings or parameters, as illustrated in Figure \ref{fig:multi_perspective_attention}.
Dialog-aware self-attention is multi-headed. For each attention head of the $l$-th Transformer layer, the attention output $\mathbf{o}^l_t$ is computed as follows:
\begin{equation}\label{eq:attn_begin}
\widetilde{\mathbf{h}}_t^{l-1} = \mathbf{m}^{l-1} || \mathbf{h}_t^{l-1}
\end{equation}
\begin{equation}
\mathbf{q}^l_t,\ \mathbf{k}^l_t,\ \mathbf{v}^l_t = \mathbf{h}_t^{l-1}\mathbf{W}^{l\top}_q,\ \widetilde{\mathbf{h}}_t^{l-1}\mathbf{W}^{l\top}_k,\ \widetilde{\mathbf{h}}_t^{l-1}\mathbf{W}^{l\top}_v
\end{equation}
\begin{equation}
\mathbf{a}_t^l = \text{RelPosAttn}(\mathbf{q}^l_t, \mathbf{k}^l_t)
\end{equation}
\begin{equation}\label{eq:mask}
\widetilde{\mathbf{a}}_t^l = \mathbf{a}_t^l- \mathbf{s}
\end{equation}
\begin{equation}\label{eq:attn_end}
\mathbf{o}_t^l = \text{softmax}(\widetilde{\mathbf{a}}_t^l) \mathbf{v}^l_t
\end{equation}
where $\mathbf{W}^{l}_q$, $\mathbf{W}^{l}_k$, and $\mathbf{W}^{l}_v$ are trainable parameters for each attention head, and $\text{RelPosAttn}(\cdot)$ are the relative position attention adopted from Transformer-XL and XLNet.
The attention mask $\mathbf{s}$ in Equation (\ref{eq:mask}) is a matrix with the same shape as the attention weights $\mathbf{a}_t^l$. The value of $\mathbf{s}_{ij}$ is set to $+\infty$ only when the attention between the $i$-th vector in $\mathbf{q}^l_t$ and $j$-th vector in $\mathbf{k}^l_t$ is masked, and set to 0 otherwise.
For the sake of convenience, we denote Equation (\ref{eq:attn_begin}) to Equation (\ref{eq:attn_end}) by a function $f(\cdot)$:
\begin{equation}
\mathbf{o}_t^l = f(\mathbf{m}^{l-1},\mathbf{h}_t^{l-1},\mathbf{s})
\end{equation}
\subsubsection{Global Self-Attention}
Global self-attention takes all the historical context and the query utterance as the reception field. It is the same as the vanilla self-attention, in which the query utterance pays attention to the whole context. This setting allows our model to attend to previously distant utterances which may also be useful \cite{majumder2019dialoguernn}. Thus, no masking is made for global self-attention:
\begin{equation}
\mathbf{s}^{global}_{ij}=0
\end{equation}
\subsubsection{Local Self-Attention}
Local self-attention only has a reception field of $\omega$ latest historical utterances, where $\omega$ is a hyperparameter. The motivation for this attention is that intuitively speaker's emotion is mostly influenced by the recent utterances. In local self-attention, we mask the attentions between the query utterance and the historical utterances outside the reception field:
\begin{equation}
\mathbf{s}^{local}_{ij} = \left\{
\begin{array}{rl}
+\infty, & j\notin \text{Idx}(\{u_{t-\omega},u_{t-\omega+1},...,u_{t-1},u_t\})\\
0, & \text{Otherwise}\\
\end{array} \right.
\end{equation}
where $\text{Idx}(\mathcal{U})$ is a function that maps the utterance tokens in $\mathcal{U}$ to the corresponding positions in the key matrix $\mathbf{k}^l_t$ .
\subsubsection{Speaker Self-Attention}
Speaker self-attention considers only the historical context spoken by the present speaker. It intends to model the intra-speaker dependency \cite{ghosal2019dialoguegcn} by identifying emotional clues in the speaker's historical utterances. In speaker self-attention, we mask the attentions between the query utterance and the utterances spoken by other speakers:
\begin{equation}
\mathbf{s}^{speaker}_{ij} = \left\{
\begin{array}{rcl}
+\infty, &j\in \text{Idx}(\{u\ |\ p(u)\neq p(u_t)\})\\
0, &\text{Otherwise}\\
\end{array} \right.
\end{equation}
\subsubsection{Listener Self-Attention}
Listener self-attention considers only the historical utterances spoken by other speakers. It intends to model the inter-speaker dependency \cite{ghosal2019dialoguegcn}, meaning that the present speaker's emotion may be influence by other speakers' words. In listener self-attention, we mask the attentions between the query utterance and the utterances made by the present speaker:
\begin{equation}
\mathbf{s}^{listener}_{ij} = \left\{
\begin{array}{rl}
+\infty, &j\in \text{Idx}(\{u\ |\ p(u)=p(u_t)\})\\
0, &\text{Otherwise}\\
\end{array} \right.
\end{equation}
The outputs of the four types of self-attention are concatenated and passed through a normalization layer followed by a feed-forward network to generate the output for this Transformer layer:
\begin{equation}
\widetilde{\mathbf{o}}^l_t = \parallel_{k=1}^{K}f_k(\mathbf{m}^{l-1},\mathbf{h}^{l-1}_t,\mathbf{s}^{c_k} )
\end{equation}
\begin{equation}
\mathbf{h}^l_t = \text{FeedForward}(\text{LayerNorm}(\widetilde{\mathbf{o}}^l_t))
\end{equation}
where $K$ is the number of self-attention heads, and $c_k \in\{global,\ local,\ speaker,\ listener\}$ is the corresponding type of dialog-aware attention for the $k$-th attention head.
\subsection{Model Training}
We take the hidden state of ``\texttt{[CLS]}'' at the last layer as the final encoding of the query utterance and the historical context, and pass it through a feed-forward neural network to get the predicted emotion:
\begin{equation}
\mathbf{h}_t = \mathbf{h}^L_{t,0}
\end{equation}
\begin{equation}
\mathbf{z}_t = \text{ReLU}(\mathbf{W}_h\mathbf{h}_t+\mathbf{b}_h)
\end{equation}
\begin{equation}
P_t = \text{softmax}(\mathbf{W}_z\mathbf{z}_t+\mathbf{b}_z)
\end{equation}
\begin{equation}
\widehat{y}_t = \text{argmax}_{k\in\mathcal{S}}(P_t[k])
\end{equation}
For the training of our model, we use the standard cross-entropy loss as the loss function:
\begin{equation}
\mathcal{L}(\theta) = -\sum_{i=1}^{M}\sum_{t=1}^{N}P_t[y_{i,t}]
\end{equation}
where $M$ is the number of conversations in the training set, and $\theta$ is the collection of trainable parameters in DialogXL.
\section{Experimental Settings}
In this section, we present the experimental settings such as implementation details, datasets, metrics, and baselines.
\subsection{Implementation Details}
We initialize the proposed DialogXL by pre-trained XLNet-Base \cite{yang2019xlnet} and employ AdamW optimizer \cite{loshchilov2018fixing} during training. Hyperparameter tuning for each dataset is conducted with hold-out validation on the validation set. The tunable hyperparameters include learning rate, number of heads for the four types of attentions in dialog-aware self-attention\footnote{The sum of the four types of attention heads is always 12.}, the max length of memory\footnote{When implementing utterance recurrence in practice, the length of memory is set not to exceed a threshold due to the limit of computational resource. Once the size of memory exceeds the threshold, the earliest hidden states will be dropped. }, and the dropout rate. The results of BERT, XLNet, and DialogXL reported in our experiments are all based on the average score of 5 random runs on the test set.
\subsection{Datasets}
We evaluate DialogXL on four multi-turn multi-party ERC datasets. The statistics of them are shown in Table \ref{tab:statistic}.
\begin{table}[t]
\centering
\resizebox{0.475\textwidth}{!}{
\begin{tabular}{l|c|c|c|c|c|c}
\hline
\multirow{2}*{Dataset} & \multicolumn{3}{|c|}{\# Conversations} &\multicolumn{3}{|c}{\# Uterrances}\\
&Train&Val&Test&Train&Val&Test\\
\hline
\hline
IEMOCAP&\multicolumn{2}{c|}{120}&31&\multicolumn{2}{c|}{5810}&1623\\
MELD&1038&114&280&9989&1109&2610\\
DailyDialog&11118&1000&1000&87170&8069&7740\\
EmoryNLP&713&99&85&9934&1344&1328\\
\hline
\end{tabular}
}
\caption{The statistics of four datasets.}
\label{tab:statistic}
\end{table}
\noindent\textbf{IEMOCAP} \cite{busso2008iemocap}: A multimodal conversational dataset for emotion recognition, with two parties included for each conversation. The emotion labels include \texttt{neutral}, \texttt{happiness}, \texttt{sadness}, \texttt{anger}, \texttt{frustrated}, and \texttt{excited}. Since this dataset has no validation set, we follow \cite{zhong2019knowledge} to use the last 20 dialogues in the training set for validation.
\noindent\textbf{MELD} \cite{poria2019meld}: A multimodal dataset for emotion recognition collected from the TV show \texttt{Friends}. The emotion labels include \texttt{neutral}, \texttt{happiness}, \texttt{surprise}, \texttt{sadness}, \texttt{anger}, \texttt{disgust}, and \texttt{fear}.
\noindent\textbf{DailyDialog} \cite{li2017dailydialog}: Human-written daily communications, with emotion labels including \texttt{neutral}, \texttt{happiness}, \texttt{surprise}, \texttt{sadness}, \texttt{anger}, \texttt{disgust}, and \texttt{fear}. Since it has no speaker information, we consider the utterance turns as the speaker turns by default.
\noindent\textbf{EmoryNLP} \cite{zahiri2017emotion}: TV show scripts collected from \texttt{Friends}, but varies from MELD in the choice of scenes and emotion labels. The emotion labels of this dataset include \texttt{neutral}, \texttt{sad}, \texttt{mad}, \texttt{scared}, \texttt{powerful}, \texttt{peaceful}, and \texttt{joyful}.
Following recent works \cite{ghosal2019dialoguegcn,zhong2019knowledge,hazarika2019emotion}, we utilize only the textual data of the above datasets for our experiments. The evaluation metrics are chosen as micro-F1 for DailyDialog\footnote{The category of neutral with significantly more samples than others are not taken into account as before.} and weighted-F1 for the other datasets.
\subsection{Baseline Methods}
We compare DialogXL with the following baselines:
\noindent\textbf{Previous methods:} CMN\cite{hazarika2018conversational}, DialogueRNN \cite{majumder2019dialoguernn}, HiGRU\cite{jiao2019higru}, DialogueGCN\cite{ghosal2019dialoguegcn}, TL-ERC\cite{hazarika2019emotion},
and KET \cite{zhong2019knowledge}.
\noindent\textbf{BERT} \cite{devlin2018bert}: The BERT baseline for ERC, initialized with the pre-trained parameters of BERT-base. We concatenate historical utterances and the query utterance in order and then feed them into BERT for classification. The hyperparameters are tuned the same as DialogXL.
\noindent\textbf{XLNet} \cite{yang2019xlnet}: The XLNet baseline with the original segment recurrence and vanilla self-attention, initialized with the pre-trained parameters of XLNet-base. The hyperparameters are tuned the same as DialogXL.
\section{Results and Analysis}
\subsection{Overall Results}
The overall results of our DialogXL and the baselines are reported in Table \ref{tab:overall}. We can clearly note that DialogXL reaches a new state of the art on all of the four datasets. Besides, we can make another two observations as follows, which help to understand the ERC task and the pros and cons of DialogXL.
First, in general, there are considerable improvements for the pre-trained language models over the others on MELD, DailyDialog, and EmoryNLP. However, the improvements of DialogXL over BERT and XLNet are not significant on these datasets. After delving into the datasets, we found that the dialogues in these datasets are relatively short (mostly 5 to 9 utterances). So the current language models, BERT and XLNet, can already encode the entire historical context and the query utterance in most cases. On these short dialogues, however, the advantages of DialogXL are not shown up completely.
Second, while inferior performance of BERT and XLNet is observed to the other baselines on IEMOCAP, the improvements of DialogXL over BERT and XLNet are significant. After examining the dataset, we realized the dialogues in IEMOCAP are much longer (around 70 utterances per dialog) than the other datasets. In this case, BERT and XLNet cannot encode too much historical context effectively, while such baselines as DialogueRNN and DialogueGCN can reach distant utterances and also encode other key features such as speaker information. Moreover, our DialogXL can both encode the historical context effectively by the utterance recurrence and
capture the speaker information by the dialog-aware self-attention, allowing it to achieve superior performance to all the baselines.
\begin{table}[t]
\centering
\resizebox{0.475\textwidth}{!}{
\begin{tabular}{l|cccc}
\hline
Model & IEMOCAP &MELD &DailyDialog &EmoryNLP\\
\hline
CMN& 56.13 &- &- &-\\
DialogueRNN&62.75&57.03&- &-\\
HiGRU&59.79*
& 56.92* &52.01* &31.88*\\
DialogueGCN&64.18&58.10&- &-\\
TL-ERC&59.30 &57.46* &52.46* &30.57*\\
KET&59.56 &58.18&53.37 &33.95*\\
BERT&60.98 &61.50&54.09 &34.17\\
XLNet&61.33 &61.65&53.62 &34.13\\
\hline
DialogXL &\textbf{65.94}&\textbf{62.41}&\textbf{54.93}&\textbf{34.73}\\
\hline
\end{tabular}
}
\caption{Overall performance on the four datasets. The scores marked by ``*'' is based on our re-implementation, because of the differences in evaluation metrics and data statistics between the corresponding work and ours.}
\label{tab:overall}
\end{table}
\subsection{Effect of the Enhanced Memory}
One of the contributions of DialogXL lies in the enhanced memory with utterance recurrence. Here, we study how the utterance recurrence and the maximum memory length contribute to the final results. We change the maximum memory length from 100 to 1000 with an interval of 100 and plot the test scores on IEMOCAP, which has sufficient utterances in each conversation. The memory waste rate of segment recurrence in XLNet is also plotted in terms of the percentage of paddings in memory. Since the proposed utterance recurrence has 0 memory waste in theory, its memory waste rate is not plotted. Three models are studied for this experiment: XLNet with the original segment recurrence, XLNet with utterance recurrence, and DialogXL.
The results are shown in Figure \ref{fig:memory}. We can note that segment recurrence always leads to a memory waste rate of over 60\% for each different memory length. The rate drops with the memory length increases, along with the growth of the three models. When the memory length exceeds 700, their performance generally stops improving any more, which indicates that increasing the maximum memory length only contributes to the test results within a certain range.
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{memory}
\caption{The results of vanilla XLNet, XLNet with utterance recurrence, and our DialogXL by different maximum memory lengths on IEMOCAP. The memory waste rate of segment recurrence in XLNet is provided for reference.}
\label{fig:memory}
\end{figure}
\subsection{Ablation Study}
In this ablation study, we analyze the impact of dialog-aware self-attention by removing each type of dialog-aware self-attention from DialogXL. The results on two representative datasets, IEMOCAP and MELD, are presented in Table \ref{tab:ablation}.
We can observe that the performance of DialogXL drops on both IEMOCAP and MELD when any type of the self-attention is removed, suggesting that all these self-attentions contribute to the improvement of DialogXL. Nevertheless, their contributions can be distinguished. When speaker self-attention or listener self-attention is removed, considerable drops are observed. But when they are both removed, the drops are more obvious. This implies the importance of the inter-/intra-speaker dependency \cite{ghosal2019dialoguegcn}.
Moreover, when local self-attention is removed, the F1 score drops the most on IEMOCAP, which contains long utterances (around 70) for each conversation. This indicates that the historical context near a query utterance is more important for this dataset. The drop on MELD is not as obvious as on IEMOCAP, because MELD has much shorter conversations (5 to 9 utterances per conversation). Finally, the removal of global self-attention leads to the least performance degradation. The reason could be twofold. First, global utterances are not as important as local utterances. Second, the speaker self-attention and listener self-attention already capture some useful information from distant utterances.
\begin{table}[t]
\centering
\resizebox{0.475\textwidth}{!}{
\begin{tabular}{l|l|l}
\hline
\multirow{2}*{Method} & \multicolumn{2}{|c}{F1 score}\\
\cline{2-3}
&IEMOCAP & MELD\\
\hline
DialogXL & \textbf{65.94}& \textbf{62.41}\\
- speaker self-attention&62.30 ($\downarrow$3.64)& 61.92 ($\downarrow$0.49)\\
- listener self-attention& 62.87 ($\downarrow$3.07)& 62.03 ($\downarrow$0.38)\\
- speaker\&listener self-attention& 61.71 ($\downarrow$4.23)& 61.70 ($\downarrow$\textbf{0.71})\\
- local self-attention& 61.66 ($\downarrow$\textbf{4.28})& 61.72 ($\downarrow$0.69)\\
- global self-attention& 63.34 ($\downarrow$2.60)& 62.15 ($\downarrow$0.26)\\
\hline
\end{tabular}
}
\caption{Results of ablation study on IEMOCAP and MELD.}
\label{tab:ablation}
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=1.9\columnwidth]{visualization}
\caption{Results of error analysis, where two query utterances are provided, along with the visualization of attention weights between the current utterance at the \text{[CLS]} token and the most attended historical utterance (selected according to the highest average attention weight among all the attention heads of the last layer). The darker colors mean larger attention weights.}
\label{fig:visualization}
\end{figure*}
\subsection{Speaker Role Embedding}
Our speaker self-attention and listener self-attention model the speaker dependencies \cite{ghosal2019dialoguegcn} by directly letting the model know which part of the utterances should be attended to. Another way to let a pre-trained language model understand the speaker dependencies in dialog is speaker role embedding \cite{bao2020plato,ham2020end}, which maps each participant to a trainable embedding vector. Here, we make a simple comparison between the two approaches of embedding different parties on IEMOCAP and DailyDialog. To this end, we replace the speaker self-attention and listener self-attention of DialogXL with the speaker role embeddings, and refer to the resulting model as DialogXL-emb. The results of comparison are shown in Table \ref{tab:speker_role}. We can observe that our explicit speaker\&listencer self-attention is more effective than the speaker role embedding approach. As a result, the proposed attention mechanism can be potentially applied to other dialog tasks as well.
\begin{table}[h]
\centering
\resizebox{0.43\textwidth}{!}{
\begin{tabular}{p{2.5cm}|p{2cm}<{\centering}|p{2cm}<{\centering}}
\hline
\multirow{2}*{Method} & \multicolumn{2}{c}{F1 score}\\
\cline{2-3}
&IEMOCAP & DailyDialog\\
\hline
DialogXL & \textbf{65.94}& \textbf{54.93}\\
DialogXL-emb&63.31 & 54.06 \\
\hline
\end{tabular}
}
\caption{Results of comparison between direct speaker role embedding and our speaker\&listener self-attention approach on the IEMOCAP and DailyDialog datasets.}
\label{tab:speker_role}
\end{table}
\subsection{Error Study}
Although our DialogXL has a novel framework and achieves a new state of the art, we still want to figure out its possible shortcomings to motivate the future research. Therefore, we carry out an error study on IEMOCAP. In short, we found that DialogXL's powerful capability of directly capturing word-level features in the historical context can be a double-edge sword. As illustrated in Figure \ref{fig:visualization}, the word-level attention mechanism based on semantic relevance can help make a good prediction (Case \#1), but it may also lead to a mistake by focusing too much on the semantic relevance between the query utterance and historical utterances (Case \#2). As a result, it seems to be necessary to combine with other mechanisms rather than merely relying on the popular attention to carry out the emotion recognition in dialogues.
Besides, we also observe from our bad cases that some of them are mentioned in previous works, such as emotional shifts (i.e., the emotion labels of two consecutive utterances from a same speaker are different) \cite{hazarika2018conversational, majumder2019dialoguernn}. Roughly, our model commits mistakes for 45\% of these cases, which calls for further investigations.
\section{Conclusion}
In this paper, we proposed an all-in-one XLNet model, namely DialogXL, for emotion recognition in conversation (ERC). To model the multi-turn multi-party conversational data, DialogXL contributes two improvements on the basis of XLNet and Transformer-XL. First, an enhanced memory was introduced to replace XLNet's vanilla memory to store historical contexts more effectively. Second, a dialog-aware self-attention mechanism was proposed to deal with the multi-turn multi-party data structures. Extensive experiments were conducted on four ERC benchmarks and the results show that the proposed model outperforms all the baselines on the datasets. The effectiveness of the two improvements is also confirmed by extensive analyses. Furthermore, we have the following three findings. First, the original segment recurrence mechanism stores more than 60\% paddings in memory, making it ineffective to encode the historical contexts for ERC. Second, the traditional speaker role embedding strategy is not as effective as our speaker\&listener self-attention, which could also be applied to other dialog tasks. Finally, an error analysis reveals that merely relying on the attention mechanism may mislead the model.
\section{Acknowledgments}
The paper was supported by the Fundamental Research Funds for the Central Universities (No.19lgpy220) and the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (No.2017ZT07X355). This work was also supported by MindSpore.
| 2ba69cdca152ab895ad7b0a7194677e409624c1b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{Background and Motivations}
Recent years have witnessed the rapid development of large-scale distributed machine learning, which well reconciles the massive computation tasks and the limited computation power of a single worker's machine.
Many workers (e.g., using servers, laptops, or even smartphones) can locally conduct model training based on assigned training data, and they feed back the computation results to the platform to complete a larger machine learning task.
However, the performance of large-scale distributed machine learning systems can be significantly affected by bottlenecks like straggling workers which finish computation much slower than other workers \cite{dean2013tail}.
A promising solution to effectively alleviate straggler bottlenecks is to use coding techniques to create proper computation redundancy \cite{lee2017speeding}.
The key idea of coded machine learning is to properly encode the data to be used for the workers' computation subtasks, such that the platform only needs the computation results from a subset $k$ workers among a total of $n$ workers who are assigned the subtasks, in order to complete the overall computation task. Hence, the fastest subset of workers will determine the overall runtime of the distributed machine learning. The minimum number of workers that the platform needs to wait for (i.e., $k$) is called \emph{recovery threshold}, which depends on the code design and subtask assignment.
\begin{figure}[tbp]
\centering
\subfigure[Encoding and assignment]{
\begin{minipage}[t]{0.42\linewidth}
\centering
\includegraphics[width=1.2 in]{cc7}
\label{fig0}
\end{minipage}}
\subfigure[Computing and decoding]{
\begin{minipage}[t]{0.54\linewidth}
\centering
\includegraphics[width=1.8 in]{cc8}
\label{fig1}
\end{minipage}}
\caption{Illustration of coded machine learning using $(n=3, k=2)$-MDS codes, which can recover the final computation result from only two (out of three) workers' computation results.}
\label{cc}
\vspace{-6mm}
\end{figure}
Fig.~\ref{cc} shows an illustrative example of coded machine learning, where we consider a system with one platform and $n=3$ workers. The platform wants to compute a matrix-vector multiplication $\mathbf{y}=\mathbf{A} \mathbf{x}$ using $(n=3, k=2)$-MDS codes.
\begin{itemize}
\item \emph{Encoding and assignment:} First, the platform splits matrix $\mathbf{A}\in \mathbb{R}^{r \times s}$ equally into two matrices $\mathbf{A}_1 \in \mathbb{R}^{\frac{r}{2} \times s}$ and $\mathbf{A}_2\in \mathbb{R}^{\frac{r}{2} \times s}$. Then, the platform assigns matrix $\mathbf{A}_1$ to worker 1, $\mathbf{A}_2$ to worker 2, and $\mathbf{A}_1+\mathbf{A}_2$ to worker 3. The platform also broadcasts the input vector $\mathbf{x}$ to the three workers.
\item \emph{Computing and decoding:} The three workers will compute subtasks $\mathbf{y}_{1}=\mathbf{A}_{1} \mathbf{x}$, $\mathbf{y}_{2}=\mathbf{A}_{2} \mathbf{x}$, and $\mathbf{y}_{3}=(\mathbf{A}_{1}+\mathbf{A}_{2}) \mathbf{x}$ in parallel, respectively, and return the results to the platform as soon as each worker finishes his own computation. The platform is able to obtain the intended outcome $\mathbf{A} \mathbf{x}$ as long as he receives any $k=2$ workers' results. For example, if worker 1 is the straggling worker which finishes the computation subtask the last, the platform can simply obtain the intended outcome from workers 2 and 3, i.e., $\mathbf{y}=\left[\mathbf{y}_{3}-\mathbf{y}_{2} ; \mathbf{y}_{2}\right]=\left[\mathbf{A}_{1} \mathbf{x}; \mathbf{A}_{2} \mathbf{x}\right]=\mathbf{A} \mathbf{x}$.
\end{itemize}
However, given significant runtime saving using codes, it may not be realistic that workers are willing to participate in coded machine learning without proper incentives, as workers are selfish and incur costs during the computation \cite{duan2012incentive}. For example, a worker needs to postpone his own background task for running the platform's assigned task, which translates to time costs locally.
There are a few challenges in incentivizing workers' participation.
First, it is challenging for the platform to incentivize only the desirable workers to participate in coded machine learning, as the platform may not know workers' computation costs and computation time. Appropriate incentives should be sufficient to cover the costs of desirable workers and be based on workers' computation time performance. The computation costs can be workers' private valuation information, which may not be easily accessed by the platform due to workers' privacy concerns. The worker's computation time depends on many factors such as computation power as well as the unpredictable and unreliable computation infrastructure, hence it may not be even accurately known by the worker himself, not to mention the platform \cite{yang2019timely}. The joint consideration of private information and stochastic information is a key feature of our model, which differs from many incentive mechanism design problems where there is only private and accurate information (to the workers) (e.g., \cite{ningwiopt,ma2018incentivizing}).
This will lead to the first key question of this paper:
\begin{question}
How to incentivize heterogeneous workers with private cost information and stochastic computation time to participate in coded machine learning?
\end{question}
Second, it is under-explored of the impact of workers' private information on the platform's incentive mechanism design and cost.
In the presence of workers' private information, the platform may obtain more information through market research. Different levels of information asymmetry require the platform to design different optimal incentive mechanisms and thus make the platform obtain different costs. It is under-explored how much cost savings the platform can obtain if he has more information about workers.
This motivates us to ask the second key question in this paper:
\begin{question}
What is the impact of platform's incomplete information about workers on the platform's incentive mechanism and cost, compared with complete information?
\end{question}
Third, the platform needs to optimize the load assignment and recovery threshold, which will affect the determination of incentives as well as the performance of coded machine learning.
If the recovery threshold is too small, the load of each worker becomes so large that the overall runtime will eventually increase. If the recovery threshold is too large, the redundancy may not be sufficient to mitigate the effect of straggling workers.
This leads to the third key question of this paper:
\begin{question}
How should the platform design the load assignment and recovery threshold for the incentivized workers?
\end{question}
\subsection{Contributions}
We summarize our key novelty and contributions below.
\begin{itemize}
\item \emph{Incentive mechanisms for coded machine learning:} To the best of our knowledge, this is the first analytical work to study economic incentives in coded machine learning.
The incentive issue determines whether coded machine learning can be implemented with enough workers' participation.
\item \emph{Easy-to-implement incentive mechanisms for heterogeneous workers:} We perform the analysis considering workers' multi-dimensional heterogeneity in computation costs and performances under incomplete information and stochastic information. As a result, searching the target incentivized workers among $M$-many types of workers would require large-scale computation. We manage to summarize workers’ multi-dimensional heterogeneity into a one-dimensional metric to guide the platform's worker selection under incomplete information.
Such a metric reduces the computation complexity from $\mathcal{O}(2^M)$ to $\mathcal{O}(M)$.
We only need to set the recovery threshold linearly proportional to the total participator number when using MDS codes, which is easy to implement.
\item \emph{Impact of incomplete information for mechanism design:}
We show that the platform may give smaller rewards to the workers with efficient computation performances, compared with workers with poor performances under complete information. However, the platform gives larger rewards to the workers with efficient computation performances under incomplete information. These reward-performance relationships are independent of the workers' distribution in each type. We also show
that the platform's increased cost due to incomplete information disappears when worker number is sufficiently large, but it does not monotonically decrease in worker number.
\end{itemize}
\vspace{-1.5mm}
\subsection{Related Work}
\vspace{-0.5mm}
\label{literature}
Most literature about coded machine learning focused on designing coding schemes to mitigate stragglers and improve efficiency in distributed computation, e.g., MDS codes \cite{lee2017speeding}, short-dot codes \cite{dutta2016short}, polynomial codes \cite{yu2017polynomial}, and s-diagonal codes \cite{wang2019computation}. These studies usually focused on the matrix computation tasks such as matrix-vector multiplication (e.g., \cite{reisizadeh2017latency}), matrix-matrix multiplication (e.g., \cite{lee2017high}), and large-sparsity matrix computation (e.g., \cite{suh2017matrix}).
Moreover, a few studies further considered the secure coded computation problem (e.g., \cite{bitar2017minimizing}) and the computation performance heterogeneity of workers (e.g., \cite{reisizadeh2019coded}).
However, existing studies make an optimistic assumption that workers are willing to participate in coded machine learning, which may not be realistic without proper incentives to the workers. To the best of our knowledge, this paper is the first attempt that analytically studies the platform's incentive mechanism design in coded machine learning.
The rest of the paper is organized as follows. The system model is described in Section \ref{model}. We study the platform's incentive mechanisms under complete information and incomplete information in Sections \ref{complete} and \ref{incomple}, respectively. We perform simulations in Section \ref{simulation} and conclude in Section \ref{conclusion}.
\section{System Model and Preliminaries }
\label{model}
We consider a distributed coded machine learning system involving a platform and $N$ workers.
In the following, we will first introduce the platform's task coding and load assignment as well as workers' heterogeneity, and then specify the information scenarios, and finally formulate the Stackelberg game between workers and the platform.
\subsection{Platform Modeling: Task Coding and Load Assignment }
\label{2a}
In this subsection, we introduce how the platform encodes the original task and assigns coded subtasks to workers.
This paper focuses on matrix-vector multiplication tasks in the widely adopted problems such as gradient descent-based algorithms, logistic regression, and reinforcement learning (e.g. \cite{reisizadeh2019coded,reisizadeh2017latency})\footnote{Note that our analysis and insights can be easily applied to matrix-matrix multiplications (which can be decomposed to several matrix-vector multiplications) and other machine learning tasks as long as we can specify the relationship between workers' computation loads and computation time.}. Given a data matrix $\mathbf{A} \in \mathbb{R}^{r \times s}$, the platform wants to compute the task $\mathbf{y}=\mathbf{A} \mathbf{x}$ for an input vector $\mathbf{x} \in \mathbb{R}^{s}$.
Regarding the task coding, we use linear combinations of the $r$ rows of the matrix $\mathbf{A}$ to generate the computation redundancy, such that the platform can recover the result $\mathbf{Ax}$ as soon as receiving any $r$ inner products from the workers.
We take $(n, k)$-MDS codes as an example. If the task is distributed across $n$ workers, for any $ k \in \{1,...,n\}$, the platform first divides matrix $\mathbf{A}$ into $k$ equal-sized submatrices in $\mathbb{R}^{\frac{r}{k} \times s}$. Then, by applying an $(n, k)$-MDS code, the platform obtains $n$ encoded submatrices with unchanged size ${\frac{r}{k} \times s}$, one for each worker. Upon receiving any $k$ workers' results, the platform can decode the result of the original task.
The \emph {computation load} of a worker's subtask is defined as the number of inner products of assigned coded rows of $\mathbf{A}$ with $\mathbf{x}$.
If worker $i$ is assigned a matrix-vector multiplication with matrix size $\ell_{i}\times s$,
his computation load is $\ell_{i}$.
In the example of $(n, k)$-MDS codes, each worker has the same computation load $\ell={r}/{k}$.
Given the computation task, the platform needs to assign appropriate subtasks for heterogeneous workers.
\subsection{Worker Modeling: Heterogeneous Costs and Performances}
We consider a population of $N$ workers in the coded machine learning system, with heterogeneous computation performances and costs as modeled below.
\subsubsection{Computation Performances}
We consider a 2-parameter shifted exponential distribution for the computation time of each worker, which is widely used in literature (e.g., \cite{reisizadeh2019coded,reisizadeh2017latency}) because of the good approximation to the experiment statistics. After being assigned $\ell_i$ rows of the matrix-vector multiplication, worker $i\in \{1,...N\}$ will finish the computation subtask in time $T_i$, which is a random variable with the following cumulative distribution function (CDF):
\begin{equation}
\label{dist}
\begin{split}
\operatorname{Pr}\left(T_{i} \leq t\right)=&1-e^{-{\mu_i}\left(\frac{t}{\ell_i}-a_i\right)}, \forall t \ge a_i\ell_i,
\end{split}
\end{equation}
where parameter $\mu_i > 0$ is worker $i$'s average computation speed and parameter $a_i>0$ tells worker $i$'s start-up time to begin the computation.
Equation \eqref{dist} shows that a worker with a larger $\mu_i$, a smaller $a_i$, or a smaller $\ell_{i}$ is more likely to finish his computation earlier.
We consider that each worker's computation time distribution in \eqref{dist} (including the form as well as parameters $\mu_i$ and $a_i$) is known by the platform and other workers in this paper. However, its realized value is unknown to both the platform and the worker himself ahead of time, due to computing resources' unpredictable noise nature\footnote{The joint consideration of heterogeneous stochastic information and the private information (i.e., computation costs to be introduced below) is a key feature of our model. It differs from many incentive mechanism design papers for networking problems, where the workers have no information uncertainty regarding their own costs or performances (e.g., \cite{ningwiopt,ma2018incentivizing}).}. A worker will return the computation results to the platform as soon as he finishes his computation. Thus, each worker's realized computation performance is known to the platform after computation.
\subsubsection{Computation Costs}
The computation may lead to many costs, such as time cost, energy consumption, and the potential negative impact to other applications. We focus on the time cost in this paper, as the commodity servers usually charge workers on time (e.g. Amazon EC2 \cite{ec2}).
We consider each worker $i$'s computation cost as $c_i$ per unit time of computation.
\subsubsection{Worker Types}
Given workers' multi-dimensional heterogeneity in computation performances and costs, we introduce \emph{worker type} to classify the large number of workers.
Workers are distinguished by the cost, average computation speed, and the start-up computation time.
We call a worker with $(c_m, \mu_m,a_m)$ as a type-$m$ worker. All $N$ workers belong to a set $\mathcal{M}=\{1,2,...,M\}$ of $M$ types. Each type $m\in \mathcal{M}$ has $N_m$ workers, with $\sum_{m=1}^{M}N_m=N$. We will see later in Section \ref{s1} that workers' three-dimensional heterogeneity greatly increases the complexity and difficulty of platform's incentive mechanism design.
\vspace{-2mm}
\subsection{Information Scenarios}
\vspace{-1mm}
To study the impact of incomplete information, we will perform the analysis in different information scenarios.
As for workers' computation time, the platform only knows each worker's computation time distribution (i.e., $\mu$ and $a$). Regarding workers' computation costs, we will consider two information cases in the following:
\vspace{-1mm}
\begin{enumerate}
\item \emph{Complete information (benchmark)}: The platform knows each worker's computation cost $c$ (and thus his type). This may not be easy for the platform to achieve, but it provides the platform's minimum cost in all information cases for comparison.
\item \emph{Incomplete information}: The platform knows the total number of workers $N$ and the specific number of each worker type $\{N_m\}_{m\in\mathcal{M}}$, but does not know each worker's computation cost $c$ (and thus his type).
\end{enumerate}
\vspace{-1mm}
Workers always have incomplete information about each other's costs but know the computation time distributions.
\vspace{-2mm}
\subsection{Two-Stage Stackelberg Game Formulation}
\vspace{-1mm}
In this subsection, we specify the strategies, payoffs, as well as costs of the platform and workers.
\subsubsection{Strategies of Platform and Workers }
As shown in Fig.~\ref{fig16}, we model the decision making of workers and the platform before the computation as a two-stage Stackelberg game.
\begin{itemize}
\item In Stage I, the platform announces the set of targeted worker types $\mathcal{S}\subseteq \mathcal{M}$, the computation loads in each computation round $\boldsymbol{\ell}\triangleq \{\ell_{m}\}_{m\in \mathcal{S}}$, and the rewards in each round $\{p_m^j\}_{m\in \mathcal{M},j\in \{1,...,k\}}$, where $p_m^j$ is the reward for a type-$m$ worker being the $j$-th to finish the computation. Thus, $p_m\triangleq\sum_{j=1}^{k}Pr_m^jp_m^j$ is the expected reward for a type-$m$ worker in each round, where $Pr_m^j$ is the probability of a type-$m$ worker being the $j$-th to finish the computation\footnote{The platform and workers know the participating worker types (i.e., targeted types in $\mathcal{S}$), the worker number of each type $N_m$, and workers' computation time distributions in \eqref{dist} under both complete and incomplete information, so both the platform and workers can derive $Pr_m^j$.}. Equivalently, the platform actually determines the expected rewards for workers $\boldsymbol{p }\triangleq \{p_m\}_{m\in \mathcal{M}}$.
\item In Stage II, after knowing the platform's decisions, each worker decides whether to participate and what type to report (which may not be his true type), as the platform will give rewards based on workers' realized performances and their reported types after the computation.
\end{itemize}
The coded machine learning applies to multiple computation rounds (e.g. in gradient descent problem).
In each round of computation, once the platform receives a decodable subset
of computation results (i.e., $r$ inner products), the platform will inform other workers to stop the computation without any waste \cite{aktas2017effective}. After each round's computation, the platform rewards workers as announced in Stage I
\begin{figure}[tbp]
\centering
\includegraphics[width=1\linewidth]{game6}
\vspace{-5mm}
\caption{Workflow of coded machine learning with incentives: the platform and workers reach an agreement by playing a Stackelberg game before computation, workers perform each round's computation, and the platform rewards workers after each round's computation.}
\label{fig16}
\vspace{-5mm}
\end{figure}
Note that the number of participating workers $n=N^{(\mathcal{S})}\triangleq\sum_{m \in \mathcal{S}}N_m$ is controlled by the incentive mechanism, and the recovery threshold $k$ depends on the load assignment as illustrated in Section \ref{2a}.
\subsubsection{Workers' Payoff}
Due to the randomness of the computation time, each worker can only maximize his expected payoff.
Each worker's expected payoff in each round is the difference between his expected reward and expected computation cost.
If a type-$m$ worker optimally reports/misreports himself as type $\tilde{m}$, his expected reward in each round is $p_{\tilde{m}}$.
In each round, all workers' expected spent time is the expected overall runtime denoted by $\mathbb{E}[T]$.
Note that even if a worker finishes earlier in a particular round, he needs to wait some random time for the rest of workers to finish and cannot turn to any other job, resulting in the same time occupation for all the workers. Thus, a type-$m$ worker's expected computation cost is $c_m\mathbb{E}[T ]$.
In summary, if a type-$m$ worker participates and reports himself as type $\tilde{m}$, his expected payoff in each round is:
\begin{equation}
\label{uhet}
\begin{split}
&\mathbb{E}[U (c_m,\mu_m,a_m,{\tilde{m} })]=p_{\tilde{m}}-c_m\mathbb{E}[T ].\\
\end{split}
\end{equation}
Workers can manipulate his type reporting to achieve the maximum payoff. Each worker will participate in the computation once he expects a non-negative payoff in \eqref{uhet}.
\subsubsection{Platform's Cost}
The platform's expected cost objective in each round involves the overall runtime and total payment to targeted workers. If the platform considers workers' incentive compatibility in designing the incentives (as guaranteed in Section \ref{incomple}), all workers will truthfully report their types and the platform's expected cost in each round is
\begin{equation}
\label{whet}
\begin{split}
&\mathbb{E}[W (\boldsymbol{p },\mathcal{S},\boldsymbol{\ell })]
=\gamma_1\mathbb{E}[T ]+\gamma_2\sum_{m\in \mathcal{S}}N_mp_m,\\
\end{split}
\end{equation}
where term $\mathbb{E}[T ]$ is the expected overall runtime and term $\sum_{m\in \mathcal{S}}N_mp_m$ is the expected total payment to targeted worker types in set $\mathcal{S}$.
Terms $\gamma_1$ and $\gamma_2$ indicate the platform's valuations on expected overall runtime and expected payment, respectively.
\begin{table}[tbp]
\caption{Key Notations}
\vspace{-4.5mm}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$N$ & Number of workers \\
\hline
$M$ & Number of worker types \\
\hline
$r$ & Number of rows of matrix $\mathbf{A}$ or simply total workload \\
\hline
$\mu_m$ & Average computation speed of type-$m$ workers \\
\hline
$a_m$ & Start-up time of type-$m$ workers\\
\hline
$c_m$ & Unit cost of type-$m$ workers\\
\hline
$p_m$& Expected reward for a type-$m$ worker\\
\hline
$\ell_m$ & Computation load in term of rows for a type-$m$ worker\\
\hline
$\mathcal{S}$ & Set of worker types that platform targets at\\
\hline
$N^{(\mathcal{S})}$ & Number of participating workers \\
\hline
$T_i$& Worker $i$'s random computation time\\
\hline
$\mathbb{E}[T]$ & Expected overall runtime\\
\hline
\end{tabular
\label{tablit}
\vspace{-5mm}
\end{center}
\end{table}
\begin{table}[tbp]
\caption{Information Scenarios}
\vspace{-4.5mm}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
\textbf{ Scenarios}&\textbf{Worker heterogeneity} & \textbf{Platform's knowledge} \\
\hline
Complete-Hetero & \tabincell{c}{Computation costs \\ and performances } & \tabincell{c}{Costs and computation\\ time distributions} \\
\hline
Incomplete-Hetero & \tabincell{c}{Computation costs\\ and performances} & \tabincell{c}{ Computation \\time distributions}\\
\hline
\tabincell{c}{Incomplete-\\HeteCostOnly} & Computation costs& \tabincell{c}{ Computation \\time distribution} \\
\hline
\end{tabular
\label{sce}
\vspace{-6mm}
\end{center}
\end{table}
For ease of reading, we list the key notations in Table \ref{tablit} and all information scenarios that we will study in Table \ref{sce}. In the following, we will study the platform's incentive mechanism design under complete information (i.e., ``Complete-Hetero'' scenario) in Section \ref{complete} and that under incomplete information (i.e., ``Incomplete-Hetero'' scenario) in Section \ref{hetero}. Moreover, we will consider a special case with workers' heterogeneity only in costs (i.e., ``Incomplete-HeteCostOnly'' scenario) in Section \ref{homo} to further explore the optimal load assignment and recovery threshold in coded machine learning with incentives.
\section{Incentive Mechanism Design under Complete Information}
\label{complete}
In this section, we study the platform's optimal incentive mechanism under the assumption that the platform knows each worker's type (i.e., ``Complete-Hetero'' scenario). Although complete information may not be practical for the platform to achieve, it serves as a benchmark for later comparison with incomplete information.
\subsection{Computation Load Assignment and Overall Runtime}
We first present the load assignment scheme and compute the corresponding overall runtime.
With workers' heterogeneous computation performances, assigning equal loads to all workers (e.g., in the MDS coding scheme) is clearly not optimal. It is challenging to compute the optimal load assignment in this case, because the expected overall runtime objective is difficult to derive under nonuniform load assignment
due to uncertain order of finish workers and workers' heterogeneous computation time distributions.
To simplify the analysis, we apply asymptotic analysis as the number of workers goes to infinity, similar to \cite{reisizadeh2019coded} which focuses on minimizing the expected overall runtime. The asymptotically optimal load assignment and the corresponding expected overall runtime are given in Lemma \ref{lm1}:
\begin{lemma}
\label{lm1}
When worker number goes to infinity, the following computation loads minimize the expected overall runtime:
\begin{equation}
\label{4}
\begin{aligned}
\ell_{m} &=\frac{r}{ \lambda_{m}\sum_{m\in \mathcal{S}} N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}},\forall m\in \mathcal{S},
\end{aligned}
\end{equation}
where $\lambda_{m}$ is the unique positive solution to $e^{\mu_{m} (\lambda_{m}-a_{m})}=\mu_{m} \lambda_{m}+1$.
The expected overall runtime under scheme \eqref{4} is
\begin{equation}
\label{time1}
\mathbb{E}[T]=\frac{r}{\sum_{m\in \mathcal{S}} N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}.
\end{equation}
\end{lemma}
Proof of Lemma \ref{lm1} is given in Appendix I of the technical report \cite{appen1}.
Term $\lambda_{m}$ is introduced for representing the closed form of the minimum point. The $1/\lambda_{m}$ indicates the computation performance of type-$m$ workers, as $\frac{\partial\lambda_{m}}{\partial\mu_{m}}\le 0$ and $\frac{\partial\lambda_{m}}{\partial a_{m}}\ge 0$. We assign a larger computation load to a worker with better computation performance.
The assignment scheme \eqref{4} can be viewed as an approximation of the finite $N$ case with approximation error $o(1)$ and is applicable to both ''Complete-Hetero'' and ''Incomplete-Hetero'' scenarios.
Next, we use backward induction to study workers' decisions in Stage II and the platform's strategies in Stage I.
\subsection{Workers' Participation Decisions in Stage II}
Based on the load assignment in \eqref{4}, the rewards, and the targeted worker type set $\mathcal{S}$ announced to workers by the platform, each worker decides whether to participate in the computation. Note that here the platform knows each worker's type under complete information, so workers cannot misreport their types.
According to \eqref{uhet} and \eqref{time1}, a type-$m$ worker's expected payoff when truthfully reporting (i.e., $\tilde{m}=m$) and handling the assigned load $\ell_m$ in \eqref{4} is:
\begin{equation}
\label{7}
\begin{split}
\mathbb{E}[U (c_m,\mu_m,a_m,m)]=p_m-{c}_m\frac{r}{\sum_{m\in \mathcal{S}} N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}.
\end{split}
\end{equation}
Each worker locally decides to participate as long as he obtains a non-negative expected payoff in \eqref{7}.
\subsection{Platform's Worker Selection and Rewards in Stage I}
\label{s1}
Considering workers' decisions, the platform determines the targeted worker type set and the rewards for workers in Stage I, by balancing workers' computation performances and costs.
\subsubsection{Problem Formulation}
According to \eqref{whet} and \eqref{time1}, the platform's expected cost objective is
\begin{equation}
\label{8}
\begin{split}
\mathbb{E}[W (\boldsymbol{p },\mathcal{S})]=\gamma_1\frac{r}{\sum_{m\in \mathcal{S}} N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}+\gamma_2\sum_{m \in \mathcal{S}}N_mp_m.\\
\end{split}
\end{equation}
Under complete information, the platform only needs to ensure that targeted workers obtain non-negative payoff \eqref{7}, i.e., satisfy Individual Rationality (IR) constraints as below:
\begin{defn}[Individual Rationality]
The incentive mechanism is individually rational if each targeted type-$m\in \mathcal{S}$ worker receives a non-negative payoff by accepting the expected reward $ {p_m }$ intended for his type, i.e.,
\begin{equation}
\label{9}
\mathbb{E}[U (c_m,\mu_m,a_m,m)]\ge 0, \forall m \in \mathcal{S}.
\end{equation}
\end{defn}
The optimization problem of the platform in the ``Complete-Hetero'' scenario is formally given below.
\begin{problem}[Complete-Hetero]
\label{c}
\begin{equation*}
\begin{split}
\min \quad&\mathbb{E}[W (\boldsymbol{p },\mathcal{S})]\\
s.t.\quad &\mathbb{E}[U (c_m,\mu_m,a_m,m)]\ge 0, \forall m \in \mathcal{S}\\
var. \quad & {\boldsymbol{p }\in [0,\infty)^M,\mathcal{S}\subseteq\mathcal{M}}\\
\end{split}
\end{equation*}
\end{problem}
To minimize the cost objective, the platform sets each targeted worker's expected reward just enough to cover his expected cost, i.e., $\mathbb{E}[U (c_m,\mu_m,a_m,m)]= 0, \forall m \in \mathcal{S}$. By substituting the derived $p_m$ into \eqref{8}, we can simplify Problem \ref{c} to only choose the optimal worker types to target as follows.
\begin{problem}[Worker Selection under Complete Information]
\label{aa}
\begin{equation*}
\begin{split}
\min_{\mathcal{S}\subseteq \mathcal{M}}\left(\gamma_1+\gamma_2\sum_{m\in \mathcal{S}}N_m{c}_m\right)\frac{r}{\sum_{m\in \mathcal{S}} N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}.\\
\end{split}
\end{equation*}
\end{problem}
Problem \ref{aa} is a combinatorial
optimization problem and should be solved by balancing workers' computation performances and costs. It is challenging to directly find the optimal subset of workers given three-dimensional heterogeneity (i.e., $\mu_m$, $a_m$, and $c_m$), which in general requires combinatorial search over all $\sum_{n=1}^{M}C_M^n=2^M-1$ possible subsets of worker types. The complexity is $\mathcal{O}(2^M)$, which is infeasible for a large number of worker types. Alternatively, we will explore to summarize the multi-dimensional heterogeneity as a single metric to guide the platform's worker selection in linear complexity.
\subsubsection{Optimal Solution}
We will use two steps to summarize the three-dimensional heterogeneity (i.e., $\mu_m$, $a_m$, and $c_m$) into a one-dimensional metric.
We first summarize the two-dimensional heterogeneity of workers' performances (i.e., $\mu_m$ and $a_m$) as
\begin{equation}
\phi_m\triangleq\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}.
\end{equation}
The $\phi_m$ can represent type-$m$ workers' computation performance as it increases in average speed $\mu_m$ and decreases in start-up time $a_m$ (as $\lambda_{m}$ decreases in $\mu_m$ and increases in $a_m$).
We further summarize the cost $c_m$ and the performance $\phi_m$ as
\begin{equation}
\Omega_m \triangleq \frac{c_m}{\phi_m}.
\end{equation}
The $\Omega_m$ is type-$m$ workers' \emph{cost-performance ratio}. Without loss of generality, we assume increasing cost-performance ratios among all the $M$ worker types: $\Omega_1\le ...\le \Omega_M$.
The following Theorem \ref{thmc} shows how the one-dimensional metric $\Omega$ guides the platform's selection in the presence of workers' three-dimensional information.
\begin{theorem}
\label{thmc}
In the ``Complete-Hetero'' scenario, there exists a threshold type $n^{ComHet}$,
\begin{equation}
\label{12}
n^{ComHet}\hspace{-1mm} = \hspace{-1mm} \max\left\{n|n\in \mathcal{M},\frac{c_n}{\phi_n}\le \frac{\gamma_1+\gamma_2\sum_{m=1}^{n}N_mc_m}{\gamma_2\sum_{m=1}^{n}N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}\right\},
\end{equation}
such that it is optimal for the platform to induce the participation of worker types in the following targeted set $\mathcal{S}^{ComHet}$:
\begin{equation}
\begin{split}
\mathcal{S}^{ComHet} =\{n|n\in \mathcal{M},n\leq n^{ComHet}\}.\\
\end{split}
\end{equation}
The optimal expected reward for a type-$m$ worker is
\begin{equation}
\begin{cases}
&\hspace{-3mm}p_m^{ComHet}={c}_m\frac{r}{\sum_{m\in \mathcal{S}^{ComHet} } N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}, \; \forall m \in \mathcal{S}^{ComHet} ,\\
&\hspace{-3mm}p_m^{ComHet}=0, \;\forall m \notin \mathcal{S}^{ComHet} .
\end{cases}
\end{equation}
\end{theorem}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.5\linewidth]{thresholdomega}
\vspace{-2.5mm}
\caption{Illustration of $\mathcal{S}^{ComHet}$ in ``Complete-Hetero'' scenario.}
\label{figb}
\vspace{-5mm}
\end{figure}
Proof of Theorem \ref{thmc} is given in Appendix II of the technical report \cite{appen1}.
As illustrated in Fig.~\ref{figb}, Theorem \ref{thmc} shows that with workers multi-dimensional heterogeneity in computation performances and costs, the platform will incentivize the worker types with a small cost-performance ratio under complete information. In other words, the platform has a trade-off between workers' computation performances and computation costs. The platform can tolerate a larger cost $c_m$ when the computation performance is better (i.e., faster average computation speed $\mu_m$ or less start-up time $a_m$).
For the worker types that the platform chooses to incentivize (i.e., types in $\mathcal{S}^{ComHet} $), the platform precisely makes the expected rewards just enough to compensate for each worker's expected computation cost under complete information.
For the worker types out of targeted set $\mathcal{S}^{ComHet} $, the platform will not incentivize them by setting a zero expected reward. We will see in next section that it is impossible for the platform to make all participating workers obtain a zero expected payoff under incomplete information, unless he only targets one type.
Moreover, we have the following remark about the complexity of this easy-to-implement approach.
\begin{remark}
The platform can compute $\mathcal{S}^{ComHet}$ with a linear complexity $\mathcal{O}(M)$, which is much more efficient than $\mathcal{O}(2^M)$.
\end{remark}
The complete information scenario servers as a benchmark to help understand the more realistic case of incomplete information scenario, which we will study in next section.
\section{Incentive Mechanism Design under Incomplete Information}
\label{incomple}
In this section, we study the optimal incentive mechanism in the more realistic incomplete information scenario.
Moreover, we will look at a special case where workers have homogeneous computation time distribution (e.g., under the same equipment or configuration) but still different costs in Section \ref{homo}, to further study the optimal load assignment and recovery threshold under incentives.
\subsection{Workers' Heterogeneity in Both Performances and Costs}
\label{hetero}
In this subsection, we focus on the general ''Incomplete-Hetero'' scenario where workers' heterogeneity lies in both computation performances and computation costs, i.e., workers have different $\mu_{i}$, $a_i$, and $c_i$.
The platform does not know each worker's marginal cost, but he knows what types there are, the number of each worker type, and each worker's computation time distribution.
\subsubsection{Workers' Participation Decisions in Stage II}
Based on the load assignment in \eqref{4}, the rewards, and the targeted worker type set $\mathcal{S}$ announced to workers by the platform, each worker decides whether to participate and (if yes) which type to report\footnote{When a worker decides to participate, he can only choose to report a type that is within the targeted worker type set announced by the platform.}.
According to \eqref{uhet} and \eqref{time1}, if a type-$m$ worker participates and reports his type as $\tilde{m}$, his expected payoff in this case is\footnote{Note that misreport will not affect the expected overall runtime. After task coding, the number of subtasks for workers is fixed, and the platform knows workers' $\mu$ and $a$ which cannot be misreported. Thus, under incomplete information, even if some workers with types not in $\mathcal{S}$ want to participate through misreporting their cost types, the number of workers who can perform the computation and these workers' computation performances will not change. }:
\begin{equation}
\label{18}
\begin{split}
\mathbb{E}[U (c_m,\mu_m,a_m,\tilde{m})]\hspace{-1mm}=p_{\tilde{m}}-{c}_m\frac{r}{\sum_{m\in \mathcal{S}} N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}.
\end{split}
\end{equation}
If a worker does not participate, his expected payoff is 0.
Each worker will participate as long as he can obtain a non-negative payoff, and he will report the type which maximizes his expected payoff to the platform.
Thus, a type-$m \in \mathcal{M}$ worker's optimal reported type (if he participates) is
\begin{equation}
\label{17}
\tilde{m}^*=\arg\max_{\tilde{m}}\mathbb{E}[U (c_m,\mu_m,a_m,\tilde{m})],
\end{equation}
and a type-$m \in \mathcal{M}$ worker's optimal participation decision is
\begin{equation}
\label{18a}
\begin{cases}
&\mathrm{Participate, \hspace{7.5mm}if }\;\mathbb{E}[U (c_m,\mu_m,a_m,\tilde{m}^*)]\ge 0,\\
&\mathrm{Not \; participate,\; if }\;\mathbb{E}[U (c_m,\mu_m,a_m,\tilde{m}^*)]< 0.\\
\end{cases}
\end{equation}
\subsubsection{Platform's Strategies in Stage I}
Taking workers' decisions into consideration, the platform determines the optimal targeted worker type set and the optimal rewards for workers.
Since the platform does not know each worker's type, the platform needs to ensure a non-negative payoff for each targeted worker type and make sure that all workers do not misreport their cost types\footnote{Revelation principle demonstrates that if a social choice function can be implemented by an arbitrary mechanism, then the same function can be implemented by an incentive-compatible-direct-mechanism (i.e. in which workers truthfully report types) with the same equilibrium outcome. Thus, requiring IC will simplify the mechanism design without affecting the optimality.}. In other words, except for the IR constraints, the platform has to make decisions under the Incentive Compatibility (IC) constraints:
\begin{defn}[Incentive Compatibility]
Then incentive mechanism is incentive compatible if each type-$m\in \mathcal{M}$ worker maximizes his own payoff by truthfully reporting his type, i.e.,
\begin{equation}
\label{IC}
\mathbb{E}[U (c_m,\mu_m,a_m,m)] \ge \mathbb{E}[U (c_m,\mu_m,a_m,{\tilde{m} })], \forall m, \tilde{m} \in \hspace{-0.6mm} \mathcal{M}.
\end{equation}
\end{defn}
Formally, the platform's optimization problem in ``Incomplete-Hetero'' scenario is:
\begin{problem}[Incomplete-Hetero]
\label{d}
\begin{equation*}
\begin{split}
\min \quad&\mathbb{E}[W (\boldsymbol{p },\mathcal{S})]\\
s.t. \quad & \rm{IR}\; \rm{Constaints\; in \; } \eqref{9},\; \rm{IC}\; \rm{Constaints\; in \; }\eqref{IC}\\
var. \quad& {\boldsymbol{p }\in [0,\infty)^M,\mathcal{S}\subseteq\mathcal{M}} \\
\end{split}
\end{equation*}
\end{problem}
Compared with Problem \ref{c}, the incomplete information further increases the complexity and difficulty of solving Problem \ref{d} through the additional $M(M-1)$ IC constraints. We present the platform's optimal strategies and the corresponding proof sketch as follows:
\begin{theorem}
\label{thmd}
In the ``Incomplete-Hetero'' scenario, there exists a threshold type $n^{IncHet}$,
\begin{equation}
\label{23}
n^{IncHet}=\arg\min_{n\in \mathcal{M}}\frac{\gamma_1+\gamma_2\frac{c_n}{\phi_n}\sum_{m=1}^nN_m\phi_m}{\sum_{m=1}^n N_m\phi_m},
\end{equation}
such that it is optimal for the platform to induce the participation of worker types in the following targeted set $\mathcal{S}^{IncHet} $:
\begin{equation}
\begin{split}
\mathcal{S}^{IncHet} =\{n|n\in \mathcal{M},n\leq n^{IncHet}\}.
\end{split}
\end{equation}
The platform's optimal expected reward for a type-$m\in \mathcal{M}$ worker is
\begin{equation}
\label{24}
p_m^{IncHet}=\phi_m\frac{rc_{n^{IncHet}}}{\phi_{n^{IncHet}}\sum_{m\in \mathcal{S}^{IncHet} } N_m\frac{\mu_{m}}{1+\mu_{m} \lambda_{m}}}, \forall m\in \mathcal{M}.
\end{equation}
\end{theorem}
Proof of Theorem \ref{thmd} is given in Appendix III of the technical report \cite{appen1}.
Theorem \ref{thmd} shows that the platform is able to derive the optimal set of workers under incomplete information in linear complexity, by summarizing the three-dimensional type information into a one-dimensional metric (i.e., cost-performance ratio $\Omega$).
Although the platform uses the same metric $\Omega$ to select workers in scenarios ``Incomplete-Hetero'' and ``Complete-Hetero'', the numbers of targeted types (i.e., $n^{ComHet}$ in \eqref{12} and $n^{IncHet}$ in \eqref{23}) are different.
One may wonder whether the platform always targets fewer worker types under incomplete information. Our simulation results in Section \ref{simulation} show that it is not always true.
In the two information scenarios, different payment rules will have different influences on the platform's expected cost, resulting in different worker selection rules in \eqref{12} and \eqref{23}.
Different from the ``Complete-Hetero'' scenario where all participating workers obtain a zero payoff, in the ``Incomplete-Hetero'' scenario, all worker types in set $\mathcal{S}^{IncHet}$ except the boundary type obtain positive expected payoffs which are the \emph{information rent} in economics. In the ``Complete-Hetero'' scenario, the platform may give smaller expected rewards to the workers with efficient computation performances, compared with workers with poor performances, as the platform knows workers' costs. However, in ``Incomplete-Hetero'' scenario, the platform sets a larger expected reward for a type with better computation performance $\phi$ as shown in \eqref{24}. Such a reward-performance relationship is independent of the workers' distribution in each type. Moreover, the rewards in Theorem \ref{thmd} ensure that worker types in the desirable set $\mathcal{S}^{IncHet}$ will participate, and
types not in $\mathcal{S}^{IncHet}$ will not participate as the expected reward cannot compensate for their expected costs.
Furthermore, we have the following insight about the optimal number of targeted types in the two scenarios:
\begin{corollary}
\label{pro}
The number of targeted types increases in the platform's valuation on expected overall runtime $\gamma_1$ and decreases in the platform's valuation on payment $\gamma_2$.
\end{corollary}
Proof of Corollary \ref{pro} is given in Appendix IV of the technical report \cite{appen1}.
If the platform attaches more attention on the computation time (larger $\gamma_1$), he is
likely to include more worker types to reduce the overall runtime; if the platform pays more attention on the payment to workers (larger $\gamma_2$), he will incentivize fewer worker types to save money.
Note that the recovery threshold $k$ is a random variable in these two scenarios. Workers' different loads and random computation time lead to the randomness of the number of workers who finish the computation before the platform receives $r$ rows (i.e., the value of $k$ in $\sum_{j=1}^{k}\ell^j=r$, where $\ell^j$ is the load of the $j$-th worker who finishes the computation).
Next, we will consider a special case where workers have the same computation time distribution but different costs to further study the optimal load assignment and recovery threshold in coded machine learning with incentives.
\subsection{Workers' Heterogeneity in Costs Only}
\label{homo}
In this subsection, we focus on the ``Incomplete-HeteCostOnly'' scenario where workers have heterogeneous computation costs $c$ and the same computation performances (i.e., computation time distribution). The platform has incomplete information about workers' costs but knows the workers' computation time distribution and worker type distribution.
Each worker $i$'s computation time (i.e., $T_i$) in this case follows \eqref{dist} with $\mu_{i}=\mu$, $a_i=a$, and $\ell_{i}=\ell, \forall i \in \{1,...,N\}$.
In this case, we consider the $(n, k)$-MDS code as introduced before, which is a widely used code for workers with homogeneous computation time distribution in literature (e.g., \cite{lee2017speeding,reisizadeh2017latency}).\footnote{Our mechanisms can also be extended to other codes similarly. MDS codes can make the expected overall runtime ${\Theta}(\log n)$ times faster than uncoded matrix multiplication (e.g., \cite{lee2017speeding}).} Then, the load assignment and the corresponding expected overall runtime are given as follows:
\begin{lemma}
\label{lm2}
Under the $(n, k)$-MDS codes, each worker has the same computation load
\begin{equation}
\label{14}
\ell=\frac{r}{k}.
\end{equation}
The expected overall runtime under \eqref{14} is
\begin{equation}
\label{time}
\begin{split}
\mathbb{E}[T ]=\frac{r}{k}\hspace{-0.5mm}\left(a\hspace{-0.5mm}+\hspace{-0.5mm}\frac{H_{n}\hspace{-0.5mm}-\hspace{-0.5mm}H_{n-k}}{\mu}\right) \simeq\frac{r}{k}\hspace{-0.5mm}\left(a\hspace{-0.5mm}+\hspace{-0.5mm}\frac{1}{\mu}\log\frac{n}{n\hspace{-0.5mm}-\hspace{-0.5mm}k}\right),
\end{split}
\end{equation}
where $H_{n} \triangleq \sum_{i=1}^{n} \frac{1}{i} $, and $H_{n} \simeq \log n$ when $n$ is large.
\end{lemma}
Proof of Lemma \ref{lm2} is given in Appendix V of the technical report \cite{appen1}.
The analysis of workers' participation decisions and the platform's strategies is similar to that in Section \ref{hetero}. Then, the platform's optimal targeted worker type set, optimal rewards, and optimal recovery threshold in the ``Incomplete-HeteCostOnly'' scenario are given in Proposition \ref{thmb}:
\begin{proposition}
\label{thmb}
Given $ {c_1} \le ...\le {c_{M}} $ in the ``Incomplete-HeteCostOnly'' scenario, there exists a threshold type $n^{IncHco}$,
\begin{equation}
\label{19}
n^{IncHco} =\arg\min_{{n\in \mathcal{M}}} \frac{\gamma_1+\gamma_2{c_n}\sum_{m=1 }^n N_m}{\sum_{m=1 }^n N_m},
\end{equation}
such that it is optimal for the platform to induce the participation of worker types in the following targeted set $\mathcal{S}^{IncHco} $:
\begin{equation}
\begin{split}
\mathcal{S}^{IncHco} =\{n|n\in \mathcal{M},n\leq n^{IncHco}\}.
\end{split}
\end{equation}
The optimal expected reward for a type-$m\in \mathcal{M}$ worker is
\begin{equation}
p_m^{IncHco}={c}_{n^{IncHco} }\frac{r}{k^*}\left(a+\frac{1}{\mu}\log\frac{1}{1-\alpha}\right),\forall m\in \mathcal{M},
\end{equation}
and the platform's optimal recovery threshold is
\begin{equation}
k^*=\alpha N^{(\mathcal{S}^{IncHco} )},
\end{equation}
where $\alpha\triangleq \left[1+\frac{1}{W_{-1}\left(-e^{-a\mu-1}\right)}\right] \in [0,1]$, $W_{-1}(\cdot)$ is the lower branch of Lambert W function\footnote{$x=-W_{-1}\left(-\frac{1}{t}\right)$ is the unique solution to $\frac{e^{x}}{x}=t, t \geq e$ and $x \geq 1$.}, and $N^{(\mathcal{S}^{IncHco})}\triangleq\sum_{m \in \mathcal{S}^{IncHco}}N_m$.
\end{proposition}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.5\linewidth]{thresholdc}
\vspace{-2.5mm}
\caption{Illustration of $\mathcal{S}^{IncHco}$ in ``Incomplete-HeteCostOnly'' scenario.}
\label{figa}
\vspace{-2.5mm}
\end{figure}
Proof of Proposition \ref{thmb} is given in Appendix VI of the technical report \cite{appen1}.
As illustrated in Fig.~\ref{figa}, Proposition \ref{thmb} shows that in the ``Incomplete-HeteCostOnly'' scenario, the platform prefers the worker types with small marginal costs, as workers have the same computation performance in this case.
The platform gives all workers the same expected reward, which ensures that worker types in the desirable set $\mathcal{S}^{IncHco}$ will participate, and
types not in $\mathcal{S}^{IncHco}$ will not participate.
More importantly, Proposition \ref{thmb} presents that the MDS codes' optimal recovery threshold $k^*$ is linearly proportional to the total participator number $N^{(\mathcal{S}^{IncHco} )}$, which provides an easy-to-implement guideline for data encoding. According to \eqref{14}, the optimal load assignment is $\ell_{m}^*=r/k^*, \forall m\in \mathcal{S}$.
Next, we will compare the platform's strategies and costs in different scenarios through numerical analysis in Section \ref{simulation}.
\section{Simulation Results}
\label{simulation}
In Section \ref{vs}, we will evaluate the performance of our proposed incentive mechanisms.
In Section \ref{more}, we will further consider a \emph{strongly incomplete information scenario} where the platform further lacks the information about workers' distributions in different types (i.e., $N_m$ of type $m$).
We will show that our incentive mechanism for incomplete information scenario is asymptotically optimal under strongly incomplete information.
\begin{table}[tbp]
\caption{Workers' Parameter Settings in Simulation}
\vspace{-4.5mm}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$(c_1,\mu_1,a_1)=(1,50,0.012)$ & $(c_2,\mu_2,a_2)=(7,100,0.024)$ \\
\hline
$(c_3,\mu_3,a_3)=(8,200,0.033)$ & $(c_4,\mu_4,a_4)=(3,10,0.031)$ \\
\hline
$(c_5,\mu_5,a_5)=(16,400,0.040)$ & $(c_6,\mu_6,a_6)=(5,20,0.081)$ \\
\hline
$(c_7,\mu_7,a_7)=(21,800,0.044)$ & $(c_8,\mu_8,a_8)=(9,40,0.123)$ \\
\hline
$(c_9,\mu_9,a_9)=(12,80,0.153)$ & $(c_{10},\mu_{10},a_{10})=(20,160,0.172)$ \\
\hline
\end{tabular
\label{3tab}
\vspace{-6mm}
\end{center}
\end{table}
When workers have three-dimensional heterogeneity in computation performances and costs, we consider $M= 10$ types of workers with parameter settings in Table \ref{3tab},
similar to that in \cite{reisizadeh2019coded}.
The parameters satisfy increasing cost-performance ratio relationship: $\Omega_1\le ...\le \Omega_M$.
There are $r=1000$ inner products to be finished by workers.
Each type has $N_m=N/M,\forall m\in \mathcal{M}$ workers. We consider both $\gamma_1=2000$ and $\gamma_2=1$ to reflect the platform's different evaluation scenarios on time cost and payment.
\subsection{Complete versus Incomplete Information}
\label{vs}
Fig.~\ref{fig10} shows the number of platform's targeted worker types characterized in Theorems \ref{thmc} and \ref{thmd} versus the total worker number $N$, under both complete and incomplete information. Both curves are decreasing in $N$. This is because the worker types of small cost-performance ratios have more workers when $N$ increases. The platform can rely less on the inefficient or costly types.
By comparing the results in two information scenarios in Fig.~\ref{fig10}, we realize that the platform does not always target at a smaller group of workers under incomplete information. For example, the platform includes $4$ worker types under incomplete information versus $3$ worker types under complete information at around $N=1400$. The platform under incomplete information afraid of large overall runtime wants to include one more type despite the mildly increased cost.
\begin{figure}[tbp]
\centering
\includegraphics[width=2.2 in]{type}
\vspace{-3mm}
\caption{Number of worker types targeted by the platform (under both complete and incomplete information) versus the total worker number $N$.}
\label{fig10}
\vspace{-3.5mm}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=2.2 in]{cost}
\vspace{-3mm}
\caption{Platform's increased cost due to the lack of information versus the total worker number $N$.}
\label{fig8}
\vspace{-1mm}
\end{figure}
\begin{figure}[tbp]
\centering
\vspace{-2mm}
\includegraphics[width=2.2 in]{payoff}
\vspace{-3mm}
\caption{Different worker types' payoffs under incomplete information versus the total worker number $N$.}
\label{fig13}
\vspace{-5mm}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=2.2 in]{diffstro}
\vspace{-3mm}
\caption{Platform's increased cost due to strongly incomplete information compared with incomplete information versus the total worker number $N$.}
\label{fig15}
\vspace{-5mm}
\end{figure}
Fig.~\ref{fig8} compares the platform's cost objectives under complete and incomplete information, by measuring increased cost due to incomplete information versus the worker number $N$. When the number of workers becomes very large (larger than $3400$ in Fig.~\ref{fig8}), the platform obtains the same cost in these two scenarios.
This is because the platform only chooses the type with smallest cost-performance ratio in both complete and incomplete information scenarios (Fig.~\ref{fig10}).
However, the platform's cost gap does not monotonically decrease in $N$. The platform may sometimes target more worker types under incomplete information (see Fig.~\ref{fig10} when $N$ equals $1400$), which increases cost difference compared to the complete information scenario.
Fig.~\ref{fig13} further shows that different types of workers' payoffs under incomplete information versus the worker number $N$. Only the targeted worker types can have positive payoffs, and such payoffs are overall decreasing in $N$ due to mutual competition and the platform's smaller set of targeted types.
\subsection{Extension of Incomplete Information}
\label{more}
In Section \ref{incomple}, we have studied the mechanism design under incomplete information without knowing each worker's computation time and cost.
Here we further evaluate the performance of the proposed mechanism in Theorem \ref{thmd} under strongly incomplete information, where the platform only knows the total number of workers and the distribution of each worker's type. Each worker has an equal probability of 10\% of belonging to each worker type.
In this strongly incomplete information scenario, as the number of workers $N$ becomes large, according to the law of large numbers, the empirical number of each type of workers approaches to the expected value computed based on the distribution.
As shown in Fig.~\ref{fig15}, when $N$ is larger than $400$, the increased cost due to strongly incomplete information reaches zero\footnote{The fluctuations are due to the random realization of the number of each type workers.}. Thus, our incentive mechanism in Theorem \ref{thmd} performs optimally under this strongly incomplete information scenario as $N$ goes to infinity.
\section{Conclusion}
\label{conclusion}
This paper studied the important incentive issues in coded machine learning. We captured the trade-off between the platform's time cost and payment in his payoff function. We proposed the optimal incentive mechanisms for the workers in both complete information and incomplete information scenarios, under workers' multi-dimensional heterogeneity in computation performances and costs.
In the presence of the high complexity of the worker selection problem, we managed to summarize workers' multi-dimensional heterogeneity into a one-dimensional metric and solved the problem in a linear complexity. We also demonstrated that the optimal recovery threshold is linearly proportional to the total participator number when using MDS codes. We showed that compared with the complete information scenario, the effect of incomplete information on the platform's cost will disappear when the total worker number is sufficiently large.
We will further study the optimal incentive mechanisms of competing platforms in future work.
\bibliographystyle{IEEEtran}
| 65702fa55fe6ad2cd6c458937c0a5ac012aad1be | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Graph Neural Networks (GNNs) extends Convolutional Neural Networks (CNNs) to graph-based data. A common problem solved by GNNs is node classification. Given a partially labelled graph structure and the data defined on each node, the goal of node classification is to accurately classify the unlabelled nodes. For example, Fig.~\ref{figure:citeseer} shows the graph representation of a node classification dataset before and after the node classification, where the colors represent labels of the nodes. On the left, we have a partially labelled graph. On the right, the graph is fully labelled from node classification. The graph structure represents a citation network, where each node represents a paper, and each undirected edge between two nodes represents a citation between the two papers.
\begin{figure}[hbp!]%
\centering
\includegraphics[width=\linewidth]{images/compare_cite.png}
\caption{Visualization \cite{GEPHI} of Citeseer \cite{node_dataset} Dataset.}
\label{figure:citeseer}
\end{figure}
While GNNs are commonly used for learning irregular data residing on non-Euclidean domain, it is challenging to evaluate how much the underlying graph structure helps with the model's performance. Node classification problems can be solved using either a GNN or a CNN\footnote{In this case, the CNN is similar to a Multilayer Perceptron (MLP) because the sliding window of convolution layers has size 1.}. One benefit of using a GNN is that it uses the given graph structure to classify nodes, while a CNN does not. However, we do not know whether using the graph structure and a GNN will provide any advantages over using a CNN and no graph structure a priori. The graph structure could provide little to no information about the labels. This motivates our question: How can we evaluate the effectiveness of the graph structure for GNNs?
To answer this question, we need to address two aspects. First, we need to define what makes a graph structure effective. We say a graph structure is relatively effective if it contains a relatively large amount of information about the label distribution. This information added by the graph structure is reflected in test accuracies. For instance, Fig.~\ref{figure:compare} shows the test accuracies of a GNN model on a citation network dataset using three different graph structures. We say the underlying graph structure of CORA-ML is more effective than the random graph, because while the edges in CORA-ML represent citations, the edges in the random graph are randomly drawn and convey no useful information. Similarly, the test accuracy of the model trained with the underlying graph structure (orange solid line) is around 20 percent higher than the model trained with the identity matrix (grey dashed line), whereas the test accuracy of the model trained with the random graph (purple dotted line) has almost no improvement from the model trained with the identity matrix.
\begin{figure}[h!]
\centering
\includegraphics[width=7cm]{images/tagcnComp.png}
\caption{Comparison of Test Accuracy of CORA-ML Dataset \cite{node_dataset} of a 2-layer TAGCN\cite{TAGCN} with 1) given dataset's graph, 2) identity matrix, 3) random Erd\H{o}s-R\`enyi graph \cite{SPM}.}
\label{figure:compare}
\end{figure}
The second question we need to address is, how do we quantitatively evaluate the effectiveness of graph structure? In this paper, we propose an information theoretic parameter called edge entropy to measure the quality of label information contained in the graph structure. We show edge entropy is a good indicator of the effectivenss of GNN over CNN using experiment results from both synthetic and real datasets.
\section{Related Work}
Node classification has been used in many different settings, but very little has been done on evaluating the effectiveness of graph structure in the node classification datasets. Much of the existing work on evaluating GNN models focuses on developing benchmarks and upper bounding the modeling capacity for graph classification tasks. Dwivedi, Joshi, Laurent, Bengio and Bresson \cite{dwivedi2020benchmarkgnns} introduced a standardized benchmarking framework for running a variety of GNN experiments. Xu, Hu, Leskovec and Jegelka \cite{powerful} showed that a Graph CNN is only as powerful as the Weisfeiler-Lehman graph isomorphism test for graph classification tasks.
Another related but different problem is graph comparison. Metrics such as the clustering coefficient \cite{SmallWorld}, the betweenness centrality \cite{centrality}, and graph spectral distance \cite{spectral} were proposed to discern the topological properties of different graph structures. This problem is different than ours, because these metrics do not consider the nodes' labels. Our challenge is to evaluate the effectiveness of the given graph structure based on both the connectivity and the label distribution.
\section{Problem Statement}
We are interested in evaluating the effectiveness of GNN over CNN by examining the effectiveness of the graph structure in node classification datasets. In this paper, we focus on a variant of GNN called Graph Convolution Neural Networks (Graph CNNs), where the convolution layer is explicitly defined using the adjacency matrix.
\subsection{Node Classification}
A node classification dataset $D$ contains the following:
\begin{itemize}
\item the graph structure, i.e., the adjacency matrix $A$
\item the data defined on each node $X$
\item the labels for all nodes $y$\textbf{}
\end{itemize}
The goal of node classification is to accurately classify unlabelled nodes based on $A,X$, and part of $y$.
\subsection{Filter-based Graph CNN}
We focus on filter-based graph CNN models such as TAGCN\cite{TAGCN} and GCN\cite{GCN}. Filter-based graph CNN models represent convolution as a filtering operation with a polynomial of the graph shift $S$, where $S$ can be either the adjacency matrix $A$ or the graph Laplacian $L$ \cite{Overview}. Formally, an adjacency matrix based graph convolution layer learns a graph filter \cite{TAGCN}:
\begin{equation}
P(A)(X) = \sum_{k=0}^{d-1} A^{k} X W_{k}
\label{eq:conv}
\end{equation}
where $P(A)$ is a polynomial of $A$, $X$ is the input graph signal, $d$ is the degree of the graph polynomial, and $\{W_{k}\}_{k=0}^{d-1}$ are learnable weights.
\subsection{Effectiveness of Graph Structures}
For a node classification problem, a graph structure is less effective if the graph structure contains little information about the label distribution. This lack of information can be seen in the test accuracies of the graph CNN model. For a less effective graph structure, the model trained with the given adjacency matrix will yield little improvement in test accuracy compared with the model trained with the identity matrix, i.e., trained with the features only.
For example, suppose we have two node classification datasets $D_{1}$ and $D_{2}$. We train a graph CNN model on both datasets using the given adjacency matrix. We also train the model for both datasets using the identity matrix instead of the adjacency matrix.
We define the improvement on $D_i$ as the difference in test accuracy when using the given adjacency matrix versus using the identity matrix.
If the improvement on $D_1$ is greater than the improvement on $D_2$, then the graph structure in $D_1$ provides more information, i.e., is more effective, than the graph structure in $D_2$. Likewise, if the improvement on $D_2$ is greater, then the graph structure in $D_2$ is more effective than the graph structure in $D_1$.
\section{Proposed Method}
To evaluate the effectiveness of the graph structure in a node classification task, we propose a parameter to evaluate the quality of label information contained in the graph structure. We will first motivate our definition with some simple examples, then we will define the parameter.
\subsection{Motivating Examples}
\label{sec:examples}
A naive attempt at evaluating the effectiveness of a graph structure is related to node clustering. Consider the graph in Fig. \ref{figure:graphs}(a). There is a blue cluster and a red cluster. Here, blue nodes always connect with blue nodes and red nodes always connect with red nodes. In this case, the edges reveal a lot of label information. For instance, suppose we want to classify an unlabelled node. If we know it has a blue neighbor, then we can confidently classify this node as blue, because the blue cluster and the red cluster are disjoint.
On the other hand, consider the graph in Fig. \ref{figure:graphs}(b). Here, the blue and red clusters are mixed. It is equally likely for a blue node to connect with either another blue node or a red node. In this case, the edges do not contain useful label information. For instance, knowing that an unlabelled node has a blue neighbor does not make that node more likely to be blue or red.
\begin{figure}[h!]
\centering
\includegraphics[width=8cm]{images/graphs2.png}
\caption{Simple graphs.}
\label{figure:graphs}
\end{figure}
A simple parameter inspired by this observation is intra-class ratio, which measures the percentage of edges that connect nodes of the same class. One might hypothesize that a graph with high intra-class ratio is easier to classify with the underlying graph structure, because the clusters are more separated. However, a counter-example for this hypothesis is shown in Fig. \ref{figure:graphs}(c). In this bipartite graph, blue nodes always connect with red nodes and red nodes always connect with blue nodes. In this case, the intra-class ratio is zero, but the graph is very helpful for node classification. For instance, if an unlabelled node has a blue neighbor, then we know it must be a red node.
Another parameter to consider is the clustering coefficient, which measures how likely nodes would form cliques in the graph. One might assume that a higher clustering coefficient indicates that accounting for the graphs will be more helpful for node classification, because the clusters are more obvious. However, Fig.~\ref{figure:graphs}(d) shows a counter-example for this observation. In this 4-clique graph, the clustering coefficient is 1, but the graph structure does not reveal label information. This is because we have a mixed cluster of both blue and red nodes. In this case, the clustering coefficient provides little help with evaluating the effectiveness of graph structure for node classification, because it ignores the labels of nodes.
To account for both the connections between different classes and the labels of nodes, we propose a more comprehensive parameter called edge entropy.
\subsection{Edge Entropy}
We propose edge entropy as a parameter to evaluate the impact of accounting for the graph structures in node classification. Edge entropy measures the quality of label information encoded in the graph.
A high edge entropy (closer to 1) indicates that accounting for the graph structure is very helpful for node classification. A low edge entropy (closer to 0) indicates that accounting for the graph structure is not very useful for node classification.
Formally, given a graph $\mathcal{G}$ with~$M$ classes of nodes, we define the \textbf{per-class edge entropy} of any class $l$ as a function $H: \{1, \hdots, M\} \to [0,1]$ such that
\begin{equation}\label{edgeEntropy}
H(l) := -\sum_{m \in \{1, \hdots, M\}} p_{lm}(n)\log_{M}(p_{lm}(n))
\end{equation}
where the first order interclass connectivity probability $p_{lm}$ is defined as
\begin{equation}\label{connProb}
p_{lm} := \frac{|\{\text{edge } w: \text{start}(w) \in \mathcal{V}_{l} \land \text{end}(w) \in \mathcal{V}_{m} \}|}{|\{\text{edge } w: \text{start}(w) \in \mathcal{V}_{l}\}|}
\end{equation}
where an edge is a member of $\mathcal{E}$ and not a self loop, $\mathcal{V}_{l}$ is the set of nodes that belong to the $l$-th class. Specifically, $p_{lm}(n)$ is the probability that a node~$v$ belongs to class~$l$ given that $v$ is a direct neighbor of a node of class~$l$.
For a dataset, we define its \textbf{edge entropy} as:
\begin{equation}\label{edgeEntropyFull}
\widehat{H} := \sum_{m \in \{1, \hdots, M\}} H(m)w_{m}
\end{equation}
where $w_{m}$ is the percentage of samples from class $m$. That is, the ratio between the number of class $m$ nodes and the number of total nodes.
For instance, we can analyze the edge entropies of the graphs in our motivating examples. The two graphs in Fig. \ref{figure:graphs}(a) and Fig. \ref{figure:graphs}(c) have $\widehat{H}=0$. The graph in Fig. \ref{figure:graphs}(b) and Fig. \ref{figure:graphs}(d) have $\widehat{H}=1$, because the edges are random. This shows that edge entropy is a good indicator of the effectiveness of graph structures for simple graph structures.
\section{Experimental Evaluation}
In this section, we show the correlation between edge entropy and accuracy gains of GNNs from accounting the underlying graph structures using both synthetic and real datasets. We also compare edge entropy with other parameters such as the clustering coefficient and the intra-class ratio. We will start by explaining the datasets and experimental setup, then we will discuss the results from graph CNN experiments.
\subsection{Synthetic Datasets}
We generate synthetic datasets with specific edge entropies. To accomplish this, we use the following approach.
Let $N$ be the number of nodes. Let $M$ be the number of classes. Let $r_i$, $i = 1,\hdots,M$ be the number of nodes with class $i$ where $\sum_{i=1}^{M}r_{i}=N$. Let $\rho$ be a sparsity factor, $0\leq \rho \leq 1$.
In order to have a specific edge entropy, the generated connected graph has to have a specific number of edges between nodes of each class. We create a $M \times M$ matrix $T$ with the desired edge entropy where $T_{i,j}$ is the number of edges that connect a class $i$ node to a class $j$ node. We normalize each row of $T$ to produce a probability matrix $P$. $P$ is the matrix of $p_{lm}$ defined for each pair of classes in \eqref{connProb}. Using the values in $P$ as $p_{lm}$ in \eqref{edgeEntropy} and \eqref{edgeEntropyFull} yields the desired edge entropy $\widehat{H}$.
For example, to get an edge entropy of approximately $\widehat{H} = 0.24$ with $M = 2$ classes with $N = 100$ nodes: $r_1 = 50$ class 1 nodes and $r_2 = 50$ class 2 nodes, a possible $T$ and $P$ are $T = \begin{bmatrix} 48 & 2 \\ 2 & 48 \end{bmatrix}$, $P = \begin{bmatrix} 0.96 & 0.04 \\ 0.04 & 0.96 \end{bmatrix}$\footnote{From \eqref{edgeEntropy} and \eqref{edgeEntropyFull}, ${H(0) = H(1) = -(0.96 \log_2{0.96} + 0.04\log_2{0.04})}$ $ = 0.24$ and $\widehat{H}=0.24 \times 0.5+0.24 \times 0.5=0.24$}.
To create the graph, we start with $N$ isolated nodes. Each node has a self loop and random features. A label is assigned to each node, based on the distribution of nodes by class, $r_i$, $i = 1,\hdots,M$.
We then consider every pair of nodes $v_i$, $v_j$, $i,j = 1,\hdots,N$. For each pair of nodes, we create a directed edge from $v_i$ to $v_j$ with probability $\rho P_{l,m}$ where $l$ is the class of $v_i$ and $m$ is the class of $v_j$.
\subsection{Other Datasets}
We used a variety of popular node classfication datasets :
\begin{itemize}
\item Citation network datasets including CORA-ML, CITESEER and PUBMED \cite{node_dataset}. Nodes are scientific papers and edges represent citations between papers. Data defined on each node is a bag of words vector. Labels represent the field of study.
\item Social network datasets such as REDDIT \cite{reddit}. Nodes represent users' posts and an edge is drawn between two posts if the same user commented on both of the posts. There is no data defined on each node. Labels represent which subreddit a post belongs to.
\item Large synthetic datasets such as SBM-PATTERN and SBM-CLUSTER. Each dataset consists of 10,000 randomly generated graphs using Stochastic Block Models. They were introduced as a benchmark dataset for evaluating GNNs in \cite{dwivedi2020benchmarkgnns}.
\end{itemize}
\subsection{Experimental Setup}
For each dataset, we trained a GNN model using both the underlying graph structure and the identity matrix. We then compare the difference between the test accuracy when we train with the underlying graph structure and the test accuracy when we train with the identity matrix. We call this difference in accuracy the \textbf{improvement} on the dataset from accounting for the graph structure.
\subsubsection{Synthetic Datasets} We generate the graphs using the following parameters: $N=3000, M=3, r_{1}=r_{2}=r_{3}=1/3$. For comparison purposes, we choose two sparsity factor $\rho_{1}=0.1, \rho_{2}=0.5$. The synthetic graph generated with $p_{1}$ is sparse, and a graph generated with $p_{2}$ is dense. We also fix two connection probability matrices such that one wo;; have low edge entropy ($\widehat{H}_{\text{low}}\approx 0.52$), and the other one will have high edge entropy ($\widehat{H}_{\text{high}}\approx 0.97$). Specifically,
\[P_{\text{low}} = \begin{bmatrix}
0.8 & 0.05 & 0.15\\
0.05 & 0.9 & 0.05 \\
0.27 & 0.03 & 0.7
\end{bmatrix}, \,
P_{\text{high}} = \begin{bmatrix}
0.4 & 0.26 & 0.34\\
0.2 & 0.5 & 0.3 \\
0.33 & 0.31 & 0.37
\end{bmatrix}
\]
Dataset dense\_low is generated with $\rho_{2}, P_{\text{low}}$. Dataset sparse\_low is generated with $\rho_{1}, P_{\text{low}}$. Dataset dense\_high is generated with $\rho_{2}, P_{\text{high}}$. Dataset sparse\_high is generated with $\rho_{1}, P_{\text{high}}$. For each dataset, we run 100 Monte Carlos trials by training a randomly initialized 2-layer TAGCN with 2nd order filters for 200 epochs, and testing with cross validation. We vary our percentage of training data from 10\% to 90\% with a 10\% increment each step.
\subsubsection{Other Datasets}
For citation network datasets, we trained a 2-layer TAGCN with polynomial order 3 and 16 hidden units using cross validation. We used an Adam optimizer with a learning rate of 0.01 and a learning rate decay factor of 5e-4.
For the Reddit dataset, we trained a 2-layer Graph Attention Network (GAT) \cite{Velickovic2018}, because we need the sampling function to handle the large size of the dataset. We used 8 heads in the first layer and 1 head in the second layer, a dropout rate of 0.6 for both layers, 200 epochs, and an Adam optimizer with a learning rate of 5e-3 and a learning rate decay factor of 5e-4. For comparison purposes we also trained the same model on citation network datasets.
For the SBM datasets, we trained a 1-layer GCN with 146 hidden units and 1000 learning epochs. We used an Adam optimizer with a learning rate of 1e-3 and a learning rate reduce factor of 0.5.
\begin{figure}[h!]
\centering
\subfloat[The upper blue line represents the test accuracy on dataset sparse\_low. The lower orange line represents the test accuracy on dataset sparse\_high.]{%
\includegraphics[clip,width=0.7\columnwidth]{images/syna.png}%
}
\subfloat[The upper blue line represents the test accuracy on dataset dense\_low. The lower orange line represents the test accuracy on dataset dense\_high]{%
\includegraphics[clip,width=0.7\columnwidth]{images/synb.png}%
}
\caption{Test accuracy of TAGCN on synthetic datasets. The red dashed line shows the improvement from accounting the graph structure is around 20\% with 30\% training samples for both sparsity factors. The shaded regions represent standard deviation of the Monte Carlos trial outcomes. This shows that edge entropy is a good indicator of the effectiveness of the graph structure in synthetic datasets.}
\label{figure:syn_acc}
\end{figure}
\begin{table}[h!]
\centering
\caption{Accuracy Improvements with 2-Layer TAGCN for Synthetic Datasets with 30 Percent Training Data}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
Dataset & Clustering & Intra & $\widehat{H}$ & Improvement\\ \hline
dense\_low & \textbf{0.45} & \textbf{0.8} & \textbf{0.521} & \textbf{51.3} \\
\hline
sparse\_low & 0.121 & \textbf{0.8} & \textbf{0.521} & 36.2 \\
\hline
dense\_high & 0.31 & 0.42 & 0.974 & 32 \\
\hline
sparse\_high & 0.077 & 0.42 & 0.974 & 19.2 \\
\hline
\end{tabular}
\label{table:synth}
\end{table}
\begin{table}[!htbp]
\centering
\caption{Accuracy Improvements with 2-layer TAGCN for Classifying Citation Networks}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
Dataset & Clustering & Intra & $\widehat{H}$ & Improvement\\ \hline
CORA-ML & 0.242 & \textbf{0.81} & $\textbf{0.390}$ & $\textbf{63.9}$\\
\hline
CiteSeer & 0.144 & 0.74 & 0.533 & 54.9 \\
\hline
Pubmed & 0.066 & 0.80 & 0.564 & 51 \\
\hline
\end{tabular}
\label{table:citations}
\end{table}
\begin{table}[!htbp]
\centering
\caption{Accuracy Improvements with 2-Layer GAT for Social Networks and Citation Networks}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
Dataset & Clustering & Intra & $\widehat{H}$ & Improvement\\ \hline
Reddit & \textbf{0.32} & 0.756 & $\textbf{0.322}$ & $\textbf{45.1}$ \\
\hline
CORA-ML & 0.242 & \textbf{0.81} & 0.390 & 23 \\
\hline
CiteSeer & 0.144 & 0.74 & 0.533 & 13.5 \\
\hline
PubMed & 0.066 & 0.80 & 0.564 & 7.4 \\
\hline
\end{tabular}
\label{table:gat}
\end{table}
\begin{table}[!htbp]
\centering
\caption{Accuracy Improvements with 1-Layer GCN for SBM Datasets}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Dataset & Clustering &Intra & $\widehat{H}$ & Improvement \\ \hline
SBM\_PATTERN & \textbf{0.427} & \textbf{0.589} & \textbf{0.811} & \textbf{34.78} \\
\hline
SBM\_CLUSTER & 0.317 &0.33 & 0.954 & 27.68 \\
\hline
\end{tabular}
\label{table:sbm}
\end{table}
\subsection{Discussions}
The results with both synthetic and real datasets show edge entropy is a good indicator of the effectiveness of accounting the graph structure in node classification. For example, Fig.~\ref{figure:syn_acc} shows that the synthetic dataset with low edge entropy always have higher improvement than the synthetic dataset with high edge entropy for both sparsity factors, $\rho_1 = 0.1, \rho_2 = 0.5$. Datasets dense\_low (Table~\ref{table:synth}), CORA-ML (Table~\ref{table:citations}), Reddit (Table~\ref{table:gat}) and SBM\_PATTERN (Table~\ref{table:sbm}) also have the lowest edge entropy and the highest improvement in their tables. This shows that a lower edge entropy corresponds to a higher improvement in test accuracy, and a higher edge entropy corresponds to a lower improvement in test accuracy.
Other parameters such as the clustering coefficient and the intra-class ratio are not consistent with the impact of accounting for graph structures in certain cases. For instance, the clustering coefficient for dataset sparse\_low in Table \ref{table:synth} is 0.121. This is lower than the clustering coefficient for dense\_high (0.31) and does not reflect the better improvement of sparse\_low (51.3\%) over dense\_high (32\%). Another example is the intra-class ratio for Reddit in Table \ref{table:gat}. The intra-class ratio for Reddit is 0.756, which is higher than the intra-class ratio of CORA-ML (0.242). This is inconsistent with Reddit's higher improvement (45.1\%) from accounting the graph structure than CORA-ML (23\%).
The underlying reason is that the clustering coefficient ignores node labels and the intra-class ratio neglects connections between different classes, as we discussed in the motivating examples.
Another interesting observation is that while the clustering coefficient is inconsistent with the improvements on synthetic datasets, a high clustering coefficient indicates a higher improvement for the real datasets studied in this paper. A possible explanation is that these problems are similar to node clustering. This might help explain why a method as simple as label propagation \cite{labelprop} can achieve state-of-the-art accuracy on some popular node classification datasets.
\section{Conclusion}
Evaluating the effectiveness of GNN over CNN on node classification datasets is an important issue, because it shows advantages of using a graph structure. In this paper, we defined what makes the graph structure effective for node classification, and we proposed edge entropy as a parameter to quantitatively evaluate the effectiveness of graph structures. We showed edge entropy is a good indicator of the effectiveness of the graph with experiment results from both synthetic and real datasets. Alternative parameters such as the clustering coefficient and the intra-class ratio are inconsistent with the accuracy gains from accounting the graph structure in some cases. In future works, we will extend the definition of edge entropy to consider longer walks of lengths greater than 1, and we will improve our simulations with other graph models such as small-world graphs and preferential attachments.
\section{Acknowledgments}
The authors thank Xujin Chris Liu, Srinivasa Pranav, Rajshekhar Das, Chaojing Duan for helpful discussions.
\bibliographystyle{IEEEtran}
| e6d7189668a91f609d6145d964d44176e83e1ce9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Device-independent quantum key distribution (DIQKD) is a cryptographic concept based on the observation that if some quantum devices violate a Bell inequality, then it is possible to distill a secret key from the devices' outputs, even when the devices are not fully characterized~\cite{BHK05,PAB+09,Sca12}. This is a stronger form of security than that offered by standard QKD protocols, which assume that the devices are performing measurements within specified tolerances~\cite{SBC+09}.
In recent years, experimental and theoretical developments have brought the possibility of a physical DIQKD demonstration closer to fruition. In particular, a series of experiments have achieved Bell inequality violations while closing the fair-sampling loophole, using NV-centre~\cite{HBD+15}, photonic~\cite{GMR+13,CMA+13,SMC+15,GVW+15,LZL+18,SLT+18}, and cold-atom~\cite{RBG+17} implementations.
On the theoretical front, several protocol modifications have been explored that improve the asymptotic keyrates and noise tolerance of DIQKD.
We focus on those studied in~\cite{HST+20} (\emph{noisy preprocessing}, in which a small amount of trusted noise is added to the device outputs),~\cite{SGP+21} (\emph{random key measurements}, in which more than one measurement basis is used to generate the key, preventing an adversary from optimally attacking both bases simultaneously), and~\cite{WAP21,SBV+21} (\emph{modified CHSH inequalities}, which certify more entropy than the standard CHSH inequality).
Furthermore, earlier device-independent security proofs~\cite{PAB+09} required the assumption that the device behaviour across multiple rounds is independent and identically distributed (IID), sometimes referred to as the assumption of \emph{collective attacks}. With the development of a result known as the \emph{entropy accumulation theorem} (EAT)~\cite{DFR20,ARV19,DF19}, this assumption can now\footnote{To be precise, other proof techniques~\cite{NPS14,VV14,JMS20} are available that do not require the IID assumption, but the asymptotic keyrates given by those techniques are lower than that of the EAT.} be removed for some DIQKD protocols --- security proofs based on the EAT are valid against general attacks (sometimes referred to as \emph{coherent attacks}), as long as the device behaviour can be modelled in a sequential manner. The EAT also provides explicit bounds on the finite-size behaviour, hence concretely addressing the question of what sample size is required for a secure demonstration of DIQKD. (Another recent result that can be applied to obtain such bounds against general attacks is the \emph{quantum probability estimation} technique~\cite{ZKB18}, but in this work we focus on the EAT.)
However, thus far there has been no comprehensive analysis which simultaneously encompasses the various approaches mentioned above. In this work, we present a finite-size security proof that can be applied to protocols which incorporate all of these proposed improvements. In addition, we show that such protocols can achieve substantially higher depolarizing noise tolerance than all previous proven results. We now highlight the main contributions of our work.
\subsection{Summary of key results}
We consider a protocol based on one-way error correction (Protocol~\ref{prot:DIQKD}) that combines noisy preprocessing and random key measurements, and we perform a finite-size analysis against general attacks (Theorem~\ref{th:DIQKD}) as well as collective attacks (Theorem~\ref{th:collective}). Our approach is similar to that used in~\cite{ARV19}; however, we slightly modify the analysis in order to relax the theoretical requirements for the error-correction step, and our bounds are tighter since we use an updated version~\cite{DF19,LLR+19} of the EAT.
Roughly speaking, the proofs rely on lower bounds on the asymptotic keyrates, and hence in Sec.~\ref{sec:1rndbnd} we also describe an algorithm to compute such bounds, which can be applied independently of our finite-size analysis. This algorithm improves over previous results in~\cite{HST+20,SGP+21,WAP21,SBV+21} by having all of the following properties simultaneously: it applies to arbitrary 2-input 2-output protocols, it accounts for noisy preprocessing and random key measurements, and it provably converges to a tight bound\footnote{In fact, similar to~\cite{WAP21}, we find numerical evidence that the bound in~\cite{SGP+21} is not entirely tight, and also that a conjecture proposed there regarding Eve's optimal attack may not be true after all; see Sec.~\ref{sec:algorithmsummary}. } (for protocols of this form).
\begin{figure}
\centering
\subfloat[\cite{HBD+15} parameters]{
\includegraphics[width=0.48\textwidth]{fig_finiteDelft}
}
\subfloat[\cite{RBG+17} parameters]{
\includegraphics[width=0.48\textwidth]{fig_finiteMunich}
}
\caption{Finite-size keyrates as a function of number of rounds in Protocol~\ref{prot:DIQKD}
without noisy preprocessing ($\p=0$),
for honest devices following the estimated parameters in~\cite{MvDR+19} for the Bell tests in~\cite{HBD+15} and~\cite{RBG+17}.
The solid curves show the results for general attacks (Theorem~\ref{th:DIQKD}), while the dashed curves show the results under the assumption of collective attacks (Theorem~\ref{th:collective}), with the error-correction protocol taken to satisfy Eqs.~\eqref{eq:optEC}--\eqref{eq:ECrate}. The colours correspond to soundness parameters (informally, a measure of how ``insecure'' the key is) of $\eps^\mathrm{sou}=10^{-3}$, $10^{-6}$, and $10^{-9}$ for black, blue, and red respectively, while the completeness parameter (the probability that the honest devices abort) is $\eps^\mathrm{com}=10^{-2}$ in all cases. The horizontal line denotes the asymptotic keyrate.
All other parameters in Theorems~\ref{th:DIQKD} and \ref{th:collective} were numerically optimized, except $\beta$.
}
\label{fig:experiments}
\vspace{2cm}
\includegraphics[width=0.6\textwidth]{fig_finitedepol}
\caption{Finite-size keyrates secure against general attacks (Theorem~\ref{th:DIQKD}) as a function of number of rounds in Protocol~\ref{prot:DIQKD}
without noisy preprocessing ($\p=0$), where the honest devices are described by depolarizing noise $\q$ (see Sec.~\ref{sec:hon}). The solid, dashed and dotted curves denote $\q=5\pct$, $6\pct$ and $7\pct$ respectively, with the error-correction protocol taken to satisfy Eqs.~\eqref{eq:optEC}--\eqref{eq:ECrate}. The soundness parameter is $\eps^\mathrm{sou}=10^{-6}$ and the completeness parameter is $\eps^\mathrm{com}=10^{-2}$. The horizontal lines denote the asymptotic keyrates. All other parameters were numerically optimized.
}
\label{fig:depol}
\end{figure}
For simplicity, the protocol we present only uses the CHSH inequality rather than the modified CHSH inequalities of~\cite{WAP21,SBV+21}, because our results suggest (see Sec.~\ref{sec:bestbnd}) that within the 2-input 2-output scenario, there may not be much prospect for improvement by using the latter, at least for the depolarizing-noise model. However, the finite-size analysis we perform can (like~\cite{BRC20}) be applied to protocols based on arbitrary Bell inequalities --- in Sec.~\ref{sec:nonCHSH}, we briefly explain the relevant adjustments in that case. Similarly, while we only performed explicit computations of the asymptotic keyrates for CHSH-based protocols, our approach as described in Sec.~\ref{sec:1rndbnd} can be applied to all 2-input 2-output Bell inequalities.
In Figs.~\ref{fig:experiments} and~\ref{fig:depol}, we plot some finite-size keyrates given by our security proofs. Fig.~\ref{fig:experiments} corresponds to honest devices with performance described by the estimated parameters in~\cite{MvDR+19} for the NV-centre~\cite{HBD+15} and cold-atom~\cite{RBG+17} loophole-free Bell tests (details in Sec.~\ref{sec:hon}). (For photonic experiments, it is currently unclear whether the protocol here provides improvements over the results in~\cite{HST+20,WAP21,SBV+21} --- there are some complications which we explain in Sec.~\ref{sec:conclusion}.)
Fig.~\ref{fig:depol} shows the finite-size keyrates for honest devices subject to depolarizing noise.
From Fig.~\ref{fig:experiments}, we see that our security proof requires the \cite{HBD+15} and \cite{RBG+17} experiments to run for approximately $n\sim10^8$ and $10^{10}$ rounds respectively in order to certify a positive finite-size keyrate against general attacks. While this is a marked improvement over the basic \cite{PAB+09}~protocol (which yields zero asymptotic keyrate for those experiments), these finite-size requirements still appear to be outside the reach of those experimental implementations.
We also see that changing the security parameters (see Sec.~\ref{sec:secdef} for details on their interpretation) by several orders of magnitude only results in fairly small changes to the keyrate, so it appears unlikely that the keyrates could be substantially improved by relaxing these security requirements.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{fig_keyrateslog}
\caption{
Lower bounds on asymptotic keyrates under depolarizing noise $\q$ (defined in Sec.~\ref{sec:hon}), on a vertical log scale. The black, blue, and red solid curves show the (net) asymptotic keyrate of Protocol~\ref{prot:preshared}, for noisy-preprocessing values of $\p=0$, $0.2$ and $0.3$ respectively, based on the bounds we computed (see Eq.~\eqref{eq:certbnds}, which also shows our results for some other values of $\p$)\protect\footnotemark.
For comparison, the dashed curves show the corresponding asymptotic keyrates of the protocol in~\cite{HST+20}, which does not use the random-key-measurement method (the $\p=0$ case is equivalent to the~\cite{PAB+09} protocol). The solid curves intersect the horizontal axis at $\q=8.39\pct$, $9.26\pct$ and $9.33\pct$, in order of increasing $\p$. The first value is a minor improvement over~\cite{SGP+21} (despite being effectively the same protocol), likely because our algorithm provably converges to a tight keyrate bound (for fixed $\p$). The last value exceeds all previous bounds~\cite{HST+20,SGP+21,WAP21,SBV+21} for depolarizing-noise tolerance, by an amount comparable to their respective improvements over the original~\cite{PAB+09} protocol. It is also close to the upper bound of $9.57\pct$ that we derive in Sec.~\ref{sec:bestbnd}.}
\label{fig:keyrates}
\end{figure}
\footnotetext{The asymptotic keyrates of Protocol~\ref{prot:DIQKD} would be given by simply halving these keyrates, which does not change their horizontal intercepts.}
To improve on these results, in Sec.~\ref{sec:mods} we describe two modifications of Protocol~\ref{prot:DIQKD}. Firstly, Protocol~\ref{prot:preshared} is a modification based on a pre-shared key, which achieves a \emph{net} key generation rate approximately double that of Protocol~\ref{prot:DIQKD}, by overcoming a crucial disadvantage of random-key-measurement protocols (namely, the sifting factor). Secondly, Protocol~\ref{prot:optcoll} is a modification that is optimized for the collective-attacks assumption --- by changing the protocol itself for this scenario rather than just the security proof, we were able to further improve the keyrate.
However, we find that even with these modifications and relaxed security requirements, it appears that the required number of rounds for a positive finite-size keyrate is still impractically large (see Fig.~\ref{fig:protIID} in Sec.~\ref{sec:mods}).
As another immediate consequence of our work, we obtain a lower bound on the asymptotic noise tolerance of our protocols against depolarizing noise. While noisy preprocessing was not applied for the plots in Fig.~\ref{fig:experiments} (because it appears to only slightly improve the keyrates for those experimental parameters; see Fig.~\ref{fig:protIID} later), our security proof indeed allows for this option, and in Fig.~\ref{fig:keyrates} we plot the asymptotic keyrates certified by our algorithm for some noisy-preprocessing values up to $\p=0.3$ (our heuristic estimates indicate that further increasing $\p$ only yields small improvements to the noise threshold; see Sec.~\ref{sec:bestbnd}).
As seen from Fig.~\ref{fig:keyrates}, our results certify that positive asymptotic keyrates are possible for depolarizing-noise values of up to $9.33\pct$ at least. In comparison, the previous best thresholds were $8.34\pct$~\cite{WAP21} (using noisy preprocessing and modified CHSH inequalities) or $8.2\pct$~\cite{SGP+21} (using the random-key-measurement approach). Our improvement over these results is hence of similar magnitude to their improvements over the basic~\cite{PAB+09} protocol, which has a threshold of $7.15\pct$.
Furthermore, our bounds are in fact close to the highest possible bounds allowed by simple convexity arguments --- we prove in Sec.~\ref{sec:bestbnd} that all protocols of this general form cannot achieve a depolarizing-noise threshold beyond $9.57\pct$, so the threshold we obtained is not far from the absolute highest possible value in this setting.
Finally, we remark that several results from this work can also be used for DI randomness expansion (DIRE)~\cite{arx_colbeckthesis,PAM+10,BRC20,LLR+19}. Specifically, the algorithm we use for computing the asymptotic DIQKD keyrates can also be used to obtain tighter bounds on the asymptotic keyrates for DIRE, after which a finite-size analysis could be performed using the EAT~\cite{BRC20,LLR+19}. Moreover, the ``key recovery'' process in the modified protocol (Protocol~\ref{prot:preshared}) can be applied in the context of DIRE, hence allowing one to improve the keyrates by using the random-key-measurement approach of~\cite{SGP+21}. We discuss this in detail in Sec.~\ref{sec:preshared}.
\subsection{Paper structure}
In Sec.~\ref{sec:prelim}, we state the definitions and notation we use in this work. In Sec.~\ref{sec:prot}, we describe the main protocol we consider, and state the main theorem which bounds the finite-size keyrates (Theorem~\ref{th:DIQKD}). We give the security proof for this finite-size bound in Sec.~\ref{sec:finite}. In Sec.~\ref{sec:1rndbnd} we present the algorithm to compute the asymptotic keyrates (i.e.~the leading-order terms in Theorem~\ref{th:DIQKD}), and describe the resulting depolarizing-noise thresholds as well as an upper bound on these thresholds. Finally, in Sec.~\ref{sec:mods} we discuss several variations, such as a modified random-key-measurement protocol which bypasses the sifting factor, and security proofs against collective attacks instead of general attacks.
\section{Preliminaries}
\label{sec:prelim}
We define some basic notation in Table~\ref{tab:notation}, and state some further definitions below.
We take all systems to be finite-dimensional, but we will not impose any bounds on the system dimensions unless otherwise specified.
\newpage
\begin{table}
\caption{List of notation}
\def1{1.5}
\setlength\tabcolsep{.28cm}
\centering
\begin{tabular}{c l}
\toprule
\textit{Symbol} & \textit{Definition} \\
\toprule
$\log$ & Base-$2$ logarithm \\
\hline
$H(\cdot)$ & Base-$2$ von Neumann entropy \\
\hline
$D(\cdot \Vert \cdot)$ & Base-$2$ quantum relative entropy \\
\hline
$\norm{\cdot}_p$ & Schatten $p$-norm \\
\hline
$\floor{\cdot}$ (resp.~$\ceil{\cdot}$) & Floor (resp.~ceiling) function \\
\hline
$X\geq Y$ (resp.~$X>Y$) & $X-Y$ is positive semidefinite (resp.~positive definite)\\
\hline
$\dop{=}(A)$ (resp.~$\dop{\leq}(A)$) & Set of normalized (resp.~subnormalized) states on register $A$ \\
\hline
$\idk_A$ & Maximally mixed state on register $A$
\\
\hline
$\upto{n}$ & Indices from $1$ to $n$, i.e.~$\{1,2,\dots,n\}$ \\
\hline
$A_{\upto{n}}$ & Registers $A_1 \dots A_n$ \\
\hline
$\no{\Omega}$ & Complement (i.e.~negation) of an event $\Omega$ \\
\hline
$\operatorname{freq}$ & See Definition~\ref{def:freq} \\
\hline
$\cdfBin{n}{p}{k}$ & See Definition~\ref{def:binom} \\
\hline
$\rho_{\land\Omega}$ and $\rho_{|\Omega}$ & See Definition~\ref{def:cond} \\
\toprule
\end{tabular}
\def1{1}
\label{tab:notation}
\end{table}
\begin{definition}\label{def:freq}
(Frequency distributions) For a string $\str{z}\in\mathcal{Z}^n$ on some alphabet $\mathcal{Z}$, $\operatorname{freq}_{\str{z}}$ denotes the following probability distribution on $\mathcal{Z}$:
\begin{align}
\operatorname{freq}_{\str{z}}(z) \defvar \frac{1}{n} \sum_{j=1}^{n} \delta_{z,z_j}.
\end{align}
\end{definition}
\begin{definition}\label{def:binom}
(Binomial distribution) Let $X\sim\operatorname{Binom}(n,p)$ denote a random variable $X$ following a binomial distribution with parameters $(n,p)$, i.e.~$X$ is the sum of $n$ IID Bernoulli random variables $X_j$ with $\pr{X_j=1}=p$.
We denote the corresponding cumulative distribution function as
\begin{align}
\cdfBin{n}{p}{k} \defvar \pr[X\sim\operatorname{Binom}(n,p)]{X \leq k}.
\end{align}
\end{definition}
\begin{definition}\label{def:cond}
(Conditioning on classical events) For a classical-quantum state $\rho \in \dop{\leq}(CQ)$ in the form
$\rho_{CQ} = \sum_c \pure{c} \otimes \omega_c$
for some $\omega_c \in \dop{\leq}(Q)$,
and an event $\Omega$ defined on the register $C$, we define the following ``conditional states'':
\begin{align}
\rho_{\land\Omega} \defvar \sum_{c\in\Omega} \pure{c} \otimes \omega_c, \qquad\qquad \rho_{|\Omega} \defvar \frac{\tr{\rho}}{\tr{\rho_{\land\Omega}}} \rho_{\land\Omega} = \frac{
\sum_{c} \tr{\omega_c}
}{\sum_{c\in\Omega} \tr{\omega_c}} \rho_{\land\Omega} .
\end{align}
We informally refer to these states as the subnormalized and normalized conditional states respectively (the latter is perhaps a slight misnomer if $\tr{\rho}<1$, but this situation does not arise in our proofs). The process of taking subnormalized conditional states is commutative and ``associative'', in the sense that for any events $\Omega,\Omega'$ we have $(\rho_{\land\Omega})_{\land\Omega'} = (\rho_{\land\Omega'})_{\land\Omega} = \rho_{\land(\Omega\land\Omega')}$; hence for brevity we will denote all of these expressions as
\begin{align}
\rho_{\land\Omega\land\Omega'} \defvar (\rho_{\land\Omega})_{\land\Omega'} = (\rho_{\land\Omega'})_{\land\Omega} = \rho_{\land(\Omega\land\Omega')}.
\end{align}
On the other hand, some disambiguating parentheses are needed when combined with taking normalized conditional states.
\end{definition}
\begin{definition}
(2-universal hashing) A \emph{2-universal family of hash functions} is a set $\mathcal{H}$ of functions from a set $\mathcal{X}$ to a set $\mathcal{Y}$, such that if $h$ is drawn uniformly at random from $\mathcal{H}$, then
\begin{align}
\pr{h(x) = h(x')} \leq \frac{1}{|\mathcal{Y}|} \qquad \forall x\neq x'.
\end{align}
\end{definition}
To ensure consistency of definitions, we now state the definitions of smoothed entropies relevant for this work, though we will not need to use the explicit expressions. We follow the presentation in~\cite{DFR20,DF19}, which can be shown to be equivalent to the definitions in~\cite{Tom16,TL17}.
\begin{definition}
For $\rho,\sigma \in \dop{\leq}(A)$, the \emph{generalized fidelity} is
\begin{align}
F(\rho,\sigma) \defvar \norm{
\sqrt{\rho}\sqrt{\sigma}
}_1 + \sqrt{(1-\tr{\rho})(1-\tr{\sigma})},
\end{align}
and the \emph{purified distance} is $\pd(\rho,\sigma)\defvar\sqrt{1-F(\rho,\sigma)^2}$.
\end{definition}
\begin{definition}
For $\rho\in\dop{\leq}(AB)$, the \emph{min- and max-entropies of $A$ conditioned on $B$} are
\begin{align}
H_\mathrm{min}(A|B)_\rho &\defvar
-\log
\min_{\substack{\sigma \in \dop{\leq}(B) \suchthat\\ \ker(\rho_B)\subseteq\ker(\sigma_B)}}
\norm{\rho_{AB}^\frac{1}{2}
(\id_A \otimes \sigma_{B})
^{-\frac{1}{2}}}_\infty^2,
\\
H_\mathrm{max}(A|B)_\rho &\defvar \log
\max_{\sigma \in \dop{\leq}(B)}
\norm{\rho_{AB}^\frac{1}{2}
(\id_A \otimes \sigma_{B})
^\frac{1}{2}}_1^2,
\end{align}
where in the first equation the
$(\id_A \otimes \sigma_{B})
^{-\frac{1}{2}}$ term
should be understood in terms of the Moore-Penrose generalized inverse.
In both equations,
the optimum is attained by a normalized state, so $\dop{\leq}(B)$ can be replaced by $\dop{=}(B)$ without loss of generality.
\end{definition}
\begin{definition}
For $\rho\in\dop{\leq}(AB)$ and $\epsilon\in\left[0,\sqrt{\tr{\rho_{AB}}}\right)$, the \emph{$\epsilon$-smoothed min- and max-entropies of $A$ conditioned on $B$} are
\begin{align}
H_\mathrm{min}^\epsilon(A|B)_\rho \defvar
\max_
{\substack{\tilde{\rho} \in \dop{\leq}(AB) \suchthat\\ \pd(\tilde{\rho},\rho)\leq\epsilon}}
H_\mathrm{min}(A|B)_{\tilde{\rho}},
\qquad
H_\mathrm{max}^\epsilon(A|B)_\rho \defvar
\min_
{\substack{\tilde{\rho} \in \dop{\leq}(AB) \suchthat\\ \pd(\tilde{\rho},\rho)\leq\epsilon}}
H_\mathrm{max}(A|B)_{\tilde{\rho}}.
\end{align}
\end{definition}
\subsection{Security definitions}
\label{sec:secdef}
The question of formalizing an appropriate security definition for the device-independent setting has not been definitively resolved yet~\cite{ARV19}, due to considerations regarding the device-reuse attacks described in~\cite{BCK13}. However, we shall proceed by following the security definitions used in~\cite{ARV19}, which were based on strong\footnote{In the sense that the definitions imply \emph{composable security}, which roughly speaking means that we can safely use this protocol in place of a simpler, more idealized functionality that generates an ideal secret key whenever it does not abort --- see~\cite{PR14} for details.} security definitions~\cite{PR14} for standard QKD, and should be sufficient under suitable constraints on the nature of the device memories (we shall briefly discuss these in the next section, when listing the assumptions). Qualitatively, the concepts involved in this definition are: \emph{completeness}, meaning that the honest devices will accept with high probability, and \emph{soundness}, meaning that the devices remain ``secure'' (possibly by aborting) even in the presence of dishonest behaviour. These notions are formalized as follows:
\begin{definition} \label{def:secure}
Consider a DIQKD protocol such that at the end, the honest parties either \emph{accept} (producing keys $K_A$ and $K_B$ of length $\ell_\mathrm{key}$ for Alice and Bob respectively) or \emph{abort} (producing an abort symbol $\perp$ for all parties).
It is said to be
$\eps^\mathrm{com}$-complete and $\eps^\mathrm{sou}$-sound if the following properties hold:
\begin{itemize}
\item (Completeness) The honest protocol implementation aborts with probability at most $\eps^\mathrm{com}$.
\item (Soundness) For any implementation of the protocol, we have
\begin{align}
\pr{\mathrm{accept}} \frac{1}{2} \norm{\sigma_{K_A K_B E'} - \left(\frac{1}{2^{\ell_\mathrm{key}}} \sum_k \pure{kk}_{K_A K_B} \right) \otimes \sigma_{E'}}_1 \leq \eps^\mathrm{sou},
\end{align}
where $\sigma$ denotes the normalized state conditioned on the protocol accepting, and $E'$ denotes all side-information registers available to the adversary at the end of the protocol.
\end{itemize}
\end{definition}
In the security proof, it is convenient to use the fact that the soundness property is implied by a pair of slightly simpler conditions, as shown in~\cite{PR14}. Specifically, to prove a DIQKD protocol is $\eps^\mathrm{sou}$-sound, it suffices to find $\eps^\mathrm{cor}_\mathrm{QKD},\eps^\mathrm{sec}_\mathrm{QKD}$ such that $\eps^\mathrm{sou}\geq\eps^\mathrm{cor}_\mathrm{QKD}+\eps^\mathrm{sec}_\mathrm{QKD}$ and the protocol is both $\eps^\mathrm{cor}_\mathrm{QKD}$-correct and $\eps^\mathrm{sec}_\mathrm{QKD}$-secret, defined as follows:
\begin{definition}
A DIQKD protocol as described above is said to be
$\eps^\mathrm{cor}_\mathrm{QKD}$-correct and $\eps^\mathrm{sec}_\mathrm{QKD}$-secret if
the following properties hold:
\begin{itemize}
\item (Correctness) For any implementation of the protocol, we have\footnote{In~\cite{ARV19}, this definition is stated slightly differently since the accept condition is implicitly absorbed into the $K_A,K_B$ condition by defining $K_A=K_B=\perp$ when the protocol aborts.}
\begin{align}
\pr{K_A \neq K_B \land \mathrm{accept}} \leq \eps^\mathrm{cor}_\mathrm{QKD}.
\end{align}
\item (Secrecy) For any implementation of the protocol, we have
\begin{align}
\pr{\mathrm{accept}} \frac{1}{2} \norm{\sigma_{K_A E'} - \idk_{K_A} \otimes \sigma_{E'}}_1 \leq \eps^\mathrm{sec}_\mathrm{QKD},
\end{align}
where $\sigma$ is as described in Definition~\ref{def:secure}, and $\idk_{K_A}$ denotes the maximally mixed state (i.e.~a uniformly random key for Alice).
\end{itemize}
\end{definition}
\section{Main protocol}
\label{sec:prot}
The overall structure of our main protocol is stated as Protocol~\ref{prot:DIQKD} below, with the details of some steps to be specified in the following subsections.
We make the following fairly standard assumptions, following the presentation in~\cite{ARV19}:
\begin{itemize}
\item Alice and Bob can prevent unwanted information from leaking outside of their respective locations (for instance, the inputs and outputs to the devices remain private for each party until/unless they are revealed in the public communication steps).
\item Alice and Bob can generate trusted (local) randomness.
\item Alice and Bob have trusted post-processing units to perform classical computations.
\item Alice and Bob perform all classical communication using an authenticated public channel.
\item All systems can be modelled as finite-dimensional quantum registers (though we do not impose any bounds on the dimensions unless otherwise specified).
\end{itemize}
Regarding the first point in particular, we remark that while the loophole-free Bell tests in~\cite{HBD+15,SMC+15,GVW+15,RBG+17} used spacelike separation to motivate the assumption that the devices do not reveal their inputs to each other, this may not be strictly necessary for a DIQKD implementation. It could be possible, and perhaps more reasonable, to instead justify this assumption by implementing some ``shielding'' measures on the devices, to prevent them from leaking unwanted information. In any case, ``shielding'' measures of this nature are likely necessary to prevent the raw \emph{outputs} of the devices from leaking to the adversary, so it may be expedient to use these measures to prevent the inputs from leaking as well.
The above assumptions will be sufficient for us to show that the protocol satisfies the completeness and soundness definitions stated above. However, when considering whether these formal definitions yield security in an intuitive (or composable) sense, some additional assumptions are needed to address the memory attack of~\cite{BCK13}. One approach would be to impose the condition that the devices do not access any registers storing ``private'' data from previous protocols, and that this condition continues to hold if the devices are reused in the future. (Note that this condition is always implicit in standard QKD, because the register being measured is inherently specified when describing the trusted measurements the devices perform.) It remains to fully formalize this condition in a suitable framework for composable security, but this would be beyond the scope of this work.
\begin{savenotes}
\begin{breakablealgorithm}
\caption{}
\label{prot:DIQKD}
The protocol is defined in terms of the following parameters (chosen before the protocol begins), which we qualitatively describe:
{\setlist[description]{leftmargin=2cm,labelindent=2cm,itemsep=0mm}
\begin{description}
\item $n$: Total number of rounds
\item $\gamma$: Probability of a test round
\item $\p$: Noisy-preprocessing bias
\item $\mathrm{EC}_\mathrm{max}$: Bound on number of bits used for error correction
\item $w_\mathrm{exp}$: Expected winning probability for the (IID) honest devices
\item $\delta_\mathrm{tol}$: Tolerated deviation from expected winning probability $w_\mathrm{exp}$
\item $\ell_\mathrm{key}$: Length of final key
\end{description}}
\noindent The honest behaviour consists of $n$ IID copies of a device characterized by $w_\mathrm{exp}$ and an error-correction parameter $h_\mathrm{hon}$ (details in Sec.~\ref{sec:hon}).\\
\begin{algorithmic}[1]
\State\label{step:1rnd}\textbf{Measurement:} For each $j \in \upto{n}$, perform the following steps:
\begin{algsubstates}
\State Alice and Bob's devices each receive some share of a quantum state.
\State Alice chooses a uniform input $X_j \in \{0,1\}$. With probability $\gamma$, Bob chooses a uniform input $Y_j \in \{2,3\}$, otherwise Bob chooses a uniform input $Y_j \in \{0,1\}$.
\State Alice and Bob supply their inputs to their devices, and record the outputs as $A_j$ and $B_j$ respectively.
\end{algsubstates}
\State Alice and Bob publicly announce their input strings $\str{X}$ and $\str{Y}$.
\State \textbf{Sifting:} For all rounds such that $Y_j \in \{0,1\}$ and $X_j \neq Y_j$, Alice and Bob overwrite their outputs with $A_j = B_j = 0$.
\State\label{step:NPP}\textbf{Noisy preprocessing:} For all rounds such that $Y_j \in \{0,1\}$ and $X_j = Y_j$, Alice generates a biased random bit $F_j$ with $\pr{F_j=1} = \p$, and overwrites her output $A_j$ with $A_j \oplus F_j$.
\State\label{step:EC}\textbf{Error correction:} Alice and Bob publicly communicate some bits $\str{L}=(\str{L}_\mathrm{EC},\str{L}_\mathrm{h})$
for error correction as follows (see Sec.~\ref{sec:EC}):
\begin{algsubstates}
\State\label{step:guess}Alice and Bob communicate some bits $\str{L}_\mathrm{EC}$ to allow Bob to produce a guess $\tilde{\str{A}}$ for $\str{A}$. If at some point the number of communicated bits reaches $\mathrm{EC}_\mathrm{max}$, Alice and Bob immediately cease communication and proceed to the next step, with Bob producing $\tilde{\str{A}}$ using only the information he has at that point.
\State\label{step:hash}Alice computes a 2-universal hash $\str{L}_\mathrm{h} = \operatorname{hash}(\str{A})$ of length $\ceil{\log(1/\eps_\mathrm{h})}$. She sends $\str{L}_\mathrm{h}$ (and the choice of hash function) to Bob.
\end{algsubstates}
\State\label{step:PA}\textbf{Parameter estimation:}
For all $j\in\upto{n}$, Bob sets
$C_j=\perp$ if $Y_j \in \{0,1\}$; otherwise he sets $C_j=0$ if $\tilde{A}_j \oplus B_j \neq X_j \cdot (Y_j-2)$ and $C_j=1$ if $\tilde{A}_j \oplus B_j = X_j \cdot (Y_j-2)$.
\State\label{step:abort}Bob checks whether $\str{L}_\mathrm{h}=\operatorname{hash}(\tilde{\str{A}})$, as well as whether the value $\str{c}$ on registers $\str{C}$ satisfies $\operatorname{freq}_\str{c}(1)\geq (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma$ and $\operatorname{freq}_\str{c}(0)\leq (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma$. If all those conditions hold, Alice and Bob proceed to the next step. Otherwise, the protocol aborts.
\State \textbf{Privacy amplification:} Alice and Bob apply privacy amplification (see Sec.~\ref{sec:PA}) on $\str{A}$ and $\tilde{\str{A}}$ respectively to obtain final keys $K_A$ and $K_B$ of length $\ell_\mathrm{key}$.
\end{algorithmic}
\end{breakablealgorithm}
\end{savenotes}
The rounds in which $Y_j \in \{2,3\}$ will be referred to as test rounds, and the rounds in which $Y_j \in \{0,1\}$ will be referred to as generation rounds (though strictly speaking, the final key in this protocol is obtained from all the rounds, not merely the generation rounds alone).
In each round,
Eve is allowed to hold some extension of the state distributed to the devices. We will use $E$ to denote the collection of all such quantum side-information she retains over the entire protocol. (We do not denote this using separate registers $E_j$ for individual rounds, because in a general scenario, Eve's side-information may not necessarily ``factorize'' into a tensor product across the individual rounds.)
We briefly highlight some aspects of this protocol that may differ slightly as compared to more commonly used QKD protocols. Firstly, we do not choose a random subset of fixed size as test rounds, but rather, each round is independently chosen to be a test or generation round, following~\cite{ARV19}. This was in order to apply the entropy accumulation theorem, which holds for processes that can be described using a sequence of maps. Furthermore, the parameter-estimation check is performed on \emph{both} $\operatorname{freq}_\str{c}(1)$ and $\operatorname{freq}_\str{c}(0)$. This was required in order to derive a critical inequality in the security proof (following~\cite{BRC20}), though in some cases it is possible to omit the $\operatorname{freq}_\str{c}(1)$ check (see Eq.~\eqref{eq:fminPEineq} and the subsequent discussion).
We now describe some of the individual steps in more detail.
\subsection{Error correction}\label{sec:EC}
We first discuss Step~\ref{step:hash}, because it will have an impact on our discussion of Step~\ref{step:guess}. Given any $\eps_\mathrm{h} \in (0,1]$,
if we consider a 2-universal family of hash functions where the output is a bitstring of length $\ceil{\log(1/\eps_\mathrm{h})}$, then the defining property of 2-universal hashing guarantees that
\begin{align}
\pr{\operatorname{hash}(\str{A})=\operatorname{hash}(\tilde{\str{A}})\middle|\str{A}\neq\tilde{\str{A}}} \leq \eps_\mathrm{h}. \label{eq:hash}
\end{align}
In other words, the probability of getting matching hashes from different strings can be made arbitrarily small, by using sufficiently long hashes. Informally speaking, this gives us some laxity in Step~\ref{step:guess}, because regardless of how much the devices deviate from the honest behaviour, the guarantee~\eqref{eq:hash} will still hold, providing a final ``check'' on how bad the guess $\tilde{\str{A}}$ could be.
Importantly, our later proof of the \emph{soundness} of the protocol will not rely on any guarantees regarding the procedure in Step~\ref{step:guess} --- only the \emph{completeness} of the protocol (i.e.~the probability that the honest devices mistakenly abort) requires such guarantees.
We now study Step~\ref{step:guess}. $\mathrm{EC}_\mathrm{max}$ is defined as follows: it is the length of $\str{L}_\mathrm{EC}$ required such that given the honest devices, Bob can use $\str{L}_\mathrm{EC}$ and $\str{B}$ to produce a guess $\tilde{\str{A}}$
satisfying
\begin{align}
\pr{\str{A} \neq \tilde{\str{A}}}_\mathrm{hon} \leq \eps^\mathrm{com}_\mathrm{EC}.
\label{eq:comEC}
\end{align}
(In this section, we will use the subscript $_\mathrm{hon}$ to emphasize quantities computed with respect to an honest behaviour.)
We stress that while some preliminary characterization of the devices can be performed beforehand to choose a suitable $\mathrm{EC}_\mathrm{max}$, this parameter \emph{must not} be changed once the protocol begins.
This immediately raises the question of what value should be chosen for $\mathrm{EC}_\mathrm{max}$ in order to achieve a desired $\eps^\mathrm{com}_\mathrm{EC}$. Theoretically, there exists a protocol~\cite{RR12}
with one-way communication that achieves~\eqref{eq:comEC} as long as we choose $\mathrm{EC}_\mathrm{max}$
such that
\begin{align}
\mathrm{EC}_\mathrm{max} \geq
H_\mathrm{max}^{\ez}(\str{A}|\str{B}\str{X}\str{Y})_\mathrm{hon} + 2\log\frac{1}{\eps^\mathrm{com}_\mathrm{EC}-\ez} + 4,
\label{eq:optEC}
\end{align}
where $\ez\in[0,\eps^\mathrm{com}_\mathrm{EC})$ is a parameter that can be optimized over.
(This bound is essentially tight for one-way protocols.)
Since the honest behaviour is IID,
the max-entropy can be bounded by using the asymptotic equipartition property (AEP) in the form stated as Corollary~4.10 of~\cite{DFR20}, which yields\footnote{In this analysis, we deviated slightly from~\cite{ARV19} by using the error-correction protocol from~\cite{RR12} instead of~\cite{RW05}, and the AEP stated in~\cite{DFR20} rather than~\cite{TCR09}. Both of these yield slight improvements in the bounds (the former at large $n$, and the latter when
$\dim(A_j)$ is not too large).
}
\begin{align}
H_\mathrm{max}^{\ez}(\str{A}|\str{B}\str{X}\str{Y})_\mathrm{hon} \leq n h_\mathrm{hon} +
\sqrt{n} \, (2\log5)\sqrt{\log\frac{2}{\ez^2}},
\label{eq:HmaxAEP}
\end{align}
where (using the decomposition $H(Q|Q'C)=\sum_c \pr{c} H(Q|Q';C=c)$ for classical $C$)
\begin{align}
h_\mathrm{hon} \defvar H(A_j|B_jX_jY_j)_\mathrm{hon}
&= \frac{1-\gamma}{4}
\sum_{z\in\{0,1\}} H(A_j | B_j;X_j = Y_j = z)_\mathrm{hon}
\nonumber \\
& \qquad
+ \frac{\gamma}{4}
\sum_{x\in\{0,1\},y\in\{2,3\}}
H(A_j| B_j ;X_j = x, Y_j = y)_\mathrm{hon},
\label{eq:ECrate}
\end{align}
where the terms in the summation over $z$ should be understood to refer to the $A_jB_j$ values \emph{after} the noisy-preprocessing step. (Any value of $j$ can be used in the above equation since the honest behaviour is IID.)
However, the protocol achieving the bound~\eqref{eq:optEC} may not be easy to implement. In practice, error-correction protocols typically achieve performance described by
\begin{align}
\mathrm{EC}_\mathrm{max} \approx \xi(n,\eps^\mathrm{com}_\mathrm{EC}) n h_\mathrm{hon},
\end{align}
where $\xi(n,\eps^\mathrm{com}_\mathrm{EC})$ lies between $1.05$ and $1.2$ for ``typical'' values of $n$ and $\eps^\mathrm{com}_\mathrm{EC}$. (More precise characterizations can be found in~\cite{TMP+17}, which gives for instance an estimate
\begin{align}
\mathrm{EC}_\mathrm{max} \approx \xi_1 n h_\mathrm{hon} + \tilde{\xi}(\eps^\mathrm{com}_\mathrm{EC},h_\mathrm{hon})\sqrt{n},
\end{align}
for a constant $\xi_1$ and a specific function $\tilde{\xi}$.) Furthermore, some protocols used in practice do not have a theoretical bound on $\eps^\mathrm{com}_\mathrm{EC}$ (for a given $\mathrm{EC}_\mathrm{max}$), only heuristic estimates.
Fortunately, as mentioned earlier, the choice of error-correction procedure in Step~\ref{step:guess} will have no effect in our proof of the soundness of Protocol~\ref{prot:DIQKD} (as long as $\mathrm{EC}_\mathrm{max}$ is a fixed parameter), only its completeness. This means that as long as we are willing to accept heuristic values for $\eps^\mathrm{com}$, we can use the heuristic values of $\eps^\mathrm{com}_\mathrm{EC}$ provided by using some ``practical'' error-correction procedure in Step~\ref{step:guess}, and the value of $\eps^\mathrm{sou}$ (i.e.~how ``secure'' the protocol is) will be completely unaffected. The critical point to remember is that $\mathrm{EC}_\mathrm{max}$ is a value to be fixed before the protocol begins, and Alice and Bob \emph{must} stop Step~\ref{step:guess} once they have reached that number of communicated bits. With this in mind, we remark that while we mainly focus on protocols using one-way error correction, this is not quite a strict requirement --- in principle, one could use a procedure involving two-way communication (such as CASCADE), as long as $\mathrm{EC}_\mathrm{max}$ includes all the communicated bits, not just those sent from Alice to Bob. Another possibility worth considering might be adaptive procedures that adjust to the noise level encountered during execution of the protocol, rather than the expected noise level (again, making sure to halt once $\mathrm{EC}_\mathrm{max}$ bits are communicated, where $\mathrm{EC}_\mathrm{max}$ is defined beforehand based on the expected behaviour).
We remark that in our situation, the registers $\str{A}\str{B}\str{X}\str{Y}$
have some substructure in the sense that they can be naturally divided into the substrings where $X_j\neq Y_j\in\{0,1\}$, $X_j=Y_j\in\{0,1\}$, and $Y_j\in\{2,3\}$, so the error-correction procedure should ideally take advantage of this substructure (for instance, no error-correction data needs to be sent regarding the rounds where $X_j\neq Y_j\in\{0,1\}$). Also, if we assume that $\pr{A_j = B_j |X_j = x, Y_j = y}_\mathrm{hon}$ is the same for all $x\in\{0,1\},y\in\{2,3\}$ (in which case it must equal $w_\mathrm{exp}$), then for the $y\in\{2,3\}$ terms in Eq.~\eqref{eq:ECrate} we have
\begin{align}
H(A_j| B_j ;X_j = x, Y_j = y)_\mathrm{hon} = \binh\left(\pr{A_j \neq B_j |X_j = x, Y_j = y}_\mathrm{hon}\right) = \binh(w_\mathrm{exp}),
\end{align}
which lies in approximately $[0.600,0.811]$ for $w_\mathrm{exp}\in[3/4,{(2+\sqrt{2})}/{4}]$.
If the protocol parameters are such that $\xi(n,\eps^\mathrm{com}_\mathrm{EC})\binh(w_\mathrm{exp})$ turns out to be fairly close to $1$,
there is not much loss incurred by simply sending the outputs of the test rounds directly rather than expending the effort to compute appropriate error-correction data.
\subsection{Privacy amplification}\label{sec:PA}
Privacy amplification is essentially centred around the \emph{Leftover Hashing Lemma}, which we state below in the form described in~\cite{TL17} (obtained via a small modification of the proof in~\cite{rennerthesis}):
\begin{proposition}
(Leftover Hashing Lemma) Consider any $\sigma \in \dop{\leq}(CQ)$ where $C$ is a classical $n$-bit register. Let
$\mathcal{H}$ be
a 2-universal family of hash functions from $\mathbb{Z}_2^n$ to $\mathbb{Z}_2^\ell$,
and let $H$ be a register of dimension $|\mathcal{H}|$.
Define the state
\begin{align}
\omega_{KCQH} \defvar \mathcal{E}\!\left(\sigma_{CQ} \otimes
\idk_H
\right),
\end{align}
where the map $\mathcal{E}$ represents the (classical) process of applying the hash function specified in the register $H$ to the register $C$, and recording the output in register $K$.
Then for any $\epsilon \in \left[0,\sqrt{\tr{\sigma_{CQ}}}\right)$, we have
\begin{align}
\frac{1}{2}\norm{\omega_{KQH} - \idk_K \otimes\omega_{QH}}_1 \leq 2^{-\frac{1}{2}(H_\mathrm{min}^\epsilon(C|Q)_\sigma - \ell + 2)} + 2\epsilon.
\label{eq:LHL}
\end{align}
\end{proposition}
Practically speaking, the privacy amplification step simply consists of Alice choosing a random function from the 2-universal family and publicly communicating it to Bob, followed by Alice and Bob applying that function to $\str{A}$ and $\tilde{\str{A}}$ respectively. The Leftover Hashing Lemma then ensures that the output of this process is close to an ideal key, as long as the conditional min-entropy of the original state was sufficiently large. (Notice that the register $H$ is included in the ``side-information'' term in Eq.~\eqref{eq:LHL}, so it can be publicly communicated.)
\subsection{Honest behaviour}
\label{sec:hon}
In general, the honest implementation consists of $n$ IID copies of a device characterized by 2 parameters, $w_\mathrm{exp}$ and $h_\mathrm{hon}$. $w_\mathrm{exp}$ is the probability with which the device wins the CHSH game when supplied with uniformly random inputs $X_j \in \{0,1\}, Y_j \in \{2,3\}$, and $h_\mathrm{hon}$ is defined in Eq.~\eqref{eq:ECrate}. While $h_\mathrm{hon}$ does not explicitly appear in the protocol description, it is implicitly used to define the parameter $\mathrm{EC}_\mathrm{max}$, as was described in Sec.~\ref{sec:EC}. (Since $h_\mathrm{hon}$ has a dependence on $\gamma$, strictly speaking it may be more precise to instead view the honest device behaviour as being parametrized by a tuple specifying all the individual entropies in Eq.~\eqref{eq:ECrate}, but for brevity we shall summarize this as the honest behaviour being parametrized by $h_\mathrm{hon}$. In the more specific models of honest devices described below, these entropies are expressed in terms of some simpler parameters.)
In Fig.~\ref{fig:experiments}, we used the following model of the honest devices, following the estimates given in~\cite{MvDR+19} for the Bell tests in~\cite{HBD+15} (resp.~\cite{RBG+17}): we characterize them via the parameters $w_\mathrm{exp} = 0.797$ (resp.~$0.777$) and $p_\mathrm{err} = 0.06$ (resp.~$0.035$), where $p_\mathrm{err}$ is a parameter such that the probabilities \emph{before} noisy preprocessing satisfy\footnote{It can be shown that this is equivalent to taking $\pr{A_j B_j |X_j = Y_j = z}_\mathrm{hon}$ to be independent of $z$ and taking the marginal distributions $\pr{A_j |X_j = Y_j = z}_\mathrm{hon}$, $\pr{B_j |X_j = Y_j = z}_\mathrm{hon}$ to be uniform, then defining $p_\mathrm{err} \defvar \pr{A_j \neq B_j |X_j = Y_j = z}_\mathrm{hon}$.}
\begin{align}
\pr{A_j B_j |X_j = Y_j = z}_\mathrm{hon} =
(1-2p_\mathrm{err})\frac{\delta_{A_j,B_j}}{2} + 2p_\mathrm{err}\, \frac{1}{4}
& \quad \text{ for all } z\in\{0,1\},
\end{align}
where $\delta_{j,k}$ is the Kronecker delta. Furthermore, we take $\pr{A_j = B_j |X_j = x, Y_j = y}_\mathrm{hon}$ to be the same for all $x\in\{0,1\},y\in\{2,3\}$ (in which case it must equal $w_\mathrm{exp}$). Under this model, the expression~\eqref{eq:ECrate} for $h_\mathrm{hon}$ can be simplified:
\begin{align}
h_\mathrm{hon}
& = \frac{1-\gamma}{2}\binh(\p+(1-2\p)p_\mathrm{err}) + \gamma \binh(w_\mathrm{exp}),
\end{align}
where the $\p+(1-2\p)p_\mathrm{err}$ term is obtained by an explicit computation~\cite{WAP21}.
In Figs.~\ref{fig:depol} and~\ref{fig:keyrates}, we modelled the honest devices as being described by a depolarizing-noise parameter $\q \in [0,1/2]$, by which we mean that the honest devices hold the two-qubit Werner state $(1-2\q)\pure{\Phi^+} + 2\q\, \id/4$ (where $\ket{\Phi^+}=(\ket{00}+\ket{11})/\sqrt{2}$). The measurements corresponding to Alice and Bob's inputs $X_j\in\{0,1\},Y_j\in\{2,3\}$ are the ideal CHSH measurements (see e.g.~\cite{PAB+09}), and the measurements corresponding to Bob's inputs $Y_j\in\{0,1\}$ are measurements in the same bases as Alice's measurements. In terms of $w_\mathrm{exp}$ and $h_\mathrm{hon}$, this yields
\begin{align}
w_\mathrm{exp}&=(1-2\q)\frac{2+\sqrt{2}}{4} + 2\q\,\frac{1}{2}, \\
h_\mathrm{hon} & = \frac{1-\gamma}{2}\binh(\p+(1-2\p)\q) + \gamma \binh(w_\mathrm{exp}),
\end{align}
where the second expression follows by the same computation as above, noting that essentially we have $p_\mathrm{err}=\q$ in this case.
\subsection{Security statement}
To describe the length of the final key, we first need to introduce some notation and ancillary functions.
First, we require an function $\lin_\p$ satisfying some properties we now describe. Consider
any tripartite quantum state $\rho_{\bar{A}\bar{B}\bar{E}}$,
and suppose there are two possible binary-outcome measurements (indexed by $x$) on register $\bar{A}$, and similarly two binary-outcome measurements (indexed by $y$) on register $\bar{B}$. Let $w$ be the probability of winning the CHSH game (with uniform inputs) for these measurements on the state $\rho_{\bar{A}\bar{B}}$.
Let $\hat{A}_x$ be a register that stores the result if measurement $x$ is performed on register $\bar{A}$ and noisy preprocessing (Step~\ref{step:NPP}) with bias $\p$ is then applied to the outcome.
We require $\lin_\p$ to be an affine function such that for any choice of $\rho_{\bar{A}\bar{B}\bar{E}}$ and measurements, we have
\begin{align}
\sum_{x\in\{0,1\}} \frac{1}{2} H(
\hat{A}_x
|\bar{E}) \geq \lin_\p(w).
\label{eq:linbnd}
\end{align}
(The factor of $1/2$ arises from
the input distributions in the protocol.)
We give details on how to obtain such a bound in Sec.~\ref{sec:1rndbnd} (in principle, the $\p=0$ case could also be obtained from the results of~\cite{SGP+21}).
Given such a function $\lin_\p$, we then define the affine function
\begin{align}
\g(w) \defvar \frac{1-\gamma}{2} \lin_\p(w) + \gamma \lin_0(w).
\end{align}
Informally, $\g$ can be interpreted as a lower bound on the entropy ``accumulated'' in one round of the protocol.
Finally, for an affine function $f$ defined on all probability distributions on some register $C$, and any subset $\mathcal{S}$ of its domain,
we will define
\begin{align}
\begin{gathered}
\operatorname{Max}(f) \defvar \max_q f(q),
\qquad
\operatorname{Min}_{\mathcal{S}}(f) \defvar \inf_{q\in\mathcal{S}} f(q),
\\
\operatorname{Var}_\mathcal{S}(f) \defvar \sup_{q\in\mathcal{S}} \left(\sum_c q(c) f(\delta_c)^2 - \left(\sum_c q(c) f(\delta_c)\right)^2\right),
\end{gathered}
\end{align}
where $\max_q$ is taken over all distributions on $C$, and $\delta_c$ denotes the distribution with all its weight on the symbol $c$.
With these definitions, we can state the security guarantees of the protocol:
\begin{theorem}\label{th:DIQKD}
Take any
$\eps^\mathrm{com}_\mathrm{EC},\eps^\mathrm{com}_\mathrm{PE},\eps_\mathrm{EA},\eps_\mathrm{PA},\eps_\mathrm{h},\eps_s,\eps_s',\eps_s''\in (0,1]$
such that $\eps_s>\eps_s'+2\eps_s''$, and any $\alpha\in(1,2)$, $\alpha'\in(1,1+2/V')$, $\beta\in[\g(0),\g(1)]$, $\gamma\in(0,1)$, $\p\in[0,1/2]$, where $V'\defvar2\log7$.
Protocol~\ref{prot:DIQKD} is $(\eps^\mathrm{com}_\mathrm{EC} + \eps^\mathrm{com}_\mathrm{PE})$-complete and $(\max\{\eps_\mathrm{EA}, \eps_\mathrm{PA} + 2\eps_s\} + 2\eps_\mathrm{h})$-sound when performed with any choice of $\mathrm{EC}_\mathrm{max}$ such that Eq.~\eqref{eq:comEC} holds, and $\delta_\mathrm{tol},\ell_\mathrm{key}$ satisfying
\begin{align}
\eps^\mathrm{com}_\mathrm{PE} &\geq \cdfBin{n}{\gammaw_\mathrm{exp}}{\floor{(w_\mathrm{exp}-\delta_\mathrm{tol})\gamma n}} + \cdfBin{n}{
1-\gamma+\gammaw_\mathrm{exp}
}{\floor{(1-\gamma+w_\mathrm{exp}\gamma-\delta_\mathrm{tol}\gamma)n}},
\label{eq:ecPE}\\
\ell_\mathrm{key} &\leq n\g(w_\mathrm{exp}-\delta_\mathrm{tol}) - n\frac{(\alpha-1)\ln 2}{2}V^2 - n(\alpha-1)^2K_\alpha^2 - n\gamma - n\left(\frac{\alpha'-1}{4}\right)V'^2
\nonumber \\ &\qquad
- \frac{\smf{\eps_s'}}{\alpha-1} - \frac{\smf{\eps_s''}}{\alpha'-1} - \left(\frac{\alpha}{\alpha-1}+\frac{\alpha'}{\alpha'-1}-2\right)\log\frac{1}{\eps_\mathrm{EA}} - 3\smf{\eps_s-\eps_s'-2\eps_s''} \nonumber\\&\qquad
- \mathrm{EC}_\mathrm{max} - \ceil{\log\left(\frac{1}{\eps_\mathrm{h}}\right)} - 2\log\frac{1}{\eps_\mathrm{PA}} + 2,
\label{eq:keylength}
\end{align}
where $\cdfBin{n}{p}{k}$ is the cumulative distribution function of a binomial distribution (Definition~\ref{def:binom}), and
\begin{align}
\begin{aligned}
\smf{\epsilon} &\defvar \log\frac{1}{1-\sqrt{1-\epsilon^2}} \leq \log\frac{2}{\epsilon^2},
\\
V &\defvar \sqrt{\operatorname{Var}_{\mathcal{Q}_{f}}(f_\mathrm{min})+2} + \log
73,
\\
K_\alpha &\defvar \frac{2^{(\alpha-1)(2\log6 + \operatorname{Max}(f_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}_{f}} (f_\mathrm{min}))} }{6(2-\alpha)^3\ln2}
\ln^3\left(2^{2\log6 + \operatorname{Max}(f_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}_{f}} (f_\mathrm{min})} + e^2\right),
\end{aligned}
\label{eq:VK}
\end{align}
with $f_\mathrm{min}$ and $\mathcal{Q}_{f}$ being a function and a set (defined explicitly in Sec.~\ref{sec:fmin}) that satisfy
\begin{align}
\begin{gathered}
\operatorname{Max}(f_\mathrm{min})
= \frac{1}{\gamma}\g(1) + \left(1-\frac{1}{\gamma}\right)\beta,
\qquad
\operatorname{Min}_{\mathcal{Q}_{f}} (f_\mathrm{min}) = \g\!\left(\frac{2-\sqrt{2}}{4}\right), \\
\operatorname{Var}_{\mathcal{Q}_{f}}(f_\mathrm{min}) \leq
\frac{2-\sqrt{2}}{4\gamma}
\min \left\{\Delta_0^2, \Delta_1^2 \right\}
+ \frac{2+\sqrt{2}}{4\gamma}
\max \left\{\Delta_0^2, \Delta_1^2 \right\}
, \text{ where } \Delta_w \defvar \beta - \g(w).
\end{gathered}
\label{eq:minmaxvarbounds}
\end{align}
\end{theorem}
The variables listed at the start of Theorem~\ref{th:DIQKD} can be considered to be variational parameters that should be chosen to optimize the keyrate as much as possible.
When optimizing these parameters in practice, for instance to produce Figs.~\ref{fig:experiments} and~\ref{fig:depol}, the exact expression for $\smf{\epsilon}$ is numerically unstable, and hence we replace it with the upper bound of $\log(2/\epsilon^2)$ (this bound is basically tight at small $\epsilon$, so it makes little difference). Furthermore, the optimization over $\beta$ also appears to be somewhat unstable. We found heuristically that the optimal value of $\beta$ appears to typically be very close to $\g(1)$, and hence for simplicity in some cases we did not optimize over it but instead simply fixed $\beta=\g(1)$ (or slightly below it, to avoid some instabilities at $\beta-\g(1)=0$). Finally, we found that direct computation of $\cdfBin{n}{p}{k}$ (e.g.~via the regularized beta function) could sometimes be slow or unstable, and in such cases we followed~\cite{LLR+19} and replaced it with the upper bound in the following theorem:
\begin{proposition}\label{prop:ZSbound}\cite{ZS13}
Let
\begin{align}
\fZS{n}{p}{k} \defvar \Phi\left(\operatorname{sign}\left(\frac{k}{n}-p\right)\sqrt{2nG\left(\frac{k}{n},p\right)}\right),
\end{align}
where $G(x,p)\defvar x\ln\frac{x}{p} + (1-x)\ln\frac{1-x}{1-p}$, and $\Phi(z)\defvar \int_{-\infty}^z
(2\pi)^{-1/2}
e^{-u^2/2} \mathrm{d}u =
\left(1+\operatorname{erf}\!\left({z}/{\sqrt{2}}\right)\right)/2$ is the cumulative distribution function of the standard normal distribution.
Then for any $k\in\upto{n-1}$,
\begin{align}
\fZS{n}{p}{k} \leq \cdfBin{n}{p}{k} \leq \fZS{n}{p}{k+1}.
\label{eq:ZSbound}
\end{align}
\end{proposition}
\noindent Replacing $\cdfBin{n}{p}{k}$ with $\fZS{n}{p}{k+1}$ and computing the latter (which is a Gaussian integral) appeared to be faster and more stable than computing $\cdfBin{n}{p}{k}$ directly. There is little loss incurred by performing this replacement --- the inequalities~\eqref{eq:ZSbound} imply $\fZS{n}{p}{k+1} \leq \cdfBin{n}{p}{k+1}$, so the effect is no larger than replacing $\cdfBin{n}{p}{k}$ by $\cdfBin{n}{p}{k+1}$, which is almost negligible in the parameter regimes studied in this work.
Note that in general, the optimal parameter values (especially for $\alpha$ and $\alpha'$) would depend heavily on $n$.
To get an estimate for the asymptotic scaling of $\ell_\mathrm{key}$, we can choose all the $\epsilon$ parameters to take some constant values satisfying the desired completeness and soundness bounds, then choose~\cite{DFR20,DF19}
\begin{align}
\begin{aligned}
\alpha-1&= \frac{1}{\sqrt{n}} \sqrt{\frac{2}{V^2 \ln 2} \left(\smf{\eps_s'} + 2\log\frac{1}{\eps_\mathrm{EA}}\right)} , \\
\alpha'-1&= \frac{1}{\sqrt{n}} \sqrt{\frac{4}{V'^2} \left(\smf{\eps_s''} + 2\log\frac{1}{\eps_\mathrm{EA}} \right)} ,
\end{aligned}
\end{align}
taking $n$ to be large enough such that $\alpha \in (1,3/2)$ and $\alpha'\in(1,1+2/V')$. Furthermore, introduce a constant
\begin{align}
\hat{K} \defvar \frac{2^{(2\log6 + \operatorname{Max}(f_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}_{f}} (f_\mathrm{min}))} }{(3/4)\ln2}
\ln^3\left(2^{2\log6 + \operatorname{Max}(f_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}_{f}} (f_\mathrm{min})} + e^2\right) ,
\end{align}
which satisfies $\hat{K} \geq K_\alpha$ for $\alpha\in(1,3/2)$.
With these choices, Eq.~\eqref{eq:keylength} can be satisfied by taking
\begin{align}
\ell_\mathrm{key} &= \bigg\lfloor
n\g(w_\mathrm{exp}-\delta_\mathrm{tol}) - n\frac{(\alpha-1)\ln 2}{2}V^2 - n(\alpha-1)^2 \hat{K}^2 - n\gamma - n\left(\frac{\alpha'-1}{4}\right)V'^2
\nonumber \\ &\qquad
- \frac{1}{\alpha-1}\left(\smf{\eps_s'} + 2\log\frac{1}{\eps_\mathrm{EA}} \right) - \frac{1}{\alpha'-1}\left(\smf{\eps_s''} + 2\log\frac{1}{\eps_\mathrm{EA}} \right) + 2\log\frac{1}{\eps_\mathrm{EA}} - 3\smf{\eps_s-\eps_s'-2\eps_s''} \nonumber\\&\qquad
- \mathrm{EC}_\mathrm{max} - \ceil{\log\left(\frac{1}{\eps_\mathrm{h}}\right)} - 2\log\frac{1}{\eps_\mathrm{PA}} + 2 \bigg\rfloor \nonumber \\
&= \bigg\lfloor n\g(w_\mathrm{exp}-\delta_\mathrm{tol}) - \sqrt{n}\,2\sqrt{\frac{V^2 \ln 2}{2} \left(\smf{\eps_s'} + 2\log\frac{1}{\eps_\mathrm{EA}}\right)} - \frac{2}{V^2 \ln 2} \left(\smf{\eps_s'} + 2\log\frac{1}{\eps_\mathrm{EA}}\right)\hat{K}^2
- n\gamma \nonumber \\ &\qquad
- \sqrt{n} \, 2\sqrt{\frac{V'^2}{4} \left(\smf{\eps_s''} + 2\log\frac{1}{\eps_\mathrm{EA}} \right)} + 2\log\frac{1}{\eps_\mathrm{EA}} - 3\smf{\eps_s-\eps_s'-2\eps_s''} \nonumber\\&\qquad
- \mathrm{EC}_\mathrm{max} - \ceil{\log\left(\frac{1}{\eps_\mathrm{h}}\right)} - 2\log\frac{1}{\eps_\mathrm{PA}} + 2 \bigg\rfloor \nonumber \\
&= \left\lfloor n\g(w_\mathrm{exp}-\delta_\mathrm{tol}) - O(\sqrt{n}) - O(1) - \mathrm{EC}_\mathrm{max} \right\rfloor.
\label{eq:estkeylength}
\end{align}
Furthermore, Eq.~\eqref{eq:ecPE} can be satisfied by choosing $\gamma=3w_\mathrm{exp} n^{-1} \delta_\mathrm{tol}^{-2} \log(2/\eps^\mathrm{com}_\mathrm{PE})$ (see Eq.~\eqref{eq:chernoffs}), in which case by taking $\delta_\mathrm{tol} \propto n^{-1/3}$ we have $\delta_\mathrm{tol},\gamma \to 0$ as $n\to\infty$, and Eq.~\eqref{eq:estkeylength} then yields
\begin{align}
\lim_{n\to\infty} \frac{\ell_\mathrm{key}}{n} =
\frac{1}{2}\left(\lin_\p(w_\mathrm{exp}) -
\sum_{z\in\{0,1\}} \frac{1}{2}H(A_j | B_j;X_j = Y_j = z)_\mathrm{hon}\right)
,
\label{eq:asympt}
\end{align}
taking $\mathrm{EC}_\mathrm{max}$ according to Eqs.~\eqref{eq:optEC}--\eqref{eq:ECrate}. This is the expected asymptotic result, e.g.~according to the Devetak-Winter bound~\cite{DW05} (with the prefactor of $1/2$ being due to the sifting), which is tight for protocols based on one-way error correction.
Given the scaling behaviour shown in Eq.~\eqref{eq:estkeylength}, it can be seen that the optimal values of the various $\epsilon$ parameters (given some desired values of $\eps^\mathrm{com}$ and $\eps^\mathrm{sou}$) may be of rather different orders of magnitude. This is because some of them appear in $O(1/\sqrt{n})$ corrections to the finite-size keyrate while others appear in $O(1/n)$ corrections. Intuitively speaking, the $\epsilon$ parameters in the latter can be chosen to be substantially smaller than the former, since the $O(1/n)$ scaling reduces their contribution to the finite-size effects.
\section{Finite-size security proof}
\label{sec:finite}
We now prove that Protocol~\ref{prot:DIQKD} indeed satisfies the security properties claimed in Theorem~\ref{th:DIQKD}.
\subsection{Completeness}
\label{sec:com}
Completeness is defined entirely with respect to the honest behaviour of the devices, hence all discussion in this section is with respect to the honest behaviour. We consider all possible ways the protocol might abort. Firstly, the hashes of $\str{A}$ and $\tilde{\str{A}}$ might not match. Note that this can only happen if $\str{A}\neq\tilde{\str{A}}$, which happens with probability at most $\eps^\mathrm{com}_\mathrm{EC}$ by the guarantees of the error-correction step.
Secondly, the register $\str{C}$ might fail to satisfy $\operatorname{freq}_\str{c}(1)\geq (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma$ or $\operatorname{freq}_\str{c}(0)\leq (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma$. To bound the probability of this event, we recall that the honest protocol consists of $n$ IID rounds such that
$\pr{C_j=1}_\mathrm{hon} = \gamma w_\mathrm{exp}$ and $\pr{C_j=0}_\mathrm{hon} = \gamma (1-w_\mathrm{exp})$ in each round. Therefore, we have
\begin{align}
\pr{\operatorname{freq}_\str{c}(1) < (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma}_\mathrm{hon}
&\leq
\pr{\operatorname{freq}_\str{c}(1)n \leq \floor{(w_\mathrm{exp}-\delta_\mathrm{tol})\gamma n}}_\mathrm{hon} \nonumber\\
&= \cdfBin{n}{\gammaw_\mathrm{exp}}{\floor{(w_\mathrm{exp}-\delta_\mathrm{tol})\gamma n}}, \label{eq:ecPEbound1}\\
\pr{\operatorname{freq}_\str{c}(0) > (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma}_\mathrm{hon}
&= \pr{\operatorname{freq}_\str{c}(
\neg 0
) \leq 1-(1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma}_\mathrm{hon} \nonumber\\
&= \pr{\operatorname{freq}_\str{c}(\neg 0) n \leq \floor{
(1-\gamma+w_\mathrm{exp}\gamma-\delta_\mathrm{tol}\gamma)
n}}_\mathrm{hon} \nonumber\\
&= \cdfBin{n}{
1-\gamma+\gammaw_\mathrm{exp}
}{\floor{(1-\gamma+w_\mathrm{exp}\gamma-\delta_\mathrm{tol}\gamma)n} },
\label{eq:ecPEbound0}
\end{align}
where $\neg 0$ represents all symbols other than $0$.
By the union bound, we hence have the desired upper bound on the probability of the protocol aborting:
\begin{align}
\pr{\str{A}\neq\tilde{\str{A}} \,\lor\, \operatorname{freq}_\str{c}(1) < (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma \,\lor\, \operatorname{freq}_\str{c}(0) > (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma}_\mathrm{hon} \leq \eps^\mathrm{com}_\mathrm{EC} + \eps^\mathrm{com}_\mathrm{PE},
\end{align}
where $\eps^\mathrm{com}_\mathrm{PE}$ is as specified in Eq.~\eqref{eq:ecPE}.
In principle, somewhat simpler expressions could be obtained using the (multiplicative) Chernoff bound:
\begin{align}
\begin{gathered}
\pr{\operatorname{freq}_\str{c}(1) < (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma}_\mathrm{hon}
\leq \left(\frac{e^{-t}}{(1-t)^{1-t}}\right)^{n\gammaw_\mathrm{exp}}
\leq e^{-\frac{n \gamma \delta_\mathrm{tol}^2}{2w_\mathrm{exp}}},
\\
\pr{\operatorname{freq}_\str{c}(0) > (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma}_\mathrm{hon}
\leq \left(\frac{e^t}{(1+t)^{1+t}}\right)^{n\gammaw_\mathrm{exp}}
\leq e^{-\frac{n \gamma \delta_\mathrm{tol}^2}{3w_\mathrm{exp}}},
\end{gathered}
\label{eq:chernoffs}
\end{align}
where $t=\delta_\mathrm{tol}/w_\mathrm{exp}\leq 1$. However, it was observed in~\cite{LLR+19} that these bounds are weaker than Prop.~\ref{prop:ZSbound} by a significant amount.
As yet another alternative, Hoeffding's inequality yields a bound of $e^{-2n \gamma^2 \delta_\mathrm{tol}^2}$,
but this is worse than the Chernoff bound whenever $6\gamma \leq 1/w_\mathrm{exp}$; furthermore, it does not yield nontrivial bounds if we choose $\gamma\propto1/n$ (for constant $\eps^\mathrm{com}_\mathrm{EC},\delta_\mathrm{tol}$), which ought to be possible in principle according to the scaling analysis in~\cite{DF19}.
\subsection{Soundness}
\label{sec:soundness}
For the purposes of understanding this proof, it may be helpful to think of it as analyzing a \emph{specific} set of states and measurements that could be occurring in a run of Protocol~\ref{prot:DIQKD} (as opposed to simultaneously considering all possible states and measurements that could be occurring). In particular, for instance,
the virtual protocol below
and the channels $\mathcal{M}_j$ in Sec.~\ref{sec:EAT} should be understood as being constructed in terms of this specific set of states and measurements. However, since we will not impose any additional assumptions on these states and measurements beyond those specified by the protocol, this will serve to prove security for all possible states and measurements that could occur in a run of the protocol.
Consider the state at the end of Step~\ref{step:PA} in Protocol~\ref{prot:DIQKD}. We now describe a virtual protocol\footnote{We stress that this ``protocol'' is not performed in practice
(and in fact cannot be, since the ``virtual parameter estimation'' step cannot be performed locally by either party).
However, it produces exactly the same state as Protocol~\ref{prot:DIQKD} on all relevant registers, and can hence be used for the security analysis.} that involves two additional registers ${\str{B}'}{\str{C}'}$, but yields exactly the same state on the registers $\str{A}\tilde{\str{A}}\str{B}\str{X}\str{Y}\str{L}\str{C}E$ (when it is implemented using the same input state and measurements as those used in a run of Protocol~\ref{prot:DIQKD}).
\let\oldthealgorithm\thealgorithm
\renewcommand{\thealgorithm}{$1'$}
\begin{savenotes}
\begin{algorithm}[H]
\caption{A virtual protocol}\label{prot:virtual}
\begin{algorithmic}[1]
\State Alice and Bob's devices each receive and store all quantum states that they will subsequently measure.
\State\label{step:EATmap}For each $j \in \upto{n}$, perform the following steps:
\begin{algsubstates}
\State Alice chooses a uniform input $X_j \in \{0,1\}$. With probability $\gamma$, Bob chooses a uniform input $Y_j \in \{2,3\}$, otherwise Bob chooses a uniform input $Y_j \in \{0,1\}$.
\State Alice and Bob supply their inputs to their devices, and record the outputs as $A_j$ and $B_j$ respectively.
\State Alice and Bob publicly announce their inputs.
\State\textbf{Sifting:} If $Y_j \in \{0,1\}$ and $X_j \neq Y_j$, Alice and Bob overwrite their outputs with $A_j = B_j = 0$.
\State\textbf{Noisy preprocessing:} If $Y_j \in \{0,1\}$ and $X_j = Y_j$, Alice generates a biased random bit $F_j$ with $\pr{F_j=1} = \p$, and overwrites her output $A_j$ with $A_j \oplus F_j$.
\State
If $Y_j \in \{0,1\}$, Bob sets $B'_j = \perp$, otherwise Bob sets $B'_j = B_j$.
\State\textbf{Virtual parameter estimation:}
Set $C'_j=\perp$ if $Y_j \in \{0,1\}$; otherwise set $C'_j=0$ if $A_j \oplus B'_j \neq X_j \cdot (Y_j-2)$ and $C'_j=1$ if $A_j \oplus B'_j = X_j \cdot (Y_j-2)$.
\end{algsubstates}
\State\textbf{Error correction:} Alice and Bob publicly communicate some bits $\str{L}$ for error correction as previously described, allowing Bob to construct a guess $\tilde{\str{A}}$ for $\str{A}$.
\State\textbf{Parameter estimation:}
For all $j\in\upto{n}$, Bob sets
$C_j=\perp$ if $Y_j \in \{0,1\}$; otherwise he sets $C_j=0$ if $\tilde{A}_j \oplus B'_j \neq X_j \cdot (Y_j-2)$ and $C_j=1$ if $\tilde{A}_j \oplus B'_j = X_j \cdot (Y_j-2)$.
\end{algorithmic}
\end{algorithm}
\end{savenotes}
\let\thealgorithm\oldthealgorithm
\addtocounter{algorithm}{-1}
\newpage
\noindent The key changes as compared to Protocol~\ref{prot:DIQKD} are as follows:
\begin{itemize}
\item All the states that the devices will measure are distributed immediately at the start (note that this is possible because in Protocol~\ref{prot:DIQKD}, the measurement choices $\str{X},\str{Y}$ are not disclosed until all measurements have been performed, and hence the distributed states cannot behave adaptively with respect to the inputs).
\item The sifting and noisy preprocessing steps are now performed immediately after each measurement, instead of after all measurements are performed. This is to allow us to subsequently apply the EAT.
\item Two additional registers were introduced: ${\str{B}'}$, which is equal to $\str{B}$ on all the test rounds but is otherwise set to a null symbol, and ${\str{C}'}$, which is analogous to $\str{C}$ but computed using $\str{A}$ in place of $\tilde{\str{A}}$. These registers were used in a \emph{virtual parameter estimation} step.
\item All parameter estimation is performed with ${\str{B}'}$ instead of $\str{B}$ (this substitution has no physical effect since $B'_j = B_j$
in all rounds used for parameter estimation).
\end{itemize}
Let $\rho$ denote the state on registers ${\str{A}\tilde{\str{A}}\str{B}{\str{B}'}\str{X}\str{Y}\str{L}\str{C}{\str{C}'} E}$ at the end of Protocol~\ref{prot:virtual}. As mentioned, the reduced state on ${\str{A}\tilde{\str{A}}\str{B}\str{X}\str{Y}\str{L}\str{C}E}$ is exactly the same as that at the end of Step~\ref{step:PA} in Protocol~\ref{prot:DIQKD}.
Since all subsequent steps in Protocol~\ref{prot:DIQKD} (i.e.~simply the accept/abort check and the privacy amplification) only involve these registers, to prove the security of Protocol~\ref{prot:DIQKD} it suffices to consider the equivalent (apart from ${\str{B}'}{\str{C}'}$) process where all the steps up to Step~\ref{step:PA} are replaced by this virtual protocol, and then the remaining steps in Protocol~\ref{prot:DIQKD} are performed. All our subsequent analysis is in terms of this process.
We define the following events on $\rho$:
{\setlist[description]{leftmargin=2cm,labelindent=2cm}
\begin{description}
\item $\Og$: $\str{A} = \tilde{\str{A}}$
(i.e.~Bob correctly guesses $\str{A}$)
\item $\Oh$: $\operatorname{hash}(\str{A}) = \operatorname{hash}(\tilde{\str{A}})$
\item $\OPE$: $\operatorname{freq}_\str{c}(1)\geq (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma$ and $\operatorname{freq}_\str{c}(0)\leq (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma$
\item $\OpPE$: $\operatorname{freq}_{{\str{c}'}}(1)\geq (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma$ and $\operatorname{freq}_{{\str{c}'}}(0)\leq (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma$
\end{description}
}
In terms of these events, the accept condition for Protocol~\ref{prot:DIQKD} is $\Oh \land \OPE$. Hence we immediately see that the protocol is $\eps_\mathrm{h}$-correct, since
\begin{align}
\pr{K_A \neq K_B \land \Oh \land \OPE}
\leq \pr{K_A \neq K_B \land \Oh}
\leq \pr{\no{\Og} \land \Oh}
\leq \pr{\Oh | \no{\Og}}
\leq \eps_\mathrm{h},
\label{eq:corrproof}
\end{align}
where the last inequality holds by the defining property of 2-universal hashing.
It remains to prove secrecy.
Denote the privacy amplification step in Protocol~\ref{prot:DIQKD} as the map $\mathcal{M}_\mathrm{PA}$, so the subnormalized state conditioned on the event of Protocol~\ref{prot:DIQKD} accepting can be written as $\mathcal{M}_\mathrm{PA}(\rho_{\land \Oh \land \OPE})$. Let $H$ be the register storing the choice of hash function used for privacy amplification, and denote $E' \defvar \str{X}\str{Y}\str{L}EH$ for brevity. We can now rewrite the secrecy condition as the requirement
\begin{align}
\frac{1}{2}\norm{\mathcal{M}_\mathrm{PA}(\rho_{\land \Oh \land \OPE})_{K_A E'} - \idk_{K_A} \otimes \mathcal{M}_\mathrm{PA}(\rho_{\land \Oh \land \OPE})_{E'}}_1 \leq \eps^\mathrm{sec}_\mathrm{QKD}.
\label{eq:esecrnew}
\end{align}
By noting that $\rho_{\land \Oh \land \OPE} = \rho_{\land \Og \land \Oh \land \OPE} + \rho_{\land \no{\Og} \land \Oh \land \OPE}$, we find that the left-hand-side can be bounded using norm subadditivity:
\begin{align}
&\frac{1}{2}\norm{\mathcal{M}_\mathrm{PA}(\rho_{\land \Oh \land \OPE})_{K_A E'} - \idk_{K_A} \otimes \mathcal{M}_\mathrm{PA}(\rho_{\land \Oh \land \OPE})_{E'}}_1 \nonumber\\
\leq{}& \frac{1}{2}\norm{\mathcal{M}_\mathrm{PA}(\rho_{\land \Og \land \Oh \land \OPE})_{K_A E'} - \idk_{K_A} \otimes \mathcal{M}_\mathrm{PA}(\rho_{\land \Og \land \Oh \land \OPE})_{E'}}_1 \nonumber\\
&\qquad + \frac{1}{2}\norm{\mathcal{M}_\mathrm{PA}(\rho_{\land \no{\Og} \land \Oh \land \OPE})_{K_A E'} - \idk_{K_A} \otimes \mathcal{M}_\mathrm{PA}(\rho_{\land \no{\Og} \land \Oh \land \OPE})_{E'}}_1 \nonumber\\
\leq{}& \frac{1}{2}\norm{\mathcal{M}_\mathrm{PA}(\rho_{\land \Og \land \Oh \land \OpPE})_{K_A E'} - \idk_{K_A} \otimes \mathcal{M}_\mathrm{PA}(\rho_{\land \Og \land \Oh \land \OpPE})_{E'}}_1
+ \eps_\mathrm{h}, \label{eq:secsplit}
\end{align}
where in the last line we have
performed the substitution $\Og \land \OPE = \Og \land \OpPE$ in the first term,
and bounded the second term using
\begin{align}
\pr{\no{\Og} \land \Oh \land \OPE} \leq \pr{\no{\Og} \land \Oh} \leq \eps_\mathrm{h},
\end{align}
as noted in Eq.~\eqref{eq:corrproof}.
We now aim to bound the first term in Eq.~\eqref{eq:secsplit}. To do so, we study three exhaustive possibilities for the state $\rho$ (the first two are not mutually exclusive, but this does not matter):
{\setlist[enumerate]{leftmargin=2cm,labelindent=2cm}
\begin{enumerate}[label*=Case \arabic*:,ref=\arabic*]
\item \label{case:ECsmall} $\pr{\Og \land \Oh|\OpPE} \leq \eps_s^2$.
\item \label{case:PEsmall} $\pr{\OpPE} \leq \eps_\mathrm{EA}$.
\item \label{case:neither} Neither of the above are true.
\end{enumerate}
}
\noindent In case~\ref{case:ECsmall}, that term is bounded by
\begin{align}
\pr{\Og \land \Oh \land \OpPE} = \pr{\Og \land \Oh |\OpPE} \pr{\OpPE} \leq \eps_s^2.
\label{eq:caseECsmall}
\end{align}
In case~\ref{case:PEsmall}, that term is bounded by
\begin{align}
\pr{\Og \land \Oh \land \OpPE} \leq \pr{\OpPE} \leq \eps_\mathrm{EA}.
\end{align}
The main challenge is case~\ref{case:neither}. To study this case, we first focus on the conditional state $\rho_{|\OpPE}$. Importantly, the relevant smoothed min-entropy of this state can be bounded by using the following theorem, which we prove in Sec.~\ref{sec:EAT} using entropy accumulation:
\begin{theorem}\label{th:rawHmin}
For all parameter values as specified in Theorem~\ref{th:DIQKD}, the state at the end of Protocol~\ref{prot:virtual} satisfies
\begin{align}
H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}E)_{\rho_{|\OpPE}} &>
n\g(w_\mathrm{exp}-\delta_\mathrm{tol}) - n\frac{(\alpha-1)\ln 2}{2}V^2 - n(\alpha-1)^2K_\alpha^2 - n\gamma - n\left(\frac{\alpha'-1}{4}\right)V'^2
\nonumber \\ &\qquad
- \frac{\smf{\eps_s'}}{\alpha-1} - \frac{\smf{\eps_s''}}{\alpha'-1} - \left(\frac{\alpha}{\alpha-1}+\frac{\alpha'}{\alpha'-1}\right)\log\frac{1}{\pr{\OpPE}} - 3\smf{\eps_s-\eps_s'-2\eps_s''}.
\label{eq:rawHmin}
\end{align}
where $V',\smf{\epsilon},V,K_\alpha$ are as defined in Theorem~\ref{th:DIQKD}.
\end{theorem}
To relate this to our state of interest,
we use~\cite{TL17} Lemma~10, which states that for a state $\sigma\in\dop{\leq}(ZZ'QQ')
$ that is classical on registers $ZZ'$, an event $\Omega$ on the registers $ZZ'$, and any $\epsilon \in [0,\sqrt{\tr{\sigma_{\land\Omega}}})$,
we have
\begin{align}
H_\mathrm{min}^\epsilon(ZQ|Q')_{\sigma_{\land\Omega}} \geq H_\mathrm{min}^\epsilon(ZQ|Q')_{\sigma} .
\label{eq:condlemma}
\end{align}
In our context, we observe that the probability of the event $\Og \land \Oh$ on the normalized state $\rho_{|\OpPE}$ is $\pr{\Og \land \Oh|\OpPE}$, which is greater than $\eps_s^2$ since we are in case~\ref{case:neither}. Hence the conditions
of the lemma are satisfied (identifying $\str{A}$ with $Z$, $\tilde{\str{A}}$ (and the error-correction hash choice) with $Z'$, $\str{X}\str{Y}\str{L}E$ with $Q'$, and leaving $Q$ empty), allowing us to obtain the bound
\begin{align}
H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}\str{L}E)_{\big(\rho_{|\OpPE}\big) {}_{\land \Og \land \Oh} } &\geq H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}\str{L}E)_{\rho_{|\OpPE}} \nonumber\\
&\geq H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}E)_{\rho_{|\OpPE}}
- \operatorname{len}(\str{L}) \nonumber\\
&\geq H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}E)_{\rho_{|\OpPE}}
- \mathrm{EC}_\mathrm{max} - \ceil{\log\left(\frac{1}{\eps_\mathrm{h}}\right)}
,\label{eq:entPA}
\end{align}
where in the second line we have applied a chain rule for smoothed min-entropy (see e.g.~\cite{WTHR11} Lemma~11 or~\cite{Tom16} Lemma~6.8).
Putting together Eq.~\eqref{eq:entPA} and Theorem~\ref{th:rawHmin}, we find that for a key of length $\ell_\mathrm{key}$ satisfying Eq.~\eqref{eq:keylength}, we have\footnote{Keeping the $\log\left(1/\pr{\OpPE}\right)$ term here yields a slightly tighter result as compared to~\cite{ARV19,MvDR+19}, which instead used $\pr{\OpPE}\leq1$ to write an inequality in place of the equality in the second-last line of Eq.~\eqref{eq:finalPA}.}
\begin{align}
&\frac{1}{2}\left(H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}\str{L}E)_{\big(\rho_{|\OpPE}\big) {}_{\land \Og \land \Oh} } - \ell_\mathrm{key} + 2\right) + \log\frac{1}{\pr{\OpPE}} \nonumber\\
\geq{}&
\frac{1}{2} \left(- \left(\frac{\alpha}{\alpha-1}+\frac{\alpha'}{\alpha'-1}\right)\log\frac{1}{\pr{\OpPE}} +
2\log\frac{1}{\eps_\mathrm{PA}} +
\left(\frac{\alpha}{\alpha-1}+\frac{\alpha'}{\alpha'-1} - 2\right)\log\frac{1}{\eps_\mathrm{EA}} \right) + \log\frac{1}{\pr{\OpPE}} \nonumber\\
={}& \log\frac{1}{\eps_\mathrm{PA}} +\frac{1}{2} \left(\frac{\alpha}{\alpha-1}+\frac{\alpha'}{\alpha'-1}-2\right) \left(\log\frac{1}{\eps_\mathrm{EA}} - \log\frac{1}{\pr{\OpPE}}\right) \nonumber\\
\geq{}& \log\frac{1}{\eps_\mathrm{PA}},
\end{align}
where the last line holds because $\pr{\OpPE} \geq \eps_\mathrm{EA}$ in case~\ref{case:neither} (and also $\frac{\alpha}{\alpha-1}+\frac{\alpha'}{\alpha'-1} -2 \geq 2
> 0
$ since $\alpha,\alpha' \in (1,2)$).
\newpage
This finally allows us to bound Eq.~\eqref{eq:secsplit}, since
we have $\eps_s^2 \leq \pr{\Og \land \Oh|\OpPE} = \operatorname{Tr}[(\rho_{|\OpPE})_{\land \Og \land \Oh}]$ in case~\ref{case:neither}, and hence we can apply the Leftover Hashing Lemma to see that\footnote{We remark that the events $\OpPE$ and $\Og \land \Oh$ were treated somewhat differently in this analysis, in that
while we introduced a parameter $\eps_\mathrm{EA}$ to divide the analysis of $\OpPE$ into separate cases, there is no analogous condition for $\Og \land \Oh$ (the condition for case~\ref{case:ECsmall} is not analogous; it is merely a technicality to allow us to apply Eq.~\eqref{eq:condlemma} and the Leftover Hashing Lemma).
This is fundamentally because conditioning on $\OpPE$ via Eq.~\eqref{eq:rawHmin} has a ``larger'' effect on min-entropy compared to conditioning on $\Og \land \Oh$ (via Eq.~\eqref{eq:condlemma} modified for normalized conditional states). When substituting these effects into the final security parameter, one finds that the former worsens the security parameter by an amount that depends on $\log(1/\pr{\OpPE})$, while the latter does not, and hence we required a bound on $\pr{\OpPE}$ in the former.}
\begin{align}
&\frac{1}{2}\norm{\mathcal{M}_\mathrm{PA}(\rho_{\land \Og \land \Oh \land \OpPE})_{K_A E'} - \idk_{K_A} \otimes \mathcal{M}_\mathrm{PA}(\rho_{\land \Og \land \Oh \land \OpPE})_{E'}}_1 \nonumber\\
={}& \pr{\OpPE} \frac{1}{2}\norm{\mathcal{M}_\mathrm{PA}((\rho_{|\OpPE})_{\land \Og \land \Oh})_{K_A E'} - \idk_{K_A} \otimes \mathcal{M}_\mathrm{PA}((\rho_{|\OpPE})_{\land \Og \land \Oh})_{E'}}_1 \nonumber\\
\leq{}& \pr{\OpPE} \left(2^{-\frac{1}{2}\left(H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}\str{L}E)_{\big(\raisebox{1.5\depth}{$\mathsmaller{\rho}$}|\OpPE\big) {\land \Og \land \Oh} } - \ell_\mathrm{key} + 2\right)} + 2\eps_s \right) \nonumber\\
={}& 2^{-\frac{1}{2}\left(H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}\str{L}E)_{\big(\raisebox{1.5\depth}{$\mathsmaller{\rho}$}|\OpPE\big)\land \Og \land \Oh } - \ell_\mathrm{key} + 2\right) - \log\left(1/\pr{\OpPE}\right)} + 2\eps_s\pr{\OpPE}
\nonumber\\
\leq{}& \eps_\mathrm{PA} + 2\eps_s. \label{eq:finalPA}
\end{align}
Since the three possible cases are exhaustive, we conclude that the secrecy condition is satisfied by choosing
\begin{align}
\eps^\mathrm{sec}_\mathrm{QKD} = \max\{\eps_s^2, \eps_\mathrm{EA}, \eps_\mathrm{PA} + 2\eps_s\} + \eps_\mathrm{h} = \max\{\eps_\mathrm{EA}, \eps_\mathrm{PA} + 2\eps_s\} + \eps_\mathrm{h}.
\end{align}
Recalling that we have already shown the protocol is $\eps_\mathrm{h}$-correct, we finally conclude that it is $(\max\{\eps_\mathrm{EA}, \eps_\mathrm{PA} + 2\eps_s\} + 2\eps_\mathrm{h})$-sound.
\subsection{Entropy accumulation}\label{sec:EAT}
This section is devoted to the proof of Theorem~\ref{th:rawHmin}. The key theoretical tool in this proof is the entropy accumulation theorem, which we shall now briefly outline in the form stated in~\cite{DF19}. To do so, we shall first introduce \emph{EAT channels} and \emph{tradeoff functions}.
\begin{definition}\label{def:EATchann}
A \emph{sequence of EAT channels} is a sequence
$\{\mathcal{M}_j\}_{j\in\upto{n}}$
where each $\mathcal{M}_j$ is a channel from a register $R_{j-1}$ to registers $D_j S_j T_j R_j$,
which satisfies the following properties:
\begin{itemize}
\item All $D_j$ are classical registers with a common alphabet $\mathcal{D}$, and all $S_j$ have the same finite dimension.
\item
For each $\mathcal{M}_j$, the value of $D_j$ is determined from the registers $S_j T_j$ alone. Formally, this means
$\mathcal{M}_j$ is of the form $\mathcal{P}_j \circ \mathcal{M}'_j$, where $\mathcal{M}'_j$ is a channel from $R_{j-1}$ to $S_j T_j R_j$, and $\mathcal{P}_j$ is a channel from $S_j T_j$ to $D_j S_j T_j$ of the form
\begin{align}
\mathcal{P}_j(\rho_{S_j T_j}) = \sum_{s\in\mathcal{S},t\in\mathcal{T}} (\Pi_{S_j,s} \otimes \Pi_{T_j,t}) \rho_{S_j T_j} (\Pi_{S_j,s} \otimes \Pi_{T_j,t}) \otimes \pure{d(s,t)}_{D_j},
\label{eq:nodisturb}
\end{align}
where $\{\Pi_{S_j,s}\}_{s\in\mathcal{S}}$ and $\{\Pi_{T_j,t}\}_{t\in\mathcal{T}}$ are families of orthogonal projectors on $S_j$ and $T_j$ respectively, and $d: \mathcal{S} \times \mathcal{T} \to \mathcal{D}$ is a deterministic function.
\end{itemize}
\end{definition}
\begin{definition}
Let $f_\mathrm{min}$ be a real-valued affine function defined on probability distributions over the alphabet $\mathcal{D}$. It is called a \emph{min-tradeoff function} for a sequence of EAT channels $\{\mathcal{M}_j\}_{j\in\upto{n}}$ if for any distribution $q$ on $\mathcal{D}$, we have
\begin{align}
f_\mathrm{min}(q) \leq \inf_{\sigma \in \Sigma_j(q)} H(S_j|T_jR)_\sigma \qquad \forall j\in\upto{n},
\end{align}
where $\Sigma_j(q)$ denotes the set of states of the form $(\mathcal{M}_j\otimes\idmap_{R})(\omega_{R_{j-1}R})$ such that the reduced state on $D_j$ has distribution $q$.
Analogously, a real-valued affine function $f_\mathrm{max}$ defined on probability distributions over $\mathcal{D}$ is called a \emph{max-tradeoff function} if
\begin{align}
f_\mathrm{max}(q) \geq \sup_{\sigma \in \Sigma_j(q)} H(S_j|T_jR)_\sigma \qquad \forall j\in\upto{n}.
\end{align}
The infimum and supremum of an empty set are defined as $+\infty$ and $-\infty$ respectively.
\end{definition}
\noindent With these definitions, we can now state the theorem:
\begin{proposition}\label{prop:EAT}
(Entropy accumulation theorem~\cite{DF19,LLR+19})
Consider a sequence of EAT channels $\{\mathcal{M}_j\}_{j\in\upto{n}}$ and a state of the form $\rho = (\mathcal{M}_n \circ\dots \circ\mathcal{M}_1 \otimes \idmap_E)\left(\rho^0_{R_0E}\right)$ satisfying the Markov conditions
\begin{align}
I(S_{\upto{j-1}} : T_j | T_{\upto{j-1}} E)_\rho
= 0 \qquad \forall j\in\upto{n}.
\label{eq:markov}
\end{align
\noindent Let $f_\mathrm{min}$ be a min-tradeoff function for $\{\mathcal{M}_j\}_{j\in\upto{n}}$ and consider any $h\in\mathbb{R}, \epsilon\in(0,1), \alpha\in(1,2)$. Then for any event $\Omega \subseteq \mathcal{D}^n$ such that
$f_\mathrm{min}(\operatorname{freq}_{\str{d}}) \geq h$ for all $\str{d}\in\Omega$, we have\footnote{This expression differs slightly from those in~\cite{DF19,LLR+19} because we have not performed the simplifications based on $\smf{\epsilon} \leq \log(2/\epsilon^2)$ and $\alpha < 2$ (though the former is only a very small improvement for typical values of $\epsilon$).
We also remark that in~\cite{LLR+19}, a modification was made to the EAT to improve its dependence on the $\operatorname{Var}$ term. However, it appears that this improvement mainly has an effect when $f_\mathrm{min}$ (which must be affine) differs from the tight bound on $H(S_j|T_jR)$. For the DIQKD protocol we study, these bounds appear to be equal in the regime of interest (see Sec.~\ref{sec:1rndresults}), and hence we do not expect any improvement by using that technique. }
\begin{align}
H_\mathrm{min}^{\epsilon}(\str{S}|\str{T}E)
>nh - n\frac{(\alpha-1)\ln 2}{2}V^2 - n(\alpha-1)^2K_\alpha^2
- \frac{\smf{\epsilon}}{\alpha-1} - \frac{\alpha}{\alpha-1}\log\frac{1}{\pr{\Omega}},
\end{align}
where $\smf{\epsilon}$ is as defined in Eq.~\eqref{eq:VK}, and
\begin{align}
\begin{aligned}
V &\defvar \sqrt{\operatorname{Var}_{\mathcal{Q}}(f_\mathrm{min})+2} + \log
(2\dim(S_j)^2+1),
\\
K_\alpha &\defvar \frac{2^{(\alpha-1)(2\log\dim(S_j) + \operatorname{Max}(f_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}} (f_\mathrm{min}))} }{6(2-\alpha)^3\ln2}
\ln^3\left(2^{2\log\dim(S_j) + \operatorname{Max}(f_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}} (f_\mathrm{min})} + e^2\right),
\end{aligned}
\end{align}
with $\mathcal{Q}$ being the set of all distributions on $\mathcal{D}$ that could be produced by applying some EAT channel to some state.
\end{proposition}
\noindent Informally, the Markov conditions
impose the requirement that the register $T_j$ does not ``leak any information'' about the previous registers $S_{\upto{j-1}}$ beyond what is already available from $T_{\upto{j-1}} E$. (Without this Markov condition, one could for instance have channels such that $T_j$ simply contains a copy of $S_{j-1}$, in which case it would still be possible to derive a nontrivial min-tradeoff function, but the conclusion of the entropy accumulation theorem would be trivially false.) There is also an analogous EAT statement regarding the max-entropy~\cite{DFR20}, using a max-tradeoff function, though we will be using it slightly differently and will elaborate further on it at that point.
We now describe how the EAT can be used to prove Theorem~\ref{th:rawHmin}. First, note that to prove Theorem~\ref{th:rawHmin} it would be sufficient to consider only the registers $\str{A}{\str{B}'}\str{X}\str{Y}{\str{C}'} E$ of the state $\rho$ (the conditioning event $\OpPE$ is determined by ${\str{C}'}$ alone). The reduced state on these registers is the same as that at the point when Step~\ref{step:EATmap} of Protocol~\ref{prot:virtual} has finished looping over $j$, since the subsequent steps do not change these registers, and thus we can equivalently study that state in place of $\rho_{|\OpPE}$.
From this point onwards, all smoothed min- or max-entropies refer to that state conditioned on $\OpPE$ (and normalized),
hence for brevity we will omit the subscript specifying the state.
Each iteration of Step~\ref{step:EATmap} of Protocol~\ref{prot:virtual} can be treated as a channel $\mathcal{M}_j$ in a sequence of EAT channels, by considering it to be a channel performing the following operations:
\begin{enumerate}
\item Alice generates $X_j$ as specified in Step~\ref{step:EATmap}.
Conditioned on the value of $X_j$, Alice's device performs some measurement on its share of the stored quantum state $R_{j-1}$ (which includes any memory retained from previous rounds), then performs sifting and noisy preprocessing on the outcome, storing the final result in register $A_j$.
\item Bob's device behaves analogously, producing the registers $Y_j$ and $B'_j$ (we will not need to consider $B_j$).
\item The value of $C'_j$ is computed from $A_j B'_j X_j Y_j$.
\end{enumerate}
We highlight that in the above description of $\mathcal{M}_j$, the only ``unknowns'' are the measurements it performs on the input state on $R_{j-1}$ --- all other operations are taken to be performed in trusted fashion.
(This is reasonable because these measurements and the stored states are the only untrusted aspects in the true protocol.) If we had simply considered completely arbitrary channels $\mathcal{M}_j$ producing the respective registers, it would not be possible to
make a nontrivial security statement about the output.
Identifying $C'_j$ with $D_j$, $A_jB'_j$ with $S_j$, and $X_jY_j$ with $T_j$ in Definition~\ref{def:EATchann}, we see that these channels $\mathcal{M}_j$ indeed form a valid sequence of EAT channels: $C'_j$ is determined from $A_j B'_j X_j Y_j$ in the manner specified by Eq.~\eqref{eq:nodisturb}. Additionally, the state they produce always fulfills the Markov conditions, because the values of $X_jY_j$ in each round are generated independently of all preceding registers.
Intuitively, it seems that we could now use the EAT to bound $H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}E)$.
However, there is a technical issue: to apply the EAT, the event $\OpPE$ must be defined entirely in terms of the (classical) registers that appear in the smoothed min-entropy term that we are bounding, which is not a condition satisfied by the registers $\str{A}\str{X}\str{Y}$ alone. This is where the register ${\str{B}'}$ comes into play, following the same approach as~\cite{ARV19}: by a chain rule for the min- and max-entropies (\cite{VDT13} or \cite{Tom16} Eq.~(6.57)),
we have for any $\eps_s' + 2\eps_s'' < \eps_s$:
\begin{align}
H_\mathrm{min}^{\eps_s}(\str{A}|\str{X}\str{Y}E) &\geq H_\mathrm{min}^{\eps_s'}(\str{A}{\str{B}'}|\str{X}\str{Y}E) - H_\mathrm{max}^{\eps_s''}({\str{B}'}|\str{A}\str{X}\str{Y}E) - 3\smf{\eps_s-\eps_s'-2\eps_s''} \nonumber\\
&\geq H_\mathrm{min}^{\eps_s'}(\str{A}{\str{B}'}|\str{X}\str{Y}E) - H_\mathrm{max}^{\eps_s''}({\str{B}'}|\str{X}\str{Y}E) - 3\smf{\eps_s-\eps_s'-2\eps_s''}.
\label{eq:chain}
\end{align}
The $H_\mathrm{max}^{\eps_s''}({\str{B}'}|\str{X}\str{Y}E)$ term admits a fairly simple bound, as follows: consider a sequence of EAT channels $\widetilde{\mathcal{M}}_j$ that are identical to $\mathcal{M}_j$ except that they do not produce the registers $A_jC'_j$. As before, these maps obey the required Markov conditions.
In addition, recall that for every round the register $B'_j$ is deterministically set to $\perp$ whenever $y\in\{0,1\}$ (which happens with probability $1-\gamma$), hence we always have
\begin{align}
H(B'_j|X_j Y_j R)_{(\widetilde{\mathcal{M}}_j\otimes\idmap_{R})(\omega_{R_{j-1}R})}
= \sum_{y} \pr{Y_j=y} H(B'_j|X_j R;Y_j =y)_{(\widetilde{\mathcal{M}}_j\otimes\idmap_{R})(\omega_{R_{j-1}R})}
\leq \gamma .
\label{eq:fmax}
\end{align}
This means we can apply the max-entropy version of the EAT\footnote{Here we shall use the results from~\cite{DFR20}, because for a constant tradeoff function, this turns out to yield a slightly better bound as compared to the version of the EAT~\cite{DF19} stated here.
Strictly speaking, the reasoning used here is not a direct application of the EAT, because once again, the event $\OpPE$ is not defined on the registers ${\str{B}'}\str{X}\str{Y}$ alone (attempting to address this by including $\str{A}$ in the conditioning registers could result in the Markov conditions not being fulfilled). Fortunately, the bound \eqref{eq:fmax} holds for our maps $\widetilde{\mathcal{M}}_j$ even without a constraint on the output distribution. Hence the reasoning is as follows, in terms of the equations and lemmas in~\cite{DFR20}: first apply Eq.~(32) \emph{without} the event-conditioning term (this is valid since \eqref{eq:fmax} holds without constraints), {then} condition on $\OpPE$ using
Lemma~B.5 (noting that $\rho_{{\str{B}'}\str{X}\str{Y}E}
= \pr{\OpPE}(\rho_{|\OpPE})_{{\str{B}'}\str{X}\str{Y}E} + \pr{\OpPEnot}(\rho_{|\OpPEnot})_{{\str{B}'}\str{X}\str{Y}E}$),
and finally apply Lemma~B.10 to obtain Eq.~\eqref{eq:Hmaxbnd}.
(Alternatively, one could use ${\str{C}'}$ instead of ${\str{B}'}$. This was done in~\cite{MvDR+19} to slightly improve the bound in the block analysis, but it does not make a difference in our analysis. However, using ${\str{C}'}$ would seem to make it harder to sharpen the slightly crude bound used to obtain Eq.~\eqref{eq:gproof}.)
} with a \emph{constant} max-tradeoff function of value $\gamma$. Letting $V'=2\log(1+2
\dim(B')
) = 2\log7$,
this yields the following bound for any $\alpha' \in (1,1+2/V')$:
\begin{align}
H_\mathrm{max}^{\eps_s''}({\str{B}'}|\str{X}\str{Y}E) < n\gamma + n\left(\frac{\alpha'-1}{4}\right)V'^2 + \frac{\smf{\eps_s''}}{\alpha'-1} + \frac{\alpha'}{\alpha'-1}\log\frac{1}{\pr{\OpPE}}.
\label{eq:Hmaxbnd}
\end{align}
The bulk of our task is to bound the $H_\mathrm{min}^{\eps_s'}(\str{A}{\str{B}'}|\str{X}\str{Y}E)$ term. To do so, we will need an appropriate min-tradeoff function, which we shall now construct.
\subsubsection{Min-tradeoff function}
\label{sec:fmin}
Consider an arbitrary state of the form $(\mathcal{M}_j\otimes\idmap_{R})(\omega_{R_{j-1}R})$.
In this section, all entropies will be computed with respect to this state, and hence for brevity we will omit the subscript specifying the state.
We first note that
\begin{align}
H(A_j B'_j | X_j Y_j \Fj R)
&= \frac{1-\gamma}{4}
\sum_{z\in\{0,1\}} H(A_j | \Fj R;X_j = Y_j = z) \nonumber \\
& \qquad + \frac{\gamma}{4}
\sum_{y\in\{2,3\}} \sum_{x\in\{0,1\}}
H(A_j B'_j | \Fj R;X_j = x, Y_j = y), \label{eq:1rndmix}
\end{align}
where we have used the fact that $H(A_j B'_j | \Fj R;X_j = x, Y_j = y) = 0$ when $(x,y)=(0,1) \text{ or } (1,0)$, and $H(A_j B'_j | \Fj R;X_j = x, Y_j = y) = H(A_j | \Fj R;X_j = x, Y_j = y)$ when $(x,y)=(0,0) \text{ or } (1,1)$.
Let $w$ denote the probability that the state wins the CHSH game, conditioned on the game being played.
Then by applying the simple but somewhat crude bound\footnote{This marks a point where the analysis could be slightly sharpened, in that if we had a tight bound on the ``two-party entropies'' $H(A_j B'_j | \Fj R;X_j = x, Y_j = y)$ rather than just the ``one-party entropies'' $H(A_j | \Fj R;X_j = x, Y_j = y)$, we could improve the second term in~\eqref{eq:gproof}. However, note that it would only improve the keyrate by $O(\gamma)$, because of the $\gamma$ prefactor on that term.} $H(A_j B'_j | \Fj R;X_j = x, Y_j = y) \geq H(A_j | \Fj R;X_j = x, Y_j = y)$ to the terms in the second sum in Eq.~\eqref{eq:1rndmix}, we get the bound\footnote{Recall that noisy preprocessing is not applied to the rounds with $Y_j \in \{2,3\}$.}
\begin{align}
H(A_j B'_j | X_j Y_j \Fj R)
\geq \frac{1-\gamma}{2} \lin_\p(w) + \frac{\gamma}{2} \sum_{y\in\{2,3\}} \lin_0(w)
= \g(w).
\label{eq:gproof}
\end{align}
We can now use the function $\g$ to construct a min-tradeoff function $f_\mathrm{min}$, with the domain of $f_\mathrm{min}$ being distributions on $C'_j$ (recall that this register is set to $\perp$ if $Y_j\in\{0,1\}$,
and otherwise is set to $0$ or $1$ if the CHSH game is lost or won respectively). First observe that the channel is an \emph{infrequent-sampling channel} in the sense described in~\cite{DF19,LLR+19}. By the argument in Appendix A.7 of~\cite{LLR+19}, a valid min-tradeoff function $f_\mathrm{min}$ for the channel is given by
the (unique) affine function specified by the following values (in a minor abuse of notation, here we interpret $\g$ as a function of a distribution instead of a winning probability):
\begin{align}
f_\mathrm{min}(\delta_c) =
\begin{cases}
\frac{1}{\gamma} \g(\delta_c) + \left(1-\frac{1}{\gamma}\right)\beta & \text{ if } c\neq\perp \\
\beta & \text{ if } c = \perp
\end{cases}
\, ,
\label{eq:fmin}
\end{align}
where $\beta\in[\g\left(\delta_0\right),\g\left(\delta_1\right)]$ is a constant that can be chosen to optimize the keyrate. (Intuitively, this function is constructed simply by noting that the maps $\mathcal{M}_j$ can only produce distributions that lie in the slice of the probability simplex specified by the constraint
$\mathrm{Pr}
[C'_j = \perp] = 1-\gamma$, and hence the min-tradeoff function is free to take any value for distributions outside of this slice, recalling that we take the infimum of an empty set to be $+\infty$. For distributions within this slice, we know that $\g$ is an affine lower bound on the entropy as a function of the winning probability, and hence we can just set $f_\mathrm{min}$ equal to $\g$ (up to a domain rescaling) on this slice. Any $f_\mathrm{min}$ constructed this way is precisely of the form described in Eq.~\eqref{eq:fmin}, with $\beta$ being a constant determining its value on all distributions outside of the $\mathrm{Pr}[C'_j = \perp] = 1-\gamma$ slice.)
As shown in~\cite{LLR+19}, the min-tradeoff function constructed this way satisfies
\begin{align}
\begin{gathered}
\operatorname{Max}(f_\mathrm{min}) =
\max\left\{
\frac{1}{\gamma}\operatorname{Max}(\g) + \left(1-\frac{1}{\gamma}\right)\beta
, \beta \right\},
\qquad
\operatorname{Min}_{\mathcal{Q}_{f}} (f_\mathrm{min}) = \operatorname{Min}_{\mathcal{Q}_{\g}} (\g), \\
\operatorname{Var}_{\mathcal{Q}_{f}}(f_\mathrm{min}) \leq \sup_{q\in\mathcal{Q}_{\g}} \sum_{c\in\{0,1\}} \frac{q(c)}{\gamma} \left(\beta - \g(\delta_c)\right)^2,
\end{gathered}
\end{align}
where $\mathcal{Q}_{f}$ denotes
the set of distributions on $C'_j$ such that $\mathrm{Pr}[C'_j = \perp] = 1-\gamma$ and $\mathrm{Pr}[C'_j = 1] \in [\gamma(2-\sqrt{2})/4, \gamma(2+\sqrt{2})/4]$, while $\mathcal{Q}_{\g}$ denotes
the set of all distributions on
the alphabet $\{0,1\}$
such that $\mathrm{Pr}[1] \in [(2-\sqrt{2})/4, (2+\sqrt{2})/4]$.
For the specific $\g$ and range of $\beta$ that we consider,
these expressions simplify to Eq.~\eqref{eq:minmaxvarbounds}, where
we have solved the optimization $\sup_{q\in\mathcal{Q}_{\g}}$ in the bound on $\operatorname{Var}_{\mathcal{Q}_{f}}(f_\mathrm{min})$ by observing that it is an affine function of the distribution $q$, and the set $\mathcal{Q}_{\g}$ we use here is essentially
a line segment (in a $1$-dimensional probability simplex).
\subsubsection{Final min-entropy bound}
The event $\OpPE$ is defined by the conditions
$\operatorname{freq}_{\str{c}'}(1)\geq (w_\mathrm{exp}-\delta_\mathrm{tol})\gamma$ and
$\operatorname{freq}_{\str{c}'}(0)\leq (1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma$. Hence for all
${\str{c}'} \in \OpPE$,
we have (since $f_\mathrm{min}$ is affine)
\begin{align}
f_\mathrm{min}(\operatorname{freq}_{\str{c}'})
&= \operatorname{freq}_{\str{c}'}(0) \left(\frac{1}{\gamma} \g\left(\delta_0\right) + \left(1-\frac{1}{\gamma}\right)\beta\right)
+ \operatorname{freq}_{\str{c}'}(1) \left(\frac{1}{\gamma} \g\left(\delta_1\right) + \left(1-\frac{1}{\gamma}\right)\beta\right)
+ \operatorname{freq}_{\str{c}'}(\perp) \beta \nonumber\\
&= \operatorname{freq}_{\str{c}'}(0) \left(\frac{1}{\gamma} \g\left(\delta_0\right) -\frac{1}{\gamma}\beta\right)
+ \operatorname{freq}_{\str{c}'}(1) \left(\frac{1}{\gamma} \g\left(\delta_1\right) -\frac{1}{\gamma}\beta\right)
+ \beta \nonumber\\
&\geq (1-w_\mathrm{exp}+\delta_\mathrm{tol}) \left(\g\left(\delta_0\right) - \beta\right)
+ (w_\mathrm{exp}-\delta_\mathrm{tol}) \left(\g\left(\delta_1\right) - \beta\right)
+ \beta \nonumber\\
&= (1-w_\mathrm{exp}+\delta_\mathrm{tol})\g\left(\delta_0\right)
+ (w_\mathrm{exp}-\delta_\mathrm{tol})\g\left(\delta_1\right)
\nonumber\\
&= \g(w_\mathrm{exp}-\delta_\mathrm{tol}),
\label{eq:fminPEineq}
\end{align}
where the inequality holds because $\beta\in[\g\left(\delta_0\right),\g\left(\delta_1\right)]$, and in the last line we use the fact that $\g$ is affine and revert to interpreting it as a function of winning probability. (We remark that if we fix $\beta=\g(1)$, then in fact this inequality can be derived using only the $\operatorname{freq}_\str{c}(0)$ condition, following~\cite{DF19}. Hence in principle one could sacrifice the option of optimizing $\beta$ in exchange for reducing the number of checks to perform in the protocol, which improves the completeness parameters.)
Therefore, we can choose $h=\g(w_\mathrm{exp}-\delta_\mathrm{tol})$ in the EAT statement (Prop.~\ref{prop:EAT}) to conclude that the state conditioned on $\OpPE$ satisfies
\begin{align}
H_\mathrm{min}^{\eps_s'}(\str{A}{\str{B}'} | \str{X}\str{Y}E) &> n
\g(w_\mathrm{exp}-\delta_\mathrm{tol})
- n\frac{(\alpha-1)\ln 2}{2}V^2 - n(\alpha-1)^2K_\alpha^2
\nonumber \\
&\qquad
- \frac{\smf{\eps_s'}}{\alpha-1} - \frac{\alpha}{\alpha-1}\log\frac{1}{\pr{\OpPE}}, \label{eq:EATbound}
\end{align}
where $\smf{\epsilon},V,K_\alpha$ are as defined in Eq.~\eqref{eq:VK}. Putting this together with Eqs.~\eqref{eq:chain} and~\eqref{eq:Hmaxbnd}, we finally obtain the bound in Theorem~\ref{th:rawHmin}.
\subsection{Scope of applicability}
\label{sec:nonCHSH}
In the above analysis, we have focused on a protocol that only uses the CHSH game. However, it would be possible to modify the analysis to account for arbitrary Bell inequalities, as was done in~\cite{BRC20}. Essentially, Alice and Bob would simply need to choose a different distribution of their input settings, corresponding to a different game being played. Furthermore, it is not strictly necessary for Alice to choose uniformly random inputs in the generation rounds --- as noted in~\cite{SGP+21}, she could instead choose some biased distribution of inputs. It would even be possible to consider applying different amounts of noisy preprocessing for the different inputs in generation rounds. All of these modifications would essentially correspond to finding an appropriate min-tradeoff function, which we describe how to do in the subsequent section. However, we show in Sec.~\ref{sec:bestbnd} that for the depolarizing-noise model at least, there is little to be gained by considering these modifications.
There is a technical issue that in order to implement the above modifications, Alice and Bob may need to know which rounds are test rounds and which rounds are generation rounds, if they need to choose different input distributions in the two cases. (In Protocol~\ref{prot:DIQKD}, this was not an issue because the CHSH game is played with uniformly random inputs, and we also used uniformly random inputs for the generation rounds, so there is no difference in the input distributions between test and generation rounds.) However, this could in principle be addressed by using a short pre-shared key in order to choose the test rounds --- we describe this in more detail in Sec.~\ref{sec:pubchann}.
\section{Single-round bound}
\label{sec:1rndbnd}
We now explain how to derive the function $\lin_\p$ as described in Eq.~\eqref{eq:linbnd}. To reduce clutter, we will use slightly different notation in this section, which should not be confused with the earlier notation. In particular, we will omit the~~$\bar{ }$~~accents from the registers $\bar{A}\bar{B}\bar{E}$, and we will be using hermitian operators $A_x,B_y$ that should not be confused with the registers $A_j,B_j$ in the earlier notation.
As described previously, consider an arbitrary state $\rho_{ABE}$ and possible measurements on the $A$ and $B$ subsystems, indexed by $x$ and $y$ respectively.
For the purpose of bounding the entropy, all measurements can be assumed projective by considering a suitably chosen simultaneous Stinespring dilation; see e.g.~\cite{arx_TSG+19}.
Let $\pvm_{a|x}$ (resp.~$\pvm_{b|y}$) denote the projector corresponding to outcome $a$ (resp.~$b$) from measurement $x$ (resp.~$y$). We first consider a somewhat more general setting than that required for the above security proof; namely, we aim to find a lower bound on the conditional entropies, knowing only that the state and measurements produce the values $\constr_j$ for some observables $\Gamma_j(\pvm_{a|x},\pvm_{b|y})\defvar \sum_{abxy} c^{(j)}_{abxy} \pvm_{a|x}\otimes\pvm_{b|y}$ defined by fixed coefficients $c^{(j)}_{abxy}\in\mathbb{R}$. (For instance, these could be the values of some Bell expressions, or even simply the entire set of output probabilities.) Furthermore, we allow for the possibility that the inputs are not uniformly random.
Making this precise, we aim to find the function
\begin{align}
\breve{\lin}_\p(\vec{\constr}) \defvar \,
\begin{gathered}
\inf_{\rho_{ABE},\pvm_{a|x},\pvm_{b|y}} \sum_{x\in\{0,1\}} \keyw_x H(
\hat{A}_x
|{E}) \\
\suchthat\; \tr{\Gamma_j(\pvm_{a|x},\pvm_{b|y}) \rho_{AB}} = \constr_j \quad \forall j
\end{gathered},
\label{eq:mainopt}
\end{align}
where $\keyw_x \geq 0$ are coefficients
that depend on the input distributions~\cite{SGP+21}.
(For the security proof of Protocol~\ref{prot:DIQKD}, we would only need to consider a single $\Gamma_j(\pvm_{a|x},\pvm_{b|y})$, which describes the probability of winning the CHSH game. Also, we simply have $\keyw_0=\keyw_1=1/2$, since we want a bound of the form~\eqref{eq:linbnd}.)
Without loss of generality, we can restrict the optimization to pure $\rho_{ABE}$.
By considering the effect of Eve using mixtures of strategies, it is easily seen that $\breve{\lin}_\p$ must be convex. It can thus be shown that for any $\vec{\constr}$ in the interior of the set of values achievable by quantum theory, this optimization is in fact equal to its Lagrange dual (see e.g.~\cite{arx_TSG+19} for a detailed explanation):
\begin{align}
\breve{\lin}_\p(\vec{\constr}) = \sup_{\vec{\lambda}}
\inf_{\rho_{ABE},\pvm_{a|x},\pvm_{b|y}} \left( \sum_{x\in\{0,1\}} \keyw_x H(\hat{A}_x|{E})\right) - \vec{\lambda} \cdot \left(\tr{\vec{\Gamma}(\pvm_{a|x},\pvm_{b|y}) \rho_{AB}} - \vec{\constr} \right) .
\label{eq:dual}
\end{align}
Since the optimization over $\vec{\lambda}$ is a supremum, it follows that for any value of $\vec{\lambda}$, we have a lower bound on $\breve{\lin}_\p(\vec{\constr})$ of the form
\begin{align}
\breve{\lin}_\p(\vec{\constr}) \geq
\vec{\lambda}\cdot\vec{\constr} + c_{\vec{\lambda}},
\label{eq:duallin}
\end{align}
where
\begin{align}
c_{\vec{\lambda}} \defvar \inf_{\rho_{ABE},\pvm_{a|x},\pvm_{b|y}} \left(\sum_{x\in\{0,1\}} \keyw_x H(\hat{A}_x|{E})\right) - \vec{\lambda} \cdot \tr{\vec{\Gamma}(\pvm_{a|x},\pvm_{b|y}) \rho_{AB}}.
\label{eq:intercept}
\end{align}
Importantly, such a lower bound is automatically affine with respect to $\vec{\constr}$, and hence we can take it as a possible choice of $\lin_\p$ for the security proof.
We thus see that in order to find an affine $\lin_\p$ for use in the security proof, it suffices to choose some $\vec{\lambda}$ and compute (or lower-bound) the value of $c_{\vec{\lambda}}$ as defined by Eq.~\eqref{eq:intercept}.
In addition, this approach yields essentially tight bounds, in the sense that taking the supremum of the bounds given by all possible $\vec{\lambda}$ returns the value of $\breve{\lin}_\p(\vec{\constr})$, by Eq.~\eqref{eq:dual}. We also note that when there are multiple constraints, for any given $\vec{\constr}$ the corresponding optimal $\vec{\lambda}$ yields a single Bell expression that certifies the same bound on the entropy as all the original constraints. When applying this bound in the EAT, the task of choosing $\vec{\lambda}$ here is equivalent to the task in~\cite{ARV19,LLR+19} of choosing a tangent point to the rate curves.
The $H(\hat{A}_x|E)$ term in the optimization can be rewritten based on the approach in~\cite{WLC18}. Specifically, we observe that the state produced on $\hat{A}_x E$ by performing Alice's measurement $x$ on $\rho_{ABE}$ and then applying noisy preprocessing could also be obtained by the following process: append an ancilla $T$ in the state
\begin{align}
\ket{\phi_\p}_T \defvar \sqrt{1-\p} \ket{0}_T + \sqrt{\p} \ket{1}_T,
\end{align}
apply a pinching channel $\mathcal{P}(\sigma_T)\defvar\sum_t \pure{t}_T \sigma_T \pure{t}_T$ to $T$, then perform a measurement on $AT$ described by the projectors
\begin{align}
\tilde{\pvm}_{a|x} \defvar \pvm_{a|x}\otimes\pure{0}_T + \pvm_{a\oplus 1|x}\otimes\pure{1}_T,
\label{eq:NPPprojs}
\end{align}
and store the outcome in $\hat{A}_x$ without further processing. However,
the pinching channel on $T$ can in fact be omitted, because the subnormalized conditional states produced on $E$ would still be the same without it (in the following, we leave some tensor factors of $\id$ implicit for brevity, and consider an arbitrary state $\sigma_{ABET}$ before applying the pinching channel):
\begin{align}
&\tr[ABT]{\tilde{\pvm}_{a|x}
\,\mathcal{P}_T (\sigma_{ABET})\,
\tilde{\pvm}_{a|x}}\nonumber\\
=&\tr[ABT]{\tilde{\pvm}_{a|x} \left(\sum_{t}
\pure{t}_T
\sigma_{ABET}
\pure{t}_T
\right) \tilde{\pvm}_{a|x}}\nonumber\\
=& \sum_{t} \tr[ABT]{(\pvm_{a\oplus t|x}\otimes\pure{t}_T) \sigma_{ABET}(\pvm_{a\oplus t|x}\otimes\pure{t}_T) }\nonumber\\
=& \sum_{t,t'} \tr[ABT]{(\pvm_{a\oplus t|x}\otimes\pure{t}_T) \sigma_{ABET}\left(\pvm_{a\oplus t'|x}\otimes\pure{t'}_T\right) }\nonumber\\
=& \tr[ABT]{\tilde{\pvm}_{a|x} \sigma_{ABET}\tilde{\pvm}_{a|x} }.
\end{align}
Hence we no longer consider the pinching channel on $T$, i.e.~we simply study the situation where we immediately perform the projective measurement~\eqref{eq:NPPprojs} on the state $\hat{\rho}_{ABET} \defvar \rho_{ABE} \otimes \pure{\phi_\p}_T$ state and store the outcome in register $\hat{A}_x$. As mentioned previously, we can take $\rho_{ABE}$ to be pure without loss of generality, in which case $\hat{\rho}_{ABET}$ is pure as well\footnote{It was important to remove the pinching channel because if $\hat{\rho}_{ABET}$ were a mixed state, then Eq.~\eqref{eq:Hduality} would be replaced by $D(\hat{\rho}_{ABT} \Vert \mathcal{Z}_x(\hat{\rho}_{ABT})) = H(\hat{A}_x|EE')$, where $E'$ purifies $\hat{\rho}_{ABET}$. While $H(\hat{A}_x|EE')$ is a valid lower bound on $H(\hat{A}_x|E)$, it also turns out to be a trivial one, because the value of $T$ is ``copied'' into $E'$,
removing the advantage of noisy preprocessing (in which Eve is not supposed to know whether the bit-flip has occurred).}, and must hence obey the following relation (derived in e.g.~\cite{Col12}):
\begin{align}
H(\hat{A}_x|E) = D(\hat{\rho}_{ABT} \Vert \mathcal{Z}_x(\hat{\rho}_{ABT})) = D(\mathcal{G}({\rho}_{AB}) \Vert \mathcal{Z}_x(\mathcal{G}({\rho}_{AB}))),
\label{eq:Hduality}
\end{align}
where
\begin{align}
\mathcal{G}(\sigma_{AB}) &\defvar \sigma_{AB} \otimes \pure{\phi_\p}_T, \\
\mathcal{Z}_x(\sigma_{ABT}) &\defvar \sum_a (\tilde{\pvm}_{a|x}\otimes\id_{B}) \sigma_{ABT} (\tilde{\pvm}_{a|x}\otimes\id_{B}) .
\end{align}
Importantly, this expression for $H(\hat{A}_x|E)$ is entirely in terms of the reduced state $\rho_{AB}$, and is convex with respect to $\rho_{AB}$. (We give an alternative expression in Appendix~\ref{app:altNPP}, which may be of use in other situations.)
We remark that the above analysis was fairly general, in that in fact it applies to DI scenarios with arbitrary numbers of inputs and outputs, as long as Alice's key-generating measurements still only have 2 outcomes.
In particular, the approaches described in~\cite{arx_TSG+19,BFF21} are able to yield bounds on these optimizations when there is no noisy preprocessing, and it may be useful to study whether the above map for describing noisy preprocessing allows one to apply those approaches to this scenario. However, we leave this question for future work.
We now specialize to 2-input 2-output scenarios. In such cases,
we have the following ``qubit reduction''~\cite{PAB+09,HST+20}: if
one finds
a convex function that lower-bounds the right-hand-side of Eq.~\eqref{eq:mainopt} with its optimization restricted to states $\rho_{ABE}$ of dimension $2\times2\times4$ and Pauli measurements,
then this function is also a lower bound on
$\breve{\lin}_\p$.
We note (by following the same arguments as before, but with the optimization domain restricted) that picking some $\vec{\lambda}$ and solving the optimization~\eqref{eq:intercept} over such states and measurements yields such a lower bound via Eq.~\eqref{eq:duallin}, which is trivially convex since it is affine. Hence it suffices to consider the optimization~\eqref{eq:intercept} restricted to such states and measurements. Furthermore, the resulting bounds are still tight in the same sense as before, in that taking the supremum over choices of $\vec{\lambda}$ yields the convex envelope of this restricted version of the optimization~\eqref{eq:mainopt} (except possibly for $\vec{\constr}$ not in the interior of the quantum-achievable values).
Since we have reduced the analysis to Pauli measurements, it is convenient to interpret the measurements as producing values in $\pm1$ instead of $\mathbb{Z}_2$, and define the corresponding hermitian observables $A_x \defvar \sum_a a \pvm_{a|x}$ and $B_y \defvar \sum_b b \pvm_{b|y}$. For 2-input 2-output scenarios, the full probability distribution $\pr{ab|xy}$ (subject to the no-signalling constraints) is completely parametrized by the 4 correlators $\expvaltr{(A_x \otimes B_y)}$ and the 4 marginals $\expvaltr{(A_x \otimes \id)}$, $\expvaltr{(\id \otimes B_y)}$. Without loss of generality, one can assume that the marginals are zero~\cite{PAB+09,HST+20}\footnote{Essentially, this is because one could consider a virtual \emph{symmetrization step} in which Alice and Bob jointly flip their outputs using a uniformly random public bit, forcing their marginals to be zero. As argued in~\cite{SR08,PAB+09,HST+20}, the entropy of their original outputs conditioned on Eve's side-information is equal to the entropy of their symmetrized outputs conditioned on Eve's side-information and the publicly communicated bit, and hence we can analyze the latter instead (note that the virtual symmetrization does not need to be physically performed; it merely serves as an intermediate construction to study the entropies).}, hence it suffices to focus on the correlators $\expvaltr{(A_x \otimes B_y)}$. In terms of the optimization~\eqref{eq:mainopt}, this means that it suffices to consider situations where there are 4 constraints, corresponding to the observables
\begin{align}
\Gamma_{xy}(\pvm_{a|x},\pvm_{b|y}) = A_x \otimes B_y,
\quad \text{ for } x,y\in\{0,1\}.
\end{align}
Considering other forms of constraints in 2-input 2-output scenarios is essentially equivalent to making specific choices of Lagrange multipliers $\vec{\lambda}$ for this 4-constraint formulation. For instance, in Protocol~\ref{prot:DIQKD} we only impose a constraint based on the CHSH value\footnote{Here, by CHSH value we mean the quantity $\constr = \tr{(A_0\otimes B_0 + A_0\otimes B_1 + A_1\otimes B_0 - A_1\otimes B_1) \rho_{AB}}$. This is related to the probability $w$ of winning the CHSH game by the simple equation $\constr = 8w-4$, so an affine function of $\constr$ is easily converted to an affine function of $w$.}, which is equivalent to restricting to Lagrange-multiplier combinations of the form $(\lambda_{00},\lambda_{01},\lambda_{10},\lambda_{11}) = (\lambda,\lambda,\lambda,-\lambda)$ for some $\lambda\in\mathbb{R}$.
We still have the freedom to choose the basis in which to express the optimization. Following~\cite{SGP+21}, we can use the measurement axes of Alice's two Pauli measurements to define the $X$-$Z$ plane on her system, taking $A_0=Z$ and $A_1 = \cos({\theta_A})Z + \sin({\theta_A})X$ for some ${\theta_A} \in [0,\pi]$
(values of ${\theta_A}$ in $[\pi,2\pi]$ can be brought into this range by rotating our axis choice by $\pi$ around the $Z$-axis).
Analogously, we can choose a basis for Bob such that $B_0=Z$ and $B_1 = \cos({\theta_B})Z + \sin({\theta_B})X$ with ${\theta_B} \in [0,\pi]$. (This is a different basis choice from the one in~\cite{PAB+09,HST+20} that allows a reduction to \emph{Bell-diagonal}
$\rho_{AB}$. That choice involves more parameters for the measurements, but fewer parameters for the state.)
With this in mind, we rewrite the optimization~\eqref{eq:intercept} as
\begin{align}
\begin{gathered}
\min_{{\theta_A}} \min_{{\theta_B}}
\min_{\rho_{AB}}
F_\mathrm{obj}({\theta_A},{\theta_B},\rho_{AB}),
\text{ where } \\
F_\mathrm{obj}({\theta_A},{\theta_B},\rho_{AB}) \defvar \left( \sum_{x\in\{0,1\}} \keyw_x D(\mathcal{G}({\rho}_{AB}) \Vert \mathcal{Z}_x(\mathcal{G}({\rho}_{AB})))\right)
- \vec{\lambda} \cdot \tr{\vec{\Gamma}(
{\theta_A},{\theta_B}
) \rho_{AB}}.
\end{gathered}
\end{align}
(The
minima are
attained because the objective function is continuous and the domain is compact.)
Furthermore, the objective function is invariant under the substitutions $\rho_{AB} \to (Y \otimes Y) \rho_{AB} (Y \otimes Y)$ and $\rho_{AB} \to \rho_{AB}^*$, which implies~\cite{PAB+09} we can restrict the optimization to $\rho_{AB}$ that are ``almost'' Bell-diagonal --- specifically, with respect to the Bell basis $\{\ket{\Phi^+}, \ket{\Psi^-}, \ket{\Phi^-}, \ket{\Psi^+}\}$ where $\ket{\Phi^\pm} = (\ket{00}\pm\ket{11})/\sqrt{2}$ and $\ket{\Psi^\pm} = (\ket{01}\pm\ket{10})/\sqrt{2}$, we can take
\begin{align}
\rho_{AB} =
\begin{pmatrix}
L_{\Phi^+} & \ell_1 & 0 & 0 \\
\ell_1 & L_{\Psi^-} & 0 & 0 \\
0 & 0 & L_{\Phi^-} & \ell_2 \\
0 & 0 & \ell_2 & L_{\Psi^+}
\end{pmatrix},
\label{eq:almostdiag}
\end{align}
for some $L_{\Phi^+} , L_{\Psi^-} , L_{\Phi^-} , L_{\Psi^+} , \ell_1 , \ell_2 \in \mathbb{R}$.
We now describe how each minimization can be tackled
when treating the parameters in the other minimizations as constants, then summarize how all these algorithms can be put together in a consistent manner, and argue that this approach indeed yields arbitrarily tight bounds.
\subsection{Minimization over Alice's measurement}
We tackle this minimization simply by applying a (uniform) continuity bound for ${\theta_A}$.
Specifically, for $\delta\in[0,\pi]$ we describe a monotone increasing function $\epsilon_\mathrm{con}(\delta)$ that bounds the change in the objective function when ${\theta_A}$ is replaced by ${\theta_A}+\delta$ (treating ${\theta_B}$ and $\rho_{AB}$ as constants).
Then for any set of intervals
of the form $\{[\theta_j-\delta_j,\theta_j+\delta_j]\}_j$ that covers the interval $[0,\pi]$, we would have
\begin{align}
\min_{{\theta_A}} F_\mathrm{obj}({\theta_A},{\theta_B},\rho_{AB}) \geq \min_j
F_\mathrm{obj}(\theta_j,{\theta_B},\rho_{AB}) - \epsilon_\mathrm{con}(\delta_j)
.
\end{align}
We apply this in practice by starting with a fairly ``coarse'' choice of intervals, then iteratively applying the process of deleting the interval that currently achieves the minimization over $j$, and replacing it with smaller intervals that cover the deleted interval.
To derive such a continuity bound, we first analyze the entropic term in the objective function, following~\cite{SBV+21} (with a minor modification to slightly improve the bound).
Take any pure initial state $\ket{\rho}_{ABE}$, and let $\sigma_{\hat{A}_1 BE}$ be the state obtained by performing a Pauli measurement along angle ${\theta_A}$ in the $X$-$Z$ plane on the $A$ register of $\ket{\rho}_{ABE}$, applying noisy preprocessing, then storing the result in the classical register $\hat{A}_1$ and tracing out $A$.
Let $\sigma'_{\hat{A}_1 BE}$ be the analogous state with ${\theta_A}$ replaced by ${\theta_A}+\delta$ for some $\delta \in [0,\pi]$. Our goal would be to bound $\left|H(\hat{A}_1| E)_{\sigma} - H(\hat{A}_1| E)_{\sigma'}\right|$.
We have $H(\hat{A}_1| E)_{\sigma}
= \sum_{\hat{a}_1} \pr{\hat{a}_1} H(\sigma_{E|\hat{A}_1=\hat{a}_1}) + H(\hat{A}_1)_{\sigma} - H(E)_{\sigma}
$, and analogously for $\sigma'$. Since the operations on $A$ do not affect $E$, we have $H(E)_{\sigma}=H(E)_{\sigma'}$. Also, for states of the form~\eqref{eq:almostdiag}, we have $\rho_{A}=\id/2$ and hence
$H(\hat{A}_1)_{\sigma} = H(\hat{A}_1)_{\sigma'} = 1$. This gives us
\begin{align}
\left|H(\hat{A}_1| E)_{\sigma} - H(\hat{A}_1| E)_{\sigma'}\right|
\leq \max_{\hat{a}_1} \left| \left(H(\sigma_{E|\hat{A}_1=\hat{a}_1}) - H(\sigma'_{E|\hat{A}_1=\hat{a}_1})\right)\right|,
\end{align}
so it suffices to bound the difference in entropies of the conditional states on $E$.
Now observe that exactly the same state $\sigma_{\hat{A}_1 B E}$ would have been produced if the initial state had been $\left(e^{i{\theta_A} Y_A/2} \otimes \id_{BE}\right)\ket{\rho}_{ABE}$ and the initial Pauli measurement were replaced by a $Z$ measurement.
Furthermore, the fact that $\rho_A=\id/2$ implies we can write $\ket{\rho}_{ABE} = \sum_a \ket{a}_A \ket{a}_{BE} / \sqrt{2}$, where $\{\ket{a}_A\}_a$ is the $Z$-eigenbasis of $A$ and $\{\ket{a}_{BE}\}_a$ are two orthonormal states on $BE$. This implies
\begin{align}
\left(e^{i{\theta_A} Y_A/2} \otimes \id_{BE}\right)\ket{\rho}_{ABE}
= \left(\id_A \otimes \left(e^{i{\theta_A} Y_{BE}/2}\right)^T\right)\ket{\rho}_{ABE}
= \frac{1}{\sqrt{2}}\sum_a \ket{a}_A \otimes
\rot{{\theta_A}}
\ket{a}_{BE},
\end{align}
where
$Y_{BE} \defvar i\ketbra{-1}{+1}_{BE}-i\ketbra{+1}{-1}_{BE}$
and $\rot{{\theta_A}}\defvar\left(e^{i{\theta_A} Y_{BE}/2}\right)^T$.
Performing the analogous analysis for
$\sigma'$,
we conclude that
\begin{align}
\begin{aligned}
\sigma_{BE|\hat{A}_1=\hat{a}_1} &=
(1-\p) \rot{{\theta_A}}\pure{\hat{a}_1}_{BE}\rot{{\theta_A}}^\dagger + \p \rot{{\theta_A}}\pure{-\hat{a}_1}_{BE}\rot{{\theta_A}}^\dagger, \\
\sigma'_{BE|\hat{A}_1=\hat{a}_1} &=
(1-\p) \rot{{\theta_A}+\delta}\pure{\hat{a}_1}_{BE}\rot{{\theta_A}+\delta}^\dagger + \p \rot{{\theta_A}+\delta}\pure{-\hat{a}_1}_{BE}\rot{{\theta_A}+\delta}^\dagger.
\end{aligned}
\end{align}
Therefore, we have
\begin{align}
F(\sigma_{E|\hat{A}_1=\hat{a}_1},\sigma'_{E|\hat{A}_1=\hat{a}_1}) &\geq F(\sigma_{BE|\hat{A}_1=\hat{a}_1},\sigma'_{BE|\hat{A}_1=\hat{a}_1}) \nonumber \\
&\geq (1-\p) \left|\bra{\hat{a}_1}\rot{{\theta_A}}^\dagger\rot{{\theta_A}+\delta}\ket{\hat{a}_1}\right| + \p \left|\bra{-\hat{a}_1}\rot{{\theta_A}}^\dagger\rot{{\theta_A}+\delta}\ket{-\hat{a}_1}\right| \nonumber \\
&\geq \left|\cos\frac{\delta}{2}\right|,
\end{align}
where the second inequality holds by concavity of fidelity,
and the third inequality is given by explicit calculation (see~\cite{SBV+21}).
This lets us apply a fidelity-based continuity bound~\cite{SBV+21}:
\begin{align}
\max_{\hat{a}_1} \left| \left(H(\sigma_{E|\hat{A}_1=\hat{a}_1}) - H(\sigma'_{E|\hat{A}_1=\hat{a}_1})\right)\right| \leq 4.023 \cos^{-1} F(\sigma_{E|\hat{A}_1=-1},\sigma'_{E|\hat{A}_1=-1}) \leq 2.012 \delta,
\label{eq:Hcontbnd}
\end{align}
where in the second inequality we used the condition $\delta \in [0,\pi]$.
(Numerical heuristics suggest that the true bound in Eq.~\eqref{eq:Hcontbnd} may simply be $\delta$, so there is some potential for improvement here, though the effect would be fairly small.
In Appendix~\ref{app:altcont} we present an approach based on trace distance instead of fidelity, but it appears to scale poorly at small $\delta$.)
As for the $\vec{\Gamma}$ term,
we note that
\newcommand{\cs}[1]{\cos(#1)}
\newcommand{\sn}[1]{\sin(#1)}
\begin{align}
&\left|\tr{\left(\vec{\lambda}\cdot\vec{\Gamma}({\theta_A}+\delta,{\theta_B})-\vec{\lambda}\cdot\vec{\Gamma}({\theta_A},{\theta_B})\right)\rho_{AB}}\right|
\nonumber\\
\leq& \norm{\vec{\lambda}\cdot\vec{\Gamma}({\theta_A}+\delta,{\theta_B}) - \vec{\lambda}\cdot\vec{\Gamma}({\theta_A},{\theta_B})}_\infty \nonumber\\
=& \norm{\left((\cs{{\theta_A}+\delta}Z + \sn{{\theta_A}+\delta}X) - (\cs{{\theta_A}}Z + \sn{{\theta_A}}X) \right) \otimes (\lambda_{10} B_0+\lambda_{11} B_1)}_\infty \nonumber\\
\leq& (|\lambda_{10}| + |\lambda_{11}|)\norm{
(\cs{{\theta_A}+\delta}Z + \sn{{\theta_A}+\delta}X) - (\cs{{\theta_A}}Z + \sn{{\theta_A}}X)
}_\infty \nonumber\\
=& (|\lambda_{10}| + |\lambda_{11}|)\sqrt{2-2\cos(\delta)},
\end{align}
where the last line follows from an explicit eigenvalue calculation.
Overall, this means that for $\delta\in[0,\pi]$ we can choose
\begin{align}
\epsilon_\mathrm{con}(\delta) =
2.012 \keyw_1
\delta + (|\lambda_{10}| + |\lambda_{11}|)\sqrt{2-2\cos(\delta)},
\label{eq:contbnd}
\end{align}
accounting for the $\keyw_1$ factor on the $H(\hat{A}_1| E)$ term. This bound is monotone increasing for $\delta\in[0,\pi]$, as required (so that it also bounds the change in entropy when the measurement angle is changed from ${\theta_A}$ to any value in the interval $[{\theta_A},{\theta_A}+\delta]$).
\subsection{Minimization over Bob's measurement}
The entropic term in the objective function has no dependence on Bob's measurement, so we only need to consider the $\vec{\Gamma}$ term.
In principle, this could be approached using the same argument as above, where we would arrive at the continuity bound
\begin{align}
&\left|\tr{\left(\vec{\lambda}\cdot\vec{\Gamma}({\theta_A},{\theta_B}+\delta)-\vec{\lambda}\cdot\vec{\Gamma}({\theta_A},{\theta_B})\right)\rho_{AB}}\right|
\leq
(|\lambda_{01}| + |\lambda_{11}|)\sqrt{2-2\cos(\delta)}.
\end{align}
However, some heuristic experiments indicate that the following approach (used in~\cite{SGP+21}) is more efficient: we can let $r_Z\defvar\cos({\theta_B})$ and $r_X\defvar\sin({\theta_B})$ and write
\begin{align}
\sum_{xy} \lambda_{xy} A_x \otimes B_y
&= \left(\sum_{x} \lambda_{x0} A_x\right) \otimes Z + \left(\sum_{x} \lambda_{x1} A_x\right) \otimes (r_Z Z + r_X X),
\end{align}
in which case the minimization over ${\theta_B}\in[0,\pi]$ is equivalent to minimizing over $(r_Z,r_X)$ that lie on the set $S_{\rightsemicirc} \defvar \{(r_Z,r_X)\mid r_Z^2+r_X^2=1 \text{ and } r_X \geq 0\}$ (i.e.~a semicircular arc). Crucially, the objective function is affine with respect to the vector $(r_Z, r_X)$. Hence if $V$ is any set of points such that $S_{\rightsemicirc}$ is contained in their convex hull $\operatorname{Conv}(V)$, we immediately have
\begin{align}
\min_{(r_Z,r_X)\in S_{\rightsemicirc}} F_\mathrm{obj}({\theta_A},(r_Z,r_X),\rho_{AB})
&\geq \min_{(r_Z,r_X)\in \operatorname{Conv}(V)} F_\mathrm{obj}({\theta_A},(r_Z,r_X),\rho_{AB}) \nonumber\\
&\geq \min_{(r_Z,r_X)\in V} F_\mathrm{obj}({\theta_A},(r_Z,r_X),\rho_{AB}),
\end{align}
because the minimum of an affine function over the convex hull of a set $V$ is always attained at an extremal point (which will be a point in $V$).
To apply this result, we start with a simple choice of the set $V$ (for instance,
in our code we use
$V = \{(1,0),(1,1),(-1,1),(-1,0)\}$) and find the point in $V$ that yields the minimum value. We then delete this point and replace it with two other points such that $S_{\rightsemicirc}$ is still contained in the convex hull, and iterate this process until a sufficiently tight bound is obtained (for instance, by checking that there is a feasible point of the optimization that is sufficiently close to the lower bound we have obtained).
\subsection{Minimization over states}
The minimization over states can be tackled by expressing it as a convex optimization and applying the \emph{Frank-Wolfe algorithm}~\cite{FW56}, as was observed in~\cite{WLC18}. For completeness, we now describe the method, with minor
modifications and clarifications for our specific scenario.
We emphasize that while this procedure is numerical, the bounds it returns are \emph{secure}, in the sense that it will never over-estimate the true value of the minimization problem.
We observe that (for fixed measurements, parametrized as described above) the minimization over states is the minimization of the convex function
\begin{align}
f_\mathrm{obj}(\rho) \defvar
\left(\sum_{x\in\{0,1\}} \keyw_x D(\mathcal{G}({\rho}) \Vert \mathcal{Z}_x(\mathcal{G}({\rho})))\right) - \vec{\lambda}\cdot \tr{\vec{\Gamma}({\theta_A},(r_Z,r_X)) \rho}.
\end{align}
This function is differentiable for all $\rho$ such that $\mathcal{G}(\rho) > 0$, with~\cite{WLC18}\footnote{The notation here follows that used in~\cite{WLC18}, in which for a function $f$ defined on matrices parametrized by their matrix entries ($\rho=\sum_{jk}\rho_{jk}\ketbra{j}{k}$), its derivative is $\nabla f(\rho) \defvar \sum_{jk} (\partial f/ \partial\rho_{jk}) \ketbra{j}{k}$.
Denoting the tangent to the graph of $f$ at $\rho$ as $t_{\rho}$, this definition of $\nabla f$ satisfies $t_{\rho}(\rho+\Delta) = f(\rho) + \tr{(\nabla f(\rho))^T \Delta}$.
}
\begin{align}
(\nabla f_\mathrm{obj}(\rho))^T = \left( \sum_{x\in\{0,1\}} \keyw_x \mathcal{G}^\dagger \!\left(\log\mathcal{G}(\rho) - \log\mathcal{Z}_x(\mathcal{G}(\rho))\right)\right) - \vec{\lambda}\cdot\vec{\Gamma}({\theta_A},(r_Z,r_X)),
\label{eq:fobjder}
\end{align}
where $\mathcal{G}^\dagger$ is the adjoint channel of $\mathcal{G}$. In practice, we will not need to explicitly compute $\mathcal{G}^\dagger$, because all subsequent arguments rely only on ``inner products'' $\tr{(\nabla f_\mathrm{obj}(\rho))^T\sigma}$, which can be rewritten as
\begin{align}
\tr{(\nabla f_\mathrm{obj}(\rho))^T\sigma} = \tr{\left( \sum_{x\in\{0,1\}} \keyw_x \left(\log\mathcal{G}(\rho) - \log\mathcal{Z}_x(\mathcal{G}(\rho))\right)\mathcal{G}(\sigma)\right) - \vec{\lambda}\cdot\vec{\Gamma}({\theta_A},(r_Z,r_X))\sigma}.
\end{align}
We also note that the optimization domain is a set defined by PSD constraints (namely, $\rho$ of the form~\eqref{eq:almostdiag} with $\rho\geq0$ and $\tr{\rho}=1$), which means that optimizing any affine function over this set is an SDP, which can be efficiently solved and bounded~\cite{BV04v8} via its dual value. Together with the explicit expression~\eqref{eq:fobjder} for the derivative of the objective function, this makes the optimization problem here a prime candidate for the {Frank-Wolfe algorithm}~\cite{FW56}, which yields \emph{secure} lower bounds on the true minimum value of the optimization.
The Frank-Wolfe algorithm is based on a simple geometric insight: for any point in the domain of a convex function $f_\mathrm{obj}$, the tangent hyperplane (or a supporting hyperplane, if $f_\mathrm{obj}$ is not differentiable at that point) to the graph of $f_\mathrm{obj}$ at that point yields an affine lower bound on $f_\mathrm{obj}$. Hence if we can minimize affine functions over the optimization domain, then any such tangent hyperplane lets us obtain a lower bound on the minimum of $f_\mathrm{obj}$ on the domain. In addition, it intuitively seems that taking the tangent at points closer to the true optimal solution should yield tighter lower bounds (though there are some technical caveats).
Thus in principle, one could simply perform some heuristic computations to get an estimate of where the true minimum lies, then take the tangent at that estimate and obtain the corresponding lower bound. For this optimization, however, we found that the results were fairly sensitive to deviations from the true optimum. To get the bounds to converge, we found it more efficient to use the standard algorithm,
which (under mild assumptions) can be proven to converge in order $O(1/k)$ after $k$ iterations:
\newcommand{\eps}{\epsilon}
\begin{savenotes}
\begin{algorithm}[htp]
\caption*{\textbf{Frank-Wolfe algorithm} (As presented in~\cite{WLC18})}
Let the domain of optimization be $\mathcal{D}$
and the acceptable gap between the feasible values and the lower bounds be $\epsilon_\mathrm{tol}>0$.
\begin{algorithmic}[1]
\State Set $k=0$
and heuristically find $\rho_0 \defvar \arg\min_{\rho\in\mathcal{D}}f_\mathrm{obj}(\rho)$.
\State\label{step:FWSDP}Solve the SDP
\begin{align}
\min_{\Delta} \tr{(\nabla f_\mathrm{obj}(\rho_k))^T \Delta} \text{ s.t. } \rho_k + \Delta \in \mathcal{D}, \label{eq:FWSDP}
\end{align}
which returns a feasible point $\Delta^*$
as well as a dual value $-\eps \leq 0$
that lower-bounds the true minimum of the SDP.
\State If $\eps \leq \epsilon_\mathrm{tol}$ then stop and return $f_\mathrm{obj}(\rho_k)-\eps$ as a lower bound on the optimization.
\State Otherwise, heuristically find
$\mu^* \defvar \arg\min_{\mu\in(0,1)}f_\mathrm{obj}(\rho_k + \mu\Delta^*)$.
\State Set $\rho_{k+1} = \rho_k + \mu^* \Delta^*$, $k\leftarrow k+1$ and return to Step~\ref{step:FWSDP}.
\end{algorithmic}
\end{algorithm}
\end{savenotes}
\noindent Geometrically, the SDP in Eq.~\eqref{eq:FWSDP} corresponds to considering the tangent to $f_\mathrm{obj}$ at $\rho_k$ and computing the maximum amount by which it can decrease (as compared to its value at $\rho_k$) over the domain $\mathcal{D}$.
As described earlier,
this bounds the maximum amount by which $f_\mathrm{obj}$ can decrease from $f_\mathrm{obj}(\rho_k)$ over $\mathcal{D}$.
In our application of the Frank-Wolfe algorithm, there is the technical issue that if $\mathcal{G}(\rho_k)$ is singular (or has negative eigenvalues, from numerical imprecision), then the derivative at $\rho_k$ (Eq.~\eqref{eq:fobjder}) is ill-behaved.
To cope with this, we used
the heuristic solution of simply replacing $\rho_k$ with $(1-\delta)\rho_k + \delta \idk$ in Step~\ref{step:FWSDP}, where $\delta
\defvar
\max(\operatorname{eigenvalues}(-\mathcal{G}(\rho_k)) \cup \{ 10^{-14} \})
$.
Note that this does not affect the \emph{security} of the result, since it merely corresponds to taking a tangent at a slightly different point, which still yields a valid lower bound. On the other hand, it could possibly affect the theoretical convergence rates, but in practice this did not appear to pose a significant problem in our setting. (In~\cite{WLC18}, this issue is addressed by analyzing a ``perturbed'' version of the optimization and applying a continuity bound, but we found that for the level of accuracy we desired in this work, the admissible perturbation values are too small to cope with the negative eigenvalues that occur.)
\subsection{Overall algorithm}
\label{sec:algorithmsummary}
Putting the above results together, we see that for any set of intervals $[\theta_j-\delta_j,\theta_j+\delta_j]$ such that $[0,\pi]\subseteq \bigcup_j[\theta_j-\delta_j,\theta_j+\delta_j]$, and set $V$ such that $S_{\rightsemicirc} \subseteq \operatorname{Conv}(V)$, we have
\begin{align}
\min_{\rho_{AB}} \min_{{\theta_B}} \min_{{\theta_A}}
F_\mathrm{obj}({\theta_A},{\theta_B},\rho_{AB})
&\geq \min_{\rho_{AB}} \min_{{\theta_B}} \min_j F_\mathrm{obj}(\theta_j,{\theta_B},\rho_{AB}) - \epsilon_\mathrm{con}(\delta_j) \nonumber \\
&= \min_j \min_{\rho_{AB}} \min_{(r_Z,r_X)\in S_{\rightsemicirc}} F_\mathrm{obj}(\theta_j,(r_Z,r_X),\rho_{AB}) - \epsilon_\mathrm{con}(\delta_j) \nonumber \\
&\geq \min_j \min_{\rho_{AB}} \min_{(r_Z,r_X)\in V} F_\mathrm{obj}(\theta_j,(r_Z,r_X),\rho_{AB}) - \epsilon_\mathrm{con}(\delta_j) \nonumber \\
&= \min_j \min_{(r_Z,r_X)\in V} \min_{\rho_{AB}} F_\mathrm{obj}(\theta_j,(r_Z,r_X),\rho_{AB}) - \epsilon_\mathrm{con}(\delta_j).
\label{eq:mincommute}
\end{align}
We can refine the intervals $[\theta_j-\delta_j,\theta_j+\delta_j]$ and set $V$ by the iterative processes described above, with the innermost minimization over $\rho_{AB}$ being handled by the Frank-Wolfe algorithm. It is clearly possible to swap the order of $\min_j$ and $\min_{(r_Z,r_X)\in V}$ in the last line; however, some heuristic plots of the objective function suggest that performing the optimizations in the order shown here is slightly faster.
To see that this expression can converge to a tight bound, observe that the first inequality in~\eqref{eq:mincommute} becomes arbitrarily tight as we choose smaller values of $\delta_j$. For the second inequality, note that
the described algorithm chooses $V$ in such a way that the minimum over $V$ approaches the minimum over $S_{\rightsemicirc}$, hence this inequality also becomes arbitrarily tight.
In general, the above approach faces the difficulty that it needs to optimize over the choice of Lagrange multipliers $\vec{\lambda}$. Given that our approach for solving~\eqref{eq:mincommute} for a specific choice of Lagrange multipliers is already highly computationally intensive (requiring about $5000$ core-hours to achieve the level of accuracy in the bounds~\eqref{eq:certbnds} below for each value of $\p$), it would be impractical to also optimize over the Lagrange multipliers while doing so. It is more feasible to first optimize the Lagrange multipliers while using a simple heuristic algorithm to estimate the minimizations, then certify the final result using our approach for solving~\eqref{eq:mincommute}. (This is essentially the same perspective as presented in~\cite{SBV+21}.)
We remark that this approach can also yield arbitrarily tight bounds for DIRE, where the goal would typically be~\cite{ARV19,BRC20} to find lower bounds on (weighted sums of) ``two-party entropies'' $H(A_x B_y|E)$. This is because by the same arguments as above, we have $H(A_x B_y|E) = D({\rho}_{AB} \Vert \mathcal{Z}_{xy}({\rho}_{AB}))$~\cite{Col12,arx_TSG+19},
where
\begin{align}
\mathcal{Z}_{xy}(\sigma_{AB}) &\defvar \sum_{ab} ({\pvm}_{a|x}\otimes{\pvm}_{b|y}) \sigma_{AB} ({\pvm}_{a|x}\otimes{\pvm}_{b|y}) .
\end{align}
(Here we omit the parts corresponding to noisy preprocessing, since it is not applied in randomness expansion.) This expression can be bounded in the same way as we have just described above, though the objective function would no longer be affine with respect to $(r_Z,r_X)$, and hence the optimization over Bob's measurements would also have to be approached using a continuity bound. Our approach should yield a substantial improvement over previous results, which are restricted to the CHSH inequality and only bound the entropy of one party's outputs~\cite{LLR+19}, or which consider the full distribution and bound the entropy of both outputs but use inequalities that are not tight~\cite{BRC20,arx_TSG+19,BFF21}. In addition, the fact that it allows for the random-key-measurement approach could yield further improvements, though there are some technicalities that we address at the end of Sec.~\ref{sec:preshared}.
It would be convenient for future analysis if it were possible to develop closed-form expressions for $\breve{\lin}_\p$ (as was done in~\cite{HST+20,WAP21,SBV+21}), rather than the computationally intensive numerical approaches shown above. However, this appears to be rather challenging, as noted in~\cite{WAP21}. In particular, we found numerical evidence against the conjecture proposed in~\cite{SGP+21} that the minimum in~\eqref{eq:mainopt} (when restricted to qubit strategies) can always be attained by states such that $\rho_{AB}$ is of rank $2$. (A similar observation was reported in~\cite{WAP21}.) More precisely, in the process of heuristically solving the optimizations in order to estimate suitable choices of $\lambda$, we discovered that if we imposed the additional restriction that $\rho_{AB}$ has rank $2$, there was a small but numerically significant difference as compared to the results without this restriction. (The states that heuristically approach the minimum in the latter indeed tend to have two very small eigenvalues, but it appears that these eigenvalues cannot be reduced exactly to zero.) This suggests that the aforementioned conjecture is not true after all, which poses a challenge for closed-form analysis because the eigenvalues involved in computing the entropy are genuinely roots of a fourth-degree polynomial (in~\cite{HST+20,SBV+21}, a key element of the analysis was to argue that it suffices to consider rank-$2$ $\rho_{AB}$, simplifying the expression for the eigenvalues).
\subsection{Resulting bounds}
\label{sec:1rndresults}
\begin{figure}
\centering
\subfloat[$\p=0$]{
\includegraphics[width=0.49\textwidth]{fig_entbnd1}
}
\subfloat[$\p=0.2$]{
\includegraphics[width=0.49\textwidth]{fig_entbnd2}
}
\caption{The solid lines are the certified lower bounds we obtained (Eq.~\eqref{eq:certbnds}), while the points indicate the results of heuristically solving the optimization~\eqref{eq:mainopt} over qubit states and measurements (with just the CHSH value as the constraint). As previously discussed, the tight bound in each case would be given by the convex envelope of the curve traced out by the points, assuming that the heuristics have found the true minimum. However, we can see that in each case that curve appears to be nonconvex over the interval $[2,2.75]$ (approximately), and its convex envelope would be affine over that interval --- specifically, it would be given by the linear interpolation between the feasible points at the ends of that interval. The certified bound is almost flush with this linear interpolation, indicating that it is basically tight over this interval.}
\label{fig:entbnds}
\end{figure}
In Fig.~\ref{fig:entbnds}, we show the affine bounds certified by our method when the only constraint imposed is the CHSH value (which is what is relevant for Protocol~\ref{prot:DIQKD}). More precisely, we consider the described optimization with a single constraint corresponding to the operator
\begin{align}
\Gamma({\theta_A},{\theta_B}) = A_0\otimes B_0 + A_0\otimes B_1 + A_1\otimes B_0 - A_1\otimes B_1,
\end{align}
with the constraint value $\constr$ being the CHSH value. (As previously mentioned, this can be implemented in the formulation where are 4 constraint operators $\Gamma_{x,y}({\theta_A},{\theta_B})$ by simply restricting to Lagrange-multiplier choices of the form $(\lambda_{00},\lambda_{01},\lambda_{10},\lambda_{11}) = (\lambda,\lambda,\lambda,-\lambda)$.) Each choice of the associated Lagrange multiplier $\lambda$ yields an affine lower bound of the form ${\lambda}{\constr} + c_{{\lambda}}$, as noted in Eq.~\eqref{eq:duallin}.
Importantly, some heuristic computations (also observed in~\cite{SGP+21} for the $\p=0$ case) suggest that the true bound $\breve{\lin}_\p({\constr})$ in this situation is in fact affine over a wide range of CHSH values --- we show this in Fig.~\ref{fig:entbnds}, which displays the results of heuristic minimizations compared to our certified bound in some cases. In particular, this range on which the bound is affine covers all currently experimentally reasonable values.
This has the implication that there is a single ${\lambda}$ that yields an affine lower bound which is equal to $\breve{\lin}_\p({\constr})$ (i.e.~it is tight) over this entire range; specifically, it is simply the value of ${\lambda}$ corresponding to the gradient of $\breve{\lin}_\p({\constr})$ in this range. This greatly simplifies our task since we only need to solve the optimization for this specific value of ${\lambda}$.
We focused on several values of noisy preprocessing, ranging from $\p=0$ to $\p=0.45$.
In each case, we first solved the minimization~\eqref{eq:intercept} heuristically for some selection of values of ${\lambda}$, in order to estimate the choice of $\lambda$ that yields a tight bound over the range of CHSH values in which $\breve{\lin}_\p({\constr})$ is affine. Then using our algorithm to get certified bounds on the corresponding $c_{\lambda}$ in~\eqref{eq:intercept}, we arrived at the final bounds
\begin{align}
\begin{gathered}
\breve{\lin}_{0}({\constr}) \geq 1.190(\constr-2) - 0.00454, \qquad
\breve{\lin}_{0.2}({\constr}) \geq 0.327(\constr-2) + 0.72063, \\
\breve{\lin}_{0.3}({\constr}) \geq 0.139(\constr-2) - 0.88051, \qquad
\breve{\lin}_{0.4}({\constr}) \geq 0.0341(\constr-2) + 0.97055, \\
\breve{\lin}_{0.45}({\constr}) \geq 0.00855(\constr-2) + 0.992487,
\end{gathered}
\label{eq:certbnds}
\end{align}
some of which are shown in Fig.~\ref{fig:entbnds} (to express the bounds as functions of winning probability $w$ instead, as we require for other figures, simply replace $\constr$ in each expression with $8w-4$). From the $\p=0.3$ bound, we obtain the depolarizing-noise threshold of $9.33\pct$ mentioned previously. However, we were unable to obtain better thresholds using the higher values of $\p$, for reasons we shall now discuss.
\subsection{Optimality of results}
\label{sec:bestbnd}
For each of the bounds in Eq.~\eqref{eq:certbnds}, there is a feasible point of the optimization for $c_{\lambda}$ which yields a value within $0.005$ (or less, for higher values of $\p$) of the certified results shown in Eq.~\eqref{eq:certbnds}, so these bounds on entropy are very close to optimal in terms of absolute error.
In terms of the depolarizing-noise thresholds that they yield, taking the convex envelope of some of the feasible points shown in Fig.~\ref{fig:entbnds}
yields the result that the thresholds for $\p=0.2$ and $\p=0.3$ cannot be improved by more than about $0.1$ percentage points, so those thresholds are very close to optimal as well.
However, larger values of $\p$ face the issue that the asymptotic keyrates become extremely low, which makes the horizontal intercepts (i.e.~the depolarizing-noise threshold) very sensitive to changes in $\breve{\lin}_{\p}$ --- even a small absolute error in this bound results in a significant change in the threshold value. Therefore, the thresholds we obtained from the certified bounds with $\p=0.4$ and $\p=0.45$ in Eq.~\eqref{eq:certbnds} were only $9.32\pct$ and $9.10\pct$ respectively, worse than the results for $\p=0.3$. Heuristic computations suggest that the true thresholds for those cases might be approximately $9.46\pct$ and $9.50\pct$ respectively, but using our algorithm to certify these values would require it to converge to tolerances that currently appear impractical. Hence a different approach may be needed to find the true thresholds for these values of $\p$. From a practical perspective though, such improvements may be of limited use, because the very low asymptotic rates mean that the finite-size keyrate would likely be zero until extremely large sample sizes.
In any case, we note that for depolarizing noise at least, the threshold value cannot be improved much further by \emph{any} protocol choices within the framework we have presented in this section, e.g.~by using the full distribution as constraints (which also encompasses the use of modified CHSH inequalities~\cite{WAP21,SBV+21}), or adjusting the values of $\keyw_x$. This is essentially because our bounds are very close to the linear interpolation between the points $(2,\binh(\p))$ and $(2\sqrt{2},1)$ (as can be seen from Fig.~\ref{fig:entbnds}). Intuitively speaking, the bound on the entropy against Eve in the depolarizing-noise scenario cannot exceed this linear interpolation (because Eve can always perform classical mixtures of strategies in order to attain every point on this linear interpolation), which means that our bounds are close to the highest bounds that are even possible in principle.
Making this reasoning quantitative, we can obtain an explicit upper bound on the depolarizing-noise threshold. (We highlight that because we will do so by constructing an extremely generic attack, this upper bound holds for \emph{all} protocols of this form, regardless of choices of parameters such as the input distributions and noisy preprocessing. To some extent, it is surprising that there even exist any protocols that achieve thresholds close to the upper bounds implied by such a generic attack.) We first recall that (as defined in Sec.~\ref{sec:hon}) the measurement statistics in the depolarizing-noise model are given by real projective measurements on the Werner state $\varrho_\q \defvar (1-2\q)\pure{\Phi^+} + 2\q\, \id/4$. It is known~\cite{AGT06,Kri79} that these measurement statistics can be reproduced by a local-hidden-variable (LHV) model as soon as the state does not violate the CHSH inequality, i.e.~for $q\geq q_2 \defvar{(2-\sqrt{2})}/{4} \approx 14.6\pct$.\footnote{In fact, this is also true in a more general context where the honest parties have \emph{any} number of real projective measurements on the Werner state, via the value of the second Grothendieck constant $K_G(2)=\sqrt{2}$~\cite{AGT06,Kri79}. For general (i.e.~not necessarily real) projective measurements the analogous value is known to satisfy $q_3 \lesssim 15.9\pct$, as follows from the best known bound on the third Grothendieck constant $K_G(3)$~\cite{HQV+17}.}
In the device-independent setting, the existence of an LHV model yields an attack for Eve that gives her full knowledge of the measurement outcomes.
This means that for any depolarizing-noise value $\q$, we can construct the following attack for Eve: first, she generates a classical ancilla bit which is equal to $0$ with probability $\pBell \defvar (\q_2-\q)/\q_2$. If the bit is equal to $0$, Alice and Bob's devices simply perform the honest measurements on the maximally entangled state $\Phi^+$. Otherwise, the devices implement an LHV model that yields the same statistics as the honest measurements performed on the Werner state $\varrho_{\q_2}$, but which gives Eve full knowledge of the outcomes. This attack indeed reproduces the statistics corresponding to depolarizing noise~$\q$, since $\pBell\pure{\Phi^+} + (1-\pBell) \varrho_{\q_2} = \varrho_{\q}$ and the second strategy produces the same statistics as the honest measurements performed on $\varrho_{\q_2}$.
In the two cases, the entropies of Alice's outputs after noisy preprocessing are $H(\hat{A}_x|\tilde{E})_\mathrm{triv}=1$ (i.e.~Eve's side-information is trivial since the Alice-Bob state is pure) and $H(\hat{A}_x|\tilde{E})_\mathrm{LHV}=\binh(\p)$ (i.e.~the uncertainty arises purely from the noisy preprocessing) respectively, where $\tilde{E}$ denotes Eve's side-information excluding the ancilla bit. Incorporating the ancilla bit into Eve's side-information $E$, this attack achieves
\begin{align}
H(\hat{A}_x|E) = \pBell H(\hat{A}_x|\tilde{E})_\mathrm{triv} + (1-\pBell) H(\hat{A}_x|\tilde{E})_\mathrm{LHV}
= \frac{\q_2-\q}{\q_2} + \frac{\q}{\q_2} \binh(p).
\end{align}
This expression hence serves as an upper bound on the best possible lower bound we could derive on the conditional entropy against Eve. Also, the conditional entropy against Bob's output (for Bob's optimal key-generation measurements on Werner states) is simply
\begin{align}
H(\hat{A}_x|B_x) = \binh(\p +(1-2\p) \q).
\label{eq:entBob}
\end{align}
Recalling that the asymptotic keyrate is given by the Devetak-Winter bound~\cite{DW05} (up to the sifting-related weights $\keyw_x$~\cite{SGP+21}), this yields a simple upper bound on the critical value of $\q$ that still allows a positive asymptotic keyrate (for protocols of this form, i.e.~applying noisy preprocessing, random key measurements, and considering the full output distribution, but only using one-way error correction):
\begin{align}
\q_\textrm{att}(\p) \defvar& \max \bigg\{ \q \, \bigg| \, \sum_x \keyw_x (H(\hat{A}_x|E) - H(\hat{A}_x|B_x)) \geq 0 \bigg\} \nonumber\\
=& \max \left\{ \q \, \middle| \, \frac{\q_2-\q}{\q_2} + \frac{\q}{\q_2} \binh(\p) - \binh(\p +(1-2\p) \q) \geq 0 \right\},
\end{align}
where in the first line the entropies refer to those of the attack we have described.
We observe numerically that this upper bound is increasing with respect to $\p$, so we have $\q_\textrm{att}(\p)\leq \q_\textrm{att}(\p\to1/2)$. In the limit $\p\to 1/2$ we can write $\p=1/2-\delta$ and expand the expression in small $\delta$ to find
\begin{equation}
\q_\textrm{att}\!\left(\p\to\frac{1}{2}\right) = \frac{1+ 4 \q_2 -\sqrt{8 \q_2+1}}{8 \q_2} = 1 - \frac{\sqrt{7+4\sqrt{2}} - 1}{2\sqrt{2}} \approx 9.57\pct.
\end{equation}
Hence for protocols of the form described in this work, it is not possible for the depolarizing-noise threshold to exceed this value. For the parameter choice $\p=0.3$, we have $\q_\textrm{att}(0.3) \approx 9.51 \pct$, which is close to the threshold of $9.33 \pct$ for which we could certify a positive keyrate.\footnote{These two thresholds cannot match exactly, because the feasible points shown in Fig.~\ref{fig:entbnds} also imply that the tight bound $\breve{\lin}_{\p}({\constr})$ is not \emph{exactly} equal to the linear interpolation between $(2,\binh(\p))$ and $(2\sqrt{2},1)$, which essentially corresponds to the attack we describe here.}
Finally, it is worth noting that while the main focus of this work is depolarizing noise applied to the statistics corresponding to the ideal CHSH measurements on $\Phi^+$, the above analysis in fact generalizes to a substantially larger family of scenarios. Specifically, the same analysis applies for depolarizing noise applied to the statistics from any number of real projective measurements on $\Phi^+$, since the results of~\cite{AGT06,Kri79} yield the required LHV models.\footnote{A more general consideration would be the scenario of depolarizing noise applied to the statistics from an arbitrary two-qubit state, but it is not immediately obvious whether the thresholds in~\cite{AGT06,Kri79,HQV+17} apply to such states as well, so we leave this for future work.} (If Bob performs suboptimal generation measurements, then Eq.~\eqref{eq:entBob} still holds as a lower bound, so the argument carries through.) In addition, replacing $\q_2$ with $\q_3$ straightforwardly provides a threshold of $\q \lesssim 10.01 \pct$ for all possible projective measurements on that state, via the respective known bound in~\cite{HQV+17}. (Note that in all such cases where there are more than 2 possible measurements, there is no ``qubit reduction'' to make it easier to derive corresponding lower bounds, and hence the lower bounds have also not been very thoroughly explored.) An interesting further consideration is the case of general POVM measurements. Here, a rather loose bound $q_\textrm{POVM} \lesssim 27.3\pct$ is also known~\cite{HQV+17} for the threshold that allows an LHV description. However, pure POVMs on qubits can involve up to four outcomes, opening up the possibility of more general noisy preprocessing --- all doubly stochastic maps on probability vectors with four outcomes. We thus leave it for future work.
The above argument relies on the fact that depolarizing-noise statistics can be obtained by simple mixtures of ``extremal'' strategies. This is not necessarily the case for other noise models, such as limited detection efficiency in photonic experiments. Therefore, the same argument cannot be directly applied to obtain upper bounds on the thresholds for such forms of noise.
\section{Possible modifications}
\label{sec:mods}
\subsection{Coordinating input choices by public communication}
\label{sec:pubchann}
The random-key-measurement protocol has the drawback that the keyrate is effectively halved, since the generation rounds have ``mismatched'' inputs approximately half the time. It would be helpful to find a way to work around this issue. One possible approach could be to observe that in~\cite{ARV19}, it was assumed that the following operations can be performed in each round: the devices receive some shares of a quantum state, then\footnote{The assumption being made here is that since the quantum states are now entirely in the possession of Alice and Bob, the test/generation decision can no longer affect the distributed state --- if the distributed state could depend on whether it is a test or generation round, the protocol would be trivially insecure.} Alice and Bob publicly communicate to come to an agreement on \emph{both} of their input choices, and finally they supply these inputs to their devices. (This was necessary in~\cite{ARV19} because
Alice and Bob's actions in that DIQKD protocol require both of them to know whether it is a test or generation round. In fact, our analysis can be viewed as the first EAT-based security proof for a ``genuinely sifting-based'' DIQKD protocol, in the sense that Alice and Bob do not coordinate which rounds are test rounds, and simply choose their inputs independently.)
If we assume that this is also possible in our scenario, then Alice and Bob could coordinate their inputs in the generation rounds instead of choosing them independently, thereby avoiding the sifting factor.
Unfortunately, it does not seem clear if such a proposal is entirely plausible in near-term experimental implementations. This is because it relies on the devices being able to store the quantum state for long enough for Alice and Bob to agree on their choice of inputs,
which is potentially challenging for current Bell-test implementations. As an alternative, we propose the following potential modification to the DIQKD protocol in~\cite{ARV19} --- instead of agreeing on the test rounds via public communication, Alice and Bob could use a small amount of pre-shared key to choose which rounds are test rounds, in the same way as in DIRE (for details on the amount of pre-shared key required, see the DIRE protocols in~\cite{ARV19,BRC20} or the discussion in Sec.~\ref{sec:preshared} below). This approach would essentially be a ``key {expansion}'' protocol that requires a small amount of pre-shared key to initialize. We remark that this is not a dramatic change in perspective, because a common method to authenticate channels (namely, message authentication codes) relies on having a small amount of pre-shared key, so the assumed existence of authenticated channels in the DIQKD protocol is likely to require some pre-shared key in any case.
However, this basic notion cannot immediately be generalized to Protocol~\ref{prot:DIQKD} here, since requiring Alice and Bob to choose uniformly distributed ``matching'' inputs in the generation rounds would require a large amount of pre-shared key (roughly $(1-\gamma)n$ bits). Fortunately, in the following section, we propose a variation which overcomes this difficulty by ``recovering'' the entropy in the pre-shared key, thereby still achieving net key expansion.
\subsection{Protocol using pre-shared key}
\label{sec:preshared}
Here we describe a variant protocol that avoids the sifting factor \emph{without} requiring the brief quantum storage described above, through the use of a fairly long pre-shared key. The limitation of this variant is that the net increase in secret key is just a (constant) fraction of the amount of pre-shared key; however, the \emph{rate} of net key generation does not have the sifting factor of $1/2$.
Informally, the idea is to simply use the pre-shared key as Alice's input string $\str{X}$, which allows Bob to choose his generation inputs to match Alice's. Just as importantly, this also allows them to (almost) entirely omit the public announcement of their inputs --- hence $\str{X}$ remains private, and with some care it can be incorporated into the final key without losing the entropy it ``contains''.
We now describe this idea in detail as Protocol~\ref{prot:preshared} below, followed by its security proof. The protocol supposes that Alice and Bob hold a pre-shared (uniform) key of $n$ bits, which we shall simply denote as $\str{X}$, since it will be exactly the string that Alice uses as her inputs. The appropriate value of $\ell_\mathrm{key}$ to choose will be described later in Theorem~\ref{th:preshared}.
\begin{savenotes}
\begin{breakablealgorithm}
\caption{}
\label{prot:preshared}
This protocol proceeds the same way as Protocol~\ref{prot:DIQKD}, except for the following changes:
\begin{itemize}
\item In each round, Alice's input $X_j$ is determined from the pre-shared key $\str{X}$, instead of being generated randomly in that round. Bob's input $Y_j$ is generated as follows: with probability $\gamma$ he chooses a uniformly random $Y_j \in \{2,3\}$, otherwise Bob chooses $Y_j = X_j$. In addition, he generates another register $Y'_j$ which equals $Y_j$ when $Y_j \in \{2,3\}$ and equals $\perp$ otherwise.
\item Alice and Bob do not publicly announce the strings $\str{X}\str{Y}$. Instead, Bob only announces the string ${\str{Y}'}$.\footnote{It might be possible to consider a slight variant which omits this step. However, knowing ${\str{Y}'}$ allows Alice to compute $\str{Y}$, which may be relevant for error correction since it allows Alice to distinguish the test and generation rounds. In any case, it seems unclear whether the entropy of ${\str{Y}'}$ can be usefully extracted even if it is kept secret, since Alice does not have access to it in that case.} Additionally, the sifting step is unnecessary, since there will be no rounds such that $Y_j \in \{0,1\}$ and $X_j \neq Y_j$.
\item Privacy amplification is performed on the strings $\str{A}\str{X}$ and $\tilde{\str{A}}\str{X}$ instead of $\str{A}$ and $\tilde{\str{A}}$.
\end{itemize}
\end{breakablealgorithm}
\end{savenotes}
To prove the security of this protocol, we can simply follow almost exactly the same security proof as for Protocol~\ref{prot:DIQKD}, with some changes we shall now describe. Firstly, the value of $h_\mathrm{hon}$ (to be used when computing $\mathrm{EC}_\mathrm{max}$) is replaced by
\begin{align}
\tilde{h}_\mathrm{hon}
= H(A_j|B_jX_jY_j)_\mathrm{hon}
&= \frac{1-\gamma}{2}
\sum_{z\in\{0,1\}} H(A_j | B_j;X_j = Y_j = z)_\mathrm{hon}
\nonumber \\& \qquad
+ \frac{\gamma}{4}
\sum_{x\in\{0,1\},y\in\{2,3\}}
H(A_j| B_j ;X_j = x, Y_j = y)_\mathrm{hon}, \label{eq:ECpreshared}
\end{align}
since the probabilities of $X_j = Y_j = z$ for $z\in\{0,1\}$ are now ${(1-\gamma)}/{2}$. (Note that no error-correction information needs to be sent from Alice to Bob regarding $\str{X}$, since both of them have a copy of that string.)
Also, since the strings used in the privacy-amplification step are now $\str{A}\str{X}$ and $\tilde{\str{A}}\str{X}$, this means that we need an equivalent of Eq.~\eqref{eq:EATbound}, with $H_\mathrm{min}^{\eps_s'}(\str{A}{\str{B}'}\str{X}|{\str{Y}'} E)$ in place of $H_\mathrm{min}^{\eps_s'}(\str{A}{\str{B}'} | \str{X}\str{Y}E)$. To obtain this, we note that we can simply construct a virtual protocol in the analogous way to Protocol~\ref{prot:virtual}, then consider the same EAT channels $\mathcal{M}_j$ as before, but instead we shall identify $C'_j$ with $D_j$, $A_jB'_jX_j$ with $S_j$, and $Y'_j$ with $T_j$ in Definition~\ref{def:EATchann}.
The Markov conditions are again fulfilled, since $Y'_j$ is generated by trusted randomness in each round and independent of all previous data. To find an appropriate min-tradeoff function for these channels, we note that the output $(\mathcal{M}_j\otimes\idmap_{R})(\omega_{R_{j-1}R})$ of channel $\mathcal{M}_j$ always satisfies
\begin{align}
H(X_j | Y'_j R) = H(X_j) = 1,
\end{align}
because $X_j$ is produced by trusted randomness independent of $Y'_j R$. Therefore, we can use the chain rule to write
\begin{align}
H(A_j B'_j X_j | Y'_j R)
&= H(X_j | Y'_j R) + H(A_j B'_j | X_j Y'_j \Fj R) \nonumber \\
&\geq 1 + (1-\gamma) \lin_\p(w) + \gamma \lin_0(w) \nonumber\\
&\defvar 1 + \tilde{\g}(w),
\label{eq:chaintrick}
\end{align}
where the function $\tilde{\g}$ (in contrast to $\g$) does {not} have the factor of $1/2$ introduced by sifting, since Alice does not ``erase'' the outputs of any rounds. We can thus construct a new min-tradeoff function $\tilde{f}_\mathrm{min}$ in the same way as in Sec.~\ref{sec:fmin}, but using $1 + \tilde{\g}(w)$ in place of $\g(w)$.\footnote{We remark that if we think of this replacement as happening in two steps, first replacing $\g$ by $\tilde{\g}$ and then adding a ``constant offset'' of $1$, then the latter has no effect on $\operatorname{Var}_{\mathcal{Q}_{f}}(f_\mathrm{min})$ or the difference $\operatorname{Max}(f_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}_{f}} (f_\mathrm{min})$, and hence does not change the finite-size correction to the keyrate except indirectly via changing the system dimensions and the range of $\beta$. However, the first step of replacing $\g$ by $\tilde{\g}$ does slightly increase the finite-size correction (since $\tilde{\g}$ has a somewhat larger range).} The rest of the proof then proceeds as before, leading to the following security statement:
\begin{theorem}\label{th:preshared}
Protocol~\ref{prot:preshared}
has the same security guarantees as those described in Theorem~\ref{th:DIQKD},
except with the following changes:
\begin{itemize}
\item $\beta$ is chosen to be in
$[1+\tilde{\g}(0),1+\tilde{\g}(1)]$
instead.
\item $h_\mathrm{hon}$ is replaced by $\tilde{h}_\mathrm{hon}$ as specified in Eq.~\eqref{eq:ECpreshared}.
\item In Eq.~\eqref{eq:keylength} for $\ell_\mathrm{key}$, $\g(w_\mathrm{exp}-\delta_\mathrm{tol})$ is replaced by $1 + \tilde{\g}(w_\mathrm{exp}-\delta_\mathrm{tol})$, and the values of $V$ and $K_\alpha$ are replaced by
\begin{align}
\begin{aligned}
\tilde{V} &\defvar \sqrt{\operatorname{Var}_{\mathcal{Q}_{f}}(\tilde{f}_\mathrm{min})+2} + \log
289,
\\
\tilde{K}_\alpha &\defvar \frac{2^{(\alpha-1)(2\log12 + \operatorname{Max}(\tilde{f}_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}_{f}} (\tilde{f}_\mathrm{min}))} }{6(2-\alpha)^3\ln2}
\ln^3\left(2^{2\log12 + \operatorname{Max}(\tilde{f}_\mathrm{min})-\operatorname{Min}_{\mathcal{Q}_{f}} (\tilde{f}_\mathrm{min})} + e^2\right),
\end{aligned}
\end{align}
where $\tilde{f}_\mathrm{min}$ is a function that satisfies
\end{itemize}
\vspace*{-.5cm}
\begin{align}
\begin{gathered}
\operatorname{Max}(\tilde{f}_\mathrm{min})
= 1 + \frac{1}{\gamma}\tilde{\g}(1) + \left(1-\frac{1}{\gamma}\right)\beta, \qquad
\operatorname{Min}_{\mathcal{Q}_{f}} (\tilde{f}_\mathrm{min})
= 1 + \tilde{\g}\!\left(\frac{2-\sqrt{2}}{4}\right), \\
\operatorname{Var}_{\mathcal{Q}_{f}}(\tilde{f}_\mathrm{min}) \leq
\frac{2-\sqrt{2}}{4\gamma}
\min \left\{\Delta_0^2, \Delta_1^2 \right\}
+ \frac{2+\sqrt{2}}{4\gamma}
\max \left\{\Delta_0^2, \Delta_1^2 \right\}
, \text{ where } \Delta_w \defvar \beta - 1 - \tilde{\g}(w).
\end{gathered}
\end{align}
\end{theorem}
Overall, recalling that Protocol~\ref{prot:DIQKD} required $n$ bits of pre-shared key, we see that the \emph{net} gain of secret key bits in Protocol~\ref{prot:preshared}
is larger than that of Protocol~\ref{prot:DIQKD} by a factor of approximately (ignoring the changes to the finite-size corrections)
\begin{align}
\frac{n \left(\tilde{\g}(w_\mathrm{exp}-\delta_\mathrm{tol}) - \tilde{h}_\mathrm{hon}\right)}{n \Big(\g(w_\mathrm{exp}-\delta_\mathrm{tol}) - h_\mathrm{hon}\Big)} \approx 2
,
\end{align}
since it avoids the sifting factor.
Informally, by keeping $\str{X}$ secret and incorporating it in the privacy amplification step, we have ``recovered'' the entropy that was present in the original pre-shared key.\footnote{Note that the underlying idea here critically relies on the fact that the bound $\lin_p$ is for the entropy of the output strings \emph{conditioned on $X_j$}. This allowed us to use the chain rule to obtain Eq.~\eqref{eq:chain}, which eventually led to the result that the seed entropy contained in $\str{X}$ simply ``adds on'' to the entropy of the output strings in the original Protocol~\ref{prot:DIQKD}. If $\lin_p$ had been a bound on, for instance,
$H(A_j | Y'_j R)$ instead of $H(A_j| X_j Y'_j R)$,
this argument would not have worked.}
In practice, including the string $\str{X}$ in privacy amplification
essentially doubles the input size for the hash function in that step, which raises its computational difficulty substantially (though not insurmountably).
One might wonder whether it would be possible to bypass this aspect --- for instance, by simply performing privacy amplification on $\str{A}$ and $\tilde{\str{A}}$ as before, then appending $\str{X}$ to the output. At first glance, this approach might appear plausible, since $\str{X}$ is not announced in Protocol~\ref{prot:preshared}. Unfortunately, it seems unclear how to certify that the publicly communicated error-correction string $\str{L}$ is independent of $\str{X}$ (in fact, it seems unlikely that this is true). Hence the idea of simply appending $\str{X}$ may not be secure. By instead incorporating it in privacy amplification in the specified manner, Protocol~\ref{prot:preshared} ensures that the entropy of $\str{X}$ is securely ``extracted'' into the final key.
As previously mentioned, the net increase in secret key given by one instance of Protocol~\ref{prot:preshared} is limited to a fraction of the amount of pre-shared key.
However, it is possible in principle to recursively run Protocol~\ref{prot:preshared} in order to achieve unbounded key expansion --- one can use the key generated by one instance of Protocol~\ref{prot:preshared} to run it again with a longer pre-shared key and larger $n$ (since the security definition is composable, the soundness parameter
will only increase
additively in this process~\cite{PR14}). We stress that in doing so, one must always incorporate the seed into the privacy-amplification step exactly as specified in Protocol~\ref{prot:preshared} --- in particular, this means that the \emph{entire} key changes with every iteration, instead of simply having some new bits appended.
Some care is necessary regarding device memory across instances of this recursive procedure --- while it does not seem to be directly vulnerable to the memory attack of~\cite{BCK13}\footnote{This is because the only public communication in Protocol~\ref{prot:preshared} that can leak any information is the error-correction string
(all other public communication is based on trusted randomness). In our security proof, we have bounded the min-entropy leakage at this step simply via the length of this string, without any assumptions about its structure, and hence we can still obtain a secure bound on the min-entropy of the input for privacy amplification in the final protocol instance. Note that this claim is strictly restricted to device reuse following the recursive process specified here --- once any key bits have been used for any other purpose, the attack again becomes a potential concern if the devices are reused.}, it is still important to ensure that the states measured in each instance of the protocol are independent of the key generated in the preceding instance, since this key is used to choose the device inputs (which must be independent of the state in order for our security arguments to hold). Again, this relies on the notion that the registers measured by the devices do not contain information about the key generated in the preceding instance.
There is another potential variant of this idea where a pre-shared key is instead used to generate \emph{both} input strings $\str{X}$ and $\str{Y}$, and the input-choice announcement is omitted entirely, with privacy amplification being performed on $\str{A}\str{X}\str{Y}$ and $\tilde{\str{A}}\str{X}\str{Y}$. This can be done by using $\kappa \binh(\gamma) n$ bits to choose the test rounds approximately according to the desired IID distribution of test rounds, then using $\kappa' \gamma n$ bits to set the value of $Y_j$ in the test rounds, where $\kappa,\kappa'>1$ are constants that can be chosen such that the approximations to the desired distributions are sufficiently accurate (see the randomness-expansion protocols in~\cite{ARV19,BRC20} for a more complete description of this process based on the \emph{interval algorithm}), and $n$ bits to set the value of $X_j$ in all rounds.\footnote{We break up the use of the seed into separate processes because it allows for better efficiency as compared to directly approximating the desired distribution of $\str{X}\str{Y}$ --- with the approach we describe, the ``inefficiency'' prefactors $\kappa,\kappa'$ of the interval algorithm only appear on the $\binh(\gamma)n,\gamma n$ terms instead of the full entropy of $\str{X}\str{Y}$.} This would hence require $(1 + \kappa \binh(\gamma) + \kappa' \gamma)n$ bits of seed randomness. A similar argument as above could then be performed by noting that (for $X_jY_j$ generated according to the ideal distribution) we have $H(X_j Y_j)
= 1 + \binh(\gamma) + \gamma
$, so most of the seed entropy can be ``recovered'', up to the losses from the $\kappa,\kappa'$ factors. However, tracking the effects of using the interval algorithm to approximate the ideal distribution is cumbersome (albeit possible), and it is unclear if this variant offers any immediate advantage over Protocol~\ref{prot:preshared} for DIQKD --- though it may be useful for protocols that use other Bell inequalities or non-uniform input distributions, as mentioned in Sec.~\ref{sec:nonCHSH}.
On the other hand, it appears that this variant may have potential for the purposes of DIRE instead. The main reason why the random-key-measurement approach in~\cite{SGP+21} could not be easily generalized to DIRE is that in order for Alice to select a uniformly random input in every round, she requires a (local) source of $n$ random bits, which is a free resource in DIQKD but not in DIRE --- if a proposed DIRE protocol consumes more random bits than it produces, then it has failed to achieve randomness \emph{expansion}\footnote{There is, however, the related but distinct task of device-independent randomness \emph{generation}~\cite{ZSB+20} (where the goal is to produce private randomness from an unbounded supply of public randomness that is independent from the devices), in which this would not be an issue.}. However, the protocol proposed in this section has the property that it ``recovers'' the entropy contained in the seed, which means that one can afford to use much larger seeds while still obtaining a net increase in secret key. Explicitly, the application of this idea to DIRE would hence be as follows: one begins with $2n$ random bits, which are then used as the input strings\footnote{For DIRE based on the CHSH inequality, Bob only requires two possible measurements instead of the four required for the DIQKD protocol here.} $\str{X}\str{Y}$ to the devices over $n$ rounds to obtain outputs $\str{A}\str{B}$. Modelling this process using EAT channels in a manner similar to above (see e.g.~\cite{LLR+19} for details), for each round we would have
\begin{align}
H(A_j B_j X_j Y_j| R)
&= H(X_j Y_j | R) + H(A_j B_j | X_j Y_j R) \nonumber \\
&= 2 + H(A_j B_j | X_j Y_j R),
\end{align}
which (given a bound on $H(A_j B_j | X_j Y_j R)$) allows one to bound the smoothed min-entropy of $\str{A}\str{B}\str{X}\str{Y}$ conditioned on $E$. By performing privacy amplification on $\str{A}\str{B}\str{X}\str{Y}$, one ``recovers'' all the entropy in the seed, due to the $H(X_j Y_j | R)$ term in the above equation. Overall, this proposed protocol allows one to use the improved entropy rate provided by the random-key-measurement approach~\cite{SGP+21}, in the context of DIRE instead of DIQKD.
\subsection{Collective attacks}
As a reference to compare our results against, we could consider whether a longer secure key could be obtained under the \emph{collective-attacks} assumption, which is the assumption that the device behaviour is IID in each round of the protocol (though Eve can still store quantum information for arbitrary periods).\footnote{To be precise, we mean that the part of the state held by Alice and Bob's devices is IID across the rounds, and in each round the devices have the same set of possible measurements. Since all purifications are isometrically equivalent, without loss of generality we can suppose that Eve also holds an IID purification of the Alice-Bob state.} To this end, we derive the following theorem, with the proof given in Appendix~\ref{app:collective}:
\begin{theorem}\label{th:collective}
Take any
$\eps^\mathrm{com}_\mathrm{EC},\eps^\mathrm{com}_\mathrm{PE},\eps_\mathrm{PA},\eps_\mathrm{h},\eps_s \in (0,1]$, $\gamma\in(0,1)$, $\p\in[0,1/2]$, and $\delta_\mathrm{IID}\in[0,w_\mathrm{exp} - \delta_\mathrm{tol})$.
Define
\begin{align}
\epsilon_\mathrm{IID} \defvar
\cdfBin{n}{1-(w_\mathrm{exp} - \delta_\mathrm{tol} - \delta_\mathrm{IID})\gamma}{\floor{(1-(w_\mathrm{exp}-\delta_\mathrm{tol})\gamma)n}}.
\end{align}
Under the collective-attacks assumption, Protocol~\ref{prot:DIQKD} is $(\eps^\mathrm{com}_\mathrm{EC} + \eps^\mathrm{com}_\mathrm{PE})$-complete and $(\max\{\epsilon_\mathrm{IID}, \eps_\mathrm{PA} + 2\eps_s\} + 2\eps_\mathrm{h})$-sound when performed with any choice of $\mathrm{EC}_\mathrm{max}$ and $\delta_\mathrm{tol}$ such that Eq.~\eqref{eq:comEC} and Eq.~\eqref{eq:ecPE} hold, and $\ell_\mathrm{key}$ satisfying
\begin{align}
\ell_\mathrm{key} \leq
n\g(w_\mathrm{exp}-\delta_\mathrm{tol}-\delta_\mathrm{IID}) - \sqrt{n} \,(2\log5)\sqrt{\log\frac{2}{\eps_s^2}}
- \mathrm{EC}_\mathrm{max} - \ceil{\log\left(\frac{1}{\eps_\mathrm{h}}\right)} - 2\log\frac{1}{\eps_\mathrm{PA}} + 2.
\label{eq:lkeycoll}
\end{align}
\end{theorem}
However, the above theorem is simply a statement for Protocol~\ref{prot:DIQKD} under the assumption of collective attacks, and that protocol does not fully exploit some implications of that assumption. For instance, in Theorem~\ref{th:collective} there is implicitly an $O(\gamma)$ subtractive penalty to the keyrates (which was also present in Theorem~\ref{th:DIQKD}\footnote{In fact, Theorem~\ref{th:DIQKD} has another $O(\gamma)$ subtractive penalty from the use of the bound~\eqref{eq:chain}, but this was due to the technical limitations of the EAT and the fact that the bounds~\eqref{eq:fmax} and \eqref{eq:gproof} are slightly suboptimal.}) caused by having to include the test-round data in the $\mathrm{EC}_\mathrm{max}$ term. Yet under the collective-attacks assumption, the test rounds are completely independent of the generation rounds, which implies that the effect of $\gamma$ should instead be to reduce the keyrate by a \emph{multiplicative} factor of $(1-\gamma)$. Importantly, in the latter case it is possible to choose arbitrarily large test probabilities $\gamma$ without necessarily making the keyrates negative, which can dramatically improve the statistical bounds for parameter estimation. To formalize this idea, we consider Protocol~\ref{prot:optcoll} below, which attempts to minimize the finite-size correction as much as possible using the most optimistic assumptions that have been discussed thus far.
\begin{savenotes}
\begin{breakablealgorithm}
\caption{}
\label{prot:optcoll}
This protocol proceeds the same way as Protocol~\ref{prot:DIQKD}, except for the following changes:
\begin{itemize}
\item Instead of independently choosing whether each round is a test or generation round, Alice chooses a uniformly random subset of size $m$ as test rounds before the protocol begins,
and we define $\gamma$ as the value $m/n$. Alice also prepares the strings $\str{X}\str{Y}$ in advance, by choosing $X_j=Y_j\in\{0,1\}$ uniformly at random in the generation rounds, and choosing $X_j\in\{0,1\}, Y_j\in\{2,3\}$ uniformly at random in the test rounds.
\item In each round, Alice and Bob briefly store their received quantum states instead of immediately measuring them. Alice then publicly announces $X_jY_j$, which Alice and Bob then use as the inputs to their devices.\footnote{Here, in our attempt to minimize the finite-size effects, we are following the~\cite{ARV19} assumption mentioned previously: Alice and Bob can briefly store their received quantum states, in a manner such that the public communication cannot affect the stored states.}
\item In the error-correction step, Alice does not send error-correction data (and a corresponding hash) for the full string $\str{A}$, but rather only the subset of it consisting of the generation rounds, denoted as $\str{A}_g$. Bob's guess for this string will be denoted as $\tilde{\str{A}}_g$. The values of $\str{A}$ in the test rounds, denoted as $\str{A}_t$, are sent directly to Bob without compression or encryption, and Bob uses this string for parameter estimation.
\item Bob's accept condition is instead to check that $\operatorname{hash}(\str{A}_g) = \operatorname{hash}(\tilde{\str{A}}_g)$ and $\operatorname{freq}_{\str{c}_t}(1)\geq w_\mathrm{exp}-\delta_\mathrm{tol}$
hold, where $\str{C}_t$ denotes the substring of $\str{C}$ corresponding to the test rounds (in particular, this means the frequencies are computed with respect to a string of length $\gamma n$, not $n$).
\item Privacy amplification is performed only on the strings $\str{A}_g$ and $\tilde{\str{A}}_g$.
\end{itemize}
\end{breakablealgorithm}
\end{savenotes}
For this protocol, the value of $\mathrm{EC}_\mathrm{max}$ is to be computed based only on the number of generation rounds,
since error correction is performed on the string $\str{A}_g$ rather than $\str{A}$. Focusing on the best possible theoretical bounds from Sec.~\ref{sec:EC}, this means we take $\mathrm{EC}_\mathrm{max}$ to be given by Eq.~\eqref{eq:optEC} with
\begin{align}
H_\mathrm{max}^{\ez}(\str{A}|\str{B}\str{X}\str{Y})_\mathrm{hon} \leq (1-\gamma)n
h_\mathrm{hon}
+ \sqrt{(1-\gamma)n} \, (2\log5)\sqrt{\log\frac{2}{\ez^2}},
\label{eq:optcollAEP}
\end{align}
where
\begin{align}
h_\mathrm{hon}
=
\sum_{z\in\{0,1\}} \frac{1}{2} H(A_j | B_j;X_j = Y_j = z)_\mathrm{hon} ,
\end{align}
since the test rounds are excluded. With this value of $\mathrm{EC}_\mathrm{max}$ in mind, we can state the security guarantees of this protocol, with the proof given in Appendix~\ref{app:optcoll}
(note that the dependence on
several security parameters
here is somewhat different as compared to the previous theorems):
\begin{theorem}\label{th:optcoll}
Take any
$\eps^\mathrm{com}_\mathrm{EC},\eps^\mathrm{com}_\mathrm{PE},\eps_\mathrm{PA},\eps_\mathrm{h},\eps_s \in (0,1]$, $\gamma\in(0,1)$, $\p\in[0,1/2]$, and $\delta_\mathrm{IID}\in[0,w_\mathrm{exp} - \delta_\mathrm{tol})$.
Define
\begin{align}
\epsilon_\mathrm{IID} \defvar
\cdfBin{\gamma n}{1-w_\mathrm{exp}+\delta_\mathrm{tol}+\delta_\mathrm{IID}}{\floor{(1-w_\mathrm{exp}+\delta_\mathrm{tol})\gamma n}}. \label{eq:eIIDoptcoll}
\end{align}
Under the collective-attacks assumption, Protocol~\ref{prot:optcoll} is $(\eps^\mathrm{com}_\mathrm{EC} + \eps^\mathrm{com}_\mathrm{PE})$-complete and $(\max\{\epsilon_\mathrm{IID}, \eps_\mathrm{PA} + 2\eps_s\} + \eps_\mathrm{h})$-sound
when performed with $\mathrm{EC}_\mathrm{max}$ defined in terms of $\eps^\mathrm{com}_\mathrm{EC}$ as described above,
and $\delta_\mathrm{tol},\ell_\mathrm{key}$ satisfying
\begin{align}
\eps^\mathrm{com}_\mathrm{PE} &\geq \cdfBin{\gamma n}{w_\mathrm{exp}}{\floor{(w_\mathrm{exp}-\delta_\mathrm{tol})\gamma n}}, \label{eq:ecPEoptcoll}\\
\ell_\mathrm{key} &\leq
(1-\gamma)n\lin_\p(w_\mathrm{exp}-\delta_\mathrm{tol}-\delta_\mathrm{IID}) - \sqrt{(1-\gamma)n} \,(2\log5)\sqrt{\log\frac{2}{\eps_s^2}}
\nonumber \\ &\qquad
- \mathrm{EC}_\mathrm{max} - \ceil{\log\left(\frac{1}{\eps_\mathrm{h}}\right)} - 2\log\frac{1}{\eps_\mathrm{PA}} + 2. \label{eq:lkeyoptcoll}
\end{align}
\end{theorem}
\begin{figure}
\centering
\subfloat[\cite{RBG+17} parameters, $\p=0$]{
\includegraphics[width=0.48\textwidth]{fig_finiteIID}
}
\subfloat[\cite{RBG+17} parameters, $\p=0.03$ (with heuristic $\lin_{\p}$)]{
\includegraphics[width=0.48\textwidth]{fig_finiteIIDheu}
}
\caption{Finite-size keyrates as a function of number of rounds in Protocol~\ref{prot:optcoll} (see Theorem~\ref{th:optcoll}),
using honest devices following the estimated parameters in~\cite{MvDR+19} for the loophole-free Bell test in~\cite{HBD+15}, for $\p=0$ and $\p=0.03$ (the latter is a rough estimate of the choice of $\p$ which yields the highest asymptotic keyrate for these experimental parameters). Note that the latter graph is computed using a heuristic estimate of $\lin_{\p}$ rather than a certified bound.
The error-correction protocol is taken to satisfy Eqs.~\eqref{eq:optEC}~and~\eqref{eq:optcollAEP}. The colours correspond to soundness parameters of $\eps^\mathrm{sou}=10^{-3}$, $10^{-6}$, and $10^{-9}$ for black, blue, and red respectively, and the completeness parameter is $\eps^\mathrm{com}=10^{-2}$ in all cases.
The horizontal line denotes the asymptotic keyrate.
All other parameters in Theorem~\ref{th:optcoll} were numerically optimized. The required number of rounds to achieve positive keyrate is substantially lower than Protocol~\ref{prot:DIQKD} (see Fig.~\ref{fig:experiments}).
}
\label{fig:protIID}
\end{figure}
In Fig.~\ref{fig:protIID}, we plot the results of Theorem~\ref{th:optcoll}, focusing on the~\cite{RBG+17} experiment. This protocol has improved finite-size performance as compared to the original Protocol~\ref{prot:DIQKD} (under the collective-attacks assumption) due to at least two factors. Firstly, we can potentially use larger $\gamma$ values, as previously mentioned (some of the points shown in the figure correspond to values ranging up to $\gamma \approx 0.3$).
Secondly, we find that for fixed values of $\gamma,n,w_\mathrm{exp},\delta_\mathrm{tol},\delta_\mathrm{IID}$, the binomial-distribution bounds for this protocol (Eqs.~\eqref{eq:eIIDoptcoll} and \eqref{eq:ecPEoptcoll}) are typically
several orders of magnitude better
than their counterparts for Protocol~\ref{prot:DIQKD} (Eqs.~\eqref{eq:eIIDbound} and \eqref{eq:ecPEbound1}). Intuitively, this arises because in Protocol~\ref{prot:DIQKD}, the number of test rounds is itself a random variable, hence increasing the variance in e.g.~the number of rounds where $C_j=1$.
Practically speaking, this means Protocol~\ref{prot:DIQKD} requires noticeably larger values of $\delta_\mathrm{tol}$ and $\delta_\mathrm{IID}$ in order to achieve given completeness and soundness parameters, hence reducing the keyrate by a nontrivial amount.
However, we see that even with the optimistic assumptions that yield Theorem~\ref{th:optcoll}, the keyrate for the estimated experimental parameters we consider only becomes positive at fairly large $n$. This indicates that substantial further work is necessary in order to achieve a demonstration of positive finite-size keyrates.
\section{Conclusion and further work}
\label{sec:conclusion}
In this work, we have performed a finite-size analysis for a protocol that combines several of the most promising approaches towards improving keyrates for DIQKD. Furthermore, we develop an algorithm that computes arbitrarily tight lower bounds on the asymptotic keyrates for protocols of this form (i.e.~allowing for noisy preprocessing and random key measurements, but restricted to one-way error correction), which applies to all 2-input 2-output scenarios. This allows us to prove a new threshold of $9.33\pct$ for noise tolerance in the depolarizing-noise model, and we show (see Sec.~\ref{sec:bestbnd}) that for one-way protocols, any further improvements on this threshold can only be fairly small. Finally, we propose a modified protocol, based on a pre-shared key, that overcomes the disadvantage of the sifting factor in random-key-measurement protocols.
We remark that while the finite-size analysis shown here is for a protocol based on the CHSH inequality, our algorithm for the asymptotic keyrates applies more generally to 2-input 2-output scenarios. If some exploration with this algorithm suggests that an improvement can be obtained by using an inequality other than CHSH, then it would not be difficult to modify the finite-size analysis for such an inequality, as was done in e.g.~\cite{BRC20} for DIRE (essentially, it would just correspond to having a different bound $\lin_{\p}$).
However, our results show that for the NV-centre experiment in~\cite{HBD+15} and the cold-atom experiment in~\cite{RBG+17}, an impractically large sample size (for those implementations) would still be needed in order to achieve a positive finite-size keyrate, even if one makes the optimistic assumption of collective attacks. A significant question that remains to be addressed is that of photonic experiments~\cite{SMC+15,GVW+15}, which achieve lower CHSH values but much larger sample sizes. Unfortunately, for photonic experiments the heuristic results in~\cite{SGP+21} suggest that the random-key-measurement approach is less useful in improving the keyrate, because the experimental parameters that achieve maximal CHSH value also tend to result in higher error probability (with respect to Bob's outputs) for one of Alice's measurements than the other. Despite this challenge, we note that there is much freedom in parameter optimization for photonic experiments~\cite{MSS20}, and given the algorithm we developed here for bounding the asymptotic keyrates, it is now possible to analyze variants such as choosing different amounts of noisy preprocessing for the two key-generating measurements. We aim to continue studying this in future work.
As another extension of our work, our algorithm for bounding the asymptotic DIQKD keyrates also applies to DIRE. Since the results it returns are arbitrarily tight and easily incorporated into the EAT, this could be combined with the finite-size analysis of~\cite{BRC20,LLR+19} to improve the results of those works --- the proof methods used there yield slightly suboptimal keyrates, in that they either bound the min-entropy rather than the von Neumann entropy~\cite{BRC20}, or they only bound the entropy of one party's outputs and are restricted to the CHSH inequality~\cite{LLR+19}. Our algorithm would yield tight bounds on the two-party von Neumann entropy for 2-input 2-output Bell inequalities; furthermore, it allows the possibility of using the random-key-measurement approach (by applying the pre-shared key proposal) to improve the keyrates.
\section*{Acknowledgements}
We are very grateful to Peter Brown for detailed information on the finite-size analysis in~\cite{BRC20,LLR+19}, as well as Omar Fawzi and Fr\'{e}d\'{e}ric Dupuis for discussions on the entropy accumulation theorem.
We also thank Koon Tong Goh and Ignatius W.\ Primaatmaja for helpful discussion and feedback.
E.~Y.-Z.~T.~and R.~R.~are supported by the Swiss National Science Foundation (SNSF) via the National Center for Competence in Research for Quantum Science and Technology (QSIT), the Air Force Office of Scientific Research (AFOSR) via grant FA9550-19-1-0202, and the QuantERA project eDICT.
P.~S.~is supported by by the Swiss National Science Foundation (SNSF).
C.~C.-W.~L is supported by the National Research Foundation (NRF) Singapore, under its NRF Fellowship programme (NRFF11-2019-0001) and Quantum Engineering Programme 1.0 (QEP-P2).
\section*{Computational platform}
Computations were performed using the MATLAB package YALMIP~\cite{yalmip} with the solver MOSEK~\cite{mosek}. Some of the calculations reported here were performed using the Euler cluster at ETH Z\"{u}rich.
\printbibliography
| 824eda53956c5413023900073a927bba1c276905 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The paper \cite{grilliette6} developed a categorical perspective of graph theory, slotting familiar graph-theoretic objects and operations into a diagram of categories and functors. Thus far, the work has considered the following structures: the category $\cat{Q}$ of quivers (i.e.\ directed multigraphs), the category $\cat{H}$ of set-system hypergraphs, the category $\cat{M}$ of (undirected) multigraphs, and the category $\cat{R}$ of incidence hypergraphs. In other works, set systems have been considered with antihomomorphisms \cite[p.\ 60]{hn1979}, which are a natural extension from both continuous functions of topological spaces and measurable functions of measurable functions.
The present work aims to extend the portrait of graph theory to include the various types of simple graphs and antihomomorphisms. To accomplish this goal, a unified simplification operation is realized as a universal construction for a comma category. Simplification naturally gives rise to the following structures: the category $\mathbf{Digra}$ of simple directed graphs, the category $\mathbf{SSys}$ of set systems, the category $\mathbf{Gra}$ of simple (undirected) graphs, and the category $\mathbf{IStr}$ of incidence structures. Also, the common graph-theoretic technique \cite[p.\ 123]{hell1979} of replacing an undirected edge with a directed 2-cycle is shown to be a categorical isomorphism between $\mathbf{Gra}$ and the category $\mathbf{SymDigra}$ of symmetric digraphs.
\begin{center}
\includegraphics[scale=1]{symmetric.pdf}
\end{center}
Including these categories completes the diagram begun in \cite{grilliette6}, giving Figure \ref{big-digram}.
\begin{figure}
\[\xymatrix{
\cat{Q}\ar@/^/[rr]^{U}\ar@/^/[dd]^{\mathcal{S}_{\cat{Q}}} & & \cat{M}\ar@/^/[rr]^{N}\ar@/^/[ll]^{\vec{D}}_{\perp} & & \cat{H}\ar@/^/[ll]^{\Del}_{\perp}\ar@/^/[dd]^{\mathcal{S}_{\cat{H}}}\\
\\
\mathbf{Digra}\ar@/^/[uu]^{\mathcal{N}_{\cat{Q}}}_{\vdash}\ar@/^1.5pc/[dr]^{N^{\diamond}_{\mathbf{Digra}}}_{\vdash}\ar@/_1.5pc/[dr]_{N^{\star}_{\mathbf{Digra}}}^{\vdash} & & & & \mathbf{SSys}\ar@/^/[uu]^{\mathcal{N}_{\cat{H}}}_{\vdash}\ar@/^/[dl]^{\Del_{\mathbf{SSys}}}\\
& \mathbf{SymDigra}\ar[ul]|-{N_{\mathbf{Digra}}}\ar@/^/[rr]^{W_{\mathbf{Gra}}} & & \mathbf{Gra}\ar@/^/[ur]^{N_{\mathbf{SSys}}}_{\dashv}\ar@/^/[ll]^{Z_{\mathbf{Gra}}}_{\cong}
}\]
\caption{Functorial diagram for $\mathbf{Digra}$, $\mathbf{SSys}$, \& $\mathbf{Gra}$}
\label{big-digram}
\end{figure}
On the other hand, this work creates a category $\cat{P}$ of set-system hypergraphs with antihomomorphisms, which allow for parallel edges. The category $\cat{P}$ generalizes the category $\mathbf{Anti}$ of set systems with antihomomorphisms, as well as categories $\mathbf{Top}$ of topological spaces and $\mathbf{Meas}$ of measurable spaces. These categories are pictured in Figure \ref{big-digram2}. The existence of the category $\cat{P}$ motivates questions about ``multi-edge'' versions of $\mathbf{Top}$ or $\mathbf{Meas}$, and what applications could involve either.
\begin{figure}
\[\xymatrix{
& \cat{P}\ar@/^/[d]^{\mathcal{T}_{\cat{P}}}_{\dashv}\\
& \mathbf{Anti}\ar@/^/[u]^{\mathcal{N}_{\cat{P}}}\ar@/^/[dr]^{\Sigma}_{\dashv}\ar@/^/[dl]^{\Gamma}_{\dashv}\\
\mathbf{Top}\ar@/^/[ur]^{\mathcal{N}_{\mathbf{Top}}} & & \mathbf{Meas}\ar@/^/[ul]^{\mathcal{N}_{\mathbf{Meas}}}
}\]
\caption{Functorial diagram for $\cat{P}$, $\mathbf{Anti}$, $\mathbf{Top}$, \& $\mathbf{Meas}$}
\label{big-digram2}
\end{figure}
Please be aware that ``simple'' in this paper will be allowing loops. Without loops, the respective irreflexive categories will not have a terminal object \cite[p.\ 7]{brown2008}. Moreover, the right adjoint to the vertex functor creates a complete graph with loops \cite[p.\ 4, 6]{grilliette6}. As these universal objects have loops, loops and 1-edges will be allowed in ``simple'' graphs.
However, in studying these categories and the functors connecting them, some surprising results appear. First, the simplification manifests as a left adjoint, making the category of simple objects is a reflective subcategory. Hence, all limit processes for simple graphs can be performed by moving to the parent category and then applying the simplification to remove parallelisms. Pedagogically, this relationship informs how products and quotients of simple graphs can be developed from the natural componentwise operations of the parent categories.
Second, simple graphs categorically act more like directed graphs than they do set systems. As mentioned above, $\mathbf{Gra}$ is isomorphic to $\mathbf{SymDigra}$ as categories. While $\mathbf{Gra}$ is a coreflective subcategory of $\mathbf{SSys}$, just as $\cat{M}$ is a coreflective subcategory of $\cat{H}$, $\mathbf{SymDigra}$ is both a reflective and a coreflective subcategory of $\mathbf{Digra}$. Therefore, the categorical structure of $\mathbf{SymDigra}$, and $\mathbf{Gra}$ by extension, is more deeply related to $\mathbf{Digra}$ than to $\mathbf{SSys}$. This relationship gives a well-founded understanding, and a categorical reason, for why moving simple graphs to directed graphs is so canonical.
Please be aware that this behavior appears to be a feature of simplicity, considering that the analogous relationship does not hold between $\cat{Q}$ and $\cat{M}$ \cite[Lemma 2.38]{grilliette6}.
Third, simplification has a natural dual notion, a ``cosimplification''. With simplification represented as a left adjoint, dualizing the process yields a right adjoint. When applied to $\cat{Q}$ and $\cat{R}$, cosimplification removes isolated vertices and loose edges. However, if one uses antihomomorphisms, cosimplification becomes traditional simplification. Thus, the characterization of cosimplification depends on the morphisms being used. This highlights the effect that a change of morphisms can have, and the importance on clearly stating what morphisms are in play.
Section \ref{simplification} defines a \emph{simple} object of a comma category and devises the generalized simplification process. Section \ref{structured} intertwines the full category of simple objects with functor-structured categories in \cite{adamek1980,katetov1963,kucera1972,solovyov2013}, which are closer to the traditional definitions of simple graphs.
Section \ref{Examples} applies these constructions to $\cat{Q}$, $\cat{H}$, and $\cat{R}$, demonstrating how simplification manifests for each. Section \ref{simplegraphs} addresses the connection between $\mathbf{Digra}$, $\mathbf{SymDigra}$, $\mathbf{Gra}$, and $\mathbf{SSys}$.
Section \ref{cosimplification} briefly defines a \emph{cosimple} object of a comma category and summarizes the cosimplification process, before applying it to $\cat{Q}$ and $\cat{R}$. Section \ref{antihomomorphisms} defines $\cat{P}$ before applying the cosimplification process.
\section{Simple Objects \& Simplification}\label{simplification}
Throughout this section, let $\xymatrix{\cat{A}\ar[r]^{F} & \cat{C} & \cat{B}\ar[l]_{G}}$ be functors, $\cat{G}:=\left(F\downarrow G\right)$ be the comma category, and $\xymatrix{\cat{A} & \cat{G}\ar[l]_{P}\ar[r]^{Q} & \cat{B}}$ be the canonical projections. In graph theory \cite[Theorem 1]{dorfler1975}, a graph is \emph{simple} if it has no parallel edges, i.e.\ if the incidence function is one-to-one. In a general comma category, the following definition is taken for a ``simple'' object.
\begin{defn}[Simple object]
An object $(A,f,B)\in\ob(\cat{G})$ is \emph{simple} if $f$ is monic in $\cat{C}$. Let $\cat{SG}$ be the full subcategory of $\cat{G}$ consisting of simple objects, and let $\xymatrix{\cat{SG}\ar[r]^{\mathcal{N}} & \cat{G}}$ be the inclusion functor. A routine check shows that $\cat{SG}$ is a replete subcategory of $\cat{G}$.
\end{defn}
A common pratice in the study of simple graphs is to regard homomorphisms as functions between the vertex sets \cite[p.\ 53]{hn1979}, as the map of edges is uniquely determined. This fact carries over to simple objects in a general comma category, provided the functor $F$ is faithful.
\begin{prop}[Maps into a simple object]\label{simple-maps}
Assume $F$ is faithful. If $\xymatrix{(A,f,B)\ar@/^/[r]^{(\phi,\psi)}\ar@/_/[r]_{(\varphi,\chi)} & \left(A',f',B'\right)}\in\cat{G}$ satisfy that $G(\psi)=G(\chi)$ and $\left(A',f',B'\right)$ is simple, then $\phi=\varphi$.
\end{prop}
\begin{proof}
Notice that $f'\circ F(\phi)
=G(\psi)\circ f
=G(\chi)\circ f
=f'\circ F(\varphi)$. As $f'$ is monic, $F(\phi)=F(\varphi)$. As $F$ is faithful, $\phi=\varphi$.
\end{proof}
\begin{cor}[Morphisms in $\cat{SG}$]\label{morphisms-simple}
Assume $F$ is faithful. A morphism $\xymatrix{(A,f,B)\ar[r]^{(\phi,\psi)} & \left(A',f',B'\right)}\in\cat{SG}$ is uniquely determined by $\psi$.
\end{cor}
In order to ``simplify'' an object of a comma category to a simple object, motivation is taken again from graph theory. In \cite{bumby1986,dorfler1975}, the simplification process is achieved by a quotient operation, tacetly using the image factorization. A simple graph itself can be regarded as a subgraph of a complete graph \cite[p.\ 316]{kung1998}.
The complete graph is realized through a right adjoint to the vertex functor, and this right adjoint $Q^{\star}$ is guaranteed to exist in a general comma category when $F$ admits a right adjoint \cite[Definition 1.11]{grilliette6}. To embed into such a ``complete'' object, $\cat{A}$ is assumed to be regular to ensure an image factorization. Under the above conditions, a ``simplified'' object can be constructed.
\begin{defn}[Simplification]\label{simple-construction}
Assume that $\cat{A}$ is regular, and that $F$ admits a right adjoint $F^{\star}$ with counit $\xymatrix{FF^{\star}\ar[r]^{\theta} & id_{\cat{C}}}$. For $(A,f,B)\in\ob(\cat{G})$, there is a unique $\xymatrix{A\ar[r]^(0.4){\hat{f}} & F^{\star}G(B)}\in\cat{A}$ such that $\theta_{G(B)}\circ F\left(\hat{f}\right)=f$. Let $\xymatrix{A\ar@{->>}[r]^(0.4){e_{\hat{f}}} & \Ran\left(\hat{f}\right)\textrm{ }\ar@{>->}[r]^{m_{\hat{f}}} & F^{\star}G(B)}\in\cat{A}$ be an image factorization of $\hat{f}$. Define
\[
\mathcal{S}(A,f,B):=\left(\Ran\left(\hat{f}\right),\theta_{G(B)}\circ F\left(m_{\hat{f}}\right),B\right)
\]
and $\eta_{(A,f,B)}:=\left(e_{\hat{f}},id_{B}\right)$. By the commutative diagram below, $\eta_{(A,f,B)}$ is an epimorphism in $\cat{G}$ from $(A,f,B)$ onto $\mathcal{S}(A,f,B)$, and $\left(m_{\hat{f}},id_{B}\right)$ is a monomorphism in $\cat{G}$ from $\mathcal{S}(A,f,B)$ into $Q^{\star}(B)=\left(F^{\star}G(B),\theta_{G(B)},B\right)$.
\end{defn}
\[\xymatrix{
F(A)\ar[dd]_{f}\ar[rrd]_{F\left(\hat{f}\right)}\ar[rr]^{F\left(e_{\hat{f}}\right)} & & F\left(\Ran\left(\hat{f}\right)\right)\ar[d]^{F\left(m_{\hat{f}}\right)}\ar[rr]^{F\left(m_{\hat{f}}\right)} & & FF^{\star}G(B)\ar[dd]^{\theta_{G(B)}}\\
& & FF^{\star}G(B)\ar[d]^{\theta_{G(B)}}\\
G(B)\ar[rr]_{G\left(id_{B}\right)} & & G(B)\ar[rr]_{G\left(id_{B}\right)} & & G(B)
}\]
By standard results, $m_{\hat{f}}$ is monic in $\cat{A}$, so preservation of monomorphisms by $F$ and monic $\theta$ are sufficient to ensure that both $\mathcal{S}(A,f,B)$ and $Q^{\star}(B)$ are always simple.
\begin{lem}[Conditions for simplicity]\label{cond-simp}
Let $\cat{A}$ be regular, and assume that $F$ preserves monomorphisms and admits a right adjoint $F^{\star}$ with a monic counit $\xymatrix{FF^{\star}\textrm{ }\ar@{>->}[r]^{\theta} & id_{\cat{C}}}$. Then, $\xymatrix{\mathcal{S}(A,f,B)\ar[rr]^(0.55){\left(m_{\hat{f}},id_{B}\right)} & & Q^{\star}(B)}\in\cat{SG}$ for all $(A,f,B)\in\ob(\cat{G})$.
\end{lem}
As one would expect, the simplification of a simple object is trivial under the condition that $F$ is faithful.
\begin{lem}[Simplification of simple]
Let $\cat{A}$ be regular, and assume that $F$ is faithful and admits a right adjoint with counit $\xymatrix{FF^{\star}\ar[r]^{\theta} & id_{\cat{C}}}$. If $(A,f,B)\in\ob(\cat{SG})$, then $\eta_{(A,f,B)}$ is an isomorphism.
\end{lem}
\begin{proof}
As $\theta_{G(B)}\circ F\left(\hat{f}\right)=f$ is monic, $F\left(\hat{f}\right)$ is monic. As $F$ is faithful, $\hat{f}$ is monic. As $m_{\hat{f}}\circ e_{\hat{f}}=\hat{f}$, $e_{\hat{f}}$ is monic. As $e_{\hat{f}}$ is a regular epimorphism and monic, it is an isomorphism.
\end{proof}
Moreover, the faithful condition yields a universal property: any map into a simple object factors through the simplification. With the conditions in Lemma \ref{cond-simp}, $\cat{SG}$ is a reflective subcategory of $\cat{G}$.
\begin{thm}[Universal property of $\mathcal{S}$]
Let $\cat{A}$ be regular, and assume that $F$ is faithful and admits a right adjoint with counit $\xymatrix{FF^{\star}\ar[r]^{\theta} & id_{\cat{C}}}$. If $\xymatrix{(A,f,B)\ar[rr]^{(\phi,\psi)} & & \mathcal{N}\left(A',f',B'\right)}\in\cat{G}$, there is a unique $\xymatrix{\mathcal{S}(A,f,B)\ar[rr]^{\left(\hat{\phi},\hat{\psi}\right)} & & \left(A',f',B'\right)}\in\cat{G}$ such that $\left(\hat{\phi},\hat{\psi}\right)\circ\eta_{(A,f,B)}=(\phi,\psi)$.
\end{thm}
\begin{proof}
Let $\xymatrix{A & R\ar[l]_{p_{1}}\ar[r]^{p_{2}} & A}\in\cat{A}$ be a kernel pair of $\hat{f}$, and $\xymatrix{A\ar[r]^(0.4){e_{\hat{f}}} & \Ran\left(\hat{f}\right)}\in\cat{A}$ be a coequalizer of $p_{1}$ and $p_{2}$. As $\hat{f}\circ p_{1}=\hat{f}\circ p_{2}$, there is a unique $\xymatrix{\Ran\left(\hat{f}\right)\ar[r]^{m_{\hat{f}}} & F^{\star}G(B)}\in\cat{A}$ such that $m_{\hat{f}}\circ e_{\hat{f}}=\hat{f}$. For $n=1,2$, observe that
\[
f'\circ F\left(\phi\circ p_{n}\right)
=G\left(\psi\right)\circ f\circ F\left(p_{n}\right)
=G\left(\psi\right)\circ\theta_{G(B)}\circ F\left(\hat{f}\circ p_{n}\right).
\]
As $\hat{f}\circ p_{1}=\hat{f}\circ p_{2}$, the equalities above yield $f'\circ F\left(\phi\circ p_{1}\right)=f'\circ F\left(\phi\circ p_{2}\right)$. As $f'$ is monic, $F\left(\phi\circ p_{1}\right)=F\left(\phi\circ p_{2}\right)$. As $F$ is faithful, $\phi\circ p_{1}=\phi\circ p_{2}$. By the universal property of $\Ran\left(\hat{f}\right)$, there is a unique $\xymatrix{\Ran\left(\hat{f}\right)\ar[r]^(0.6){\hat{\phi}} & A'}\in\cat{A}$ such that $\hat{\phi}\circ e_{\hat{f}}=\phi$. Consider the diagram below.
\[\xymatrix{
F(A)\ar[dd]_{f}\ar[rrd]_{F\left(\hat{f}\right)}\ar[rr]^{F\left(e_{\hat{f}}\right)} & & F\left(\Ran\left(\hat{f}\right)\right)\ar[d]^{F\left(m_{\hat{f}}\right)}\ar[rr]^{F\left(\hat{\phi}\right)} & & F\left(A'\right)\ar[dd]^{f'}\\
& & FF^{\star}G(B)\ar[d]^{\theta_{G(B)}}\\
G(B)\ar[rr]_{G\left(id_{B}\right)} & & G(B)\ar[rr]_{G\left(\psi\right)} & & G\left(B'\right)\\
}\]
Notice that
\[\begin{array}{rcl}
f'\circ F\left(\hat{\phi}\right)\circ F\left(e_{\hat{f}}\right)
& = & f'\circ F\left(\phi\right)
=G(\psi)\circ f
=G(\psi)\circ\theta_{G(B)}\circ F\left(\hat{f}\right)\\
& = & G(\psi)\circ\left(\theta_{G(B)}\circ F\left(m_{\hat{f}}\right)\right)\circ F\left(e_{\hat{f}}\right).\\
\end{array}\]
Since $e_{\hat{f}}$ is epic and $F$ is cocontinuous, $F\left(e_{\hat{f}}\right)$ is epic, giving that $f'\circ F\left(\hat{\phi}\right)=G(\psi)\circ\left(\theta_{G(B)}\circ F\left(m_{\hat{f}}\right)\right)$. Thus, $\left(\hat{\phi},\psi\right)$ is a morphism in $\cat{G}$ from $\mathcal{S}(A,f,B)$ to $\left(A',f',B'\right)$. By construction, $\left(\hat{\phi},\psi\right)\circ\eta_{(A,f,B)}=(\phi,\psi)$, and uniqueness follows from Proposition \ref{simple-maps}.
\end{proof}
\begin{cor}[Reflective subcategory]\label{reflective}
Let $\cat{A}$ be regular, and assume that $F$ is faithful, preserves monomorphisms, and admits a right adjoint with a monic counit $\xymatrix{FF^{\star}\ar[r]^{\theta} & id_{\cat{C}}}$. If $\xymatrix{(A,f,B)\ar[rr]^{(\phi,\psi)} & & \mathcal{N}\left(A',f',B'\right)}\in\cat{G}$, there is a unique $\xymatrix{\mathcal{S}(A,f,B)\ar[rr]^{\left(\hat{\phi},\hat{\psi}\right)} & & \left(A',f',B'\right)}\in\cat{SG}$ such that $\mathcal{N}\left(\hat{\phi},\hat{\psi}\right)\circ\eta_{(A,f,B)}=(\phi,\psi)$. Consequently, $\cat{SG}$ is a reflective subcategory of $\cat{G}$.
\end{cor}
\section{Relationship to Structured Categories}\label{structured}
Consider the case when $\cat{A}=\cat{C}$ and $F=id_{\cat{A}}$. Letting $\xymatrix{\cat{B}\ar[r]^{G} & \cat{A}}$ be any functor, the comma category $\cat{G}:=\left(id_{\cat{A}}\downarrow G\right)$ would consist of objects $(A,f,B)$, where $\xymatrix{A\ar[r]^{f} & G(B)}\in\cat{A}$, like a generalized coalgebra as in \cite{rutten2000}.
If $(A,f,B)$ is in the full subcategory $\cat{SG}$ of $\cat{G}$ consisting of simple objects, then $(A,f)$ represents a subobject of $G(B)$ in $\cat{A}$. Moreover, as $id_{\cat{A}}$ is faithful, Corollary \ref{morphisms-simple} shows that a morphism in $\cat{SG}$ is determined uniquely by its $\cat{B}$-coordinate. Indeed, $\cat{SG}$ seems to generalize the functor-structured categories discussed in \cite{adamek1980,katetov1963,kucera1972,solovyov2013}, a connection which will be made explicit.
With the notation above, let $\cat{A}=\mathbf{Set}$ be the category of sets with functions. Letting $\mathbf{Spa}(G)$ be the category of $G$-spaces with $G$-maps, the following pair of functors are defined to form an equivalence between $\cat{SG}$ and $\mathbf{Spa}(G)$.
\begin{defn}[Comma \& space functors]\label{comma-space-equivalence}
For $\xymatrix{(B,\alpha)\ar[r]^{\psi} & \left(B',\alpha'\right)}\in\mathbf{Spa}(G)$, define $\xymatrix{\Comma(B,\alpha)\ar[rr]^{\Comma(\psi)} & & \Comma\left(B',\alpha'\right)}\in\cat{SG}$ by
\begin{itemize}
\item $\Comma(B,\alpha):=\left(\alpha,\iota_{\alpha},B\right)$, where $\iota_{\alpha}$ is the inclusion map of $\alpha$ into $G(B)$;
\item $\Comma(\psi):=\left(\tilde{\psi},\psi\right)$, where $\tilde{\psi}(x):=G(\psi)(x)$.
\end{itemize}
For $\xymatrix{(A,f,B)\ar[r]^{(\phi,\psi)} & \left(A',f',B'\right)}\in\cat{SG}$, define $\xymatrix{\Space(A,f,B)\ar[rr]^{\Space(\phi,\psi)} & & \Space\left(A',f',B'\right)}\in\mathbf{Spa}(G)$ by
\begin{itemize}
\item $\Space(A,f,B):=\left(B,\alpha_{f}\right)$, where $\alpha_{f}$ is the set-theoretic range of $f$;
\item $\Space(\phi,\psi):=\psi$.
\end{itemize}
\end{defn}
Routine calculations show that both $\Comma$ and $\Space$ define functors, and that $\Space\Comma=id_{\mathbf{Spa}(G)}$. All that remains is to demonstrate a natural isomorphism between $\Comma\Space$ and $id_{\cat{SG}}$, which is constructed below using the set-theoretic image factorization.
\begin{defn}[Comma-space isomorphism]
Given $(A,f,B)\in\ob(\cat{SG})$, define $q_{f}:A\to\alpha_{f}$ by $q_{f}(x):=f(x)$, and let $\zeta_{(A,f,B)}:=\left(q_{f},id_{B}\right)$ and $\zeta:=\left(\zeta_{(A,f,B)}\right)_{(A,f,B)\in\cat{SG}}$.
\end{defn}
A quick check shows that $\zeta$ is a natural transformation from $id_{\cat{SG}}$ to $\Comma\Space$. Since $f$ is one-to-one for $(A,f,B)\in\ob(\cat{SG})$, $q_{f}$ is bijective, showing that $\zeta_{(A,f,B)}$ is an isomorphism. Moreover, note that $\mathbf{Set}$ is regular and that $id_{\mathbf{Set}}$ satisfies the conditions of Corollary \ref{reflective}, giving the result below.
\begin{thm}[Spaces within commas]\label{spaces-in-commas}
Let $\xymatrix{\cat{B}\ar[r]^{G} & \mathbf{Set}}$ be a functor. Then, $\mathbf{Spa}(G)$ is equivalent to the reflective subcategory of simple objects within $\left(id_{\mathbf{Set}}\downarrow G\right)$.
\end{thm}
\section{Examples}\label{Examples}
\subsection{Quivers \& Digraphs}
Let $\cat{E}$ be the finite category drawn below.
\[\xymatrix{
1\ar@/^/[r]^{s}\ar@/_/[r]_{t} & 0
}\]
The category $\cat{Q}:=\mathbf{Set}^{\cat{E}}$ is the category of directed multigraphs, or quivers, as seen in \cite[p.\ 2]{bumby1986}. Let $\Delta:\mathbf{Set}\to\mathbf{Set}\times\mathbf{Set}$ be the diagonal functor with its right adjoint $\Delta^{\star}$. The category $\cat{Q}_{1}:=\left(id_{\mathbf{Set}}\downarrow \Delta^{\star}\Delta\right)$ is the category of quivers as seen in \cite{bondy-murty,knauer}. The isomorphism between $\cat{Q}$ and $\cat{Q}_{1}$ follows from the universal property of the product in $\mathbf{Set}$ within the diagram below for each $Q=\left(\vec{V}(Q),\vec{E}(Q),\sigma_{Q},\tau_{Q}\right)\in\ob\left(\cat{Q}\right)$.
\[\xymatrix{
& & \vec{E}(Q)\ar[dll]_{\sigma_{Q}}\ar[drr]^{\tau_{Q}}\ar@{..>}[d]^{\exists!\vec{\epsilon}_{Q}}\\
\vec{V}(Q) & & \vec{V}(Q)\times\vec{V}(Q)\ar[ll]^{\pi_{\vec{V}(Q)}^{1}}\ar[rr]_{\pi_{\vec{V}(Q)}^{2}} & & \vec{V}(Q)\\
}\]
\begin{defn}[Isomorphism of $\cat{Q}$ and $\cat{Q}_{1}$]
For a set $X$, let $\xymatrix{X & X\times X\ar[l]_(0.6){\pi_{X}^{1}}\ar[r]^(0.6){\pi_{X}^{2}} & X}\in\mathbf{Set}$ be the coordinate projections. For $\xymatrix{Q\ar[r]^{\phi} & Q'}\in\cat{Q}$, define $\xymatrix{W_{\cat{Q}}(Q)\ar[r]^{W_{\cat{Q}}(\phi)} & W_{\cat{Q}}\left(Q'\right)}\in\cat{Q}_{1}$ by
\begin{itemize}
\item $W_{\cat{Q}}(Q):=\left(\vec{E}(Q),\vec{\epsilon}_{Q},\vec{V}(Q)\right)$, where $\vec{\epsilon}_{Q}(e):=\left(\sigma_{Q}(e),\tau_{Q}(e)\right)$;
\item $W_{\cat{Q}}(\phi):=\left(\vec{E}(\phi),\vec{V}(\phi)\right)$.
\end{itemize}
For $\xymatrix{Q\ar[r]^{\phi} & Q'}\in\cat{Q}_{1}$, define $\xymatrix{Z_{\cat{Q}}(Q)\ar[r]^{Z_{\cat{Q}}(\phi)} & Z_{\cat{Q}}\left(Q'\right)}\in\cat{Q}$ by
\begin{itemize}
\item $Z_{\cat{Q}}(Q):=\left(\vec{V}(Q),\vec{E}(Q),\pi_{\vec{V}(Q)}^{1}\circ\vec{\epsilon}_{Q},\pi_{\vec{V}(Q)}^{2}\circ\vec{\epsilon}_{Q}\right)$;
\item $Z_{\cat{Q}}(\phi):=\left(\vec{V}(\phi),\vec{E}(\phi)\right)$.
\end{itemize}
Routine calculations show that $W_{\cat{Q}}Z_{\cat{Q}}=id_{\cat{Q}_{1}}$ and $Z_{\cat{Q}}W_{\cat{Q}}=id_{\cat{Q}}$.
\end{defn}
Let $\cat{SQ}_{1}$ be the full subcategory of $\cat{Q}_{1}$ of simple objects, and $\xymatrix{\cat{SQ}_{1}\ar[r]^{\mathcal{N}_{\cat{Q}_{1}}} & \cat{Q}_{1}}$ be the inclusion functor. By Corollary \ref{reflective}, $\mathcal{N}_{\cat{Q}_{1}}$ admits a left adjoint $\mathcal{S}_{\cat{Q}_{1}}$ with the following action on objects:
\[
\mathcal{S}_{\cat{Q}_{1}}\left(\vec{E}(Q),\vec{\epsilon}_{Q},\vec{V}(Q)\right)
=
\left(\Ran\left(\vec{\epsilon}_{Q}\right),m_{\vec{\epsilon}_{Q}},\vec{V}(Q)\right),
\]
where $\Ran\left(\vec{\epsilon}_{Q}\right)=\left\{\vec{\epsilon}_{Q}(e):e\in\vec{E}(Q)\right\}$ and $m_{\vec{\epsilon}_{Q}}(v,w)=(v,w)$.
Let $\mathbf{Digra}:=\mathbf{Spa}\left(\Delta^{\star}\Delta\right)$, which is the category of ``directed graphs'' and ``digraph homomorphisms'' from \cite[p.\ 122]{hell1979}. Letting $\xymatrix{\mathbf{Digra}\ar@/^/[r]^{\Comma_{\cat{Q}}} & \cat{SQ}_{1}\ar@/^/[l]^{\Space_{\cat{Q}}}}$ be the functors defined in Definition \ref{comma-space-equivalence}, Theorem \ref{spaces-in-commas} gives an equivalence.
Setting all of the above categories and functors in line together yields the following progression from digraphs as spaces, to directed graphs as objects of a comma category, to quivers as objects of a functor category.
\[\xymatrix{
\mathbf{Digra}\ar@/^/[rr]^{\Comma_{\cat{Q}}}_{\simeq} & & \cat{SQ}_{1}\ar@/^/[ll]^{\Space_{\cat{Q}}}\ar@/^/[rr]^{\mathcal{N}_{\cat{Q}_{1}}}_{\intercal} & & \cat{Q}_{1}\ar@/^/[ll]^{\mathcal{S}_{\cat{Q}_{1}}}\ar@/^/[rr]^{Z_{\cat{Q}}}_{\cong} & & \cat{Q}\ar@/^/[ll]^{W_{\cat{Q}}}
}\]
Therefore, the following result is an immediate consequence.
\begin{thm}[Digraphs \& quivers]
The category $\mathbf{Digra}$ is equivalent to the reflective subcategory of simple quivers within $\cat{Q}$. Moreover, $\mathbf{Digra}$ is complete and cocomplete with limits performed by passing to $\cat{Q}$ and then applying the simplification.
\end{thm}
To reduce notation, let $\mathcal{N}_{\cat{Q}}:=Z_{\cat{Q}}\mathcal{N}_{\cat{Q}_{1}}\Comma_{\cat{Q}}$ and $\mathcal{S}_{\cat{Q}}:=\Space_{\cat{Q}}\mathcal{S}_{\cat{Q}_{1}}W_{\cat{Q}}$. Figure \ref{quiver-example} gives an example to illustrate that $\mathcal{S}_{\cat{Q}}$ does indeed remove parallelisms in a quiver.
\begin{figure}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$Q$ & $\mathcal{S}_{\cat{Q}}(Q)$\\\hline
\includegraphics[scale=1]{quiver1.pdf}
&
\includegraphics[scale=1]{quiver2.pdf}\\\hline
\end{tabular}
\end{center}
\label{quiver-example}
\caption{Simplification of a quiver}
\end{figure}
\subsection{Set-System Hypergraphs \& Set Systems}
Let $\mathcal{P}:\mathbf{Set}\to\mathbf{Set}$ be the covariant power-set functor. The category $\cat{H}:=\left(id_{\mathbf{Set}}\downarrow \mathcal{P}\right)$ is the category of set-system hypergraphs from \cite[p.\ 153-154]{dorfler1975}.
Let $\cat{SH}$ be the full subcategory of $\cat{H}$ of simple objects, and $\xymatrix{\cat{SH}\ar[r]^{\mathcal{N}_{\cat{H}_{1}}} & \cat{H}}$ be the inclusion functor. By Corollary \ref{reflective}, $\mathcal{N}_{\cat{H}_{1}}$ admits a left adjoint $\mathcal{S}_{\cat{H}_{1}}$ with the following action on objects:
\[
\mathcal{S}_{\cat{H}_{1}}\left(E(G),\epsilon_{G},V(G)\right)
=
\left(\Ran\left(\epsilon_{G}\right),m_{\epsilon_{G}},V(G)\right),
\]
where $\Ran\left(\epsilon_{G}\right)=\left\{\epsilon_{G}(e):e\in E(G)\right\}$ and $m_{\epsilon_{G}}(A)=A$.
Let $\mathbf{SSys}:=\mathbf{Spa}(\mathcal{P})$. An object $G$ of $\mathbf{SSys}$ consists of a set $V(G)$ and a family of subsets $\beta_{G}\subseteq\mathcal{P}V(G)$, which is precisely a ``set system'' or ``hypergraph'' as defined in \cite[p.\ 53]{hn1979}. A morphism $\xymatrix{G\ar[r]^{f} & H}\in\mathbf{SSys}$ is a ``hypergraph homomorphism'' as defined in \cite[p.\ 53]{hn1979}. Letting $\xymatrix{\mathbf{SSys}\ar@/^/[r]^{\Comma_{\cat{H}}} & \cat{SH}\ar@/^/[l]^{\Space_{\cat{H}}}}$ be the functors defined in Definition \ref{comma-space-equivalence}, Theorem \ref{spaces-in-commas} gives an equivalence.
Setting all of the above categories and functors in line together yields the following progression from set systems as spaces to set-system hypergraphs as objects of a comma category. Moreover, the limit properties of $\mathbf{SSys}$ are inherited from $\cat{H}$.
\[\xymatrix{
\mathbf{SSys}\ar@/^/[rr]^{\Comma_{\cat{H}}}_{\simeq} & & \cat{SH}\ar@/^/[ll]^{\Space_{\cat{H}}}\ar@/^/[rr]^{\mathcal{N}_{\cat{H}_{1}}}_{\intercal} & & \cat{H}\ar@/^/[ll]^{\mathcal{S}_{\cat{H}_{1}}}
}\]
\begin{thm}[Set systems \& hypergraphs]
The category $\mathbf{SSys}$ is complete and cocomplete with limits performed by passing to $\cat{H}$ and then applying the simplification.
\end{thm}
To reduce notation, let $\mathcal{N}_{\cat{H}}:=\mathcal{N}_{\cat{H}_{1}}\Comma_{\cat{H}}$ and $\mathcal{S}_{\cat{H}}:=\Space_{\cat{H}}\mathcal{S}_{\cat{H}_{1}}$. Figure \ref{set-system-example} gives an example to illustrate that $\mathcal{S}_{\cat{H}}$ does indeed remove parallelisms in a set-system hypergraph.
\begin{figure}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$G$ & $\mathcal{S}_{\cat{H}}(G)$\\\hline
\includegraphics[scale=1]{setsys1.pdf}
&
\includegraphics[scale=1]{setsys2.pdf}\\\hline
\end{tabular}
\end{center}
\caption{Simplification of a set-system hypergraph}
\label{set-system-example}
\end{figure}
\subsection{Incidence Hypergraphs \& Incidence Structures}
Let $\cat{D}$ be the finite category drawn below.
\[\xymatrix{
0 & 2\ar[l]_{y}\ar[r]^{z} & 1\\
}\]
This category $\cat{R}:=\mathbf{Set}^{\cat{D}}$ is the category of incidence hypergraphs from \cite{grilliette7,grilliette6}. The category $\cat{R}_{1}:=\left(id_{\mathbf{Set}}\downarrow \Delta^{\star}\right)$ is isomorphic to $\cat{R}$ via the diagram below, for each $G=\left(\check{V}(G),\check{E}(G),I(G),\varsigma_{G},\omega_{G}\right)$.
\[\xymatrix{
& & I(G)\ar[dll]_{\varsigma_{G}}\ar[drr]^{\omega_{G}}\ar@{..>}[d]^{\exists!\iota_{G}}\\
\check{V}(G) & & \check{V}(G)\times\check{E}(G)\ar[ll]^{\pi_{\check{V}(G)}}\ar[rr]_{\pi_{\check{E}(G)}} & & \check{E}(G)\\
}\]
\begin{defn}[Isomorphism of $\cat{R}$ and $\cat{R}_{1}$]
For a sets $X$ and $Y$, let $\xymatrix{X & X\times Y\ar[l]_(0.6){\pi_{X}}\ar[r]^(0.6){\pi_{Y}} & Y}\in\mathbf{Set}$ be the coordinate projections. For $\xymatrix{G\ar[r]^{\phi} & G'}\in\cat{R}$, define $\xymatrix{W_{\cat{R}}(G)\ar[r]^{W_{\cat{R}}(\phi)} & W_{\cat{R}}\left(G'\right)}\in\cat{R}_{1}$ by
\begin{itemize}
\item $W_{\cat{R}}(G):=\left(I(G),\iota_{G},\left(\check{V}(G),\check{E}(G)\right)\right)$, where $\iota_{G}(i):=\left(\varsigma_{G}(i),\omega_{G}(i)\right)$;
\item $W_{\cat{R}}(\phi):=\left(I(\phi),\left(\check{V}(\phi),\check{E}(\phi)\right)\right)$.
\end{itemize}
For $\xymatrix{G\ar[r]^{\phi} & G'}\in\cat{R}_{1}$, define $\xymatrix{Z_{\cat{R}}(G)\ar[r]^{Z_{\cat{R}}(\phi)} & Z_{\cat{R}}\left(G'\right)}\in\cat{R}$ by
\begin{itemize}
\item $Z_{\cat{R}}(G):=\left(\check{V}(G),\check{E}(G),I(G),\pi_{\check{V}(G)}\circ\iota_{G},\pi_{\check{E}(G)}\circ\iota_{G}\right)$;
\item $Z_{\cat{R}}(\phi):=\left(\check{V}(\phi),\check{E}(\phi),I(\phi)\right)$.
\end{itemize}
Routine calculations show that $W_{\cat{R}}Z_{\cat{R}}=id_{\cat{R}_{1}}$ and $Z_{\cat{R}}W_{\cat{R}}=id_{\cat{R}}$.
\end{defn}
Let $\cat{SR}_{1}$ be the full subcategory of $\cat{R}_{1}$ of simple objects, and $\xymatrix{\cat{SR}_{1}\ar[r]^{\mathcal{N}_{\cat{R}_{1}}} & \cat{R}_{1}}$ be the inclusion functor. By Corollary \ref{reflective}, $\mathcal{N}_{\cat{R}_{1}}$ admits a left adjoint $\mathcal{S}_{\cat{R}_{1}}$ with the following action on objects:
\[
\mathcal{S}_{\cat{R}_{1}}\left(I(G),\iota_{G},\left(\check{V}(G),\check{E}(G)\right)\right)
=
\left(\Ran\left(\iota_{G}\right),m_{\iota_{G}},\left(\check{V}(G),\check{E}(G)\right)\right),
\]
where $\Ran\left(\iota_{G}\right)=\left\{\iota_{G}(i):i\in I(G)\right\}$ and $m_{\iota_{G}}(v,e)=(v,e)$.
Let $\mathbf{IStr}:=\mathbf{Spa}\left(\Delta^{\star}\right)$, which is precisely the category of incidence structures and incidence-preserving homomorphisms from \cite[p.\ 7-8]{bumby1986}. Letting $\xymatrix{\mathbf{IStr}\ar@/^/[r]^{\Comma_{\cat{R}}} & \cat{SR}_{1}\ar@/^/[l]^{\Space_{\cat{R}}}}$ be the functors defined in Definition \ref{comma-space-equivalence}, Theorem \ref{spaces-in-commas} gives an equivalence.
Setting all of the above categories and functors in line together yields the following progression from incidence structures as spaces, to incidence hypergraphs as objects of a comma category, to incidence hypergraphs as objects of a functor category.
\[\xymatrix{
\mathbf{IStr}\ar@/^/[rr]^{\Comma_{\cat{R}}}_{\simeq} & & \cat{SR}_{1}\ar@/^/[ll]^{\Space_{\cat{R}}}\ar@/^/[rr]^{\mathcal{N}_{\cat{R}_{1}}}_{\intercal} & & \cat{R}_{1}\ar@/^/[ll]^{\mathcal{S}_{\cat{R}_{1}}}\ar@/^/[rr]^{Z_{\cat{R}}}_{\cong} & & \cat{R}\ar@/^/[ll]^{W_{\cat{R}}}
}\]
Therefore, the following result is an immediate consequence.
\begin{thm}[Incidence structures \& hypergraphs]
The category $\mathbf{IStr}$ is equivalent to the reflective subcategory of simple incidence hypergraphs within $\cat{R}$. Moreover, $\mathbf{IStr}$ is complete and cocomplete with limits performed by passing to $\cat{R}$ and then applying the simplification.
\end{thm}
To reduce notation, let $\mathcal{N}_{\cat{R}}:=Z_{\cat{R}}\mathcal{N}_{\cat{R}_{1}}\Comma_{\cat{R}}$ and $\mathcal{S}_{\cat{R}}:=\Space_{\cat{R}}\mathcal{S}_{\cat{R}_{1}}W_{\cat{R}}$. Figure \ref{incidence-example} gives an example to illustrate that $\mathcal{S}_{\cat{R}}$ does indeed remove parallelisms in an incidence hypergraph. Please note that unlike $\mathcal{S}_{\cat{Q}}$ and $\mathcal{S}_{\cat{H}}$, which remove parallel edges, $\mathcal{S}_{\cat{R}}$ removes parallel incidences.
\begin{figure}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$G$ & $\mathcal{S}_{\cat{R}}(G)$\\\hline
\includegraphics[scale=1]{istr1.pdf}
&
\includegraphics[scale=1]{istr2.pdf}\\\hline
\end{tabular}
\end{center}
\caption{Simplification of an incidence hypergraph}
\label{incidence-example}
\end{figure}
\section{Simple Graphs \& Symmetric Digraphs}\label{simplegraphs}
Let $\mathbf{Gra}$ be the full subcategory of $\mathbf{SSys}$ consisting of all conventional graphs, and let $\xymatrix{\mathbf{Gra}\ar[r]^{N_{\mathbf{SSys}}} & \mathbf{SSys}}$ be the inclusion functor. The relationship between $\mathbf{Gra}$ and $\mathbf{SSys}$ is directly analogous to the deletion adjunction from \cite[Definition 2.32]{grilliette6} and can be directly replicated.
\begin{defn}[Simple deletion]
Given a set system $H$, define a graph $\Del_{\mathbf{SSys}}(H):=\left(V(H),\beta_{\Del_{\mathbf{SSys}}(H)}\right)$, where
\[
\beta_{\Del_{\mathbf{SSys}}(H)}:=\left\{A\in\beta_{H}:1\leq\card(A)\leq 2\right\}.
\]
Let $\xymatrix{\Del_{\mathbf{SSys}}(H)\ar[r]^(0.65){j_{H}} & H}\in\mathbf{SSys}$ be the canonical inclusion homomorphism from $\Del_{\mathbf{SSys}}(H)$ into $H$.
\end{defn}
\begin{thm}[Characterization of $\Del_{\mathbf{SSys}}$]
If $\xymatrix{N_{\mathbf{SSys}}(G)\ar[r]^(0.65){\phi} & H}\in\mathbf{SSys}$, there is a unique $\xymatrix{G\ar[r]^(0.3){\hat{\phi}} & \Del_{\mathbf{SSys}}(H)}\in\mathbf{Gra}$ such that $j_{H}\circ N_{\mathbf{SSys}}\left(\hat{\phi}\right)=\phi$. Thus, $\mathbf{Gra}$ is a coreflective subcategory of $\mathbf{SSys}$.
\end{thm}
Equivalently, one could repeat the work of Section \ref{Examples} to show that $\mathbf{Gra}$ is equivalent to the reflective subcategory of simple objects in the category $\cat{M}$ of multigraphs. The details of this argument are omitted for brevity.
On the other hand, \cite[p.\ 123]{hell1979} states that a simple graph can be viewed as a digraph $G$, where $(v,w)\in\alpha_{G}$ if and only if $(w,v)\in\alpha_{G}$, i.e.\ when $\alpha_{G}$ is a symmetric relation on $\vec{V}(G)$. Under this definition, simple graphs would be regarded as a subcategory of $\mathbf{Digra}$. The relationship is directly analogous to the associated digraph adjunction from \cite[Definitions 2.35 \& 2.36]{grilliette6}. Let $\mathbf{SymDigra}$ be the full subcategory of $\mathbf{Digra}$ consisting of all symmetric digraphs, and let $\xymatrix{\mathbf{SymDigra}\ar[rr]^{N_{\mathbf{Digra}}} & & \mathbf{Digra}}$ be the inclusion functor.
\begin{defn}[Symmetric closure]
Given a digraph $G$, define $N^{\diamond}_{\mathbf{Digra}}(G):=\left(\vec{V}(G),\alpha_{N^{\diamond}_{\mathbf{Digra}}(G)}\right)$, where $\alpha_{N^{\diamond}_{\mathbf{Digra}}(G)}:=\left\{(v,w):(v,w)\in\alpha_{G}\textrm{ or }(w,v)\in\alpha_{G}\right\}$. Let $\xymatrix{G\ar[r]^(0.3){\kappa_{G}} & N_{\mathbf{Digra}}N^{\diamond}_{\mathbf{Digra}}(G)}\in\mathbf{Digra}$ be the canonical inclusion homomorphism of $G$ into $N^{\diamond}_{\mathbf{Digra}}(G)$.
\end{defn}
\begin{thm}[Characterization of $N^{\diamond}_{\mathbf{Digra}}$]
If $\xymatrix{G\ar[r]^(0.3){\phi} & N_{\mathbf{Digra}}(H)}\in\mathbf{Digra}$, there is a unique $\xymatrix{N^{\diamond}_{\mathbf{Digra}}(G)\ar[r]^(0.65){\hat{\phi}} & H}\in\mathbf{SymDigra}$ such that $N_{\mathbf{Digra}}\left(\hat{\phi}\right)\circ\kappa_{G}=\phi$. Thus, $\mathbf{SymDigra}$ is a reflective subcategory of $\mathbf{Digra}$.
\end{thm}
However, $N_{\mathbf{Digra}}$ also admits a right adjoint by swapping the connective ``or'' to ``and''. As the symmetric closure is the smallest symmetric digraph containing a given digraph, this ``symmetric interior'' is the largest symmetric subdigraph within the given digraph. The universal property is routine to check.
\begin{defn}[Symmetric interior]
Given a digraph $G$, define $N^{\star}_{\mathbf{Digra}}(G):=\left(\vec{V}(G),\alpha_{N^{\star}_{\mathbf{Digra}}(G)}\right)$, where $\alpha_{N^{\star}_{\mathbf{Digra}}(G)}:=\left\{(v,w):(v,w)\in\alpha_{G}\textrm{ and }(w,v)\in\alpha_{G}\right\}$. Let $\xymatrix{N_{\mathbf{Digra}}N^{\star}_{\mathbf{Digra}}(G)\ar[r]^(0.8){\nu_{G}} & G}\in\mathbf{Digra}$ be the canonical inclusion homomorphism of $N^{\star}_{\mathbf{Digra}}(G)$ into $G$.
\end{defn}
\begin{thm}[Characterization of $N^{\star}_{\mathbf{Digra}}$]
If $\xymatrix{N_{\mathbf{Digra}}(G)\ar[r]^(0.65){\phi} & H}\in\mathbf{Digra}$, there is a unique $\xymatrix{G\ar[r]^(0.3){\tilde{\phi}} & N^{\star}_{\mathbf{Digra}}(H)}\in\mathbf{SymDigra}$ such that $\nu_{G}\circ N_{\mathbf{Digra}}\left(\tilde{\phi}\right)=\phi$. Thus, $\mathbf{SymDigra}$ is a coreflective subcategory of $\mathbf{Digra}$.
\end{thm}
To reconcile these two viewpoints, one defines the ``corresponding symmetric digraph'' of an undirected graph by following \cite[p.\ 123]{hell1979}, associating $(v,w)$ and $(w,v)$ to $\{v,w\}$. This construction yields a functor, which is an isomorphism between $\mathbf{SymDigra}$ and $\mathbf{Gra}$.
\begin{defn}[Isomorphism of $\mathbf{Gra}$ \& $\mathbf{SymDigra}$]
For $\xymatrix{G\ar[r]^{\phi} & G'}\in\mathbf{SymDigra}$, define $\xymatrix{W_{\mathbf{Gra}}(G)\ar[rr]^{W_{\mathbf{Gra}}(\phi)} & & W_{\mathbf{Gra}}\left(G'\right)}\in\mathbf{Gra}$ by
\begin{itemize}
\item $W_{\mathbf{Gra}}(G):=\left(\vec{V}(G),\beta_{W_{\mathbf{Gra}}(G)}\right)$, where $\beta_{W_{\mathbf{Gra}}(G)}:=\left\{\{v,w\}:(v,w)\in\alpha_{G}\right\}$;
\item $W_{\mathbf{Gra}}(\phi):=\phi$.
\end{itemize}
For $\xymatrix{G\ar[r]^{\phi} & G'}\in\mathbf{Gra}$, define $\xymatrix{Z_{\mathbf{Gra}}(G)\ar[rr]^{Z_{\mathbf{Gra}}(\phi)} & & Z_{\mathbf{Gra}}\left(G'\right)}\in\mathbf{SymDigra}$ by
\begin{itemize}
\item $Z_{\mathbf{Gra}}(G):=\left(V(G),\alpha_{Z_{\mathbf{Gra}}(G)}\right)$, where $\alpha_{Z_{\mathbf{Gra}}(G)}=\left\{(v,w):\{v,w\}\in\beta_{G}\right\}$;
\item $Z_{\mathbf{Gra}}(\phi):=\phi$.
\end{itemize}
Routine calculations show that $Z_{\mathbf{Gra}}W_{\mathbf{Gra}}=id_{\mathbf{SymDigra}}$ and $W_{\mathbf{Gra}}Z_{\mathbf{Gra}}=id_{\mathbf{Gra}}$.
\end{defn}
Consequently, the conventional category of simple graphs exists as a coreflective subcategory of $\mathbf{SSys}$, but also as a reflective and coreflective subcategory of $\mathbf{Digra}$. Therefore, both digraphs and set systems can be considered natural extensions of classical graph theory, as can general quivers and set-system hypergraphs. Interestingly, one can quickly show that $N_{\mathbf{SSys}}$ does not preserve products, so it cannot admit a left adjoint. The existence of both right and left adjoints to $N_{\mathbf{Digra}}$ seems to indicate that simple graphs are more closely related to digraphs than to set systems.
The diagram in Figure \ref{big-digram} seats these categories and functors together with those discussed in \cite{grilliette6}, giving a visual representation of these relationships.
\section{Cosimple Objects \& Cosimplification}\label{cosimplification}
All of the definitions and results of Section \ref{simplification} can be dualized, giving rise to ``cosimple'' objects and a ``cosimplification'' operation.
\begin{defn}[Cosimple object]
An object $(A,f,B)\in\ob(\cat{G})$ is \emph{cosimple} if $f$ is epic in $\cat{C}$. Let $\cat{TG}$ be the full subcategory of $\cat{G}$ consisting of cosimple objects, and let $\xymatrix{\cat{TG}\ar[r]^{\mathscr{N}} & \cat{G}}$ be the inclusion functor. A routine check shows that $\cat{TG}$ is a replete subcategory of $\cat{G}$.
\end{defn}
\begin{defn}[Cosimplification]
Assume that $\cat{B}$ is coregular, and that $G$ admits a left adjoint $G^{\diamond}$ with unit $\xymatrix{id_{\cat{C}}\ar[r]^{\delta} & GG^{\diamond}}$. For $(A,f,B)\in\ob(\cat{G})$, there is a unique $\xymatrix{G^{\diamond}F(A)\ar[r]^(0.6){\tilde{f}} & B}\in\cat{B}$ such that $G\left(\tilde{f}\right)\circ\delta_{F(A)}=f$. Let $\xymatrix{G^{\diamond}F(A)\ar@{->>}[r]^{e_{\tilde{f}}} & \coran\left(\tilde{f}\right)\textrm{ }\ar@{>->}[r]^(0.65){m_{\tilde{f}}} & B}\in\cat{B}$ be a coimage factorization of $\tilde{f}$. Define
\[
\mathcal{T}(A,f,B):=\left(A,G\left(e_{\tilde{f}}\right)\circ\delta_{F(A)},\coran\left(\tilde{f}\right)\right)
\]
and $\vartheta_{(A,f,B)}:=\left(id_{A},m_{\tilde{f}}\right)$.
\end{defn}
\[\xymatrix{
F(A)\ar[rr]^{F\left(id_{A}\right)}\ar[dd]_{\delta_{F(A)}} & & F(A)\ar[rr]^{F\left(id_{A}\right)}\ar[d]_{\delta_{F(A)}} & & F(A)\ar[dd]^{f}\\
& & GG^{\diamond}F(A)\ar[d]_{G\left(e_{\tilde{f}}\right)}\ar[drr]^{G\left(\tilde{f}\right)}\\
GG^{\diamond}F(A)\ar[rr]_(0.45){G\left(e_{\tilde{f}}\right)} & & G\left(\coran\left(\tilde{f}\right)\right)\ar[rr]_(0.6){G\left(m_{\tilde{f}}\right)} & & G(B)\\
}\]
\begin{thm}[Universal property of $\mathcal{T}$]
Let $\cat{B}$ be coregular, and assume that $G$ is faithful and admits a left adjoint with unit $\xymatrix{id_{\cat{C}}\ar[r]^{\delta} & GG^{\diamond}}$. If $\xymatrix{\mathscr{N}(A,f,B)\ar[rr]^{\left(\phi,\psi\right)} & & \left(A',f',B'\right)}\in\cat{G}$, there is a unique $\xymatrix{(A,f,B)\ar[rr]^{\left(\tilde{\phi},\tilde{\psi}\right)} & & \mathcal{T}\left(A',f',B'\right)}\in\cat{G}$ such that $\vartheta_{(A,f,B)}\circ\left(\tilde{\phi},\tilde{\psi}\right)=(\phi,\psi)$.
\end{thm}
\begin{cor}[Coreflective subcategory]\label{coreflective}
Let $\cat{B}$ be coregular, and assume that $G$ is faithful, preserves epimorphisms, and admits a left adjoint with an epic unit $\xymatrix{id_{\cat{C}}\ar[r]^{\delta} & GG^{\diamond}}$. If $\xymatrix{\mathscr{N}(A,f,B)\ar[rr]^{\left(\phi,\psi\right)} & & \left(A',f',B'\right)}\in\cat{G}$, there is a unique $\xymatrix{(A,f,B)\ar[rr]^{\left(\tilde{\phi},\tilde{\psi}\right)} & & \mathcal{T}\left(A',f',B'\right)}\in\cat{TG}$ such that $\vartheta_{(A,f,B)}\circ\mathscr{N}\left(\tilde{\phi},\tilde{\psi}\right)=(\phi,\psi)$. Consequently, $\cat{TG}$ is a coreflective subcategory of $\cat{G}$.
\end{cor}
As the proofs are merely the categorical duals of those found in Section \ref{simplification}, the details will be omitted in favor of examples. Note that $\mathbf{Set}$ is coregular, that the diagonal functor $\Delta$ admits both a left adjoint $\Delta^{\diamond}$ and a right adjoint $\Delta^{\star}$, and all three are faithful. Thus, the operator $\mathcal{T}$ can be constructed for both $\cat{Q}_{1}$ and $\cat{R}_{1}$. Unfortunately, the unit of each adjunction is not epic, so $\mathcal{T}$ does not guarantee a cosimple object in either category.
\begin{ex}[Cosimplification of quivers]
A cosimple object in $\cat{Q}_{1}$ would be a quiver, where there is directed edge for every ordered pair of vertices, i.e.\ a loaded quiver. Therefore, with the exception of the initial object, a quiver is cosimple in $\cat{Q}_{1}$ if and only if it is injective with respect to monomorphisms \cite[Proposition 3.2.1]{grilliette3}. The functor $\Delta^{\star}\Delta$ admits a left adjoint $\Delta^{\diamond}\Delta$ with unit $\delta_{X}:X\to\Delta^{\star}\Delta\Delta^{\diamond}\Delta(X)$ by $\delta_{X}(x):=\left((0,x),(1,x)\right)$, which is not epic. For $Q\in\ob\left(\cat{Q}_{1}\right)$, the output of $\mathcal{T}_{\cat{Q}_{1}}$ is
\begin{itemize}
\item $\vec{V}\mathcal{T}_{\cat{Q}_{1}}(Q)=\left\{\sigma_{Q}(e):e\in\vec{E}(Q)\right\}\cup\left\{\tau_{Q}(e):e\in\vec{E}(Q)\right\}$,
\item $\vec{E}\mathcal{T}_{\cat{Q}_{1}}(Q)=\vec{E}(Q)$,
\item $\vec{\epsilon}_{\mathcal{T}_{\cat{Q}_{1}}}(e)=\vec{\epsilon}_{Q}(e)$.
\end{itemize}
As seen in Figure \ref{quiver-example2}, $\mathcal{T}_{\cat{Q}_{1}}$ has removed all isolated vertices from the quiver. Moreover, $\mathcal{T}_{\cat{Q}_{1}}(Q)$ is not injective, so it is not cosimple in $\cat{Q}_{1}$.
\end{ex}
\begin{figure}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$Q$ & $\mathcal{T}_{\cat{Q}_{1}}(Q)$\\\hline
\includegraphics[scale=1]{quiver3.pdf}
&
\includegraphics[scale=1]{quiver4.pdf}\\\hline
\end{tabular}
\end{center}
\caption{Action of $\mathcal{T}_{\cat{Q}_{1}}$}
\label{quiver-example2}
\end{figure}
\begin{ex}[Cosimplification of incidence hypergraphs]
A cosimple object in $\cat{R}_{1}$ would be an incidence hypergraph, where there is an incidence for every vertex-edge pair. Therefore, with the exception of the initial object, an incidence hypergraph is cosimple in $\cat{R}_{1}$ if and only if it is injective with respect to monomorphisms \cite[Proposition 2.3.2]{grilliette6}. The functor $\Delta^{\star}$ admits a left adjoint $\Delta$ with unit $\delta_{X}:X\to\Delta^{\star}\Delta(X)$ by $\delta_{X}(x):=(x,x)$, which is not epic. For $G\in\ob(\cat{R})$, the output of $\mathcal{T}_{\cat{R}_{1}}$ is
\begin{itemize}
\item $\check{V}\mathcal{T}_{\cat{R}_{1}}(G)=\left\{\varsigma_{G}(i):i\in I(G)\right\}$,
\item $\check{E}\mathcal{T}_{\cat{R}_{1}}(G)=\left\{\omega_{G}(i):i\in I(G)\right\}$,
\item $I\mathcal{T}_{\cat{R}_{1}}(G)=I(G)$,
\item $\iota_{\mathcal{T}_{\cat{R}_{1}}}(i)=\iota_{G}(i)$.
\end{itemize}
As seen in Figure \ref{incidence-example2}, $\mathcal{T}_{\cat{R}_{1}}$ has removed all isolated vertices and loose edges from the incidence hypergraph. Moreover, $\mathcal{T}_{\cat{R}_{1}}(G)$ is not injective, so it is not cosimple in $\cat{R}_{1}$.
\end{ex}
\begin{figure}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$G$ & $\mathcal{T}_{\cat{R}_{1}}(G)$\\\hline
\includegraphics[scale=1]{istr3.pdf}
&
\includegraphics[scale=1]{istr4.pdf}\\\hline
\end{tabular}
\end{center}
\caption{Action of $\mathcal{T}_{\cat{R}_{1}}$}
\label{incidence-example2}
\end{figure}
\section{Set-System Hypergraphs with Antihomomorphisms}\label{antihomomorphisms}
Moreover, all of the definitions and results of Section \ref{structured} can be dualized, giving rise to the functor-costructured categories of \cite[Definition 21]{solovyov2013}. As the proofs are merely the categorical duals of those found in Section \ref{structured}, the details will be omitted in favor an illustrative example.
Let $\mathcal{Q}:\mathbf{Set}\to\mathbf{Set}^{\textrm{op}}$ be the contravariant power-set functor. Define $\cat{P}:=\left(\mathcal{Q}\downarrow id_{\mathbf{Set}^{\textrm{op}}}\right)$ with domain functor $\hat{V}:\cat{P}\to\mathbf{Set}$ and codomain functor $\hat{E}:\cat{P}\to\mathbf{Set}^{\textrm{op}}$. An object $G$ of $\cat{P}$ consists of two sets, $\hat{E}(G)$ and $\hat{V}(G)$, and a function $\xymatrix{\mathcal{Q}\hat{V}(G)\ar@{<-}[r]^(0.6){\epsilon_{G}} & \hat{E}(G)}$. Hence, $G$ is a set-system hypergraph as in $\cat{H}$.
However, the morphisms of $\cat{P}$ are quite different from those in $\cat{H}$. A homomorphism $\xymatrix{G\ar[r]^{\varphi} & G'}\in\cat{P}$ consists of two functions, $\hat{V}(\varphi):\hat{V}(G)\to\hat{V}\left(G'\right)$ and $\hat{E}(\varphi):\hat{E}\left(G'\right)\to\hat{E}(G)$, such that $\epsilon_{G}\circ\hat{E}(\varphi)=\mathcal{Q}\hat{V}(\varphi)\circ\epsilon_{G'}$.
\[\xymatrix{
\mathcal{Q}\hat{V}(G)\ar@{<-}[d]_{\epsilon_{G}}\ar@{<-}[rr]^{\mathcal{Q}\hat{V}(\varphi)} & & \mathcal{Q}\hat{V}\left(G'\right)\ar@{<-}[d]_{\epsilon_{G'}}\\
\hat{E}(G)\ar@{<-}[rr]_{\hat{E}(\varphi)} & & \hat{E}\left(G'\right)\\
}\]
Indeed, $\hat{E}(\varphi)$ maps edges of $G'$ to edges of $G$, rather than the reverse in $\cat{H}$. A morphism of $\cat{P}$ will be termed an \emph{antihomomorphism} of set-system hypergraphs for reasons that will be illuminated shortly.
Let $\cat{TP}$ be the full subcategory of $\cat{P}$ of cosimple objects, and $\xymatrix{\cat{TP}\ar[r]^{\mathscr{N}_{\cat{P}_{1}}} & \cat{P}}$ be the inclusion functor. By Corollary \ref{coreflective}, $\mathscr{N}_{\cat{P}_{1}}$ admits a right adjoint $\mathcal{T}_{\cat{P}_{1}}$ with the following action on objects:
\[
\mathcal{T}_{\cat{P}_{1}}\left(\hat{V}(G),\epsilon_{G},\hat{E}(G)\right)
=
\left(\hat{V}(G),e_{\epsilon_{G}},\coran\left(\epsilon_{G}\right)\right),
\]
where $\coran\left(\epsilon_{G}\right)=\left\{\epsilon_{G}(f):f\in\hat{E}(G)\right\}$ and $e_{\epsilon_{G}}(A)=A$. Thus, cosimplification in $\cat{P}$ removes parallelisms, like simplification in $\cat{H}$.
Let $\mathbf{Anti}:=\mathbf{Cospa}(\mathcal{Q})$. An object $G$ of $\mathbf{Anti}$ consists of a set $\hat{V}(G)$ and a family of subsets $\beta_{G}\subseteq\mathcal{Q}\hat{V}(G)$, a set system. A morphism $\xymatrix{G\ar[r]^{f} & G'}\in\mathbf{Anti}$ is a function $f:\hat{V}(G)\to\hat{V}\left(G'\right)$ such that $f^{-1}(A)\in\beta_{G}$ for all $A\in\beta_{G'}$, which is an ``antihomomorphism'' of set systems as defined in \cite[p.\ 60]{hn1979}. As such, the category $\mathbf{Top}$ of topological spaces and the category $\mathbf{Meas}$ of measurable spaces are both replete subcategories of $\mathbf{Anti}$.
Letting $\xymatrix{\mathbf{Anti}\ar@/^/[r]^(0.6){\CoComma_{\cat{P}}} & \cat{TP}\ar@/^/[l]^(0.4){\CoSpace_{\cat{P}}}}$ be the functors defined in the dual of Definition \ref{comma-space-equivalence}, the dual of Theorem \ref{spaces-in-commas} gives an equivalence. If $\mathcal{N}_{\mathbf{Top}}$ and $\mathcal{N}_{\mathbf{Meas}}$ are the inclusion functors of $\mathbf{Top}$ and $\mathbf{Meas}$ into $\mathbf{Anti}$, respectively, each admits a right adjoint given by generating a topology or $\sigma$-algebra from the set system \cite[Propositions 2.1 \& 4.9]{folland}. Let $\Gamma$ and $\Sigma$ be the functors from $\mathbf{Anti}$ to $\mathbf{Top}$ and $\mathbf{Meas}$, respectively, which generate the appropriate structure. Setting all of the above functors and categories in line together yields the following progression from topological and measurable spaces, to set systems as spaces, to set-system hypergraphs with antihomomorphisms. Moreover, the creation of the Borel $\sigma$-algebra can be represented with the composite functor $\Sigma\mathcal{N}_{\mathbf{Top}}$.
\[\xymatrix{
\mathbf{Top}\ar@/^/[drr]^{\mathcal{N}_{\mathbf{Top}}}\\
& & \mathbf{Anti}\ar@/^/[rr]^{\CoComma_{\cat{P}}}_{\simeq}\ar@/^/[dll]^{\Sigma}_{\perp}\ar@/^/[ull]^{\Gamma}_{\perp} & & \cat{TP}\ar@/^/[ll]^{\CoSpace_{\cat{P}}}\ar@/^/[rr]^{\mathscr{N}_{\cat{P}_{1}}}_{\perp} & & \cat{P}\ar@/^/[ll]^{\mathcal{T}_{\cat{P}_{1}}}\\
\mathbf{Meas}\ar@/^/[urr]^{\mathcal{N}_{\mathbf{Meas}}}
}\]
| 7e694d07fcfe562d6f6e871ed880dc4e0e1f34c7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Topological quantization has captivated a generation of physicists since the discovery of the quantum Hall effect \cite
QHE}. In the more recent years, its standing as a novel quantum phenomenon was further strengthened by the observation of discretely varying Hall conductances in the quantum anomalous Hall \cite{Haldane_QAHE,liu2016quantum} and quantum spin Hall effects \cite{KM_QSHE1}.
Rigorously established through linear response theory, the link between topological winding numbers and quantized conductivity has indeed earned a place as a classic result of quantum condensed matter physics. Indeed quantized Hall response and nontrivial Chern topology are now widely regarded to be almost synonymous.
Yet, the concept of quantized response has so far eluded classical systems. Without a quantum mechanical ground state, classical systems are not amendable to conventional linear response theory, which expresses quantized responses as perturbations upon a ground state. While classical metamaterials like photonic crystals, acoustic structures and electrical circuits can possess an integer number of topologically protected boundary modes~\cite{huber2016topological,nash2015topological,lu2014topological,imhof2018topolectrical}, their response behaviors are not
based on the number of accessible channels, but on analog solutions of differential equations that are by no means quantized.
In this work, we introduce the paradigm of quantized response based on the winding of the spectrum in the complex energy plane, rather than the winding of eigenstates in momentum space.
For a long time, the complex spectral winding number has been exploited in predicting the non-Hermitian pumping under the open boundary conditions (OBCs) [known as the non-Hermitian skin effect (NHSE)] \cite{yao2018edge,okuma2020topological,zhang2019correspondence,Lee2019anatomy}, leading to the breaking of bulk-boundary correspondence and various anomalous topological phenomena
\cite{xiong2018does,shen2018topological
kawabata2019symmetry,
yao2018edge
Lee2019anatomy,
kunst2018biorthogonal,Yao2018nonH2D,Yin2018nonHermitian,Hui2018nonH,li2019geometric,song2019realspace,song2019non,okuma2019topological,mu2020emergent,
jiang2019interplay,longhi2019topological,
Song2019BBC
Lee2019hybrid,li2020topological,
helbig2020generalized,xiao2020non,
okuma2020topological,zhang2019correspondence,
lee2020unraveling,li2020critical,yang2019auxiliary,yi2020non
}. However, to date no directly measurable quantity has been associated with the spectral winding. Notably, spectral winding as a topological feature is not directly related to quantum physics and is hence a \emph{classical} concept. Furthermore, the notion of classical response can be also a property of the Green's function alone. While nontrivial complex energy winding is in principle well-defined for quantum systems, it is most physically relevant in classical settings like mechanical, photonic, plasmonic and electrical systems where non-Hermiticity does not present significant measurement difficulties~\cite{ghatak2020observation,xiao2020non,helbig2020generalized}. Indeed, electrical circuits are governed by circuit Laplacians whose complex eigenvalues merely indicate phase shifts or steady-state impedances, rather than ephemeral excitations.
{Specifically, what we find quantized is the response of the logarithm of the Green's function components with respect to an imaginary flux-like local parameter that continuously adjust the system boundary conditions, forming some quantum-Hall-like plateaus quantized according to the spectral winding number. This discovery was inspired by the observation that deforming a non-Hermitian system from periodic to open boundaries (PBCs to OBCs)~\cite{xiong2018does,Lee2019anatomy,Lee2019hybrid,lee2020unraveling,mu2020emergent}, which is closely related to complex flux insertion, always reduces the spectral winding number one by one till it reaches zero.} In a classical setup subject to a steady-state drive
{e.g. a circuit lattice with an input current}, the quantized quantity \textcolor{black}{can be} the logarithmic impedance experienced by the response field, e.g. the voltage. This intriguing result is rooted in the way non-Bloch eigenstates explore the interior of spectral loops, and can \textcolor{black}{be understood with an unexpected type of non-Hermitian pumping that is scale-free}
~\cite{li2020impurity}.
\begin{comment}
\textcolor{red}{
3) Introduce the theory in the existing draft, and emphasize that $d ln|G_{xy}|/d\beta$ (or higher m generalizations) can indeed be quantized by the spectral winding in the non-herm setting. Explain the $\beta$ needs to be some sort of imaginary flux that can physically be implemented by tuning the boundary hoppings. Invoke the theory of the NHSE as when needed, but not as a main focus.
4) Describe a specific demonstration proposal in an electrical circuit, where the impedance is just $G_{xy}$, where $E_r$ is an offset corresponding to the admittance of identical ground connections. I suppose the measurement will be more complicated for $m>1$. But this should still be extremely practical to do. Simulated impedances will be shown to be quantized in the sense of Eq 7.
5) Discussion on scope and generalizations
}
\end{comment}
\begin{comment}
Topological phases of matter feature a variety of quantized physical responses
that manifest their underlying topological invariants~\cite{hasan2010colloquium,qi2011topological,bernevig2013topological,asboth2016short,shun2018topological}.
It is hence important to investigate the analogous physical implications of bulk topological invariants in non-Hermitian lattice systems
\cite{xiong2018does,shen2018topological,gong2018topological,ghatak2019new,kawabata2019symmetry,
yao2018edge
Lee2019anatomy,
kunst2018biorthogonal,Yao2018nonH2D,Yin2018nonHermitian,Hui2018nonH,li2019geometric,song2019realspace,song2019non,okuma2019topological,herviou2019defining,mu2020emergent,
jiang2019interplay,longhi2019topological,
Song2019BBC,borgnia2020nonH,zhang2020bulk,
edvardsson2019non,ezawa2019non,Luo2019HOnonH,Lee2019hybrid,li2020topological,
hofmann2020reciprocal,helbig2020generalized,xiao2020non,
okuma2020topological,zhang2019correspondence,
lee2020unraveling,li2020critical,yang2019auxiliary,yi2020nonH,NPReview}.
Of central interest here is the spectral winding on the complex plane for systems under the periodic boundary condition (PBC), a unique non-Hermitian topological feature. This type of topology can be characterized by a spectral winding number of the complex spectrum with respect to a reference energy point \cite{xiong2018does,shen2018topological,gong2018topological,ghatak2019new,kawabata2019symmetry}, also referred as point-gap topology.
As one significant result from ongoing studies of non-Hermitian topological systems, a nonzero spectral winding number of point-gap topology can be used to predict the emergence of the so-called non-Hermitian skin effect (NHSE) under open boundary conditions (OBC) \cite{yao2018edge,okuma2020topological,zhang2019correspondence}.
Nevertheless, the NHSE itself can only distinguish between nontrivial and trivial point-gap topology: the consequences of different nonzero spectral winding numbers as topological invariants have been elusive.
For systems under the semi-infinite boundary condition (SIBC), the analogous spectral winding number has been found to yield, remarkably, the degeneracy of eigenmodes at the reference energy \cite{okuma2020topological,pan2020point}. However, noting that a realistic lattice system {cannot be infinitely large,}
it remains to unveil how different nontrivial point-gap topology may be connected with the physical response of a realistic system.
In this work, we attack the open questions identified above by directly investigating the PBC-OBC spectral evolution as well as one physical response to this evolution. Given that the complex spectrum of a non-Hermitian system can be continuously deformed by tuning the boundary conditions from PBC to OBC \cite{xiong2018does,Lee2019anatomy,Lee2019hybrid,lee2020unraveling,mu2020emergent}, we advocate to look into the impact of different spectral winding behaviors on the complex PBC-OBC spectral evolution. Our key motivation is the following: because the system under OBC cannot have spectral winding (since the flux threading can be gauged away under OBC), the complex spectral evolution starting from regimes with different winding numbers should behave differently so as to to reach the same zero winding in the end. Indeed, the number of eigenstates flowing through a given reference energy point can be precisely given by the corresponding spectral winding number.
We have further considered a directional signal amplification scenario, i.e. a signal entering the system being amplified toward one nonreciprocal direction of the system. This amplification changes continuously as we gradually tune the boundary from PBC to OBC. Using a representative lattice model as an
example, we discover that the ratio of the change in the logarithm of a signal amplification factor to small variation in the boundary tuning parameter exhibits quantum-Hall-like plateaus, with the height of these plateaus proportional to the spectral winding numbers of the point-gap topology under PBC.
\end{comment}
\section{Results}
\subsection{Classical vs. Quantum response}
{In conventional literature, quantized response usually refers to quantized linear response in a quantum setting, where an occupied quantum state is driven by a time-dependent perturbation. In this work, however, our focus is to highlight that it is still possible to extract a quantized quantity from the response of a purely classical system, where there is no notion of an occupied ground state. In the following, we shall first clarify the distinction between quantum and classical response.}
{For concreteness, first consider the quantum setting described by a Hamiltonian $\hat H(t)=\hat H_0 + \epsilon(t)\hat H'$, where $\hat H_0$ is the equilibrium Hamiltonian, $\hat H'$ the perturbed operator and $\epsilon(t)$ a time-dependent factor that controls the perturbation strength. The linear response of a chosen operator $\hat \phi$ (which has zero expectation when $\epsilon(t)=0$) is its expectation $\hat \phi(t)$ after $\epsilon(t)$ is switched on. In frequency space, we write
\begin{equation}
\langle \hat \phi\rangle(\omega)=D(\omega)\epsilon(\omega),
\end{equation}
where
\begin{eqnarray}
D(\omega)& =& -i\int_0^\infty \langle[\hat \phi(t),\hat H'(0)]\rangle \,dt\notag\\
& =& -iZ^{-1}\int_0^\infty \text{Tr}\left(e^{-\beta H_0}[\hat \phi(t),\hat H'(0)]\right) \,dt\notag\\
&=& \frac1{Z}\sum_{nm}\frac{\langle n|\hat \phi|m\rangle\langle m|\hat H'|n\rangle (e^{-\beta E_n}-e^{-\beta E_m})}{\omega +i\delta -(E_m-E_n)}
\label{Kubo}
\end{eqnarray}
where $Z=\text{Tr}\left(e^{-\beta \hat H_0}\right)$, $\beta $ is the inverse temperature and the trace is taken over all eigenstates $|n\rangle$ satisfying $\hat H_0|n\rangle = E_n|n\rangle$. Very importantly, it has been assumed that the system is in a ground state where the eigenstates $|n\rangle$ are occupied according to the Boltzmann probability distribution $\propto e^{-\beta E_n}$.
{By expressing the trace as a momentum-space integral which shall correspond to a topological invariant,}
a topologically quantized response coefficient
can be obtained, for instance, with the DC limit of $D(\omega)/i\omega$ in the case of Hall response.
}
{
By contrast, in the classical settings that we shall focus on i.e. photonic waveguides, acoustic lattices and electronic circuit, the system do not settle into a ground state, and Eq.~\ref{Kubo} is inapplicable.
The classical response corresponds to the flow of arbitrary amounts of optical, phononic or electric current, and not the modified (quantized) occupancy of quantum mechanical eigenstates.} Consider an external coherent drive $\vec{\epsilon}=(\epsilon_1,\epsilon_2,...\epsilon_L)$ with different amplitudes applied to each of the $L$ sites of a lattice \cite{xue2020non}. For a {steady-state} drive with a fixed frequency $\omega$, $\vec{\epsilon}(t)=\vec{\epsilon}(\omega){\rm exp}(-i\omega t)$, and the resultant classical response field at the same frequency can be written as $\vec{\phi}(t)=\vec{\phi}(\omega){\rm exp}(-i\omega t)$ with the response field amplitude given by
\begin{eqnarray}
\vec{\phi}(\omega)=G(\omega,\gamma)\vec{\epsilon}(\omega),~G(\omega,\gamma)=\frac{1}{\omega+i\gamma-H},\label{eq:green}
\end{eqnarray}
analogous to Eq.~\ref{Kubo} which is exclusively for quantum settings. Here $G$ is the Green's function matrix and $\gamma$ represents an overall gain/loss in the system.
For a signal entering the system from a single site $x$, $\vec{\epsilon}$ only possesses one nonzero component $\epsilon_x$, and the induced field at another site $y$ is $\phi_{y}=G_{yx}\epsilon_x$. {In particular, the directional signal amplification of a signal entering one end of a 1D chain and measured at the other end is described by the two matrix elements $G_{1N}$ and $G_{N1}$ \cite{Wanjura2020,xue2020non}}.
\begin{comment}Given an excitation expressed as $|\epsilon(t)\rangle$ in the state space of the system, we define its \emph{classical} response $|A(t)\rangle$ as the time-averaged evolution of $|\epsilon(t)\rangle$:
\begin{eqnarray}
|A(t)\rangle&=& -i\int^t_{-\infty}e^{-i\hat H_\beta(t-t')}\epsilon(t')dt'\notag\\
&=& \int^\infty_{-\infty}G_\beta(t-t')\epsilon(t')dt'
\label{classicalres}
\end{eqnarray}
where $\beta$ is a tunable parameter in the Hamiltonian. In frequency space, Eq.~\ref{classicalres} takes the familiar form
\begin{equation}
|A(\omega)\rangle = G_\beta(\omega)|\epsilon(\omega)\rangle
\end{equation}
with $G_\beta(\omega) = (\omega+i\delta - \hat H_\beta)^{-1}$.
\end{comment}
\subsection{Motivation for quantized Green's function response}
{
In the spectral representation, the Green's function Eq.~\ref{eq:green} takes the form
\begin{eqnarray}
G=\sum_n\frac{1}{\omega+i\gamma-E_n}|\Psi_n^R\rangle\langle\Psi^L_n|,\label{eq:green_sr}
\label{G}
\end{eqnarray}
where $|\Psi^{L/R}_n\rangle$ are the left/right eigenstates corresponding to the $n$-th eigenenergy $E_n$. Its matrix elements $G_{xy}$
can be computed by evaluating $|\Psi^{L/R}_n\rangle$ at $x$ and $y$. Eq.~\ref{G} is valid in both classical and quantum settings, since the Hamiltonian is just the operator that describes time evolution, and is well-defined regardless of whether position and momentum commute. }
{Ordinarily, we do not expect the matrix elements of $G$ (or functions of them) to respond to an external influence $\beta$ in a quantized manner, since there is no reason why the derivatives of $(\omega-E_n)^{-1}$ and the eigenstates should conspire to add up to an integer. However, when translation symmetry is broken, the eigenstates can potentially become exponentially localized like $\sim e^{\kappa x}$, such that the matrix elements of $G$ are dominated by the largest spatial decay rate $\kappa=\kappa_\text{max}$, with $G_{xy} \sim e^{\kappa_\text{max}(y-x)}$. In such special scenarios, the response of $\ln G_{xy}$ for a fixed interval $x-y$ is \emph{wholly} dependent on how $\kappa_\text{max}$ varies with the external influence. In particular, if $\kappa_\text{max}$ were to vary with an external parameter in a quantized manner, so will the response quantity $\ln G_{xy}$. }
In this work, we discover that $\ln G_{xy}$ indeed possess such a quantized response if the external influence $\beta$ were to be
an impurity parameter tuning the boundary conditions, which coincides with tuning an imaginary flux when the latter is sufficiently weak \cite{Lee2019anatomy}.
This quantized quantity is furthermore equal to the winding number of the energy spectrum in the complex energy plane.
In the following sections, we shall elaborate on these rather surprising findings, and show the classical quantized response can be measured. While this quantization applies to both classical and quantum systems, we shall call it the quantized \emph{classical} response to distinguish it from the topological Hall response that exclusively exists in quantum systems.
\begin{comment}
For a given non-Hermitian Hamiltonian $H$, the Green's function $G=1/(\omega-H)$ describes the response of the system to a periodic driving at a frequency $\omega$. In the spectral representation, the Green's function reads
\begin{eqnarray}
G=\sum_n\frac{1}{\omega-E_n^R}|\Psi^R_n\rangle\langle\Psi^L_n|,
\end{eqnarray}
with $|\Psi^R_n\rangle$ the $n$-th right eigenstate of $H$ with eigenenergy $E_n^R$, and $\langle\Psi^L_n|$ the corresponding left eigenstate.
In a lattice system, the elements $G_{xy}$ describes the correlation between lattice sites at positions $x$ and $y$.
We are particular interested in the change of the Green's function as a response of tuning certain parameters of the system.
Under the PBC, the eigenstates always distribute evenly in the lattice system, thus the variation of the absolute value of $|G_{xy}|$ comes from only the changes of eigenenergies. In this scenario, it may be difficult to obtain a quantized response to a varying parameter, as the eigenenergies enter the Green's function as a part of the denominator. On the other hand, when the discrete translational symmetry is broken,
exponentially localized eigenstates may arise, generally taking the form of $\sim e^{-\kappa x}$ with $\kappa$ the decaying exponent, which usually has a linear dependence on certain system parameters $\beta$.
\\
Specifically in the situation considered in latter sections, the decaying exponent of the eigenstates takes the form of $\kappa\sim \beta/N$.
Thus the correlation between sites $x$ and $y$ reads $G_{xy}\sim e^{\kappa (y-x)}\sim e^{\beta(y-x)/N}$,
and we shall have $d \ln |G_{1N}|/d \beta = 1$ as a quantized response of the non-Hermitian point-gap topology of concerned systems.
\textcolor{red}{[MAY NEED TO MOVE THE FOLLOWING SENTENCES LATER] [Linhu: or just remove them?]}
\textcolor{blue}{As we shall elaborate later, by considering an off-diagonal block instead of a single element of the Green's function, the quantized response we consider is able to distinguish not only between nontrivial and trivial point-gap topology, but also between cases with different nonzero spectral winding numbers. }
\textcolor{blue}{
This Green's function can be used to depict system's response at $E_r=\omega+i\gamma$, leading to the hope that the winding number $\nu(E_r)$ at different reference energy points can be physically probed by scanning both the external frequency $\omega$ and the gain/loss parameter $\gamma$.
}
\end{comment}
\subsection{Point-gap topology and PBC-OBC spectral evolution}
{To elucidate the role of spectral winding and motivate a suitable notion of classical response, we consider a generic} one-band non-Hermitian system with $N$ lattice sites, described by the following tight-binding Hamiltonian
\begin{eqnarray}
H=\sum_{x=1}^N\sum_{j=-r}^{l} t_j\hat{c}^{\dagger}_j\hat{c}_{x+j},\label{eq:H_general}
\end{eqnarray}
with $t_j$ the hopping amplitude across $|j|$ lattice sites, $\hat{c}_x$ the annihilation operator of a (quasi-)particle at the $x$th lattice site,
$\hat{c}_{x+N}=\hat{c}_x$ representing the PBC,
and $r$ ($l$) the maximal range of the hopping toward right (left) direction.
The associated momentum-space Hamiltonian is given by
\begin{equation}
H(z)=\sum_{j=-r}^{l} t_j z^j={P_{r+l}(z)}/{z^r},
\end{equation}
with $k$ the quasi-momentum, $z:=e^{ik}$, and $P_{r+l}(z)$ a $(r+l)$th-order polynomial of $z$.
For any $t_j\neq t_{-j}^*$ (assuming $t_j=0$ if $j>l$ or $j<-r$), the Hamiltonian becomes non-Hermitian and possesses a point-gap topology, characterized by a nonzero spectral winding number w.r.t. a reference energy $E_r$ enclosed by the PBC spectrum~\cite{gong2018topological,okuma2020topological,zhang2019correspondence},
\begin{equation}
\nu(E_r)=\oint_\mathcal{C}\frac{dz}{2\pi }\frac{d}{dz}{\rm arg}\det[H(z)-E_r],\label{eq:winding}
\end{equation}
with $\mathcal{C}$ being the Brillouin zone (BZ), i.e. $k$ varying from $0$ to $2\pi$.
Simply put, $\nu(E_r)$ gives the number of times the PBC spectrum winds around $E_r$, as illustrated by a representative example in Fig.~\ref{fig:I_vs_beta}(a), corresponding to the Hamiltonian of Eq. \ref{eq:H_general} with $r=l=2$ .
As a side note, $\nu(E_r)$ also reflects the degeneracy of eigenmodes at $E_r$ when the system is placed under {semi-infinite boundary conditions} (SIBC) \textcolor{black}{\cite{okuma2020topological}}.
\begin{figure}
\includegraphics[width=1\linewidth]{fig2_6.pdf}
\caption{
(a) PBC (cyan loops) and OBC (gray dots) spectra of the Hamiltonian in Eq.~ (\ref{eq:H_general}) with $r=l=2$. The spectral winding number for each different regime is indicated on the panel.
(b) {The quantity $I_\beta(E_r)$ (Eq.~\ref{Ibeta}) vs $\beta$ for the three cases marked in (a). It diverges exactly $\nu$ times when $E_r$ sits in a region of spectral winding $\nu$.}
Parameters are set at $t_1=1$, $t_{-1}=0.5$, $t_2=2$, and $t_{-2}=0$, with $N=100$.
}
\label{fig:I_vs_beta}
\end{figure}
Unlike the loop-like PBC spectrum depicting a nontrivial point-gap topology, the OBC spectrum must not cover any finite area in the complex plane, and so generically \textcolor{black}{must take the form of curves within the PBC spectral loops} \cite{xiong2018does,Lee2019anatomy}, e.g. the Y-shape lines formed by the gray dots in Fig.~\ref{fig:I_vs_beta}(a).
That is, any reference energy $E_r$ inside a PBC spectral loop is enclosed $\nu(E_r)$ times by the PBC spectrum as $k$ varies from 0 to $2\pi$, but the same $E_r$ cannot be enclosed by the OBC spectrum. The important qualitative insight is hence the following: if we continuously deform the system from PBC to OBC, the evolving spectrum gradually collapses from the PBC loop spectrum to the OBC line spectrum, and is therefore expected to pass $E_r$ for $\nu(E_r)$ times.
To further appreciate this understanding, we consider the real-space Hamiltonian of Eq.~(\ref{eq:H_general}) with the following substitution only at the system's boundary, i.e.,
$t_j\rightarrow e^{-\beta} t_j$,
if $x+j>N$ or $x+j<1$. This introduces an additional scaling factor, or an impurity, to the boundary couplings of a 1D chain with $N$ unit cells.
If one now continuously varies $\beta$ from $0$ to infinity, a PBC-OBC spectral evolution can be examined in detail.
As shown in Fig.~\ref{fig:I_vs_beta}(a), by considering only $t_{\pm 1}$ and $t_{\pm 2}$, we already have a representative and intriguing model whose spectral winding number can be either 1 or 2 in the topologically nontrivial regime. Next we consider the tuning of the boundary coupling via $t_{\pm1}\rightarrow e^{-\beta} t_{\pm1}$ for $x=N$ and $t_{\pm2}\rightarrow e^{-\beta} t_{\pm2}$ for $x=N,N-1$.
Let the $n$-th right eigenvalue of this model system with such PBC-OBC interpolations be $E_n(\beta)$. Then the spectral evolution is all captured by $E_n(\beta)$ vs $\beta$.
{To motivate the connection with the Green's function, and to capture incidences when $E_n(\beta)$ comes close to $E_r$, we also define a quantity
\begin{equation}
I_\beta(E_r)=\sum_n |1/(E_n(\beta)-E_r)|,
\label{Ibeta}
\end{equation}}
i.e. the absolute sum of
the inverse energy spacings between the evolving spectrum $E_n(\beta)$ and a reference energy $E_r$.
For an actual system always of finite size, $E_n(\beta)$ is discretized, but it can still be made to be very close to the PBC reference eigenvalue $E_r$ for a sufficiently large system.
As such, the quantity $I_\beta(E_r)$ can be a diagnosis tool to examine how many
times $E_n(\beta)$ visits (the proximity of) $E_r$ as $\beta$ is tuned. Moreover, the OBC limit can be essentially reached once $\beta$ is beyond a critical value $\beta=\beta_{\rm OBC}\sim {N\eta}$ with $\eta$ the effective localizing length of the eigenstates \cite{koch2020bulk,li2020impurity}. With these understandings, one infers that as $\beta$ varies from $0$ to $\beta_{\rm OBC}$, $I_\beta(E_r)$ is expected to display high peaks whenever the complex spectral evolution passes through $E_r$. As explained above, the total number of such local peaks then reflects the spectral winding number $\nu(E_r)$.
In Fig. \ref{fig:I_vs_beta}(b) we illustrate $I_\beta(E_r)$ as a function of $\beta$ for several $E_r$
denoted by the stars of different colors in Fig.~\ref{fig:I_vs_beta}(a), corresponding to different spectral winding numbers $\nu(E_r)$. It is indeed observed that the number of peaks of $I_\beta(E_r)$ directly reflects the spectral winding number $\nu(E_r)$.
In the Methods section, we offer more insights based on the so-called generalized Brillouin zone (GBZ), to better understand
why $\nu(E_r)$ can be captured by the number of singularities encountered throughout the complex spectral evolution.
\subsection{Quantized response in signal amplification}
{While the previously defined quantity $I_\beta(E_r)$ is useful in diagnosing the spectral winding, it is not directly measurable. Below, we show how it inspires another analogous quantity that is directly associated with signal amplification.}
\textcolor{black}{In particular, the quantity introduced below displays quantized plateaus that precisely match spectral winding numbers, making it possible to distinguish between one nontrivial point-gap topology from another. This is a true advance as compared with earlier interesting attempts where signal amplification was only used to probe NHSE under OBC~\cite{Wanjura2020,xue2020non,yi2020nonH}. }
\begin{comment}
Consider an external coherent drive $\vec{\epsilon}=(\epsilon_1,\epsilon_2,...\epsilon_L)$ with different amplitudes applied to different lattice sites \cite{xue2020non}. For a drive with a fixed frequency, $\vec{\epsilon}(t)=\vec{\epsilon}(\omega){\rm exp}(-i\omega t)$,
the resultant response field at the same frequency can be written as $\vec{\phi}(t)=\vec{\phi}(\omega){\rm exp}(-i\omega t)$ with the response field amplitude given by
\begin{eqnarray}
\vec{\phi}(\omega)=G(\omega,\gamma)\vec{\epsilon}(\omega),~G(\omega,\gamma)=\frac{1}{\omega+i\gamma-H},\label{eq:green}
\end{eqnarray}
where $G$ is the Green's function matrix.
$\gamma$ represents an overall gain/loss of the system. This Green's function can be used to depict system's response at $E_r=\omega+i\gamma$, leading to the hope that the winding number $\nu(E_r)$ at different reference energy points can be physically probed by scanning both the external frequency $\omega$ and the gain/loss parameter $\gamma$.
For a signal entering from a single site $j$, i.e. $\vec{\epsilon}$ has only one nonzero component $\epsilon_j$, the induced field at site $j'$ is $\phi_{j'}=G_{j'j}\epsilon_j$.
Thus the directional signal amplification of a signal entering from one end and measured at the other end of a 1D chain is described by the two matrix elements of $G$, i.e. $G_{1N}$ and $G_{N1}$.
\end{comment}
For our system with the PBC-OBC interpolation parameter $\beta$ ($0 <\beta<\beta_{\rm OBC}$), what enters into the expression of the {Green's function $G$ in Eq. \ref{eq:green_sr}} is $1/[E_r-E_n(\beta)]$ again (as in $I_\beta(E_r)$) and hence the Green's function should be able to capture the complex spectral evolution. More importantly, it is found that the associated eigenmodes under PBC-OBC interpolation pile up at its boundary in a manner qualitatively different from that of NHSE \cite{li2020impurity}.
{With further derivation in the Method section, }
we find
\begin{eqnarray}
|G_{1N}|\propto e^{\beta},~\frac{d\ln |G_{1N}|}{d\beta}=1
\label{G1N}
\end{eqnarray}
for a one-dimensional system with only the nearest-neighbor couplings, provided $t_1>t_{-1}$ and $E_r$ falls within the loop-spectrum of the system at a given $\beta$ (See Supplemental Materials).
For $t_{-1}>t_{1}$, one can analogously obtain $\frac{d\ln |G_{N1}|}{d\beta}=1$, corresponding to a winding number $\nu(E_r)=-1$. {We emphasize that the variation in Eq.~\ref{G1N} can be directly measured in a steady-state response experiment. In an electrical circuit setting, this task can be done via impedance measurements, as will be elaborated later.}
For more general cases with couplings across $r$ ($l$) lattice sites to the right (left), the system can be effectively understood as $m={\rm Max}[r,l]$ different sub-chains, with $t_{\pm m}$ viewed as the nearest neighbor coupling on each sub-chain, and the rest understood as inter-sub-chain or intra-chain longer-range couplings. For example, lattice sites $j$, $j+m$, $j+2m$, $\cdots$, with each pair of neighbors coupled by $t_{\pm m}$, form the $j$-th sub-chain, for $j=0, 1, 2, \cdots, m-1$.
The above picture based on $m$ sub-chains becomes most useful if $t_{\pm m}$ is much larger than all other hopping parameters. Under this assumption these sub-chains are nearly decoupled from each other in the bulk, and their respective edges are connected through the couplings modified by $e^{-\beta}$.
This then suggests a unit-cell structure with $m$ sublattices, which greatly simplifies our consideration of signal amplification. Indeed, the overall amplification can be captured by the $m\times m$ block at the top-right (bottom-left) corners of the overall Green's function $G$, denoted as $G_{\leftarrow,m\times m}$ ($G_{\rightarrow,m\times m}$), corresponding to measuring the output at the first (last) $m$ sites of a signal entering from the last (first) $m$ sites.
Since each sub-chain yields an amplification factor proportional to $e^{\beta}$, one qualitatively expects that $|G_{\leftarrow,m\times m}|\equiv \det[G_{\leftarrow,m\times m}]\propto e^{m\beta}$ or $|G_{\rightarrow,m\times m}|\equiv \det[G_{\rightarrow,m\times m}]\propto e^{m\beta}$, depending on the amplification direction. This indicates that {
\begin{equation}
\nu_{\leftarrow,m}\equiv d\ln |G_{\leftarrow,m\times m}|/d\beta
\label{nuu}
\end{equation}
or analogously }$\nu_{\leftarrow,m}\equiv d\ln |G_{\leftarrow,m\times m}|/d\beta$ is quantized at $m$.
On the other hand, the PBC spectrum of the system winds $m$ times around the origin of the complex plane, a fact obviously true since the associated momentum-space Hamiltonian is dominated by the terms $e^{\pm i mk}$.
As theoretically shown in Supplementary Material \cite{suppmat}, this correspondence between quantized response and spectral winding indeed holds, if $E_r$ falls inside the PBC spectrum loop with the spectral winding number $\nu(E_r)$ given by $\pm m$, where the $\pm$ sign is to be determined by the actual direction of amplification.
Figure \ref{fig:Green} presents computational results for $\ln \left|G_{\leftarrow,2\times 2}\right|$ and its derivative with respect to $\beta$ as functions of $\beta$, denoted $\nu_{\leftarrow, 2}$, again for the same system as that in Fig. \ref{fig:I_vs_beta}. We also compare these results with $I(E_r)$ defined previously for a chosen reference energy point $E_r=\omega+i\gamma$, with $\nu(E_r)=2$.
It is observed that $\nu_{\leftarrow, 2}$ as a measurable physical response
shows three clear plateaus quantized at $\nu_{\leftarrow,2}=2,1,0$. Echoing with the jumps between these plateaus during the spectral evolution, $I_{\beta}(E_r)$ shows local peaks whenever the spectral evolution passes through $E_r$.
As shown in Fig.~\ref{fig:Green}(c),
these transitions during the spectral evolution as a result of increasing $\beta$ match precisely with the $\beta$ values for which the complex spectrum touches the reference energy point and hence the spectral winding number is about to jump. That is, the transitions between these quantized plateaus have a clear topological origin and are hence identified as topological transitions. As a side remark, observing all the plateaus in our example here happens to require a broad regime of $\beta$ and hence cases with very weak boundary coupling. This is however not a concern because the first plateau at small $\beta$ is the one to reflect the topology under PBC.
\begin{figure}
\includegraphics[width=1\linewidth]{plateau.pdf}
\caption{
(a) Amplification ratio $|G_{\leftarrow,2\times 2}|$ and (b) $\nu_{\leftarrow,2}$, the derivative of $\ln|G_{\leftarrow,2\times 2}|$ over $\beta$, as functions of $\beta$, along with $I_\beta(E_r)$, whose peaks match well with the jumps between different plateaus.
(c) Spectrum at different values of $\beta$, corresponding to the five dashed lines in (b), respectively.
Red stars indicate the reference energy $E_r=\omega+i\gamma$, with $\omega=-0.96$ and $\gamma=1$.
Insets zoom in on the regime around $E_r$ to give a clearer view of the relation between the shown spectrum and $E_r$.
Parameters are set at $t_1=1$, $t_{-1}=0.5$, $t_2=2$, $t_{-2}=0$, and $N=100$.
}
\label{fig:Green}
\end{figure}
The results presented in Fig.~\ref{fig:Green} are particularly stimulating. Indeed, therein neither the next-nearest-neighbor coupling nor the nearest-neighbor coupling is dominating. Yet quantized plateaus at $m=1$ and $m=2$ are still obtained. Returning to our decoupled sub-chain picture above, this indicates that for different reference energy points, the behavior of this system is topologically equivalent to that of one single chain or that of two weakly coupled sub-chains. With this perspective, we propose to examine $\nu_{\leftarrow,m}$ vs. different choices of $m$ in order to fully map out the phase boundaries, without any prior knowledge of spectral winding. We thus proceed to find the maximal $\nu_{\leftarrow,m}$ by scanning $m$, for sufficiently small $\beta$. The obtained value is then expected to yield $\nu(E_r)$ of the studied system under PBC. To this end we define $\nu_\leftarrow={\rm Max}[\nu_{\leftarrow,1},\nu_{\leftarrow,2}]$, given that our model system at most has effectively two sub-chains (one can similarly define $\nu_\rightarrow$).
We present in Fig.~\ref{fig:G_diagram}(a) our results of $\nu_\leftarrow$ at $\beta=0$ for different $\omega$ and $\gamma$ , with the phase boundaries identified there in excellent agreement with the actual PBC spectrum shown in Fig.~\ref{fig:I_vs_beta}(a). The only subtlety is that we also obtain a negative value
$\nu_\leftarrow=-1$ in the topological trivial regime (theoretically this result is also explained in Supplementary Material \cite{suppmat}). However, in these regimes both $\nu_\leftarrow$ and $\nu_\rightarrow$ are found to be negative, the amplification factor is far less than unity, and hence there is actually no signal amplification after all. One may just exploit this additional feature to locate regimes with zero spectral winding.
\begin{figure}
\includegraphics[width=1\linewidth]{diagram_impedance.pdf}
\caption{The topological phase diagram and simulation of impedance measurement in a circuit realization.
(a) The topological phase diagram, in excellent agreement with that shown in Fig.~\ref{fig:I_vs_beta} is mapped out from examining physical response functions $\nu_{\leftarrow,m}$ with $m=1$ or $m=2$, for the model depicted in the main text, with parameters set at $t_1=1$, $t_{-1}=0.5$, $t_2=2$, $t_{-2}=0$, and $N=100$.
At the phase boundaries, $E_r=\omega+i\gamma$ is an eigenvalue of the system, leading to the divergence of the Green's function matrix.
(b) Quantization of $\frac{d \ln| Z_{m}|}{d\beta}$, which is directly obtainable from simulated circuit measurement data of the impedance between the first $m$ and last $m$ nodes of the circuit lattice, here computed with $50$ nodes.
$Z_{m}$ (Eq.~\ref{Zm2} is the $m\times m$ matrix of impedances between the measured nodes, and $\beta$ is the parameter controlling the strengths of the couplings connecting the two ends.
\textcolor{black}{The complex admittance parameter $\Omega$ indicated on panel (b) is connected with parameters $\omega$ and $\gamma$ plotted in panel (a) with $\omega={\rm Re}(\Omega)$ and $\gamma={\rm Im}(\Omega)$.}
{The first plateau encountered when $\beta$ is increased from zero gives the {nonzero} topological winding $\nu$ (green, yellow for $\nu=1,2$). }
{When $\nu=0$, the logarithmic gradients of both $Z_{1,2}$ are close to or smaller than $0$ (blue), corresponding to the absence of signal amplification toward either direction.}
}
\label{fig:G_diagram}
\end{figure}
\subsection{Measurement of quantized response in electrical circuits}
We next elaborate on how the quantized response can be directly extracted by measuring the impedance in an electrical circuit setting. Instead of external perturbations, a circuit is most naturally driven by a steady-state AC or DC current, with voltage response given by Kirchhoff's law $\bold I = L\bold V$, where $L$ is the circuit Laplacian matrix and the components of $\bold I$ and $\bold V$ are respectively the input currents and voltages at each node.
Suppose that the circuit is then grounded by identical circuit components with complex admittances $-\Omega$. In this case, the full (grounded) Laplacian becomes $J=L-\Omega \,\mathbb{I}$, and the voltage distribution due to the input current are given by~\cite{lee2018topolectrical,lee2020imaging}
\begin{equation}
\bold V_i = [J^{-1}]_{ij}\bold I_j = G_{ij}\bold I_j=\left[\sum_n\frac{|\Psi^R_n\rangle\langle \Psi_n^L|}{\Omega-E_n}\right]_{ij}\bold I_j
\end{equation}
where the eigenvalues $E_n$ and L/R eigenstates $|\Psi^{L/R}_n\rangle$ are that of the Laplacian $J$. Notably, the quantity in the square parentheses agree exactly with our definition of the Green's function (Eq.~\ref{G}), with $\Omega$ taking the role of $\omega+i\gamma$. We extract the physical response through the impedance $Z_{ij}$ between the $i$-th and $j$-th nodes, which is related to the Green's function via $Z_{ij}=G_{ii}+G_{jj}-G_{ij}-G_{ji}$. By varying the identical grounding admittances $\Omega$ via a combination of RLC components with $\pm\pi/2$ relative phase shifts, we will be able to effectively access the response from different regions of the complex spectral plane.
Quantized classical response can be obtained from a circuit whose Laplacian exhibits nontrivial spectral winding. This requires effectively asymmetric couplings between nodes, which has already been demonstrated in existing topolectrical experiments through combinations of capacitors, inductors and INICs (negative impedance converter with current inversion) comprising operation amplifiers~\cite{hofmann2019chiral,helbig2020generalized,ezawa2019electric}, as further elaborated in the Supplementary Material. The boundary couplings can also be adjusted to effect the variation with $\beta$ through tunable inductors connected in series with the asymmetric couplings~\cite{li2020impurity}. Arbitrarily large spectral winding numbers can always be achieved by coupling sufficiently distant nodes, which can be much more feasibly done in electrical circuits compared to other platforms.
In analogy to $\nu_{\leftarrow,m}$ from Eq.~\ref{nuu}, we can define $Z_m=\text{det}\,Z_{ij}|_{i\leq m, N-j<m}$, the determinant of the $m\times m$ matrix of impedances between the first $m$ nodes at one end with the last $m$ at the other end of the circuit chain. Although it is not exactly equivalent to $\nu_{\leftarrow,m}$,
it is expected to vary with $\beta$ in a similar manner, since the impedance $Z_{ij}$ is dominated by the component of the Green's function that produces the directional amplification. Therefore,
by keeping track of the effective $\beta(\omega)$ and $\Omega(\omega)$, the logarithmic gradient of the impedance determinant $\frac{d \ln| Z_{m}|}{d\beta}$ will be expected to exhibit quantized jumps as $\Omega(\omega)$ crosses the boundary between regions of different topological winding $\nu$. For $m\leqslant 2$, which includes the model we had considered (Figs.~1, 2 and 3), we explicitly have
\begin{equation}
Z_{m=2}=Z_{1,N-1}Z_{2,N}-Z_{1,N}Z_{2,N-1},
\label{Zm2}
\end{equation}
whose gradients are dominated by those of terms like $G_{1,N-1}G_{2,N}$ and $G_{1,N}G_{2,N-1}$ in the presence of directional amplification (toward the first lattice site). As demonstrated via the simulated measurements in Fig.~\ref{fig:G_diagram}(b), the gradient $\frac{d \ln| Z_{m=1,2}|}{d\beta}$ indeed exhibits plateaus quantized at the winding number $\nu$ (blue, green, yellow for $\nu=0,1,2$ respectively) where $\Omega$ is tuned to. For higher topological winding, we take the plateau that first occurs when $\beta$ is increased from $0$, i.e. the plateau closest to periodic boundary conditions.
\begin{comment}
To read $\nu$, we take the plateau that first occurs when $\beta$ is increased from $0$, i.e. the plateau closest to periodic boundary conditions.
\textcolor{red}{[NOTE: I have chosen some other $\Omega$, just to show that the impedance measurement also works. I don't quite understand the above statement. In Fig. \ref{circuit}, the winding numbers of some $\Omega$ do not correspond to the plateaus. E.g. the PBC winding number is $2$ for $\Omega=-1.75+i$, and $1$ for $\Omega=3.3$. For the first case, the first transition point is too close to $\beta=0$, so it is difficult to see it. For the second case, we will see the quantization at $1$ if we consider $Z_{m=1}$.]}
\begin{figure}
\includegraphics[width=0.99\linewidth]{Imp_plot.pdf}
\caption{Quantization of $\frac{d \ln| Z_{m=2}|}{d\beta}$, which is directly obtainable from impedance measurements between the first $m$ and last $m$ nodes. Show are data from simulated measurements from the circuit corresponding to the model in Figs 1-3 with $t_1=1$,$t_{-1}=0.5$ and $t_2=2$. When $\beta$ is increased from $0$, the curves first reaches plateaus quantized exactly at $\nu$, the PBC winding number of the model at the indicated $\Omega$ values. \textcolor{red}{Consider combining with Fig 3, I used matching colors}
}
\label{circuit}
\end{figure}
\end{comment}
\section{Conclusions}
{In this work, we have introduced the new paradigm of quantized classical response, where a quantized response coefficient can always be extracted from how the Green's function \textcolor{black}{varies} with the evolution of an imaginary flux-like parameter.
Being based on the topological winding properties of the Green's function in the complex energy plane, \textcolor{black}{this quantization} does not assume the existence of any quantum mechanical ground state, and applies to all systems, classical and quantum.
Specifically, we show that the spectral winding number is directly detectable as a steady-state response coefficient to changes in the boundary condition. \textcolor{black}{Such correspondence between spectral winding numbers and quantized response is arguably broader in scope than in the case of momentum-space topology, because spectral winding does not even require translational invariance.} \textcolor{black}{Our results are relevant to a number of current experimental platforms of non-Hermitian systems \cite{hofmann2020reciprocal, helbig2020generalized,xiao2020non,Wanjura2020,li2020critical,NPReview}.}
In the context of classical electrical circuits, we have shown that a quantized response can be easily extracted from extremely experimentally accessible impedance measurements.
}
\begin{acknowledgements}
{\it Acknowledgement.-}
We would like to thank Da-Jian Zhang for helpful discussions.
L. L. acknowledges funding support by
the Key-Area Research and Development Program of GuangDong Province under Grants No. 2019B030330001
and the Startup Grant of Sun Yat-sen University (No. 71000-18841245).
J.G. acknowledges funding support by the Singapore NRF
Grant No. NRF-NRFI2017-04 (WBS No. R-144-000-378-281). C.H.L acknowledges funding support by the Singapore MOE Tier-1 start-up grant (WBS R-144-000-435-133).
\end{acknowledgements}
\section{Methods}
\subsection{Insights based on generalized Brillouin zone}
\textcolor{black}{Here} we offer more insights based on the so-called generalized Brillouin zone (GBZ), to better understand
why $\nu(E_r)$ can be captured by the complex spectral evolution.
According to the non-Bloch band theory, the OBC spectrum can be described by the PBC one in a GBZ, using a complex deformation of the quasi-momentum $k\rightarrow k+i\kappa_{\rm OBC}(k)$ \cite{yao2018edge,Lee2019anatomy,lee2018tidal,lee2020unraveling,yokomizo2019non}.
The PBC-OBC spectral evolution can then be effectively described by $k\rightarrow k+i\kappa(k)$ with $\kappa(k)$ varying from $0$ to $\kappa_{\rm OBC}(k)$, with $\kappa_{\rm OBC}(k)$ having the minimal magnitude to yield the OBC spectrum \cite{yokomizo2019non,Lee2019anatomy}. The PBC-OBC spectral evolution can hence be understood as arising from tuning $\kappa(k)$ and hence deforming the BZ to the GBZ, as shown in Fig.~\ref{fig:GBZ}.
Moreover, an winding number can be defined as,
\begin{equation}
\nu(E_r)=\oint_{\rm GBZ}\frac{dz}{2\pi }\frac{d}{dz}{\rm arg}\det[H(z)-E_r],\label{eq:winding_GBZ}
\end{equation}
analogous to that of Eq.~(\ref{eq:winding}), with integration in the GBZ instead of the BZ.
This winding number must be zero when the OBC spectrum is reached because, again, the OBC spectrum cannot enclose any finite area \cite{okuma2020topological,zhang2019correspondence}.
Following the Cauchy principle, the spectral winding number is found to be
$\nu(E_r)=N_{\rm zero}-N_{\rm pole}$,
where $N_{\rm zero}$ and $N_{\rm pole}$ are the counting of zeros and poles enclosed by the integration path (BZ or GBZ) weighted by their respective orders.
The conclusion is hence as simple as follows. If we continuously tune $\kappa(k)$, the PBC-OBC spectral evolution must pass through different zeros of $[\frac{P_{r+l}(z)}{z^r}-E_r]$ [colored dots in Fig.~\ref{fig:GBZ}] for a total of $\nu(E_r)$ times, such that the spectral winding number reduces from $\nu(E_r)$ to $0$ eventually when the integration path approaches the GBZ.
Thus, during the complex spectral evolution, the spectrum under $k\rightarrow k+i\kappa(k)$ must pass the reference energy $E_r$ for a total of $\nu(E_r)$ times, constituting a rather formal argument to justify our treatment in the main text.
\begin{figure}
\includegraphics[width=0.7\linewidth]{GBZ.pdf}
\caption{A typical example of the Brillouin zone (BZ) and the generalized Brillouin zone (GBZ).
The BZ and GBZ are given by cyan and gray dots respectively, in the complex plane of $z:=e^{ik}e^{-\kappa(k)}$. Here $\kappa(k)$ represents a complex deformation of the momentum $k$.
The black dash loop indicates an evolving GBZ with $\kappa(k)$ between $0$ and $\kappa_{\rm OBC}(k)$, the value that gives the spectrum under the open boundary condition. Blue, red, and yellow dots are the zeros of $H(z)-E_r$, where $H(z)$ is the PBC Hamiltonian of the system and $E_r$ is the chosen reference energy for calculating the winding number. The system is chosen as the same as that in Fig. \ref{fig:I_vs_beta} in the main text,
i.e. $H(z)=2z^2+z+1/2z$.
}
\label{fig:GBZ}
\end{figure}
\begin{comment}
\subsection{Derivation of the Hamiltonian from the master equation}
\textcolor{red}{to be revised and completed}
\textcolor{blue}{Having establishing the correspondence between the spectral winding number and the PBC-OBC evolution,
we note look at the directional amplification in 1D bosonic lattice systems \textcolor{red}{[cite a few experimental papers here?]}, which has been used to probe NHSE under OBC, thus recognized as a tool to distinguish between trivial and nontrivial point-gap topology~\cite{Wanjura2020,xue2020non}.
Our approach below further makes it possible to distinguish between nontrivial point-gap topology with different spectral winding numbers.
To manifest the Hamiltonian system in Fig. \ref{fig:I_vs_beta} (i.e. Eq. \ref{eq:H_general} with $r=l=2$), we start from a Hermitian system with the nearest and next-nearest couplings, described by the Hamiltonian
\begin{eqnarray}
H_0&=&\sum_x [(t_1^0\hat{a}_x\hat{a}_{x+1}+t_2^0\hat{a}_x\hat{a}_{x+2}+h.c.)\nonumber\\
&&+\omega_0\hat{a}_x^\dagger \hat{a}_x+\epsilon_x\hat{a}^\dagger_x+\epsilon^*_x\hat{a}_x],
\end{eqnarray}
with $\hat{a}_x$ the annihilation operator of a bosonic mode at site $x$ \cite{xue2020non}.
The bosonic modes are assumed to have the same bare frequency $\omega_0$, and are coupled with each other through the couplings $t_1^0$ and $t_2^0$.
The time evolution of the system's density matrix $\rho$ follow the master equation
\begin{eqnarray}
\dot{\rho}=-i[H_0,\rho]+\sum_\mu\left(L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}\right),
\end{eqnarray}
where $\{L_{\mu}\}=\{L_x^g,L_x^l\}$ is the dissipators describing the coupling of the system to the environment, and $H_0$ is the Hamiltonian of the system with an external coherent drive $\vec{\epsilon}=(\epsilon_1,\epsilon_2,...\epsilon_L)$ as an input signal to be amplified/
}
\textcolor{blue}{
Consider a drive with a fixed frequency, $\vec{\epsilon}(t)=\vec{\epsilon}(\omega){\rm exp}(-i\omega t)$,
the resultant response field at the same frequency can be written as $\vec{\phi}(t)=\vec{\phi}(\omega){\rm exp}(-i\omega t)$ with the response field amplitude given by
\begin{eqnarray}
\vec{\phi}(\omega)=G(\omega,\gamma)\vec{\epsilon}(\omega),~G(\omega,\gamma)=\frac{1}{\omega-H_{\gamma}},\label{eq:green}
\end{eqnarray}
where $G$ is the Green's function matrix, and $H_{\gamma}=H-i\gamma$ an effective non-Hermitian Hamiltonian describing the evolution of $\vec{\phi}(t)$ \cite{xue2020non}. To simulate the concerned system, we requires $H_{\gamma}=H-i\gamma$, i.e. the Hamiltonian of Eq. \ref{eq:H_general} at $r=l=2$ with the spectral winding number up to $2$, and an extra overall gain/loss of $i\gamma$ to tune the reference energy $E_r$.
With detailed in Supplemental Materials, we find that this can be done by properly choosing dissipators $\{L_{\mu}\}=\{L_x^g,L_x^l\}$ connecting nearest and next-nearest neighboring sites, and considering a shift of the measured frequency $\omega\rightarrow\omega-\omega_0$. In this way,}
this Green's function can be used to depict system's response at $E_r=\omega+i\gamma$, leading to the hope that the winding number $\nu(E_r)$ at different reference energy points can be physically probed by scanning both the external frequency $\omega$ and the gain/loss parameter $\gamma$.
For a signal entering from a single site $j$, i.e. $\vec{\epsilon}$ has only one nonzero component $\epsilon_j$, the induced field at site $j'$ is $\phi_{j'}=G_{j'j}\epsilon_j$.
Thus the directional signal amplification of a signal entering from one end and measured at the other end of a 1D chain is described by the two matrix elements of $G$, i.e. $G_{1N}$ and $G_{N1}$.
\end{comment}
\subsection{Directional signal amplification versus the PBC-OBC spectral evolution}
\subsubsection{Cases with only nearest-neighbor coupling}
As mentioned in the main text, under an external drive $\vec{\epsilon}(t)=\vec{\epsilon}(\omega){\rm exp}(-i\omega t)$ and an overall on-site gain/loss parameter $\gamma$, the resultant response field $\vec{\phi}(t)$ can be written as $\vec{\phi}(t)=\vec{\phi}(\omega){\rm exp}(-i\omega t)$, with
\begin{eqnarray}
\vec{\phi}(\omega)=G(\omega,\gamma)\vec{\epsilon}(\omega),~G(\omega,\gamma)=\frac{1}{E_r-H},
\end{eqnarray}
where $E_r=\omega+i\gamma$, $G$ is the Green's function matrix.
The amplification factors for a signal toward the left and the right are described by the matrix elements $G_{1N}$ and $G_{N1}$, respectively.
For a non-Hermitian $H$, the Green's function matrix can be expressed in the spectral representation \cite{zirnstein2019bulk,xue2020non}
\begin{eqnarray}
G(\omega,\gamma)=\frac{1}{E_r-H}=\sum_n\frac{1}{E_r-E_n^R}|\Psi^R_n\rangle\langle\Psi^L_n|,
\end{eqnarray}
with $|\Psi^R_n\rangle$ the $n$th right eigenstate of $H$ with eigenenergy $E_n^R$, and $\langle\Psi^L_n|$ the corresponding left eigenstate.
To be more explicit, consider the Hatano-Nelson model under the PBC-OBC interpolation, described by the following Hamiltonian
\begin{equation}
H_\beta = \sum_{x=1}^{N-1}(t_1\hat{c}^{\dagger}_{x}\hat{c}_{x+1}+t_{-1}\hat{c}^{\dagger}_{x+1}\hat{c}_x)
+e^{-\beta}(t_{1}\hat{c}^{\dagger}_{N}\hat{c}_{1}+t_{-1}\hat{c}^{\dagger}_{1}\hat{c}_N),
\label{hn_model}
\end{equation}
also with the assumption $t_1>t_{-1}$ without loss of generality.
Let the $n$-th right eigenstate be $|\Psi_n^R\rangle=\sum_{x=1}^{\infty}\psi_{x,n}^R\hat{c}^{\dagger}_{x}|0\rangle$, with $|0\rangle$ the vacuum state. Using the eigenvalue-eigenstate equation $H_\beta |\Psi_n^R\rangle=E_n^R|\Psi_n^R\rangle$, one obtains
\begin{eqnarray}
t_{1} \psi^R_{x+1,n}+t_{-1} \psi^R_{x-1,n}=E_n^R\psi^R_{x,n}~\label{eq:bulk_HN}
\end{eqnarray} for $x\in[2,N-1]$,
and
\begin{eqnarray}
e^{-\beta}t_{1} \psi^R_{1,n}+t_{-1} \psi^R_{N-1,n}&=&E_n^R\psi^R_{N,n},\label{eq:boundary1_HN}\\
t_{1} \psi^R_{2,n}+e^{-\beta}t_{-1} \psi^R_{N,n}&=&E_n^R\psi^R_{1,n}.\label{eq:boundary2_HN}
\end{eqnarray}
If we further assume $t_1\gg t_{-1}$ and the following exponentially decaying eigensolutions:
\begin{eqnarray}
\psi_{x,n}^R=C_n e^{-M_n^R (x-1)}
\end{eqnarray}
with $C_n$ the normalization constant and $M_n^R>0$, then one obtains
\begin{eqnarray}
t_{1} e^{-M_n^R}=E_n^R, M_n^R=\frac{\beta-i2n\pi}{N}.
\label{loopR}
\end{eqnarray}
Note that here $|E_n^{R}/t_1|$ is determined by the ratio $\beta/N$. That is, the eigenvalues $E_n^{R}$ will be distributed on a circle on the complex plane, whose radius depends only on $t_1$ as well as the ratio $\beta/N$.
Likewise, the left eigenstates under the same assumptions satisfy
$H^\dagger_\beta|\Psi_{n}^L\rangle=E_n^L|\Psi_{n}^L\rangle$ and $E_n^L=(E_n^R)^*$,
and they are found to be
\begin{eqnarray}
\psi_{x,n}^L&=&C_n^* e^{-M_n^L (N-x)}, t_1 e^{-M_n^L} =E_n^L, \nonumber\\
M_n^L&=&\frac{\beta+i2n\pi}{N}.
\end{eqnarray}
From the biothorgonal condition $\langle \Psi_{n}^L|\Psi_{n}^R\rangle=1$, we then obtain the normalization constant
\begin{eqnarray}
C_n=\frac{e^{-i\frac{N-1}{N}\pi n}}{\sqrt{Ne^{-\beta(N-1)/N}}}.
\end{eqnarray}
\begin{figure*}
\includegraphics[width=1\linewidth]{G_NN.pdf}
\caption{(a) Spectra under PBC (cyan loop), OBC (gray dots), and (black circles) the PBC-OBC interpolation through $\beta$. The red star indicates the reference energy $E_r=\omega+i\gamma$ for further results in (b) and (c). Here $E_r$ falls right on the spectrum of $H_\beta$, as the value of $\beta=\beta_c\approx 6.2$ is read out from (b) and (c).
(b) and (c) show the amplification ratios $|G_{1N}|$ and $|G_{N1}|$ [(b2) and (c2)] and the derivatives of their logarithms over $\beta$ [(b3) and (c3)] as functions of $\beta$, where the quantity $I_\beta(E_r)$ indicates with its peaks when $E_r$ is passed through by the $H_\beta$ spectrum.
The parameters are $t_1=1$ and $t_{-1}=0.5$.
}
\label{fig:G_NN}
\end{figure*}
With preparations above the matrix element of the Green's function $G$ can then be found as follows:
\begin{eqnarray}
G_{1N}&=&\sum_n\frac{1}{E_r-E_n^R}\psi_{1,n}^R(\psi_{N,n}^L)^*\nonumber\\
&=&\sum_n\frac{1}{E_r-t_1 e^{-\frac{\beta-i2n\pi}{N}}}
\frac{1}{N (e^{-\frac{\beta-i2n\pi}{N}})^{N-1}}\label{eq:G1N_1}\\
&=&\sum_n\frac{e^{\beta}e^{-\frac{\beta-i2n\pi}{N}}}{E_r-t_1 e^{-\frac{\beta-i2n\pi}{N}}}
\frac{1}{N}\label{eq:G1N_2}.
\end{eqnarray}
We may now attempt to rewrite the discrete sum in Eq.~(\ref{eq:G1N_2}) in terms of a loop integral of a complex variable $z$, by defining $k_n:=2n\pi/N$ and $z:=e^{-\beta/N}e^{ik_n}$.
To that end one must implicitly assume that $N$ under consideration is sufficiently large so as to use an integral to replace the discrete sum. With this in mind, for a given $\beta$, $\beta/N$ is assumed to be vanishingly small, and hence essentially we are working in the regime of $|z|\rightarrow 1$. Under these conditions, the sum in Eq.~(\ref{eq:G1N_2}) can then be evaluated by the following integral
\begin{eqnarray}
G_{1N}=\oint_{|z|= e^{-\beta/N}} \frac{1}{2\pi i } \frac{e^\beta}{E_r-t_{1} z}dz,
\end{eqnarray}
which is found to be
\begin{eqnarray}
G_{1N} =- \frac{e^{\beta}}{t_1 }, \label{eq:G1N}
\end{eqnarray}
if $z_0 \equiv E_r/t_1$ satisfies $|z_0| <e^{-\beta/N}$, i.e. the pole of the integrand falls within the the integral loop.
This condition leads to
\begin{eqnarray}
|E_r|<t_{1} e^{-\beta/N}=|E^R_n|,\label{eq:Er_enclosed}
\end{eqnarray}
meaning that the reference energy $E_r$ falls within the loop-spectrum of $H_\beta$.
The above detailed theoretical considerations indicate that, so long as $E_r$ is enclosed by the loop spectrum of $H_\beta$, we have
\begin{eqnarray}
\frac{d \ln |G_{1N}|}{d \beta}=1,
\end{eqnarray}
which is just the claim in the main text regarding how to use a quantized physical response to detect the spectral winding number $\nu(E_r)$, as computationally verified in Fig.~\ref{fig:G_NN}(b).
Next we investigate what happens if we let $\beta$ exceed $\beta_c$, the critical $\beta$ value for which a reference energy point $E_r$ falls exactly on the loop spectrum of $H_\beta$ (as shown in Fig.~\ref{fig:G_NN}(a)). Let us first recall the result from Eq.~(\ref{loopR}), which indicates that the radius of the loop spectrum of $H_\beta$ scales with $\beta/N$. Specifically, for a given lattice size $N$ and a given reference point $E_r$ under investigation, one immediately obtains that
\begin{eqnarray}
\beta_c =- N \ln |E_r/t_1|,
\label{betac}
\end{eqnarray} which is clearly proportional to $N$. As such, to probe the regime of $\beta\geq \beta_c$ , $\beta$ should at least linearly increase with $N$ as well. This being the case, we can no longer approximate the discrete sum in Eq.~(\ref{eq:G1N_2}) as a loop integral with $N\rightarrow \infty$ in mind, simply because it has the factor $e^{\beta}$ in its numerator, which diverges with $N$ because in the regime of interest $\beta$ diverges with an increasing $N$.
Based on the discussions above, in the regime of $\beta\geq \beta_c$, what is under investigation becomes the evaluation of the same discrete sum, but with $\beta$ scaling proportionally with $N$, and hence both becoming sufficiently large in the event of using a loop integral to replace this discrete sum. Though we still use the same complex variable $z=e^{-\beta/N}e^{ik_n}$ to invoke a possible loop integral, we see that $|z|$ is now far from unity. To reflect this, we now use the alternative expression of the discrete sum in Eq.~(\ref{eq:G1N_1}) and then rewriting it as the following integral, with the integrand having {one pole of order $N$ at $z=0$, and another $1$st-order pole at $z=E_r/t_1$, }
\begin{eqnarray}
G_{1N}&=&\oint_{|z|=e^{-\beta/N}} \frac{1}{2\pi i } \frac{1}{z^N(E_r-t_{1} z)}dz.
\label{exp2}
\end{eqnarray}
Assuming that $|z_0|=|E_r/t_1|>e^{-\beta/N}$, namely, the reference energy $E_r$ is outside the loop, so $z=0$ is the only pole of the integrand {enclosed by the integration path}, we have
\begin{eqnarray}
G_{1N}= \frac{t_1^{N-1}}{E_r^N}, \label{eq:G1N_3}
\end{eqnarray}
a value independent of $\beta$, which is again consistent with the computational results in Fig.~\ref{fig:G_NN}(b2).
Further, using the previous expression of $\beta_c$ from Eq.~(\ref{betac}), we find that in this case $G_{1N}=\frac{t_1^{N-1}}{E_r^N}=e^{\beta_c}/t_1$. Interestingly, though this magnitude $e^{\beta_c}/t_1$ exponentially larger than $e^{\beta}/t_1$ obtained earlier for $\beta<\beta_c$, this amplification factor is saturated and no longer depends on $\beta$ in the regime of $\beta>\beta_c$.
As a side note, one might wonder why we cannot also use the loop integral in Eq.~(\ref{exp2}) to treat the first case, namely,
a fixed $\beta$ in the regime of $\beta<\beta_c$ but with $N$ approaching sufficiently large values. As said earlier, in this case we essentially perform the summation under the condition of $|z|=1$. Under this condition we always have $z^N=1$ and hence the expression in Eq.~(\ref{exp2}) is no longer useful.
\begin{comment}
Fixed $\beta$ gives $z=1$ when $N\rightarrow \infity$. But Eq. S12 has a $z^N$ term, which tends to a constant $e^\beta$. So S12 does not give correct integral in this case.
Fixed $\beta/N$ means that $e^{\beta}/e^{\beta_c}\rightarrow 0$ if $\beta<\beta_c$ and $N\rightarrow \infty$. So S19 predicts zero when $\beta<\beta_c$. We can multiply $E_r^N/t_1^{N-1}$ to the integral to see this.
\end{comment}
\begin{comment}
The assumed equivalence between the summation and the integral above is subtle and worthy of further remarks. Strictly speaking, it is mathematically valid only in the limit $N\rightarrow \infty$, under which we have $e^{-\beta/N}\simeq 1$.
\textcolor{magenta}{However, for an $E_r$ falling within the loop spectrum of $H$ under PBCs, Eq.~(\ref{eq:Er_enclosed}) yields $|E_r|<t_1$ at PBCs (with $\beta=0$) and $|E_r|=t_1e^{-\beta_c/N}$ at the transition point, meaning that $\beta_c$ is comparable with $N$.
By rewritting Eq. (\ref{eq:G1N_2}) as
\begin{eqnarray}
G_{1N}=\sum_n F(k_n),
\end{eqnarray}
we can see that
$F(k_n)$ tends to infinity when $\beta\gtrsim \beta_c\simeq N$ and $N\rightarrow \infty$, thus the summation cannot be replaced by an integral when $\beta$ is close or larger than the critical value $\beta_c$.
}
\textcolor{magenta}{Numerically, we find that $G_{1N}$ remains a constant when $\beta>\beta_c$, as shown in Fig. \ref{fig:G_NN}(b2).
This can also be seen by directly rewriting Eq. (\ref{eq:G1N_1}) in the form of an integration, which takes a similar form as that under OBCs in Ref. \cite{zirnstein2019bulk,zirnstein2020exponentially,xue2020non}.
The explicit form reads
\begin{eqnarray}
G_{1N}&=&\oint_{|z|=e^{-\beta/N}} \frac{1}{2\pi i } \frac{1}{z^N(E_r-t_{1} z)}dz\nonumber\\
&=& \frac{t_1^{N-1}}{E_r^N},\label{eq:G1N_3}
\end{eqnarray}
i.e. a constant $G_{1N}$ when $|z_0|=|E_r/t_1|>e^{-\beta/N}$, consistent with the results in Fig. \ref{fig:G_NN}(b2).}
\end{comment}
In the same fashion, we can now proceed to examine $G_{N1}$, which depicts how the signal is amplified or suppressed in the other direction. The matrix element $G_{N1}$ is found to be the following,
\begin{eqnarray}
G_{N1}&=&\sum_n\frac{1}{E_r-E_n^R}\psi_{N,n}^R(\psi_{1,n}^L)^*\nonumber\\
&=&\sum_n\frac{1}{E_r-t_1 e^{-\frac{\beta-i2n\pi}{N}}}
\frac{(e^{-\frac{\beta-i2n\pi}{N}})^{N-1}}{N }
\label{eq:GN1_1}\\
&=&\sum_n\frac{e^{-\beta}e^{\frac{\beta-i2n\pi}{N}}}{E_r-t_1 e^{-\frac{\beta-i2n\pi}{N}}}
\frac{1}{N}.\label{eq:GN1_2}
\end{eqnarray}
Rewriting Eq.~(\ref{eq:GN1_2}) in terms of a loop integral, we have
\begin{eqnarray}
G_{N1}&=&\oint_{|z|=e^{-\beta/N}} \frac{1}{2\pi i } \frac{e^{-\beta}}{z^2(E_r-t_{1} z)}dz.
\label{loop2}
\end{eqnarray}
The integrand has a 1st-order pole at $z_0=E_r/t_1$, and a second-order pole at $z_1=0$. Therefore we have
\begin{eqnarray}
G_{N1}= \frac{e^{-\beta}t_1}{E_r^2}
\end{eqnarray}
if $|E_r|>t_1e^{-\beta/N}$ (reference energy falls outside the loop of integral), and
\begin{eqnarray}
G_{N1}= 0
\end{eqnarray}
if $|E_r|<t_1e^{-\beta/N}$ (reference energy $E_r$ falls inside the loop of integral).
Here because of the factor $ e^{-\beta}$ in Eqs.~(\ref{eq:GN1_2}) and Eq.~({\ref{loop2}), the replacement of the discrete sum by the loop integral is always valid by assuming a sufficiently large $N$, i.e., regardless of whether $\beta$ is assumed to be fixed or assumed to scale linearly with $N$. Thus, results obtained above for both $\beta>\beta_c$ and $\beta<\beta_c$ are valid, which are indeed consistent with our numerical results.
{Note also that for fixed $\beta$, our numerical results for finite systems give a small but nonzero $G_{N1}$ when $\beta<\beta_c$ (e.g. $|G_{N1}|\approx e^{-5}$ in Fig.~\ref{fig:G_NN}(c2)), which vanishes when further increasing $N$ (not shown).}
\begin{comment}
Similar to how we treat $G_{1N}$, one may also directly use Eq. (\ref{eq:GN1_1}) instead in the form of an integration and obtain
a different expression of loop integral. After some careful inspection, the results
\begin{eqnarray}
G_{N1}&=&\oint_{|z|=e^\frac{-\beta}{N}} \frac{1}{2\pi i } \frac{z^{N-2}}{E_r-t_{1} z}dz,\nonumber\\
&=&\frac{E_r^{N-2}}{t_1^{N-1}
\end{eqnarray}
i.e. $G_{N1}$ independent of $\beta$ if $|z_0|=|E_r/t_1|<e^{-\beta/N}$.
\textcolor{red}{The numerical results in Fig.~\ref{fig:G_NN}(c2) also show a constant $G_{N1}$ when $\beta<\beta_c$ and a decreasing one when $\beta$ gets larger, with $\beta_c$ the critical value where the gradient of $\ln|G_{1N}|$ jumps from $-1$ to $0$, i.e. when the spectrum evoution touches $E_r$.}
\end{comment}
Overall, we obtain that
the gradient of $\ln|G_{N1}|$ with respect to $\beta$ is again quantized, with
\begin{eqnarray}
\frac{d \ln |G_{N1}|}{d \beta}=-1,
\end{eqnarray}
if $E_r$ is NOT enclosed by the loop spectrum of $H_\beta$, corresponding to $\beta>\beta_c$ [Fig. \ref{fig:G_NN}(c)].
If we further increase $\beta$ the system shall approach the OBC limit when $\beta= \beta_{\rm OBC}\approx\alpha N$ with $\alpha=\ln (\sqrt{t_1/t_{-1}})$, {where the spectrum falls on the same lines as the OBC spectrum~\cite{kunst2018biorthogonal,koch2020bulk}. Nevertheless, this limit is still not exactly like the OBCs, as the two boundaries are still weakly connected. For example, due to the (weak) boundary couplings, a flux threading cannot be gauged away and can lead to fluctuation of eigenergies, unlike in real OBC cases.
Indeed, from our numerical results, we indeed see that $G_{N1}$ keeps decreasing after $\beta$ exceeds $\beta_{\rm OBC}$, and becomes a constant when $\beta\gtrsim 2\beta_{\rm OBC} -\beta_c$ [Fig. \ref{fig:G_NN}(c3)].
\begin{comment}
This hints that for a signal moving through the whole 1D chain, the system approaches the OBC limit at $\beta \approx\alpha N+\beta_c$, instead of $\beta \approx\alpha N$ for the eigensolutions. Here $\beta_c$ is also associated with the reference energy $E_r=\omega+i\gamma$, with $\omega$ the frequency of the signal and $i\gamma$ an overall gain/loss of the system.
On the other hand, the ``amplification" ratio $G_{N1}$ is of the scale around $e^{-50}$, which is hardly physically relevant
This seems to suggest that for a signal moving through the whole 1D chain, the system approaches the OBC limit at $\beta \approx\alpha N+\beta_c$, instead of $\beta \approx\alpha N$ for the eigensolutions. Here $\beta_c$ is also associated with the reference energy $E_r=\omega+i\gamma$, with $\omega$ the frequency of the signal and $i\gamma$ an overall gain/loss of the system. On the other hand, the ``amplification" ratio $G_{N1}$ is of the scale around $e^{-50}$, which is hardly physically relevant.}
Note that $\beta_c$ is given by $N\ln |t_1/E_r|$ from the above calculation, which is an approximating value at $t_1\gg t_{-1}$. This condition is not satisfied in our example with $t_1=2t_{-1}$. In Fig. \ref{fig:G_NN}, $\beta_c$ is obtain numerically from the peak of $I_\beta(E_r)$ and the jump of $d\ln|G_{1N}|/d\beta$.
\end{comment}
\begin{comment}
\clearpage
\subsection{To explain the problem between summation and integral}
The general idea is that the two summations/integrals work for different cases, with fixed $\beta$ and fixed $\beta/N$ respectively. Then we need to look at different parameter regimes to see how they behave.
\subsubsection{$G_{1N}$}
First we consider the case with fixed $\beta$ for $G_{1N}$. To write the summation into an integral, we need $N\rightarrow \infty$, hence we shall also take $\beta/N \rightarrow 0$. So the integral path in Eq. S14 shall be $|z|=1$, not $|z|=e^{-\beta/N}$. In this case, we cannot directly write Eq. S12 into the integral of Eq. S19, because S12 has a term of $e^{(N-1)\beta/N} \approx e^{\beta}$. But with $|z|=1$, this term is assumed to be zero. So S19 is incorrect for a fixed $\beta$. Here the conclusion is that for any fixed $\beta$, we have $\ln|G_{1N}|\approx\beta$ if $N$ is large enough, and if $E_r$ falls within the PBC loop. Numerically we indeed see this for some finite but large $N$.
\bigskip
On the other hand, in our SFA results, we can see that systems with different sizes have the same spectrum as long as $\beta/N$ is fixed. This is also consistent with the OBC limit with $e^{-\beta}\approx e^{-\kappa N}$. And we are looking at the transition when the spectrum with $\beta=\beta_c$ passes a given $E_r$, meaning that $\beta_c/N$ is a constant depending on $E_r$. So we need to consider a fixed $\beta/N$ to study the transition. In this case, we cannot use Eq. S13 as now $e^{\beta}$ diverges with $N\rightarrow\infty$. But we can use Eq. S12 as $\beta/N$ is absorbed by $z$.
For the reason why S19 predicts zero amplification in the regime with $\beta<\beta_c$ (with the plateau at $\nu=1$), it is because now when we increase $N$, the considered $\beta$ value increases (as we have fixed $\beta/N$), and so does the critical value $\beta_c$. So now $\beta/\beta_c$ is also a constant (because $\beta_c/N$ is a constant). Note that the amplification ratio is $|G_{1N}|=e^{\beta}/t_1$ when $\beta<\beta_c$, and $|G_{1N}|=|t_1^{N-1}/E_r^N|=e^{\beta_c}/t_1$ when $\beta>\beta_c$. Then $e^{\beta}/e^{\beta_c}$ tends to zero when $N\rightarrow \infty$, meaning that the amplification ratio $e^{\beta}$ is ignorable for this integral when $\beta<\beta_c$. So this integral gives zero when $\beta<\beta_c$.
\subsubsection{$G_{N1}$}
Next, we consider the signal moving toward the other direction, i.e. S20 and S21 in our paper. Now the coefficient in S21 (sum) and S22 (integral) is $e^{-\beta}$, so it works well even when beta is large. When $\beta<\beta_c$, S22 gives zero. Our numerical result is actually also very close to zero [in Fig. S1(c2) it is about $e^{-5}$]. When increasing the system's size, the starting value in Fig. S1(c2) will also decrease. When $N\rightarrow \infty$ with fixed $\beta$, we shall have zero amplification ratio for any fixed $\beta$ (as $\beta_c\rightarrow\infty$).
\bigskip
Similarly, S20 (sum) and S25 (integral) work for fixed $\beta/N$. It now predicts the constant value of $G_{N1}$ when $\beta<\beta_c$. For $\beta>\beta_c$, $\beta/\beta_c$ is a constant larger than $1$ when increasing $N$, as $\beta_c/N$ is a constant. Now the ``amplification" ratio is $|G_{N1}|=e^{-\beta}t_1/E_r^2$ when $\beta>\beta_c$, and $|G_{N1}|=|E_r^{N-2}/t_1^{N-1}|=e^{-\beta_c}t_1/E_r^2$ when $\beta<\beta_c$.
So $e^{-\beta}/e^{-\beta_c}$, i.e. the ratio between the ``amplification" for $\beta>\beta_c$ and $\beta<\beta_c$, tends to zero when $N\rightarrow \infty$, hence $e^{-\beta}$ at $\beta>\beta_c$
is ignorable in this integral.
\end{comment}
\subsubsection{Cases with only $m$th-nearest neighbor coupling}
Consider now a 1D non-Hermitian chain with only the $m$th-nearest neighbor couplings:
\begin{eqnarray}
H_\beta &=& \sum_{x=1}^{N-m}(t_{m}\hat{c}^{\dagger}_{x}\hat{c}_{x+m}+t_{-m}\hat{c}^{\dagger}_{x+m}\hat{c}_x)\nonumber\\
&&+e^{-\beta}\sum_{x=N-m+1}^{N}(t_{m}\hat{c}^{\dagger}_{x}\hat{c}_{x+m-N}+t_{-m}\hat{c}^{\dagger}_{x+m-N}\hat{c}_x).\nonumber\\
\label{m_hn_model}
\end{eqnarray}
Here we first assume $N/m$ is an integer,
thus the system is decoupled into $m$ identical 1D sub-chains, and the eigenstates satisfy
\begin{eqnarray}
t_{m} \psi^R_{x+m,n}+t_{-m} \psi^R_{x-m,n}&=&E_n^R\psi^R_{x,n}\label{eq:bulk_m_HN}
\end{eqnarray}
for $x\in[2,N-m]$, and
\begin{eqnarray}
e^{-\beta}t_{m} \psi^R_{s,n}+t_{-m} \psi^R_{N-2m+s,n}&=&E_n^R\psi^R_{N-m+s,n},\label{eq:boundary1_m_HN}\\
t_{m} \psi^R_{m+s,n}+e^{-\beta}t_{-m} \psi^R_{N-m+s,n}&=&E_n^R\psi^R_{s,n},\label{eq:boundary2_m_HN}
\end{eqnarray}
with $s=1,2,...,m$ labelling different sub-chains.
As in the previous discussion for the case of $m=1$, we assume $t_m\gg t_{-m}$ to obtain some simple analytical results.
Here we replace the labels $x$ and $n$ with $x_s$ and $n_s$ for each sub-chain. We take the ansatz
\begin{eqnarray}
\psi^R_{x_s,n_s}=C_{n_s}e^{-M_{n_s}^R(x_s-1)},
\end{eqnarray}
with $n_s$ only takeing values from $1$ to $N_m=N/m$, given that each sub-chain contains only $N_m$ lattice sites, and $x_s= (x-s+m)/m$ ranging from $1$ to $N_m$, with $x$ being $s,m+s,2m+s,...,N-m+s$, and more importantly,
\begin{eqnarray}
t_{m} e^{-M_{n_s}^R}=E_{n_s}^R, M_{n_s}^R=\frac{\beta-i2n_s\pi}{N_m}.
\end{eqnarray}
Similarly, the left eigenstates are given by
\begin{eqnarray}
\psi_{x_s,n_s}^L&=&C_{n_s}^* e^{-M_{n_s}^L (N_m-x_s)}, t_{m} e^{-M_{n_s}^L} =E_{n_s}^L, \nonumber\\
M_{n_s}^L&=&\frac{\beta+i2{n_s}\pi}{N_m}.
\end{eqnarray}
Again, from the biothorgonal condition $\langle \Psi_{n}^L|\Psi_{n}^R\rangle=1$, we have the normalization constants
\begin{eqnarray}
C_{n_s}=\frac{e^{-i\frac{N_m-1}{N_m}\pi {n_s}}}{\sqrt{N_me^{-\beta(N_m-1)/N_m}}}.
\end{eqnarray}
Note that in the Green's function matrix, the element $G_{1N}$ shall always be zero as each sub-chain is decoupled from the others.
In this case, the directional amplification of each sub-chain corresponds to the element
$G_{s(N-m+s)}$, with
\begin{eqnarray}
G_{s(N-m+s)}&=&\sum_{n_s}\frac{1}{E_r-E_{n_s}^R}\psi_{x_s=1,{n_s}}^R(\psi_{x_s=N_m,{n_s}}^L)^*\nonumber\\
&=&\sum_{n_s}\frac{1}{E_r-t_{m} e^{-\frac{\beta-i2n_s\pi}{N_m}}}
\frac{1}{N_m(e^{-\frac{\beta-i2n_s\pi}{N_m}})^{N_m-1}}\nonumber\\
\\
&=&\sum_{n_s}\frac{e^{\beta}e^{-\frac{\beta-i2n_s\pi}{N_m}}}{E_r-t_{m} e^{-\frac{\beta-i2n_s\pi}{N_m}}}\frac{1}{N_m}.
\end{eqnarray}
As $n_s$ takes value from $1$ to $N_m=N/m$, here we need to define $k_s=i2mn_s\pi/N$, so that the summation can be replaced by an integral with $k_s$ varying from $0$ to $2\pi$. Similar to the case with only nearest-neighbor coupling, we then have
\begin{eqnarray}
\frac{d\ln |G_{s(N-m+s)}|}{d\beta}=1
\end{eqnarray}
for each sub-chain, when $E_r$ is enclosed by the loop-like spectrum of each sub-chain.
This result indicates that each sub-chain has its own spectral winding number $\nu_s(E_r)=1$, but for the original 1D chain with $m$th-nearest neighbor couplings, the element $G_{1N}$ and $G_{N1}$ are zero as sites $1$ and $N$ belong to different decoupled sub-chains.
On the other hand, the $m$-sub-chain picture here also indicates an effective unit-cell structure with $m$ sublattices, even though the sublattices are physically equivalent on the lattice.
Thus the directional amplification of the overall system comprised by these sub-chains/sublattices shall be described by the combination of that of each sub-chain, corresponding to the corner blocks of the overall Green function matrix,
\onecolumngrid
\begin{eqnarray}
G_{\leftarrow,m\times m}=
\left(\begin{array}{cccc}
G_{1(N-m+1)} & G_{1(N-m+2)} &\cdots & G_{1N}\\
G_{2(N-m+1)} & G_{2(N-m+2)} &\cdots & G_{2N}\\
\vdots & \vdots &\vdots &\vdots \\
G_{m(N-m+1)} & G_{m(N-m+2)} &\cdots & G_{mN}\\
\end{array}\right), \\
G_{\rightarrow,m\times m}=
\left(\begin{array}{cccc}
G_{(N-m+1)1} & G_{(N-m+1)2} &\cdots & G_{(N-m+1)m}\\
G_{(N-m+2)1} & G_{(N-m+2)2} &\cdots & G_{(N-m+2)m}\\
\vdots & \vdots &\vdots &\vdots \\
G_{N1} & G_{N2} &\cdots & G_{Nm}\\
\end{array}\right),
\end{eqnarray}
\twocolumngrid\noindent for signals moving toward the left- and the right side, respectively.
In the above schematic scenario where the sub-chains are fully decoupled from each other, only the diagonal elements of the above two matrices are nonzero, i.e.
$(G_{\leftarrow,m\times m})_{ab}\propto \delta_{ab}e^{\beta}$ when $E_r$ is enclosed by the spectrum on the complex plane, and $(G_{\rightarrow,m\times m})_{ab}\propto \delta_{ab}e^{-\beta}$ when $E_r$ is NOT enclosed by the spectrum on the complex plane.
This being the case, we arrive at
\begin{eqnarray}
\det[G_{\leftarrow,m\times m}]\propto e^{m\beta},
~\det[G_{\rightarrow,m\times m}]\propto e^{-m\beta},
\end{eqnarray}
in the above two cases repectively.
Thus we have
\begin{eqnarray}
\nu_{\leftarrow,m}:=\frac{d\ln |G_{\leftarrow,m\times m}|}{d\beta}=m,
\end{eqnarray}
corresponding to the spectral winding number $\nu(E_r)=m$ for the case with only the $m$th-nearest neighbor couplings.
Furthermore, this conclusion shall also be valid when $N/m$ is not an integer. In such cases, the system still possesses the $m$-sub-chain picture, only that the two ends of one sub-chain are connected to those of other sub-chains now. This conclusion is also verified by our numerical calculations. As also mentioned in the main text, the physical analysis here is also valid if the system has many different types of coupling coexisting, but is still dominated by one type of coupling. The topological robustness of the response can then retain the quantization.
\clearpage
\onecolumngrid
\begin{center}
\textbf{\large Supplementary Materials}\end{center}
\setcounter{equation}{0}
\setcounter{figure}{0}
\renewcommand{\theequation}{S\arabic{equation}}
\renewcommand{\thefigure}{S\arabic{figure}}
\renewcommand{\cite}[1]{\citep{#1}}
\section{Cases with couplings across different ranges}
In the main text, we have already witnessed an intriguing example where both nearest-neighboring and next-nearest-neighbor couplings present in the system. In that case, with respect to different reference energy points, the spectral winding can have different nonzero values and this is also manifested as different quantized responses. Benchmarking this with our decoupled sub-chain picture, this indicates that there is competition in different regimes, and the underlying topological robustness associated with either single-chain physics or two coupled sub-chain physics
still yields, remarkably, quantized responses. The next question is then, if we add more and more complexity to a non-Hermitian lattice system with many coexisting hopping length scales, can we still observe quantization and hence a clear correspondence between spectral winding and the signal amplification. Our answer is yes based on more computational tests.
\begin{figure*}[h]
\includegraphics[width=1\linewidth]{spectrum_G_t2t3.pdf}
\caption{(a) PBC (blue loop) and OBC (black dots) spectra of a system with many coexisting hopping length scales, and the PBC-OBC spectral evolution between them (blue-purple curves), tuned by the parameter $\beta$. Four reference energy values $E_r$ are chosen to inspect the possible quantization of the gradient of the logarithm of directional amplification with respect to $\beta$, each with a different winding number $\nu(E_r)$, as indicated by the four colored marks.
(b)-(e) show the summed reverse energy spacing $I_\beta(E_r)=\sum_n |1/(E_n(\beta)-E_r)|$, logarithm of the amplification ratio for the block of $G$ associated with the winding number $\nu(E_r)$, as well as $\nu_{\leftarrow,m}$ and $\nu_{\rightarrow,m}$, as functions of $\beta$. In (e) we have $\nu(E_r)=0$, hence we only illustrate amplification ratios $G_{\leftarrow,1\times 1}$ and $G_{\rightarrow,1\times 1}$, which are both less than unity, indicating no amplification for a signal moving either toward the left or the right.
Parameters are $t_{2}=2$, $t_{-2}=1$, $t_3=0$, $t_{-3}=1$, and $t_{\pm m}=0$ for all other values of $m$, with $N=300$ lattice sites.
}
\label{fig:t2t3}
\end{figure*}
Consider then a lattice model with nonzero terms
$\{t_{-r},t_{-r+1},...,t_{l-1},t_{l}\}$, and none of them is dominating over the rest. We can still view this system
as one comprised of $m$ sub-chains, with $m\leqslant {\rm Max}[r,l]$. We can investigate if the same response functions $\nu_{\leftarrow,m}$ or
$\nu_{\rightarrow,m}$ can reflect the spectral winding behaviors. To proceed specifically, consider the following Hamiltonian as an example,
\begin{eqnarray}
H=\sum_{x=1}^N\sum_{j=-r}^{l} t_j\hat{c}^\dagger_j\hat{c}_{x+j},
\end{eqnarray}
with $r=l=3$, and the boundary coupling tuned via $t_j\rightarrow e^{-\beta}t_j$ when it connects sites at different ends of this 1D chain.
The rather complicated spectral winding behavior is shown in Fig.~\ref{fig:t2t3}(a), indicating winding numbers -1, 1, 2, and 0.
As shown in Fig.~\ref{fig:t2t3}(b)-(d),
the obtained $\nu_{\leftarrow,m}$ or $\nu_{\rightarrow,m}$ for $m=1$ or $m=2$ still shows relatively clear plateaus for cases with nonzero $\nu(E_r)$, with the transitions of these plateaus in excellent agreement with the critical $\beta$ values for which the spectral winding numbers make jumps. In panel (c3), the quantization in $\nu_{\leftarrow,1}$ is clearly seen. A careful reader might notice that in panel (b3) and (d3), the plateaus of the obtained response function $\nu_{\rightarrow,1}$ are so not clearly quantized. To double check if this is merely a finite-size effect, we
have increased the size of the model system and then much better quantized plateaus are indeed observed, as presented in Fig.~\ref{fig:size}.
As to the case of $\nu(E_r)=0$ labeled by the yellow star in Fig.~\ref{fig:t2t3}(a), no amplification is obtained for a signal moving toward either the left- or the right-hand side, as indicated by the always-negative $\ln|G_{\leftarrow,1\times 1}|$ and $\ln|G_{\rightarrow,1\times 1}|$ in Fig.~\ref{fig:t2t3}(e).
\begin{figure}[h]
\includegraphics[width=0.6\linewidth]{size.pdf}
\caption{(a) $\nu_{\rightarrow,1}$ and (b) $\nu_{\leftarrow,2}$ as functions of $\beta/N$ for the two cases in Fig.~\ref{fig:t2t3}(b) and (d) respectively, with different numbers of lattice $N=300$, $400$, and $600$.
As our previous analytical results suggest that the transition value $\beta_c$ is proportional to $N$,
here the variable $\beta$ is rescaled by a factor of $1/N$, so as to map the transition points for systems with different sizes to the same parameter.
The plateaus are seen to be flatter (hence better quantization) for larger systems.
Other parameters are $t_1=1$ and $t_{-1}=0.5$.
}
\label{fig:size}
\end{figure}
\section{Illustrative details of circuit for probing the response of the Hatano-Nelson model}
To be concrete, we provide explicit details for the realization of the simplest case of the circuit representing the Hatano-Nelson model with topological winding $\nu=1$. It consists of a chain of nodes connected by unbalanced couplings which simultaneously give rise to non-Hermiticity and non-reciprocity. It is well-established that such couplings can be realized with INICs (negative impedance converters with current inversion)~\cite{hofmann2019chiral,helbig2020generalized}, which contains operation amplifiers that break the reciprocity.
\begin{comment}
can be realized a configuration capacitors and a negative impedance converter with current inversion (INIC)~\cite{li2020impurity}\textcolor{red}{cite more}. That is, for a configuration as shown in \textcolor{red}{a figure,} its Laplacian reads
\begin{eqnarray}
J_{ab}=i\omega \left(
\begin{array}{cc}
C_b-C_a&C_a-C_b\\
-C_a-C_b&C_a+C_b
\end{array}
\right),
\end{eqnarray}
and the total Laplacian for the LC circuit in \textcolor{red}{a figure}, together with the $\Omega$ grounding elements, takes the form of
\begin{align}
J=& \biggl[e^{-\beta(\omega)}\left(e^\alpha|N\rangle\langle 0|+e^{-\alpha}|0\rangle\langle N|\right)
+\sum_{x=0,\pm}^{N-1}e^{\pm\alpha}|x\rangle\langle x\pm 1|
\notag\\
&-\sum_{x=0}^{N}\left(2 \cosh\alpha-\omega_0^2/\omega^2+\Omega(\omega)\right)|x\rangle\langle x|\biggl]\times i\omega C
\label{J}
\end{align}
\begin{eqnarray}
-e^{-\beta(\omega)}[(C_2+C_1)|N\rangle\langle 1|+(C_2-C_1)|1\rangle\langle N|]\nonumber\\
-\sum_{x=1}^{N-1}(C_2+C_1)|x\rangle\langle x+ 1|+(C_2-C_1)|x+1\rangle\langle x|\nonumber\\
-\sum_{x=1}^{N-2}(C_4+C_3)|x\rangle\langle x+ 2|+(C_4-C_3)|x+2\rangle\langle x|
\end{eqnarray}
\end{comment}
To realize the tuning of the spectral reference point $\Omega$ and end-to-end couplings $e^{-\beta(\omega)}$, we also include tunable inductors and additional RLC elements as according to Ref.~\cite{li2020impurity}. Its Laplacian, together with these tunable elements, takes the form
\begin{align}
J=& \biggl[e^{-\beta(\omega)}\left(e^\alpha|N\rangle\langle 0|+e^{-\alpha}|0\rangle\langle N|\right)+\sum_{x=0,\pm}^{N-1}e^{\pm\alpha}|x\rangle\langle x\pm 1|
\notag\\
&-\sum_{x=0}^{N}\left(2 \cosh\alpha-\omega_0^2/\omega^2+\Omega(\omega)\right)|x\rangle\langle x|\biggl]\times i\omega C
\label{J}
\end{align}
where $\alpha=\tanh^{-1}\frac{C_1}{C_2}$, $C=\sqrt{C_2^2-C_1^2}$, $\omega_0^{-2}=l_{gr}C$ and $\Omega = \frac{-C_\Omega+iR_\Omega/\omega}{C}$. $C_1$ and $C_2$ are capacitors involved in internode couplings as according to Ref.~\cite{li2020impurity}, and $R_\Omega$, $l_{gr},C_\Omega$ are RLC elements that connect each node to the ground. By varying the choice of $C_\Omega$ and $R_\Omega$, $\Omega$ and hence different points of the complex eigenvalue plane can be sampled. In this circuit, the $\nu=1$ and $\nu=0$ regions are separated by the curve $\Omega=2\cosh(\alpha+ik)-2\cosh\alpha-\omega_0^2/\omega^2$. The effective flux $\beta(\omega)$ can be adjusted by tuning the AC frequency $\omega$, and is given by $\beta(\omega)=\ln[1-\omega^2l(C_2-C_1)]$, where $l$ is an inductor involved in the internode couplings which controls the sensitivity of the tuning. Clearly, by including additional couplings between more distant nodes~\cite{li2019emergence,lee2020imaging,ezawa2019electric2,lu2019probing,zhang2020topolectrical,song2020experimental,stegmaier2020topological}, this circuit construction can be extended to models with additional further couplings, such as that shown in the main text with $t_1=1,t_{-1}=0.5,t_2=2,t_{-2}=0$.
| d9de85a7a04ebe3061bf4667b49a6cf11e36e445 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Recommender systems can help users in fulfilling their intentions by suggesting the most relevant actions to follow. A wide variety of systems has been developed, while most of the approaches benefit from modeling of long-term user profiles whether of sequential or static nature. In many real-world applications, long-term user profiles are not available, and therefore, suggestions have to be made solely based on the observed actions during an ongoing session. Typically, session-based recommendation algorithms suggest a user the immediate next actions. The practical relevance of the problem leads to a number of proposals for algorithms mostly based on the analysis of item-level dependencies in interactional data.
When making recommendations based only on the click-stream data from an ongoing session, it is reasonable to take into account characteristics usually coming with the items. Users are likely to browse alternatives when they visit a website with the intention of purchasing a specific type of product. The alternative relations among products can be recognized by similar text descriptions. It has been shown that incorporating rich features into the recommendation process helps to deal with the sparsity and the item cold start, though it is not trivial \cite{hidasi2016parallel, tanjim2020attentive}.
Most of the existing session-based recommenders assume sessions to be associated with a single intention, however, it may change within an individual session multiple times. Detecting and adapting to these changes is one of the open challenges. Based on the idea of interactional context, a precondition can be devised that allows the delineation of coherent sub-sequences of semantically related actions (Fig.~\ref{fig:interactional-context}). For this purpose, we employ a technique for semantic analysis of natural language to extract high-level concepts from the item descriptions. We used it to help reduce unnecessary information that may introduce uncertainty into the intention change detection task. Within an individual session, users often browse items that complement the ones viewed earlier in the session. Hence the changes of semantics are intuitively expected to reflect patterns.
\begin{figure}[h]
\centering
\includegraphics[width=0.6\textwidth]{assets/multi-contextual-session_deps.pdf}
\caption{Given that context is a relational property, an action can be related to its close neighborhood based on an arbitrarily devised precondition of coherence. As a result, a session is split into several sub-sequences of actions, within which context is seemingly constant and among which drifts can be easily detected. There are sequential dependencies expected, both on the level of actions and the level of sub-sequences.}
\label{fig:interactional-context}
\Description{figure description}
\end{figure}
This paper is devoted to the hypothesis that boosting the score of semantically related items will improve the ranking of a recommendation list generated by a baseline method. The idea underlying our approach is to make a list of top-k recommendations and then re-rank the items in a way that items semantically related to the current intention will be moved to the top of the list while the order among them is maintained.
The main contributions presented in this paper are as follows:
\begin{itemize}
\item we extend a simple k-NN based recommender with a semantic factor that helps deliver recommendations on the user's current intention, the proposed cSkNN method improves Mean Reciprocal Rank (MRR) in all conducted experiments,
\item we compare effects of several linguistic-based generalization configurations on the recommended items ranking,
\item we evaluate two versions of semantic adaptation in the course of activity and compare their results to the potential given by the generalization configuration.
\end{itemize}
\section{Related works}
Traditional collaborative-filtering methods cannot be used as a session-based recommender because no past user behavior is available. A vital solution is the approach based on item co-occurrences in available click-stream data \cite{linden2003amazon}. Recommending items that are often clicked together has been proven to be effective in real-world applications and such nearest-neighbor-based algorithms of almost trivial nature are often used as competitive baselines. Session-based kNN method (SkNN), compared to Item-based kNN, considers an entire session when calculating the similarity, not only the last action. Several SkNN extensions providing sequential awareness are competitive with considerably more complex approaches, even those neural-network-based \cite{ludewig2018evaluation, jannach2017recurrent}.
Sequential recommenders are based on Markov model to utilize sequential dependencies. For example, Markov model makes predictions of latent topics derived from song tags that are then used to post-filter the initial ranking produced by a traditional kNN algorithm \cite{hariri2012context}. The simple Markov chains methods calculate transition probabilities depending only on the previous state. Higher orders may be used to model more complex dependencies, however, it dramatically increases the computational complexity.
Recurrent neural networks (RNN) are better suited to model complex dependencies and are used in many areas. Gated recurrent units (GRU) can handle longer dependencies and achieve significant improvements over traditional methods \cite{hidasi2015session}.
Deep RNNs have the ability to capture sequential dependencies among a variety of entities and to involve rich features such as text descriptions or images \cite{hidasi2016parallel}, so naturally, the most of the research has been devoted to them in the past few years. Rich item representations have been utilized in Neural Attentive Session-based Recommendation model (NARM) to focus attention on items similar to session purpose \cite{li2017neural}. There are also works that use such representations to divide a session. Mixture-channel purpose
routing networks (MCPRNs) model multiple intents in an individual session \cite{wang2019modeling} or semantic changes in attentive bi-GRU with song tags \cite{sachdeva2018attentive}. Attentive sequential
model of latent intent (ASLI) method uses categories to model the intent \cite{tanjim2020attentive}.
\section{Proposed approach}
The underlying idea of proposed cSkNN is to make a list of top-k recommendations and re-rank the items according to their semantics. In this section, we elaborate on our motivation and describe our method.
\subsection{Motivation}
Contextual information is an additional and important modeling dimension that helps to enhance the prediction accuracy. Existing approaches mostly focus on a \textit{representational view} of the context. It is a descriptor of an environment in the form of stable information that can be delineated in advance and then simply observed before activity or during its initialization, e.g., weather, time, or a device’s sensory inputs. However, we lack the richness of such relevant contextual information in e-commerce domain so we take the position of an alternative \textit{interactional view} that differs from the former fundamentally. Instead of context being considered as a bit of information, it is treated as a relational property that is actively produced and maintained in the course of activity \cite{dourish2004we}. A session-based recommender is a kind of a protagonist of such a phenomenological position since a session can be seen as context in the sense of a property relating actions together.
However, context is perpetually being altered so when context is represented with a session as a whole, it may contain misleading information that is not up-to-date after several actions. Since a system cannot monitor all factors inducing changes in context, a temporal factor is often used as an acceptably effective proxy for these hidden factors. Researchers proposed a variety of functions that decay weight of actions in time or attention mechanisms to assign significance to the most recent actions; and/or to actions related to them \cite{li2017neural, garg2019sequence, guo2018adjustable, campos2014time}. The drawbacks of using temporal information as a solitary indication of a change are that there is no indication of a preference shift direction, and also that widely-used decaying functions lack the ability of adaptation to sudden changes within an individual session.
Labeling each action with a high-level concept conveys a reliable coherence precondition that is needed for the drift detection. Generalization of actions that enforces high-level similarities and repress low-level differences (i.e., noise) will indicate the exact boundaries of sub-sequences as depicted in Figure~\ref{fig:interactional-context}. Representations learned by the distributional analysis of the interactional data inherently give rise to generalizations. However, this approach is less useful if a sufficient amount of user-item interactions is not available and also introduces bias from user interactions (as such information is already available on the level of items).
Therefore, we opt to exploit a concise nature of item titles in e-commerce catalogs, from which high-level concepts can be obtained via semantic analysis. Despite difficulties emerging from the akin nature of the text, dependency parsing enables a higher control over the process of semantic generalization compared to the analysis based on the distributional hypothesis or a popular bag-of-words approach.
Intuitively, transitions among coherent sub-sequences follow patterns. For example, users tend to see complementary accessories (e.g., belt) along with items that exhibit the origin purchase intention (e.g., trousers), and also users might browse items that together constitute a whole outfit. A sequential pattern of concepts “trousers” → “belt” would have higher support and confidence than sequences of particular items that correspond to those concepts. Since all such patterns on the item-level are aggregated to a single generalized pattern on the concept-level. This helps to tackle the sparsity and possibly the cold start problem.
As SkNN methods are already competitive with the more sophisticated approaches, we devote more effort to their development. We evaluate a version with a simple extension (S-SkNN) that makes the method aware of sequential dependencies \cite{ludewig2018evaluation}. The S-SkNN method is used for predicting both items and semantics used to post-filter a set of recommended items.
\subsection{Semantic clusters as a coherence precondition}
We make a division of items regarded as having particular shared characteristics expressed by natural language into mutually exclusive \textit{semantic clusters}. These will satisfy the coherence precondition that allows splitting a session into several coherent sub-sequences among which is easy to detect interest drifts.
Inferring intentions from items is more difficult than inferring them from categories \cite{tanjim2020attentive} because they ease interaction sparsity issues and also carry explicit information about item meaning for a user. In e-commerce, human-made categories have a role of such semantic clusters, however, they are mostly widely defined and contain hundreds or thousands of items (e.g., \textit{summer dress}). The larger clusters the more items of a higher variability they contain, hence it implies a sort of hierarchical organization in which a variety of items are aggregated into a single cluster. This allows bound actions to their neighbors into semantic windows as depicted in Figure~\ref{fig:configurations}. Since the length of windows is directly affected by the size of clusters, we need to strike a balance between their size and the level of abstraction that allows both to detect drifts of interest and effective semantic filtering of recommendations. The size of semantic clusters has to be relatively smaller than the usual category size, however, not too small, because it would reduce the effectiveness of sparsity reduction and make the semantics prediction harder.
Besides the size also the constitution of clusters is important because by merely enlarging clusters we can cause items that do not satisfy the convenient coherence condition to merge into a single cluster. Generalization based on natural language processing can help to overcome this issue. Dependency parsing is the task of recognizing a syntactic structure in a sentence, hence it allows a controlled extraction of words that carry a semantically essential part of item characteristics.
The main goal of the product title is to attract the attention of potentially interested users, and therefore, titles are written rather concisely and mostly contain words describing the main characteristics of the product along with several adjuncts. Although we can exploit such nature, it also often leads to syntactically or grammatically incorrect constructions that considerably complicate the semantic analysis. We identified several issues and inconveniences, such as multiword and multi-language expressions or inflections of words, however, in our experiments we put the only reasonable effort in resolving the most critical ones. We believe that more effort would improve the end results, so we will discuss some of the identified issues briefly in Section \ref{sec:discussion}.
\begin{table}[h]
\caption{Definitions of generalization configurations and example inputs for item title {\itshape ``Red striped dress with silver zip on back''}}
\label{tab:confs}
\begin{tabular}{cll}
\toprule
Conf.&Extraction rule&Example\\
\midrule
\#1 & root + first degree nmod & dress zip\\
\#2 & root + second degree nmod & dress zip back\\
\#3 & root amods + root + first degree nmod & red stripes dress zip\\
\#4 & all words & red stripes dress silver zip back \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{assets/dependency_tree.pdf}
\caption{Visualisation of a dependency tree with for example input title “Red striped dress with silver zip on back”}
\label{fig:configurations}
\Description{figure description}
\end{figure}
The idea is to extract salient words rather linguistically than statistically as via TF-IDF method. Dependency parsing help to keep only highly distinctive descriptors and to reduce low-level noise, e.g., infrequent colors, causing the items to be merged in an undesirable way. The key part of parsing is the nominal phrase root extraction since it has a function to express a high-level concept which is in collocation with modifiers that add descriptive information to it \cite{de2014universal}. We focus on nominal dependents (\textit{nmod}) because they function as a non-core argument that brings about a relatively dense packaging of referential information, e.g., \textit{dress with zip on back}, which results in an extreme reliance on implicit meaning providing a utility for our task. Another important function is an adjectival modifier (\textit{amod}) which modifies the meaning of a nominal, e.g., \textit{striped dress}. Usually, it describes qualities of items such as material, color, or other distinctive appearance or functions.
Table \ref{tab:confs} describes four different rules of the generalization process from the standpoint of dependency parsing which control what words are extracted and used to create high-level concepts for items. We call them configurations in this paper from now on.
We need to create semantic clusters to use their \textit{IDs} during recommendation using a k-NN method. After words are normalized, we use \textit{word2vec} model trained on all item titles to create semantic relations among similar concepts which is useful in the process of generalization during the clustering process. Each item is then represented by an average of vectors of extracted words concatenated with an item category vector. This allows to distinguish among likewise named concepts, e.g., t-shirt in women and in men category, and the nature of the category tree also allowed to capture relations based on the parts of an outfit, e.g., upper or lower body.
We used a density-based clustering method DBSCAN to group items together with their surroundings given by an $\epsilon$ parameter.
The radius given by $\epsilon$ directly influences the number of clusters and their sizes, therefore, it is another factor that strongly influences the generalization process alongside the word extraction rules.
We used different $\epsilon$ values according to the distributions of nearest neighbors' distances for each configuration.
As the result, we have 29 different mappings and for each of them, we have a function $cmap:I \rightarrow C$, where $I$ is a set of all items and $C$ is a set of all semantic clusters for a particular semantic cluster configuration.
\subsection{Recommendation methodology}
To make recommendations, we use S-SkNN approach with an extension providing sequential awareness: given the current session $s$, its neighbors $N_s$ and a pairwise similarity function, the recommendation score for an item $i \in I$ is defined as:
$$score_{S-SkNN}(i,s) = \sum_{n \in N_s} sim(s,n) \cdot w_n(s) \cdot 1_n(i)$$
where the similarity function is a dot product of binary-encoded sessions,
the indicator function $1_n(i)$ returns $1$ if session $n$ contains item $i$ and $0$ otherwise.
If an item $s_x$ is the most recent item of the current session and it also appears in the neighbor session $n$, then the weight is defined as $w_n(s)=x/|s|$, where the index $x$ indicates the position of $s_x$ within the session.
Hence, the weight increase depends on the position of an item that appears in both sessions, while the more recent position in the current session gains a higher weight. More information can be found in \cite{ludewig2018evaluation}.
To re-rank recommendations and favor semantically related items, we extend the scoring function by the semantic factor:
$$score_{cSkNN}(i,s) =score_{S-SkNN}(i,s) + c_s(i)$$
where $c_s(i)$ returns $1$ if an item $i$ is relevant to semantics given by $s$ and $0$ otherwise.
Since the S-SkNN score is normalized into a range of $<0,1>$, the $c_s(i)$ factor assures the semantically relevant items to appear at the beginning of the recommendation list.
We evaluate two different implementations of the semantic factor functions which differ in the complexity of generating a semantic cluster candidate.
\subsection{Semantic candidates by the last action \label{sec:last-semantics}}
Based on the assumption that semantics does not change along with several actions, we evaluate a version of a model that favors items semantically related to the last known action.
As a downside, the semantic factor function suggests wrong semantics at each beginning of a coherent sub-sequence. On the other hand, it is able to adapt immediately and there is some assurance that it will be correct for several following actions while it depends on the quality of semantic clusters.
In our interaction data, the average length of semantically coherent sub-sequences is roughly from $1.5$ to $3.8$, depending on the clusters' configuration.
In this version, $c_s(i)$ returns $1$ if $cmap(i)$ equals to $cmap(s_{|s|})$, where $s_{|s|}$ is the last known action in the session $s$.
\subsection{Prediction of semantic candidates \label{sec:predicted-semantics}}
Semantics is derived not only from the last known action but rather from the entire session.
During the training phase, another S-SkNN instance is trained on the train sessions mapped on semantic clusters from a particular $C$ configuration using the $cmap(i)$ function. The instance can be then used for the prediction of the next immediate semantic cluster.
To make an effect with semantic filtering, we have to rely on the top-1 prediction according to semantic-level dependencies, which is used to improve the score of semantically related items in the items’ recommendation list.
Therefore, $c_s(i)$ returns $1$ if $cmap(i)$ equals to a semantic cluster given by top-1 prediction using $score_{S-SkNN}(c, s')$, where $c \in C$ and $s'$ is the current session mapped onto the configuration $C$.
\section{Experiments}
\subsection{Dataset}
We conducted our experiments on a fashion e-commerce dataset that contains approx. 572k actions in 100k sessions in a period of 14 days.
The number of users is unknown since sessions are anonymous, therefore, only short-term dependencies within an individual session can be analyzed.
The dataset is very sparse, the density calculated by session-item interactions is 0.011\%.
The dataset contains approx. 155k products organized into 366 categories of shallow hierarchical structure with the top-level categories dividing items by the gender and age diversity dimensions, i.e., women, men, girls, and boys. The lower 3 levels divide items by the parts of the outfit, e.g., shirts, shoes, or accessories. Although our semantic analysis is based on all of the items, only 39k of them are used in the interaction data. As the texts are in the Czech language, we used UDPipe tool with Czech model \cite{udpipe:2017} for dependency parsing.
\subsection{Evaluation protocol}
We use an evaluation scheme in which the task is to predict the immediate next item given the first $n$ actions of a session.
For each session, we iteratively increment $n$ and make predictions. This simulates user activity.
Each time a prediction is made, we measure two popular top-k metrics in recommender systems, namely: hit rate (HR@k) and Mean Reciprocal Rank (MRR@k) for a set of \textit{k} values.
HR measures the proportion of cases having the true item amongst the top-k items, which is a suitable metric for certain scenarios where absolute order of recommendations does not matter.
MRR is important in cases where the order of recommendations matters, which is our case since the main purpose of our task is to re-rank items and to prefer items by the semantics of the current user intent. Both the reciprocal rank and hit rate are set to zero if the rank is above \textit{k}.
The dataset is divided into random train and test subsets on the level of sessions in a ratio of 80:20, each configuration is evaluated on the same train-test sample.
For the comparison, the following types of recommenders are reported:
\begin{itemize}
\item {\bfseries baseline}: a vanilla S-SkNN model without the semantic factor with a random sampling of 1000 neighboring sessions to speed up the prediction. Therefore, we cached the scores outputs given by $score_{S-SkNN}(i,s)$ and used them in the evaluation of all semantic configurations and methodology versions.
\item {\bfseries last semantics}: a cSkNN model using the last seen semantics as described in Section \ref{sec:last-semantics}.
\item {\bfseries predicted semantics}: a cSkNN model using the semantics predicted by the entire session as described in Section~\ref{sec:predicted-semantics}. Different S-SkNN instance is trained for each semantic clusters configuration separately. Only 25\% of the most recent actions in the train set is used to train the semantics prediction since more of the data did not lead to significant improvement because generalization produces a large number of identical sequences.
\item {\bfseries potential}: to show the maximum theoretical improvement, we employ a version of the $c_s(i)$ function whereby it uses the true next immediate semantic cluster via a leak from the future.
\end{itemize}
For all schemes with the semantic factor, we re-rank top-50 recommendations generated by the baseline model.
\subsection{Results}
The evaluation showed that semantic post-filtering increases the performance even though the semantic prediction accuracy is relatively low. One of the most important takeaways is that the generalization via NLP has a higher potential for improving recommendations compared to the popular bag-of-words approach which is manifested by configuration \#4. Table~\ref{tab:conf-all-k} shows that MRR is improved in higher rates compared to HR, i.e., the method does not find more of the relevant items but ranks them better.
Figure~\ref{fig:conf-effects} shows noticeable differences between configurations which convey high-level concepts given by only \textit{nmods} (\#1, \#2) and the other configurations that use also \textit{amods} (\#3, \#4).
Configurations \#1 and \#2 provide consistently higher MRR@1 uplift than \#3 and \#4.
We believe it is because the former omit low-level features (qualities such as colors etc.) that can be derived from the item-level dependencies but cause undesirable merging on the level of semantics.
Just as is the case of \#3, if conceptually unrelated items are merged because of their identical appearance quality aspects, clusters must be large enough to create coherent sub-sequences in the browsing behavior.
Although large clusters are easier to predict since their quantity declines with the growing size, the trade-off is unfavorable in this setup because the larger clusters the lower potential for the performance uplift.
Configurations \#3 and \#4 display MRR growth (Fig.~\ref{fig:conf-effects}a,b) stopping at the point where clusters become too large and their potential consequently starts to decrease significantly (Fig.~\ref{fig:conf-effects}c).
The last semantics scheme improves performance in all cases. It benefits from the length of intrasession coherence, however, results show a strong negative correlation between the coherence and the performance uplift for \#1 and \#2; and weak positive correlation for \#3 and \#4.
This means that conceptual coherence is inherently given by user behavior and increasing it artificially by employing larger semantic clusters is not inevitably beneficial as can be also seen in Table~\ref{tab:conf-results-overall}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{assets/eps_plot.pdf}
\caption{Comparison of generalization effects produced by the combination of different NLP parsing configurations and $\epsilon$ parameter of DBSCAN clustering. Line charts show different values given by the growth of $\epsilon$ for each generalization configuration. Configurations using high-level concepts (\#1 and \#2) have stronger linear correlation of $\epsilon$ with metric uplift in all semantic recommendation schemes.}
\label{fig:conf-effects}
\Description{}
\end{figure}
On the other hand, we noticed many cases of concepts alternating in sequences, e.g, $[a,a,b,a,b,a]$, what incapacitates advantages of the last semantics scheme, therefore, improving coherence by also incorporating information about interactions among the concepts prior to the clustering process could help.
In this scheme, the $c_s(i)$ function accuracy calculated as HR@1 is only $\sim$35\% for the best clustering configurations.
Despite flaws that arise from dynamics, this scheme ensures providing the correct semantic cluster within the coherent window regardless of the cluster's size or popularity which is not the case in the predicted semantics scheme.
Also, the predicted semantics scheme improves performance though in smaller rates. The \mbox{top-1} cluster prediction accuracy is roughly the same as in the last semantics scheme with small fluctuations around $\pm$1-2\%. Nevertheless metrics uplift is in most cases lower, because large clusters that provide lower utility are often predicted since the SkNN models tend to suffer from the popularity bias. Moreover, it might have trouble predicting the same cluster several times in case of a longer coherent semantic window. Therefore, this scheme is less suitable for the next-item prediction but it could be useful in different use case scenarios as discussed in Section~\ref{sec:discussion}.
In our settings, the semantic post-filtering has the potential to double the performance (Fig.~\ref{fig:conf-effects}c). However, the best performing mappings utilized only up to $\sim$20\% of the potential as shown in Table~\ref{tab:conf-results-overall}. Hence, there is still the opportunity for improvement of the semantic factor function $c_s(i)$.
\begin{table}[t]
\caption{The results for the best mapping in each configuration and its characteristics in the means of average cluster size and average length of coherent sub-sequences. The table shows achieved performance uplifts and compares them to the maximum theoretical potential given by the dedicated scheme. Also the rate of the potential utilization is shown. Note that at k=1, MRR and HR are equal.}
\label{tab:conf-results-overall}
\begin{tabular}{lccclcc}
\toprule
\multicolumn{1}{l}{} & Conf. & \thead{Avg. \\ coherence} & \thead{Avg. \\ cluster size} & \thead{MRR@1 \\ HR@1} & \thead{Potential \\ MRR@1} & \thead{Utilized \\ potential} \\
\midrule
baseline && 1.15 & 1 & 3.20\% & - & - \\
\hline
last
& \#1 & 1.77 & 18 & 4.09\% (+27.93\%) & 7.38\% & 21.37\% \\
& \#2 & 1.78 & 19 & \textbf{4.13\% (+29.12\%)} & 7.28\% & 22.82\% \\
& \#3 & 1.63 & 14 & 3.79\% (+18.48\%) & 7.95\% & 12.45\% \\
& \#4 & 2.07 & 19 & 3.81\% (+19.05\%) & 6.40\% & 19.04\% \\
\hline
predicted
& \#1 & 1.78 & 21 & \textbf{3.81\% (+18.98\%)} & 7.35\% & 14.62\% \\
& \#2 & 1.78 & 19 & 3.76\% (+17.62\%) & 7.28\% & 13.81\% \\
& \#3 & 3.54 & 135 & 3.64\% (+13.93\%) & 4.27\% & 41.53\% \\
& \#4 & 3.62 & 319 & 3.66\% (+14.46\%) & 4.08\% & 52.66\% \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[t]
\caption{The results for the best mapping in each configuration. The table shows achieved relative performance uplifts compared to the baseline. The best results are highlighted for each metric and scheme. Note that at k=1, MRR and HR are equal.}
\label{tab:conf-all-k}
\begin{tabular}{lcrrrrr}
\toprule
& Conf. & MRR@1 & MRR@10 & MRR@20 & HR@10 & HR@20 \\
\midrule
baseline & & 3.20\% & 5.53\% & 5.62\% & 10.84\% & 12.16\% \\
\hline
last & \#1 & +27.93\% & +16.56\% & +15.86\% & +4.27\% & +1.14\% \\
& \#2 & \textbf{+29.12\%} & \textbf{+17.19\%} & \textbf{+16.47\%} & \textbf{+4.39\%} & +1.57\% \\
& \#3 & +18.48\% & +9.30\% & +8.92\% & +3.59\% & \textbf{+1.63\%} \\
& \#4 & +19.05\% & +10.28\% & +9.85\% & +3.69\% & \textbf{+1.63\%} \\
\hline
predicted
& \#1 & \textbf{+18.98\%} & \textbf{+11.29\%} & \textbf{+10.71\%} & +3.56\% & +0.84\% \\
& \#2 & +17.62\% & +10.31\% & +9.79\% & +3.18\% & +1.48\% \\
& \#3 & +13.93\% & +8.86\% & +8.50\% & +3.32\% & \textbf{+1.57\%} \\
& \#4 & +14.46\% & +9.36\% & +8.94\% & \textbf{ +3.59\%} & +1.54\% \\
\bottomrule
\end{tabular}
\end{table}
\section{Discussion and conclusions \label{sec:discussion}}
In this paper, we presented a proof-of-concept method that allows better adaptation of the next-item recommendations to the user interest via the semantics derived from product titles. The main contribution is the finding that the item generalization based on semantic analysis appears to provide a high potential for improving the recommendations. We extended a popular session-based k-NN recommendation model to prototype our idea, though it is not limited to any kind of model.
We employed the dependency parsing technique to identify head-dependent relations in a noun phrase describing an item and to extract only the most salient words that convey high-level concepts. The results of our experiments conducted on a fashion e-commerce dataset showed that omitting adjunct words describing low-level quality aspects of items, such as colors, leads to better items ranking in recommendation list when using text features to improve model performance.
The experiments showed that NLP based semantic analysis has significant potential in improving recommendations, from which the large part is not utilized yet. In the future, we plan to design the interest change detection as a binary task that can use behavioral data or other events, such as category view or add to cart, to switch between the last and the predicted semantics schemes. Furthermore, we plan implementation for the neural-network-based method, for which a latent representation can also be envisioned that would allow us to handle headwords and modifiers in a more flexible way.
We believe that the quality of semantic clusters and their positive effects on the performance can be improved even further. We identified several complications emerging from the nature of the text that had a bad influence on dependency parsing. The crucial part is to identify the root headword correctly, which is problematic when the root is a multiword expression, such as \textit{skinny jeans}. Also, the correct root headword may be misleading in some cases, e.g., \textit{pair of socks} would be inconveniently matched with \textit{pair of gloves} because of the common root \textit{pair}. Sometimes redundant words are used, e.g., \textit{woman dress} in a woman category is no different from products that do not contain the word \textit{woman}. The semantic analysis must be able to deal with copywriters' different styles, e.g., \textit{red dress} and \textit{dress in red color} have a different constitution but essentially the same meaning. We did not put effort into resolving these so a more comprehensive semantic analysis might provide a higher improvement.
Although we evaluated the proposed method on the next-item prediction task, it can be also used in different scenarios, e.g., as a shopping cart page recommendation strategy. After a user adds an item to the cart, top-n next-concept prediction can be used to split recommended items into \textit{n} recommendation lists for each concept separately, while we can rely upon that the sequential dependencies from the preceding browsing behavior would be also considered.
\begin{acks}
This work was partially supported by the Slovak Research and Development Agency under the contract No. APVV-15-0508 and the Scientific Grant Agency of the Slovak Republic, grant No. VG 1/0667/18. The authors would like to thank for financial contribution from the STU Grant scheme for Support of Young Researchers.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
| e31e5fc93294535c3d4b84c6a126f0e2f7771f7d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Fabrication and Device Properties}
\subsection{Material Properties}
Our resonators are fabricated out of a 100-$\rm nm$ silicon nitride film that is deposited on top a 550-$\mu \rm m$-thick silicon handle wafer~\cite{Wafer} via low-pressure chemical vapor deposition (LPCVD). The material properties of silicon nitride depend upon the specifics of the LPCVD process. The reported Young's modulus and density values are in the range $250~{\rm GPa} \le E \le 350 ~ {\rm GPa}$ and $2600~ {\rm kg / m^3} \le \rho_s \le 3400 ~ {\rm kg / m^3}$, respectively \cite{Senturia:2002, Drummond:1995, Santucci:2001, Stoffel:1996, Larsen:2013, Fauver:1998, Petersen:1982, Edwards:2004, Yang:2008, Yang:2002, Tabata:1989}. Furthermore, there is an unknown stress in the silicon nitride film, which results in a tension force $S$ in the suspended beams.
In Section \ref{section:Determining Material Properties from Natural Frequencies}, we show that beam theory and the measured eigenfrequencies of the beams allow us to estimate a single set of optimal material properties that can be used to describe all of the devices used in our experiments. We have found that this approach is more insightful than trying to estimate the specific material properties of each beam individually. The values we have found from this analysis are $E=300~ \rm GPa$, $\rho_s=2750~ \rm kg/m^3$, $\sigma = 86.5~ \rm MPa$, and $S= 7.64~ \rm \mu N$, as listed in Tables~\ref{tab_devices} and \ref{fit_parameters}.
\subsection{Fabrication Process}
A flowchart of the fabrication process is given in Fig.~\ref{fig:Supp_fabrication}. Transducers and contact pads are patterned using electron beam lithography (EBL). A 100-nm-thick gold film with a 5-nm chromium adhesion layer underneath is deposited by thermal evaporation. After lift-off, a second step of EBL is performed to define the beam structures. A 60-nm-thick copper film is deposited via electron beam evaporation, which serves as a dry etch mask. Inductively coupled plasma etching is used to etch the silicon nitride layer anisotropically; then, the beam is suspended by an isotropic etch step that removes the silicon beneath. After the removal of the copper etch mask, the devices are placed in a chip carrier with 50-$\Omega$ strip lines and wirebonds are made between the contact pads and the striplines. Figure~\ref{fig:Supp_SEM} shows SEM images of a completed chip and one of the resonators, a 15-$\rm \mu m$-long beam.
\begin{figure}[b]
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_Fabrication.pdf}
\caption{Flowchart for the fabrication process. (a) PMMA resist is spun on the silicon nitride layer, (b) then transducers are patterned using EBL, followed by deposition of gold. (c) After lift-off, the transducers and contact pads are completed. (d) A second step of EBL is used to pattern the nanomechanical structures. (e) To mask the beam structures during dry etch, a copper etch mask is deposited. (f) Silicon nitride is etched anisotropically down towards the silicon layer, and then (g) silicon is etched isotropically to suspend the beams. (h) After the etch mask is removed, the fabrication is complete.}
\label{fig:Supp_fabrication}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_SEM.pdf}
\caption{(a) A $10~\rm mm \times 10 ~\rm mm$ chip that has 16 NEMS devices with lengths 15~${\rm\mu m} \le l \le$ 60~${\rm\mu m}$. (b) Close-up and tilted SEM image of a 15-$\rm \mu m$-long beam. The gap between the suspended beam and the substrate is approximately 2 $\rm \mu m $. (c) Magnified top view of a transducer.}
\label{fig:Supp_SEM}
\end{center}
\end{figure}
\subsection{Device Properties}
In this section, we provide more details on the device dimensions and mechanical properties and parameters. The dimensions of the beams reported in Table~\ref{tab_devices} are determined by imaging them in SEM at high magnification. The in-plane dimensions are found to be close to the design dimensions with small deviations due to fabrication imperfections. The thickness of the silicon nitride beams has more uncertainty and is measured to be $\rm 92.8 \pm 0.8~nm$. In the theoretical calculations, we have used the value $\rm 92.8~nm$. In order to find the mechanical parameters, we measure the first few resonant frequencies by driving the resonators in air; we also measure the thermal vibrations of the fundamental mode of each resonator in air and find the modal masses and spring constants by the equipartition theorem. We observe a close agreement between the mean-squared fluctuation amplitudes in air and water, as expected from the equipartition theorem. This is a good sanity check and proves that the spring constant does not change in water. All the relevant device and material parameters in this work can be found in Tables \ref{tab_devices}, \ref{tab_device_frequencies} and \ref{fit_parameters}.
\begin{table}[b!]
\caption{\label{tab_devices} Device parameters and experimentally-obtained mechanical properties of the devices used in this study. The mean-squared fluctuation amplitudes measured at the centers of the beams in air and water, $\left\langle {{{W}_{air}}^2} \right\rangle$ and $\left\langle {{{W}_{water}}^2} \right\rangle $, respectively, are listed for comparison.}
\begin{ruledtabular}
\begin{tabular}{cccccccccc}
Device & $l \times b \times h$ & $k_1$ & $m_1$ & $\alpha_1$ & $\omega_1 / 2\pi$ & $U$ & $S$ & $\left\langle {{{W}_{air}}^2} \right\rangle $ & $\left\langle {{{W}_{water}}^2} \right\rangle $ \\
& ($\mu \rm m^3$) & (N/m) & (pg) & & (MHz) & & ($\mu \rm $N) & ($ \rm m^2 \times 10^{-21}$) & ($ \rm m^2 \times 10^{-21}$) \\
\hline
$60 \mu \rm m$ & $60 \times 0.95 \times 0.093$ & 0.93 & 10.82 & 0.476 & 1.48 & 720 & 7.64 & 4.36 & 4.41 \\
$50 \mu \rm m$ & $50 \times 0.95 \times 0.093$ & 1.24 & 9.63 & 0.469 & 1.80 & 500 & 7.64 & 3.29 & 3.33 \\
$40 \mu \rm m$ & $40 \times 0.95\times 0.093$ & 1.42 & 6.53 & 0.462 & 2.35 & 320 & 7.64 & 2.86 & 2.75 \\
$30 \mu \rm m$ & $30 \times 0.95\times 0.093$ & 1.66 & 3.33 & 0.451 & 3.55 & 180 & 7.64 & 2.46 & 2.56 \\
$20 \mu \rm m$ & $20 \times 0.95\times 0.093$ & 2.09 & 1.83 & 0.434 & 5.38 & 80 & 7.64 & 1.95 & 1.94 \\
$15 \mu \rm m$ & $15 \times 0.95\times 0.093$ & 3.42 & 1.26 & 0.423 & 8.28 & 45 & 7.64 & 1.19 & 1.18 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\subsubsection{Theoretical Description of the Natural Frequencies and Mode Shapes for a Beam with Tension}
\label{Beam Equation and Eigenmodes}
Due to the pre-stressed nature of the silicon nitride layer, the suspended beams are under tension. The tension causes the dynamics of the beams to deviate from that of a typical Euler-Bernoulli (EB) beam without tension and approach that of a string under tension. The dynamics can be described by adding a tension term to the EB beam equation, which can be used to find the eigenfrequencies and mode shapes~\cite{Bokaian1990,stachiv:2014} in the usual manner.
We will follow the analytical approach provided by Bokaian~\cite{Bokaian1990} to describe the dynamics of a doubly clamped-beam under tension in the absence of a fluid. The result of this analysis will be the eigenfrequencies (natural frequencies) $\omega_n$ and mode shapes $\phi_n$ of the beam. While we only need the mechanical parameters of the fundamental mode for the fluid experiments, we have used the higher mode frequencies to determine the material properties of silicon nitride, as we describe in Section \ref{section:Determining Material Properties from Natural Frequencies}.
Our intention in the following is to provide only the necessary details needed to arrive at the expressions for $\omega_n$ and $\phi_n$, and to clearly establish our notation and conventions. We refer the reader to Ref.~\cite{Bokaian1990} for more details. For a long and slender doubly-clamped beam that is under tension and undergoing small flexural oscillations the dynamics can be described by
\begin{equation}
{\frac{E I}{l^4} \frac{\partial^4 {\cal W}({x^*},t)}{\partial {x^*}^4}} - {\frac{S}{l^2} \frac{\partial^2 { \cal W}({x^*},t)}{\partial {x^*}^2} }+ {\mu \frac{\partial^2 {\cal W}({x^*},t)}{\partial t^2}} = 0.
\label{eq:beam-equation}
\end{equation}
The beam geometry is determined by its length $l$, width $b$, and thickness $h$. A long and slender beam indicates $l \gg b, h$, and small deflections indicate that the maximum beam deflection is much smaller than the thickness of the beam. Both of these conditions are satisfied by the nanomechanical devices that we explore here. In our notation, $x^* = x/l$ is the dimensionless axial coordinate along the beam where $x$ is the dimensional coordinate. In addition, ${\cal W}(x^*,t)$ is the transverse displacement (in the $z$ direction) of the beam at axial position $x^*$ and time $t$; $E$ is the Young's modulus, $I = b h^3/12$ is the area moment of inertia, $\mu = \rho_s b h$ is the mass per unit length of the beam, and $S$ is the applied axial force. The minus sign on the second term indicates that the beam is under tension. The boundary conditions for the doubly clamped beam are ${\cal W}(0,t)={\cal W}(1,t)=\frac{\partial {\cal W}(0,t)}{\partial x^*}=\frac{\partial {\cal W}(1,t)}{\partial x^*}=0$.
In the main text, we focus upon experimental measurements at a specific axial location on the beam ${x_0}^*$. To simplify the notation for this case, we define the time varying oscillations at location ${x_0}^*$ as $W(t) = {{\cal W}}({x_0}^*,t)$, where we have used ${x_0}^* = 1/2$. However, in the following we will keep the discussion general and use ${\cal W}(x^*,t)$.
We next assume that the oscillating solutions are of the form
\begin{equation}
{\cal W}({x^*},t) = {Y_n({x^*})} {e^{i \omega_n t}}
\end{equation}
where $n$ is the mode number, $Y_n({x^*})$ is the $n$\textsuperscript{th} mode shape, and $\omega_n$ is the $n$\textsuperscript{th} natural frequency. Substituting this expression into Eq.~(\ref{eq:beam-equation}) yields
\begin{equation}
{\frac{E I}{l^4} \frac{d^4 Y_n({x^*})}{d {x^*}^4}} - {\frac{S}{l^2} \frac{d^2 Y_n({x^*})}{d {x^*}^2}} - {\mu \omega_n^2 Y_n({x^*})} = 0,
\label{eq:beam-ode}
\end{equation}
whose solution can be expressed as
\begin{equation}
\label{eq:mode_shape}
{Y_n({x^*})} = {c_{1,n} \sinh(M_n {x^*})} + {c_{2,n} \cosh(M_n {x^*})} + {c_{3,n} \sin(N_n {x^*})} + {c_{4,n} \cos(N_n {x^*})},
\end{equation}
which is in terms of the constants $M_n = ( U + \sqrt{U^2 + {\Omega_n}^2})^{1/2}$ and $N_n = ( -U + \sqrt{U^2 + {\Omega_n}^2})^{1/2}$. The amount of tension in the beam is described by the nondimensional tension parameter $U$, where
\begin{equation} \label{eq:U-defn}
U= \frac{S}{2 E I/l^2}
\end{equation}
represents the ratio of the axial load on the beam to its rigidity. The nondimensional natural frequency for mode $n$ is $\Omega_n$, where
\begin{equation}
\Omega_n = \frac{\omega_n}{\beta/l^2}
\label{eq:Omega_n}
\end{equation}
and $\beta = (EI/\mu)^{1/2}$ is a constant determined by material properties and linear dimensions.
Boundary conditions, $Y(0) = Y(1) = d Y(0)/d{x^*} = d Y(1)/d{x^*} = 0$, allow for determining the values of the constants as
\begin{eqnarray}
c_{1,n} &=& {1}, \\
c_{2,n} &=& {\frac{M_n \sin (N_n) - N_n \sinh (M_n)}{N_n \left[ \cosh(M_n) - \cos(N_n)\right]}}, \\
c_{3,n} &=& { - \frac{M_n}{N_n}}, \\
c_{4,n} &=& - c_{2,n}.
\end{eqnarray}
The characteristic equation relating $U$ and $\Omega_n$ is
\begin{multline}
{\Omega_n} + {U \sinh\left( U + \sqrt{U^2 + {\Omega_n}^2}\right)^{1/2}} {\sin\left( -U + \sqrt{U^2 + {\Omega_n}^2}\right)^{1/2}} \\ - {\Omega_n \cosh \left( U + \sqrt{U^2 + {\Omega_n}^2}\right)^{1/2}} {\cos \left( -U + \sqrt{U^2 + {\Omega_n}^2}\right)^{1/2}} = {0}.
\label{eq:characteristic-equation}
\end{multline}
Given a value of $U$, all of the natural frequencies $\Omega_n$ are found as solutions of Eq.~(\ref{eq:characteristic-equation}). In practice, we determine the values of $\Omega_n$ numerically as the successive roots of Eq.~(\ref{eq:characteristic-equation}). We will find it convenient to use the normalized mode shapes $\phi_n({x^*}) = Y_n(x^*)/{\tilde{N}_n}^{1/2}$ where $\tilde{N}_n$ is chosen such that
\begin{equation}
{\int_{0}^{1}} {\phi_n({x^*})} {\phi_m({x^*})} {d{x^*}} = {\delta_{nm}}
\end{equation}
with $\delta_{nm}=1$ for $n=m$ and $\delta_{nm}=0$ for $n \ne m$. Note that the orthonormal eigenfunctions $\phi_n$ are dimensionless.
\begin{figure}
\begin{center}
\includegraphics[width=3.375in]{Figures/Supplementary/Supp_5_mode_shapes.pdf}
\end{center}
\caption{The first five normalized mode shapes $\phi_n$ as a function of ${x}/{l}$ for a doubly-clamped beam under tension. For this figure we have used $U=320$ which corresponds to the parameters of the 40-$\mu$m-long beam (Table~\ref{tab_devices} and \ref{fit_parameters}).}
\label{fig:modes}
\end{figure}
\subsubsection{Determining Material Properties from Natural Frequencies} \label{section:Determining Material Properties from Natural Frequencies}
Since we do not know the precise values of $S$, $E$ and $\rho_s$, we use an error minimization approach to determine these from theory and experimental observables of the beams in air, namely the eigenfrequencies. Theoretical predictions of the eigenfrequencies in terms of $\Omega_n$ and $U$ are given by Eq.~(\ref{eq:characteristic-equation}) where $\Omega_n$ depends on $\rho_s$ (Eq.~(\ref{eq:Omega_n})) and $U$ depends on $S$ and $E$ (Eq.~(\ref{eq:U-defn})). We take the following approach. We select trial values for $E$ and $\rho_s$. Using the linear dimensions of the beam, the measured $\omega_1$, and the trial values of $E$ and $\rho_s$, we compute the tension $S$ in the beam by numerically solving Eq.~(\ref{eq:characteristic-equation}). We repeat this for all the fundamental frequencies of all the beams and find average $S$ and $U$ values. Once $U$ is determined for these trial values of $E$ and $\rho_s$, we find all of the remaining eigenfrequencies as the successive roots of Eq.~(\ref{eq:characteristic-equation}). Then, we form an error function defined as \begin{equation}
\varepsilon (E, \rho_s) = \sum\limits_k \frac{|\omega_{k}^{(e)} - \omega_{k}^{(t)}|^2}{|\omega_{k}^{(e)}|^2},
\end{equation}
where $k$ runs through all (available) modes of all the beams, and $\omega_{k}^{(e)}$ and $\omega_{k}^{(t)}$ are the experimental and theoretical eigenfrequencies. We repeat this process and calculate $\varepsilon$ by sweeping the $\rho_s$-$E$ plane within the range of acceptable values from the literature which are $2600~ {\rm kg / m^3} \le \rho_s \le 3400 ~ {\rm kg / m^3}$ and $250~{\rm GPa} \le E \le 350 ~ {\rm GPa}$~\cite{Senturia:2002, Drummond:1995, Santucci:2001, Stoffel:1996, Larsen:2013, Fauver:1998, Petersen:1982, Edwards:2004, Yang:2008, Yang:2002, Tabata:1989}. We find that $E= 300 ~\rm GPa$ and $\rho_s=2750~\rm kg/m^3$ values provide a global minimum for the function $\varepsilon (E, \rho_s)$ for the average tension value $S\approx 7.64 ~\rm \mu N$. This tension value corresponds to a stress of $\sigma \approx 86.5 ~\rm MPa$. These values thus provide useful approximations for the material proprieties of our silicon nitride sample. Table~\ref{table:frequencies} lists experimentally measured eigenfrequencies and the theoretical eigenfrequencies (given in parentheses). The theoretical predictions of the frequencies are determined using the optimal values of $E$, $\rho_s$ and the average value of $S$.
The tension in the beams manifests itself in several ways. First, tension increases the natural frequencies of a beam resonator. Second, the relation between the resonance frequencies deviate from that of EB beam. Figure~\ref{fig:Supp_Tension}(a) shows the ratio of $\omega_n/\omega_1$ for our beams as well as the two limiting cases. Orange symbols correspond to an EB beam where $\omega_n/ \omega_1\approx 1,2.75,5.40,\ldots$ for $n=1,2,3,\ldots$. The black symbols correspond to a string under tension where $\omega_n/\omega_1 = n$. The ratios $\omega_n/\omega_1$ for all of our experimentally measured beams lie in between these two limits and are shown by the different open symbols. This behavior is due to the fact that $U$ depends quadratically on the beam length as shown in Eq.~(\ref{eq:U-defn}).
\begin{table}[b]
\caption{\label{tab_device_frequencies} Eigenfrequencies (natural frequencies) of the fundamental and higher modes of the devices used in this study. Values in parentheses are the theoretical estimations using the optimal material properties.}
\begin{ruledtabular}
\begin{tabular}{cccccccccc}
Device & $\omega_1 / 2\pi$ & $\omega_2 / 2\pi$ & $\omega_3 / 2\pi$ & $\omega_4 / 2\pi$ & $\omega_5 / 2\pi$ \\
& (MHz) & (MHz) & (MHz) & (MHz) & (MHz) \\
\hline
$60 \mu \rm m$ & 1.48 (1.56) & 2.93 (3.16) & 4.46 (4.82) & 6.10 (6.57) & 7.82 (8.44) \\
$50 \mu \rm m$ & 1.80 (1.90) & 3.78 (3.86) & 5.78 (5.92) & 7.89 (8.15) & 10.14 (10.56) \\
$40 \mu \rm m$ & 2.35 (2.42) & 4.66 (4.96) & 7.14 (7.70) & 9.94 (10.74) & 13.04 (14.13) \\
$30 \mu \rm m$ & 3.55 (3.35) & - (6.96) & - (11.04) & - (15.77) & - (21.26) \\
$20 \mu \rm m$ & 5.38 (5.42) & 11.68 (11.68) & 18.67 (19.42) & - (28.94) & - (53.97) \\
$15 \mu \rm m$ & 8.28 (7.85) & - (17.61) & - (30.39) & - (46.67) & - (66.64) \\
\end{tabular}
\end{ruledtabular}
\label{table:frequencies}
\end{table}
Finally, the modeshapes also depend on $S$, but only weakly. The first five normalized mode shapes calculated for a 40-$\rm \mu m$-long beam are shown in Fig.~\ref{fig:modes}. We note that it is experimentally not possible for us to measure the modeshapes with enough precision such that we can distinguish between an EB beam and a string under tension. All of the relevant parameters used in finding the mode shapes are also given in Table~\ref{fit_parameters}.
\begin{figure}
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_properties.pdf} \\
\caption{(a) Ratio of the $n^{\rm th}$ mode frequency to the fundamental mode frequency, $\omega_n/\omega_1$ for four of the devices used in the experiments. Theoretical ratios for an Euler-Bernoulli beam without tension (orange) and a string under tension (black) are also shown. (b) Fundamental mode frequencies $f_1=\omega_1/2\pi$ and (c) effective masses $m_1$ for all the devices as a function of their lengths.}
\label{fig:Supp_Tension}
\end{center}
\end{figure}
\subsubsection{The Spring Constant of the Fundamental Eigenmode} \label{eigen}
In the main text, we describe how the spring constant of the fundamental mode $k_1$ for each beam is determined using the classical equipartition theorem. Figure~\ref{fig:Supp_Noise_integrate}(a) shows an example of how we calculate the mean-squared thermal displacement for the fundamental mode of a 60-$\rm \mu m$ device from the power spectal density (PSD) of its Brownian noise in air. Since the thermal resonance is sharply peaked, we can find the area under the curve by integration, which provides the mean-squared displacement $\langle {W_{air}}^2 \rangle$. In Fig.~\ref{fig:Supp_Noise_integrate}(a), this area is the shaded region under the curve (red). Then, using the classical equipartition theorem, we find the spring constant as $k_1 = \frac { k_B T} {\langle {W_{air}}^2 \rangle}$, where $k_B$ is the Boltzmann constant and $T$ is the temperature. The variation of the measured spring constants $k_1$ with length $l$ for our devices are shown in Fig~\ref{fig:Supp_Noise_integrate}(b) by the data points. In addition, the values for $k_1$ are listed in Table~\ref{tab_devices}.
As a comparison, the static spring constant of a doubly-clamped beam without tension when a point load is applied perpendicularly at the center of the beam, $x=l/2$, is given by~\cite{Cleland2003}
\begin{equation}
k_0 = \frac{192 E I}{l^3}.
\label{elastic}
\end{equation}
When the beam is under significant stress, Eq.~(\ref{elastic}) becomes less accurate, since the axial load must be taken into account. Lachut \textit{et al.}~\cite{Lachut2012} formulated a method to calculate the spring constant of a beam under stress, taking into account the axial load. For a beam with stress of $\sigma_s$, the ratio of the increase $\delta k$ in the spring constant to the stress-free spring constant $k_0$ is given by
\begin{equation}
\label{lachut}
{\frac{\delta k}{k_0} }= {\frac{3}{10}}{ \frac{\left(1 - \nu_{p}\right) \sigma_s}{E h}} {\bigg(\frac{l}{h}\bigg)^2},
\end{equation}
where $\nu_p$ is the Poisson ratio. The stress $\sigma_s$ is the result of the axial load $S$, where $S= (1 - \nu_p)\sigma_s b $. The spring constant is then given by $k=k_0 + \delta k$. Figure~\ref{fig:Supp_Noise_integrate}(b) includes the predicted spring constants using Eq.~(\ref{elastic}) (blue, dotted line) and using Eq.~(\ref{lachut}) (red, solid line), using the linear dimensions, material properties, and tension values in Table~\ref{fit_parameters} along with $\nu_p =0.28$.
\begin{figure}
\begin{center}
\includegraphics[width=4.5 in]{Figures/Supplementary/Supp_equi.pdf}
\caption{(a)~Thermal noise peak of a 60-$\rm \mu m$-long nanomechanical beam resonator in air. The shaded area shows the integral of the PSD and is used in the equipartition theorem to calculate the spring constant of the device. (b) Experimentally measured spring constants $k_1$ as a function of the device length $l$ (symbols). The blue dotted curve is the theoretical prediction in the absence of tension given by Eq.~(\ref{elastic}). The red solid line includes tension (Eq.~(\ref{lachut})).}
\label{fig:Supp_Noise_integrate}
\end{center}
\end{figure}
\subsubsection{The Effective Mass of the Fundamental Eigenmode}
\label{subsection:effective_mass}
The effective mass $m_1$ of the fundamental mode of a beam can be found using $m_1=k_1/ {\omega_1}^2$. In modal analysis, the effective mass (and the spring constant) can be defined only up to a constant. However, since we experimentally measure the resonance frequencies $\omega_1$ and determine the spring constants $k_1$ using the equipartition theorem, there is no ambiguity in our values, and we are able to present quantitative measurements for $m_1$. Figure~ \ref{fig:Supp_Tension}(c) shows the effective mass $m_1$ as a function of the beam length $l$. In addition, the numerical values of measured $m_1$ are listed in Table~\ref{tab_devices}.
In our experiments we use a point measurement at the optical spot to quantify the beam dynamics. In the following, we describe how this point measurement is connected with our lumped single-degree-of-freedom model. We equate the kinetic energy of the fundamental mode as measured at position $x=x_0$ of the beam to the kinetic energy of the entire beam with physical mass $m=\rho_s l b h$ that is oscillating at frequency $\omega_1$ as~(cf.~\cite{Cleland2003,Cross2006}). This yields
\begin{equation}
\label{eff_mass}
\frac{1}{2} m_1 \dot{{\cal W}}(x_0,t)^2 = \frac{1}{2} {\int_{0}^{l}} \mu \dot {\cal W}(x,t)^2 dx,
\end{equation}
where $m_1$ is the effective mass of the fundamental mode when measured at position $x_0$. If we assume $\mu = m/l$ is constant and describe the beam motion using the fundamental mode shape as ${\cal W}(x,t) = a_1 Y_1(x) e^{i \omega_1 t}$ where $a_1$ is an arbitrary constant determining the amplitude of the motion, this expression can be simplified to
\begin{equation}
\frac{m_1}{m} = \frac{1}{ Y_1(x_0)^2 l} {\int_{0}^{l}} Y_1(x)^2 dx,
\end{equation}
which is equivalent to
\begin{equation}
\frac{m_1}{m} = \frac{1}{ Y_1({x_0}^*)^2} {\int_{0}^{1}} Y_1(x^*)^2 dx^*,
\end{equation}
where we now use the nondimensional coordinate $x^*$ and ${x_0}^* = x_0/l$. Using the normalized dimensionless mode shape $\phi_1$, this can be written as
\begin{equation}
\frac{m_1}{m} = \frac{1}{\phi_1({x_0}^*)^2} {\int_{0}^{1}} \phi_1(x^*)^2 dx^*.
\end{equation}
Lastly, the integral is unity because of the orthogonality relationship for $\phi_n$, which gives the final result
\begin{equation}
\frac{m_1}{m} = \frac{1}{ \phi_1({x_0}^*)^2}.
\label{eq:moverm}
\end{equation}
In the following, we will define $\alpha_1 = 1/\phi_1({x_0}^*)^2$ which yields the following useful relationship between the effective mass of the fundamental mode when measured at location ${x_0}^*$ and the actual mass of the beam: $m_1 = \alpha_1 m$.
Since all of our experimental measurements are at the center of the beams, we will always consider $\alpha_1$ to be determined at ${x_0}^*=1/2$. For an Euler-Bernoulli doubly-clamped beam without tension, $\alpha_1 = 0.3965$, and for a string under tension, $\alpha_1 = 1/2$. For the devices used in this work, $0.423\le \alpha_1 \le 0.476$, corresponding to the shortest and longest beams, respectively. The numerical values for $\alpha_1$ for all our beams are provided in Table~\ref{fit_parameters}. These $\alpha_1$ values are calculated using the fundamental mode shapes of the beam with tension (Section \ref{Beam Equation and Eigenmodes}) using the tension value $S= 7.64~\rm \mu N$.
We can theoretically predict a value of ${m_1}$ using ${m_1} = \alpha_1 m$, which requires the mode shape, beam geometry, and beam material properties. This theoretical value is also included in Table~\ref{fit_parameters} as the calculated effective mass. The difference between the experimentally measured mass and the calculated effective mass increases with beam length and is due to our limited knowledge of the precise device geometry and material properties. For example, the calculated effective mass is an estimation based on device dimensions and does not account for fabrication imperfections nor additional factors such as the masses of the metal films and the contributions of the overhangs at the base of the anchors (see Figure~\ref{fig:Supp_SEM}(b)-(c)).
\section{Measurements}
\subsection{Experimental Setup and Protocols}
\subsubsection{Optical Interferometer}
A homodyne optical interferometer with path stabilization is used for detecting the nanomechanical motion of the beams both in noise and driven measurements. A diagram of the optical setup can be found in Figure~\ref{fig:Supp_Setup}. A HeNe laser with a wavelength of $\lambda = 632.8~\rm nm$ and peak power of 1 mW is operated in its intensity stabilization mode. One photodetector (Thorlabs PDA8A) is used for path stabilization while a second photodetector (New Focus 1801) is used to detect the oscillations of the beams. The displacement sensitivity of the interferometer is estimated to be $\sim25~\rm fm /\sqrt Hz$ in the range 1-50 MHz with {40 $\mu \rm W$} incident on the photodetector. In the noise measurements, we can further improve the displacement resolution to $\sim5~\rm fm /\sqrt Hz$ by averaging and numerically subtracting the background noise from the nanomechanical noise --- as discussed below in Section \ref{Noise_measurements}. The optical spot is $\sim800$ nm in diameter, and the typical power incident on a beam during a measurement is $~20 ~\mu \rm W$.
\begin{figure}
\begin{center}
\includegraphics[width=4.5in]{Figures/Supplementary/Supp_Setup.pdf}
\caption{Optical detection setup used for the experiments. HWP: half wave plate, QWP: quarter wave plate, PBS: polarized beam splitter, PD: Photodetector, PID: Proportional-integral-derivative controller. For thermal fluctuation measurements, the detection PD is connected to a spectrum analyzer; for driven measurements, it is connected to a lock-in amplifier. The second PD with a PID controller and a movable mirror is used for path stabilization.}
\label{fig:Supp_Setup}
\end{center}
\end{figure}
\subsubsection{Experimental Protocols}
The devices are glued to a chip carrier; the measurements are performed with the chip carrier placed on a piezoelectric $XYZ$ stage. The laser, shown in Fig.~\ref{fig:Supp_Setup}, is focused on a beam at the measurement location and the $XYZ$ stage is moved along the $x$ axis for the measurements. The beam length is much larger than the optical spot radius, and the measurements are essentially at a single point on the beam. For liquid experiments, we fill the chip carrier with the liquid (deionized water or isopropyl alcohol) and then seal the carrier with a thin glass lid to prevent evaporation. We note that the depth of the liquid bath is much larger than any device dimensions. The liquid inside the cavity does not show any visible signs of evaporation even after several days.
\subsection{Noise Measurements} \label{Noise_measurements}
\begin{figure}[b!]
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_Noise.pdf}
\caption{(a) Background subtraction on noise data of a 50-$\rm \mu m$ beam in water. The background (black) is subtracted from the measured nanomechanical noise (green) to find the subtracted noise (red). This subtracted noise is then converted to displacement noise by calibrating against the wavelength of the laser. Measurement bandwidth here is 20 $\rm kHz$. (b) A color-map of the noise response of the of 50-$\rm \mu m$ beam immersed in water, showing the noise amplitude as a function of frequency and location of the measurement on the beam. In order to construct the color-map, we have collected noise measurements, such as those in (a), along the beam length at equal steps, starting from one anchor to the other. The data are interpolated using Matlab's surf function. The color-bar corresponds to the root-mean squared displacement amplitude (within the 20 kHz bandwidth) in units of $\rm pm$. (c) Experimentally measured mode shapes of the first three out-of plane modes for the 50-$\rm \mu m$ beam. These are cross-sections from the color plot along the beam position at frequencies of 0.23 MHz, 0.63 MHz, and 1.41 MHz.}
\label{fig:Supp_Noise}
\end{center}
\end{figure}
In noise measurements, a spectrum analyzer is used to detect the optical signals on the photodetector. We numerically subtract the background signal from the measurement signal to remove the low-frequency laser noise. Figure~\ref{fig:Supp_Noise}(a) shows an example of this background subtraction process. Brownian motion signals are recorded by focusing the laser on a location of interest on the nanomechanical beam (green trace in Fig.~\ref{fig:Supp_Noise}(a)). We average $\sim10^4$ data traces for each measurement at a $20$-kHz resolution bandwidth. The measurements typically take several hours each. We then repeat the procedure with identical experimental parameters on the anchor of the beam on top of the silicon nitride layer to obtain the background signal (black trace in Fig.~\ref{fig:Supp_Noise}(a)). The assumption is that noise is uncorrelated and therefore noise powers can be added (or subtracted). The subtracted signal is then calibrated using the laser wavelength to find the PSD in units of $\rm m^2/Hz$.
The frequency and position dependent thermal response of a 50-$\rm \mu m$-long nanomechanical beam is shown in Fig.~\ref{fig:Supp_Noise}(b). For this color-map in water, the PSD of the thermal fluctuations of the beam is measured along the beam length at equal intervals by moving the $XYZ$ stage. The root-mean square (rms) amplitude at each point is found by multiplying the PSD with the resolution bandwidth and then taking the square root. The color-map is then generated using MATLAB's surf function. Figure~\ref{fig:Supp_Noise}(c) shows the measured mode shapes of the first three thermal modes of the beam in water along the beam length. These are essentially slices along the frequency axis in the color-map of Fig.~\ref{fig:Supp_Noise}(b) taken at frequencies of 0.23 MHz, 0.63 MHz, and 1.41 MHz. While noisy, the measured modeshapes match the eigenmodes of the undamped beam shown in Fig.~\ref{fig:modes} closely. We note that the displacement noise measured in the experiment is reported as rms and always remains positive, in contrast to the theoretical mode shape.
\subsection{Electrical Properties of the Electrothermal Actuators}
Electrothermal actuators are used to excite the harmonic response and measure the susceptibility of the nanomchanical beam resonators. Bargatin \textit{et al.} \cite{Bargatin2007a} showed that the thermal time constant for such an actuator is on the order of ${\tau_{th}}^{-1} = {2\pi\times40~ \rm MHz}$. If one drives the actuators at shorter time scales than ${\tau_{th}}$, the actuation does not work efficiently. In our experiments the maximum drive frequency is 5 $\rm MHz$, well below this thermal cut-off frequency. Moreover, the above-mentioned time constant was determined in vacuum where the heat transfer from the actuator is dominated by conduction through the solid. In our experiments, the devices are immersed in a liquid, where the medium enhances the heat transfer and facilitates more rapid cooling.
The electrothermal actuators are designed so that their resistances are close to $50~\Omega$. This ensures that the actuation power is dissipated in the resistor, making the actuation process less susceptible to parasitic circuit elements. In order to confirm that the actuation power (and hence the force) remains constant over the frequency range of our experiments, we have measured the voltage standing wave ratio (VSWR) of the electrothermal actuators. The VSWR is typically expressed as \cite{Joines2012}
\begin{equation}
{\rm VSWR} =\frac{1+\left | S_{11} \right|}{1-\left |S_{11} \right |},
\end{equation}
where $S_{11}$ is the reflection coefficient and quantifies the power reflected from the load. The actuator electrical resistances throughout the devices are around $20\pm 5 ~\Omega$, with the uncertainty coming from fabrication imperfections and design parameters. Figure~\ref{fig:Supp_VSWR}(a) shows the measured VSWR for an electrothermal transducer that is $24.4~\Omega$ and a $50~\Omega$ resistor for comparison, in the frequency range ($0.1-5~ \rm MHz$) of all the experiments. The VSWR drops from $\approx 2.19$ to $\approx 2.14$ over this frequency range, which corresponds to less than a 1\% change in the reflected power. Thus, the power dissipated on the transducer is, for all practical purposes, constant.
\begin{figure}
\begin{center}
\includegraphics[width=4.5in]{Figures/Supplementary/Supp_VSWR.pdf}
\caption{(a) VSWR measurements for a 24.4 $\Omega$ electrode and a 50 $\Omega$ reference resistance in the range of $0.1-5~ \rm MHz$ plotted in a semi-logarithmic graph. (b) Close up of the transducer VSWR. The VSWR decreases from 2.19 to 2.14 in the measurement range which corresponds to a $\le 1\%$ change in the dissipated power.}
\label{fig:Supp_VSWR}
\end{center}
\end{figure}
The two-dimensional sheet resistance of the gold transducer layer is estimated to be $ R_s = 0.24~ \rm \Omega/ \square$, and the resistance of the u-shaped actuator wire that sits close to the anchor of the beam is calculated as $\approx 3.5~\Omega$. A close-up top view of the actuator is shown in Fig.~\ref{fig:Supp_SEM}(c). The parasitic capacitance of the actuators have been estimated from other measurements of similar chips to be around $\sim 100 ~\rm pF$.
\subsection{Driven Measurements}
In the driven measurements, a signal generator drives the transducers with a sinusoidal voltage waveform. The frequency of the sinusoidal signal is swept while the amplitude is kept constant. To detect the signal, a lock-in amplifier is used in $2f$ mode. The power dissipated on a resistor due to an applied oscillating voltage, $V_0 \cos \left({\frac {\omega_0t}{2}}\right)$, is proportional to $\frac{V_0^2}{2} + \frac{ V_0^2}{2} \cos \omega_0 t $, where the constants $V_0$ and $\omega_0$ are the drive voltage and angular frequency, respectively. The first term in the power dissipation expression causes an increase in the average temperature of the resistor and the second term causes the temperature to oscillate at twice the drive frequency. The temperature oscillations result in nanomechanical oscillations, as described in \cite{Bargatin2007a}.
In the main text, we have calibrated the transducer in water using the response of the resonator at low frequency. This data set in water is shown in Fig.~\ref{fig:Supp_drive}(a) as well as another data set that is taken on the same resonator in air. In air, the drive data are calibrated using the amplitude at resonance --- taking advantage of the sharp resonance peak (see Fig.~1(b) in the main text). The amplitude of the mode at the resonance frequency $\omega=\omega_1$ can be expressed as ${{W}_{\omega_1} } = F_0 Q_1/k_1$. Here, the quality factor $Q_1$ of the peak has been found to be $Q_1 \approx 15$ from fitting the resonance lineshape to a Lorentzian. With $k_1$, $Q_1$ and ${{W}_{\omega_1} }$ available from experiments, we determine the lumped force $F_0$ acting on this mode for each drive voltage $V_0$. The results are shown in Fig.~\ref{fig:Supp_drive}(a).
\begin{figure}
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_drive.pdf}
\caption{(a)~Calculated drive force amplitude $F_0$ \textit{vs.} the drive voltage amplitude $V_0$ for the 60-$\mu \rm m$ beam in air (blue, squares) and in water (black, circles). Air calibration is based on the amplitude at the fundamental resonance frequency, whereas water calibration is based on the amplitude at $\omega \rightarrow 0$. The dashed blue and solid black lines are fits to $F_0 = A{V_0}^2$ with $A=2.06 \times 10^{-8}~\rm N/V^2$ and $A=1.30 \times 10^{-8}~\rm N/V^2$ for air and water, respectively. (b) The experimentally measured mode shape of the fundamental mode of a 50-$\mu \rm m$ beam in water (black, squares) that has been actuated at both ends of the beam. (c)~The experimentally measured mode shape of the fundamental mode of the 20-$\mu \rm m$ beam (violet, circles) driven by only one actuator near $x=0$. In panels (b)-(c), the red solid line is the theoretical mode shape of a string with tension and the blue dotted line is the theoretical mode shape of an Euler-Bernoulli beam with tension.}
\label{fig:Supp_drive}
\end{center}
\end{figure}
Both force calibration data sets in Fig.~\ref{fig:Supp_drive}(a) depend on the square of the applied voltage, as expected. The efficiency of the transducer is slightly higher in air, i.e., the anchors heat up to a higher temperature in air compared to water. This is reasonable because the thermal conductivity is higher in water. However, given that that there is no dramatic increase in transducer efficiency going from water to air, we conclude that the thermal conduction in these structures mostly takes place through the solid.
Even though the structures are heating up to approximately the same temperature in air and water, the oscillation amplitudes in water are significantly smaller than those in air because of the viscous damping. We also note that the damping is larger for longer beams. In order to alleviate this, the longer beams (40, 50, and 60~$\mu m$) are driven using both transducers on the two ends of the beam as shown in Fig.~\ref{fig:Supp_SEM}(b) \cite{Bello:2020}. Figure~\ref{fig:Supp_drive}(b) shows the fundamental mode shape of a 50-$\rm \mu m$-long beam measured when both actuators are driving the nanomechanical motion. Here, the mode shape appears symmetric and close to the theoretical predictions. When only one transducer is used, the mode shape becomes slightly distorted toward the end of the beam where the drive transducer is located. An example of a distorted fundamental mode shape of on a 20-$\rm \mu m $ beam is shown in Fig.~\ref{fig:Supp_drive}(c). This distortion is due to fact that the waves generated by the single transducer decay along the length of the beam because of the high viscosity before they can form the familiar standing wave patterns. While this mode shape distortion is not significant for shorter beams, we anticipate it to be a source of error when comparing the theory and experimental results. We discuss this and other sources of errors further in Section~\ref{section_three}.
\subsection{ Measurements in Isopropyl Alcohol (IPA)}
In addition to water measurements, we have measured the Brownian force PSD for a 40-$\mu \rm m$ beam immersed in isopropyl alcohol (IPA) as shown in Fig.~\ref{fig:IPA}(a). IPA is more viscous than water ($\nu_f = 2.78 \times 10^{-6}~ \rm m^2/s $) but has a lower density ($\rho_f =786~ \rm kg/m^3$). Thus, the mechanical response of a given resonator is more damped in IPA as compared to its response in water. This can be seen by comparing Fig.~\ref{fig:IPA}(a) with Fig. 2(c) in the main text. The experimental and theoretical curves for the PSD of the Brownian force $G_F$ are shown in Fig. \ref{fig:IPA}(b). The theoretical estimations are found using Eqs.~(\ref{G_d_supp}) and~(\ref{chi_d_supp}) which are discussed further below. All of the necessary fit parameters are included in Table~\ref{fit_parameters}.
\begin{figure}
\begin{center}
\includegraphics[width=4.5in]{Figures/Supplementary/Supp_IPA.pdf}
\end{center}
\caption{(a) Amplitude-squared susceptibility $\left|{\hat \chi (\omega )} \right|^2$ (black) and PSD of the Brownian fluctuations (blue) for a 40-$\mu \rm m$ beam in IPA. Both quantities are measured at the center of the beam. The symbols show the experimental data, and the continuous lines are predictions of the single-mode theory with no free parameters. (b) PSD of the Brownian force acting on the beam in IPA. The continuous lines show the normalized experimental thermal (blue) and driven (black) responses for each beam from (a) as well as the theoretical PSD of the Brownian force (black).}
\label{fig:IPA}
\end{figure}
\subsection{Mass Loading and Frequency Shift}
In the fluid-structure interaction problem discussed below in Section \ref{subsection:cylinder_approximation}, one obtains an in-phase solution that couples to the mass term and gives mass loading. This added mass depends upon the oscillation frequency and decreases the frequency of the peak in thermal and driven measurements in fluids. In Fig.~\ref{fig:omega_water}, we show our measurements of the peak frequencies of different beams and modes in water, all obtained from driven responses. The frequency of each peak in water, $\omega_{nf}$, is normalized by the frequency of the same peak in air, $\omega_n$. The data are presented as a function of the frequency-dependent Reynolds number ${\rm Re}_\omega$, where ${\rm Re}_\omega$ is based on $\omega_{nf}$ (see Eq.~(\ref{Reynolds}) below). Because all these beams have the same width, ${\rm Re}_\omega$ is simply proportional to $\omega_{nf}$. In the region $0<{\rm Re}_\omega \le 4$, the data plot allows us to estimate any unknown $\omega_{nf}$ accurately.
\begin{figure}
\begin{center}
\includegraphics[width=3.375in]{Figures/Supplementary/Supp_omega_water.pdf}
\end{center}
\caption{ Ratio of the $n^{\rm th}$ mode frequency in water ($\omega_{nf}$) to that in air ($\omega_{n}$) for each beam as a function of frequency-dependent Reynolds number, where ${\rm Re}_\omega= \frac{\omega_{nf}b^2}{4\nu_f}$.}
\label{fig:omega_water}
\end{figure}
\section{Theory}
\label{section_three}
In this section, we briefly derive the expressions that describe the dynamics of the beam with tension in a viscous fluid. We provide our results in terms of a lumped single-degree-of-freedom description. The theoretical description discussed in detail in Refs.~\cite{Cross2006,Clark2010} can be directly applied to our case with only a few modifications. In essence, the type of beam and the presence of tension affect the mode shape and the frequency of oscillation. Once the mode shape is known, it is possible to determine the fundamental natural frequency $\omega_1$ and the parameter $\alpha_1$ which can be used to determine the effective mass $m_1$ and spring constant $k_1$. With $\omega_1$, $m_1$ and $k_1$ determined, the previously-derived expressions~\cite{Cross2006,Clark2010} can be used to obtain the PSD of the displacement fluctuations of the beam and the susceptibility.
\subsection{A Theoretical Description of the Fundamental Mode}
\label{1DOF}
A point measurement made at a specific location on the beam can be described as a lumped single-degree-of-freedom system. In our work, we have made measurements in the middle ($x=l/2$) of the beam. In this description, the nanomechanical beam in its fundamental mode probed at its center can be described as a particle of effective mass $m_1$ with modal displacement (or amplitude) ${W}(t)$ that is attached to a linear spring of stiffness $k_1$. Furthermore, the beam is exposed to an externally-applied driving force $F_d(t)$ and to a force due to the surrounding viscous fluid $F_f(t)$. This can be expressed as
\begin{equation}
m_1\ddot{W} + k_1 W = F_f(t) + F_d(t),
\label{eqm}
\end{equation}
where $k_1=m_1 {\omega_1}^2$. In order to compute the susceptibility $\hat \chi(\omega)$, we use an unit impulse for the drive force $F_d(t) = \delta(t)$ and transform into Fourier space to obtain
\begin{equation}
- m_1 \omega^2 \hat{W}(\omega) + k_1 \hat{W}(\omega) = \hat{F}_f(\omega) + 1,
\end{equation}
where we have used the conventions
\begin{equation}
{{W} (t)} = {{1} \over {2 \pi} }\int_{-\infty}^{\infty} \hat{W}(\omega)e^{-i \omega t} d\omega,
\label{Fourier1}
\end{equation}
\begin{equation}
{\hat{W}(\omega)} = {\int_{-\infty}^{\infty}} {{W}(t)e^{i \omega t}} {dt}.
\label{Fourier2}
\end{equation}
If the fluid force is separated into its real and imaginary components, these will contribute to the mass and damping terms, respectively, which we can represent as
\begin{equation}
-{m_f(\omega)\omega^2\hat{W}(\omega)} - {i\omega\gamma_f(\omega)\hat{W}(\omega)} + {k_1\hat{W}(\omega)} = 1,
\label{eqm_Fourier}
\end{equation}
where $m_f(\omega)$ is the frequency-dependent mass that includes the contribution from the fluid and $\gamma_f(\omega)$ is the frequency-dependent damping due to the fluid. Specific expressions for the oscillating cylinder and oscillating blade approximations are discussed in Section~\ref{subsection:cylinder_approximation}. Solving for $\hat{W}(\omega)$ and noting that this is the Fourier transform of the susceptibility, we have the desired result
\begin{equation}
\hat{\chi}(\omega) = \frac{1}{k_1 - m_f(\omega )\omega ^2 - i\omega \gamma _f(\omega )}.
\label{x_w}
\end{equation}
The squared magnitude of the susceptibility is then
\begin{equation}
|\hat{\chi}(\omega)|^2 = \frac{1}{\left[k_1 - m_f(\omega )\omega ^2\right]^2 + \left[\omega \gamma _f(\omega)\right]^2}.
\label{xw2}
\end{equation}
The PSD of the beam fluctuations when driven by Brownian motion is directly related to the susceptibility by~\cite{Cross2006}
\begin{equation}
G_{W}(\omega) = \frac{4 k_B T}{\omega} {{\hat \chi}''}(\omega),
\end{equation}
where ${{\hat \chi}''}(\omega) $ is the imaginary component of $\hat{\chi}(\omega)$. This yields
\begin{equation}
G_{W} (\omega) = \frac{4 k_B T \gamma_f}{(-m_f \omega^2 + k)^2 + (\omega \gamma_f)^2}.
\label{eq:gw}
\end{equation}
By the the fluctuation-dissipation theorem, the PSD of the force noise (Brownian force) driving the beam can be expressed as
\begin{equation}
G_F(\omega) = 4 k_B T \gamma_f(\omega),
\end{equation}
which results in
\begin{equation}\label{eq:SHO_noise_Gw_GF}
G_{W} (\omega) = \frac{G_F(\omega)}{(-m_f \omega^2 + k_1)^2 + (\omega \gamma_f)^2}.
\end{equation}
Lastly, we note that Eq.~(\ref{eq:SHO_noise_Gw_GF}) can also be expressed as
\begin{equation}
G_{W} (\omega) = |\hat{\chi}(\omega)|^2 G_F(\omega).
\end{equation}
Therefore, for the single mode description we have the result
\begin{equation}
G_F (\omega) = \frac{G_{W}(\omega)}{|\hat{\chi}(\omega)|^2}.
\end{equation}
In the main text, we use the ratio of the experimentally measured $G_W(\omega)$ and $|\hat{\chi}(\omega)|^2$ to explore the spectral properties of $G_F (\omega)$.
\subsection{Treating the Beam as an Oscillating Blade in Fluid}
\label{subsection:cylinder_approximation}
In order to find the frequency dependence of the damping and mass from the last section, we look to fluid dynamics, i.e., we seek to evaluate the fluid force acting on the oscillating beam. We start by making the following assumptions: the fluid is incompressible; the beam is long and thin ($l \gg b, h$); and its displacements are small compared to any of its linear dimensions. With these assumptions, the beam can be approximated as an infinitely long cylinder oscillating transverse to its axis with the same displacement amplitude as the beam. The radius of the cylinder is taken as $b/2$ in order to seamlessly transition back to the beam solution. The mathematical treatment of the oscillating cylinder problem is provided in \cite{Rosenhead}, see also Refs.~\cite{sader:1998,Cross2006} for its application to oscillating beams.
The hydrodynamic force acting on the cylinder is proportional to its oscillatory velocity, ${\omega} \hat{W}$, and can be written as~\cite{sader:1998,Cross2006}
\begin{equation}
{{\hat F}_f}(\omega ) = m_{cyl,e}{\omega ^2} \Gamma_c ({\rm{R}}{{\rm{e}}_\omega}) \hat{W}(\omega ).
\label{eq:fluid_force}
\end{equation}
Here, ${\Gamma_c}({\rm{R}}{{\rm{e}}_\omega })$ is the complex hydrodynamic function for the oscillating cylinder and has the form
\begin{equation}
{\Gamma}_{c}({\rm{Re}_{\omega}}) = {1} + \frac{ {4 i K_1 (-i \sqrt{i \rm{Re_{\omega}}})}}{\sqrt{i \rm{Re_{\omega}}} K_0(- i \sqrt{i \rm{Re_{\omega}}})},
\label{Hydrodynamic}
\end{equation}
where $K_0$ and $K_1$ are zeroth and first order modified Bessel functions of the second kind. The argument of $\Gamma_c$ is expressed in terms of the frequency-dependent Reynolds number
\begin{equation}
{{\rm Re}_{\omega}} = \frac{\omega b^2}{4 \nu_f},
\label{Reynolds}
\end{equation}
where $\nu_f$ is the kinematic viscosity of the fluid and $\rm Re_\omega$ plays the role of a nondimensional frequency in the theory. The mass term $m_{cyl,e}$ is the effective mass of the oscillating cylinder of fluid of diameter $b/2$ where $m_{cyl,e} = \alpha_n m_{cyl}$ with $m_{cyl} = \rho_l \frac{\pi}{4} b^2 l$ the actual mass of the fluid cylinder. The coefficient $\alpha_n$ is described in Section~\ref{subsection:effective_mass}. Since we are focusing on mode 1, we set $n=1$ from here on.
To account for the rectangular cross-section of the beams, we further apply a correction factor to $\Gamma_c(\rm {Re_{\omega}})$ and find the hydrodynamic function of an oscillating blade $\Gamma_b$ immersed in a liquid~\cite{sader:1998}
\begin{equation}
\Gamma_b(\rm {Re_{\omega}}) = {{\tilde{\Omega}} (\rm {Re_{\omega}}) {\Gamma_c}(\rm Re_{\omega})},
\label{correction}
\end{equation}
where $\tilde{\Omega}$ is a complex valued correction factor that is tabulated in Ref.~\cite{sader:1998}.
\begin{figure}
\begin{center}
\includegraphics[width=3.375in]{Figures/Supplementary/Supp_Gf.pdf}
\caption{Dimensionless PSD of the Brownian force as a function of the frequency-based Reynolds number ${\rm Re}_\omega$. Eq.~(\ref{supp_GF_eq}) can be manipulated to show that $G_F$ $\propto {\rm Re}_\omega \Gamma''(\rm {Re_{\omega}})$. Theoretical predictions for the cylinder (red dashed line) and blade (black solid line) descriptions are shown.}
\label{fig:Supp_Gf}
\end{center}
\end{figure}
Separating ${\hat{F}_f(\omega)}$ to its real and imaginary parts, as we did in Eq.~(\ref{eqm_Fourier}), we express the frequency dependent mass and damping as~\cite{Cross2006}
\begin{equation} \label{eq:mass_1}
m_f(\omega) = m_1 + m_{cyl,e} = m_1 + \alpha_1 {\rho_f \frac{\pi}{4} b^2 l {\Gamma}_b'(\rm {Re_{\omega}})}
\end{equation}
and
\begin{equation} \label{eq:damping_1}
{\gamma_f(\omega)} = \alpha_1 \rho_f {\frac{\pi}{4} b^2 l \omega {\Gamma}_b''(\rm {Re_{\omega}})}.
\end{equation}
The mass loading parameter $T_0$ is the ratio of the mass of a cylinder of fluid of diameter $b$ to the actual mass of the beam which can be expressed as
\begin{equation}
T_0 = \frac{m_{cyl}}{m_b} = {\frac{\pi}{4} \frac{\rho_f b}{\rho_s h}},
\label{eq:mass-loading}
\end{equation}
where $\rho_f$ and $\rho_s$ are fluid and solid densities, respectively. Substituting into Eqs.~(\ref{eq:mass_1}) and (\ref{eq:damping_1}), we write
\begin{equation}
\label{m_f}
{m_f}(\omega ) = {m_1}\left( {1 + {T_0}{\Gamma _b'}({\rm{R}}{{\rm{e}}_\omega })} \right)
\end{equation}
and
\begin{equation}
\label{gamma_f}
{\gamma_f(\omega)} = {m_1 T_0 \omega \Gamma_b''(\rm {Re_{\omega}})}.
\end{equation}
The PSD of the Brownian force $G_F$ then becomes
\begin{equation}\label{supp_GF_eq}
G_F(\omega) = 4 k_B T m_1 T_0 \omega \Gamma_b''({\rm Re}_\omega).
\end{equation}
The PSD of the Brownian force is proportional to ${\rm Re_{\omega}} \Gamma_b''(\rm {Re_{\omega}})$ for the cylinder and blade approximations and is shown in Fig.~\ref{fig:Supp_Gf}. Substituting the PSDs of the thermal force noise $G_F$, frequency dependent mass $m_f$, and damping $\gamma_f$ into Eq.~(\ref{eq:gw}) and (\ref{xw2}), we arrive at explicit expressions for the PSD of Brownian fluctuations of the beam
\begin{equation} \label{G_d_supp}
{G_{W}(\omega)} = \frac{4 k_B T m_1 T_0 \omega \Gamma_b''({\rm Re}_\omega)}{{{{\left[ {{k_1} - {m_1}\left( {1 + {T_0}\Gamma_b'({\rm Re}_\omega)} \right){\omega ^2}} \right]}^2} + {{{\omega ^2} \left[ {{m_1}{T_0}\omega \Gamma_b''({\rm Re}_\omega)} \right]}^2}}}
\end{equation}
and for the squared-magnitude of the susceptibility
\begin{equation} \label{chi_d_supp}
{\left| {\hat\chi (\omega )} \right|^2} = \frac{1}{{{{\left[ {{k_1} - {m_1}\left( {1 + {T_0}\Gamma_b'({\rm Re}_\omega)} \right){\omega ^2}} \right]}^2} + {{{\omega ^2} \left[ {{m_1}{T_0}\omega \Gamma_b''({\rm Re}_\omega)} \right]}^2}}}.
\end{equation}
These expressions can be cast in a slightly different form as
\begin{equation}
G_{W}(\omega) = \frac{1}{k_1^2} \frac{ 4 k_B T \alpha_1 m_b T_0 \omega \Gamma_b''(\omega)}{\left[1 - \tilde{\omega}^2 (1 + T_0 \Gamma_b'(\omega)) \right]^2 + \left[\tilde{\omega}^2 T_0 \Gamma_b''(\omega)\right]^2}.
\end{equation}
and
\begin{equation}
|\hat{\chi}(\omega)|^2 = \frac{1}{k_1^2} \frac{1}{\left[1 - \tilde{\omega}^2 (1 + T_0 \Gamma_b'(\omega)) \right]^2 + \left[\tilde{\omega}^2 T_0 \Gamma_b''(\omega)\right]^2}
\end{equation}
where $\tilde{\omega} = \omega /\omega_1$ to yield the expressions provided in~\cite{Cross2006}.
In experiment, the beam is driven thermoelastically. The drive can be modeled as a sinusoidally varying force applied near the ends of the beam as
\begin{equation}
F_d(x,t) = \begin{cases} \frac{F_0}{l} \sin (\omega_d t ) ~~~~ 0 \le x \le \xi_L ~~ \text{and} ~~ \xi_R \le x \le 1\\
0 ~~~~~~~~~~~~~~~~ \text{otherwise}
\end{cases}
\label{eq:driveforce}
\end{equation}
where $\xi_L$ and $\xi_R$ indicate the spatial extent over which the drive is applied.
In our experiments we have used $\xi_L = 0.05$ and $\xi_R = 0.95$. Following the approach described in Ref.~\cite{Clark2010} the amplitude response can be expressed as
\begin{equation}
W_{\omega}(\omega)^2 = \left( \frac{\pi F_0 \psi_1}{\phi_1(x_0)} \right)^2 \frac{1}{k_1^2} \frac{1}{\left( 1 - \tilde{\omega}^2 \left( 1 + T_0 \Gamma_b' \right) \right)^2 + \left( \tilde{\omega}^2 T_0 \Gamma_b'' \right)^2 }
\end{equation}
where
\begin{equation}
\psi_1 = \int_0^{\xi_L} \phi_1(x/l) dx + \int_{\xi_R}^1 \phi_1(x/l) dx
\end{equation}
which captures the coupling of the driving force to the fundamental mode. For the single mode description we can express this result as
\begin{equation}
{W_{\omega}}^2 = \left( \frac{\pi F_0 \psi_1}{\phi_1(x_0/l)} \right)^2 \left| \hat{\chi}(\omega) \right|^2,
\end{equation}
where it is apparent that the square of the amplitude is proportional to the magnitude of the susceptibility squared.
\subsection{Discussion and Possible Sources of Error}
The quantitative values used in our study are given in Tables~\ref{tab_devices},~\ref{tab_device_frequencies}, and~\ref{fit_parameters}. In Table~\ref{fit_parameters}, we have gathered together many of the important parameters used in our study for the purposes of further discussion. The top section of Table~\ref{fit_parameters} includes the material properties of the beam and the surrounding fluids that we have used. To recapitulate, the Young's modulus $E$, the density of the silicon nitride beam $\rho_s$, and the tension force in the beam $S$ are found using the optimization approach described in Section~\ref{section:Determining Material Properties from Natural Frequencies}. This provides a single set of values of $(E,\rho_s,S)$ that we use to describe all of the beams in our study. The simplification of using a single set of optimized values to describe all of the beams is expected to result in some error in the parameter values for an individual beam.
The middle section of Table~\ref{fit_parameters} includes several important nondimensional parameters. The mass loading parameter $T_0$ is evaluated using Eq.~(\ref{eq:mass-loading}). The mass loading parameter increases linearly with the density of the surrounding fluid. This is evident by the increase in $T_0$ when using water as the surrounding fluid when compared with the $T_0$ value for IPA. The nondimensional tension parameter $U$ is evaluated using Eq.~(\ref{eq:U-defn}). Since $S$ is a constant for our study, the value of $U$ increases quadratically with the length of the beam $l$. We have also included several values of the frequency based Reynolds number $Re_\omega$ given by Eq.~(\ref{Reynolds}). The Reynolds number provides insight into the relative importance of inertial and viscous effects. Each value of the Reynolds number is evaluated at the frequency of the peak of the fundamental mode $\omega_{1f}$ of the PSD of the Brownian fluctuations for that fluid. For air we use $\omega_1$ since $\omega_1 \approx \omega_{1f}$ for this case. In all cases, the Reynolds number remains small where $0.1 \le Re_\omega \le 4$ which indicates the importance of viscosity as expected for small beams in liquid.
The bottom section of Table~\ref{fit_parameters} includes the experimentally measured, and theoretically predicted, values for the effective mass $m_1$, natural frequency $\omega_1$, and the effective spring constant $k_1$ of the fundamental mode for the different beams. The experimental values of $m_1, \omega_1$, and $k_1$ are directly measured in experiment using air as the surrounding fluid. The frequency of the fundamental peak determines $\omega_1$ and the equipartition of energy is used to determine an experimental value of $k_1$. Using these two measured value of $\omega_1$ and $k_1$, the experimental measurement of the effective mass is determined from $m_1 = k_1/\omega_1^2$.
The theoretical values of $\alpha_1$, $m_1'$, $\omega_1'$, and $k_1'$ are found using the theory of an Euler-Bernoulli beam with tension that is described in Section~\ref{Beam Equation and Eigenmodes}. We use the primed variables, $m_1'$, $\omega_1'$, and $k_1'$, to clearly distinguish between these theoretical predictions and the experimentally determined values, $m_1$, $\omega_1$, and $k_1$. The theoretical value of $\alpha_1$ is found using $\phi_1(1/2)$ in Eq.~(\ref{eq:moverm}). A useful limiting case is that of a string under tension which yields $\alpha_1 = 1/2$. It is clear from our results that as the beam gets longer, the value of $U$ increases, and therefore the value of $\alpha_1$ approaches a value of 1/2. Once $\alpha_1$ is determined, it is straight forward to compute the theoretical values $m_1'$, $\omega_1'$, and $k_1'$. The theoretical prediction for the effective mass is $m_1' = \alpha_1 m_b\rho_s l b h$. The theoretical prediction for the fundamental frequency is given by Eq.~(\ref{eq:Omega_n}). Lastly, the prediction for the effective spring constant is $k_1'=\alpha_1 \rho_s l b h {\omega_1'}^{2}$.
It can be seen that there is some disagreement between the experimental and theoretical values of the fundamental frequency and a larger disagreement in the values of the effective mass and spring constant. The source of this disagreement depends upon several factors. An important contribution is our choice of the averaged material properties that we use to describe all of the beams. For example, it is possible to have lower errors in $m_1'$, $\omega_1'$, and $k_1'$ if we were instead to fit for the parameters for each individual beam while using an adjustable parameter, such as an effective length for the beam, and also allowing for Young's modulus to have a value beyond what is expected for a silicon nitride beam. However, this complicates the description by creating parameter values for each beam and we have not followed this approach here. We note that $m_1'$, $\omega_1'$, and $k_1'$ are not used in generating the theoretical curves shown in Fig.~(2)-(3) of the main text, and they are computed here only to provide some insight into the ability of an Euler-Bernoulli beam with tension to capture the dynamics of our experimental beams. In addition, there are several factors that we have not included in our analysis. These include the stiffness and the mass of the metallic electrodes as well as the presence of fabrication imperfections such as undercuts.
The theoretical predictions of the susceptibility squared $|\hat \chi(\omega)|^2$ and the PSD of the Brownian fluctuations $G_{W}(\omega)$ presented in Fig.~(2) of the main text require as input $m_1$, $\omega_1$, $k_1$, and $T_0$. Here, we have used the values for $m_1$, $\omega_1$, and $k_1$ that were determined from experiments conducted in air for each individual beam. The mass loading parameter $T_0$ depends upon the beam width and thickness, the density of the beam, and the density of the fluid. Of all of these quantities, we are most uncertain of the density of the beam $\rho_s$. We have used the value of $\rho_s$ that has been determined using the approach described in Section~\ref{section:Determining Material Properties from Natural Frequencies}.
There are several additional possible sources of disagreement between the experimental and theoretical curves of $|\hat \chi(\omega)|^2$ and $G_{W}(\omega)$ near the fundamental mode of oscillation that we would like to highlight. The first is the interaction of the oscillating viscous boundary layer and surrounding solid boundaries~\cite{Clark:2008,Lissandrello2012}. This effect is often called squeeze film damping. This additional source of damping causes a reduction in the quality factor of the oscillator. The effect of squeeze film damping can be quantified using the oscillating boundary layer thickness $\delta_s$ that is generated by the oscillating beam (cf.~\cite{green:2005,clarke:2006,Clark:2008}). The boundary layer thickness scales as $\delta_s \propto \sqrt{\nu_f/\omega}$, which captures the dependence of $\delta_s$ upon the kinematic viscosity $\nu_f$ and the frequency of oscillation $\omega$. An estimate shows that $\delta_s \sim 1~\mu \rm m$ at 100 kHz, which is of the same order as the gap distance between the beam and the closest solid surface. An important aspect is that $\delta_s$ varies as the inverse square root of the frequency. Therefore, the oscillations of the fundamental mode will have a larger boundary layer thickness than the higher modes, and in addition, the longer beams will have the largest boundary layers. For example, the boundary layer thickness decreases to $\delta_s \sim 300 ~\rm nm$ at 1 MHz. In terms of our results, the presence of this additional damping would cause the peak region near the fundamental mode for curves of $|\hat \chi(\omega)|^2$ and $G_{W}(\omega)$ to widen. This effect should be most prominent for our longest beam with $l=60$ $\mu$m. From Fig. 2(a) of the main text it is evident that the experimental curves are wider than the theory curves which would be expected due to the additional squeeze film damping in the experiment that is not included in the theory.
Another source of error is in the mode shapes of the experimental beams. The experimentally measured mode shapes are distorted when compared with the theoretically predicted mode shapes. This is evident for the fundamental mode as shown in Fig.~\ref{fig:Supp_drive}(b)-(c). Figure~\ref{fig:Supp_drive}(b) shows the 50 $\mu$m beam that has been driven at both ends of the beam. The mode shape is significantly wider than what is predicted theoretically. Figure~\ref{fig:Supp_drive}(c) shows the experimental mode shape of the 20 $\mu$m beam which has been driven only at one end of the beam near $x=0$. In this case, the fundamental mode shape is skewed toward the edge of the beam near the electrothermal drive. In this case, it can be seen that when a point measurement is made at the center, the measured amplitude will be $\sim 20\%$ smaller than the predicted one due to the distortion, resulting in a less accurate linear response measurement.
\begin{table}
\caption{\label{fit_parameters} Parameters used for the calculation of the theoretical $G_{W} (\omega)$ and $\left| {\hat\chi (\omega ) } \right|^2$.}
\begin{ruledtabular}
\begin{tabular}{lcccccccc}
Parameter & Symbol or Formula & Units& 60 $\mu \rm m$ & 50 $\mu \rm m$ & 40 $\mu \rm m$ & 30 $\mu \rm m$ & 20 $\mu \rm m$ & 15 $\mu \rm m$ \\
\hline
Young's Modulus & $E$ & GPa & 300 & 300 & 300 & 300 & 300 & 300\\
Beam Density & $\rho_s$ & $\rm kg/m^3$ & 2750 & 2750 & 2750 & 2750 & 2750 & 2750\\
Axial Load & $S$ & $\rm \mu N$ & 7.64 & 7.64 & 7.64 & 7.64 & 7.64 & 7.64 \\
Temperature & $T$ & K & 295.15 & 295.15 & 295.15 & 295.15 & 295.15 & 295.15\\
Air Kinematic Viscosity & $\nu_f$ & $\rm m^2/s \times 10^{-7}$ & 156 & 156 & 156 & 156 & 156 & 156\\
Water Kinematic Viscosity & $\nu_f$ & $\rm m^2/s \times 10^{-7}$ & 9.55 & 9.55 & 9.55 & 9.55 & 9.55 & 9.55\\
IPA Kinematic Viscosity & $\nu_f$ & $\rm m^2/s \times 10^{-7}$ & - & - & 27.8 & - & - & -\\
Air Density & $\rho_f$ & $\rm kg/m^3$ & - & - & 1.18 & - & - & - \\
Water Density & $\rho_f$ & $\rm kg/m^3$ & 997.8 & 997.8 & 997.8 & 997.8 & 997.8 & 997.8\\
IPA Density & $\rho_f$ & $\rm kg/m^3$ & - & - & 786 & - & - & - \\
\hline
Mass Loading Parameter in Water & $T_0$ & - & 2.91 & 2.91 & 2.91 & 2.91 & 2.91 & 2.91\\
Mass Loading Parameter in IPA & $T_0$ & - & -& - & 2.29 & - & - & - \\
Tension Parameter & $U$ & - & 720 & 500 & 320 & 180 & 80 & 45 \\
Reynolds Number in Air & ${\rm Re}_{\omega}(\omega_{1})$ & - & 0.13 & 0.16 & 0.21 & 0.32 & 0.49 & 0.75 \\
Reynolds Number in Water & ${\rm Re}_{\omega}(\omega_{1f})$ & - & 0.34 & 0.55 & 0.68 & 1.36 & 2.52 & 3.78 \\
Reynolds Number in IPA & ${\rm Re}_{\omega}(\omega_{1f})$ & - & - & - & 0.11 & - & - & - \\
\hline
Effective Mass (experiment) & $m_1$ & pg & 10.82 & 9.63 & 6.53 & 3.33 & 1.83 & 1.26 \\
Fundamental Frequency (experiment) & $\frac{\omega_1}{2\pi}$ & MHz & 1.48 & 1.80 & 2.35 & 3.55 & 5.38 & 8.28 \\
Spring Constant (experiment) & $k_1$ & $\rm N/m$ & 0.93 & 1.24 & 1.42 & 1.66 & 2.09 & 3.42\\
Effective Mass Coefficient (theory) & $\alpha_1$ & - & 0.476 & 0.469 & 0.462 & 0.451 & 0.434 & 0.423 \\
Effective Mass (theory) & $m_1'$ & pg & 6.93 & 5.70 & 4.49 & 3.29 & 2.11 & 1.54\\
Fundamental Frequency (theory) & $\frac{\omega_1'}{2\pi}$ & MHz & 1.56 & 1.90 & 2.42 & 3.35 & 5.42 & 7.85 \\
Spring Constant (theory) & $k_1'$ & N/m & 0.66 & 0.81 & 1.04 & 1.46 & 2.45 & 3.75\\
\end{tabular}
\end{ruledtabular}
\end{table}
\clearpage
\section{Appendix: Additional Data}
In this appendix, we present additional data plots using different axes and non-dimensionalizations. All of the relevant information is included in the figure captions.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_Appendix_1.pdf}
\caption{Figure~2 from main text plotted in linear scale. Amplitude-squared susceptibility $\left|{\hat \chi} \right|^2$ (black) and PSD of the displacement fluctuations $G_{W}$ (blue) for each beam in water; beam length is indicated in the lower left of each sub-figure. Both quantities are measured at the center of the beam. The arrows show the approximate positions of the second ($\omega_{2f}/2\pi$) and third mode ($\omega_{3f}/2\pi$) peaks in fluid (when in range).}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_Appendix_2.pdf}
\caption{Figure~2 from main text plotted as a function of reduced frequency, $\omega / \omega_1$, where $\omega_1$ is the fundamental frequency in air. The arrows show the approximate positions of the second ($\omega_{2f}/\omega_1$) and third mode ($\omega_{3f}/\omega_1$) peaks in fluid (when in range).}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_Appendix_3.pdf}
\caption{Normalized amplitude-squared susceptibility $\left|{\hat \chi} \right|^2$ (black) and PSD of the displacement fluctuations $G_{W}$ (blue) as a function of reduced frequency, $\omega / \omega_1$, for all the devices in water.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_Appendix_4.pdf}
\caption{Semi-logarithmic plots showing the measured dimensionless dissipation for all the devices in water as a function of reduced frequency, $\omega / \omega_1$.}
\label{fig:Supp_A4}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=6.75in]{Figures/Supplementary/Supp_Appendix_5.pdf}
\caption{Linear plots of the data in Figure~\ref{fig:Supp_A4}.}
\end{center}
\end{figure}
\clearpage
| be88300ab6924fcfd976336841a89f9195030d17 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\vspace*{0.4cm}
Structural Ramsey theory originated at the beginning of the 1970's in a series of
papers (see \cite{JN95}). Given structures $\pmb{A}$ and $\pmb{B}$, let $\big(\mathop{}_{\pmb{A}}^{\pmb{B}}\big)$ denote the set of
all copies of $\pmb{A}$ in $\pmb{B}$.
We write $\pmb{C}\longrightarrow (\pmb{B})^{\pmb{A}}_{l,m}$ to denote the following property:
For every finite coloring $c: \big(\mathop{}_{\pmb{A}}^{\pmb{C}}\big)
\longrightarrow l$, there is $\pmb{B}^\prime\in \big(\mathop{}_{\pmb{B}}^{\pmb{C}}\big)$ such that $c$ takes no more than $m$ colors on
$\big(\mathop{}_{\pmb{A}}^{\pmb{B^\prime}}\big)$.
Let $\mathcal{K}$ be a class
of structures. The (small) Ramsey degree of $\pmb{A}$ in $\mathcal{K}$ is the smallest positive integer $m$, if it exists, such
that for every $\pmb{B}\in \mathcal{K}$ and every positive integer $l\geq 2$ there exists $\pmb{C}\in \mathcal{K}$ such that
$\pmb{C}\longrightarrow (\pmb{B})^{\pmb{A}}_{l,m}$.
The class $\mathcal{K}$ is said to be a {\em Ramsey class} if the Ramsey degree of every $A\in \mathcal{K}$ is
1. Ramsey classes are the main topic of interest of structural Ramsey theory.
Many Ramsey classes are known.
Examples relevant for our
presentation include the classes of finite ordered graphs, finite ordered $k$-clique free graphs with
$k\geq 3$, finite ordered oriented graphs, finite ordered tournaments, and finite partial orders with a linear extension.
Given an infinite structure $\pmb{S}$ and a finite substructure $\pmb{A}$, the {\em big Ramsey degree}
of $\pmb{A}$ in $\pmb{S}$ is the smallest positive integer $m$, if it exists, such that $\pmb{S}\longrightarrow (\pmb{S})^{\pmb{A}}_{l,m}$ for every $l\geq 2$.
Research on big Ramsey degrees has gained recent momentum
due to the seminal paper of Kechris, Pestov, and Todor\v{c}evi\'{c} in \cite{AK05}, and the results
by Zucker in \cite{AZ19}
connecting big Ramsey degrees for countable structures with topological dynamics, answering a question in \cite{AK05}.
The history of big Ramsey degrees for countably infinite structures has its beginnings in an example of Sierpi\'{n}ski, who constructed a 2-coloring of pairs of rationals such that every subset forming a dense linear order retains both colors.
Later, Galvin proved that for every finite coloring of pairs of rationals, there is a subset forming a dense linear order on which the coloring takes no more than two colors, thus proving that the big Ramsey degree for pairs of rationals is exactly two.
This line of work has developed over the decades, notably with Laver proving upper bounds for all finite sets of rationals, and culminating in Devlin's calculations of the exact big Ramsey degrees for finite sets of rationals in
\cite{DD79}.
The area of big Ramsey degrees on countably infinite structures has seen considerable growth
in the
past two decades,
beginning notably with
Sauer's proof in \cite{NS06} that every finite graph has finite big Ramsey degree in the Rado graph, which is the
Fra\"{\i}ss\'{e} limit of the class of all the finite graphs, and the immediately following result of
Laflamme, Sauer and Vuksanovic in \cite{CL06} characterizing the exact big Ramsey degrees of the Rado graph.
Other recent work on big Ramsey degrees of countable structures include
ultrametric spaces (Nguyen Van Th\'{e}, \cite{LN08}),
the dense local order
(Laflamme, Nguyen Van Th\'{e} and Sauer, \cite{CL10}),
the ultrahomogeneous $k$-clique free graphs
(Dobrinen, \cite{ND17, ND19}),
and, very recently, the following:
\cite{Balkoproofs}, \cite{Balko7},
\cite{CDP}, \cite{Hubicka20},
\cite{MasulovicFBRD18},
\cite{MasulovicRDBVS20},
\cite{ZuckerForb}.
For more background in this area, we refer the reader to the excellent Habilitation of Nguyen Van Th\'{e}
\cite{NVT}
and a more recent expository paper of the first author \cite{DobIfCoLog}.
Results on big Ramsey degrees for uncountable structures are even more sparse than for countable structures.
Ramsey theorems for perfect sets mark a beginning of this line of inquiry, and most of these theorems have at their core either the Milliken theorem (\cite{KR81}),
or the Halpern-L\"{a}uchli theorem (\cite {JH66}) on which Milliken's theorem is based.
For example, Blass proved in \cite{AB81} the following partition theorem for perfect sets of $\mathbb{R}$, which was conjectured by Galvin (see \cite{FG68}), who proved it for $n\leq 3$.
\begin{theorem}[Blass \cite{AB81}]\label{thm.Blass} For every perfect subset $P$ of $\mathbb{R}$ and every finite continuous
coloring of $[P]
^n$, there is a perfect set $Q\subseteq P$ such that $[Q]^n$ has at most $(n- 1)!$
colors.
\end{theorem}
In the proof of this theorem, Blass defined patterns for finite subsets
of a perfect tree $T$ such that for every finite continuous coloring of finite subsets of the
nodes in $T$, one can make all subsets with a fixed pattern monochromatic by going to
a perfect subtree. Todor\v{c}evi\'{c} (see \cite{ST10}, Corollary 6.47) provided a simpler proof of Blass'
theorem using the Milliken space, as the perfect trees in Blass' argument can be
replaced by strong subtrees.
Given a set $X$, a subset $Y$ is called an $n$-subset of $X$ if $Y$ is a subset of $X$ of size $n$. Let
$[X]^n=\{Y\subseteq X : |Y|=n\}$ be the set of all $n$-subsets of $X$. For a graph $G$, let $V(G)$
denote its vertex set and $E(G)\subseteq [V(G)]^2$ denote its edge relation, that is, $E(G)$ is an irreflexive and symmetric binary relation. An inverse limit of finite
ordered graphs is called {\em universal} if every inverse limit of finite
ordered graphs order-embeds
continuously into it. Geschke (see \cite{SG13}) proved the existence of a universal inverse limit graph. Moreover, Huber-Geschke-Kojman (see \cite{SH19}) gave the definition of a universal inverse limit graph with no mention of an inverse system.
\begin{definition}[Huber-Geschke-Kojman \cite{SH19}]
A {\em universal inverse limit} of finite
ordered graphs
is a triple $G=\langle V, E, <\rangle$, such
that the following conditions hold.
\begin{enumerate}
\item
$V$ is a compact subset of $\mathbb{R}\backslash\mathbb{Q}$, $E\subseteq [V]
^2$, and $<$ is the
restriction of the standard order on $\mathbb{R}$ to $V$.
\item
(Modular profiniteness) For every pair of distinct vertices $u, v\in V$, there is a partition of $V$ to finitely
many closed intervals such that
\begin{enumerate}
\item[(a)]
$u, v$ belong to different intervals from the partition;
\item[(b)]
for every interval $I$ in the partition, for all $x\in V\backslash I$ and for all $y, z\in I$,
$(x, y)\in E$ if and only if $(x, z)\in E$.
\end{enumerate}
\item[(3)]
(Universality) Every nonempty open interval of $V$ contains induced copies of all finite ordered
graphs.
\end{enumerate}
\end{definition}
Based on the way Blass proved Theorem \ref{thm.Blass} by partitioning finite subsets into patterns
in \cite{AB81}, Huber, Geschke and Kojman proved in \cite{SH19} the following partition theorem for universal inverse limits of finite ordered graphs by partitioning the
isomorphism class of finite ordered graph $H$ into $T(H)$ many strong isomorphism
classes, called types.
This theorem tells us that the universal inverse limit graphs have
finite big Ramsey degrees under finite Baire-measurable colorings.
\begin{theorem}[Huber-Geschke-Kojman \cite{SH19}]\label{thm.HGK} For every finite ordered graph $H$ there is
$T(H)< \omega$ such that for every universal inverse limit graph $G$, and
for every finite Baire-measurable coloring of the set
$\big(\mathop{}_{H}^{G}\big)$ of all copies of $H$ in $G$, there is
a closed copy $G^\prime$ of $G$ in $G$ such that the set $\big(\mathop{}_{H}^{G^\prime}\big)$ of all copies of $H$ in $G^\prime$ has at most $T(H)$
colours.
\end{theorem}
The following notation will be used throughout.
The set of natural numbers, $\{0,1,2,3,\dots\}$, will be denoted by $\omega$.
Let $\omega^{<\omega}$ be the set of all finite sequences of natural numbers. Let $\subseteq$ denote
the initial segment relation. For an element $s\in \omega^{<\omega}$, let $|s|$ denote the length of $s$. We call a downward closed subset $T$ of $\omega^{<\omega}$ a tree, ordered by $\subseteq$. Every element
$t$ of a tree $T$ is called a node. Given a tree $T$, let $[T]$ be the set of all infinite branches of
$T$, i.e.,
$[T]=\{x\in \omega^{\omega} : (\forall \ n< \omega) \ x\upharpoonright n\in T\}$, where $x\upharpoonright n$ is its initial
segment of length $n$. $T^\prime$ is called a subtree of $T$ if $T^\prime\subseteq T$ and $T^\prime$ is a tree. For a tree $T$ and $t\in T$,
$s$ is called an immediate successor of
$t$ if $s$ is a minimal element of $T$ above $t$. The set of immediate successors of
$t$ in $T$ is denoted by $\mbox{succ}_T(t)$. Let $T_t$ be the set of all nodes in $T$ comparable to $t$, i.e., $T_t =\{s\in T : t\subseteq s \vee s\subseteq
t\}$. For $n\in \omega$, we let $T(n)=\{t\in T : |t|=n\}$.
In order to state the results of Huber-Geschke-Kojman and of Zheng, we need to introduce the following notation and structures.
Let $\mathrm{R}$ denote the Rado graph, i.e., the unique (up to isomorphism) countable universal homogeneous
graph. We assume that the set of vertices of $\mathrm{R}$ is just the set $\omega$
of natural numbers. For $n\in \omega$, let $\mathrm{R}_n$ be
the induced subgraph of $\mathrm{R}$ on $\{0,\dots,n\}$.
\begin{definition}[\cite{SH19}] Let $T_{\max}\subseteq \omega^{<\omega}$ be the nonempty tree such that for each $t\in T_{\max}$,
$$\mbox{succ}_{T_{\max}}(t)=\{t^\frown\langle0\rangle, t^\frown\langle1\rangle, \dots, t^\frown\langle|t|\rangle\}.$$
For $t\in T_{\max}$, we define $G_t$ to be the ordered graph on the vertex set $\mbox{succ}_{T_{\max}}(t)$ with
lexicographical ordering, such that $G_t$ is isomorphic to $\mathrm{R}_{|t|}$.
\end{definition}
Note that $[T_{\max}]$ is a subset of $\omega^{\omega}$.
Given $x, y\in [T_{\max}]$ with $x\neq y$, let $x\cap y\in \omega^{<\omega}$ be the
common initial segment of $x$ and $y$, i.e. $x\cap y= x\upharpoonright \min\{n : x(n)\neq y(n)\}$. The tree $T_{\max}$ and the ordered graphs $G_t \ (t\in T_{\max})$ induce an ordered graph $G_{\max}$ on the
vertex set $[T_{\max}]$, ordered lexicographically, with the edge relation defined as follows.
For $x, y\in [T_{\max}], (x, y)\in E(G_{\max})$ if and only if
$(x\upharpoonright (|x\cap y|+1), y\upharpoonright (|x\cap y|+1))\in E(G_{x\cap y})$. Suppose that $T$ is a subtree of $T_{\max}$ and $t\in T$. Let
$G^T_t$ denote the induced subgraph of $G_t$ on the vertex set $\mbox{succ}_{T}(t)$. We define $G(T)$ to
be the induced subgraph of $G_{\max}$ on $[T]$.
A subtree $T$ of $T_{\max}$ is called a {\em $G_{\max}$-tree} if for every finite ordered graph $H$ and every $t\in T$, there is $s\in T$
with $t\subseteq s$ such that $H$ embeds into $G^T_s$.
In particular, $T_{\max}$ is a $G_{\max}$-tree.
Let $(\mathcal{R}, \leq, r)$ be a triple satisfying the following: $\mathcal{R}$ is a nonempty set, $\leq$ is a quasi-ordering on $\mathcal{R}$, and $r : \mathcal{R}\times \omega\longrightarrow \mathcal{AR}$ is a
mapping giving us the sequence ($r_n(\cdot)=r(\cdot,n)$) of approximation mappings, where
$$\mathcal{AR}=\{r_n(A):A\in\mathcal{R}\mathrm{\ and\ } n\in\omega\}.
$$
For $a\in \mathcal{AR}$ and $A\in \mathcal{R}$,
$$[a, A]=\{B\in \mathcal{R} : (B\leq A)\wedge (\exists n)(r_n(A)=a)\}.$$
The topology on $\mathcal{R}$ is given by the basic open sets $[a, A]$. This topology is called
the {\em Ellentuck topology} on $\mathcal{R}$. Given the Ellentuck topology
on $\mathcal{R}$, the notions of nowhere dense, and hence of meager are defined in the natural
way. Thus, we may say that a subset $\mathcal{X}$ of $\mathcal{R}$ has the {\em property of Baire} iff $\mathcal{X}=\mathcal{O}\triangle \mathcal{M}$
for some Ellentuck open set $\mathcal{O}\subseteq \mathcal{R}$ and Ellentuck meager set $\mathcal{M }\subseteq \mathcal{R}$.
\begin{definition}[\cite{ST10}] A subset $\mathcal{X}$ of $\mathcal{R}$ is {\em Ramsey} if for every $\emptyset\neq [a, A]$, there is a
$B\in [a, A]$ such that $[a, B]\subseteq \mathcal{X}$ or $[a, B]\cap \mathcal{X}=\emptyset$. $\mathcal{X}\subseteq \mathcal{R}$ is {\em Ramsey null} if for
every $\emptyset\neq [a, A]$, there is a
$B\in [a, A]$ such that $[a, B]\cap \mathcal{X}=\emptyset$.
A triple $(\mathcal{R},\leq,r)$ is a {\em topological Ramsey space} if every property of Baire subset
of $\mathcal{R}$ is Ramsey and if every meager subset of $\mathcal{R}$ is Ramsey null.
\end{definition}
In \cite{YY18}, Zheng constructed a collection of topological Ramsey spaces of trees. For each type $\tau$ of finite ordered graphs, the space $(\mathcal{G}_\infty(\tau), \leq, r)$ consists of $G_{\max}$-trees of a
particular shape. The new spaces $\mathcal{G}_\infty(\tau)$ not only depend on the fact that the class of finite ordered graphs is the Ramsey class, but also, similarly to the Milliken space, are based
on the Halpern-L\"{a}uchli theorem.
Moreover, she presented an application of the topological Ramsey spaces $\mathcal{G}_\infty(\tau)$ to inverse limit graph theory.
Similarly to how Todor\v{c}evi\'{c} proved
Blass' Theorem \ref{thm.Blass}, Zheng used the new spaces $\mathcal{G}_\infty(\tau)$ to prove
the following Theorem \ref{thm.Zheng3.1} (Theorem 3.1 in \cite{SH19}),
which is a key step to show the above Theorem
\ref{thm.HGK}
in \cite{SH19}.
\begin{theorem}[Theorem 3.1 in \cite{SH19}]\label{thm.Zheng3.1}
Let $T$ be an arbitrary $G_{\max}$-tree. For every type $\tau$ of a
finite induced subgraph of $G_{\max}$, and for every
continuous coloring $c :
\big(\mathop{}_{ \ \ \tau}^{G(T)}\big)\longrightarrow 2$,
there is a $G_{\max}$-subtree $S$ of $T$ such that c is constant on $\big(\mathop{}_{ \ \ \tau}^{G(S)}\big)$.
\end{theorem}
In this paper, we extend Zheng's methods to build a collection of topological Ramsey spaces of trees in the setting of
Fra\"{\i}ss\'{e} classes of finite ordered structures with finitely many binary relations satisfying the Ramsey property.
Based on these topological Ramsey spaces and the work of Huber-Geschke-Kojman on
inverse limits of finite ordered graphs, we prove the following theorem.
Here, $\pmb{F_{\max}}$ is a universal limit structure encoded
in a particular way on
the set of infinite branches of a certain finitely branching tree $T_{\max}$
(see Definitions \ref{defn.T_max} and \ref{defn.F_max}).
\begin{theorem}\label{thm.1.7}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class, in a finite signature,
of finite ordered binary relational structures with the Ramsey property. For every $\pmb{H}\in \mathcal{K}$, there is a finite number $T(\pmb{H}, \pmb{F_{\max}})$ such that for every universal inverse limit structure
$\pmb{G}$, for every finite Baire-measurable coloring of the set $\big(\mathop{}_{\pmb{H}}^{\pmb{G}}\big)$
of all copies of $\pmb{H}$ in $\pmb{G}$, there is a closed
copy $\pmb{G^\prime}$ of $\pmb{G}$ contained in $\pmb{G}$ such that the set $\big(\mathop{}_{\pmb{H}}^{\pmb{G^\prime}}\big)$
of all copies of $\pmb{H}$ in $\pmb{G^\prime}$ has no more than $T(\pmb{H}, \pmb{F_{\max}})$ colors.
\end{theorem}
This means that for each such Fra\"{\i}ss\'{e} class,
its universal inverse limit structures have finite big Ramsey degrees under finite Baire-measurable colorings.
For the following classes, we characterize the big Ramsey degrees in terms of types.
\begin{theorem}\label{thm.1.8}
Let $\mathcal{K}$ be a \Fraisse\ class in a finite binary relational signature such that one of the following hold:
\begin{enumerate}
\item
$\mathcal{K}$ is an ordered expansion of a free amalgamation class;
\item
$\mathcal{K}$ is the class of finite ordered tournaments;
\item
$\mathcal{K}$ is the class of finite partial orders with a linear extension.
\end{enumerate}
Let $\pmb{G}$ be
a universal inverse limit structure for $\mathcal{K}$ contained in
$\pmb{F_{\max}}$.
Then for each $\pmb{H}\in \mathcal{K}$, each type representing $\pmb{H}$ in $\pmb{G}$ persists in each closed subcopy of $\pmb{G}$.
It follows that
the big Ramsey degree $T(\pmb{H}, \pmb{F_{\max}})$
for finite
Baire-measurable
colorings of
$\big(\mathop{}_{ \ \ \pmb{H}}^{\pmb{F_{\max}}}\big)$
is exactly the number of types in $T_{\max}$ representing a copy of $\pmb{H}$.
\end{theorem}
\section{Ordered binary relational Fra\"{\i}ss\'{e} classes and $\pmb{F_{\max}}$-trees}
Let us first review some basic facts of the Fra\"{\i}ss\'{e} theory for finite ordered binary relational structures which are
necessary to this paper.
More general background on Fra\"{\i}ss\'{e} theory can be found in \cite{AK05}.
We shall call $L=\{<, R_0,\dots,R_{k-1}\}$ an {\em ordered binary relational signature} if it consists
of the order relation symbol $<$ and finitely many binary relation symbols $R_{\ell}$, $\ell<k$ for some $k<\omega$.
A structure for $L$ is of the form $\pmb{A}=\langle A, <^{\pmb{A}}, R_0^{\pmb{A}},\dots,R_{k-1}^{\pmb{A}}\rangle$, where $A\neq \emptyset$ is the universe of $\pmb{A},\ <^{\pmb{A}}$ is a linear
ordering of $A$, and each $R_{\ell}^{\pmb{A}}\subseteq A\times A$.
An embedding between structures $\pmb{A}, \pmb{B}$ for $L$ is an injection
$\pi : A\longrightarrow B$ such that for any two $a, a^\prime\in A,\ a<^{\pmb{A}} a^\prime\Longleftrightarrow
\pi(a)<^{\pmb{B}} \pi(a^\prime)$ and
for each $\ell<k$,
$(a_1, a_2)\in R_\ell^{\pmb{A}}\Longleftrightarrow (\pi(a_1),\pi(a_2))\in R_\ell^{\pmb{B}}$.
If $\pi$ is the identity, we say
that $\pmb{A}$ is a substructure of $\pmb{B}$. An isomorphism is an onto embedding. We
write $\pmb{A}\leq \pmb{B}$ if $\pmb{A}$ can be embedded in $\pmb{B}$ and $\pmb{A}\cong \pmb{B}$ if $\pmb{A}$ is isomorphic to $\pmb{B}$.
A class $\mathcal{K}$ of finite structures is called {\em hereditary} if $\pmb{A}\leq \pmb{B}\in \mathcal{K}$ implies $\pmb{A}\in \mathcal{K}$.
It satisfies the {\em joint embedding property} if for any $\pmb{A},\ \pmb{B}\in \mathcal{K}$, there is
$\pmb{C}\in \mathcal{K}$ with $\pmb{A}\leq \pmb{C}$ and $\pmb{B}\leq \pmb{C}$. We say that $\mathcal{K}$ satisfies the {\em amalgamation property}
if for any embeddings $f : \pmb{A}\longrightarrow \pmb{B},\ g : \pmb{A}\longrightarrow \pmb{C}$ with $\pmb{A},\ \pmb{B},\ \pmb{C}\in \mathcal{K}$, there is
$\pmb{D}\in \mathcal{K}$ and embeddings $r : \pmb{B}\longrightarrow \pmb{D}$ and $s : \pmb{C}\longrightarrow \pmb{D}$, such that $r\circ f= s\circ g$.
A class of finite structures $\mathcal{K}$ is called a {\em Fra\"{\i}ss\'{e} class} if it is hereditary, satisfies joint embedding and
amalgamation, contains only countably many structures, up to isomorphism,
and contains structures of arbitrarily large (finite) cardinality.
A Fra\"{\i}ss\'{e} class satisfies the {\em free amalgamation property} (or {\em has free amalgamation})
if $\pmb{D}$, $r$, and $s$ in the amalgamation property can be chosen
so that
$r[B]\cap s[C]=r\circ f[A]=s\circ g[A]$, and
$\pmb{D}$ has no additional relations on its universe other than those inherited from $\pmb{B}$ and $\pmb{C}$.
Let $\pmb{A}$ be a structure for $L$. For each $X\subseteq A$, there is a smallest substructure containing $X$, called the substructure generated by $X$. A substructure is called finitely generated if it is generated by a finite set. A structure is locally finite if all its finitely generated substructures are finite. The age of $\pmb{A}$, Age$(\pmb{A})$ is the class of all finitely generated structures in $L$ which can be embedded in $\pmb{A}$.
We call $\pmb{A}$ {\em ultrahomogeneous} if every
isomorphism between finitely generated substructures of $\pmb{A}$ can be extended to an automorphism
of $\pmb{A}$.
A locally finite, countably infinite, ultrahomogeneous
structure is called a {\em Fra\"{\i}ss\'{e} structure}.
There is a canonical one-to-one correspondence between
Fra\"{\i}ss\'{e} classes of finite structures and Fra\"{\i}ss\'{e} structures,
discovered by Fra\"{\i}ss\'{e}.
If $\pmb{A}$ is a Fra\"{\i}ss\'{e} structure, then Age$(\pmb{A})$ is a
Fra\"{\i}ss\'{e} class of finite structures. Conversely, if $\mathcal{K}$ is a Fra\"{\i}ss\'{e} class of relational structures, then there is a unique Fra\"{\i}ss\'{e}
structure, called the Fra\"{\i}ss\'{e} limit of $\mathcal{K}$, denoted by Flim$(\mathcal{K})$, whose age is exactly
$\mathcal{K}$.
\begin{definition}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures. We say that $\mathcal{K}$ satisfies the {\em Ramsey property} if $\mathcal{K}$ is a Ramsey class.
That is, for each
$\pmb{A},\pmb{B}\in\mathcal{K}$
such that $\pmb{A}\le \pmb{B}$ and
for every positive integer $l\geq 2$, there exists $\pmb{C}\in \mathcal{K}$ such that
$\pmb{C}\longrightarrow (\pmb{B})^{\pmb{A}}_{l}$.
\end{definition}
Given an ordered binary relational signature
$L=\{<,R_0,\dots, R_{k-1}\}$,
let $L^-$ denote $\{R_0,\dots, R_{k-1}\}$.
An $L^-$-structure $\pmb{A}$ is called {\em irreducible} if for any two elements $x,y \in A$, there is some relation $R\in L^-$ such that either $R^{\pmb{A}}(x,y)$
or $R^{\pmb{A}}(y,x)$
holds.
Given a set $\mathcal{F}$ of finite $L^-$-structures, let Forb$(\mathcal{F})$ denote the class of finite $L^-$-structures $\pmb{A}$
such that no member of $\mathcal{F}$ embeds into $\pmb{A}$.
It is well-known that a \Fraisse\ class in signature $L^-$ has free amalgamation if and only if it is of the form Forb$(\mathcal{F})$ for some set $\mathcal{F}$ of finite irreducible $L^-$-structures.
It follows from results of
Ne\v{s}et\v{r}il and R\"{o}dl in \cite{JV77,JV83} that all \Fraisse\ classes in signature $L$ for which the $L^-$-reduct has free amalgamation
has the Ramsey property.
For $k\geq 3$, a graph $G$ is called {\em $k$-clique free} if for any $k$ vertices in $G$, there is at least
one pair with no edge between them; in other words, no $k$-clique embeds into $G$ as
an induced subgraph.
An oriented graph $G=\langle V(G), E(G)\rangle$ is a relational structure, where $V(G)$
denotes its vertex set and $E(G)\subseteq V(G)\times V(G)$ denotes its directed edge relation, that is, $E(G)\subseteq V(G)\times V(G)$ is an irreflexive binary relation such that for all $x, y\in V(G)$,
$(x, y)\in E(G)$ implies $ (y, x)\notin E(G)$. A tournament $G$ is an
oriented graph such that for all $x\neq y$, either $(x, y)\in E(G)$ or $(y, x)\in E(G)$.
A partial order with a linear extension is a structure $\pmb{P}=\langle P,<^{\pmb{P}},R^{\pmb{P}}\rangle$ where $R^{\pmb{P}}$ is a partial ordering on $P$,
$<^{\pmb{P}}$ is a linear ordering on $P$, and whenever $x\ne y$ and $R^{\pmb{P}}(x,y)$ holds, then also $x<^{\pmb{P}}y$ holds.
\begin{example}
Let $\mathcal{OG}$, $\mathcal{OG}_k$,\ $\mathcal{OOG}$, $\mathcal{OT}$, and $\mathcal{OPO}$ denote the Fra\"{\i}ss\'{e} classes of all finite ordered graphs, finite ordered $k$-clique free graphs ($k\geq 3$), finite ordered oriented graphs, finite ordered tournaments, and finite partial orders with a linear extension, respectively.
Each of these classes
has the Ramsey property.
The Ramsey property for
$\mathcal{OG}$, $\mathcal{OG}_k$, $\mathcal{OOG}$, and $\mathcal{OT}$,
are special cases of a theorem of
Ne\v{s}et\v{r}il-R\"{o}dl (\cite{JV77,JV83});
the Ramsey property for
that $\mathcal{OG}$ and $\mathcal{OT}$ follow from independent work of Abramson and Harrington in \cite{AH78}.
The Ramsey property for $\mathcal{OPO}$ was announced by Ne\v{s}et\v{r}il and R\"{o}dl in \cite{NR84}, and the first proof was published
by Paoli, Trotter, and Walker
in \cite{PTW85}.
\end{example}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property.
We may assume the universe of Flim$(\mathcal{K})$
is $\omega$, so that the universe is well-ordered.
For $n\in \omega$, let Flim$(\mathcal{K})_n$ be the initial segment of Flim$(\mathcal{K})$ on $\{0,\dots,n\}$.
\begin{definition}\label{defn.T_max}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property, and let $T_{\max}\subseteq \omega^{<\omega}$ be the nonempty tree such that for each $t\in T_{\max}$,
$$
\mbox{succ}_{T_{\max}}(t)=\{t^\frown\langle0\rangle, t^\frown\langle1\rangle, \dots, t^\frown\langle|t|\rangle\}.
$$
For $t\in T_{\max}$, we define $\pmb{F_t}\in \mathcal{K}$ to have universe
$F_t:=\mbox{succ}_{T_{\max}}(t)$, ordered by the
lexicographical ordering, such that $\pmb{F_t}$ is isomorphic to Flim$(\mathcal{K})_{|t|}$.
\end{definition}
\begin{definition}\label{defn.F_max}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property.
The tree $T_{\max}$ and $\pmb{F_t}\in \mathcal{K} \ (t\in T_{\max})$ induce a structure $\pmb{F_{\max}}$ on the universe $F_{\max}:=[T_{\max}]$, ordered lexicographically, with the binary relations $R^{\pmb{F_{\max}}}_\ell$, $\ell<k$
(where $k$ is the cardinality of the signature $L$),
as follows:
$$
\forall \ x, y\in [T_{\max}],\ (x, y)\in
R_\ell^{\pmb{F_{\max}}}\Longleftrightarrow
(x\upharpoonright (|x\cap y|+1), y\upharpoonright (|x\cap y|+1))\in R_\ell^{\pmb{F_{x\cap y}}}.
$$
\end{definition}
\begin{lemma}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property.
If $\pmb{F}\in \mathcal{K}$ and $t\in T_{\max}$, then there is $s\in T_{\max}$
with $t\subseteq s$ such that $\pmb{F}$ embeds into $\pmb{F_s}$.
\end{lemma}
\begin{proof}
Since $\pmb{F_t}$ is isomorphic to Flim$(\mathcal{K})_{|t|}$,
it suffices
to prove that each $\pmb{F}\in \mathcal{K}$ embeds into Flim$(\mathcal{K})$ on universe $\omega$.
Now the age of Flim$(\mathcal{K})$ is exactly $\mathcal{K}$. So each $\pmb{F}\in \mathcal{K}$ embeds into Flim$(\mathcal{K})$ on universe $\omega$.
\end{proof}
\begin{definition}\label{defn.2.6}
(1) Suppose that $T$ is a subtree of $T_{\max}$ and $t\in T$. Let
$\pmb{F^T_t}$ denote the induced substructure of $\pmb{F_t}$ on the universe $F^T_t:=\mbox{succ}_{T}(t)$. We define $\pmb{F(T)}$ to
be the induced substructure of $\pmb{F_{\max}}$ on the universe $F(T):=[T]$.
(2) Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property. A subtree $T$ of $T_{\max}$ is called an
{\em $\pmb{F_{\max}}$-tree}
if for every $\pmb{F}\in \mathcal{K}$ and every $t\in T$, there is $s\in T$
with $t\subseteq s$ such that $\pmb{F}$ embeds into $\pmb{F^T_s}$.
In particular, $T_{\max}$ is an $\pmb{F_{\max}}$-tree.
\end{definition}
\begin{definition}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property. A sequence $(T_j)_{j\in \omega}$ is a {\em fusion sequence} with witness $(m_j)_{j\in \omega}$ if the
following hold:
\begin{enumerate}
\item
$(m_j)_{j\in \omega}$ is a strictly increasing sequence of natural numbers.
\item
For all $j, l\in \omega$, if $j< l$, then $T_l$
is an $\pmb{F_{\max}}$-subtree of $T_j$ such that
$T_j(m_j)=T_l(m_j)$.
\item
For every $\pmb{F}\in \mathcal{K}$ , every $j\in \omega$, and every $t\in T_j(m_j)$,
there is $l> j$ such that $t$ has an extension $s$ in $T_l$ such that $|s|< m_l$ and
$\pmb{F}$ embeds into $\pmb{F^{T_l}_s}$.
\end{enumerate}
\end{definition}
One can check that if $(T_j)_{j\in \omega}$ is a fusion sequence witnessed by
$(m_j)_{j\in \omega}$, then the fusion $\bigcap_{j\in \omega}
T_j=
\bigcup_{j\in \omega}
(T_j\cap \omega^{\leq m_j})$ is an $\pmb{F_{\max}}$-tree.
\section{Types}
\begin{definition}
Let $T$ be an $\pmb{F_{\max}}$-tree and $\pmb{F}$ a finite induced substructure of $\pmb{F(T)}$.
We define $\triangle(\pmb{F})$ and $\pmb{F}^{\vee}$ as follows:
$$\triangle(\pmb{F})=\max\{|x\cap y| : x, y\in F\wedge x\neq y\},$$
$$\pmb{F}^{\vee}=\{x\upharpoonright (\triangle(\pmb{F})+1) : x\in F\}.$$
\end{definition}
\begin{example}
Let $\mathcal{K}=\mathcal{OG}$ and $\pmb{H}\in \mathcal{K}$ as in Figure 1, where $H=\{x, y, z\}$, $x=0000\dots$,
$y=0100\dots$, and $z=0111\dots$. Then $\triangle(\pmb{H})=2$ and $\pmb{H}^{\vee}=\{u, v, w\}$, where $u=000, v=010$ and $w=011$.
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{1}\\Figure 1:\ $\triangle(\pmb{H})$ and $\pmb{H}^{\vee}$
\end{center}
\end{example}
\begin{definition}
Let $\pmb{F}$ and $\pmb{F^{\prime}}$ be finite induced substructure of $\pmb{F_{\max}}$. We
say $\pmb{F}$ and $\pmb{F^{\prime}}$ are {\em strongly isomorphic} if there exists an isomorphism $\varphi : \pmb{F}\longrightarrow \pmb{F^{\prime}}$
such that $\forall \ \{x_0, y_0\}, \{x_1, y_1\}\in [F]^2$,
$$|x_0\cap y_0|\leq |x_1\cap y_1|\Longleftrightarrow |\varphi(x_0)\cap \varphi(y_0)|\leq |\varphi(x_1)\cap \varphi(y_1)|.$$
\end{definition}
Clearly, strong isomorphism is an equivalence relation. By a {\em type} we mean a strong
isomorphism equivalence class. In particular,
there are only finitely many types inside an isomorphism class.
Suppose that $\pmb{F}$ and $\pmb{H}$ are structures. Let $\big(\mathop{}_{\pmb{H}}^{\pmb{F}}\big)$
be the set of all
induced substructures $\pmb{H^{\prime}}$ of $\pmb{F}$ isomorphic to $\pmb{H}$. If $\pmb{F}$ is a induced substructure of $\pmb{F_{\max}}$ and $\tau$ is a type,
we let $\big(\mathop{}_{\tau}^{\pmb{F}}\big)$
be the set of all induced substructures of $\pmb{F}$ of type $\tau$.
\begin{example}
Let $\mathcal{K}=\mathcal{OG}$ and $\pmb{H}\in \mathcal{K}$ be as in Figure 2, where $H=\{v_0, v_1, v_2\}$.
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{2}\\Figure 2
\end{center}
Then there are 3 types for $\pmb{H}$ as in Figure 3.
\begin{center}
\centering
\includegraphics[totalheight=2.0in]{3}\\Figure 3:\ 3 types for $\pmb{H}$
\end{center}
\noindent Let $\pmb{H}\in \mathcal{K}$ be as in Figure 4. Then there are 2 types for $\pmb{H}$ as in Figure 5.
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{4}\\Figure 4
\end{center}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{5}\\Figure 5:\ 2 types for $\pmb{H}$
\end{center}
\vspace*{0.2cm}
\noindent Let $\pmb{H}\in \mathcal{K}$ be as in Figure 6. Then there is only 1 type for $\pmb{H}$ as in Figure 7.
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{6}\\Figure 6
\end{center}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{7}\\Figure 7:\ 1 type for $\pmb{H}$
\end{center}
\end{example}
\begin{example}
Let $\mathcal{K}=\mathcal{OT}$ and $\pmb{F}\in \mathcal{K}$ be as in Figure 8.
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{8}\\Figure 8
\end{center}
\noindent Then there are 3 types for $\pmb{F}$ as in Figure 9.
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{9}\\Figure 9:\ 3 types for $\pmb{F}$
\end{center}
\noindent If $\pmb{F}\in \mathcal{K}$ is as in Figure 10, then there are 2 types for $\pmb{F}$ as in Figure 11.
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{10}\\Figure 10
\end{center}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{11}\\Figure 11:\ 2 types for $\pmb{F}$
\end{center}
Let $\pmb{F}\in \mathcal{K}$ be as in Figure 12.
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{12}\\Figure 12
\end{center}
Then there is only 1 type for $\pmb{F}$ as in Figure 13.
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{13}\\Figure 13:\ 1 type for $\pmb{F}$
\end{center}
\end{example}
\begin{example}
Suppose that $\mathcal{K}=\mathcal{OPO}$ and $\pmb{H}\in \mathcal{K}$ as in Figure 14.
\begin{center}
\centering
\includegraphics[totalheight=0.5in]{14}\\Figure 14
\end{center}
\noindent Here
\begin{center}
\centering
\includegraphics[totalheight=0.25in]{15}
\end{center}
denotes that $R(v_i, v_j)$, where $R$ is a partial order. Then there are 3 types for $\pmb{H}$ as in Figure 15.
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{16}\\Figure 15:\ 3 types for $\pmb{H}$
\end{center}
\vspace*{0.3cm}
\noindent Let $\pmb{H}\in \mathcal{K}$ be as in Figure 16. Then there are 2 types for $\pmb{H}$ as in Figure 17.
\begin{center}
\centering
\includegraphics[totalheight=0.5in]{17}\\Figure 16
\end{center}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{18}\\Figure 17:\ 2 types for $\pmb{H}$
\end{center}
\vspace*{0.3cm}
\noindent Let $\pmb{H}\in \mathcal{K}$ be as in Figure 18. Then there is only 1 type for $\pmb{H}$ as in Figure 19.
\vspace*{0.3cm}
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{19}\\Figure 18
\end{center}
\vspace*{0.2cm}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{20}\\Figure 19:\ 1 type for $\pmb{H}$
\end{center}
\end{example}
\vspace*{0.1cm}
\section{Topological Ramsey spaces for coding inverse limit structures for finitely many binary relations}
This section is essentially work of Zheng from Section 3 in \cite{YY18}.
Her work is straightforwardly extended from the context of finite ordered graphs to
the broader context of Fra\"{\i}ss\'{e} classes of finite ordered binary relational structures satisfying the Ramsey property.
We include it in this paper for the reader's convenience, making a few modifications.
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures satisfying the Ramsey property, with signature $L=\{<, R_0,\dots,R_{k-1}\}$, where each $R_{\ell}$ $(\ell<k)$ is a binary relation.
We fix a type $\tau$ and build a topological Ramsey space $\mathcal{F}_\infty(\tau)$.
We may denote this space by $\mathcal{F}_\infty$ when the type is clear
from the context.
Let $m+1$ be the number of elements for a finite ordered structure in $\tau$.
Let $\{\pmb{F_i} : i< \omega\}$ enumerate the set of all finite structures in $\mathcal{K}$, up to isomorphism, labelled
so that for every $i < j < \omega$, $|F_i| \le |F_j|$.
Let $S$ be a tree.
A node $t\in S$ is a {\em splitting node} if $|\mbox{succ}_S (t)| > 1$. We
say $S$ is {\em skew} if $S$ has at most one splitting node at each level, i.e.,
$$
\forall \ n\in \omega,\ |\{t\in S\cap \omega^n : |\mbox{succ}_S(t)| > 1\}|\le 1.
$$
Notice that if
$S$ is a skew tree and $i>0$ is given,
then each node
$t\in S$ for which $\pmb{F^S_t}
\cong \pmb{F_i}$ is a splitting node, since the structure $\pmb{F_i}$ has universe of size at least two.
Thus, any two nodes in the set
$\{t\in S : \pmb{F^S_t}
\cong \pmb{F_i}\}$ have different lengths, so the nodes in this set can be enumerated in order of increasing length. This will be useful in part (iii) of (2) in the next definition.
\begin{definition} Let $\tau$ be a type and $m+1$ be the number of elements for a finite ordered structure in $\tau$. We define the space $(\mathcal{F}_\infty, \leq, r)$ as follows.
Let $S$ be a member of $\mathcal{F}_\infty$ if $S$ is a skew subtree of $T_{\max}$ and when we enumerate the set of
splitting nodes $\{s\in S : |\mbox{succ}_S(s)|> 1\}$ as $\{s_i\}_{i<\omega}$ in the order of length,
(1) there is a finite structure $\pmb{F}\in \tau$ such that
$$\pmb{F}^{\vee}= S\cap \omega^{|s_{m-1}|+1};$$
(2) for all $i> 0$,
\quad (i) $\forall \ s\in S\cap \omega^{>|s_{m-1}|}, \forall \ u\in \mbox{succ}_S(s)$,
$$\pmb{F^S_{s}}\cong \pmb{F_i}\Rightarrow \exists!\ t\in S\
(u\subseteq t\wedge \pmb{F^S_t}\cong \pmb{F_{i+1}});$$
\quad (ii) for every pair $s,\ t\in S\cap \omega^{>|s_{m-1}|}$,
$$\pmb{F^S_s}\cong \pmb{F_i}\wedge \pmb{F^S_t}\cong \pmb{F_{i+1}}\Rightarrow |s| < |t|;$$
\quad (iii) if $\{t\in S : \pmb{F^S_t}
\cong \pmb{F_i}\}$ is enumerated in order of increasing length as $\{t_j\}_j$, then
there is $l< \omega$ such that $\{t_j\upharpoonright l\}_j$ is strictly increasing in lexicographical
ordering.
\end{definition}
When we say that $\{s_i\}_{i<\omega}$ is the set of splitting nodes in $S$, we tacitly assume
that the length $|s_i|$ is strictly increasing in $i$.
For $S, U\in \mathcal{F}_\infty$, we write $S\leq U$ if and only if $S\subseteq U$. For $l< \omega$ and $S\in \mathcal{F}_\infty$ with the set of splitting nodes $\{s_i\}_{i<\omega}$, we define the finite
approximation $r_l(S)$ as follows: let $$r_0(S)= \emptyset \ \mbox{and} \ r_{l+1}(S) = S\cap \omega^{\leq |s_l|+1}.$$
We specify a few
more definitions that are often used in topological Ramsey spaces. Let $\mathcal{F}_{< \infty}$ denote
the set of all finite approximations, i.e.
$$\mathcal{F}_{< \infty}=\{r_l(S) : S\in \mathcal{F}_{\infty}\wedge l\in \omega\}.$$
For $a, b\in \mathcal{F}_{< \infty}$, let $a \leq_{\small\mbox{fin}} b$ if $a\subseteq b$. Let $|a|=n$ if there is $S\in \mathcal{F}_{\infty}$ with $r_n(S)=a$.
For $a, b\in \mathcal{F}_{< \infty}$, we write $a\sqsubseteq b$ if there are $l< p<\omega$ and $S\in \mathcal{F}_{\infty}$ such that
$a=r_l(S)$ and $b=r_p(S)$.
For $a\in \mathcal{F}_{< \infty}$ and $S\in \mathcal{F}_{\infty}$,
depth$_S(a)=\min\{n : a\leq_{\small\mbox{fin}}r_n(S)\}$,
where by convention, $\min \emptyset=\infty$.
We equip the space $\mathcal{F}_{\infty}$ with the Ellentuck topology,
with basic open sets of the form
$$[a, S]=\{X\in \mathcal{F}_{\infty} : (X\leq S)\wedge (\exists l)(r_l(X)=a)\},$$ for $a\in \mathcal{F}_{< \infty}$ and $S\in \mathcal{F}_{\infty}$. For $l< \omega$, let
$$\mathcal{F}_l=\{r_l(X) : X\in \mathcal{F}_{\infty}\},$$
$$[l, S]=[r_l(S), S], \mbox{and}$$
$$r_l[a, S]=\{r_l(X) : X\in [a, S]\}.$$
The height of an element $a\in \mathcal{F}_{< \infty}$ is height$(a)=\max_{s\in a} |s|$. In general, $|a|\leq\mbox{height}(a)$.
Now we show that $(\mathcal{F}_{\infty}, \leq, r)$ is a topological Ramsey space by proving that $\mathcal{F}_{\infty}$ is
closed
as a subspace of
$(\mathcal{F}_{< \infty})^\omega$,
and satisfies the axioms {\bf (A1)-(A4)} as defined in pages 93--94 of \cite{ST10}.
It is straightforward to check {\bf (A1)-(A3)}.
Moreover, $\mathcal{F}_{\infty}$ is a closed subset of $(\mathcal{F}_{< \infty})^\omega$ when we identify $S\in \mathcal{F}_{\infty}$
with $(r_n(S))_{n< \omega}\in (\mathcal{F}_{< \infty})^\omega$ and equip $\mathcal{F}_{< \infty}$ with the discrete topology and $(\mathcal{F}_{< \infty})^\omega$ with the product topology.
\begin{definition}
Let $T\subseteq \omega^{<\omega}$ be a (downwards closed) finitely branching tree with no terminal nodes, and let
$N$ an infinite subset of $\omega$.
A set $U$ is called a {\em strong subtree}
of $\bigcup_{n\in N} T\cap \omega^n$ if there is an infinite set $M\subseteq N$ such
that the following conditions hold.
\begin{enumerate}
\item
$U\subseteq \bigcup_{n\in M} T\cap \omega^m$ and $U\cap \omega^m\neq \emptyset$ for all $m\in M$. In this case, we say that $M$
witnesses that $U$ is a strong subtree.
\item
If $m_1< m_2$ are two successive elements of $M$ and if $u\in U\cap \omega^{m_1}$, then every
immediate successor of $u$ in $\bigcup_{n\in N} T\cap \omega^n$ has exactly one extension
in $U\cap \omega^{m_2}$.
\end{enumerate}
\end{definition}
\begin{theorem}[Halpern-L\"{a}uchli \cite{JH66}]\label{thm.HL}
For each $i< d$,
let $T_i\subseteq \omega^{<\omega}$ be a finitely branching tree with no terminal nodes, let $N\in [\omega]^\omega$, where $d$ is any positive integer, and let
$c :
\bigcup_{n\in N}
\prod_{i<d}T_i\cap \omega^n\longrightarrow p$
be a finite coloring, where $p$ is any positive integer. Then there is an infinite subset $M\subseteq N$ and infinite
strong subtrees $U_i\subseteq \bigcup_{n\in N} T_i\cap \omega^n$ witnessed by $M$ such that $c$ is monochromatic on
$\bigcup_{m\in M}
\prod_{i<d}U_i\cap \omega^m$.
\end{theorem}
We observe that
every $S\in \mathcal{F}_{\infty}$ is an $\pmb{F_{\max}}$-tree and
every $\pmb{F_{\max}}$-tree contains some
$S\in \mathcal{F}_{\infty}$ as a subtree.
This is because for all $n< \omega$, we can find infinitely many $m\in \omega$
such that $\pmb{F_n}$ embeds into $\pmb{F_m}$
(in fact, this holds for all but finitely many $m\in \omega$.)
We define $(\pmb{A_n})_{n<\omega}$ to be a sequence of finite structures in $\mathcal{K}$
such that for each $n< \omega$, the finite structures $\pmb{F_0},\dots, \pmb{F_n}$ embed into $\pmb{A_n}$.
Now we will prove {\bf (A4)} in Lemma 4.4.
For $\mathcal{K}=\mathcal{OG}$,
Zheng proved in
\cite{YY18}
that $(\mathcal{F}_{\infty}, \leq, r)$ is a topological Ramsey space.
So the axiom {\bf (A4)} holds for finite ordered graphs.
It should be pointed out that the proof of Lemma
\ref{lem.4.4}
follows from Zheng's proof for finite ordered graphs. However, for the convenience of the reader, we also present the proof here.
\begin{lemma}\label{lem.4.4}
The axiom {\bf (A4)} holds for $(\mathcal{F}_{\infty}, \leq, r)$, i.e. for $a\in \mathcal{F}_{< \infty}$ and $S\in \mathcal{F}_{\infty}$, if
\em{depth}$_S(a)< \infty$ and $\mathcal{O}\subseteq \mathcal{F}_{|a|+1}$, then
there is $U\in [$\mbox{depth}$_S(a), S]$ such that $r_{|a|+1}[a, U]\subseteq \mathcal{O}$ or $r_{|a|+1}[a, U]\cap \mathcal{O}=\emptyset$.
\end{lemma}
\begin{proof}Let $m_0=\mbox{height}(a)$. Then $a$ has $|a|$ many splitting nodes, and each element in $\mathcal{F}_{|a|+1}$ has $|a|+1$ many splitting nodes.
In particular, we
can find $u\in a\cap \omega^{m_0}$ and $j<\omega$ such that for every $b\in \mathcal{F}_{|a|+1}$ with $a\sqsubseteq b$, there is a unique splitting node
$t\in b_u$, and $\pmb{F^b_t}
\cong \pmb{F_j}$.
Let $a\cap \omega^{m_0}$ be enumerated as $u= v_0, v_1,\dots, v_d$.
Step 1.
Let us construct a subtree $X\subseteq S$ along
with a strictly increasing
sequence $(m_l)_{l<\omega}$ starting with $m_0=$ height$(a)$ such that the following conditions
hold for every $l< \omega$ and every $s\in X\cap \omega^{m_l}$:
\begin{enumerate}
\item[(a)] $X_s$ has a unique splitting node $t$ of length in $[m_l, m_{l+1})$.
\item[(b)]
For the $t$ from condition (a), $\pmb{F^X_t}\cong \pmb{A_{j+l}}$.
\end{enumerate}
Let $t\in S_u$. The set of tuples of {\em nephews} of $t$ is defined to be
$$\{(t_1,\dots, t_d) : t_i\in X_{v_i}\cap \omega^{|t|+1} \
\mbox{for} \ 1\leq i\leq d\}.$$
Suppose $(t_i)_{1\leq i\leq d}$ is a tuple of nephews of $t$. Each finite structure $\pmb{F}\in
\big(\mathop{}_{\pmb{F_j}}^{\pmb{F^X_t}}\big)$
together with
$(t_i)_{1\leq i\leq d}$ determines an element $b_{\pmb{F},(t_i)}\in \mathcal{F}_{|a|+1}$, where the set of $\subseteq$-maximal nodes
in $b_{\pmb{F},(t_i)}$ is $F\cup \{t_i : 1\leq i\leq d\}$. With this notion of $b_{\pmb{F}, (t_i)}$ defined, we can state
another requirement for $X$.
(c) For every $t\in X$ and every tuple $(t_i)_{1\leq i\leq d}$ of nephews of $t$, the set
$\Big\{b_{\pmb{F},(t_i)} : \pmb{F}\in \big(\mathop{}_{\pmb{F_j}}^{\pmb{F^X_t}}\big)\Big\}$ is either included or disjoint from $\mathcal{O}$.
We recursively construct sets $X(m_l)\subseteq \omega^{m_l}$.
Then $X$ will be the downward closure of
$\bigcup_{l<\omega} X(m_l)$ and $X\cap \omega^{m_l}=X(m_l)$. Start with $X(m_0)=\{v_i : i\leq d\}$.
Assume we have constructed $X(m_l)$. The number of extensions in $X(m_{l+1})$ for
each $s\in X(m_l)$ is prescribed.
In particular, for each $t\supseteq u$ in $X(m_{l+1})$, the set of (tuples of)
nephews of $t$ will be finite and of the same size, independent of $t$.
Let this size be $k< \omega$.
Since $\mathcal{K}$ is a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property, there is a finite
ordered structure $\pmb{H}\in \mathcal{K}$ such that
$\pmb{H}\longrightarrow (\pmb{A_{j+l}})^{\pmb{F_{j}}}_{2^{k}}$
Since $S$ is an $\pmb{F_{\max}}$-tree, for each $s\in X(m_l)$ extending $v_i \ (1\leq i\leq d)$, $s$ has an
extension $t(s)\in S$ such that there is
$\pmb{F(s)}\in \big(\mathop{}_{\pmb{A_{j+l}}}^{\pmb{F^S_{t(s)}}}\big)$.
On the other hand, for each
$s\in X(m_l)$ extending $u$, $s$ has an extension $t(s)\in S$ such that $\pmb{H}$ embeds into $\pmb{F^S
_{t(s)}}$.
For each tuple $(t_i)_{1\leq i\leq d}$ of nephews of $t(s)$, there is a natural coloring $c : \big(\mathop{}_{ \ \pmb{F_{j}}}^{\pmb{F^S_{t(s)}}}\big)\longrightarrow 2$
depending on whether $b_{\pmb{F}, (t_i)}$ is in $\mathcal{O}$.
Thus, there are at most $k$ many
2-colorings.
These colorings can be encoded in a single $2^n$-coloring of $\big(\mathop{}_{ \ \pmb{F_{j}}}^{\pmb{F^S_{t(s)}}}\big)$.
Then there is $\pmb{H(s)}\in \big(\mathop{}_{\pmb{A_{j+l}}}^{\pmb{F^S_{t(s)}}}\big)$ such that the set $\big(\mathop{}_{ \ \pmb{F_{j}}}^{\pmb{H(s)}}\big)$
is monochromatic.
Let
$$
m_{l+1}=\max\{|t(s)| : s\in X(m_l)\}+1.
$$
Suppose that $X(m_{l+1})\subseteq S\cap \omega^{m_{l+1}}$ has the property that every node in $(\bigcup_{s\in D} F(s))$
$\cup (\bigcup_{s\in X(m_l)\backslash D} H(s))$ has a unique
extension in $X(m_{l+1})$, where $D=\{s\in X(m_l) : s$ extends $v_i,\ 1\leq i\leq d\}$. Thus $X$ satisfies (a), (b) and (c).
This finishes the construction of $X$ and $(m_l)_{l<\omega}$.
Step 2.
We use the Halpern-L\"{a}uchli theorem to shrink $X$ to an $\pmb{F_{\max}}$-tree $T$ such that
the set $\{b\in \mathcal{F}_{|a|+1} : b\subseteq T \}$ is either included in or disjoint from $\mathcal{O}$.
We define a coloring $$c :
\bigcup_{l< \omega}
\prod_{i\leq d}X_{v_i}\cap \omega^{m_l}\longrightarrow 2
$$
as follows:
Let $c(v_0, v_1,\dots, v_d)=1$.
Suppose $0<l<\omega$ and $(t_i)_{i\leq d}\in X_{v_i}\cap \omega^{m_l}$. Since $S$ is a skew tree, so is $X$.
Let
$$
sl(t_0)=\max\{|t| : t\sqsubseteq t_0 \ \mbox{and succ}_X(t)>1\}+1.
$$
Then $(t_i\upharpoonright sl(t_0))_{1\leq i\leq d}$ is a tuple of
nephews for $t_0\upharpoonright sl(t_0)-1$. Let
$$\begin{array}{ll}
c(t_0, t_1,\dots, t_d)=\left\{\begin{array}{ll}
1,&\Big\{b_{\pmb{F}, (t_i\upharpoonright sl(t_0))} : \pmb{F}\in \big(\mathop{}_{ \ \ \ \ \ \pmb{F_{j}}}^{\pmb{F^X_{t_{0}\upharpoonright sl(t_0)-1}}}\big)
\Big\}\subseteq \mathcal{O},\\
0, &\Big\{b_{\pmb{F}, (t_i\upharpoonright sl(t_0))} : \pmb{F}\in \big(\mathop{}_{ \ \ \ \ \ \pmb{F_{j}}}^{\pmb{F^X_{t_{0}\upharpoonright sl(t_0)-1}}}\big)\Big\}
\cap \mathcal{O}=\emptyset.\end{array}\right.
\end{array}$$ By (c), this colouring is well-defined. By Theorem 4.3,
there are a strictly increasing sequence $(n_j)_{j<\omega}\subseteq (m_l)_{l<\omega}$ and strong
subtrees $$Y_i\subseteq \bigcup_{l< \omega}
X_{v_i}\cap \omega^{m_l}$$ witnessed by $(n_j)_{j<\omega}$ such that $c$ is monochromatic on
$\bigcup_{j< \omega}
\prod_{i\leq d}Y_{i}\cap \omega^{n_j}$.
Let $T$ be the downward closure of $\bigcup_{i\leq d}Y_i$.
\begin{claim}
If $l< \omega$, $i\leq d$, $s\in Y_i\cap \omega^{m_l}$ and $t$ corresponds to $s$ as in (a) and (b), then $t\in T$
and $\pmb{F_t^{T}}=\pmb{F_t^X}$.
\end{claim}
\begin{proof}
Since $s\in Y_i\cap \omega^{m_l}$, there is some $p< \omega$ such that $n_p= m_l$. As $Y_i$ is a strong
subtree of $\bigcup_{l^\prime< \omega}X_{v_i}\cap \omega^{m_{l^\prime}}$, every immediate successor of $s$ in $\bigcup_{l^\prime< \omega}X_{v_i}\cap \omega^{m_{l^\prime}}$
has exactly one extension in $Y_i\cap \omega^{n_{p+1}}$. It follows from the construction of $X$ that every node in succ$_X(t)$ has exactly one
extension in $X\cap \omega^{m_{l+1}}$. Moreover, the immediate successors of $s$ in $\bigcup_{l^\prime< \omega}X_{v_i}\cap \omega^{m_{l^\prime}}$
are precisely the
extensions in $X\cap \omega^{m_{l+1}}$ of nodes in succ$_X(t)$. Then $t$ is in the downward closure of $Y_i$, and thus $t\in T$. So $\pmb{F_t^{T}}=\pmb{F_t^X}$.
\end{proof}
Then it is straightforward to check that $T$ is an $\pmb{F_{\max}}$-tree such that
the set $\{b\in \mathcal{F}_{|a|+1} : b\subseteq T\}$ is either included in or disjoint from $\mathcal{O}$. By the
observation before this lemma, we can further shrink $T$ to $U\in \mathcal{F}_{\infty}$ satisfying the
conclusion of the lemma.\end{proof}
\begin{theorem}\label{thm.FclasstRs}
For each Fra\"{\i}ss\'{e}
class of finite ordered binary relational structures with the Ramsey property and each type $\tau$,
the space $(\mathcal{F}_\infty(\tau), \leq, r)$
is a topological Ramsey space.
\end{theorem}
\vspace*{0.2cm}
\section{Finite big Ramsey degrees
for ordered binary relational universal inverse limit structures}
\vspace*{0.4cm}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures
satisfying the Ramsey property, with signature $L=\{<,R_0,\dots, R_{k-1}\}$ where each $R_i$, $i<k$, is a binary relation.
In this section,
we prove that for each such $\mathcal{K}$, the universal inverse limit structure has finite big Ramsey degrees under finite Baire-measurable colorings.
The proofs in this section are straightforward via
the
topological Ramsey spaces
from Theorem \ref{thm.FclasstRs}
(which is based on work of Zheng in \cite{YY18})
and the work of Huber-Geschke-Kojman on
inverse limits of finite ordered graphs in \cite{SH19}.
\begin{definition}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class of finite ordered binary relational structures with the Ramsey property. A {\em universal inverse limit} of finite ordered structures in $\mathcal{K}$
is a triple $\pmb{G}=\langle G, <^{\pmb{G}}, R^{\pmb{G}}_0,\dots, R^{\pmb{G}}_{k-1}\rangle$, such
that the following conditions hold.
1. $G$ is a compact subset of $\mathbb{R}\backslash\mathbb{Q}$ without isolated points,
$ <^{\pmb{G}}$ is the
restriction of the standard order on $\mathbb{R}$ to $G$, and
$ R^{\pmb{G}}_i\subseteq [G]
^2$ for each $i<k$.
2. For every pair of distinct elements $u, v\in G$, there is a partition of $G$ to finitely
many closed intervals such that
(a) $u, v$ belong to different intervals from the partition;
(b) for every interval $I$ in the partition, for all $x\in G\backslash I$ and for all $y, z\in I$,
$(x, y)\in R^{\pmb{G}}_i$ if and only if $(x, z)\in R^{\pmb{G}}_i$, for each $i<k$.
3. Every nonempty open interval of $G$ contains induced copies of all finite ordered
structures in $\mathcal{K}$.
\end{definition}
For
every $\pmb{F_{\max}}$-tree $T$, it can be seen from Definition
\ref{defn.2.6}
that $\pmb{F(T)}$ is a universal inverse limit
structure.
So it follows from the universality that we can consider colourings of finite induced substructures of $\pmb{F(T)}$.
\begin{definition}
Let $\tau$ be a type and $\pmb{H}\in \tau$. The $\mathcal{F}_\infty$-envelope of $\pmb{H}$ is
$$\mathcal{C}_{\pmb{H}} =\{U\in \mathcal{F}_\infty : (\exists \ l)(r_l(U)=\dn \pmb{H}^\vee)\},$$
where $\dn H^\vee=\{a\in \omega^{<\omega} : (\exists \ x\in H)(a\subseteq x\upharpoonright (\triangle(\pmb{H})+1))\}$.
\end{definition}
\begin{lemma}\label{lem.5.3}
Let $\tau$ be a type and $T\in \mathcal{F}_\infty$. Define a map
$c_1 : [\emptyset, T]\longrightarrow \big\{\pmb{H}^\vee : \pmb{H}\in \big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)\big\}$ as follows:
$$\forall \ U\in [\emptyset, T],\ \mbox{if} \ U\in \mathcal{C}_{\pmb{H}}, \ c_1(U)=\pmb{H}^\vee.$$
Then $c_1$ is well-defined and continuous, where we equip the range with the discrete topology.
\end{lemma}
\begin{proof}
Let $m+1$ be the number of elements for each $\pmb{H}\in \tau$. Then $\dn \pmb{H}^\vee$ has $m$ splitting nodes. Thus
$$\forall \ U\in \mathcal{F}_\infty, \forall \ l,
(r_l(U)=\dn \pmb{H}^\vee\Rightarrow l=m).$$ Let $U, V\in [\emptyset, T]$. Then there are $\pmb{H}, \pmb{K}\in \tau$ such that $U\in \mathcal{C}_{\pmb{H}}$ and
$V\in \mathcal{C}_{\pmb{K}}$. If $U=V$, then $\dn \pmb{H}^\vee=r_m(U)=r_m(V)=\dn \pmb{K}^\vee$, and thus $\pmb{H}^\vee=\pmb{K}^\vee$. So $c_1$ is well-defined.
Suppose that $\pmb{H}\in \tau$ and $U\in (c_1)^{-1}(\pmb{H}^\vee)$. We have that $U\in \mathcal{C}_{\pmb{H}}$. Then the set $[m, U]$ is an open set containing $U$ and $[m, U]\subseteq (c_1)^{-1}(\pmb{H}^\vee)$. Thus $c_1$ is continuous.\end{proof}
We equip $\omega^{\omega}$ with the first-difference metric topology, which has basic open sets of
the form $[s]=\{x\in \omega^{\omega} : s\subseteq x\}$ for $s\in \omega^{<\omega}$. For $n\in \omega$, let $[\pmb{F_{\max}}]^n$
denote the set of all induced substructures of $\pmb{F_{\max}}$ of size $n$.
\begin{definition} For $n\geq 1$, we define a topology on $[\pmb{F_{\max}}]^n$
as follows: A
set $\mathcal{U}\subseteq [\pmb{F_{\max}}]^n$
is open if for all $\pmb{H}\in \mathcal{U}$, there are open neighborhoods
$U_1, \dots, U_n$ of the elements of $H$ such that all $\pmb{H^\prime}\in [\pmb{F_{\max}}]^n$
that have exactly
one vertex in each $U_i$ are also in $\mathcal{U}$. This topology is separable and induced
by a complete metric. A coloring of $n$-tuples from $[\pmb{F_{\max}}]^n$ is continuous if it
is continuous with respect to this topology.
\end{definition}
\begin{lemma}
Let $\tau$ be a type and $T\in \mathcal{F}_\infty$. For every continuous coloring
$c : \big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)
\longrightarrow 2$, there exists an $\pmb{F_{\max}}$-subtree $S$ of $T$ such that $c$ depends only on $\pmb{H}^\vee$, i.e., for $\pmb{H}, \pmb{K}\in
\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$, if $\pmb{H}^\vee=\pmb{K}^\vee$, then $c(\pmb{H})=c(\pmb{K})$.
\end{lemma}
\begin{proof}
For $\pmb{H}\in \big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$, by
definition of $\triangle(\pmb{H})$, the map $x\longmapsto x\upharpoonright (\triangle(\pmb{H})+1))$ is a bijection from
the universe $H$ of $\pmb{H}$ onto $\pmb{H}^\vee$.
Let $t_1,\dots, t_l$ denote
the elements of $\pmb{H}^\vee$. For all $\bar{x}=(x_1,\dots, x_l)\in [T_{t_1}]\times \cdot\cdot\cdot\times [T_{t_l}]$,
the induced substructure $\pmb{\overline{H}}$ of $\pmb{F(T)}$ on the set $\{x_1,\dots, x_l\}$ is isomorphic to $\pmb{H}$.
By the continuity of $c$, for all such $x$ there are open neighborhoods $x_1\in U^{\bar{x}}_{1},\dots, x_l\in U^{\bar{x}}_{l}$
such that for all $(y_1,\dots, y_l)\in U^{\bar{x}}_{1}
\times\cdot\cdot\cdot\times U^{\bar{x}}_{l}$
for the induced
substructure $\pmb{H^\prime}$ of $\pmb{F(T)}$ on the vertices $y_1,\dots, y_l$, we have $c(\pmb{\overline{H}})=c(\pmb{H^\prime})$.
We may assume that the $U^{\bar{x}}_{i}$
are basic open sets, i.e., sets of the form $[T_r]$
for some $r\in T$. Since the space $[T_{t_1}]\times\cdot\cdot\cdot\times [T_{t_l}]$ is compact, there is a finite
set $A \subseteq
[T_{t_1}]\times\cdot\cdot\cdot\times [T_{t_l}]$ such that
$$[T_{t_1}]\times\cdot\cdot\cdot\times [T_{t_l}]=
\bigcup\limits_{\bar{x}\in A}\prod\limits_{i=1}^{l}U^{\bar{x}}_{i}.$$
Hence there is $m\in \omega$ such that for all induced substructures $\pmb{H^\prime}$ of $\pmb{F(T)}$ with $\pmb{H^\prime}\upharpoonright (\triangle(\pmb{H})+1)=\pmb{H}^\vee$, the color
$c(\pmb{H^\prime})$ only depends on $\pmb{H^\prime}\upharpoonright m$, where $m$ is the maximal length of the $r$'s with
$[T_r]= U^{\bar{x}}_{i}$
for some $\bar{x}\in F$ and $i\in \{1,\dots, l\}$.
Since for each $m\in \omega$, there are only finitely many sets of the form $\pmb{H}\upharpoonright m$,
where $H\in \big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$, there is a function $f : \omega\longrightarrow \omega$ such that for
every finite induced substructure $\pmb{H}$ of $\pmb{F(T)}$ with $\triangle(\pmb{H})+1=n$, the color $c(\pmb{H})$
only depends on $\pmb{H}\upharpoonright f(n)$.
Now let $S$ be an $\pmb{F_{\max}}$-subtree of $T$ such that
whenever $s\in S$ is a splitting node of $S$ of length $n$, then $S$ has no splitting node $t$ whose length is in the interval $(n, f(n)]$. Now for all $H\in \big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$, the color $c(\pmb{H})$ only depends on $\pmb{H}^\vee$.\end{proof}
\begin{theorem}
Let $T$ be an $\pmb{F_{\max}}$-tree. For every type $\tau$ of a finite induced substructure of $\pmb{F_{\max}}$, and every continuous coloring $c : \big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)
\longrightarrow 2$, there is an
$\pmb{F_{\max}}$-subtree $S$ of $T$ such that $c$ is monochromatic on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.
\end{theorem}
\begin{proof}We can shrink $T$ and assume
$T\in \mathcal{F}_\infty$. By Lemma 5.5, $c$ depends only on $\pmb{H}^\vee$. We may think of $c$ as a map as follows:
$$c : \Big\{\pmb{H}^\vee : \pmb{H}\in \big(\mathop{}_{\tau}^{\pmb{F(T)}}\big)\Big\}\longrightarrow 2.$$
Define $\bar{c} : [\emptyset, T ]\longrightarrow 2$ by $\bar{c}=c\circ c_1$. By Lemma 5.3, $\bar{c}$ is also a continuous
map.
By Theorem \ref{thm.FclasstRs},
there is some $S\leq T$ such that $\bar{c}$ is monochromatic on $[\emptyset, S]$.
Suppose that $\pmb{H}\in \big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.
Then the universe $H$ of $\pmb{H}$ is contained in
$[S]$, so $\pmb{H}^\vee\subseteq S$.
Hence there is a
$U\in [\emptyset, S]$ such that $U\in \mathcal{C}_{\pmb{H}}$,
and thus, $c(\pmb{H})=\bar{c}(U)$.
Therefore $c$ is monochromatic on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.
\end{proof}
The next lemma is a straightforward extension of Lemma 3.8 in \cite{SH19}.
\begin{lemma} Let $T$ be an $\pmb{F_{\max}}$-tree. For every type $\tau$ of a finite induced substructure of $\pmb{F_{\max}}$, and every Baire-measurable coloring $c : \big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)
\longrightarrow 2$, there is an
$\pmb{F_{\max}}$-subtree $S$ of $T$ such that $c$ is continuous on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.
\end{lemma}
\begin{proof}Since $c : \big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)
\longrightarrow 2$ is Baire-measurable,
$c^{-1}(0)$ and $c^{-1}(1)$ have the property of Baire. Then there exist open sets $U$, $V$ in $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$ and meager sets $M$, $N$ in $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$ such that
$$c^{-1}(0)=U\triangle M \ \mbox{and} \ c^{-1}(1)=V\triangle N,$$
where $\triangle$ denotes the symmetric difference. Let $(N_n)_{n\in \omega}$ be a sequence of
closed nowhere dense subsets of $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$
such that $M\cup N\subseteq \bigcup_{n\in \omega}N_n$. We would like to construct an $\pmb{F_{\max}}$-subtree $S$ of $T$ such that $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$ is disjoint from $\bigcup_{n\in \omega}N_n$. In this case, we have
$$c^{-1}(0)\cap \big(\mathop{}_{\tau}^{\pmb{F(S)}}\big)=U\cap \big(\mathop{}_{\tau}^{\pmb{F(S)}}\big) \ \mbox{and} \ c^{-1}(1)\cap \big(\mathop{}_{\tau}^{\pmb{F(S)}}\big)=V\cap \big(\mathop{}_{\tau}^{\pmb{F(S)}}\big).$$
It follows that $c$ is continuous on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.
In order to find an $\pmb{F_{\max}}$-subtree $S$ that is disjoint from $\bigcup_{n\in \omega}N_n$, we construct a fusion sequence $(T_j)_{j\in \omega}$ of $\pmb{F_{\max}}$-subtrees of $T$ with witness a strictly
increasing sequence $(m_j)_{j\in \omega}$ of natural numbers.
Put $S=\bigcap_{j\in \omega}T_j$. Then $S$ is an $\pmb{F_{\max}}$-subtree of $T$.
Suppose $T_j$ and $m_j$ have already been chosen. We assume that for all
$t\in T_j(m_j)$ and all $s\in T$ with $t\subseteq s$, we have $s\in T_j$. For a certain $t\in T_j(m_j)$, we have to find a splitting node $s$ with $t\subseteq s$
such that for a certain finite ordered structure $\pmb{H}$, $\pmb{H}$ embeds into $\pmb{F
^{T_{j+1}}_s}$.
Since $T_j$ is
an $\pmb{F_{\max}}$-tree, there is $m> m_j$ and an extension $s$ of $t$ with $|s|
< m$ such that $\pmb{H}$ embeds into $\pmb{F^{T_j}_s}$.
Suppose that $\pmb{H}$ is a finite substructure of $\pmb{F(T_j)}$ of type $\tau$ such that $\triangle(\pmb{H})< m$.
We list elements of $H$ as $t_0, t_1,\dots, t_p$.
The set $\pmb{H}\upharpoonright m$ determines an open
subset $O$ of $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$.
Since $\bigcup_{n\leq j}N_n$ is closed and nowhere dense in $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$,
$O$ contains a nonempty open subset that is disjoint from $\bigcup_{n\leq j}N_n$.
It
follows that for $i\in \{0,1,\dots,p\}$, $t_i\upharpoonright m$ has an extension $s_i\in T_j$ such that
the open subset of $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)$
determined by $s_0, s_1, \dots, s_p$
is disjoint from $\bigcup_{n\leq j}N_n$.
We may assume that $s_0, s_1, \dots, s_p$ have the same length $m_{j+1}> m$.
Let $X\subseteq T_j(m_{j+1})$
be a set that contains exactly one extension of every
element of $T_j(m_{j})$ and in particular the elements $s_0,s_1,\dots,s_p$. Let
$$T_{j+1}=\{t\in T_j : \exists s\in X ( s\subseteq t\vee t\subseteq s)
\}.$$
Then $T_{j+1}$ is an $\pmb{F_{\max}}$-tree.
Whenever $\pmb{H^\prime}$
is a finite substructure of $\pmb{F(T_{j+1})}$
of type $\tau$ with
$\pmb{H^\prime}\upharpoonright m=\pmb{H}\upharpoonright m$, then $\pmb{H^\prime}\upharpoonright m_{j+1}=\{s_0,s_1,\dots,s_p\}$.
In particular, $\pmb{H^\prime}\notin \bigcup_{n\leq j}N_n$. This finishes the recursive definition of the sequences $(T_j)_{j\in \omega}$
and $(m_j)_{j\in \omega}$.
Finally, let $S=\bigcap_{j\in \omega}T_j$.
One can check that $S$ is an $\pmb{F_{\max}}$-tree. Let $n\in \omega$ and let $\pmb{H}$ be a
finite substructure of $\pmb{F(S)}$ of type $\tau$. Then there is $j\in \omega$ such that $\triangle(\pmb{H})< m_j$. We
can choose $j\geq n$.
Note that $S(m_j)=T(m_j)$. Since $S\subseteq T_{j+1}$, by the construction of $T_{j+1}$, $\pmb{H}\notin
\bigcup_{n\leq j}N_n$.
In particular, $H\notin N_n$. This shows that
$\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$
is disjoint from $\bigcup_{n\in \omega}N_n$. It follows that $c$ is continuous on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.
\end{proof}
\begin{theorem}\label{thm.5.8}
Let $T$ be an $\pmb{F_{\max}}$-tree.
For every type $\tau$ of a finite induced substructure of
$\pmb{F_{\max}}$
and every Baire-measurable coloring $c :
\big(\mathop{}_{ \ \ \tau}^{\pmb{F(T)}}\big)\longrightarrow 2$, there is an $\pmb{F_{\max}}$-subtree $S$ of $T$ such that $c$ is constant on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.
\end{theorem}
\begin{proof}
By Lemma 5.7, there is an $\pmb{F_{\max}}$-subtree $U$
of $T$ such that $c$ is continuous on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(U)}}\big)$.
By Theorem 5.6, there is an $\pmb{F_{\max}}$-subtree $S$ of $U$
such
that $c$ is constant on $\big(\mathop{}_{ \ \ \tau}^{\pmb{F(S)}}\big)$.\end{proof}
\begin{theorem}\label{thm.5.9}
Let $\mathcal{K}$ be a Fra\"{\i}ss\'{e} class, in a finite signature, of finite ordered binary relational structures with the Ramsey property.
For every $\pmb{H}\in \mathcal{K}$,
there is a finite number $T(\pmb{H}, \pmb{F_{\max}})$ such that the following holds:
For every universal inverse limit structure
$\pmb{G}$ and for each finite Baire-measurable coloring of the set $\big(\mathop{}_{\pmb{H}}^{\pmb{G}}\big)$
of all copies of $\pmb{H}$ in $\pmb{G}$, there is a closed
copy $\pmb{G^\prime}$ of $\pmb{G}$ contained in
$\pmb{G}$ such that the set $\big(\mathop{}_{\pmb{H}}^{\pmb{G^\prime}}\big)$
of all copies of $\pmb{H}$ in $\pmb{G^\prime}$ has no more than $T(\pmb{H}, \pmb{F_{\max}})$ colors.
In particular, $T(\pmb{H}, \pmb{F_{\max}})$ is at most the number of types associated to $\pmb{H}$.
\end{theorem}
\begin{proof}
We list all types for $\pmb{H}$ as $\tau_0, \tau_1,\dots, \tau_{m-1}$. Since $\pmb{G}$ is a universal inverse limit structure, $\pmb{F_{\max}}$ embeds
continuously into it. Let $c$ be a finite Baire-measurable coloring of the set $\big(\mathop{}_{\pmb{H}}^{\pmb{G}}\big)$. Now work with the tree $T_{\max}$ coding $\pmb{F_{\max}}$.
Iterating Theorem \ref{thm.5.8}, there are $\pmb{F_{\max}}$-trees $T_{\max}\geq T_{\tau_0}\geq \cdots \geq T_{\tau_{m-1}}$
so
that for each $i< m$, $c$ is constant on $\big(\mathop{}_{ \ \ \tau_i}^{\pmb{F(T_{\tau_i})}}\big)$.
We take $\pmb{G^\prime}=\pmb{F(T_{\tau_{m-1}}})$. Then $\pmb{G^\prime}$ is a closed
copy of $\pmb{G}$ contained in $\pmb{G}$,
and the set $\big(\mathop{}_{\pmb{H}}^{\pmb{G^\prime}}\big)$
of all copies of $\pmb{H}$ in $\pmb{G^\prime}$ has no more than $m$ colors.
\end{proof}
\vspace*{0.1cm}
\section{Exact big Ramsey degrees for some ordered binary relational inverse limit structures}
\vspace*{0.2cm}
In this section, we find the exact big Ramsey degrees in the
inverse limit structures $\pmb{F_{\max}}$ of the following Fra\"{\i}ss\'{e} classes in an ordered binary relational signature:
Free amalgamation classes, the class of
finite ordered tournaments $\mathcal{OT}$,
and the class of finite partial orders with a linear extension $\mathcal{OPO}$.
We shall do so by first showing in Lemmas
\ref{lem.6.1},
\ref{lem.6.4},
and
\ref{lem.6.7}
that for any finite substructure $\pmb{H}$ of
$\pmb{F_{\max}}$, there is a larger finite structure $\pmb{\overline{H}}$ containing $\pmb{H}$ as an induced substructure such that
any copy of $\pmb{\overline{H}}$ in the inverse limit structure $\pmb{F_{\max}}$ must have exactly one meet in the tree $T_{\max}$.
Then in Theorem
\ref{thm.main},
we shall prove by induction on number of splitting nodes that each type persists in any subcopy of $\pmb{F_{\max}}$.
This proves that the exact big Ramsey degree for a given finite substructure $\pmb{H}$ of $\pmb{F_{\max}}$
is exactly the number of types $\tau$ representing
$\pmb{H}$ in the tree $T_{\max}$.
Given a structure $\pmb{G}\le \pmb{F_{\max}}$, let $T_{\pmb{G}}=\{x\upharpoonright n : x\in \pmb{G}, n\in \omega\}$. Then $T_{\pmb{G}}$ is a subtree of $T_{\max}$.
Given a tree $T$, its {\em stem}, denoted stem$(T)$, is the minimal splitting node in $T$.
\begin{lemma}\label{lem.6.1}
Let $\mathcal{K}$ be any
\Fraisse\ class with free amalgamation
in an ordered binary relational signature.
Then for each $\pmb{H}\in \mathcal{K}$,
there is a structure $\pmb{\overline{H}}\in \mathcal{K}$ containing a copy of $\pmb{H}$, where $\pmb{\overline{H}}$ has the following property:
Given a universal inverse limit structure
$\pmb{G}$ for $\mathcal{K}$ contained in
$\pmb{F_{\max}}$, every copy $\pmb{\overline{I}}$ of $\pmb{\overline{H}}$ in
$\pmb{G}$ has induced a subtree $T_{\pmb{\overline{I}}}$ of $T_{\pmb{G}}$ such that
the type of $T_{\pmb{\overline{I}}}$ has exactly one splitting node.
It follows that the
immediate successors of stem$(T_{\pmb{\overline{I}}})$ in $T_{\pmb{\overline{I}}}$ have a copy of $\pmb{\overline{I}}$.
\end{lemma}
\begin{proof}
Let $\mathcal{K}$ be
as in the hypotheses,
let $\pmb{G}$ be the universal inverse limit for $\mathcal{K}$ contained in
$\pmb{F_{\max}}$, and fix $\pmb{H}$ a finite substructure of
$\pmb{G}$.
Then $\pmb{H}$ is in $\mathcal{K}$.
Let $m$ be the size of the universe of $\pmb{H}$.
We construct a finite ordered structure
$\pmb{\overline{H}}\in \mathcal{K}$ of size $2m+1$,
containing a copy of $\pmb{H}$ as a substructure on the odd indexed vertices,
as follows.
Let $R$ denote the
binary relation symbol $R_0$ in the signature of
$\mathcal{K}$.
\begin{enumerate}
\item
Let $\overline{H}=\{v_0,v_1,v_2, \dots , v_{2m}\}$.
\item
$\pmb{\overline{H}}\upharpoonright\{v_1, v_3,\dots, v_{2m-1}\}$ is isomorphic to $\pmb{H}$.
\item
For $i\in \{0, 2,\dots, 2m-2\}$,
$R^{\pmb{\overline{H}}}(v_i, v_{i+2})$ holds.
If $R^{\pmb{\overline{H}}}$ is a symmetric relation, then also
$R^{\pmb{\overline{H}}}(v_{i+2}, v_i)$ holds;
otherwise, $\neg R^{\pmb{\overline{H}}}(v_{i+2}, v_i)$ holds.
\item
No other relations are added to $\pmb{\overline{H}}$.
\end{enumerate}
Then $\pmb{\overline{H}}$ contains a copy of $\pmb{H}$. Now we check that $\pmb{\overline{H}}$ satisfies the property in this lemma.
Let $\pmb{\overline{I}}$ be a copy of $\pmb{\overline{H}}$ in $\pmb{G}$ with $\overline{I}=\{u_0,u_1,u_2, \dots, u_{2m}\}$.
\begin{claim}\label{claim.6.2}
Let $\pmb{J}$ be the induced substructure
of $\pmb{\overline{I}}$ on universe $J=\{u_0,u_2, \dots , u_{2m}\}$.
Then the associated subtree $T_{\pmb{J}}$ of $T_{\pmb{G}}$ has exactly one splitting node.
\end{claim}
\begin{proof}
Without loss of generality, it suffices to prove that $u_0\cap u_2=u_2\cap u_4$.
Assume to the contrary that $u_0\cap u_2\neq u_2\cap u_4$. Then either
$u_0\cap u_2\subset u_2\cap u_4$ or else $u_2\cap u_4\subset u_0\cap u_2$,
where $\subset$ denotes proper subset.
If
$u_0\cap u_2\subset u_2\cap u_4$
(see Figure 20),
then
$u_0\cap u_2= u_0\cap u_4$;
let $s$ denote this node and let $l-1$ denote its length.
Then $u_2\upharpoonright l= u_4\upharpoonright l$ is a successor of $s$,
and
the relation $R(u_0\upharpoonright l ,u_4\upharpoonright l)$
holds in $T_{\max}$ since $R(u_0\upharpoonright l, u_2\upharpoonright l)$ holds in $T_{\max}$.
Hence,
$R^{\pmb{I}}(u_0, u_4)$ holds.
Similarly, if $u_2\cap u_4\subset u_0\cap u_2$,
then $u_0\cap u_4=u_2\cap u_4$, and it follows that
$R^{\pmb{I}}(u_0, u_4)$ holds (see Figure 20).
But this contradicts the fact that $R^{\pmb{I}}(u_0,u_4)$ does not hold, by (3) in the defnition
of $\pmb{\overline{H}}$ above.
Therefore, it must be the case that
$u_0\cap u_2= u_2\cap u_4$.
Thus, every copy of $\pmb{J}$ has induced subtree $T_{\pmb{J}}$ in $T_{\pmb{G}}$ with exactly one splitting node.
\end{proof}
\begin{center}
\centering
\includegraphics[totalheight=2.0in]{21}\\Figure 20
\end{center}
\begin{claim}\label{claim.6.3}
The subtree $T_{\pmb{\overline{I}}}$ of $T_{\pmb{G}}$
induced by $\pmb{\overline{I}}$ has exactly one splitting node.
\end{claim}
\begin{proof}
By Claim \ref{claim.6.2}, without loss of generality, it suffices to prove that $u_0\cap u_1=u_1\cap u_2$. Assume that $u_0\cap u_1\neq u_1\cap u_2$.
Then either
$u_0\cap u_1\subset u_1\cap u_2$ or $u_1\cap u_2\subset u_0\cap u_1$.
Since $u_0<u_1<u_2$ in the linear order $<$ on the universe of
$\pmb{F_{\max}}$,
it follows that the set $\{u_0,u_1,u_2\}$ has type equal to one of the two types in Figure 21 (the solid lines).
Since $R^{\pmb{\overline{I}}}(u_0, u_2)$ holds,
at least one of
$R^{\pmb{\overline{I}}}(u_0, u_1)$ or
$R^{\pmb{\overline{I}}}(u_1, u_2)$
holds (the dashed lines in Figure 21).
But this contradicts the fact that $ \neg R^{\pmb{\overline{I}}}(u_0, u_1)$ and
$\neg R^{\pmb{\overline{I}}}(u_1, u_2)$ hold, by (4) in the definition of $\pmb{\overline{H}}$.
Thus, the induced subtree $T_{\pmb{\overline{I}}}$ has exactly one
splitting node.\end{proof}
\begin{center}
\centering
\includegraphics[totalheight=2.0in]{22}\\Figure 21
\end{center}
By Claim \ref{claim.6.3}, the induced subtree $T_{\pmb{\overline{I}}}$ of $T_{\pmb{G}}$
has exactly one splitting node, namely its stem.
It follows from the definition of the relations on
$\pmb{F_{\max}}$ that
the immediate successors of stem$(T_{\pmb{\overline{I}}})$ in $T_{\pmb{\overline{I}}}$ is isomorphic to
$\pmb{\overline{H}}$.
\end{proof}
Next, we prove a similar lemma for the class of finite ordered tournaments, $\mathcal{OT}$.
The proof is similar to the previous lemma, the main difference being that the construction of $\pmb{\overline{H}}$ must take into account the fact that any two vertices of a tournament must have some directed edge relation between them.
\begin{lemma}\label{lem.6.4}
Let $\mathcal{K}$ be $\mathcal{OT}$.
Then for each $\pmb{H}\in \mathcal{K}$,
there is a structure $\pmb{\overline{H}}\in \mathcal{K}$ containing a copy of $\pmb{H}$
and $\pmb{\overline{H}}$ has the following property:
Given a universal inverse limit structure
$\pmb{G}$ for $\mathcal{K}$
contained in $\pmb{F_{\max}}$,
every copy $\pmb{\overline{I}}$ of $\pmb{\overline{H}}$ in
$\pmb{G}$ has induced a subtree $T_{\pmb{\overline{I}}}$ of $T_{\pmb{G}}$ such that
the type of $T_{\pmb{\overline{I}}}$ has exactly one splitting node.
It follows that the
immediate successors of stem$(T_{\pmb{\overline{I}}})$ in $T_{\pmb{\overline{I}}}$ have a copy of $\pmb{\overline{I}}$.
\end{lemma}
\begin{proof}
Fix any $\pmb{H}\in \mathcal{K}$, and let
$m$ be the size of the universe of
$\pmb{H}$.
Recall that the relation $R$ here is a directed edge.
We construct a finite ordered structure
$\pmb{\overline{H}}\in \mathcal{K}$ containing a copy of
$\pmb{H}$ as an induced substructure as follows:
\begin{enumerate}
\item
Let $\overline{H}=\{v_0,v_1,v_2, \dots , v_{2m}\}$.
\item
$\pmb{\overline{H}}\upharpoonright\{v_1, v_3,\dots, v_{2m-1}\}$ is isomorphic to $\pmb{H}$.
\item
For $i, j\in \{0, 2,\dots, 2m\}$ with $i< j$, if $j=i+2$
then
$R^{\pmb{\overline{H}}}(v_i, v_j)$ holds.
\item
For $i, j\in \{0, 2,\dots, 2m\}$ with $i< j$, if $j\neq i+2$
then
$R^{\pmb{\overline{H}}}(v_j, v_i)$ holds.
\item
For all $i\in \{0, 2,\dots, 2m\}$ and $j\in \{1, 3,\dots, 2m-1\}$,\ if $i< j$ then
$R^{\pmb{\overline{H}}}(v_j,v_i)$ holds.
\item
For all $i\in \{0, 2,\dots, 2m\}$ and $j\in \{1, 3,\dots, 2m-1\}$,\ if $j< i$ then
$R^{\pmb{\overline{H}}}(v_i,v_j)$ holds.
\end{enumerate}
By (2), $\pmb{\overline{H}}$ contains a copy of $\pmb{H}$. Now we check that $\pmb{\overline{H}}$ satisfies the property in this lemma.
Let $\pmb{\overline{I}}$ be any copy of $\pmb{\overline{H}}$ in $\pmb{G}$, say with universe $\overline{I}=\{u_0,u_1,u_2, ..., u_{2m}\}$.
\begin{claim}\label{claim.6.5}
Let $\pmb{J}$ be the induced substructure
of $\pmb{\overline{I}}$ on universe $J=\{u_0,u_2, ..., u_{2m}\}$.
Then
the associated subtree $T_{\pmb{J}}$ of $T_{\pmb{G}}$ has exactly one splitting node.
\end{claim}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{23}\\Figure 22
\end{center}
\begin{proof}
Without loss of generality, it suffices to prove that $u_0\cap u_2=u_2\cap u_4$.
Assume that $u_0\cap u_2\neq u_2\cap u_4$.
Then
$u_0\cap u_2\subset u_2\cap u_4$ or $u_2\cap u_4\subset u_0\cap u_2$.
Since $R^{\pmb{J}}(u_0, u_2)$ and
$R^{\pmb{J}}(u_2, u_4)$ hold,
it follows that
$R^{\pmb{J}}(u_0, u_4)$ holds (see the dashed arrows in Figure 22).
This contradicts the fact that $R^{\pmb{J}}(u_4, u_0)$ holds, by (4) in the definition of $\pmb{\overline{H}}$ (solid arrows from $u_4$ to $u_0$ in Figure 22).
Thus, $u_0\cap u_2= u_2\cap u_4$.
It follows that
$T_{\pmb{J}}$ has exactly one splitting node, namely its stem.
\end{proof}
\begin{claim}\label{claim.6.6}
The subtree $T_{\pmb{\overline{I}}}$ of $T_{\pmb{G}}$ induced by $\pmb{\overline{I}}$ has exactly one splitting node.
\end{claim}
\begin{proof}
By Claim \ref{claim.6.5}, without loss of generality, it suffices to prove that $u_0\cap u_1=u_1\cap u_2$.
Assume that $u_0\cap u_1\neq u_1\cap u_2$.
Then either
$u_0\cap u_1\subset u_1\cap u_2$ or
$u_1\cap u_2\subset u_0\cap u_1$.
Since $u_0<u_1<u_2$ in the linear order $<$ on the universe of
$\pmb{F_{\max}}$,
it follows that the set $\{u_0,u_1,u_2\}$ has type equal to one of the two types in Figure 23 (the solid arrows).
Since $R^{\pmb{\overline{I}}}(u_0, u_2)$,
either
$ R^{\pmb{\overline{I}}}(u_1, u_2)$ or
$R^{\pmb{\overline{I}}}(u_0, u_1)$ holds
(see the dashed arrows in Figure 23).
This contradicts the facts that $ R^{\pmb{\overline{I}}}(u_1, u_0)$ and $R^{\pmb{\overline{I}}}(u_2, u_1)$ hold, by (5) and (6) of the definition of
$\pmb{\overline{H}}$.
Thus, the induced subtree $T_{\pmb{\overline{I}}}$ has exactly one
splitting node.
\end{proof}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{24}\\Figure 23
\end{center}
By Claim \ref{claim.6.6}, $\pmb{\overline{H}}$ satisfies the property in this lemma.
\end{proof}
Finally,
we prove a similar lemma for the class of finite
partial orders with a linear extension,
$\mathcal{OPO}$.
\begin{lemma}\label{lem.6.7}
Let $\mathcal{K}$ be $\mathcal{OPO}$.
Then for each $\pmb{H}\in \mathcal{K}$,
there is a structure $\pmb{\overline{H}}\in \mathcal{K}$ containing a copy of $\pmb{H}$
and $\pmb{\overline{H}}$ has the following property:
Given a universal inverse limit structure
$\pmb{G}$ for $\mathcal{K}$
contained in $\pmb{F_{\max}}$,
every copy $\pmb{\overline{I}}$ of $\pmb{\overline{H}}$ in
$\pmb{G}$ has induced a subtree $T_{\pmb{\overline{I}}}$ of $T_{\pmb{G}}$ such that
the type of $T_{\pmb{\overline{I}}}$ has exactly one splitting node.
It follows that the
immediate successors of stem$(T_{\pmb{\overline{I}}})$ in $T_{\pmb{\overline{I}}}$ have a copy of $\pmb{\overline{I}}$.
\end{lemma}
\begin{proof}
Fix any $\pmb{H}\in \mathcal{K}$, and let
$m$ be the size of the universe of
$\pmb{H}$.
If $\pmb{H}$ has universe of size one, then there is nothing to prove, so assume that the universe of $\pmb{H}$ has size $m\ge 2$.
The relation $R$ here is a partial order, where $R(v,w)$ denotes that $v$ is $R$-less than or equal to $w$.
We construct a finite ordered structure
$\pmb{\overline{H}}\in \mathcal{K}$ containing a copy of
$\pmb{H}$ as an induced substructure as follows:
\begin{enumerate}
\item
Let $\overline{H}=\{v_0,v_1,v_2, \dots , v_{2m+2}\}$.
\item
$\pmb{\overline{H}}\upharpoonright\{v_2, v_4,\dots, v_{2m}\}$ is isomorphic to $\pmb{H}$.
\item
$R(v_0,v_3)$ and $R(v_{2m-1},v_{2m+2})$ hold.
\item
For each $i \in \{1,3,\dots, 2m-3\}$,
$R^{\pmb{\overline{H}}}(v_i, v_{i+4})$ holds.
\item
The $R$-relations above are closed under transitivity of $R$, and no other $R$ relations are added.
\end{enumerate}
By (2), $\pmb{\overline{H}}$ contains a copy of $\pmb{H}$.
Note that there are no $R$-relations between any vertices in $\{v_2,v_4,\dots,v_{2m}\}$ and any vertices $\{v_0,v_1,v_3,v_5,\dots,v_{2m+1},v_{2m}\}$.
For example, if $\pmb{H}\in \mathcal{OPO}$ with 3 vertices as in Figure 24, then $\pmb{\overline{H}}$ has 9 vertices as in Figure 25. The copy of $\pmb{H}$ in $\pmb{\overline{H}}$ is on vertices $\{v_2, v_4, v_6\}$.
\begin{center}
\centering
\includegraphics[totalheight=0.4in]{25}\\Figure 24
\end{center}
\begin{center}
\centering
\includegraphics[totalheight=1.2in]{26}\\Figure 25
\end{center}
Now we check that $\pmb{\overline{H}}$ satisfies the property in this lemma.
Let $\pmb{\overline{I}}$ be any copy of $\pmb{\overline{H}}$ in $\pmb{G}$, say with universe $\overline{I}=\{u_0,u_1,u_2, \dots, u_{2m+2}\}$.
\begin{claim}\label{claim.6.8}
Let $\pmb{J}$ be the induced substructure
of $\pmb{\overline{I}}$ on universe $J=\{u_0,u_1,u_3, \dots,$ $u_{2m-1}, u_{2m+1}, u_{2m+2}\}$.
Then
the associated subtree $T_{\pmb{J}}$ of $T_{\pmb{G}}$ has exactly one splitting node.
\end{claim}
\begin{proof}
It suffices to
prove that
any three successive vertices in $J$ have the same meet.
Without loss of generality, it suffices to prove that
$u_0\cap u_1=u_1\cap u_3$,
as the same argument shows that
for any $0\le i\le m-2$,
$u_{2i+1}\cap u_{2i+3} = u_{2i+3}\cap u_{2i+5}$, and that
$u_{2m-1}\cap u_{2m+1}= u_{2m+1}\cap u_{2m+2}$.
Assume that $u_0\cap u_1\neq u_1\cap u_3$.
Then
$u_0\cap u_1\subset u_1\cap u_3$ or $u_1\cap u_3\subset u_0\cap u_1$.
Note that $R^{\pmb{J}}(u_0, u_3)$ holds
and
$\neg R^{\pmb{J}}(u_0, u_1)$ and $\neg R^{\pmb{J}}(u_1, u_3)$ hold.
If
$u_0\cap u_1\subset u_1\cap u_3$
then $R^{\pmb{J}}(u_0,u_1)$, a contradiction.
If
$u_1\cap u_3\subset u_0\cap u_1$, then
$R^{\pmb{J}}(u_1,u_3)$, also a contradiction
(see Figure 26).
Thus, $u_0\cap u_1= u_1\cap u_3$.
It follows that
$T_{\pmb{J}}$ has exactly one splitting node, namely its stem.
\end{proof}
\begin{center}
\centering
\includegraphics[totalheight=2in]{27}\\Figure 26
\end{center}
\begin{claim}\label{claim.6.9}
The subtree $T_{\pmb{\overline{I}}}$ of $T_{\pmb{G}}$ induced by $\pmb{\overline{I}}$ has exactly one splitting node.
\end{claim}
\begin{proof}
We will first prove that
for all $0\le i\le m-2$,
$u_{2i+1}\cap u_{2i+2}=u_{2i+2}\cap u_{2i+5}$,
and that
$u_{2m-1}\cap u_{2m}=u_{2m}\cap u_{2m+2}$.
Since the argument is the same for each of these cases,
it suffices to prove that $u_1\cap u_2= u_2\cap u_5$.
In fact, the same argument as that in Claim \ref{claim.6.8} applies here, for the structure
$\pmb{\overline{I}}$ restricted to the vertices $\{u_1,u_2, u_5\}$ is
isomorphic to the one in the proof of Claim \ref{claim.6.8}:
$R^{\pmb{\overline{I}}}(u_1,u_5)$,
$\neg R^{\pmb{\overline{I}}}(u_1,u_2)$, and
$\neg R^{\pmb{\overline{I}}}(u_2,u_5)$.
Since the linear order $<$ extends $R^{\pmb{\overline{I}}}$,
$\neg R^{\pmb{\overline{I}}}(u_2,u_1)$,
$\neg R^{\pmb{\overline{I}}}(u_5,u_1)$, and
$\neg R^{\pmb{\overline{I}}}(u_5,u_2)$
all hold.
Therefore, $u_1\cap u_2= u_2\cap u_5$ (see Figure 27).
Given any three vertices $u_{2i+1}, u_{2i+2}, u_{2i+3}$, where $0\le i\le m-1$,
let $w=u_{2i+5}$ if $i<m-1$ and $w= u_{2m+2}$ if $i=m-1$.
By the above argument,
$$
u_{2i+1}\cap u_{2i+2}= u_{2i+2}\cap w=u_{2i+1}\cap w.
$$
By Claim \ref{claim.6.8},
$$
u_{2i+1}\cap u_{2i+3}=u_{2i+3}\cap w=u_{2i+1}\cap w.
$$
Therefore,
$$
u_{2i+1}\cap u_{2i+2} =
u_{2i+1}\cap w=
u_{2i+2}\cap u_{2i+3}.
$$
Hence, also $u_{2i+1}\cap u_{2i+2}= u_{2i+1}\cap u_{2i+3}$.
Thus, any three successive vertices in
$\pmb{\overline{I}}$ have a common meet, meaning that
$\pmb{\overline{I}}$ has exactly one
splitting node.
\end{proof}
\begin{center}
\centering
\includegraphics[totalheight=2.2in]{28}\\Figure 27
\end{center}
By Claim \ref{claim.6.9}, $\pmb{\overline{H}}$ satisfies the property in this lemma.
\end{proof}
The next theorem shows that the upper bounds proved in
Theorem \ref{thm.5.9} are exact for ordered binary relational free amalgamation classes, for ordered tournaments, and for partial orders with a linear extension.
\begin{theorem}\label{thm.main}
Let $\mathcal{K}$ be any
\Fraisse\ class
in an ordered binary relational signature
such that either $\mathcal{K}$ has free amalgamation
or $\mathcal{K}$ is one of
$\mathcal{OT}$ or $\mathcal{OPO}$.
Let $\pmb{G}$ be
a universal inverse limit structure for $\mathcal{K}$ contained in
$\pmb{F_{\max}}$.
Then for each $\pmb{H}\in \mathcal{K}$, each type representing $\pmb{H}$ in $\pmb{G}$ persists in each closed subcopy of $\pmb{G}$.
It follows that
the big Ramsey degree $T(\pmb{H}, \pmb{F_{\max}})$
for finite Baire-measurable colorings of
$\big(\mathop{}_{ \ \ \pmb{H}}^{\pmb{F_{\max}}}\big)$
is exactly the number of types in $T_{\max}$ representing a copy of $\pmb{H}$.
\end{theorem}
\begin{proof}
Let $\pmb{G}\subseteq \pmb{F_{\max}}$ be a universal inverse limit structure for $\mathcal{K}$.
We will prove by induction on the number of splitting nodes that every type $\tau$ for each finite structure in $\mathcal{K}$ persists in $\pmb{G}$.
Suppose $\tau$ is a type for a structure in $\mathcal{K}$ which has no splitting nodes.
Then $\tau$ codes a single element,
so there is a copy of $\tau$ in $T_{\pmb{G}}$.
Now assume that $n\ge 1$ and for each type $\tau$ with less than $n$ many splitting nodes,
the type $\tau$ appears in $T_{\pmb{G}}$.
For the induction step,
let $\tau$ be a type for a structure in $\mathcal{K}$ with exactly $n$ splitting nodes.
Let $s$ denote the splitting node of longest length in $\tau$, and
let $\sigma=\tau\upharpoonright |s|$.
By the induction hypothesis, there is a copy of $\sigma$ in $T_{\pmb{G}}$.
Then there is a subtree $U$ of $T_{\pmb{G}}$ such that $U$ has type $\sigma$.
Let $\varphi : \sigma\longrightarrow U$ be the strong isomorphism from $\sigma$ to $U$,
and let $u=\varphi(s)$.
Suppose that $\pmb{F}$ is the finite structure in $\mathcal{K}$ at the immediate successors of $s$ in $\tau$.
Let
$\pmb{\overline{F}}\in \mathcal{K}$
be the structure
containing a copy of $\pmb{F}$ satisfying the properties in Lemma \ref{lem.6.1}, \ref{lem.6.4}, or \ref{lem.6.7}, respectively.
Since $\pmb{G}$ is a universal inverse limit structure for $\mathcal{K}$, there is a copy
$\pmb{\overline{H}}$ of $\pmb{\overline{F}}$ in the open interval $N_u$.
Taking $t=\mbox{stem}(T_{\pmb{\overline{H}}})$,
then succ$_{T_{\pmb{\overline{H}}}}(t)$ contains a copy of $\pmb{\overline{F}}$,
and thus succ$_{T_{\pmb{\overline{H}}}}(t)$
contains a copy $\pmb{H}$ of $\pmb{F}$.
Let $X$ denote the set of nodes in succ$_{T_{\pmb{\overline{H}}}}(t)$ forming the universe of
$\pmb{H}$.
Let $Y$ be a set of nodes $\{y_z:z\in U\setminus \{u\}\}$ of length $|t|+1$ such that for each $z\in U\setminus \{u\}$,
$y_z\supseteq z$.
Then $U\cup \{t\}\cup X\cup Y$ is
a copy of $\tau$.
Hence, $\tau$ persists in $\pmb{G}$.
Now
given a structure $\pmb{H}\in\mathcal{K}$,
let $\tau_0,\dots,\tau_{m-1}$ list the collection of all types
for copies of $\pmb{H}$ in $\pmb{F_{\max}}$.
Let $c :\big(\mathop{}_{ \ \ \pmb{H}}^{\pmb{F_{\max}}}\big)\longrightarrow m$
be defined by $c(\pmb{J})=i$ if and only if $T_{\pmb{J}}$ has type $\tau_i$.
Note that $c$ is in fact continuous, hence Baire-measurable.
By the above argument, there is a substructure $\pmb{G}$ of $\pmb{F_{\max}}$ which is again a universal inverse limit structure with the property that for each $i<m$, there is a copy of $\pmb{H}$ in $\pmb{G}$ with type $\tau_i$.
Therefore, all colors $i<m$ persist in $\pmb{G}$.
Therefore, $T(\pmb{H},\pmb{F_{\max}})\ge m$.
By Theorem \ref{thm.5.9}, we know that $T(\pmb{H},\pmb{F_{\max}})\le m$ for finite Baire-measurable colorings.\
Therefore, $T(\pmb{H},\pmb{F_{\max}})$ is exactly the number of types associated to $\pmb{H}$.
\end{proof}
\begin{remark}
Theorem \ref{thm.main} characterizes the exact big Ramsey degrees
under the finite Baire-measurable colorings
for some ordered structures with one (non-order) binary relation.
It seems likely that similar methods can be developed
to characterize the exact big Ramsey degrees for all structures considered in this paper in terms of the number of types.
\end{remark}
\vspace{0.2cm}
{\bf Acknowledgements}\ \ The authors would like to thank the anonymous referee for carefully checking the original draft and giving us many helpful
suggestions for improvements.
\bibliographystyle{amsplain}
| ab59063ea67ab413dc4cfd0c1a68144b15f0d987 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{introduction}
Chiral magnets are special examples of magnetic materials where
the energy functional which describes the system contains a term
with a preferred chirality for the magnetization vector
\cite{NT,YOKPHMNT,MBJPRNGB,RHMBWVKW}.
This chirality preference comes from the parity (inversion) violating
(noncentrosymmetric) interaction called the Dzyaloshinskii-Moriya
(DM) interaction consisting
of an inner product of the magnetization vector $\vec{n}$
with the curl $\nabla \times \vec{n}$ of the
magnetization vector \cite{Dzyaloshinskii,Moriya}.
The chiral magnet models are considered in various
spatial dimensions with a variety of potentials for the
magnetization vector besides the DM interaction term and the
exchange term (square of derivative of magnetization vector).
The ground state of the system is a translationally invariant
homogeneous configuration, if the DM interaction is weak compared
to the potential for the magnetization vector.
If the DM interaction becomes more important than the potential,
however, spatially modulated solutions become more favorable,
and can become the ground state, which is spatially inhomogeneous
and breaks the translation invariance.
The spiral phase is one such interesting phase, with
various terminologies used for various specific cases
in the existing literature for magnetic systems.
One of the first and simplest examples is the case
of no potential (with only exchange term and a DM interaction),
which is called a chiral helimagnet.
When the potential consists only of a Zeeman term (a
term proportional to the component $n_3$ of the magnetization vector),
the spiral solution is called a chiral soliton lattice.
These specific spiral solutions are discussed in detail in~\cite{KO2015,TOPSK_2018}.
Both of these terms refer to spiral solutions in a chiral
magnet without an anisotropy term (quadratic in the component $n_3$
of the magnetization vector) in the potential.
In this work, we will use the terminology of spiral solutions
for generic cases\cite{LSB}, and use the term chiral soliton lattice
for the specific case of no anisotropy, although
chiral soliton lattice is sometimes used for generic spiral
solutions~\cite{MRG2016}.
In high energy physics, a similar situation occurs,
with chiral soliton lattice states possible due to both magnetic
effects~\cite{Brauner:2016pko} and vortical effects~\cite{Nishimura:2020odq}.
Another related one-dimensional modulated state is the nematic
phase of chiral nematic liquid crystals \cite{deGennes,Dierking,Collings}.
The Frank free energy which describes liquid crystals is known
to become equivalent to the free energy of a chiral magnet in a
suitable limit~\cite{FKATDS} and skyrmion configurations can be
observed experimentally in liquid crystals.
This correspondence is also commented on in~\cite{KO2015}.
Another interesting inhomogeneous phase is the skyrmion lattice
phase~\cite{LSB}.
Skyrmion solutions were first studied in \cite{BY} in two dimensions.
Among the spatially modulated solutions, magnetic skyrmions are
intrinsically two dimensional \cite{BY,BH}.
In contrast, the spiral solutions are modulated along only one
spatial dimension, and persist in the one dimensional system.
In some special situations, spiral solutions have been constructed
exactly~\cite{Han_2010, KO2015}.
In the case of antiferromagnetic materials, various interaction
terms due to the sublattice structure are considered phenomenologically,
and various inhomogeneous solutions have been extensively
studied and the phase diagram obtained~\cite{BRWM,BS1999}.
For a different combinations of potential and the DM interaction
term in one spatial dimension, instanton solutions have been
exhaustively studied~\cite{Hongo:2019nfr}.
Effective low-energy field theories in the spiral phase has
been worked out to yield anisotropic dispersion
relations\cite{Hongo:2020xaw}.
When one considers chiral magnets in three spatial dimensions, a
richer phase structure emerges including modulated solutions
in three spatial directions such as the cone or elliptic cone
phases \cite{Rowland_2016}.
If we restrict ourselves to one spatial dimension, we obtain
the one-dimensional chiral magnet.
It can offer a simpler system where we can study spiral solutions
extensively.
The field equations
for a one dimensional model of a chiral
magnet are closely related to the double sine-Gordon model.
When there is no DM interaction, the model has been studied
extensively, to obtain in particular domain wall (kink) solutions
and their thermodynamic properties for the double sine-Gordon chain
\cite{Condat1983}.
The double sine-Gordon model also appears in high energy
contexts such as two Higgs doublet models
\cite{Eto:2018hhg,Eto:2018tnk} and $^3P_2$ neutron superfluids
\cite{Chatterjee:2016gpm}, but without an interaction of the DM type.
One should note, however, that the presence of the DM interaction
is crucially important to understand the energetics of spiral solutions
and the phase structure of chiral magnets.
The double sine-Gordon domain wall solutions have also been
discussed in the context of two dimensional chiral magnets,
in particular at their
one-dimensional edges~\cite{MRG2016}.
The purpose of our paper is to introduce a model of a chiral
magnet in one-spatial dimension to obtain the phase diagram
exactly, and clarify its relation to energetics of domain wall solutions.
Alongside the usual energy term for the Heisenberg ferromagnet
the model includes a Dzyaloshinskii-Moriya (DM) interaction term
with the coefficient $\kappa$ \cite{Dzyaloshinskii,Moriya} and
a potential which is a sum of a Zeeman term (linear in $n_3$)
with the coefficient $B$ and an anisotropy term (quadratic in
$n_{3}$) with the coefficient $A$.
We find that the boundary energy functional needed to obtain the
Euler-Lagrange equation (through the variational principle) does
not contribute to the solutions of interest to us here, such as spiral
or domain wall solutions.
We exhaustively work out the static solutions of the model, and
find that spiral solutions have a continuous spectra of average
energy density, and the lowest energy spiral solution gives the
ground state in the spiral phase (for small potential parameters
$A, B$). On the other hand, homogeneous (polarized or canted
ferromagnetic) solutions give the ground state for large $A, B$.
We demonstrate that the phase boundary between the spiral phase
and the homogeneous (ferromagnetic) phases is characterized by
the emergence of zero energy domain wall solutions.
In the homogeneous (ferromagnetic) phase, the domain wall solution
is a soliton as a finite positive energy excited state.
In the spiral phase region, the domain wall solutions,
as solitons above the homogeneous background, have negative
energy signaling the instability of the homogeneous solution.
The zero energy domain wall solutions are obtained as the
(infinite period) limit of the spiral solutions when the phase
boundary is approached from the spiral phase region.
The exact phase boundary in the $A/\kappa^2, B/\kappa^2$ plane with $2A\leq \vert B\vert$
is worked out explicitly. That is the transition between the spiral phase and the homogeneous, polarised, ferromagnetic phase which we show to be a second order transition.
We also obtain exact domain wall solutions explicitly
for all parameter regions and confirm our general argument
on the phase diagram.
After finishing this work we became aware of the recent
paper~\cite{Paterson}, as well as the older work \cite{IL1983}.
The former obtains the phase boundary between homogeneous phases
and the spiral phase using explicit spiral solutions in terms of
elliptic functions, although they consider a different
physical situation, namely with an external elastic strain applied
to the chiral magnet.
We have approached the problem from a complementary
point of view, with general arguments using inequalities to show
the region of the spiral phase.
In particular, we have clarified the role of domain wall solutions
explicitly to work out the phase boundary, and also have included
more detail about the domain wall limit of the spiral solutions
at the phase boundary.
We have also explicitly derived the order of phase transition
across the phase boundary.
Another related work is \cite{Chovan} where a similar model, with just an anisotropy term and no magnetic field is considered. The model is studied numerically and two spiral phases are discussed; a ``flat" spiral similar to what we construct here, and a ``non-flat" spiral. The difference between the two types of spirals is apparent in a spherical coordinate decomposition of the magnetisation field. For flat spirals the angular variable $\Phi(x)$ is a constant, while for a non-flat spiral it is allowed to vary. There is a numerical evidence that the non-flat spiral has lower energy than the flat spiral close to the transition to the homogeneous phase, and that it determines the location of the phase transition. The non-flat spiral has no analytic expression, here we focus on the flat spiral, just called spiral throughout, configurations which can be constructed explicitly. This is why we restrict to $2A\leq\vert B\vert$, as for $2A>\vert B\vert$ the non-flat spiral is the ground state and the phase transition cannot be found analytically. We still discuss details about spiral and domain wall solutions in this parameter region.
Our model can be obtained as a dimensional reduction of the most
popular model of two-dimensional chiral
magnets~\cite{BY,BH,BRS,Schroers1,RSN,DM,Melcher}.
Hence one can expect that all the solutions that we consider
are also solutions of chiral magnet models in two
or higher spatial dimensions.
We would only need to additionally study other solutions specific to two spatial
dimensions in order to determine the phase diagram of a two dimensional model.
A similar problem was studied for the case of noncentrosymmetric
uniaxial antiferromagnetic materials \cite{BRWM}, where the authors
considered various interaction terms arising from sublattice
magnetization vectors.
As a result, physical consequences such as the diagram are
different, although the mathematical structure is similar.
The paper is organized as follows.
In Sec.~\ref{sec:1dmodel} we introduce the one dimensional model,
discuss its homogeneous phases and the vanishing boundary contributions.
Sec.~\ref{sec:EOM} gives the field equations
and a classification
of the different types of solutions.
In Sec.~\ref{sec:spiral_solution}, the lowest energy spiral
solution is shown to give the ground state in the spiral phase, and
the phase boundary between the homogeneous phase is determined explicitly.
The exact spiral solutions in terms of elliptic functions are
also given for $B=0$ or $A=0$ cases.
In Sec.~\ref{sec:exactdomainwalls}
we present the exact domain wall solutions for the model.
Sec.~\ref{sec:summary} contains a summary of the work
presented in the paper and a discussion of some open questions.
Appendix~\ref{sec:der_spiral} is devoted to the evaluation of the
integrals needed to obtain the second derivative of the average
energy density of the spiral ground state near the phase boundary.
Finally Appendix~\ref{sec:exact_sol_AorB} gives some details of
exact spiral solutions for $B=0$ or $A=0$.
\section{One dimensional chiral magnet}\label{sec:1dmodel}
\subsection{Model}\label{sec:field-eq}
Throughout this paper we study chiral magnets in one
spatial dimension, in particular their phases and exact spiral
and domain wall solutions.
The energy density of the particular model we consider
is given in terms of
a real three component magnetization vector field
$\vec{n}:\mathbb{R}^{1}\to S^{2}$ with the constraint $\vec{n}^{2}=1$
\begin{equation}
\mathcal{E}=\frac{1}{2}\left(\frac{ d\vec{n}}{dx}\right)^{2}
+\kappa\vec{n}\cdot \nabla_{-\alpha}\times \vec{n} +U(n_{3})
+\mathcal{E}_{\rm boundary},
\label{1D energy density}
\end{equation}
where the potential is
\begin{equation}
U(n_{3})=-Bn_{3}+An_{3}^{2}.
\label{eq:potential term}
\end{equation}
We leave to the next subsection a discussion of the
boundary term $\mathcal{E}_{\rm boundary}$, which does not
affect the field equations but is needed to make the variational
problem well-defined.
The rotated gradient $\nabla_{-\alpha}$ is
defined in terms of $\nabla=(\frac{d}{dx}, 0, 0)^T$ and a rotation
matrix $R(-\alpha)$ by an angle $-\alpha\in (-2\pi, 0]$ about the 3-axis
\begin{equation}
\nabla_{-\alpha}=R(-\alpha)\nabla=
\left(
\begin{array}{c}
\cos\alpha
\\
-\sin\alpha
\\
0
\end{array}
\right) \frac{d}{dx}.
\label{eq:DMderivative}
\end{equation}
Using this the Dzyaloshinskii-Moriya (DM) interaction term
\cite{Dzyaloshinskii,Moriya} can be written explicitly as
\begin{equation}
\vec{n}\cdot \nabla_{-\alpha}\times \vec{n}
=\cos\alpha w_{B}+\sin\alpha w_{N},
\end{equation}
Here $w_{B}$ and $w_{N}$ are the Bloch and Ne\'{e}l DM terms which
in one dimension become
\begin{equation}
w_{B}=n_{3}\frac{d n_{2}}{dx}-n_{2}\frac{d n_{3}}{dx},
\qquad w_{N}=n_{3}\frac{d n_{1}}{dx}-n_{1}\frac{d n_{3}}{dx},
\end{equation}
respectively.
This model has the symmetry that the $B>0$ and $B<0$ parameter
regions are related by sending $B\to -B$ and $n_{3}\to -n_{3}$.
As such we chose to work with $B\geq 0$.
As observed in \cite{BRS} working in terms of a rotation by
$\alpha$ around the $3$-axis, $R(\alpha)$, we can rewrite the DM term as
\begin{equation}
\vec{n}\cdot \nabla_{-\alpha}\times \vec{n}
=
\vec{n}^{\alpha}\cdot \nabla\times \vec{n}^{\alpha},
\end{equation}
with
\begin{equation}
\vec{n}^{\alpha}=R(\alpha)\vec{n}. \label{rotated magnetisation vector}
\end{equation}
Since the first and third term in Eq.~\eqref{1D energy density}
are invariant under $R(\alpha)$, Eq.~\eqref{1D energy density}
can be rewritten in terms of $\vec{n}^{\alpha}$ and $\nabla$
replacing $\vec{n}$ and $\nabla_{-\alpha}$.
\subsection{Boundary terms}
A total derivative term in the action is a boundary
term which does not affect the equations of motion.
However, it can in general contribute to the energy similarly
to the DM interaction term.
It has been realized in Ref.~\cite{BRS} that there are subtle
contributions from infinity (the boundary) to the energy of
the skyrmion (hedgehog) solutions in the case of two-dimensional
chiral magnets along the solvable line $2A=B$.
Subsequently it has been realized that the additional total
derivative term in the energy density is needed to make the
variational principle well-defined, in deriving the field equations
for the skyrmion \cite{RSN}.
We call this term the boundary energy functional.
In the two dimensional solvable case, the boundary energy functional gives a finite
and crucial negative energy contribution to the skyrmion energy.
This fact leads to the instability of the polarized ferromagnetic
background solution and a phase transition to other phases, such as
the skyrmion lattice phase, below the critical value of potential
parameters \cite{RSN}.
When one varies the energy functional to obtain the equation of
motion, it requires a partial integration for the DM interaction
term.
The resulting surface term can be canceled by adding a boundary
energy functional of the form \cite{RSN}
\begin{equation}
\mathcal{E}_{\rm boundary}=\kappa \epsilon_{ijk}\nabla_i
\left(n^{\alpha,{\rm bound}}_j n^{\alpha}_k\right),
\label{eq:boundary-term}
\end{equation}
where $\vec{n}^{\alpha,{\rm bound}}$ is the boundary value of the
magnetization vector.
It is important to realize that the derivative is acting also on
the boundary value $\vec{n}^{\alpha,{\rm bound}}(x)$, since
it can depend on the position along the boundary, namely (two)
isolated points in our one-dimensional model.
Let us consider the total energy in an interval $-x_{1}\le x \le x_{2}$
with the boundaries located at $x=-x_1$ and $x=x_2$.
The contribution from the boundary term is given by
\begin{equation}
\begin{split}
E_{\rm boundary}
&=\kappa \int_{-x_1}^{x_2}dx\partial_1
\left(n_3^{{\rm bound}}(x)n^{\alpha}_2(x)
-n_2^{\alpha,{\rm bound}}(x)n_3(x)\right)\\
&=0,
\end{split}
\label{eq:energy-boundary-term}
\end{equation}
since the magnetization vector $\vec{n}^\alpha(x)$ takes the
boundary value $\vec{n}^{\alpha, {\rm boundary}}(x)$ at the
boundary.
In the case of spiral solutions, we are interested in the total
energy in one period, taking $x_2+x_1$ to be the period.
In the case of domain wall solutions, we should choose $x_1, x_2\to \infty$.
Thus we find no contribution from the boundary term for
the class of solutions that we consider here.
As such we will drop the boundary term from the energy density
from now on.
\subsection{Homogeneous phases}
\begin{figure}[htbp]
\begin{center}\includegraphics[width=0.4\textwidth]{figures/potential_phase_diagram3}
\caption{The three ferromagnetic phases of the potential $U(n_{3})
=-Bn_{3}+An_{3}^{2}$.}
\label{phase fig}
\end{center}
\end{figure}
Our field equations admit not only one-dimensionally
modulated solutions but also homogeneous configurations as
solutions.
They are the ground states in parameter regions with vanishing
or small DM term.
Homogeneous solutions are given by the stationary points of
the potential, and are in common with the model in
two or higher spatial dimensions
which has the same potential as in Eq.~\eqref{eq:potential term}, see for instance Refs.~\cite{BRS, RSN}.
Restricting to homogeneous solutions and
comparing the energy density results in the familiar phase diagram shown
in Fig.~\ref{phase fig}.
There are three distinct regions:
\begin{enumerate}
\item
region I: $-2A\le B \le 2A$, with the minimum at $n_3=\frac{B}{2A}$, and
\begin{equation}
{\cal E}_{\rm canted}=U_{\rm min}=-\frac{B^2}{4A}.
\label{eq:min_canted}
\end{equation}
This is the canted ferromagnetic phase.
The minimum configuration of $\vec{n}$ is the
circle $(n_{1})^{2}+ (n_{2})^{2}=1-\frac{B^{2}}{4A^{2}}$.
\item
region II: $B\ge 0$, and $B\ge 2A$, with minimum at $n_{3}= 1$, and
\begin{equation}
{\cal E}_{+}=U_{\rm min}=A-B.
\label{eq:min_pos_pol}
\end{equation}
This is the positively polarized ferromagnetic
phase.
\item
region III: $B\le 0$, and $B\le 2A$, with the minimum of the potential
\begin{equation}
{\cal E}_{-}=U_{\rm min}=A+B,
\label{eq:min_neg_pol}
\end{equation}
at $n_{3}= -1$.
This is the negatively polarized ferromagnetic
phase.
\end{enumerate}
The phase boundary lines are located at $B=2A\ge 0$ (between I and II),
$B=0, A\le 0$ (between II and III), and $B=2A \le 0$ (between II and III).
The order of the phase transition was found previously
\cite{Hongo:2019nfr}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/u_minvsA_b15}
\caption{The energy density $\mathcal{E}(A,B)=U_{\rm min}$ of the
homogeneous ground state as a function of $A$ for a fixed $B\not=0$.
It exhibit the second order transition between polarized
($2A<|B|$) and the canted ($|B|<2A$) ferromagnetic phase.
}
\label{fig:u_minvsA}
\end{center}
\end{figure}
The phase transition between regions I and II is of second order, since the energy density and its first
derivative are continuous, while the second derivative
is discontinuous.
Similarly the phase transition between regions I and III
is also of second order.
In Fig.~\ref{fig:u_minvsA}, we illustrate the second
order phase transition between II and I ($B>0$) or III and I
($B<0$).
The phase transition between the regions II and III across the
line $B=0, A<0$ is of first order, since the
energy density is identical on both sides but its first
derivative is discontinuous
.
Moreover, there are two distinct minima at $n_3=1$ and $n_3=-1$
with the same energy.
In Fig.~\ref{fig:u_minvsB}, we illustrate the first
order phase transition between II and III (for a fixed $A<0$)
in
the left panel and the second order phase transition between
II and I, and one between I and III (for a fixed $A>0$) in the right panel.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/u_minvsB_a-05}
\includegraphics[width=0.45\textwidth]{figures/u_minvsB_a2}
\caption{The energy density $\mathcal{E}(A,B)=U_{\rm min}$
of the homogeneous ground state as a function of $B$ for fixed $A$.
The left panel is for fixed $A<0$, exhibiting the first order
phase transition between region II and III.
The right panel is for $A>0$, exhibiting the second order phase
transitions between regions II (positively polarized) and I
(canted), and I and III (negatively polarized) phases. }
\label{fig:u_minvsB}
\end{center}
\end{figure}
One can expect that homogeneous configurations tend to
have lower energy when the potential term $U$ is more important
than other terms, in particular the DM term.
It is known that spiral solutions
become the ground states for parameter regions where the
DM interaction is more important than the potential.
We will determine the precise phase boundary.
\section{The field equation}\label{sec:EOM}
\subsection{Conservation of ``mechanical energy'' }
\label{sec:conservation-law}
To find the solutions for the energy density in
Eq.~\eqref{1D energy density},
it is easiest to first solve the constraint, $(\vec{n}^\alpha)^2=1$,
using the independent angular variables $\Theta, \Phi$
\begin{equation}
\vec{n}^{\alpha}=(\sin\Theta\cos\Phi, \sin\Theta\sin\Phi, \cos\Theta)^T.
\label{eq:angle-variable}
\end{equation}
The energy density in Eq.~\eqref{1D energy density} is given in
terms of the independent fields $\Theta(x)$ and $\Phi(x)$ as
\begin{equation}
\begin{split}
\mathcal{E}&=\frac{1}{2}\left\{\left(\frac{ d\Theta}{dx}\right)^{2}
+\left(\sin\Theta\frac{d\Phi}{dx}\right)^2\right\}+\kappa\left(\sin\Phi\frac{ d\Theta}{dx}
+\frac{1}{2}\sin 2\Theta \frac{d}{dx}\sin\Phi\right)
+U(\Theta) ,
\end{split}
\label{eq:energy-density-angle}
\end{equation}
with the potential
\begin{equation}
U(\Theta)=-B\cos\Theta+A\cos^{2}\Theta.
\label{eq:potential-angle}
\end{equation}
Minimising the energy functional leads to the Euler-Lagrange or field equations for $\Theta$
and $\Phi$ being
\begin{equation}
-\frac{d^2\Theta}{dx^2}+\frac{1}{2}\sin2\Theta\left(\frac{d\Phi}{dx}\right)^2
-2\kappa\sin^2\Theta\frac{d\sin\Phi}{dx
+\frac{\partial U}{\partial\Theta}=0,
\label{eq:theta-EOM}
\end{equation}
\begin{equation}
-\frac{d}{dx}\left(\sin^2\Theta\frac{d\Phi}{dx}\right)
+2\kappa\sin^2\Theta\cos\Phi\frac{d\Theta}{dx}=0.
\label{eq:phi-EOM}
\end{equation}
These equations are written down for the case $B=0$ and $\alpha=\frac{3\pi}{2}$ in~\cite{Chovan}, where they are numerically studied, including beyond the constant $\Phi$ approximation that we make here\footnote{As a comment on notation in~\cite{Chovan} they use $\lambda$ for the DM parameter $\kappa$ and their $\gamma^{2}=2A$ is equivalent to our anisotropy parameter.}. It is worth reiterating here that the flat spirals we consider here have $\Phi(x)=$ constant, while a non-flat spiral would not have this restriction.
As a first step, we solve Eq.~\eqref{eq:phi-EOM} for $\Phi(x)$
by assuming a constant value. We then find
\begin{equation}
\Phi(x)=\left(2n\pm \frac{1}{2}\right)\pi, \quad n\in\mathbb{Z}.
\label{eq:phi-solution}
\end{equation}
Then the DM interaction term $\kappa$ disappears from the
field equation
for $\Theta(x)$ and Eq.~\eqref{eq:theta-EOM} becomes
\begin{equation}
\frac{d^2\Theta}{dx^2}
+\frac{\partial U}{\partial\Theta}.
\label{eq:theta-EOM2}
\end{equation}
If we regard $x$ as ``time'', Eq.~\eqref{eq:theta-EOM2} is
the equation of motion of a particle with unit mass moving in the
periodic potential $-U(\Theta)$.
Since the potential is periodic, the particle is constrained
to move on a circle, $S^1$. This classical mechanics
analogy is useful for classifying solutions and finding
their properties.
Multiplication of \eqref{eq:theta-EOM2} by $\frac{d\Theta}{dx}$ gives
a conservation law corresponding to translational invariance
\begin{equation}
\frac{1}{2}\left(\frac{d\Theta}{dx}\right)^{2}-U(\Theta)
=C_{0}, \label{eq: first integral}
\end{equation}
where the constant $C_0$ is a conserved quantity (an integral of motion)
corresponding to the ``mechanical energy'' (sum of kinetic and ``potential energy'' $-U(\Theta)$) in classical mechanics.
\subsection{Classification of solutions }\label{sec:classification}
We
denote the value of the maximum of the potential as $U_{\rm max}$.
\begin{figure}[htbp]
\begin{center}\includegraphics[width=0.6\textwidth]{figures/mechanical_energy_mag_dominated_3}
\caption{A plot of the potential $-U_{\rm min}(\Theta)$ for
the ``mechanical analog" when $B>2A$. The
upper solid horizontal line (red) is $-U_{\rm min}=-A+B$
and the lower horizontal dashed line (blue) is $-U_{\rm max}$.
Motions (from right to left) with the ``mechanical energy"
$C_0>-U_{\rm min}$ give spiral
solutions. Motions with $C_0=-U_{\rm min}$ connect two neighboring
$\Theta_{\rm min}$ and give domain wall solutions.
Motions with $C_0<-U_{\rm min}$ give oscillating solutions.
}
\label{fig:mech_analogy_region2}
\end{center}
\end{figure}
Since a symmetry $B\to -B, n_3\to-n_3$ allows us to obtain
solutions for $B\le 0$ from those of $B\ge0$, we concentrate
here on regions I and II.
The conservation law implies that solutions exist if and only if
$C_0\ge -U_{\rm max}$, because of the positivity $(\frac{d\Theta}{dx})^2\ge 0$.
We depict a typical figure
of $-U(\Theta)$ for the region
II in Fig
\ref{fig:mech_analogy_region2}.
In the captions we discuss the qualitatively different
types of solutions in terms of the value of $C_0$.
Using the constant solution \eqref{eq:phi-solution} for $\Phi$,
the energy density becomes
\begin{equation}
\mathcal{E}_{\rm sol}=\kappa\frac{d\Theta}{dx}
+\frac{1}{2}\left(\frac{d\Theta}{dx}\right)^{2}+U(\Theta).
\label{eq:energy-density-Phi-solution}
\end{equation}
The energy of the solution depends on the DM interaction term,
even though the field equation \eqref{eq:theta-EOM2} does not.
However, the DM term does not contribute to the
energy for homogeneous oscillating solutions
after averaging over one period.
The inequality $\frac{1}{2}\left(\frac{d\Theta}{dx}\right)^{2}
+U(\Theta)\ge U_{\rm min}$ implies that the energy density of
the homogeneous ground state at $\Theta_{\rm min}$ is lower than
the average energy densities of the oscillating solutions and other
homogeneous solutions.
We will study spiral solutions and domain wall solutions
in order to find the ground state of a one dimensional chiral
magnet in the following.
For the spiral and domain wall solutions with $C_0>-U_{\rm min}$,
Eq.~\eqref{eq: first integral} implies that $\frac{d\Theta}{dx}$
never vanishes. Hence its sign never changes,
and the solutions are monotonic.
Since the energy density \eqref{eq:energy-density-Phi-solution}
favors the negative sign for $\frac{d\Theta}{dx}$, we should choose
the negative chirality solution with $\frac{d\Theta}{dx}<0$
in order to obtain the ground state of the chiral magnet.
Therefore we need to solve the following first order equation
\begin{equation}
\frac{d\Theta}{dx}=- \sqrt{2(C_0+U(\Theta))}.
\label{eq: theta-equation}
\end{equation}
By making use of the conservation of ``mechanical energy''
from Eq.~\eqref{eq: first integral}, we can rewrite the energy density
as
\begin{equation}
\mathcal{E}_{\rm sol}=\kappa\frac{d\Theta}{dx}
+\left(\frac{d\Theta}{dx}\right)^{2}-C_0
=\kappa\frac{d\Theta}{dx}+2U(\Theta)+C_0.
\label{eq:energy-density-solution}
\end{equation}
Explicit solutions are found by integrating the first order
equation \eqref{eq: theta-equation}.
We observe that the solutions are given in terms of
elliptic functions.
In the next subsection, however, we derive general properties of
spiral solutions and will calculate the phase boundary
between the spiral and
homogeneous phases without needing to find explicit solutions.
\section{Properties of spiral solutions}
\label{sec:spiral_solution}
\subsection{Spiral solutions without the potential}
\label{sec:spiral_solution_no_pot}
As a warm up for studying more general spirals let us first consider
the case without a potential, $A=B=0, U(\Theta)=0$. This has been treated in the literature before \cite{BH} and we discuss it here to set up our conventions for future sections.
In this case, we need $C_0\ge 0$ and obtain $\frac{d\Theta}{dx}=-\sqrt{2C_0}$
from \eqref{eq: theta-equation}. The general solutions
contain one additional parameter, $x_0$ as a Nambu-Goldstone
(NG) mode for the spontaneously broken translational symmetry,
\begin{equation}
\Theta(x)= -\sqrt{2C_0} (x-x_0).
\label{eq:plane-wave}
\end{equation}
The magnetisation vector is
$n_3+in_2^\alpha=e^{-i\sqrt{2C_0}(x-x_0)}$ rotating along $x$ with the
momentum $ -\sqrt{2C_0}<0$, namely a negative chirality plane wave as
a spiral solution.
The period $L$ of this spiral solution is
$L=2\pi/\sqrt{2C_0}$.
Although all the configurations with different values of $C_0$ are solutions of the equations of motion,
they have different energy densities
\begin{equation}
\mathcal{E}=-\kappa\sqrt{2C_0}+C_0.
\label{eq:energy-density-no-pot}
\end{equation}
As illustrated in the left panel of Fig.~\ref{fig:C0-E_AB=0},
the energy $\mathcal{E}$ of spiral solutions as a function
of $C_0$ has a minimum
at $C_{0,{\rm min}}=\frac{\kappa^{2}}{2}$
giving the lowest energy spiral solution as
\begin{equation}
\mathcal{E}_{\rm min}=-C_{0,{\rm min}}=-\frac{\kappa^2}{2}.
\end{equation}
This energy density is lower than that of the homogeneous
solution ($\mathcal{E}_{\rm homogeneous}=0$), thus
$A=B=0$ is in the spiral phase.
For later use, we define the excess energy $f(C_0; A=B=0)$ in one
period above the constant energy density $-C_0$ as
\begin{equation}
f(C_0; A=B=0) =\int_0^{L} dx (\mathcal{E}+C_0)
=-2\pi\kappa+2\pi\sqrt{2C_0}.
\label{eq:f-no-pot}
\end{equation}
As illustrated in the right panel of
Fig.~\ref{fig:C0-E_AB=0}, $f(C_0; A=B=0)$
increases
monotonically and vanishes at $C_0=C_{0, {\rm min}}=\frac{\kappa^{2}}{2}$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/average_energy_density_2d_2}
\includegraphics[width=0.45\textwidth]{figures/energy_2d_v2}
\caption{Left: Energy density $\mathcal{E}$ of
spiral solutions as a function of $C_{0}$ for $A=B=0$.
The lowest energy spiral solution occurs at $C_0=C_{0, {\rm min}}=\kappa^2/2$,
giving the energy density $\mathcal{E}_{\rm min}=-\kappa^2/2$
for the spiral ground state.
Right: The excess energy $f(C_0; A=B=0)$ in one period as a function
of $C_{0}$ for $A=B=0$. It increases monotonically and
vanishes at $C_0=C_{0, {\rm min}}=\kappa^2/2$.}
\label{fig:C0-E_AB=0}
\end{center}
\end{figure}
\subsection{Lowest energy flat spiral solution }
\label{sec:lowest_spiral_solution}
Although spiral solutions (including domain wall solutions)
are monotonic functions of $\Theta$,
they are periodic as a function of magnetization vector $\vec{n}$.
We define the period $L$ of spiral solutions as the distance
corresponding to translation by $2\pi$ in $\Theta$. Since
Eq.~\eqref{eq: theta-equation} gives a one-to-one correspondence
between $x$ and $\Theta$, once the NG mode $x_0$ is fixed
by $\Theta(x_0)=0$ for instance.
We can now change variable from $x$ to $\Theta$ by using
Eq.~\eqref{eq: theta-equation}. The period $L$ of a spiral
solution is given by
\begin{equation}
L=\int_{0}^{L}dx=\int_0^{2\pi}d\Theta\left[-\frac{d\Theta}{dx}\right]^{-1}
=\int_{0}^{2\pi} \frac{d\Theta}{\sqrt{2(C_0+U(\Theta))}}.
\label{eq:period}
\end{equation}
The average energy density $\langle\mathcal{E}\rangle$ is given by
\begin{equation}
\langle\mathcal{E}\rangle=\frac{1}{L}\int_{0}^{L}dx\,\mathcal{E}
=\frac{f(C_0;A,B)}{L}-C_0,
\label{eq:energy-solution}
\end{equation}
where
\begin{equation}
\begin{split}
f(C_0;A,B)&=\int_{0}^{L}dx\left[\kappa\frac{d\Theta}{dx}+
\left(\frac{d\Theta}{dx}\right)^2\right]\\
&=
-2\pi\kappa +\int_{0}^{2\pi} d\Theta
\sqrt{2(C_0+U(\Theta))},
\end{split}
\label{eq:energy-solution2}
\end{equation}
is the energy in one period due to the excess energy
density above the constant value $-C_0$. As in Eq.~\eqref{eq:f-no-pot} we call $f(C_0;A,B)$
the excess energy.
Only the first term of the excess energy comes
from the DM interaction, this gives a negative energy contribution only when
the solution has a non-zero winding ($-2\pi$) after one
period of translation.
This feature makes it a possibility for spiral solutions to have
lower average energy density than the homogeneous ground state.
For fixed values of parameters $\kappa, A, B$, our spiral solutions
have two parameters (moduli) $C_0$ and $x_0$. The average energy
density depends only on $C_0$ and is
independent of $x_0$. Therefore we need to look for the lowest energy solution among spiral
solutions, by minimizing the average energy density as a function
of $C_0$. We find the identities
\begin{equation}
\frac{\partial f(C_0;A,B)}{\partial C_0}=
\int_{0}^{2\pi}
\frac{d\Theta}{\sqrt{2(C_0+U(\Theta))}}
=L,
\label{eq:derivative-f}
\end{equation}
\begin{equation}
\frac{\partial L}{\partial C_0}=
-\int_{0}^{2\pi}
\frac{d\Theta}{(2(C_0+U(\Theta)))^{3/2}},
\label{eq:derivative-L}
\end{equation}
some of which are also given in \cite{Chovan}.
Thus we obtain
\begin{equation}
\frac{\partial \langle\mathcal{E}\rangle}{\partial C_0}
=\frac{-1}{L^2}\frac{\partial L}{\partial C_0}f(C_0;A,B).
\label{eq:derivative-average-energy}
\end{equation}
Eq.~\eqref{eq:derivative-f} shows that $f(C_0;A,B)$ is monotonically
increasing.
Because of \eqref{eq:derivative-L}, the average energy density
reaches a minimum when
\begin{equation}
f(C_0;A,B)=-2\pi\kappa+\int_0^{2\pi}d\Theta\sqrt{2(C_0+U(\Theta))}=0,
\label{eq:minimum-energy-cond}
\end{equation}
is satisfied, provided this value of $C_{0,{\rm min}}$ is in the allowed
region of spiral solutions $C_0 \ge -U_{\rm min}$.
This minimum energy condition \eqref{eq:minimum-energy-cond}
gives $C_{0,{\rm min}}$ as a function of $A$ and $B$.
Eq.~\eqref{eq:energy-solution} implies that the average energy
density of the lowest energy spiral solution is given by this
$C_{0,{\rm min}}$ as
\begin{equation}
\langle\mathcal{E}\rangle_{\rm min}
=-C_{0,{\rm min}}.
\label{eq:min-energy-spiral}
\end{equation}
Hence the lowest energy spiral solution always has
lower energy than the homogeneous ground state with the energy
density $U_{\min}$
\begin{equation}
\langle\mathcal{E}\rangle_{\rm min}
=-C_{0,{\rm min}}\le U_{\rm min}.
\label{eq:min-energy-spiral2}
\end{equation}
We conclude that a chiral magnet is in the spiral phase,
once the minimum energy condition is satisfied in the allowed
region for spiral solutions $C_0> -U_{\rm min}$.
\subsection{Boundary between spiral and homogeneous phases }
\label{sec:phase_boundary}
As given in Eq.~\eqref{eq:f-no-pot}, the function $f$
at $A=B=0$ becomes $f(C_0; A=B=0)=-2\pi\kappa+2\pi\sqrt{2C_0}$.
This behaves asymptotically as
$f(C_0;A=B=0) \sim 2\pi \sqrt{2C_0} > 0$ for $C_0\to \infty$
whereas
f(C_0=-U_{\rm min}=0;A=B=0)=-2\pi\kappa^2 <0$.
Hence the excess energy $f(C_0;A=B=0)$ has a zero in the
region $C_0>-U_{\rm min}$, ensuring the existence of the lowest
energy spiral solution for $A=B=0$.
Let us now study the case of nonvanishing values of $A,B$, such that $2A\leq B$.
We observe that the excess energy $f(C_0;A,B)$ at
$C_0\gg A, B, \kappa^2$ becomes asymptotically
\begin{equation}
f(C_0;A,B) \sim 2\pi \sqrt{2C_0} > 0.
\label{eq:asymptotic_f}
\end{equation}
Since the excess energy $f(C_0;,B)$ is monotonically increasing,
the zero of the excess energy $f(C_0;A,B)=0$ occurs in the
allowed region if and only if $f(C_0=-U_{\rm min};A,B)<0$.
When $C_0$ approaches the lower bound at $-U_{\rm min}$,
we find
\begin{equation}
L \to \infty,
\label{eq:infinite_period}
\end{equation}
since the integrand diverges at $\Theta = \Theta_{\rm min}$,
as $\sqrt{C_0+U(\Theta)} \to \sqrt{\frac{U''(\Theta_{\rm min})}{2}}
|\Theta-\Theta_{\rm min}|$ with $U''=\frac{d^{2}U}{d\Theta^{2}}$.
On the other hand, the excess energy remains finite as
$C_0\to -U_{\rm min}$.
Therefore the average energy density \eqref{eq:energy-solution}
becomes
\begin{equation}
\left.\langle\mathcal{E}\rangle\right|_{C_0=-U_{\rm min}}
=U_{\rm min},
\label{eq:energy_density_Umin}
\end{equation}
which is the same as that of the homogeneous solution.
We denote the excess energy at this $C_0=-U_{\rm min}$ as
\begin{align}
E_{\rm period}(A,B)&=f(C_0=-U_{\rm min};A,B)\nonumber\\
&=-2\pi\kappa+\int_0^{2\pi}d\Theta\sqrt{2(U(\Theta)-U_{\rm min})}.
\label{eq:excess-energy-min}
\end{align}
Summarizing the above, we find that the chiral magnet is in the
spiral phase if\\ $E_{\rm period}(A,B)<0$, and
in the homogeneous phase if $E_{\rm period}(A,B)>0$.
The phase boundary between spiral phase and the homogeneous phases
is given by
\begin{equation}
0=E_{\rm period}(A,B)
=-2\pi\kappa+\int_0^{2\pi}d\Theta\sqrt{2(U(\Theta)-U_{\rm min})}.
\label{eq:phase-boundary-line}
\end{equation}
Since the period is infinite, the spiral solution in this limit
becomes a domain wall solution connecting two adjacent values
of $\Theta_{\min}$. We will describe more explicitly the
domain wall solutions as a limit of spiral solutions in a later
section.
Since $-U_{\rm min}$ is the energy density of the homogeneous solution,
the excess energy above this energy density integrated over one period
$E_{\rm period}(A,B)$ is precisely the domain wall energy as a
finite energy soliton solution on the homogeneous background.
Our results show that the domain wall solution as a single
soliton above the homogeneous background has positive energy
and is an ordinary excitation in the homogeneous phase.
It becomes zero energy at the boundary between the homogeneous and
spiral phases.
It has a negative energy in the spiral phase region, and
signals the instability of the homogeneous background solution
and its decay to the lowest energy spiral solution as the true
ground state.
One can intuitively visualize the deformation of the configuration
as a condensation of these negative energy solitons settling down
to the spiral ground state.
\subsection{Explicit formula for phase boundaries}\label{sec:phaseboundaries}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/AB_phase_plot_5}
\caption{The phase diagram in the
$\frac{A}{\kappa^{2}}, \frac{B}{\kappa^{2}}$ plane.
The spiral phase is below the blue curve.
The(positively) polarized ferromagnetic phase is above
the blue curve. The two ferromagnetic phases are separated by
the green line $B=2A$.
Below the tricritical point $B=2A=\kappa^2$, the dashed line $B=2A$
does not correspond to a phase boundary. The dashed red line shows where flat spirals in the anisotropy dominated case have zero energy. This is not a phase boundary as the flat spiral is not the ground state.
}
\label{fig:Phase boundaries}
\end{center}
\end{figure}
When $2A\le B$, the homogeneous phase in this region is the
polarized ferromagnetic
phase $n_3=+1$.
The minimum of the potential is $U_{\rm min}=A-B$ which occurs at
$\Theta_{\rm min}=2{\mathbb Z}\pi$.
We find that the boundary in Eq.~\eqref{eq:phase-boundary-line}
between
the polarized ferromagnetic phase and the spiral phase is given by
\begin{align}
2\pi\kappa
&=
\int_0^{2\pi}d\Theta
\sqrt{2(-B\cos\Theta+A\cos^2\Theta -A+B)}
\nonumber \\
&=
4\sqrt{B-2A}
+\frac{4B}{\sqrt{2A}}\arcsin\sqrt{\frac{2A}{B}}.
\label{eq:phase-boundary-case2}
\end{align}
This boundary is depicted as the (blue) curve to the left of
$B=2A$ line in Fig.~\ref{fig:Phase boundaries}.
This expression is manifestly real for $0\le A$.
This formula is also valid for $A\le 0$ by an analytic continuation
of $A>0\to A<0$.
We find a convenient (manifestly real) expression in the region $A\le 0$
to be
\begin{align}
2\pi\kappa
&=
4\sqrt{B-2A}
+\frac{4B}{\sqrt{-2A}}\arcsinh\sqrt{\frac{-2A}{B}}
\label{eq:phase-boundary-case2-2}
\\
&=
4\sqrt{B-2A}
+\frac{2B}{\sqrt{-2A}}\log\left(\frac{\sqrt{B-2A}+\sqrt{-2A}}
{\sqrt{B-2A}-\sqrt{-2A}}\right),
\nonumber
\end{align}
where the last expression agrees with the one derived in the
case of $A\le 0$ previously \cite{BH}, apart from the difference
in conventions.
This boundary curve is also depicted as the (blue) curve
to the left of the
$B=2A$ line in Fig.~\ref{fig:Phase boundaries}.
For completeness, in order to see the region of negative $B$,
we note that the phase diagram is symmetric under $B \to -B$.
Although partial results for particular parameter values or regions
have been obtained for the phase boundary between homogeneous
ferromagnetic phases and spiral phase, our results give the
complete boundary curve explicitly for the first time.
\subsection{Order of phase transition between spiral and homogeneous phases}
In this subsection, we will demonstrate that the phase transition
between the spiral and homogeneous phases is of second order.
The phase transition is of the $n$-th order if the $l$-th
order derivatives of the free energy at the phase boundary
are continuous for $l<n$ and discontinuous for $l=n$.
At zero temperature, the free energy density is given by the
average energy density of the ground state.
In the homogeneous phase, the energy density is given by the
minimum $U_{\rm min}$ of the potential.
The average energy density of spiral phase
$\langle\mathcal{E}\rangle_{\rm spiral}$ is given by that
of the lowest energy spiral solution in Eq.~\eqref{eq:min-energy-spiral}
\begin{equation}
\langle\mathcal{E}\rangle_{\rm spiral}
=-C_{0,{\rm min}}(A, B),
\label{eq:spiral_energy}
\end{equation}
where the mechanical energy $C_{0,{\rm min}}$ of the lowest
energy spiral solution is determined as a function of the potential
parameters $A, B$ by the minimum energy condition in
Eq.~\eqref{eq:minimum-energy-cond} as
\begin{align}
2\pi\kappa&=\int_0^{2\pi}d\Theta\sqrt{2(U(\Theta)+C_{0,{\rm min}})}\nonumber \\
&=\int_0^{2\pi}d\Theta\sqrt{2(-B\cos\Theta+A\cos^2\Theta+C_{0,{\rm min}})}
.
\label{eq:boundary-line}
\end{align}
As noted in Eq.~\eqref{eq:energy_density_Umin}, the energy
density of the spiral phase tends continuously to that
of the homogeneous phase at the phase boundary
\begin{equation}
\langle\mathcal{E}\rangle_{\rm spiral}\to
U_{\rm min}(A, B)=\langle\mathcal{E}\rangle_{\rm homogeneous} .
\end{equation}
Since the energy density along the boundary is common for homogeneous
and spiral phases, all of their derivatives tangential to the
phase boundary are continuous between the
homogeneous and spiral phases.
Therefore, we need to consider only the derivative perpendicular
to the boundary.
Let us parametrize the difference between the energy density of the spiral and
homogeneous phases in terms of the deviation $\Delta C\ge 0$ defined as
\begin{equation}
C_{0,{\rm min}}(A, B)=-U_{\rm min}(A, B)+\Delta C(A, B),
\label{eq:DeltaC}
\end{equation}
\begin{equation}
\langle\mathcal{E}\rangle_{\rm spiral}
-
\langle\mathcal{E}\rangle_{\rm homogeneous}
=-\Delta C(A, B).
\label{eq:diff_spiral_homo}
\end{equation}
The limiting procedure of the parameters $A, B$ approaching the
phase boundary is equivalent to $\Delta C(A,B)\to 0$.
We can obtain the first derivative of $\Delta C$
in terms of $A, B$ by differentiating the
minimum energy condition in Eq.~\eqref{eq:boundary-line}
\begin{equation}
0=\int_0^{2\pi}d\Theta
\frac{-\frac{\partial U_{\rm min}}{\partial A}
+\frac{\partial\Delta C}{\partial A}
+\cos^2\Theta}{\sqrt{2(U(\Theta)-U_{\rm min}+\Delta C)}},
\label{eq:1st-derivative-A}
\end{equation}
\begin{equation}
0=\int_0^{2\pi}d\Theta
\frac{-\frac{\partial U_{\rm min}}{\partial B}
+\frac{\partial\Delta C}{\partial B}
-\cos\Theta}{\sqrt{2(U(\Theta)-U_{\rm min}+\Delta C)}}.
\label{eq:1st-derivative-B}
\end{equation}
We can rewrite these relations as
\begin{equation}
-\frac{\partial\Delta C}{\partial A}
=\frac{L_0^{01}}{L}, \quad
-\frac{\partial\Delta C}{\partial B}
=\frac{L_0^{10}}{L},
\label{eq:1stBder_Ble2A}
\end{equation}
where the period $L$ and weighted integrals $L_n^{kl}$ are defined as
\begin{equation}
L=L_0^{00}=\int_0^{2\pi}
\frac{d\Theta}{\sqrt{2(U(\Theta)-U_{\rm min}+\Delta C)}} ,
\label{eq:period-integral-rep}
\end{equation}
and
\begin{equation}
L_n^{kl}=\int_0^{2\pi}d\Theta
\frac{\left(-\frac{\partial U_{\rm min}}{\partial B}-\cos\Theta\right)^k
\left(\cos^2\Theta-\frac{\partial U_{\rm min}}{\partial A}\right)^l}
{[2(U(\Theta)-U_{\rm min}+\Delta C)]^{n+1/2}}.
\label{eq:weighted-integral}
\end{equation}
As is derived in Appendix \ref{sec:der_spiral},
$L\to \infty$ in the limit of $\Delta C\to 0$, whereas
$L_0^{10}, L_0^{01}$ are finite.
Therefore, we obtain at the phase boundary
\begin{equation}
\frac{\partial \Delta C
}{\partial A}\to 0, \quad
\frac{\partial \Delta C
}{\partial B}\to 0,
\label{eq:1st_der_continuous}
\end{equation}
implying that the first derivative of energy density
is continuous at the phase boundary.
Similarly we can obtain the second derivative of $\Delta C$ by
differentiating \eqref{eq:1st-derivative-B} and
\eqref{eq:1st-derivative-A} in terms of $A, B$ again.
We find an exact expression for
$\frac{\partial^2\Delta C}{\partial A_i \partial A_j}$ with $A_1=A, A_2=B$
using the integral in Eq.~\eqref{eq:weighted-integral} .
As described in Appendix \ref{sec:der_spiral}, we find that the
second derivative diverges at the boundary
\begin{equation}
-\left(
\begin{array}{cc}
\frac{\partial^2 \Delta C}{\partial A^2}
& \frac{\partial^2 \Delta C}{\partial A \partial B} \\
\frac{\partial^2 \Delta C}{\partial B \partial A}
& \frac{\partial^2 \Delta C}{\partial B^2}
\end{array}
\right)
\sim
- \frac{L_1^{00}}{L^3}
\left[
\left(
\begin{array}{c}
L_0^{01} \\
L_0^{10}
\end{array}
\right)
\left(
\begin{array}{cc}
L_0^{01} & L_0^{10}
\end{array}
\right)
\right]_{\rm boundary}
,
\label{eq:C-der-bound}
\end{equation}
with the asymptotic behavior as $\Delta C\to 0$
\begin{equation}
\frac{L_1^{00}}{L^3}\sim \frac{1}{\Delta C\left(\log\frac{1}{\Delta C}\right)^{3}}
\to +\infty.
\end{equation}
Therefore the second derivative of the difference of energy density
$\langle\mathcal{E}\rangle_{\rm spiral}-
\langle\mathcal{E}\rangle_{\rm homogeneous}=-\Delta C$
is not continuous.
Hence we conclude that the phase transition between homogeneous
phases and the spiral phase is of second order.
The rank of the coefficient matrix is unity.
This corresponds to the fact that the derivative is nonzero only along
the normal to the phase boundary.
The boundary is defined by Eq.~\eqref{eq:phase-boundary-line}
corresponding to $\Delta C=0$.
By varying this condition we find that along the boundary
\begin{equation}
0=\frac{dA}{dB}\left[L_1^{01}\right]_{\rm boundary}
+\left[L_0^{10}\right]_{\rm boundary}.
\label{eq:derivative_boundary}
\end{equation}
The relation shows that the second derivative vanishes
along the tangential direction to the boundary.
\subsection{Exact spiral solutions}
\label{sec:exact_spiral_solution}
As we have seen in the previous sections, there is a spiral phase
in the regions below the boundary line defined in
Eq.~\eqref{eq:phase-boundary-line}.
In order to illustrate our results more explicitly, we here present
exact spiral solutions for the simple typical cases.
The solutions are given in terms of elliptic functions and the technical details are relegated to Appendix~\ref{app:positiveA_spiral}. Here we just present the period and energy of the solutions. Throughout this section $K(k)$ and $E(k)$ are the complete elliptic integral of
the first and the second kind defined as
\begin{align}
K(k)&=\int_{0}^{\frac{\pi}{2}}\frac{d\varphi}{\sqrt{1-k^{2}\sin^{2}\varphi}},\\
E(k)&=\int_{0}^{\frac{\pi}{2}}d\varphi\sqrt{1-k^{2}\sin^{2}\varphi},
\end{align}
respectively.
\subsubsection{Spirals without external magnetic fields (\texorpdfstring{$B=0$}{B=0})}
In the case of nonzero anisotropy ($A\not=0$) without
the Zeeman term ($B=0$), the first order equation for $\Theta(x)$
in Eq.~\eqref{eq: theta-equation} becomes
\begin{equation}
\frac{d\Theta}{dx}=- \sqrt{2(C_0+A\cos^2\Theta)}.
\label{eq: theta-equation_B=0+}
\end{equation}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]
{figures/minimum_average_energy_anisotropy_dominated_2}
\includegraphics[width=0.45\textwidth]
{figures/derivative_of_energy_by_anisotropy_2}
\caption{Case of $B=0$. Red solid curves and blue solid curves correspond to
spiral phase and homogeneous ferromagnetic phase, respectively.
Left: Average energy density
$\langle \mathcal{E}\rangle_{min}=-C_{0}(A,0)$ as a function
of $\frac{A}{\kappa^{2}}$, determined by Eqs.~\eqref{eq:elliptic_k_B=0-},
\eqref{eq:A-k-relation_B=0-}, and
\eqref{eq:lowest_energy_spiral_B=0-} for $A<0$.
Right: Derivative of average energy density as a function of
$\frac{A}{\kappa^{2}}$.
The second derivative diverges at the phase boundary
$A=- \frac{(\pi\kappa)^{2}}{8}$ showing the second order phase transition.
}
\label{fig:spiral_energy_density_B=0}
\end{center}
\end{figure}
\paragraph{$A<0$ case:}
The elliptic modulus $k$ and the period $L$ in this case are given by
\begin{equation}
k=\sqrt{\frac{-A}{C_0}}, \quad
\label{eq:elliptic_k_B=0-}
L=\frac{4}{\sqrt{2C_0}} K(k) .
\end{equation}
The average energy density of the spiral solution
$\langle \mathcal{E}\rangle =\frac{f(C_0;A>0,B=0)}{L}-C_0$
is minimized by the condition of vanishing excess energy
\begin{equation}
f(C_0;A<0,B=0)=-2\pi\kappa
+4\sqrt{2C_0}E(k)=0 ,
\label{eq: excess_energy_B=0-}
\end{equation}
which determines $C_0(A)$ as a function of $A$ through
\begin{equation}
A=-\frac{\pi^{2}\kappa^{2}}{8}\left(\frac{k}{E(k)}\right)^2.
\label{eq:A-k-relation_B=0-}
\end{equation}
and \eqref{eq:elliptic_k_B=0-}.
The minimum energy density
\begin{equation}
\langle \mathcal{E}\rangl
=-C_0(A).
\label{eq:lowest_energy_spiral_B=0-}
\end{equation}
The critical point occurs at $A=-\frac{\pi^2}{8}$.
At this point, the energy density of the lowest energy spiral solution
becomes equal to that of the ferromagnetic state.
The average energy density of the lowest energy spiral solutions
and its derivatives are plotted in Fig.~\ref{fig:spiral_energy_density_B=0}
exhibiting the second order phase transitions at
$A=\pm \frac{\pi^{2}\kappa^{2}}{8}$.
\subsubsection{Spiral solutions without anisotropy (\texorpdfstring{$A=0$}{A=0})}
The solution for $A=0$ has been studied previously~\cite{KO2015}.
In this case the first order equation for $\Theta(x)$ in Eq.~\eqref{eq: theta-equation}
becomes
\begin{equation}
\frac{d\Theta}{dx}=- \sqrt{2(C_0-B\cos\Theta)}.
\label{eq: theta-equation_A=0}
\end{equation}
The solutions are elliptic functions with the elliptic modulus $k$
and the period $L$
\begin{equation}
k=\sqrt{\frac{2B}{C_0+B}}, \quad
\label{eq:elliptic_k_A=0}
L=2\sqrt{\frac{2}{C_0+B}} K(k) .
\end{equation}
The average energy density of the spiral solution
$\langle \mathcal{E}\rangle =\frac{f(C_0;A>0,B=0)}{L}-C_0$
is minimized by the condition of vanishing excess energy
\begin{equation}
f(C_0;A=0,B)=-2\pi\kappa
+4\sqrt{2(C_0+B)}E(k) ,
\label{eq: excess_energy_A=0}
\end{equation}
which determines $C_0(B)$ as a function of $B$ through
\begin{equation}
B=\frac{\pi^{2}\kappa^{2}}{16}\left(\frac{k}{E(k)}\right)^2,
\label{eq:B-k-relation_A=0}
\end{equation}
and \eqref{eq:elliptic_k_A=0}.
The minimum energy density is given by
\begin{equation}
\langle \mathcal{E}\rangle =-C_0(B),
\label{eq:lowest_energy_spiral_A=0}
\end{equation}
where $C_0(B)$ is defined by \eqref{eq: excess_energy_A=0}.
This energy density $-C_0(B)\le 0$ is lower than the energy density of
the positively polarized ferromagnetic state with the
energy density $-B$.
As $C_0$ approaches the lowest allowed value $-U_{\rm min}=B$,
the period $L$ becomes infinite, and the
lowest energy spiral solution has the same average energy
density as the ferromagnetic state.
This critical point $B=\frac{\pi^{2}}{16}\kappa^{2}$ occurs
at the $C_0\to 0$ limit of the minimum energy condition
\eqref{eq: excess_energy_A=0}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]
{figures/min_average_energy_density_magfield-dominated}
\includegraphics[width=0.45\textwidth]
{figures/derivative_of_energy_by_magfield}
\caption{Case of $A=0$. Red solid curves and blue solid curves correspond to
spiral phase and homogeneous ferromagnetic phase, respectively.
Left: Average energy density
$\langle \mathcal{E}\rangle_{min}=-C_{0}(A=0,B)$ as a function
of $\frac{B}{\kappa^{2}}$, determined by Eqs.~\eqref{eq:elliptic_k_A=0},
\eqref{eq:B-k-relation_A=0}, and \eqref{eq:lowest_energy_spiral_A=0}.
Right: Derivative of average energy density as a function of
$\frac{B}{\kappa^{2}}$.
The second derivative diverges at the phase boundary
$B= \frac{(\pi\kappa)^{2}}{16}$ showing the second order phase transition.
}
\label{fig:min average energy density}
\end{center}
\end{figure}
As an example we draw the energy density, its first and second derivatives
as a function of $A$ in the case of $B=0$ in
Fig.~\ref{fig:spiral_energy_density_B=0}.
Similar plots of the energy density, and its first and second
derivatives are given as a function of $B$ for the case of
$A=0$ in Fig.~\ref{fig:min average energy density}.
\subsection{Domain walls as infinite period limit of spirals}
For generic values of parameters $A, B$, one can still integrate
\eqref{eq: theta-equation}, since it is an elliptic integral.
We can express the one-to-one mapping between $x$ and $\Theta$
of spiral solutions in terms of Jacobi elliptic functions,
although the mapping is similar to, but algebraically more involved
than the explicit spiral solutions in simple cases:
Eq.~\eqref{eq: solution_ellipticFunc_A=0} in the case of $A=0$,
and Eq.~\eqref{eq: solution_ellipticFunc_B=0-} in the case of $B=0, A<0$.
One should note that the functional form of the profile function
$\Theta(x)$ of the
domain wall solutions or spiral solutions are determined solely
by the parameters of the potential $A,B$ and $C_0$, but are
independent of the DM interaction parameter $\kappa$.
The DM interaction parameter comes in only when we evaluate
the average energy density of the spiral solutions or domain wall solutions.
Therefore the lowest energy spiral solutions and the critical point
explicitly depends on the DM parameter $\kappa$.
Similarly the domain wall solutions can exist irrespective of the value
of $\kappa$, but they become zero energy solitons at the
phase boundary between spiral and homogeneous ferromagnetic
phases.
Let us examine the infinite period limit of the lowest energy
spiral solutions.
As we approach the phase boundary, the period of spiral solutions
tends to infinity, and the solutions become domain wall solutions,
as we argued.
At the boundary we find the average energy density
becomes identical to that of the homogeneous solution.
Moreover, the total energy of the domain wall vanishes at the boundary.
We will explicitly see these properties by studying domain wall
solutions in all parameter regions across the phase boundary
in the following section.
The domain wall solutions $\Theta(x)$ can be expressed in terms
of elementary functions.
This corresponds to the fact that the elliptic function describing
a spiral solution becomes an elementary function in the limit
of infinite period.
Let us take the limit of infinite period for illustrative examples
of spiral solutions in the case of $A=0$ and $B=0$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/sgdw2}
\caption{This is the profile function $\Theta$ from Eq.~\eqref{eq:sg profile and magnetisation} plotted against the dimensionless variable $\kappa x$ for the critical value $\frac{B}{\kappa^{2}}=\frac{\pi^{2}}{16}$ at $A=0$.
}
\label{fig:anisotropy domain wall}
\end{center}
\end{figure}
\paragraph{$A=0, B>0$ Case :}
One period of the spiral solution \eqref{eq: solution_ellipticFunc_A=0}
reduces to two sets of the following domain wall solution in the
limit of infinite period $k\to 1$ corresponding to
$B=\frac{(\pi\kappa)^{2}}{16}$
\begin{equation}
\begin{split}
\Theta&=-4\arctan\left(\exp(\sqrt{B}x)\right),\\
\vec{n}^{\alpha}&=\left(0, 2\frac{\tanh\left(\sqrt{B}x\right)}
{\cosh\left(\sqrt{B}x\right)},
1-2\sech^{2}\left(\sqrt{B}x\right)\right)^{T}.
\end{split}\label{eq:sg profile and magnetisation}
\end{equation}
We note that this is a domain wall solution for any $B>0$ with
$A=0$.
The total energy of the domain wall is zero at the phase transition
point $B=\frac{(\pi\kappa)^{2}}{16}$, positive in the positively
polarized ferromagnetic phase $B>\frac{(\pi\kappa)^{2}}{16}$, and
negative in the spiral phase $B<\frac{(\pi\kappa)^{2}}{16}$.
We plot this on the right in
Fig.~\ref{fig:anisotropy domain wall}.
It should be noted that while we have referred to the configurations as domain walls for $A=0, B>0$ they do not connect two different magnetic domains. In fact they interpolate between a magnetic domain and itself, since their phase rotates by a full $2\pi$. Some times these configurations would be called kinks, however, we would reserve that for when they are genuine soliton solutions with positive energy above the ferromagnetic phase. As such we stick to calling them domain walls.
\section{Domain walls in chiral magnets}\label{sec:exactdomainwalls}
\subsection{Exact domain wall solutions}
\label{sec:domain-wall}
Since the field equation~\eqref{eq:theta-EOM2} involves
only the potential parameters $A, B$ and is independent of the DM
interaction parameter $\kappa$, the shapes of domain wall solutions
are identical to those of the double sine-Gordon model.
For instance, Ref.~\cite{Condat1983} gave domain wall solutions
for the double sine-Gordon model, but in an entirely different
physical context.
The functional form of our exact solutions agree with those in
Ref.~\cite{Condat1983} after adjusting for the different conventions.
However, our energy functional involves the DM term, which gives
a crucial negative energy contribution for the energy of domain
wall solution as solitons.
This fact is the basis of our finding of zero energy domain
wall solutions at the boundary between the homogeneous phase
and the spiral phase.
We will derive exact domain wall solutions and evaluate their energy,
in order to be reasonably self-contained.
A domain wall is a soliton with a localized energy.
We will demonstrate explicitly that it has
positive energy in the homogeneous phase, negative energy in
the spiral phase, and zero energy at the phase boundary.
The domain wall solutions are found to correspond
to the infinite period limit of the lowest energy spiral solution.
These features show that domain wall solutions
are solitons for excited states in the homogeneous phase,
but are instability modes for the homogeneous background solution
in the spiral phase.
This picture agrees with the fact that the lowest energy spiral
solution gives the ground state in the spiral phase region,
which has lower average energy density than the homogeneous
solution.
The domain wall solutions of our interest are defined as those
solutions of the first order equation
\eqref{eq: theta-equation} with the mechanical energy
tuned as $C_0=-U_{\rm min}$.
The first order equation for $\Theta(x)$ in Eq.~\eqref{eq: theta-equation}
becomes
\begin{equation}
\frac{d\Theta}{dx}=- \sqrt{2(-B\cos\Theta+A\cos^2\Theta+U_{\rm min})},
\label{eq: theta-equation_DW}
\end{equation}
which gives a monotonically decreasing function $\Theta(x)$
connecting two adjacent minima $\Theta_{\rm min}$ of the
potential.
\subsection{Domain wall solutions for \texorpdfstring{$0\le B \le 2A$}{0< B <2A} }
\label{sec:domain-wall-Bless2A}
In the parameter region
$0\le B \le 2A$, the minimum of the potential
is $U_{\rm min}=-\frac{B^{2}}{4A}$, and the homogeneous phase is given
by the canted ferromagnetic solution with $n_3=\frac{B}{2A}$.
The first order field equation~\eqref{eq: theta-equation_DW}
becomes
\begin{equation}
\frac{d\Theta}{dx}=- \sqrt{2A}\left|\cos\Theta-\frac{B}{2A}\right|.
\label{eq: theta-equation_DW_Ble2A}
\end{equation}
Since there are two degenerate minima in one period $2\pi$,
we have two different types of domain walls:
the long wall connecting $\Theta=2\pi-\arccos\frac{B}{2A}$ at $x=-\infty$
to $\arccos\frac{B}{2A}$ at $x=\infty$, and the short wall connecting
$\Theta=\arccos\frac{B}{2A}$ at $x=-\infty$
to $-\arccos\frac{B}{2A}$ at $x=\infty$.
\subsubsection{Long wall }
\label{sec:long-wall}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.48\textwidth]{figures/long_dw_profile2}
\includegraphics[width=0.48\textwidth]{figures/BPS_dw_profile2}
\caption{On the left is a plot of the profile function $\Theta$ as a
function of the dimensionless quantity $\kappa x$ for the long wall for $0\le B \le 2A$ in
Eq.~\eqref{eq: long_DW_Bless2A}.
We take $\frac{A}{\kappa^{2}}=3,\frac{B}{\kappa^{2}}=2$ as an example, and the
domain wall to be centred at $x=0$. On the right is the profile of the domain wall solution $\Theta$ and in Eq.~\eqref{eq:DW_B=2A}
plotted against the dimensionless quantity $\kappa x$
for $B=2A=\frac{1}{2}\kappa^2$, with the domain wall to be centred at $x=0$.}
\label{fig:anisotropy dominated domain wall2}
\end{center}
\end{figure}
The long wall is obtained by integrating \eqref{eq: theta-equation_DW_Ble2A}
with $\frac{B}{2A}-\cos\Theta\ge 0$.
By choosing $\Theta=\pi$ at $x=0$ (midpoint of the domain wall),
we obtain the solution connecting $\Theta=2\pi-\arccos\frac{B}{2A}$
at $x=-\infty$ to $\arccos\frac{B}{2A}$ at $x=\infty$ as
\begin{align}
x&=
- \frac{1}{\sqrt{2A}}\int_\pi^\Theta
\frac{d\Theta'}{\frac{B}{2A}-\cos\Theta'}
\nonumber \\
&=\sqrt{\frac{2A}{4A^2-B^2}}\log\left(
\frac{\tan\frac{\Theta}{2}+\sqrt{\frac{2A-B}{2A+B}}}
{\tan\frac{\Theta}{2}-\sqrt{\frac{2A-B}{2A+B}}}\right)
\nonumber \\
&=\sqrt{\frac{2A}{4A^2-B^2}}\; 2\; {\rm arccoth}\left(
\sqrt{\frac{2A+B}{2A-B}}\tan\frac{\Theta}{2}\right)
.
\label{eq: long_DW_Bless2A}
\end{align}
On the left in Fig.~\ref{fig:anisotropy dominated domain wall2}, we plot
the domain wall profile $\Theta$ as a function of $x$.
In the limit of $2A \to B+0$, the domain wall solution reduces to
\begin{eqnarray}
x= \frac{1}{\sqrt{B}}\cot\frac{\Theta}{2},
\label{eq:DW_B=2A}
\end{eqnarray}
where we choose the branch $2\pi\ge\Theta\ge\pi$ for $-\infty\le x\le 0$,
and $\pi\ge\Theta\ge0$ for $0\le x\le \infty$.
This solution is depicted on the right in Fig.~\ref{fig:anisotropy dominated domain wall2}.
\subsubsection{Short wall}
\label{sec:short-wall}
The short wall is obtained by integrating \eqref{eq: theta-equation_DW_Ble2A}
with $\cos\Theta-\frac{B}{2A}\ge 0$.
By choosing $\Theta=0$ at $x=0$ (midpoint of the domain wall),
we obtain the solution connecting $\Theta=\arccos\frac{B}{2A}$
at $x=-\infty$ to $-\arccos\frac{B}{2A}$ at $x=\infty$ as
\begin{align}
x&=
- \frac{1}{\sqrt{2A}}\int_0^\Theta
\frac{d\Theta'}{\cos\Theta'-\frac{B}{2A}}
\nonumber \\
&=\sqrt{\frac{2A}{4A^2-B^2}}\log\left(
\frac{\sqrt{\frac{2A-B}{2A+B}}-\tan\frac{\Theta}{2}}
{\sqrt{\frac{2A-B}{2A+B}}+\tan\frac{\Theta}{2}}\right)
\nonumber \\
&=-\sqrt{\frac{2A}{4A^2-B^2}}\; 2\; {\rm arctanh}\left(
\sqrt{\frac{2A+B}{2A-B}}\tan\frac{\Theta}{2}\right)
.
\label{eq: short_DW_Bless2A}
\end{align}
In Fig.~\ref{fig:anisotropy dominated domain wall}, we plot
the domain wall profile $\Theta$ as a function of the dimensionless quantity $\kappa x$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/short_dw_profile}
\caption{This is a plot of the profile function $\Theta$ as a
function of the dimensionless quantity $\kappa x$ for the short wall for $0\le B \le 2A$ in
Eq.~\eqref{eq: short_DW_Bless2A}.
We take $\frac{A}{\kappa^{2}}=4,\frac{B}{\kappa^{2}}=2$ as an example, and the domain wall to be
centered at $\kappa x=0$. }
\label{fig:anisotropy dominated domain wall}
\end{center}
\end{figure}
In the limit of $B\to0$, the long wall connects $\Theta=\frac{3\pi}{2}$
to $\frac{\pi}{2}$, whereas the short wall connects $\Theta=\frac{\pi}{2}$ to
$-\frac{\pi}{2}$.
We find that the long and short walls at $B\to 0$ are precisely
identical to those walls emerging from the infinite period
limit of the spiral solution at $B=0$.
\subsubsection{Energy of long and short domain walls}
\label{sec:energy_long-short-wall}
For the long wall, the value of $\Theta$ spans from
$2\pi-\arccos\frac{B}{2A}$ at $x=-\infty$ to $\arccos\frac{B}{2A}$
at $x=\infty$.
The total energy of the long domain wall on the homogeneous
solution (with energy density $U_{\rm min}$) as background is
given by integrating over the energy density in
Eq.~\eqref{eq:energy-density-solution} with $C_0=-U_{\rm min}$:
\begin{align}
{E}_{\rm long}&=\int_{-\infty}^{\infty}dx
\left({\mathcal E}_{\rm sol}-U_{\rm min}\right)
\nonumber \\
&=\int_{-\infty}^{\infty}dx
\left(\kappa\frac{d\Theta}{dx}
+2(U(\Theta)-U_{\rm min})\right) .
\label{eq:energy-long-wall1}
\end{align}
Changing variable from $x$ to $\Theta$, we obtain by using
the domain wall equation in Eq.~\eqref{eq: theta-equation_DW}
with $\frac{d\Theta}{dx}=-\sqrt{2(U(\Theta)-U_{\rm min})}
=-\sqrt{2A}(\frac{B}{2A}-\cos\Theta)$:
\begin{align}
{E}_{\rm long}&=\int_{2\pi-\arccos\frac{B}{2A}}^{\arccos\frac{B}{2A}}
d\Theta \left(\frac{d\Theta}{dx}\right)^{-1}
\left(\kappa\frac{d\Theta}{dx}+2A\left(\frac{B}{2A}-\cos\Theta\right)^2
\right)
\label{eq:energy-long-wall2}
\\
&=-2\kappa\left(\pi-\arccos\frac{B}{2A}\right)
+\frac{2B}{\sqrt{2A}}\left(\pi-\arccos\frac{B}{2A}\right)
+2\sqrt{2A-\frac{B^2}{2A}} .
\nonumber
\end{align}
Similarly, the total energy of the short domain wall on the homogeneous
solution as the background is given by integrating over the energy density.
The only difference from the long wall is that $\Theta$ varies
from $\arccos\frac{B}{2A}$ at $x=-\infty$ to $-\arccos\frac{B}{2A}$
at $x=\infty$, and $\cos\Theta-\frac{B}{2A}\ge0$.
By noting $\frac{d\Theta}{dx}=-\sqrt{2(U(\Theta)-U_{\rm min})}
=-\sqrt{2A}(\cos\Theta-\frac{B}{2A})$, we find
\begin{align}
{E}_{\rm short}&=\int_{-\arccos\frac{B}{2A}}^{\arccos\frac{B}{2A}}
d\Theta \left(\frac{d\Theta}{dx}\right)^{-1}
\left(\kappa\frac{d\Theta}{dx}+2A\left(\cos\Theta-\frac{B}{2A}\right)^2
\right)
\label{eq:energy-short-wall} \\
&=-2\kappa\arccos\frac{B}{2A}
-\frac{2B}{\sqrt{2A}}\arccos\frac{B}{2A}
+2\sqrt{2A-\frac{B^2}{2A}} .
\nonumber
\end{align}
In order to obtain domain wall solutions as the infinite period limit of a spiral solution, we need to consider domain walls
connecting minimum of the potential which are $\Delta\Theta=2\pi$ apart.
Such limiting domain walls precisely correspond to the sum of
a pair of long and short domain walls.
The sum of the energy of the short and long domain wall is given by
\begin{align}
\label{eq:sum_energy-long-short}
E_{\rm period}&={E}_{\rm long}+E_{\rm short}
\\
&=-2\pi\kappa
+\frac{4B}{\sqrt{2A}}\left(\frac{\pi}{2}-\arccos\frac{B}{2A}\right)
+4\sqrt{2A-\frac{B^2}{2A}} .
\nonumber
\end{align}
Thus we see that the energy of a pair of long and short domain
wall solutions is positive in the (homogeneous) ferromagnetic
phase, negative in the spiral phase, and zero along the dashed red line in Fig.~\ref{fig:Phase boundaries}.
Considering both the $B\to 2A$ and the $B\to 0$ limits of the long and short walls we see that they have the same energy in the $B\to 0$ limit, while the long wall has lower energy in the $B\to 2A$ limit. In fact for $B\to 2A$ the short domain wall disappears and has zero energy.
\subsection{Domain wall solutions for \texorpdfstring{$B>0$}{B>0} and \texorpdfstring{$B>2A$}{B>2A} }
\label{sec:domain-wall-2AlessB}
In the parameter region
$B>0$ and $B>2A$, the minimum of the potential
is $U_{\rm min}=A-B$, and the homogeneous phase is given
by the polarized ferromagnetic solution with $n_3=1$.
The first order equation \eqref{eq: theta-equation_DW}
becomes
\begin{equation}
\frac{d\Theta}{dx}=- \sqrt{2(1-\cos\Theta)(B-A-A\cos\Theta)}.
\label{eq: theta-equation_DW_2AleB}
\end{equation}
In the region
$B>0$ and $B>2A$, the minimum of the potential occur at
$\Theta=2\pi{\mathbb Z}$.
Let us consider a domain wall connecting $\Theta=2\pi$ at $x=-\infty$
to $\Theta=0$ at $x=\infty$.
By choosing the midpoint of the domain wall $\Theta=\pi$ at $x=0$, we find
\begin{align}
x&=
- \int_\pi^\Theta
\frac{d\Theta'}{\sqrt{2(1-\cos\Theta')(B-A-A\cos\Theta')}}
\nonumber \\
&=\frac{1}{\sqrt{B-2A}}\log\left(
\sqrt{1+\left(1-\frac{2A}{B}\right)\cot^2\frac{\Theta}{2}}
+\sqrt{1-\frac{2A}{B}}\cot\frac{\Theta}{2}\right).
\label{eq: DW_2AlessB}
\end{align}
In Fig.~\ref{fig:B dominated domain wall}, we plot
the domain wall profile $\Theta$ as a function of $x$.
In the limit of $2A \to B-0$, the domain wall solution reduces to
an identical form to the domain wall obtained in the limit
of $2A\to B+0$
from Eq.~\eqref{eq:DW_B=2A}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/magfield_dominated_dw2}
\caption{This is a plot of the profile function $\Theta$ as a
function of the dimensionless variable $\kappa x$ for a domain wall in the magnetic field dominated
region $B\geq 2A$ in Eq.~\eqref{eq: DW_2AlessB}.
We take $A=\frac{\kappa^2}{2},B=2\kappa^2$ as an example, and
the domain wall to be
centered at $x=0$.
}
\label{fig:B dominated domain wall}
\end{center}
\end{figure}
Since the energy density of the homogeneous solution (polarized
ferromagnetic ground state) is $U_{\rm min}=A-B$, the total energy
of the domain wall on the homogeneous background is given by
\begin{align}
{E}_{\rm DW}
&=\int_{-\infty}^{\infty}dx
\left(\kappa\frac{d\Theta}{dx}+2(U(\Theta)-U_{\rm min})\right)
\nonumber \\
&=
-2\pi\kappa+4\sqrt{B-2A}
+\frac{4B}{\sqrt{2A}}\arcsin\sqrt{\frac{2A}{B}},
\label{eq:energy-domain-wall_2AleB}
\end{align}
in the case of $A> 0$, and
\begin{equation}
{E}_{\rm DW}
=
-2\pi\kappa+4\sqrt{B-2A}
+\frac{4B}{\sqrt{-2A}}\arcsinh\sqrt{\frac{-2A}{B}},
\label{eq:energy-domain-wall_2AleB-2}
\end{equation}
in the case of $A\le0$.
We see that the zero energy condition for the domain wall gives
the boundary between the homogeneous phase and spiral phase,
as given in Eqs.~\eqref{eq:phase-boundary-case2} and
\eqref{eq:phase-boundary-case2-2}.
Moreover, the domain wall is a positive energy soliton as an
excited state in the homogeneous phase.
In the spiral phase, on the other hand,
the domain wall on the (homogeneous) ferromagnetic background
solution gives a negative energy solution, signaling the instability
of the ferromagnetic (homogeneous) solution.
These observations confirm our conclusion that the homogeneous
solution is unstable in the spiral phase and decays into the
spiral solution, giving the ground state.
\section{Conclusion and discussion }\label{sec:summary}
We have studied solutions of the field equations
for chiral magnets in one spatial dimension,
in order to determine their phase diagram.
There are three distinct homogeneous phases: the canted polarized
ferromagnetic phase with $n_3\not=\pm 1$, the positive polarized
ferromagnetic phase with $n_3=1$, and the negative polarized phase
with $n_3=-1$.
These homogeneous ferromagnetic phases are realized for large
values of potential parameters, whereas the spiral phase is
realized for small values of the potential parameters.
We have explicitly determined the exact phase boundary between spiral and homogeneous
ferromagnetic phases. As mentioned previously, there is numerical evidence in \cite{Chovan} that a lower energy non-flat spiral state is possible when $B=0$. This state cannot be studied analytically which is why we have not discussed it here, and we have focused on the parameter region $2A\leq \vert B\vert$ where the flat spiral is the ground state.
The lowest energy spiral solution has an infinite period
at the phase boundary, and becomes a domain wall solution with
zero energy.
We have constructed domain wall solutions and found that they have
positive energy in the homogeneous ferromagnetic phases and
have negative energy in the spiral phase, exhibiting the instability
of the homogeneous ferromagnetic solution in the spiral phase region.
We have also found that the phase transitions between spiral phase and
homogeneous ferromagnetic phases are of second order.
Our results should be useful when discussing the phase diagram of
chiral magnets in spatial dimensions two or higher~\cite{BY,BH,
BRS,Schroers1,RSN,DM,Melcher}.
The homogeneous and spiral solutions are also present
in higher spatial dimensions.
However, We need to examine possible instabilities due to the fluctuations
around these solutions, in particular those emerging from additional
dimensions.
More importantly, we need to consider other inhomogeneous solutions
that are intrinsically two or higher dimensional, such as
skyrmions or merons and their lattices.
Another interesting question to be addressed is the low energy
effective field theory on the spiral ground state.
Since the spiral solutions are spatially inhomogeneous, the
low energy effective field theory is not invariant under translation
nor rotation. Therefore the dispersion relation of fluctuations
is expected to exhibit asymmetry in this regards.
It has been already computed for the simplest spiral ground state
for the chiral magnet without the potential, and an interesting
anisotropic dispersion relation was found in Ref.~\cite{Hongo:2020xaw}.
Moreover, the frequency of the fluctuations depends on the
structure of the time derivative, which is different for
ferromagnetic, antiferromagnetic, or ferrimagnetic materials
\cite{Kobayashi:2015pra}.
For a single domain wall, the first order time derivative
introduces a type-B NG mode due to
spontaneously broken translational and U(1) symmetries
\cite{Kobayashi:2014xua}.
It is also interesting to consider a more general potential.
The more general interaction terms arising from antiferromagnetic
materials have been considered phenomenologically \cite{BRWM}.
In this work, we considered only the effect of various solitonic
objects such as domain walls and spiral solutions
in a
mean-field approximation.
It is an interesting future task to consider quantum effects,
including nonperturbative effects in the chiral magnet.
\begin{acknowledgments}
This work is supported in part by the Japan Society for the Promotion
of Science (JSPS)
Grant-in-Aid for Scientific Research (KAKENHI) Grant Number
18H01217 (M.~N. and N.~S.).
\end{acknowledgments}
| bb4203b845ef1de88ed86fcc152324808007d4e2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec1}
In this paper, we are concerned with the Cauchy problem for the following two-component b-family system on $\mathbb{R}$:
\begin{eqnarray}\label{eq1}
\left\{\begin{array}{ll}
m_t=um_x+ k_1u_xm+k_2\rho\rho_x,\\
\rho_t=k_3(u\rho)_x,\\
m=u-u_{xx},\\
u(0, x)=u_0, \rho(0, x)=\rho_0. \end{array}\right.
\end{eqnarray}
which was introduced by Guha in \cite{Guha 2007}. As shown in \cite{Guha 2007}, there are two cases about this system: (i) $k_1=b$, $k_2=2b$ and $k_3=1$; (ii) $k_1=b+1,$ $k_2=2$ and $k_3=b$ with $b\in \mathbb{R}.$
If $k_1=2$ and $k_3=1$, then system (\ref{eq1}) becomes the following two-component Camassa-Holm system
\begin{eqnarray}\label{eq2}
\left\{\begin{array}{ll}
m_t=um_x+ 2u_xm+\sigma\rho\rho_x,\\
\rho_t=(u\rho)_x,\\
m=u-u_{xx}, \end{array}\right.
\end{eqnarray}
here $\sigma=\pm 1.$ System (\ref{eq2}) was derived by Constantin and Ivanov \cite{Constantin 2008} in the context of shallow water theory, and then has attracted much more attention. The local well-posedness for system (\ref{eq2}) in Sobolev and Besov space were established in \cite{Escher 2007, Constantin 2008, Guan 2010, Gui 2011}. The global existence of strong solutions and wave-breaking mechanism were investigated in
\cite{Escher 2007, Guan 2010, Gui 2011, Gui 2010}, and the global weak solution
has been obtained in \cite{Guan 2011}. Moreover, when $\rho=0$, (\ref{eq2}) reduces to the classical Camassa-Holm equation modeling the unidirectional propagation of shallow water waves over a flat bottom.
The Cauchy problem of the Camassa-Holm equation was studied in the
series of papers \cite{Constantin 1998, 1Constantin 1998, Constantin 2000, 1Constantin 2000, Danchin 2001, Li 2016, Guo 2019}. Danchin\cite{Danchin 2001, Danchin 2003} showed the local existence and uniqueness of strong solutions to Camassa-Holm equation with initial data in $B_{p, r}^s$ for $s>\max\{1+\frac{1}{p}, \frac{3}{2}\}$, $1\leq p\leq \infty$, $1\leq r<\infty$. The continuous dependence of the solution on the initial data has been supplemented by Li and Yin in \cite{Li 2016}. Very recently, Li, Yu and Zhu \cite{Li 2020} have sharpened the results in \cite{Li 2016} by showing that the solution map is not uniformly continuous (see \cite{Himonas 2009} for earlier results in Sobolev space $H^s(\mathbb{R})$ with $s>\frac{3}{2}$).
If $k_1=3$ and $k_3=2$, then system (\ref{eq1}) becomes the following two-component Degasperis-Procesi system
\begin{eqnarray}\label{eq3}
\left\{\begin{array}{ll}
m_t=um_x+ 3u_xm+\sigma\rho\rho_x,\\
\rho_t=2(u\rho)_x,\\
m=u-u_{xx}, \end{array}\right.
\end{eqnarray}
which was first proposed in \cite{Popowicz 2006} as a natural generalization of the Degasperis-Procesi equation in the context of supersymmetry. The local well-posedness of system (\ref{eq3}) was established in \cite{Jin 2010, Escher 2011, Yan 2012}, the precise blow-up scenario and some blow-up rate of strong solutions were also presented in \cite{Jin 2010, Yan 2012}. Moreover, when $\rho=0$, (\ref{eq3}) reduces to the famous Degasperis-Procesi equation. The first result of nonuniform dependence for Degasperis-Procesi equation in the periodic case with $s\geq 2$ can be found by Christov and Hakkaev
\cite{Christov 2009}. This result was lately improved to the best by Himonas and Holliman in \cite{Himonas 2011}, where non-uniform dependence on the initial data for Degasperis-Procesi equation was proven on both the circle and line for $s>\frac{3}{2}$ using the method of approximate solutions in tandem with a twisted $L^2$-norm that is conserved by the DP equation. Recently, the non-uniform continuity has been once extended to the Besov space \cite{Wu 2021}.
For $\rho\neq 0$, and $b\in \mathbb{R}$, the Cauchy problem of system (\ref{eq1}) in Sobolev space was first established by Liu and Yin \cite{Liu 2011} for $(u_0, \rho_0)\in H^s \times H^{s-1}$ with $s\geq 2$. Later, Lv and Wang \cite{Lv 2014} proved that the solution map is not uniformly continuous for $s>\frac 52$ by using the approximate solutions.
The local well-posedness space was further enlarged, and established in Besov space $B_{p, r}^s\times B_{p, r}^{s-1}$ with $s>\max\{1+\frac{1}{p}, \frac{3}{2}\}$, $1\leq p, r\leq \infty$ (however, for $r=\infty$, the continuity of the data-to-solution map is established in a
weaker topology)\cite{Zhu 2013, Yan 2020}. Some aspects concerning blow-up scenario, global solutions, persistence properties and propagation speed, see
the discussions in \cite{Liu 2011, Zong 2012}.
In the present paper, motivated by \cite{Li 2020, Wu 2021}, we aim at showing that the solution map of (\ref{eq1}) is not uniformly continuous depending on the initial data in Besov spaces $B_{p, r}^s\times B_{p, r}^{s-1}$ with $s>\max\{1+\frac{1}{p}, \frac{3}{2}\}$, $1\leq p, r< \infty$. However, one problematic issue is that we here deal with a coupled system with these two components of the solution in different Besov spaces. On the other hand, compared with Novikov equation with cubic nonlinearity \cite{1Li 2020}, quadratic nonlinearity for system (\ref{eq1}) weakens the attenuation we need. These two aspects make the proof of several required nonlinear estimates somewhat delicate.
For studying the non-uniform continuity of a two-component b-family system, it is more convenient to express (\ref{eq1}) in the following equivalent nonlocal form
\begin{eqnarray}\label{eq4}
\left\{\begin{array}{ll}
u_t-uu_x=f(u)+g(\rho),\\
\rho_t-k_3u\rho_x=k_3\rho u_x,\\
u(0, x)=u_0, \rho(0, x)=\rho_0, \end{array}\right.
\end{eqnarray}
where $f(u)=f_1(u)+f_2(u)$ and
\begin{eqnarray*}
f_1(u)=\partial_x(1-\partial_x^2)^{-1}(\frac{k_1}{2}u^2), \quad f_2(u)=\partial_x(1-\partial_x^2)^{-1}(\frac{3-k_1}{2}u_x^2), \quad g(\rho)=\partial_x(1-\partial_x^2)^{-1}(\frac{k_2}{2}\rho^2).
\end{eqnarray*}
Our main result is stated as follows.
\begin{theorem}\label{the1.1} Let $s>\max\{1+\frac{1}{p}, \frac{3}{2}\}$, $1\leq p, r < \infty$. The solution map $(u_0, \rho_0)\rightarrow (u(t), \rho(t))$ of the initial value problem (\ref{eq4}) is not
uniformly continuous from any bounded subset of $B_{p, r}^s(\mathbb{R})\times B_{p, r}^{s-1}(\mathbb{R})$ into $\mathcal{C}([0, T]; B_{p, r}^s(\mathbb{R})\times B_{p, r}^{s-1}(\mathbb{R}))$. More precisely,
there exist two sequences of solutions $(u_n^1(t), \rho_n^1(t))$ and $(u_n^2(t), \rho_n^2(t))$ such that the corresponding initial data satisfy
\begin{eqnarray*}
\|u_n^1(t), u_n^2(t)\|_{B_{p, r}^s}+\|\rho_n^1(t), \rho_n^2(t)\|_{B_{p, r}^{s-1}}\lesssim 1,\\
\end{eqnarray*}
\mbox{and}
\begin{eqnarray*}
\lim_{n\rightarrow \infty}(\|u_n^1(0)-u_n^2(0)\|_{B_{p, r}^s}+\|\rho_n^1(0)-\rho_n^2(0)\|_{B_{p, r}^{s-1}})=0,
\end{eqnarray*}
but
\begin{eqnarray*}
\liminf_{n\rightarrow \infty} \|u_n^1(t)-u_n^2(t)\|_{B_{p, r}^s}\gtrsim t, \quad \liminf_{n\rightarrow \infty} \|\rho_n^1(t)-\rho_n^2(t)\|_{B_{p, r}^{s-1}}\gtrsim t, \quad t\in [0, T_0],
\end{eqnarray*}
with small positive time $T_0$ for $T_0\leq T$.
\end{theorem}
\begin{remark}\label{rem1}
Since $B_{2, 2}^s=H^s$, our result covers and extends the previous non-uniform continuity of solutions on initial data in Sobolev spaces $H^s(\mathbb{R})\times H^{s-1}(\mathbb{R})$ for $s>\frac{5}{2}$ \cite{Lv 2014} to Besov spaces.
\end{remark}
{\bf Notations}: Given a Banach space $X$, we denote the norm of a function on $X$ by $\|\|_{X}$, and \begin{eqnarray*}
\|\cdot\|_{L_T^\infty(X)}=\sup_{0\leq t\leq T}\|\cdot\|_{X}.
\end{eqnarray*}
For $\mathbf{f}=(f_1, f_2,...,f_n)\in X$,
\begin{eqnarray*}
\|\mathbf{f}\|_{X}^2=\|f_1\|_{X}^2+\|f_2\|_{X}^2+...+\|f_n\|_{X}^2.
\end{eqnarray*}
The symbol
$A\lesssim B$ means that there is a uniform positive constant $C$ independent of $A$ and $B$ such that $A\leq CB$.
\section{Littlewood-Paley analysis}\label{sec2}
\setcounter{equation}{0}
In this section, we will review the definition of Littlewood-Paley decomposition and nonhomogeneous Besov space, and then list some useful properties. For more details, the readers can refer to \cite{Bahouri 2011}.
There exists a couple of smooth functions $(\chi,\varphi)$ valued in $[0,1]$, such that $\chi$ is supported in the ball $\mathcal{B}\triangleq \{\xi\in\mathbb{R}:|\xi|\leq \frac 4 3\}$, $\varphi$ is supported in the ring $\mathcal{C}\triangleq \{\xi\in\mathbb{R}:\frac 3 4\leq|\xi|\leq \frac 8 3\}$. Moreover,
$$\forall\,\, \xi\in\mathbb{R},\,\, \chi(\xi)+{\sum\limits_{j\geq0}\varphi(2^{-j}\xi)}=1,$$
$$\forall\,\, \xi\in\mathbb{R}\setminus\{0\},\,\, {\sum\limits_{j\in \mathbb{Z}}\varphi(2^{-j}\xi)}=1,$$
$$|j-j'|\geq 2\Rightarrow\textrm{Supp}\,\ \varphi(2^{-j}\cdot)\cap \textrm{Supp}\,\, \varphi(2^{-j'}\cdot)=\emptyset,$$
$$j\geq 1\Rightarrow\textrm{Supp}\,\, \chi(\cdot)\cap \textrm{Supp}\,\, \varphi(2^{-j}\cdot)=\emptyset.$$
Then, we can define the nonhomogeneous dyadic blocks $\Delta_j$ as follows:
$$\Delta_j{u}= 0,\,\, \text{if}\,\, j\leq -2,\quad
\Delta_{-1}{u}= \chi(D)u=\mathcal{F}^{-1}(\chi \mathcal{F}u),$$
$$\Delta_j{u}= \varphi(2^{-j}D)u=\mathcal{F}^{-1}(\varphi(2^{-j}\cdot)\mathcal{F}u),\,\, \text{if} \,\, j\geq 0.$$
\begin{definition}[\cite{Bahouri 2011}]\label{de2.1}
Let $s\in\mathbb{R}$ and $1\leq p,r\leq\infty$. The nonhomogeneous Besov space $B^s_{p,r}(\mathbb{R})$ consists of all tempered distribution $u$ such that
\begin{align*}
\|u\|_{B^s_{p,r}(\mathbb{R})}\triangleq \Big|\Big|(2^{js}\|\Delta_j{u}\|_{L^p(\mathbb{R})})_{j\in \mathbb{Z}}\Big|\Big|_{\ell^r(\mathbb{Z})}<\infty.
\end{align*}
\end{definition}
In the following, we list some basic lemmas and properties about Besov space which will be frequently used in proving our main result.
\begin{lemma}(\cite{Bahouri 2011})\label{lem2.1}
(1) Algebraic properties: $\forall s>0,$ $B_{p, r}^s(\mathbb{R})$ $\cap$ $L^\infty(\mathbb{R})$ is a Banach algebra. $B_{p, r}^s(\mathbb{R})$ is a Banach algebra $\Leftrightarrow B_{p, r}^s(\mathbb{R})\hookrightarrow L^\infty(\mathbb{R})\Leftrightarrow s>\frac{1}{p}$ or $s=\frac{1}{p},$ $r=1$.\\
(2) For any $s>0$ and $1\leq p,r\leq\infty$, there exists a positive constant $C=C(s,p,r)$ such that
$$\|uv\|_{B^s_{p,r}(\mathbb{R})}\leq C\Big(\|u\|_{L^{\infty}(\mathbb{R})}\|v\|_{B^s_{p,r}(\mathbb{R})}+\|v\|_{L^{\infty}(\mathbb{R})}\|u\|_{B^s_{p,r}(\mathbb{R})}\Big).$$
(3) Let $m\in \mathbb{R}$ and $f$ be an $S^m-$ multiplier (i.e., $f: \mathbb{R}^d\rightarrow \mathbb{R}$ is smooth and satisfies that $\forall \alpha\in \mathbb{N}^d$, there exists a constant $\mathcal{C}_\alpha$ such that $|\partial^\alpha f(\xi)|\leq \mathcal{C}_\alpha(1+|\xi|)^{m-|\alpha|}$ for all $\xi \in \mathbb{R}^d$). Then the operator $f(D)$ is continuous from $B_{p, r}^s(\mathbb{R}^d)$ to $B_{p, r}^{s-m}(\mathbb{R}^d)$.\\
(4) Let $1\leq p, r\leq \infty$ and $s>\max\{1+\frac{1}{p}, \frac{3}{2}\}$. Then we have
$$\|uv\|_{B_{p, r}^{s-2}(\mathbb{R})}\leq C\|u\|_{B_{p, r}^{s-2}(\mathbb{R})}\|v\|_{B_{p, r}^{s-1}(\mathbb{R})}.$$
\end{lemma}
\begin{lemma}\label{lem2.2}(\cite{Bahouri 2011, Li 2017})
Let $1\leq p,r\leq \infty$. Assume that
\begin{eqnarray*}
\sigma> -\min\{\frac{1}{p}, 1-\frac{1}{p}\} \quad \mathrm{or}\quad \sigma> -1-\min\{\frac{1}{p}, 1-\frac{1}{p}\}\quad \mathrm{if} \quad \mathrm{div\,} v=0.
\end{eqnarray*}
There exists a constant $C=C(p,r,\sigma)$ such that for any solution to the
following linear transport equation:
\begin{equation*}
\partial_t f+v\cdot\nabla f=g,\qquad
f|_{t=0} =f_0,
\end{equation*}
the following statements hold:
\begin{align*}
\sup_{s\in [0,t]}\|f(s)\|_{B^{\sigma}_{p,r}}\leq Ce^{CV_{p}(v,t)}\Big(\|f_0\|_{B^\sigma_{p,r}}
+\int^t_0\|g(\tau)\|_{B^{\sigma}_{p,r}}d \tau\Big),
\end{align*}
with
\begin{align*}
V_{p}(v,t)=
\begin{cases}
\int_0^t \|\nabla v(s)\|_{B^{\sigma-1}_{p,r}}ds, &\quad \mathrm{if} \;\sigma>1+\frac{1}{p}\ \mathrm{or}\ \{\sigma=1+\frac{1}{p} \mbox{ and } r=1\},\\
\int_0^t \|\nabla v(s)\|_{B^{\sigma}_{p,r}}ds,&\quad\mathrm{if} \; \sigma=1+\frac{1}{p} \quad \mathrm{and} \quad r>1,\\
\int_0^t \|\nabla v(s)\|_{B^{\frac{1}{p}}_{p,\infty}\cap L^\infty}ds,&\quad\mathrm{if} \; \sigma<1+\frac{1}{p}.
\end{cases}
\end{align*}
\end{lemma}
\section{Non-uniform continuous dependence}\label{sec3}
\setcounter{equation}{0}
In this section, we will give the proof of our main theorem. For brevity, we sometimes use $u_{0, n}^i$, $\rho_{0, n}^i$ to denote $u^i_n(0)$ and $\rho^i_n(0)$ respectively, $i=1, 2.$
Let $\hat{\phi}\in \mathcal{C}^\infty_0(\mathbb{R})$ be an even, real-valued and non-negative funtion on $\R$ and satisfy
\begin{numcases}{\hat{\phi}(x)=}
1, &if $|x|\leq \frac{1}{4}$,\nonumber\\
0, &if $|x|\geq \frac{1}{2}$.\nonumber
\end{numcases}
Define the high frequency function $f_n$ and the low frequency function $g_n$ by
$$f_n=2^{-ns}\phi(x)\sin \bi(\frac{17}{12}2^nx\bi), \qquad g_n=2^{-n}\phi(x), \quad n\gg1.$$
It has been showed in \cite{Li 2020} that $\|f_n\|_{B_{p, r}^\sigma}\lesssim 2^{n(\sigma-s)}$.
Let
$$(u_n^1(0),\; \rho_n^1(0))=(f_n, 2^nf_n), \quad (u_n^2(0),\; \rho_n^2(0))=(f_n+g_n, 2^nf_n+g_n),$$
then it is easy to verify that
\begin{align}
\|u_n^1(0), \; u_n^2(0)\|_{B_{p, r}^{s+\sigma}}\lesssim 2^{n\sigma}\; \mathrm{for}\; \;\sigma\geq -1 \;\;\;\mathrm{and} \;\;\;\|\rho_n^1(0),\; \rho_n^2(0)\|_{B_{p, r}^{s+l}}\lesssim 2^{n(l+1)}\;\mathrm{for}\; \; l\geq -2,\label{eq3.1}
\end{align}
Consider the system (\ref{eq4}) with initial data $(u_n^1(0), \rho_n^1(0))$ and $(u_n^2(0), \rho_n^2(0))$, respectively. According to the local well-posedness result in \cite{Zhu 2013, Yan 2020}, there exists corresponding solution $(u_n^1, \rho_n^1)$, $(u_n^2, \rho_n^2)$ belonging to
$\mathcal{C}([0, T]; B_{p, r}^s\times B_{p, r}^{s-1})$ and has common lifespan $T\thickapprox 1$. Moreover, by Lemma \ref{lem2.1}-\ref{lem2.2}, there holds
\begin{eqnarray}
&\|u_n^1\|_{L_T^\infty(B_{p, r}^{s+k})}+\|\rho_n^1\|_{L_T^\infty(B_{p, r}^{s+k-1})}\lesssim \|u_n^1(0)\|_{B_{p, r}^{s+k}}+\|\rho_n^1(0)\|_{B_{p, r}^{s+k-1}}\lesssim2^{nk},\quad k\geq -1,\label{eq3.2}\\
& \|u_n^2\|_{L_T^\infty(B_{p, r}^{s+l})}+\|\rho_n^2\|_{L_T^\infty(B_{p, r}^{s+l-1})}\lesssim \|u_n^2(0)\|_{B_{p, r}^{s+l}}+\|\rho_n^2(0)\|_{B_{p, r}^{s+l-1}}\lesssim2^{nl},\quad l\geq -1.\label{eq3.3}
\end{eqnarray}
In the following, we shall firstly show that for the selected high frequency initial data $(u_n^1(0), \rho_n^1(0))$, the corresponding solution $(u_n^1, \rho_n^1)$ can be approximated by the initial data. More precisely, that is
\begin{proposition}\label{pro1}
Under the assumptions of Theorem \ref{the1.1}, we have
\begin{eqnarray}
\|u_n^1-u_n^1(0)\|_{L_T^\infty(B_{p, r}^s)}+ \|\rho_n^1-\rho_n^1(0)\|_{L_T^\infty(B_{p, r}^{s-1})}\lesssim 2^{-\frac{n}{2}(s-\frac 32)}.\label{eq3.4}
\end{eqnarray}
\end{proposition}
\noindent{\bf Proof} \;
Denote $\epsilon=u_n^1-u_n^1(0)$, $\delta=\rho_n^1-\rho_n^1(0)$, then we can derive from (\ref{eq4}) that $(\epsilon, \varrho)$ satisfies
\begin{eqnarray}\label{eq3.5}
\left\{\begin{array}{ll}
\epsilon_t-u_n^1\partial_x\epsilon=(u_n^1-u^1_{0, n})\partial_xu^1_{0, n}
+[f(u_n^1)-f(u^1_{0, n})]+[g(\rho_n^1)-g(\rho^1_{0, n})]\\
\qquad\qquad\qquad+f(u^1_{0, n})+g(\rho^1_{0, n})+u^1_{0, n}\partial_xu^1_{0, n},\\
\delta_t-k_3u_n^1\partial_x\delta=k_3(u_n^1-u^1_{0, n})\partial_x\rho^1_{0, n}
+k_3u^1_{0, n}\partial_x\rho^1_{0, n}+k_3\rho_n^1\partial_xu_n^1,\\
\epsilon(0, x)=0, \delta(0, x)=0, \end{array}\right.
\end{eqnarray}
Applying Lemma \ref{lem2.2} yields
\begin{align}
\|\epsilon\|_{B_{p, r}^{s-1}}&\lesssim \int_0^t\|\partial_xu_n^1\|_{B_{p, r}^{s-1}}\|\epsilon\|_{B_{p, r}^{s-1}}d\tau+\int_0^t \|(u_n^1-u^1_{0, n})\partial_xu^1_{0, n}\|_{B_{p, r}^{s-1}}d\tau \nonumber\\
&\quad +\int_0^t \|f(u_n^1)-f(u^1_{0, n}), \;g(\rho_n^1)-g(\rho^1_{0, n})\|_{B_{p, r}^{s-1}}d\tau \nonumber\\
&\quad+t\|f(u^1_{0, n}),\;g(\rho^1_{0, n}),\;u^1_{0, n}\partial_xu^1_{0, n}\|_{B_{p, r}^{s-1}},\label{eq3.6}\\
\|\delta\|_{B_{p, r}^{s-2}}&\lesssim \int_0^t\|\partial_xu_n^1\|_{B_{p, r}^{s-1}}\|\delta\|_{B_{p, r}^{s-2}}d\tau+\int_0^t \|(u_n^1-u^1_{0, n})\partial_x\rho^1_{0, n}\|_{B_{p, r}^{s-2}}d\tau \nonumber\\
&\quad +\int_0^t \|\rho_n^1\partial_xu_n^1-\rho^1_{0, n}\partial_xu_{0, n}^1\|_{B_{p, r}^{s-2}}d\tau +t\|u^1_{0, n}\partial_x\rho^1_{0, n}, \;\rho^1_{0, n}\partial_xu_{0, n}^1\|_{B_{p, r}^{s-2}}.\label{eq3.7}
\end{align}
Using Lemma \ref{lem2.1} and the fact that $B_{p, r}^{s-1}(\mathbb{R})$ is a Banach algebra when $s>\max\{1+\frac{1}{p}, \frac{3}{2}\}$, we have
\begin{align*}
\|(u_n^1-u^1_{0, n})\partial_xu^1_{0, n}\|_{B_{p, r}^{s-1}}
&\lesssim \|u_n^1-u^1_{0, n}\|_{B_{p, r}^{s-1}}
\|\partial_xu^1_{0, n}\|_{B_{p, r}^{s-1}}\lesssim \|u_n^1-u^1_{0, n}\|_{B_{p, r}^{s-1}}\|u^1_{0, n}\|_{B_{p, r}^s},\\
\|f(u_n^1)-f(u^1_{0, n})\|_{B_{p, r}^{s-1}}&\lesssim\|u_n^1-u^1_{0, n}\|_{B_{p, r}^{s-1}}\|u_n^1, \; u^1_{0, n}\|_{B_{p, r}^s},\\
\|g(\rho_n^1)-g(\rho^1_{0, n})\|_{B_{p, r}^{s-1}}&\lesssim\|\rho_n^1-\rho^1_{0, n}\|_{B_{p, r}^{s-2}}\|\rho_n^1, \;\rho^1_{0, n}\|_{B_{p, r}^{s-1}},\\
\|(u_n^1-u^1_{0, n})\partial_x\rho^1_{0, n}\|_{B_{p, r}^{s-2}} &\lesssim\|\partial_x\rho^1_{0, n}\|_{B_{p, r}^{s-2}}\|u_n-u_{0, n}\|_{B_{p, r}^{s-1}}\lesssim \|u_n-u_{0, n}\|_{B_{p, r}^{s-1}}\|\rho^1_{0, n}\|_{B_{p, r}^{s-1}},\\
\|u^1_{0, n}\partial_x\rho^1_{0, n}\|_{B_{p, r}^{s-2}} &\lesssim\|u^1_{0, n}\partial_x\rho^1_{0, n}\|_{B_{p, r}^{s-\frac 32}}\lesssim\|u^1_{0, n}\|_{L^\infty}\|\rho^1_{0, n}\|_{B_{p, r}^{s-\frac 12}}+\|u^1_{0, n}\|_{B_{p, r}^{s-\frac 32}}\|\partial_x\rho^1_{0, n}\|_{L^\infty},\\
&\lesssim 2^{-ns}2^{n(-\frac 12 +1)}+2^{-\frac 32n}2^n2^{-ns}2^n\lesssim 2^{-n(s-\frac 12)},\\
\|\rho^1_{0, n}\partial_xu_{0, n}^1\|_{B_{p, r}^{s-2}}&\lesssim\|\rho^1_{0, n}\partial_xu_{0, n}^1\|_{B_{p, r}^{s-\frac 32}}\lesssim2^{n(s-\frac 32)}\|\rho^1_{0, n}, \partial_xu_{0, n}^1\|_{L^\infty}\|\rho^1_{0, n}, \partial_xu_{0, n}^1\|_{L^p}\lesssim 2^{-n(s-\frac 12)}.
\end{align*}
Again using Lemma \ref{lem2.1} and the Banach algebra property of $B_{p, r}^{s-1}$, one has
\begin{align*}
\|f_1(u_{0, n}^1)\|_{B_{p, r}^{s-1}}&\lesssim\|(u_{0, n}^1)^2\|_{B_{p, r}^{s-2}}\lesssim\|(u_{0, n}^1)^2\|_{B_{p, r}^{s-\frac 32}}\lesssim \|u_{0, n}^1\|_{L^\infty}\|u_{0, n}^1\|_{B_{p, r}^{s-\frac 32}}\lesssim 2^{-n(s+\frac 32)},\\
\|f_2(u_{0, n}^1)\|_{B_{p, r}^{s-1}}&\lesssim\|(\partial_xu_{0, n}^1)^2\|_{B_{p, r}^{s-2}}\lesssim \|\partial_xu_{0, n}^1\|_{L^\infty}\|\partial_xu_{0, n}^1\|_{B_{p, r}^{s-\frac 32}}\lesssim 2^{-n(s-\frac 12)},\\
\|g(\rho_{0, n}^1)\|_{B_{p, r}^{s-1}}&\lesssim\|(\rho_{0, n}^1)^2\|_{B_{p, r}^{s-2}}\lesssim\|(\rho_{0, n}^1)^2\|_{B_{p, r}^{s-\frac 32}}\lesssim \|\rho_{0, n}^1\|_{L^\infty}\|\rho_{0, n}^1\|_{B_{p, r}^{s-\frac 32}}\lesssim 2^{-n(s-\frac 12)},\\
\|u_{0, n}^1\partial_xu_{0, n}^1\|_{B_{p, r}^{s-1}}
&\lesssim \|u_{0, n}^1\|_{L^\infty} \|\partial_xu_{0, n}^1\|_{B_{p, r}^{s-1}}+\|u_{0, n}^1\|_{B_{p, r}^{s-1}}\|\partial_xu_{0, n}^1\|_{L^\infty}\lesssim 2^{-ns}.
\end{align*}
For the term
\begin{eqnarray*}
\rho_n^1\partial_xu_n^1-\rho^1_{0, n}\partial_xu_{0, n}^1=(\rho_n^1-\rho^1_{0, n})\partial_xu_n^1
+ \rho^1_{0, n}\partial_x(u_n^1-u^1_{0, n}).
\end{eqnarray*}
Following the same procedure of estimates as above, we find that
\begin{align*}
\|(\rho_n^1-\rho^1_{0, n})\partial_xu_n^1\|_{B_{p, r}^{s-2}}&\lesssim\|\rho_n^1-\rho^1_{0, n}\|_{B_{p, r}^{s-2}}\|\partial_xu_n^1\|_{B_{p, r}^{s-1}}\lesssim\|\rho_n^1-\rho^1_{0, n}\|_{B_{p, r}^{s-2}}\|u_n^1\|_{B_{p, r}^s},\\
\|\rho^1_{0, n}\partial_x(u_n^1-u^1_{0, n})\|_{B_{p, r}^{s-2}}&\lesssim\|\partial_x(u_n^1-u^1_{0, n})\|_{B_{p, r}^{s-2}}\|\rho^1_{0, n}\|_{B_{p, r}^{s-1}}\lesssim\|u_n^1-u^1_{0, n}\|_{B_{p, r}^{s-1}}\|\rho^1_{0, n}\|_{B_{p, r}^{s-1}}.
\end{align*}
Denote
\begin{eqnarray*}
X_s=\|\epsilon\|_{B_{p, r}^s}+\|\delta\|_{B_{p, r}^{s-1}},
\end{eqnarray*}
taking the above estimates into (\ref{eq3.6})-(\ref{eq3.7}), we get
\begin{eqnarray*}
X_{s-1}\lesssim \int_0^tX_{s-1}(\|u_n^1,\; u^1_{0, n}\|_{B_{p, r}^s}+\|\rho_n^1,\;\rho^1_{0, n}\|_{B_{p, r}^{s-1}})+t2^{-n(s-\frac 12)},
\end{eqnarray*}
since $\{u_n^1, \rho_n^1\}$ is bounded in $B_{p, r}^s\times B_{p, r}^{s-1}$,
which together with the Gronwall Lemma imply
$$ X_{s-1}\lesssim 2^{-n(s-\frac 12)}.$$
Combining with (\ref{eq3.2}) for $k=1$ and the interpolation inequality, we obtain that
$$X_s\lesssim X_{s-1}^{\frac{1}{2}}X_{s+1}^{\frac{1}{2}}\lesssim 2^{-\frac{n}{2}(s-\frac 12)}2^{\frac{n}{2}}\lesssim 2^{-\frac{n}{2}(s-\frac 32)}.$$
Thus we have complete the proof of Proposition \ref{pro1}.
In order to obtain the non-uniformly continuous dependence property for the system (\ref{eq4}), we will show that for the constructed initial data $(u_n^2(0), \rho_n^2(0))$ with small perturbation, it can not approximate to the solution
$(u_n^2, \rho_n^2)$.
\begin{proposition}\label{pro2}
Under the assumptions of Theorem \ref{the1.1}, we have
\begin{eqnarray}\label{eq3.9}
\|u_n^2-u_{0, n}^2-t\mathbf{v}_0^n\|_{B_{p, r}^s}+\|\rho_n^2-\rho_{0, n}^2-t\mathbf{w}_0^n\|_{B_{p, r}^{s-1}}\lesssim t^2+2^{-n\min\{s-\frac 32,\; \frac 12\}},
\end{eqnarray}
here, $\mathbf{v}_0^n=u_{0, n}^2\partial_xu_{0, n}^2,$ $\mathbf{w}_0^n=k_3u_{0, n}^2\partial_x\rho_{0, n}^2$.
\end{proposition}
\noindent{\bf Proof} \; Firstly, due to (\ref{eq3.1}) and making full use of the product estimates in Lemma \ref{lem2.1}, for $\sigma\geq -1$, we have
\begin{align}
\|\mathbf{v}_0^n\|_{B_{p, r}^{s+\sigma}}&\lesssim \|u_{0, n}^2\|_{L^\infty}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s+\sigma}}+ \|u_{0, n}^2\|_{B_{p, r}^{s+\sigma}}\|\partial_xu_{0, n}^2\|_{L^\infty}\nonumber\\
&\lesssim (2^{-ns}+2^{-n})2^{n(\sigma+1)}+2^{n\sigma}(2^{-ns}2^n+2^{-n})\lesssim 2^{n\sigma}\label{eq3.9}\\
\|\mathbf{w}_0^n\|_{B_{p, r}^{s+\sigma}}&\lesssim \|u_{0, n}^2\|_{L^\infty}\|\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s+\sigma}}+ \|u_{0, n}^2\|_{B_{p, r}^{s+\sigma}}\|\partial_x\rho_{0, n}^2\|_{L^\infty}\nonumber\\
&\lesssim (2^{-ns}+2^{-n})2^{n(\sigma+2)}+2^{n\sigma}(2^n2^{-ns}2^n+2^{-n})\lesssim 2^{n(\sigma+1)},\label{eq3.10}\\
\|\mathbf{w}_0^n\|_{B_{p, r}^{s-2}}&\lesssim \|\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-2}}\|u_{0, n}^2\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}.\label{eq3.11}
\end{align}
Let
\begin{eqnarray*}
\left\{\begin{array}{ll}
z_n=u_n^2-u_{0, n}^2-t\mathbf{v}_0^n,\\
\omega_n=\rho_n^2-\rho_{0, n}^2-t\mathbf{w}_0^n,\end{array}\right.
\end{eqnarray*}
then we can derive from (\ref{eq4}) that $(z_n, \omega_n)$ satisfies
\begin{eqnarray}\label{eq3.12}
\left\{\begin{array}{ll}
\partial_tz_n-u_n^2\partial_xz_n=(z_n+t\mathbf{v}_0^n)\partial_xu_{0, n}^2+tu_n^2\partial_x\mathbf{v}_0^n+f(u_n^2)+g(\rho_n^2)\\
\partial_t\omega_n-k_3u_n^2\partial_x\omega_n=k_3(z_n+t\mathbf{v}_0^n)\partial_x\rho_{0, n}^2+k_3tu_n^2\partial_x\mathbf{w}_0^n+k_3\rho_n^2\partial_xu_n^2,\\
z_n(0, x)=0, \; \omega(0, x)=0.
\end{array}\right.
\end{eqnarray}
Applying Lemma \ref{lem2.1}, using (\ref{eq3.3}), (\ref{eq3.9})-(\ref{eq3.10}), we arrive at
\begin{align}
\|z_n\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-1}}&\lesssim \|z_n\|_{B_{p, r}^{s-1}}\|\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-1}}\lesssim 2^n\|z_n\|_{B_{p, r}^{s-1}}, \label{eq3.13}\\
\|z_n\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-2}}&\lesssim \|\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-2}}\|z_n\|_{B_{p, r}^{s-1}}\lesssim \|z_n\|_{B_{p, r}^{s-1}}, \label{eq3.14}\\
\|\mathbf{v}_0^n\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-1}}&\lesssim \|\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\|\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}2^n\leq C, \label{eq3.15}\\
\|\mathbf{v}_0^n\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-2}}&\lesssim \|\partial_x\rho_{0, n}^2\|_{B_{p, r}^{s-2}}\|\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}, \label{eq3.16}\\
\|u_n^2\partial_x\mathbf{w}_0^n\|_{B_{p, r}^{s-1}}&\lesssim \|u_n^2\|_{B_{p, r}^{s-1}}\|\partial_x\mathbf{w}_0^n\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}2^n\leq C, \label{eq3.17} \\
\|u_n^2\partial_x\mathbf{w}_0^n\|_{B_{p, r}^{s-2}}&\lesssim \|\partial_x\mathbf{w}_0^n\|_{B_{p, r}^{s-2}}\|u_n^2\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}. \label{eq3.18}
\end{align}
It needs to pay more attention to deal with the term $k_3\rho_n^1\partial_xu_n^1$ and it can be can be decomposed as
\begin{equation*}
\rho_n^2\partial_xu_n^2=\rho_n^2\partial_xz_n+(\omega_n+\rho_{0, n}^2)\partial_xu_{0, n}^2+t(\rho_n^2\partial_x\mathbf{v}_0^n+\mathbf{w}_0^n\partial_xu_{0 ,n}^2).
\end{equation*}
With (\ref{eq3.3}), (\ref{eq3.9})-(\ref{eq3.10}) at hand, by Lemma \ref{lem2.1}, we find that
\begin{align}
\|\rho_n^2\partial_xz_n\|_{B_{p, r}^{s-1}}&\lesssim \|\rho_n^2\|_{B_{p, r}^{s-1}}\|\partial_xz_n\|_{B_{p, r}^{s-1}}\lesssim \|z_n\|_{B_{p, r}^s}, \label{eq3.19}\\
\|\rho_n^2\partial_xz_n\|_{B_{p, r}^{s-2}}&\lesssim \|\partial_xz_n\|_{B_{p, r}^{s-2}}\|\rho_n^2\|_{B_{p, r}^{s-1}}\lesssim \|z_n\|_{B_{p, r}^{s-1}}, \label{eq3.20}\\
\|\omega_n\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}&\lesssim \|\omega_n\|_{B_{p, r}^{s-1}}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}\lesssim \|\omega_n\|_{B_{p, r}^{s-1}}, \label{eq3.21}\\
\|\omega_n\partial_xu_{0, n}^2\|_{B_{p, r}^{s-2}}&\lesssim \|\omega_n\|_{B_{p, r}^{s-2}}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}\lesssim \|\omega_n\|_{B_{p, r}^{s-2}}, \label{eq3.22}\\
\|\rho_{0, n}^2\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}&\lesssim \|\rho_{0, n}^2\|_{L^\infty}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}+ \|\rho_{0, n}^2\|_{B_{p, r}^{s-1}}\|\partial_xu_{0, n}^2\|_{L^\infty}\nonumber\\
&\lesssim (2^{-ns}2^n+2^{-n})\cdot 1+1\cdot(2^{-ns}2^n+2^{-n})\lesssim 2^{-n\min\{s-1,\;1\}}\label{eq3.23}\\
\|\rho_{0, n}^2\partial_xu_{0, n}^2\|_{B_{p, r}^{s-2}}&\lesssim \|\rho_{0, n}^2\partial_xu_{0, n}^2\|_{B_{p, r}^{s-\frac 32}}\lesssim\|\rho_{0, n}^2\|_{L^\infty}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-\frac 32}}+ \|\rho_{0, n}^2\|_{B_{p, r}^{s-\frac 32}}\|\partial_xu_{0, n}^2\|_{L^\infty}\nonumber\\
&\lesssim (2^{-ns}2^n+2^{-n})\cdot 2^{-\frac n2}+2^{-\frac n2}\cdot(2^{-ns}2^n+2^{-n})\lesssim 2^{-n\min\{s-\frac 12,\;\frac 32\}},\label{eq3.24}\\
\|\rho_n^2\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}&\lesssim \|\rho_n^2\|_{B_{p, r}^{s-1}}\|\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\leq C, \label{eq3.25}\\
\|\rho_n^2\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-2}}&\lesssim \|\rho_n^2\|_{B_{p, r}^{s-2}}\|\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}, \label{eq3.26}\\
\|\mathbf{w}_0^n\partial_xu_{0 ,n}^2\|_{B_{p, r}^{s-1}}&\lesssim \|\mathbf{w}_0^n\|_{B_{p, r}^{s-1}}\|\partial_xu_{0 ,n}^2\|_{B_{p, r}^{s-1}}\leq C, \label{eq3.27}\\
\|\mathbf{w}_0^n\partial_xu_{0 ,n}^2\|_{B_{p, r}^{s-2}}&\lesssim \|\mathbf{w}_0^n\|_{B_{p, r}^{s-2}}\|\partial_xu_{0 ,n}^2\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}, \label{eq3.28}
\end{align}
Applying Lemma \ref{lem2.2} to the second equation of (\ref{eq3.12}), using the fact that $\{u_n^2\}$ is bounded in $L^\infty_T(B_{p, r}^s)$, firstly with (\ref{eq3.14}), (\ref{eq3.16}), (\ref{eq3.18}), (\ref{eq3.20}), (\ref{eq3.22}), (\ref{eq3.24}), (\ref{eq3.26}), (\ref{eq3.28}), we infer that
\begin{equation}\label{eq3.29}
\|\omega_n\|_{B_{p, r}^{s-2}}\leq C\int_0^t(\|z_n\|_{B_{p, r}^{s-1}}+\|\omega_n\|_{B_{p, r}^{s-2}})d\tau+Ct^22^{-n}+C2^{-n\min\{s-\frac 12,\; \frac 32\}},
\end{equation}
and again combining with (\ref{eq3.13}), (\ref{eq3.15}), (\ref{eq3.17}), (\ref{eq3.19}), (\ref{eq3.21}), (\ref{eq3.23}), (\ref{eq3.25}), (\ref{eq3.27}), we obtain that
\begin{equation}\label{eq3.30}
\|\omega_n\|_{B_{p, r}^{s-1}}\leq C\int_0^t(\|z_n\|_{B_{p, r}^s}+\|\omega_n\|_{B_{p, r}^{s-1}})d\tau+C\int_0^t2^n\|z_n\|_{B_{p, r}^{s-1}}d\tau+Ct^2+C2^{-n\min\{s-1,\; 1\}}.
\end{equation}
In the following, we shall estimate $z_n$ in $B_{p, r}^{s-1}$ and $B_{p, r}^s$, respectively.
With the aid of Lemma \ref{lem2.1} and (\ref{eq3.3}), (\ref{eq3.9}), one has
\begin{align}
\|z_n\partial_xu_{0, n}^2\|_{B_{p, r}^s}&\lesssim \|z_n\|_{L^\infty}\|\partial_xu_{0, n}^2\|_{B_{p, r}^s}
+\|z_n\|_{B_{p, r}^s}\|\partial_xu_{0, n}^2\|_{L^\infty}\nonumber\\
&\lesssim \|z_n\|_{B_{p, r}^{s-1}}\|\partial_xu_{0, n}^2\|_{B_{p, r}^s}
+\|z_n\|_{B_{p, r}^s}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}\nonumber\\
&\lesssim 2^n\|z_n\|_{B_{p, r}^{s-1}}+\|z_n\|_{B_{p, r}^s},\label{eq3.31}\\
\|z_n\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}&\lesssim \|z_n\|_{B_{p, r}^{s-1}}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}\lesssim \|z_n\|_{B_{p, r}^{s-1}},\label{eq3.32}\\
\|\mathbf{v}_0^n\partial_xu_{0, n}^2\|_{B_{p, r}^s}
&\lesssim \|\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\|\partial_xu_{0, n}^2\|_{B_{p, r}^s}
+\|\mathbf{v}_0^n\|_{B_{p, r}^s}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}\leq C,\label{eq3.33}\\
\|\mathbf{v}_0^n\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}
&\lesssim \|\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\|\partial_xu_{0, n}^2\|_{B_{p, r}^{s-1}}\lesssim 2^{-n},\label{eq3.34}\\
\|u_{0, n}^2\partial_x\mathbf{v}_0^n\|_{B_{p, r}^s}
&\lesssim \|u_{0, n}^2\|_{B_{p, r}^{s-1}}\|\partial_x\mathbf{v}_0^n\|_{B_{p, r}^s}
+\|u_{0, n}^2\|_{B_{p, r}^s}\|\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\leq C,\label{eq3.35}\\
\|u_{0, n}^2\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}&\lesssim\|u_{0, n}^2\|_{B_{p, r}^{s-1}}\|\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\lesssim 2^{-n}. \label{eq3.36}
\end{align}
For the term $f(u_n^2)=f_1(u_n^2)+f_2(u_n^2),$ we have from Lemma \ref{lem2.1} and (\ref{eq3.3}) that
\begin{equation}\label{eq3.37}
\|f_1(u_n^2)\|_{B_{p, r}^s}\lesssim \|(u_n^2)^2\|_{B_{p, r}^{s-1}}\lesssim \|u_n^2\|^2_{B_{p, r}^{s-1}}\lesssim 2^{-2n},
\end{equation}
while it needs to be more careful to deal with $f_2(u_n^2)$. By making full use of the structure of $u_n^2$, we find that
\begin{align*}
f_2(u_n^2)=&\underbrace{\frac {3-k_1}{2}\partial_x(1-\partial_x^2)^{-1}(\partial_x(u_n^2+u_{0, n}^2)\partial_xz_n)}_{f_{2, 1}}\\
&\;+\underbrace{\frac {3-k_1}{2}\partial_x(1-\partial_x^2)^{-1}(t\partial_x(u_n^2+u_{0, n}^2)\partial_x\mathbf{v}_0^n)}_{f_{2, 2}}\\
&\;+\underbrace{\frac {3-k_1}{2}\partial_x(1-\partial_x^2)^{-1}((\partial_xu^2_{0, n})^2)}_{f_{2, 3}},
\end{align*}
and
\begin{align}
\|f_{2, 1}\|_{B_{p, r}^s}&\lesssim \|\partial_x(u_n^2+u_{0, n}^2)\partial_xz_n\|_{B_{p, r}^{s-1}}\lesssim \|z_n\|_{B_{p, r}^s}, \label{eq3.38}\\
\|f_{2, 1}\|_{B_{p, r}^{s-1}}&\lesssim \|\partial_x(u_n^2+u_{0, n}^2)\partial_xz_n\|_{B_{p, r}^{s-2}}\lesssim \|z_n\|_{B_{p, r}^{s-1}}, \label{eq3.39}\\
\|f_{2, 2}\|_{B_{p, r}^s}&\lesssim \|\partial_x(u_n^2+u_{0, n}^2)\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-1}}\leq C,\label{eq3.40}\\
\|f_{2, 2}\|_{B_{p, r}^{s-1}}&\lesssim \|\partial_x(u_n^2+u_{0, n}^2)\partial_x\mathbf{v}_0^n\|_{B_{p, r}^{s-2}}\lesssim 2^{-n},\label{eq3.41}\\
\|f_{2, 3}\|_{B_{p, r}^s}&\lesssim \|(\partial_xu^2_{0, n})^2\|_{B_{p, r}^{s-1}}\lesssim\|\partial_xu^2_{0, n}\|_{L^\infty}\|\partial_xu^2_{0, n}\|_{B_{p, r}^{s-1}}\nonumber\\
&\lesssim(2^{-ns}2^n+2^{-n})\cdot 1\lesssim 2^{-n\min\{s-1, \; 1\}}, \label{eq3.42}\\
\|f_{2, 3}\|_{B_{p, r}^{s-1}}&\lesssim \|(\partial_xu^2_{0, n})^2\|_{B_{p, r}^{s-\frac 32}}\lesssim\|\partial_xu^2_{0, n}\|_{L^\infty}\|\partial_xu^2_{0, n}\|_{B_{p, r}^{s-\frac 32}}\nonumber\\
&\lesssim(2^{-ns}2^n+2^{-n})\cdot 2^{-\frac n2}\lesssim 2^{-n\min\{s-\frac 12, \; \frac 32\}}. \label{eq3.43}
\end{align}
$g(\rho_n^2)$ can be performed in a similar way. Firstly, we obtain that
\begin{align*}
g(\rho_n^2)=&\underbrace{\frac {k_2}{2}\partial_x(1-\partial_x^2)^{-1}(\omega_n(\rho_n^2+\rho_{0, n}^2))}_{g_1}+\underbrace{\frac {k_2}{2}\partial_x(1-\partial_x^2)^{-1}(t\mathbf{w}_0^n(\rho_n^2+\rho_{0, n}^2))}_{g_2}\\
&\;+\underbrace{\frac {k_2}{2}\partial_x(1-\partial_x^2)^{-1}((\rho^2_{0, n})^2)}_{g_3},
\end{align*}
and
\begin{align}
\|g_1\|_{B_{p, r}^s}&\lesssim \|\omega_n(\rho_n^2+\rho_{0, n}^2)\|_{B_{p, r}^{s-1}}\lesssim \|\omega_n\|_{B_{p, r}^{s-1}}, \label{eq3.44}\\
\|g_1\|_{B_{p, r}^{s-1}}&\lesssim \|\omega_n(\rho_n^2+\rho_{0, n}^2)\|_{B_{p, r}^{s-2}}\lesssim \|\omega_n\|_{B_{p, r}^{s-2}}, \label{eq3.45}\\
\|g_2\|_{B_{p, r}^s}&\lesssim \|\mathbf{w}_0^n(\rho_n^2+\rho_{0, n}^2)\|_{B_{p, r}^{s-1}}\leq C,\label{eq3.46}\\
\|g_2\|_{B_{p, r}^{s-1}}&\lesssim \|\mathbf{w}_0^n(\rho_n^2+\rho_{0, n}^2)\|_{B_{p, r}^{s-2}}\lesssim 2^{-n},\label{eq3.47}\\
\|g_3\|_{B_{p, r}^s}&\lesssim \|(\rho^2_{0, n})^2\|_{B_{p, r}^{s-1}}\lesssim\|\rho^2_{0, n}\|_{L^\infty}\|\rho^2_{0, n}\|_{B_{p, r}^{s-1}}\nonumber\\
&\lesssim(2^{-ns}2^n+2^{-n})\cdot 1\lesssim 2^{-n\min\{s-1, \; 1\}}, \label{eq3.48}
\end{align}
\begin{align}
\|g_3\|_{B_{p, r}^{s-1}}&\lesssim \|(\rho^2_{0, n})^2\|_{B_{p, r}^{s-\frac 32}}\lesssim\|\rho u^2_{0, n}\|_{L^\infty}\|\rho^2_{0, n}\|_{B_{p, r}^{s-\frac 32}}\nonumber\\
&\lesssim(2^{-ns}2^n+2^{-n})\cdot 2^{-\frac n2}\lesssim 2^{-n\min\{s-\frac 12, \; \frac 32\}}. \label{eq3.49}
\end{align}
Applying Lemma \ref{lem2.2} firstly together with (\ref{eq3.32}), (\ref{eq3.34}), (\ref{eq3.36}), (\ref{eq3.37}), (\ref{eq3.39}), (\ref{eq3.41}), (\ref{eq3.43}), (\ref{eq3.45}), (\ref{eq3.47}), (\ref{eq3.49}) to the first equation of (\ref{eq3.12}), using the fact that $\{u^2_n\}$ is bounded in $L^\infty_T(B_{p, r}^s)$, we infer that
\begin{equation}\label{eq3.50}
\|z_n\|_{B_{p, r}^{s-1}}\leq C\int_0^t(\|z_n\|_{B_{p, r}^{s-1}}+\|\omega_n\|_{B_{p, r}^{s-2}})d\tau+Ct^22^{-n}+C2^{-n\min\{s-\frac 12,\; \frac 32\}},
\end{equation}
and again combining with (\ref{eq3.31}), (\ref{eq3.33}), (\ref{eq3.35}), (\ref{eq3.37}), (\ref{eq3.38}), (\ref{eq3.40}), (\ref{eq3.42}), (\ref{eq3.44}), (\ref{eq3.46}), (\ref{eq3.48}), we obtain
\begin{equation}\label{eq3.51}
\|z_n\|_{B_{p, r}^s}\leq C\int_0^t(\|z_n\|_{B_{p, r}^s}+\|\omega_n\|_{B_{p, r}^{s-1}})d\tau+ C\int_0^t2^{n}\|z_n\|_{B_{p, r}^{s-1}}d\tau+Ct^2+C2^{-n\min\{s-1, \; 1\}}.
\end{equation}
Using Gronwall Lemma to (\ref{eq3.23}) and (\ref{eq3.43}) imply
\begin{align*}
\|z_n\|_{B_{p, r}^{s-1}}+\|\omega_n\|_{B_{p, r}^{s-2}}\leq Ct^22^{-n}+C2^{-n\min\{s-\frac 12,\; \frac 32\}},
\end{align*}
which together with (\ref{eq3.30}) and (\ref{eq3.51}) yield that
\begin{equation*}
\|z_n\|_{B_{p, r}^s}+\|\omega_n\|_{B_{p, r}^{s-1}}\leq Ct^2+C2^{-n\min\{s-\frac 32,\; \frac 12\}}.
\end{equation*}
Thus, we have finished the proof of Proposition \ref{pro2}.
{\bf Proof of Theorem \ref{the1.1}} It is obvious that
\begin{align*}
\|u_{0, n}^2-u_{0, n}^1\|_{B_{p, r}^s}&=\|g_n\|_{B_{p, r}^s}\leq C2^{-n},\\
\|\rho_{0, n}^2-\rho_{0, n}^1\|_{B_{p, r}^{s-1}}&=\|g_n\|_{B_{p, r}^{s-1}}\leq C2^{-n},
\end{align*}
which means that
\begin{equation*}
\lim_{n\rightarrow \infty}(\|u_{0, n}^2-u_{0, n}^1\|_{B_{p, r}^s}+ \|\rho_{0, n}^2-\rho_{0, n}^1\|_{B_{p, r}^{s-1}})=0.
\end{equation*}
However, according to Proposition \ref{pro1} and Proposition \ref{pro2}, we get
\begin{align}\label{eq3.52}
\|\rho_n^2-\rho_n^1\|_{B_{p, r}^{s-1}}&=\|\omega_n+t\mathbf{w}_0^n+g_n+\rho_{0, n}^1-\rho_n^1\|_{B_{p, r}^{s-1}}\nonumber\\
&\gtrsim t\|\mathbf{w}_0^n\|_{B_{p, r}^{s-1}}-2^{-n}-t^2-2^{-n\min\{s-\frac 32,\; \frac 12\}}-2^{-\frac n2(s-\frac 32)}.
\end{align}
Notice that
\begin{align*}
\mathbf{w}_0^n&=k_3u_{0, n}^2\partial_x\rho_{0, n}^2=k_3(f_n+g_n)\partial_x(2^nf_n+g_n)\\
&=k_3f_n\partial_x(2^nf_n)+k_3g_n\partial_x(2^nf_n)+k_3f_n\partial_xg_n+k_3g_n\partial_xg_n.
\end{align*}
With the aid of Lemma \ref{lem2.1} and the Banach algebra property of $B_{p, r}^{s-1}$, we find that
\begin{align*}
\|f_n\partial_x(2^nf_n)\|_{B_{p, r}^{s-1}}&\lesssim \|f_n\|_{L^\infty}\|\partial_x(2^nf_n)\|_{B_{p, r}^{s-1}}+\|f_n\|_{B_{p, r}^{s-1}}\|\partial_x(2^nf_n)\|_{L^\infty}\\
&\lesssim 2^{-ns}2^n+2^{-n}2^n2^{-ns}2^n\lesssim 2^{-n(s-1)},\\
\|f_n\partial_xg_n\|_{B_{p, r}^{s-1}}&\lesssim \|f_n\|_{B_{p, r}^{s-1}}\|\partial_xg_n\|_{B_{p, r}^{s-1}}\lesssim 2^{-2n},\\
\|g_n\partial_xg_n\|_{B_{p, r}^{s-1}}&\lesssim \|g_n\|_{B_{p, r}^{s-1}}\|\partial_xg_n\|_{B_{p, r}^{s-1}}\lesssim 2^{-2n}.
\end{align*}
However, using the fact that $\Delta_j\big(g_n\partial_x(2^nf_n)\big)=0, j\neq n$ and $\Delta_n\big(g_n\partial_x(2^nf_n)\big)=g_n\partial_x(2^nf_n)$ for $n\geq 5$, direct calculation
shows that for $n\gg 1$,
\begin{align*}
&\|g_n\partial_x(2^nf_n)\|_{B_{p, r}^{s-1}}=2^{n(s-1)}\|g_n\partial_x(2^nf_n)\|_{L^p}\\
=&\|2^{-n}\phi(x)\partial_x\phi(x)\sin(\frac{17}{12}2^nx)+\frac{17}{12}\phi^2(x)\cos(\frac{17}{12}2^nx)\|_{L^p}\\
\gtrsim &\|\frac{17}{12}\phi^2(x)\cos(\frac{17}{12}2^nx)\|_{L^p}-2^{-n}\rightarrow \frac{17}{12} \big(\frac{\int_0^{2\pi}|\cos x|^pdx}{2\pi}\big)^{\frac 1p}\|\phi^2(x)\|_{L^p},
\end{align*}
by the Riemann Theorem.
Taking the above estimates into (\ref{eq3.52}) yields
\begin{align*}
\liminf_{n\rightarrow \infty}\|\rho_n^2-\rho_n^1\|_{B^{s-1}_{p,r}}\gtrsim t\quad\text{for} \ t \ \text{small enough}.
\end{align*}
Similarly, we have
\begin{align*}
\liminf_{n\rightarrow \infty}\|u_n^2-u_n^1\|_{B^s_{p,r}}\gtrsim t\quad\text{for} \ t \ \text{small enough}.
\end{align*}
This completes the proof of Theorem \ref{the1.1}.
\section*{Acknowledgments}
This work is supported by the National Natural Science Foundation of China (Grant No.12001163).
| 08609f92f46cd71e119c64a9b541087d5ed5b519 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\vspace{-0.3cm}
One of the main goals in modern heavy-ion physics is the study of QCD matter at extreme conditions of high temperature and/or high net-baryon number density ($\epsilon \approx$~1~GeV/fm$^3$). Under these conditions, which can be generated in ultra-relativistic heavy-ion collisions at laboratories using accelerators such as the Large Hadron Collider (LHC) at CERN, a colour-deconfined state of strongly-interacting matter, the quark--gluon plasma (QGP)~\cite{Shuryak:1977ut} is created. At RHIC and LHC energies, QCD matter is studied at nearly vanishing net-baryon density (baryochemical potential $\mu_{\rm B} \approx 0$), for which lattice QCD calculations predict a smooth crossover between hadronic matter and a QGP, where chiral symmetry is restored, at a pseudo-critical temperature $T_{\rm C} \approx 156.5 \pm 1.5$~MeV~\cite{Bazavov:2018mes}. Furthermore, the QGP is assumed to have existed up to a few microseconds after the Big Bang~\cite{Boyanovsky:2006bf} and might exist in the core of neutron stars~\cite{Alford:2013pma}. Therefore, studies of the QGP will also provide input for e.g.\ theoretical models of the early universe and thus address some of the fundamental questions of mankind. \\
An ultra-relativistic heavy-ion collision can be subdivided into the following stages: i) initial collisions, which occur during the passing of the Lorentz-contracted nuclei; ii) creation and thermalisation of the QGP during 0.1--1~fm/$c$; iii) expansion, collective motion, and cooling of the QGP; iv) hadronisation once the temperature drops below $T_{\rm C}$; v) after 5--10 fm/$c$~\cite{Braun-Munzinger:2015hba} chemical freeze-out (inelastic collisions stop and hadron yields become fixed) and vi) kinetic-freeze out occur (elastic collisions stop and momenta no longer change). The final-state particles or their decay particles can then be measured in a detector to infer the characteristics of the QGP. It is not possible to detect the QGP directly, because neither quarks nor gluons are colour-neutral objects. However, different experimental measurements are more sensitive to the various stages of the collision, and by comparing the results with theoretical calculations, various properties of the QGP can be determined. \\
In order to draw firm conclusions on effects induced by the hot medium (‘final-state effects’), the same measurements are performed in pp collisions as a baseline and in proton--nucleus (pA) collisions to study initial-state effects in cold nuclear matter (CNM), such as the modifications of the parton distribution functions (PDF) in the nucleus with respect to that in the free proton. However, it has turned out that these measurements have provided in the last years far more interesting results than being a pure reference (see below). \\
The effects of the hot medium can be quantified using the nuclear modification factor $R_{\rm AA}$, which is defined as the ratio of the $p_{\rm T}$ distributions measured in AA collisions with respect to that in pp collisions:
$R_{\rm AA} = \frac{1}{\langle N_{\rm coll} \rangle}\frac{{\rm d}N_{\rm AA}/{\rm d}p_{\rm T}}{{\rm d}N_{\rm pp}/{\rm d}p_{\rm T}}$ ,
where ${\rm d}N_{\rm AA}/{\rm d}p_{\rm T}$ and ${\rm d}N_{\rm pp}/{\rm d}p_{\rm T}$ are the $p_{\rm T}$-differential yields of a given particle species in AA and pp collisions, respectively, and $\langle N_{\rm coll} \rangle$ is the average number of binary nucleon--nucleon collisions in the overlap region of the colliding nuclei.
If in-medium effects are absent and cold nuclear matter effects are small, the ratio $R_{\rm AA}$ is unity in the region where particle production from hard scattering processes dominates ($p_{\rm T} \gtrsim$~2~GeV/$c$).
\vspace{-0.2cm}
\section{Initial State}
\vspace{-0.2cm}
Photons with high transverse momentum as well as W and Z bosons (due to their large mass, they are produced before the QGP is created), are penetrating probes and provide information about the initial state and the nuclear parton distribution functions. Due to the high collision energy at the LHC, these are copiously produced. The $R_{\rm AA}$ of isolated photons in AA collisions is consistent with unity~\cite{Sirunyan:2020ycu} (see Fig.~\ref{Fig::photons}), since they do not carry colour charge and thus do not interact strongly. The data are better described by calculations with EPPS16 and nCTEQ15 nuclear PDFs than by calculations without nuclear modifications. Measurements by the ATLAS collaboration in p--Pb collisions at forward, mid- and backward rapidity are described by calculations that include modestly modified nPDFs~\cite{Aaboud:2019tab}. The measurement of Z bosons by the ALICE collaboration in Pb--Pb collisions shown in Fig.~\ref{Fig::photons} exhibits a clear deviation from the expectation using free-nucleon PDFs as a function of rapidity, yielding a 3.4$\sigma$ effect integrated over rapidity~\cite{Acharya:2020puh}. Measurements of Z boson production in p--Pb collisions at forward and backward rapidity by the LHCb collaboration agree with calculations considering PDF sets with and without nuclear modification within current uncertainties~\cite{LHCb-CONF-2019-003}. These measurements can be used to better constrain initial state effects and nPDF global fits.
\begin{figure}[tbp]
\includegraphics[width=0.42\textwidth]{figures/CMS-HIN-18-016_Figure_006.pdf}
\hspace{1.0cm}
\includegraphics[width=0.46\textwidth]{figures/ZPbPb5tev_RaaVsRapidity-99740.pdf}
\caption{Nuclear modification factor of (left) isolated photons as a function of photon energy~\cite{Sirunyan:2020ycu} and (right) Z bosons vs rapidity~\cite{Acharya:2020puh} in Pb--Pb collisions. \vspace{-0.2cm}} \label{Fig::photons}
\end{figure}
\vspace{-0.2cm}
\section{Soft Probes}
\vspace{-0.3cm}
The particle yields of light-flavour hadrons and nuclei from pions up to (anti-)hypertriton and (anti-)$^4$He, as measured by the ALICE collaboration in central Pb--Pb collisions at $\sqrt{s_{\rm NN}} =$ 2.76~TeV, span nine orders of magnitude and are well described by a thermal equilibrium model~\cite{Andronic:2017pug} underlining the thermal nature of particle production in ultra-relativistic heavy-ion collisions. The grand canonical thermal fit describes the particle abundances with a universal temperature $T_{\rm chem} = 156 \pm 1.5$~MeV~\cite{Andronic:2017pug}, which coincides with the pseudo-critical temperature of the chiral crossover transition from lattice QCD calculations (see above).
The azimuthal distributions of particle production with respect to the measured symmetry plane orientations can be expanded in terms of Fourier coefficients $v_{\rm n}$. The `elliptic flow' coefficient $v_{2}$ is shown in Fig.~\ref{Fig::flow1} for different particle species in central and semi-central collisions. For all shown particle species large anisotropies due to collective effects are visible. For semi-central collisions the coefficients are larger due to the initial conditions of the collision. Moreover, at low momenta a mass ordering effect is visible characteristic for hydrodynamic collective expansion. Both deuterons and $^3$He are loosely bound composite objects and the data can also be used to understand the hadronisation mechanism. The $^3$He data are well described by a model that includes coalescence of nucleons, a hydrodynamic evolution of the fireball and a hadronic afterburner~\cite{Zhao:2018lyf}. The calculations also reveal that the produced matter has a small shear viscosity to entropy density ratio, i.e.\ the created fireball is an almost perfect liquid. \\
\begin{wrapfigure}{l}{0.65\textwidth}
\vspace{-0.8cm}
\begin{flushleft}
\includegraphics[width=0.65\textwidth]{figures/flowPID.png}
\caption{Elliptic flow as a function of $p_{\rm T}$ for various particle species~\cite{Acharya:2019ttn,Acharya:2020lus}.} \label{Fig::flow1}
\end{flushleft}
\vspace{-0.6cm}
\end{wrapfigure}
The $v_{2}$ coefficient is shown in Fig.~\ref{Fig::flowsystems} as a function of charged particle multiplicity in pp, p--Pb, Xe--Xe and Pb--Pb collisions. In the large collision systems, at large multiplicities, due to the initial geometry and strong interactions strong signs of collective effects are seen, which are well described by hydrodynamic calculations. In small collision systems (pp, p--Pb, peripheral Pb--Pb), the sizeable coefficients, which are similar in the different systems at the same multiplicity, could be indicative of collective effects. However, the data are neither described by hydrodynamic calculations nor by PYTHIA8 calculations, although the latter show a trend similar to data. Several detailed studies are ongoing to investigate the long-range correlations seen in the small systems and the continuous evolution of strangeness enhancement across collision systems~\cite{ALICE:2017jyt}. The ATLAS collaboration e.g.\ measured in pp collisions the elliptic flow as a function of charged particle multiplicity for inclusive and Z-tagged collisions (the impact parameter of the collision may be on average smaller)~\cite{Aaboud:2019mcw}, where neither a significant multiplicity dependence nor an influence due to the presence of the hard-scattering process is observed. In contrast to high-multiplicity (HM) pp collisions~\cite{Khachatryan:2015lva,Acharya:2019vdf}, no indication of collective behaviour is observed in elementary HM ep and e$^+$e$^-$ collisions as measured by the ZEUS~\cite{ZEUS:2019jya}, shown in Fig.~\ref{Fig::flowsystems}, and BELLE collaborations~\cite{Abdesselam:2020snb}. So far no signs of energy loss have been seen in small collision systems (see below). For further investigations, future measurements with large data samples and very HM pp collisions~\cite{ALICE-PUBLIC-2020-005} and collisions of light ions~\cite{Huss:2020whe} are foreseen.
\begin{figure}[btp]
\includegraphics[trim=0cm 0cm 0cm 0.2cm, clip=true,width=0.6\textwidth]{figures/flowdiffsystems.pdf}
\hspace{0.3cm}
\includegraphics[trim=0cm 0cm 0cm 0.2cm, clip=true,width=0.36\textwidth]{figures/Zeus.png}
\caption{(left) Elliptic flow coefficient $v_{2}$ as a function of charged particle multiplicity in pp, p--Pb, Xe--Xe and Pb--Pb collisions~\cite{Acharya:2019vdf}. (right) Two-particle correlation $C(\Delta \eta, \Delta\varphi)$ for HM ep collisions~\cite{ZEUS:2019jya}.} \label{Fig::flowsystems}
\end{figure}
\vspace{-0.2cm}
\section{Hard Probes (parton energy loss, jets, quarkonia)}
\vspace{-0.3cm}
\begin{figure}[htbp]
\includegraphics[width=0.5\textwidth]{figures/2020-05-25-2020-05-25-Pi0_RpPb_Hadrons_PbPb.pdf}
\hspace{0.4cm}
\includegraphics[width=0.43\textwidth]{figures/CMSjet.png}
\caption{(left) Nuclear modification factor of $\pi^{0}$ mesons and charged hadrons in p--Pb and Pb--Pb collisions. (right) Ratio of Z boson-tagged charged hadron spectra in Pb--Pb and pp collisions~\cite{CMS:2020rxf}. } \label{Fig::RAA}
\end{figure}
In the large collision systems, such as in central Pb--Pb collisions, the yield of charged hadrons (see Fig.~\ref{Fig::RAA}) and of jets with large transverse momenta up to $p_{\rm T}~=$~900 GeV/$c$, as measured by the ATLAS collaboration~\cite{Aaboud:2018twu}, is strongly suppressed compared to pp collisions scaled by the number of binary collisions. The comparison with results from p--Pb collisions, where this effect is absent as seen in Fig.~\ref{Fig::RAA}, and theoretical calculations show that the suppression in Pb--Pb collisions is due to a final state effect, where the parton traversing the medium looses energy due to radiative and collisional energy loss. The large data samples at the LHC allow for many differential studies. For example, the CMS collaboration measured for the first time Z-tagged charged particle yields in Pb--Pb collisions, which are, as shown in Fig.~\ref{Fig::RAA}, compared to the expected yields in pp collisions strongly suppressed at high $p_{\rm T}$, but increase towards low $p_{\rm T}$~\cite{CMS:2020rxf}. The jet substructure was studied by the ALICE collaboration~\cite{ALICE:pubnotejets} as a function of the jet resolution parameter in Pb--Pb and pp collision using the grooming technique, where the large angle soft gluon radiation is removed to identify a hard splitting within the jet. The fully corrected data show that the jet substructure is modified by the hot medium, the selected hard splitting being narrower, which is described by e.g.\ models that include incoherent energy loss~\cite{ALICE:pubnotejets}.
\begin{figure}[tbp]
\includegraphics[width=0.43\textwidth]{figures/CMS_RAA.png}
\hspace{0.5cm}
\includegraphics[trim=0cm 0cm 1.0cm 1.0cm, clip=true,width=0.48\textwidth]{figures/flow_charm_beauty.eps}
\caption{Nuclear modification factor $R_{\rm AA}$
of charged hadrons (mainly pions), prompt D mesons, B$^{\pm}$, non-prompt J/$\psi$ as a function of $p_{\rm T}$ in Pb--Pb collisions at $\sqrt{s_{\mathrm{NN}}}=5.02$~TeV~\cite{Sirunyan:2017xss}. Elliptic flow $v_{\rm 2}$ of pions, prompt D mesons and electrons from charm- and beauty-hadron decays in the same collision system~\cite{Acharya:2020pnh,Acharya:2020qvg,Aad:2020grf}.\vspace{-0.5cm}
} \label{Fig::RAAandflow}
\end{figure}
Heavy quarks (charm and beauty) are created in initial hard scattering processes and thus experience the whole spatial and temporal evolution of a heavy-ion collision, thus providing essential information on the interactions of partons with the hot medium and its properties. Gluons are argued to loose more energy than quarks due to their stronger colour coupling to the medium. In addition, several mass-dependent effects (dead-cone effect and mass dependent spatial diffusion coefficient) are expected to influence the amount of energy loss of the heavy quarks. Thus a hierarchy of the parton energy loss is predicted, namely $\Delta E_{\rm gluon} > \Delta E_{\rm charm} > \Delta E_{\rm beauty}$, resulting in an expected ordering: $R_{\rm AA}^{\pi} \le R_{\rm AA}^{\rm D} \le R_{\rm AA}^{\rm B}$ for pions (mostly originating from gluons), D, and B mesons. Measurements (e.g. those by the CMS collaboration in Fig.~\ref{Fig::RAAandflow}) show that $R_{\rm AA}^{\rm hadron} \approx R_{\rm AA}^{\rm D}$ for $p_{\rm T} \geq 5$ GeV/$c$. The $R_{\rm AA}$ is not only influenced by the energy loss, but also by the parton $p_{\rm T}$ spectrum and the fragmentation into hadrons. Theoretical calculations including these effects described the data well~\cite{Djordjevic:2013pba}. However, comparing with J/$\psi$ mesons from B-hadron decays and new measurements of prompt and non-prompt D mesons by the ALICE collaboration a quark mass depend energy loss at intermediate $p_{\rm T}$ is indeed observed. At high transverse momentum ($p_{\rm T}$ $\gtrsim$ 20~GeV/$c$), all hadrons including prompt J/$\psi$~\cite{Khachatryan:2016ypw} show within uncertainties surprisingly the same values of $R_{\rm AA}$. Differential new complimentary measurements by the ALICE and ATLAS collaborations (see Fig.~\ref{Fig::RAAandflow}) show that heavy quarks participate in the collective motion~\cite{Acharya:2020pnh,Acharya:2020qvg,Aad:2020grf}. Also here, at small $p_{\rm T}$, a mass-ordering effect becomes visible with electrons from beauty-hadron decays showing a positive $v_{2}$ with 3.8$\sigma$ significance indicating a partial thermalisation of beauty quarks in comparison to a large degree or complete thermalisation of charm quarks. At intermediate $p_{\rm T}$, the coefficients are similar for pions and D mesons, maybe due to a coalesence effect of a light and a charm quark. At high $p_{\rm T}$, the distributions merge suggesting that the path-length dependent energy loss dominates.
\begin{wrapfigure}{l}{0.48\textwidth}
\vspace{-0.4cm}
\includegraphics[width=0.5\textwidth]{figures/lambda_c_multiexp.eps}\vspace{-0.2cm}
\caption{$\Lambda_{\rm c}$/D$^0$ ratio as a function of $p_{\rm T}$ in pp and heavy-ion collisions~\cite{Acharya:2020uqi,Sirunyan:2019fnc,Adam:2019hpq}. } \label{Fig::Lambdac}
\vspace{-0.6cm}
\end{wrapfigure}
To better understand the aforementioned mechanisms, yields of hadrons containing heavy quarks are also studied. The ALICE and the STAR collaboration measured the ratio of strange to non-strange D mesons in Pb--Pb collisions at $\sqrt{s_{\mathrm{NN}}}=5.02$~TeV and Au--Au collisions at $\sqrt{s_{\mathrm{NN}}}=0.2$~TeV. The ratios are higher by about a factor 2 in central collisions than the ones measured in pp collisions or PYTHIA calculations for the STAR result and well described by the statistical hadronisation model. The CMS collaboration finds the same behaviour in the beauty sector (B$_{\rm s}$/B$^+$). The data hint at an enhanced production of strange hadrons due to the strangeness-rich QGP. Thanks to the large available data samples, baryon-to-meson ratios can also be studied in the charm sector, which might provide insight into the hadronisation mechanism of charm baryons. The ratio of $\Lambda_{\rm c}$/D$^0$ in pp collisions shown in Fig.~\ref{Fig::Lambdac} is much larger compared to the one in e$^+$e$^-$ collisions ($\Lambda_{\rm c}$/D$^0$ $\approx 0.11$) and shows an even larger value for heavy-ion collisions as measured by the ALICE, CMS and STAR collaborations~\cite{Acharya:2020uqi,Sirunyan:2019fnc,Adam:2019hpq}. The measurements are qualitatively described by the statistical hadronisation model~\cite{Andronic:2017pug} and the Catania model with hadronisation via a combination of fragmentation and coalescence~\cite{Plumari:2017ntm}.\\
Quarkonia were proposed as a signature of the deconfinement in the QGP. It was predicted that quarkonium production would be suppressed due to a screening of the heavy-quark potential in the colour-deconfined medium~\cite{Matsui:1986dk}. With increasing temperature of the hot medium, the quarkonium states with decreasing radius were predicted to melt subsequently, thus providing a sort of `thermometer' of the QGP. Another competing mechanism is the production via (re)generation during the QGP phase or at hadronisation which depends e.g.\ for the J/$\psi$ on the $c\overline{c}$ production cross section ($ N_{\rm{J}/\psi} \propto N^2_{c}$)~\cite{Grandchamp:2001pf,BraunMunzinger:2000px}. The LHC data allow the production mechanism of J/$\psi$ to be elucidated. While the $R_{\rm AA}$ of J/$\psi$ decreases with increasing charged-particle density in Au--Au collisions at $\sqrt{s_{\mathrm{NN}}}=0.2$~TeV (see Fig.~\ref{Fig::Quarkonia}) the corresponding measurement at LHC energies shows at LHC full energy no suppression which is well in line with the regeneration scenario due to the large $c\overline{c}$ cross section at the LHC and is a clear signature of deconfinement. The differential studies vs $p_{\rm T}$ show an increase of the $R_{\rm AA}$ with decreasing $p_{\rm T}$, which is well described by theoretical transport models and the statistical hadronisation model that include the regeneration mechanism that dominates at low $p_{\rm T}$. The observed strong sign of collective effects of J/$\psi$ mesons~\cite{Acharya:2020jil}, including also the observed positive $v_{2}$ coefficient with a significance of 2.5$\sigma$ at midrapidity, confirms the regeneration scenario.\\
On the other hand, the suppression mechanism dominates for the bottomonium family as observed by an increasing suppression of $\Upsilon$(1S), $\Upsilon$(2S), and $\Upsilon$(3S) in measurements by the ALICE, ATLAS, CMS and STAR collaborations~\cite{ATLAS-CONF-2019-054,Sirunyan:2018nsz,Acharya:2020kls,Adamczyk:2016dzv}, see e.g.\ Fig.~\ref{Fig::Quarkonia}. The data are described by theoretical models that include the suppression mechanism and feedown from higher lying resonances.
\begin{figure}[htbp]
\includegraphics[width=0.5\textwidth]{figures/2020-07-22-RAA_Npart_comp_Ele_dNchdeta_5020_2018.pdf}
\hspace{0.6cm}
\includegraphics[width=0.38\textwidth]{figures/fig_09a.pdf}
\caption{$R_{\rm AA}$ of J/$\psi$ (left) and $\Upsilon$ states (right)~\cite{ATLAS-CONF-2019-054} as a function of charged-particle density and average number of participants, respectively.}\label{Fig::Quarkonia}
\end{figure}
\vspace{-0.2cm}
\section{Rare probes}
\vspace{-0.3cm}
The large data samples recorded at the LHC, allow for new measurements of very rare probes. Studies of the $\chi_{\rm c1}$(3872) in heavy-ion collisions could provide additional insight into the production mechanism and the nature of this exotic particle. Figure~\ref{Fig::exotic} presents the ratio of the yield of $\chi_{\rm c1}$(3872) to $\psi$(2S) in pp and Pb--Pb collisions measured by the ATLAS and CMS collaborations. The ratio is larger in the heavy-ion collision, also because the CMS collaboration reported a significant suppression of $\psi$(2S) in Pb--Pb collisions. In future, it will be important to extend the measurement to low $p_{\rm T}$, where the regeneration mechanism dominates as seen for the J/$\psi$ meson (see above). \\
Evidence of top quark production in Pb--Pb collisions was for the first time reported by the CMS collaboration~\cite{CMS-PAS-HIN-19-001}. The measurement is performed via the semileptonic decay channel of W bosons with and without the presence of b-quark jets. The observed cross section $\sigma_{\rm t\overline{t}}$ as shown in Fig.~\ref{Fig::exotic} is compatible, but somewhat lower than pQCD calculations and the result from pp collisions correspondingly scaled. The measurement of top quark production in heavy-ion collisions might have the potential to probe the time structure of the QGP~\cite{Apolinario:2017sob} and thus will be an important study for the future, maybe for the LHC Run 5.
\begin{figure}[htbp]
\includegraphics[trim=0cm 0cm 0cm 0.9cm, clip=true,width=0.45\textwidth]{figures/Figure_002.pdf}
\hspace{0.5cm}
\includegraphics[width=0.5\textwidth]{figures/CMS-HIN-19-001_Figure_004.pdf}
\caption{(left) Ratio of the fully-corrected yields of $\chi_{\rm c1}$(3872) and $\psi$(2S) in pp and Pb--Pb collisions~\cite{CMS-PAS-HIN-19-005} and (right) t$\overline{\rm t}$ cross sections in pp and Pb--Pb collisions~\cite{CMS-PAS-HIN-19-001}.} \label{Fig::exotic}
\vspace{-0.5cm}
\end{figure}
\vspace{-0.2cm}
\section{Impact beyond the physics of heavy-ion collisions}
\vspace{-0.3cm}
\begin{wrapfigure}{l}{0.5\textwidth}
\includegraphics[width=0.49\textwidth]{figures/dcross.png}
\caption{Inelastic antideuteron cross section vs. momentum at which the interaction occurs~\cite{Acharya:2020cee}.}\label{figure:dcross}
\vspace{-0.6cm}
\end{wrapfigure}
Results from high-energy heavy-ion collisions also have impact on other fields of physics. Measurements of the cross sections of antinuclei provide important input for dark matter searches. The precise $p_{\rm T}$-differential antiproton cross section measurements in p--He collisions~\cite{Aaij:2018svt}, a measurement in fixed-target mode performed by the LHCb collaboration (He is injected into the beam pipe close to the interaction point), help to constrain theoretical models. The ALICE collaboration has measured for the first time the inelastic antideuteron--nucleus cross section at small momenta (see Fig.~\ref{figure:dcross}) using the detector as an absorber~\cite{Acharya:2020cee}. These kind of measurements will be extended to $^3{\overline{\rm He}}$ and $^4{\overline{\rm He}}$ in LHC Run~3 and 4.
\vspace{-0.4cm}
\section{Future of heavy-ion physics}
\vspace{-0.3cm}
The many new important results from the LHC and RHIC lead to an improved understanding of the interaction of partons with the hot medium, its properties and initial-state effects. They also constrain the nPDF global fits and provide input for other fields of physics. \\
With the ongoing upgrades (ALICE, ATLAS, CMS, LHCb) for LHC Run~3 and 4 an unprecedented level of precision will be reached and the search for rare probes significantly extended. Further, the European strategy for particle physics encourages the heavy-ion programme at CERN in the HL-LHC era. There is a plan for a next-generation LHC heavy-ion experiment: ALICE 3 constructed almost entirely from silicon~\cite{Adamova:2019vkf} for LHC Run~5 and beyond. In addition, projects such as sPHENIX and STAR at RHIC (incl.\ Beam Energy Scan), the fixed-target programme at CERN, facilities in the high net-baryon number density frontier (NICA, FAIR, ...) and the electron-ion collisions at EIC will provide further insight into different kinematic regimes of the QCD phase diagram.
\vspace{-0.4cm}
\begin{spacing}{0.9}
\setlength{\bibsep}{0.35\baselineskip}
\begin{footnotesize}
\bibliographystyle{JHEP}
| 3467eae65d37dbcaffab4e59213ce216011597e4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Galaxies are usually classified into ellipticals and spirals according to their morphology. Elliptical galaxies, by definition, have regular elliptical shapes of stellar distribution. They have very little ongoing star formation, and thus they are red in color and old in stellar age. In contrast, spiral galaxies have disk shapes of stellar distribution,
\Authorfootnote
\noindent
though a significant fraction of them possess bulges at the centers. There is usually ongoing star formation in spiral galaxies, so they look blue in color and have young stellar populations \cite{ref1}.
These properties of galaxies can be well understood in the standard paradigm of hierarchical structure formation~\cite{ref2,ref3,ref4}. In the theory, cosmic structures grow by gravitational instability from the initial tiny quantum fluctuations generated during the inflationary epoch. Dark Matter (DM) halos, which are defined as the dark matter objects with density about $200$ times of the mean density of the Universe, are formed through accreting surrounding smaller DM halos and diffuse matter including DM and gas. The gas, when accreted into a halo, is heated by shocks \cite{shock}. The hot gas cools by radiating its energy, and the cold gas spirals into the halo center to form a disk (spiral) galaxy. With
the growth of a halo, surrounding galaxies can also be accreted into the halo and become its satellites. Because of dynamical friction, some satellite galaxies, especially relatively massive ones, spiral into the center and coalesce with the centra galaxy~\cite{ref5,ref6}. A merger of a central disk galaxy with a satellite of comparable mass (say, a mass ratio $\geq 0.2$) may significantly change its internal structure, either producing an elliptical galaxy or a disk galaxy with a significant bulge\cite{ref7,ref8,ref9}. Based on the observed relation between black hole mass and bulge mass \cite{ref10}, we expect there are supermassive black holes (SMBHs) within elliptical galaxies or bulges. The strong energy output and/or material outflow produced by an SMBH can heat and/or blow out the surrounding cold gas, thus suppressing the reservoir from which stars form and making the host galaxy look red. Current hydrodynamical simulations of galaxies that have incorporated these (and other) physical processes have successfully reproduced a wide range of observed properties of galaxies including the stellar mass, morphology, color, and stellar age~\cite{ref11,ref12}.
\begin{figure}[H]
\centering
\includegraphics[scale=0.28]{sersic2d.pdf}
\caption{HSC $z$-band image (left panel), CLAUDS $u^{*}$-band image(middle panel) and $u^{*}-z$ color magnitude map (right panel) of the galaxy. Physical length of 15 $kpc$ corresponds to 1.977 $arcsec$ and the FWHM of the seeing in the $z$ and $u^*$ images is 0.822 $arcsec$ and 1.046 $arcsec$.}\label{fig:f1}
\end{figure}
Massive galaxies with stellar mass larger than $10^{11.0}M_{\odot}$ are expected to form in DM halos of group and cluster mass~\cite{ref13}. Because of frequent mergers happening in such rich environments, one expects that most such massive galaxies are red ellipticals with significant amounts of bulge. Only in a rare case with little merger happened in the past, because no bulge has formed, thus no SMBH has formed. The central galaxy may keep forming stars, thus becoming a blue disk galaxy. In any case, one would not expect to form a central red disk galaxy in a massive DM halo, as there is no mechanism to stop gas cooling and star formation if no bulge (thus no SMBH) has formed.
\begin{table}[H]
\centering
\caption{Physical properties of SDSS 022248.95-045914.9.}
\begin{tabular}{cc}
\toprule
Properties & value \\
\midrule
RA ($deg$) & $35.703888$ \\
DEC ($deg$) & $-4.987444$ \\
Redshift & $0.76$ \\
rest-frame $NUV-r$ ($mag$) & $5.3$ \\
sSFR ($yr^{-1}$) & $10^{-11.38}$ \\
$M_*$ ($M_{\odot}$) (SED) & $10^{11.64 \pm 0.10}$ \\
$M_*$ ($M_{\odot}$) (PCA\textsuperscript{a}) & $10^{11.67 \pm 0.17}$ \\
S\'ersic index & $1.22 \pm 0.06$ \\
axis ratio $b/a$ & $0.85$ \\
half-light radius $R_e$ (kpc) & $8.81$ \\
$NUV$ ($mag$)& $26.9$ \\
$u$ ($mag$) & $24.7$ \\
$g$ ($mag$) & $22.8$ \\
$r$ ($mag$) & $21.3$ \\
$i$ ($mag$) & $20.0$ \\
$z$ ($mag$) & $19.5$ \\
$K_s$ ($mag$) & $18.4$\\
\bottomrule
\multicolumn{2}{l}{\textsuperscript{a}\footnotesize{From Wisconsin group \cite{ref16}}}
\end{tabular}
\label{tab:t1}
\end{table}
\section{Properties of the galaxy}
In this paper, we report a massive red central disk galaxy SDSS J022248.95-045914.9 with stellar mass reaching $10^{11.6}M_{\odot}$. The details of the galaxy are shown in Table \ref{tab:t1}. The $z$-band image from Deep field of HyperSuprime-Cam Subaru Strategic Program (HSC-SSP) \cite{ref18}, $u^{*}$-band image from CFHT Large Area U-band Deep Survey (CLAUDS) \cite{clauds}, and $u^{*}-z$ color map of the galaxy are shown in Figure \ref{fig:f1}. The galaxy was discovered during our study of physical properties of massive galaxies in the CMASS sample of the Baryon Oscillation Spectroscopic Survey (BOSS) survey \cite{ref14} in the XMM-LSS region \cite{ref33}. We perform Spectral Energy Distribution (SED) fitting on multi-band data of the galaxy from $FUV$ to $K_s$ collected by Moutard et al. \cite{ref15} using \textsc{Le PHARE} \cite{le1,le2}. We use the stellar population synthesis models of BC03 \cite{bc03} with the Chabrier \cite{imf} Initial Mass Function (IMF) and a delayed star formation history. We get the stellar mass $M_{\ast}=10^{11.64}M_{\odot}$ and the specific star formation rate (sSFR) $10^{-11.38}\ {\rm yr^{-1}}$ and the rest-frame $NUV-r$ color is $5.3$ based on the SED fitting. As the BOSS survey has taken a spectrum for the galaxy, we show it in Figure \ref{fig:f2}. From the figure, we can easily see the best fitting spectrum is typical for an old stellar population: there is no visible emission line and the $D4000$ break is about $1.9$. The stellar mass released by the Wisconsin group \cite{ref16} using Principal Component Analysis method based on the spectrum is $10^{11.67\pm 0.17}M_{\odot}$, which is well consistent with the stellar mass obtained by our SED fitting. From both the photometry and the spectrum, we can conclude that the galaxy is very massive and its star formation has been quenched. These are the typical features of a massive elliptical galaxy.
\begin{figure}[H]
\centering
\includegraphics[scale=0.34]{spec.pdf}
\caption{Blue line shows the original spectrum from BOSS and the red line is the best fitting model. Typical emission lines are shown via vertical black lines. Wavelength is in rest-frame.}\label{fig:f2}
\end{figure}
However, our analysis of the morphology for the galaxy surprisingly shows that it is a disk galaxy. The radial luminosity distribution of a galaxy can be described by the S\'ersic form \cite{ref17}:
\begin{equation}
\label{eq:1}
I(R) = I_e\exp\bigg\{-b_n\bigg[\bigg(\frac{R}{R_e}\bigg)^{1/n}-1\bigg]\bigg\},
\end{equation}
where $b_n$ satisfies:
\begin{equation}
\label{eq:2}
\gamma(2n; b_n)=\frac{1}{2}\Gamma(2n).
\end{equation}
$I_e$ is the intensity at the half-light radius $R_e$, and $n$, called the S\'ersic index, describes the compactness of a luminosity profile. $\Gamma$ and $\gamma$ are respectively the Gamma function and lower incomplete Gamma function. Generally speaking, disk galaxies have exponential profiles with $n = 1$, and ellipticals follow the de Vaucouleurs profiles with $n = 4$. We take the $z$ band photometric image from the public data release $2$ of HyperSuprime-Cam Subaru Strategic Program (HSC-SSP) \cite{ref18}. We fit the image with the S\'ersic form in 2D using \textsc{galfit} \cite{ref19} with the point spread function (PSF) taken into account. The best-fitting S\'ersic profile of the galaxy is shown in Figure \ref{fig:f3}, which describes the observed radial luminosity profile very well. The S\'ersic index is $n = 1.22 \pm 0.06$, which indicates the galaxy is a disk galaxy. This is consistent with the visual appearance, which does not show any significant bulge component. To directly measure the bulge/disk ratio, we also performed two component fitting on the galaxy, using an exponential plus de Vaucouleurs profile. The fitting results do not indicate any bulge component, and we can set an upper limit that the bulge contains less than $1/15$ of the stellar light of the galaxy. In fact, the luminosity profile is indeed well described by the exponential profile with $n=1$ (Green dot line in Figure \ref{fig:f3}).
\begin{figure}[H]
\centering
\includegraphics[scale=0.3]{sersic.pdf}
\caption{Radial profile of the $z$ band image and models of the galaxy. Red dots with error bars are the data from HSC $z$ band. Blue solid line shows the profile of the best fitting S\'ersic model. Parameters from the best fitting are: Integrated magnitude: $z=19.64 \pm 0.01\ {\rm mag}$; effective radius: $R_e=1.035 \pm 0.020\ {\rm arcsec}$; Axis ratio: $b/a=0.850 \pm 0.010$; Position Angle: $PA= -78.46 \pm 3.87\ {\rm deg}$; S\'ersic index: $n= 1.22 \pm 0.06$. Green dot line shows the best fitting exponential function.}\label{fig:f3}.
\end{figure}
As we will discuss below, the environment effects, such as the ram pressure and/or tidal stripping may help disk galaxies quench their star formation. It is therefore vital to check whether the galaxy is a central or satellite galaxy. We use the photometric catalog of the second public data release (PDR-2) of the VIMOS Public Extragalactic Redshift Survey (VIPERS) \cite{ref15}. This catalog contains all sources brighter than $22.5$ $mag$ in the $i$ band in the W1 field which covers the galaxy we study. The $K_s$ band magnitudes from the VIDEO observation \cite{ref20} are also available in the region around the galaxy. We consider a field of view of comoving radius $1\ {\rm Mpc}$ around the galaxy, which corresponds to $\theta=0.02081\ deg$ in the sky using the cosmological parameters from the Planck satellite \cite{ref21}. We use the photometric redshifts $z_{ph}$ to search for the neighbors. Since the error of $z_{ph}$ is typically about $0.05$ for galaxies with $i<22.5$ $mag$ and $K_s<22.0$ $mag$ \cite{ref15}, we plot all the galaxies of $K_s\leq 21.0$ $mag$ within a cylinder of radius $\theta$ and length $\Delta z_{ph} =0.4$ centered on redshift $z_s=0.76$ of the red disk galaxy. As shown in Figure \ref{fig:f4}, there are $8$ neighbor galaxies, all of which are at least $1.2$ mag fainter than our red disk ($K_s=18.4$ $mag$), which confirms that the red disk galaxy is a central galaxy. Since the $K_s$ luminosity is approximately proportional to the stellar mass, we also find that the existence of $8$ satellites with the stellar mass larger than $1/10$ of the central red disk??s stellar mass is fully consistent with the observed conditional stellar mass function \cite{ref22}.
\begin{figure}[H]
\centering
\includegraphics[scale=0.45]{check_central.pdf}
\caption{Galaxies of $K_s<21$ $mag$ within $|\Delta z_{ph}| = 0.2$ in the sky area of a comoving radius $1\ {\rm Mpc}$ centered on the red disk galaxy. Colors of the dots reflect the $K_s$ band magnitude of the neighbor galaxies.}\label{fig:f4}
\end{figure}
Thus, the galaxy we find is very likely a massive central red disk. But it is very challenging to explain its formation in current theories of galaxy formation. According to the halo mass and stellar mass relation \cite{ref13}, we expect the galaxy is at the center of the host halo of mass about $5\times 10^{14}M_{\odot}$. As described in the beginning, such a galaxy should find it hard to maintain its disk structure since frequent merging is inevitable with other spiraling-in galaxies in the rich environment (cluster of galaxies). Even in very rare of a galaxy with very little merging in its recent past, one would expect ongoing star formation activities unless there is enough feedback from the central SMBH. However, as we have not detected the bulge component at the center of the red disk, we would expect that the feedback from the SMBH is not an efficient way to quench the whole galaxy. This qualitative reasoning is verified by state-of-the-art hydrodynamical simulations that implement the physical processes we have described. We have searched for massive red disks ($NUV-r>4.5$ or sSFR $<10^{-11}\ {\rm yr^{-1}}$) of $M_{\ast} >10^{11.3}M_{\odot}$ and $n<1.5$ (in rest-frame $r$ band) in the Illustris-1 simulation \cite{ref23,ref24} at $z_s=0.5$, but we found a null result. To test the sensitivity to the details of simulations, we have also inspected the Illustris TNG100-1 \cite{ref25} simulation at $z_s = 0$ \footnote{Because the color of galaxies is redder and the merging is less frequent in a fixed time period at $z_s=0$ than $z_s=0.7$, we would expect to form more red disk galaxies at $z_s=0$ than at an earlier epoch.}, and no massive red disk is found either.
\section{Discussion and conclusion}
It is interesting to point out that a few studies of red spiral galaxies have appeared recently in the literature. Red disks with smaller mass ($\sim10^{10.0}M_{\odot}$) are found at intermediate local densities such as the infall regions of clusters~\cite{ref26,ref27}. The tidal force and/or gas ram pressure in such intermediate density environments could be strong enough to destroy the cold gas reservoir, but still too weak to affect the disk structures. Statistical studies of nearby massive red disks ($10^{10.5}\sim10^{11.0}M_{\odot}$) show larger bulge/total mass ratios than their blue counterparts~\cite{ref28,ref29}. Various quenching mechanisms have been theoretically proposed to quench a massive galaxy without AGN feedback such as halo quenching \cite{shock} and angular-momentum quenching \cite{angular}, which should be investigated further. In other studies, relations beyond the standard scaling rules are proposed to explain the existence of massive red disks. For example, some studies show that massive disks may exist in smaller halos comparing to other galaxies with the same stellar mass \cite{ref30}. Gas is exhausted by star formation in these galaxies and no quenching mechanism is needed to explain their red color. Moreover, deviations from the BH-bulge mass and the DM-baryon mass relations have also been proposed \cite{ref31,ref32}. Compared to the red galaxies found in the above studies, the galaxy reported in this paper has the highest redshift and, more importantly, the largest mass, making it much harder to explain its quenching.
In conclusion, we find a giant red central disk. The existence of this galaxy challenges current theories of galaxy formation, as there is no known physical mechanism to stop its star formation. Since there are some rare cases whereby mergers, which may cause very low S\'ersic indices, can not be distinguished from imaging data \cite{ref33}, we plan to confirm the galaxy's disk morphology with kinematic data from Integral Field Unit (IFU) observations. Meanwhile, we will search the whole HSC wide field for massive red disks and employ statistical analysis to check whether their properties deviate from standard scaling rules or not.
\Acknowledgements{The work is supported by NSFC (11533006, 11621303, 11890691, 11933003, 11833005). We are grateful to Yong Shi, Xiaoyang Xia, Caina Hao, Yongzhong Qian, Yanmei Chen, and Yunchong Wang for their helpful discussions. This research uses data obtained through the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories, Chinese Academy of Sciences, and the Special Fund for Astronomy from the Ministry of Finance. This work is based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Science de l'Univers of the Centre National de la Recherche Scientifique (CNRS) of France, and the University of Hawaii. The work has made use of the public data of the Canada-France-Hawaii Telescope Legacy Survey, the Sloan Digital Sky Survey, the Hyper Suprime-Cam Subaru Strategic Program, and the VIMOS Public Extragalactic Redshift Survey.}
\InterestConflict{The authors declare that they have no conflict of interest.}
| 5bd862e4e1781cca7508c38548ce65e582f9ed13 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Proof of~\Cref{thm:lower-bound-intro}
In this section, we prove the following theorem:
\begin{theorem}[Tolerant testing of DTs $\Rightarrow$ Proper learning of DTs]
\label{thm:lower-bound}
Let $c > 0$ be an absolute constant and $\mathcal{A}$ be an algorithm with the following guarantee. Given query access to $f: \bn \to \bits$ and parameters $s \in \N $ and $\eps \in (0,1)$, the algorithm $\mathcal{A}$:
\begin{itemize}
\item[$\circ$] Accepts w.h.p.~if $f$ is $\eps$-close to a size-$s$ decision tree;
\item[$\circ$] Rejects w.h.p.~if $f$ is $(c\eps)$-far from all size-$s$ decision trees.
\end{itemize}
Then there is an algorithm $\mathcal{B}$ with the following guarantee. Given parameters $s' \in \N$ and $\eps' \in (0,1)$, and query access to a function $g : \bn\to\bits$ that is computed by a size-$s'$ decision tree, $\mathcal{B}$ makes $\poly(s', n, 1/\eps')$ calls to $\mathcal{A}$, each with parameters $s \leq s'$ and $\eps \ge \poly(1/s',\eps')$, and produces a decision tree which is $\eps'$-close to $g$ with high probability. Furthermore, the auxiliary computation that $g$ does takes time $\poly(n, s', 1/\eps')$.
\end{theorem}
\Cref{thm:lower-bound-intro} follows as a special case of \Cref{thm:lower-bound}.
We prove~\Cref{thm:lower-bound} in two steps:
\begin{enumerate}
\item A tolerant tester implies an algorithm for estimating the distance of any function to the class of size-$s$ decision trees. This is well known~\cite{PRR06} and applies to any function class, not just decision trees.
\item An algorithm for estimating distance to decision trees implies a proper learner for decision trees. Here, we take advantage of the structure of decision trees.
\end{enumerate}
For a function $g : \bn\to\bits$ and $s\in \N$, we write $\opt_s(g)$ to denote the distance of $g$ to the closest size-$s$ decision tree.
\begin{lemma}[Tolerant testing $\Rightarrow$ distance estimation~\cite{PRR06}]
\label{lem:distance-estimator}
Let $c$ and $\mathcal{A}$ be as in \Cref{thm:lower-bound}. There exists an estimator $\mathcal{E}$ with the following guarantee. Given query access to $g: \bn \to \bits$ and parameters $s' \in \N$ and $\gamma \in (0,1)$, the estimator $\mathcal{E}$ makes $c/\gamma$ calls to $\mathcal{A}$ and returns an $\boldeta$ that with high probability satisfies
\[ \boldeta \leq \opt_s(g) \leq c \cdot \boldeta + \gamma.
\]
Furthermore, the auxiliary computation of $g$ takes time $O(c/\gamma)$.
\end{lemma}
\Crefname{enumi}{Step}{Steps}
\begin{proof}
The algorithm $\mathcal{E}$ runs $\mathcal{A}$ with $\eps = \frac{\gamma}{c}, \frac{2\gamma}{c}, \frac{3\gamma}{c},\ldots, 1$, and sets $\boldeta$ to be the largest $\eps$ for which $\mathcal{A}(g, s, \eps)$ rejects.
Since $\mathcal{A}(g, s, \boldeta)$ rejected,
\[ \boldeta < \opt_s(g)\]
with high probability. Furthermore, since $\mathcal{A}(g, s, \boldeta + \frac{\gamma}{c})$ accepted,
\begin{align*}
\opt_s(g) &< c \cdot \big(\boldeta + \lfrac{\gamma}{c}\big) \\
&= c \cdot \boldeta + \gamma
\end{align*}
with high probability. Finally, we note that $\mathcal{E}$ indeed makes $c/\gamma$ calls to $\mathcal{A}$, and aside from those calls, it only needs to make a single pass over the output of those calls and return the largest $\eps$ that led to a rejection, which takes time $O(c/\gamma)$.
\end{proof}
We are now ready to state our algorithm, $\textsc{BuildDT}$ (\Cref{fig:BuildDT}), for properly learning size-$s'$ decision trees. $\textsc{BuildDT}$ will additionally take in a depth parameter $d$ that will facilitate our analysis of it (looking ahead, $d$ will be chosen to be $O(\log(s'/\eps'))$ in our proof of~\Cref{thm:lower-bound}).
\begin{figure}[h]
\captionsetup{width=.9\linewidth}
\begin{tcolorbox}[colback = white,arc=1mm, boxrule=0.25mm]
\vspace{3pt}
$\textsc{BuildDT}(f,s,d,\gamma)$:
\begin{enumerate}[align=left]
\item[\textbf{Input:}] Query access to $f: \bn \to \bits$, parameters $s,d\in \N$ and $\gamma \in (0,1)$.
\item[\textbf{Output:}] A size-$s$ depth-$d$ decision tree $T$.
\end{enumerate}
\begin{enumerate}
\item \label{alg-base-case} If $s = 1$ or $d = 0$, return $\sign(\E[f])$.
\item For each $i \in [n]$ and integers $s_0, s_1 \geq 1$ satisfying $s_0 + s_1 = s$:
\begin{enumerate}
\item \label{subroutine} Use $\mathcal{E}$ from~\Cref{lem:distance-estimator} to obtain estimates $\boldeta(x_i=0,s_0)$ and $\boldeta(x_i=1,s_1)$ that satisfy:
\begin{align*}
\boldeta(x_i=0,s_0) \leq \opt_{s_0}(f_{x_i = 0}) \leq c \cdot \boldeta(x_i=0,s_0) + \gamma; \\
\boldeta(x_i=1,s_1) \leq \opt_{s_1}(f_{x_i = 1}) \leq c \cdot \boldeta(x_i=1,s_1) + \gamma.
\end{align*}
\item Store $\mathrm{error}(i, s_1, s_2) \leftarrow \frac{1}{2} \big(\boldeta(x_i=0,s_0) + \boldeta(x_i=1,s_1)\big)$.
\end{enumerate}
\item \label{minerror} Let $(i^\star, s_0^\star, s_1^\star)$ be the tuple that minimizes $\mathrm{error}(i, s_0, s_1)$. Output the tree with $x_{i^\star}$ as its root, $\textsc{BuildDT}(f_{x_{i^\star} = 0}, s_0^\star, d-1,\gamma)$ as its left subtree, and $\textsc{BuildDT}(f_{x_{i^\star} = 1}, s_1^\star, d-1,\gamma)$ as its right subtree.
\end{enumerate}
\end{tcolorbox}
\caption{$\textsc{BuildDT}$ computes a size-$s$ depth-$d$ decision tree that approximates a target function $f: \bn \to \bits$.}
\label{fig:BuildDT}
\end{figure}
\begin{lemma}[Error of $\textsc{BuildDT}$]
\label{lemma:build-dt-error}
For all functions $f: \bn \to \bits$ and parameters $s,d \in \N$ and $\gamma\in (0,1)$, the algorithm $\textsc{BuildDT}(f,s,d,\gamma)$ outputs a decision tree $T$ satisfying
\begin{align}
\label{eq:buildDT-accuracy}
\dist(T, f) \leq c^d \cdot \opt_s(f) + \gamma \cdot \frac{c^d - 1}{c-1} + \frac{s}{2^{d+2}}.
\end{align}
\end{lemma}
\begin{proof}
We proceed by induction on $s$ and $d$. If $s = 1$, then in \Cref{alg-base-case}, \textsc{BuildDT}\ outputs the best decision tree of size $1$. Therefore, $\dist(T,f) \leq \opt_s(f)$, satisfying \Cref{eq:buildDT-accuracy}. If $d = 0$ and $s \geq 2$, then $\frac{s}{2^{d+2}} \geq \frac{2}{4} = \frac1{2}$. Furthermore, in \Cref{alg-base-case}, \textsc{BuildDT}\ always outputs a tree with error at most $\frac{1}{2}$. Therefore,
\[ \dist(T, f) \leq \frac{s}{2^{d+2}} \leq c^d \cdot \opt_s(f) + \gamma \cdot \frac{c^d - 1}{c-1} + \frac{s}{2^{d+2}}.\]
Finally, we consider the case where $d \geq 1$ and $s \geq 2$. Let $T_{\opt}$ be the size-$s$ decision tree that is $\opt_s(f)$ close to $f$. Let $x_{i_{\opt}}$ the root of $T_{\opt}$, and $s_{0, \opt}$, $s_{1, \opt}$ the sizes of the left and right subtrees of $T_{\opt}$ respectively. Since the estimates computed in \Cref{subroutine} are underestimates of or equal to the true error (i.e.~$\mathrm{error}(i_{\opt}, s_{0, \opt}, s_{1, \opt}) \leq \opt_s(f)$), and
since $i^\star, s_0^\star, s_1^\star$ are chosen in~\Cref{minerror} to minimize the estimated error, we have
\[
\mathrm{error}(i^\star, s_0^\star, s_1^\star) \leq \mathrm{error}(i_{\opt}, s_{0, \opt}, s_{1, \opt}) \leq \opt_s(f).
\]
Finally, we bound $\dist(T, f)$. Let $T_0$ and $T_1$ be the left and right subtrees of $T$. Then,
\begin{align*}
\dist(T, f) &= \lfrac{1}{2} \big(\dist(T_0, f_{x_{i^\star} = 0}) + \dist(T_1, f_{x_{i^\star} = 1}) \big) \\
&\leq \lfrac{1}{2} \ds \Big(c^{d-1} \cdot \opt_{s_0^\star}(f_{x_{i^\star} = 0}) + \gamma \cdot \frac{c^{d-1} - 1}{c-1} + \frac{s_0^\star}{2^{d+1}} \\
& \ \ \ \ + c^{d-1} \cdot \opt_{s_1^\star}(f_{x_{i^\star} = 1}) + \gamma \cdot \frac{c^{d-1} - 1}{c-1} + \frac{s_1^\star}{2^{d+1}} \Big) \tag{Inductive hypothesis} \\
&= c^{d-1} \cdot \lfrac{1}{2}\, \ds \big(\opt_{s_0^\star}(f_{x_{i^\star} = 0}) + \opt_{s_1^\star}(f_{x_{i^\star} = 1})\big) + \gamma \cdot \frac{c^{d-1} - 1}{c-1} + \frac{s_0^\star + s_1^\star}{2^{d+2}} \\
&\leq c^{d-1} \cdot \lfrac{1}{2} \ds \big((c \cdot \boldeta(x_{i^\star}=0,s_0^\star) + \gamma) + (c \cdot \boldeta(x_{i^\star} = 1, s_1^\star) + \gamma)\big) + \gamma \cdot \frac{c^{d-1} - 1}{c-1} + \frac{s}{2^{d+2}}\\
&= c^{d} \cdot \lfrac{1}{2} \ds \,\big(\boldeta(x_{i^\star} = 0, s_0^\star) + \boldeta(x_{i^\star} = 1, s_1^\star)\big) + c^{d-1} \cdot \gamma + \gamma \cdot \frac{c^{d-1} - 1}{c-1} + \frac{s}{2^{d+2}} \\
&= c^{d} \cdot
\mathrm{error}(i^\star, s_0^\star, s_1^\star) + \gamma \cdot \frac{c^{d-1}(c-1) + c^{d-1} - 1}{c-1} + \frac{s}{2^{d+2}} \\
&\leq c^d\cdot \opt_s(f) + \gamma \cdot \Big(\frac{c^d - 1}{c-1}\Big) + \frac{s}{2^{d+2}}.
\end{align*}
The desired result holds by induction.
\end{proof}
For readability, \Cref{lemma:build-dt-error} assumes that $\textsc{BuildDT}$ is able to compute $\mathrm{round}(\E[f])$ in \Cref{alg-base-case}. To make $\textsc{BuildDT}$ efficient, we would only estimate $\E[f]$ by querying $f$ on uniform random inputs $\bx \in \bn$. If those estimates are computed to accuracy $\eps'$, then each leaf of our tree can have up to $\eps'$ additional error. This is not an issue since it increases the total error of $T$, which is simply the average of the error at each leaf, by only $\eps'$.
Finally, we prove \Cref{thm:lower-bound}:
\begin{proof}[Proof of \Cref{thm:lower-bound}]
Our goal is to properly learn a size-$s'$ decision tree $g: \bn \to \bits$ to accuracy $\eps'$. To do so, we run $\textsc{BuildDT}(g, s',d,\gamma)$, with $d$ set to
\begin{align*}
d = \log(s'/\eps') - 1,
\end{align*}
and $\gamma$ set to
\[ \gamma = \frac{\eps'}{2 \cdot \max(2, c)^d} = \frac{\eps'}{2} \cdot (2^{-d})^{\max(1, \log c)}
= \frac{\eps'}{2} \cdot \left(\frac{\eps'}{s'}\right)^{\max(1, \log c)}.
\]
By \Cref{lemma:build-dt-error}, for $T$ the tree $\textsc{BuildDT}$ outputs,
\begin{align*}
\dist(T,g) &\leq c^d \cdot \opt_{s'}(g) + \gamma \cdot \frac{c^d - 1}{c-1} + \frac{s'}{2^{d+2}} \\
&\leq 0 + \frac{\eps'}{2 \cdot \max(2, c)^d} \cdot \max(2, c)^d + \frac{s'}{2^{\log(s'/\eps') + 1}} \\
& \leq \frac{\eps'}{2} + \frac{\eps'}{2} = \eps'.
\end{align*}
Hence, $\textsc{BuildDT}$ produces the desired output. We next argue that it is efficient. During the recursion, $\textsc{BuildDT}$ is called at most $s'$ times in total. Each such call makes $O(ns')$ calls to $\mathcal{E}$. By \Cref{lem:distance-estimator}, those calls to $\mathcal{E}$ each make $c/\eps$ calls to $\mathcal{A}$. Hence, the total number of calls to $\mathcal{A}$ is
\begin{align*}
O\left(\frac{n(s')^2}{\gamma}\right) = O\left(\frac{n (s')^2}{\eps'} \cdot \left(\frac{s'}{\eps'}\right)^{\max(1 ,\log c)}\right) = \poly(n,s',1/\eps').
\end{align*}
The total auxiliary computation of $\textsc{BuildDT}$ is bounded by the same quantity. Finally, each call to $\mathcal{A}$ is made with parameters $s$ and $\eps$ where $s \leq s'$ and $\eps = \frac{\eps'}{2} \cdot \left(\frac{\eps'}{s'}\right)^{\max(1, \log c)} \ge \poly(1/s',\eps')$.
\end{proof}
\Crefname{enumi}{Item}{Item}
\subsection{Algorithmic component of~\Cref{thm:reconstruct}}
\label{sec:algorithmic}
\subsubsection{Query-efficient simultaneous score estimation}
\label{sec:score-estimator}
We begin by designing a query-efficient subroutine that simultaneously estimates the scores of all~$n$ variables of a function $f$. The fact that we are able to do so with $O(\log n)$ queries, as opposed to $\Omega(n)$ as would be required by a naive approach, will be a key component in the query efficiency of our reconstructor.
\begin{theorem}[Score estimator]
\label{thm:score-estimator}
There is an algorithm which, given query access to a function $f: \bits^n \to \bits$, noise rate $p \in (0,1)$, accuracy parameter $\tau \in (0,1)$, and confidence parameter $\delta \in (0,1)$, for
\[ q = O\left(\frac{\log n + \log(1/\delta)}{\tau^2} \right)
\]
makes $O(q)$ queries, runs in $O(q n)$ time, and returns estimates $\boldeta_1, \ldots, \boldeta_n$ such that, with probability at least $1 - \delta$, satisfies
\[ \big| \boldeta_i - \mathrm{Score}_i(f, p) \big | < \tau \quad \text{for all $i \in [n]$}.
\]
\end{theorem}
We prove~\Cref{thm:score-estimator} by first giving a $2$-query algorithm, $\textsc{UnbiasedEstimator}$ (\Cref{fig:UnbiasedEstimator}), that runs in $O(n)$ time and outputs unbiased estimates of all $n$ scores. The algorithm of \Cref{thm:score-estimator} takes the mean of multiple runs of that unbiased estimator, with its guarantees following from a simple concentration bound.
\begin{figure}[h]
\captionsetup{width=.9\linewidth}
\begin{tcolorbox}[colback = white,arc=1mm, boxrule=0.25mm]
\vspace{3pt}
$\textsc{UnbiasedEstimator}(f,p)$:
\begin{enumerate}[align=left]
\item[\textbf{Input:}] Query access to a function $f: \bits^n \to \bits$ and a noise rate $p \in (0,1)$.
\item[\textbf{Output:}] Unbiased estimates of $\mathrm{Score}_i(f, p)$ for all $i \in [n]$.
\end{enumerate}
\begin{enumerate}
\item Choose $\bx \in \bits^n$ uniformly at random and generate a $p$-noisy copy $\by$ of $\bx$.
\item For each $i \in [n]$, return the estimate
\begin{align*}
\boldeta_i = \Ind\big[f(\bx) \neq f(\by)\big] \cdot \left (1 - \frac{1}{1 - \frac{p}{2}} \cdot \Ind[\bx_i = \by_i]\right).
\end{align*}
\end{enumerate}
\end{tcolorbox}
\caption{$\textsc{UnbiasedEstimator}$ computes unbiased estimates of the scores of all variables of a function $f$.}
\label{fig:UnbiasedEstimator}
\end{figure}
\begin{lemma}[Analysis of $\textsc{UnbiasedEstimator}$]
\label{lemma:unbiased-estimates}
For any $f: \bits^n \to \bits$ and $p \in (0,1)$, let $\boldeta_1, \ldots, \boldeta_n$ be the outputs of $\textsc{UnbiasedEstimator}(f, p)$. Then
\begin{align*}
\Ex_{\bx, \by}[\boldeta_i] = \mathrm{Score}_i(f, p) \quad \text{for all $i\in [n]$.}
\end{align*}
\end{lemma}
\begin{proof}
We first note that $\Pr[f(\bx) \neq f(\by)]$ is $\NS_p(f)$ by definition. Therefore, it is enough for us to prove that
\begin{align}
\label{eq:unbiased-estimator}
\Ex_{\bb \in \bits}\big[\NS_{p}(f_{x_i = \bb})\big] = \frac{1}{1 - \frac{p}{2}} \cdot \Prx_{\bx, \by}\left[f(\bx) \neq f(\by) \text{ and } \bx_i = \by_i\right].
\end{align}
Given the above equation, the desired result holds by linearity of expectation and the definition of score. Consider the distribution over $(\bx, \by)$ conditioned on the event that $b = \bx_i = \by_i$. That distribution is equivalent to if we picked $\bx$ randomly from the domain of $f_{x_i = b}$ and selected $\by$ by rerandomizing each coordinate in that domain with probability $p$. Therefore,
\begin{align*}
\NS_{p}(f_{x_i = b}) &= \Prx_{\bx, \by}[f(\bx) \neq f(\by) \,|\, b= \bx_i = \by_i] \\
&= \frac{1}{\Prx_{\bx, \by}[b= \bx_i = \by_i]}] \cdot \Prx_{\bx, \by}[f(\bx) \neq f(\by) \text{ and } b= \bx_i = \by_i].
\end{align*}
We now prove \Cref{eq:unbiased-estimator}:
\begin{align*}
\Ex_{\bb \in \bits}\big[\NS_{p}(f_{x_i = \bb})\big] &= \Ex_{\bb \in \bits}\left[\frac{1}{\ds\Prx_{\bx, \by}[\bb= \bx_i = \by_i]}] \cdot \Prx_{\bx, \by}[f(\bx) \neq f(\by) \text{ and } \bb= \bx_i = \by_i]\right] \\
&=\frac{1}{\frac{1}{2} \cdot \ds\Prx_{\bx, \by}[ \bx_i = \by_i]}\Ex_{\bb \in\bits}\left[\Prx_{\bx, \by}[f(\bx) \neq f(\by) \text{ and } \bb= \bx_i = \by_i]\right] \\
&=\frac{1}{\frac{1}{2} \cdot (1 - \frac{p}{2})} \cdot \lfrac{1}{2} \cdot \ds\Prx_{\bx, \by}\left[f(\bx) \neq f(\by) \text{ and } \bx_i = \by_i\right] \\
&= \frac{1}{1 - \frac{p}{2}} \cdot \Prx_{\bx, \by}\left[f(\bx) \neq f(\by) \text{ and } \bx_i = \by_i\right].
\end{align*}
\Cref{lemma:unbiased-estimates} then holds by linearity of expectation.
\end{proof}
We now prove \Cref{thm:score-estimator}.
\begin{proof}[Proof of \Cref{thm:score-estimator}]
The algorithm runs $\textsc{UnbiasedEstimator}(f, p)$ $q$ times and then outputs the means of each returned estimates. Each estimate from $\textsc{UnbiasedEstimator}$ is bounded between $-1$ and $1$. By Hoeffding's inequality, for any $i \in [n]$,
\begin{align*}
\Pr\big[\big| \boldeta_i - \mathrm{Score}_i(f, p) \big | \geq \tau\big] \leq -\exp_e\left(-\frac{q \cdot \tau^2}{2}\right).
\end{align*}
For $q$ as in \Cref{thm:score-estimator}, the above probability is at most $\delta/n$. By union bound, all estimates are accurate within $\pm\tau$ with probability at least $1 - \delta$.
Finally, this algorithm uses only $2q = O(q)$ queries. Each run of $\textsc{UnbiasedEstimator}$ estimator takes $O(n)$ time to construct the query and compute all the estimates, so the entire algorithm takes $O(q n)$ time.
\end{proof}
\subsubsection{Proof of \Cref{thm:reconstruct}}
\label{sec:proof-of-reconstruct}
We prove \Cref{thm:reconstruct} by providing an algorithm, $\textsc{Reconstructor}$ (\Cref{fig:BuildStrand}), which assumes query access to a function $f : \bn\to\bits$ and provides fast query access to a tree $T$ meeting the criteria of \Cref{thm:reconstruct}. We build off a simple observation that also underlies \cite{BGLT-NeurIPS2}: to determine the output of a decision tree $T$ on a particular input~$z$, it suffices to build the root-to-leaf path corresponding to~$z$, which can be exponentially faster than building the entire tree. Our algorithm is different from~\cite{BGLT-NeurIPS2}'s; as mentioned in the introduction their algorithm is tailored to monotone functions, and is known to fail for non-monotone ones. We on the other hand leverage the specific structure of $T$ established in~\Cref{sec:structural} together with the query-efficient score estimator from~\Cref{sec:score-estimator} in our design and analysis of $\textsc{Reconstructor}$.
$\textsc{Reconstructor}$ maintains a partial tree $T^\circ$ containing all the root-to-leaf paths in $T$ corresponding to queries received so far. In the pseudocode for $\textsc{Reconstructor}$, we use the notation $T^\circ_{\mathrm{internal}}(\alpha) \in [n] \cup \{\varnothing\}$ to indicate the variable queried in $[n]$ at internal node $\alpha$ of the partial tree $T^\circ$, or $\varnothing$ if that node has not yet been built. Similarly, $T^\circ_{\mathrm{leaf}}(\alpha) \in \{-1,1,\varnothing\}$ indicates the value at leaf $\alpha$ in $T^\circ$, or $\varnothing$ if that value has not yet been decided.
\Crefname{enumi}{Step}{Steps}
\begin{figure}[h]
\captionsetup{width=.9\linewidth}
\begin{tcolorbox}[colback = white,arc=1mm, boxrule=0.25mm]
\vspace{3pt}
$\textsc{Reconstructor}(f, s, \eps, \delta)$:
\begin{enumerate}[align=left]
\item[\textbf{Input:}] Query access to a function $f: \bits^n \to \bits$, size parameter $s$, error parameter $\eps$, and failure probability $\delta$.
\item[\textbf{Output:}] Query access to a decision tree $T$ that satisfies $\dist(f, T) \leq O(\opt_s) + \eps$ with probability at least~$1-\delta$.
\end{enumerate}
\begin{enumerate}
\item Set parameters $d, p$, and $\tau$ as in~\Cref{lem:BGLT-approx}.
\item Initialize $T^\circ$ to be the empty partial tree.
\item Upon receiving an input $z \in \bits^n$:
\begin{enumerate}
\item Initialize $\alpha$ to be the root of $T^\circ$.
\item Repeat $d$ times.
\begin{enumerate}
\item \label{step:score-estimate} If $T^\circ_{\mathrm{internal}}(\alpha)$ is $\varnothing$ use the estimator from \Cref{thm:score-estimator} to compute estimates of $\mathrm{Score}_i(f_{\alpha}, p)$ with additive accuracy $\pm \frac{\tau}{2}$ and failure probability $O(\frac{\delta}{2^d})$ for all $i \in [n]$ and set $T^\circ_{\mathrm{internal}}(\alpha)$ to the variable with highest estimated score.
\item For $i = {T^\circ_{\mathrm{internal}}(\alpha)}$, If $\overline{z}_{i}$ is $1$, set $\alpha$ to its right child. Otherwise, set $\alpha$ to its left child.
\end{enumerate}
\item \label{step:leaf-estimate} If $T^\circ_{\mathrm{leaf}}(\alpha)$ is $\varnothing$, use random samples to estimate $\E[f_{\ell}]$ to additive accuracy $\pm \frac{\eps}{4}$ with failure probability $O(\frac{\delta}{2^d})$ and set $T^\circ_{\mathrm{leaf}}(\alpha)$ to whichever of $\bits$ that estimate is closer to.
\item Output $T^\circ_{\mathrm{leaf}}(\alpha)$.
\end{enumerate}
\end{enumerate}
\end{tcolorbox}
\caption{$\textsc{Reconstructor}$ gives efficient query access to a decision tree is close to $f$ with high probability.}
\label{fig:BuildStrand}
\end{figure}
\Cref{thm:reconstruct} follows from the following two lemmas, showing the correctness and efficiency of $\textsc{Reconstructor}$ respectively.
\begin{lemma}[Correctness of $\textsc{Reconstructor}$] \label{lemma:reconstructor-correctness}
For any $f: \bits^n \to \bits$, $s \in \N$, $\eps \in (0, \frac{1}{2})$, $\delta \in (0,1)$, and sequence of inputs $z^{(1)}, \ldots, z^{(m)}\in \bn$, the outputs of $\textsc{Reconstructor}$ are consistent with some decision tree $T$ where
\begin{itemize}
\item[$\circ$] $T$ has size $s^{O((\log s)^2/\eps^3)}$,
\item[$\circ$] $\dist(T,f) \le O(\opt_s) + \eps$ with probability at least $1-\delta$.
\end{itemize}
\end{lemma}
\begin{proof}
The outputs of $\textsc{Reconstructor}$ are always consistent with $T^\circ$ and the depth of $T^\circ$ is always capped at $d$. Let $T$ be the tree that $T^\circ$ would be if every $x \in \bits^n$ were given as an input to $\textsc{Reconstructor}$. Then, $T$ has size at most $2^d = s^{O((\log s)^2/\eps^3)}$, and every output is consistent with $T$.
If all score estimates in \Cref{step:score-estimate} are accurate to $\pm \frac{\tau}{2}$ and expectation estimates is \Cref{step:leaf-estimate} are accurate to $\pm \frac{\eps}{4}$, then $T$ meets the criteria of \Cref{lem:BGLT-approx} and therefore $\dist(T, f) \leq O(\opt_s) +\eps$. The number of time scores are estimated in \Cref{step:score-estimate} is at most the number of internal nodes of~$T$, which is $2^d - 1$. Similarly, the number of expectation estimates in \Cref{step:score-estimate} is at most the number of leaves of $T$, which is $2^d$. By union bound over the possible failures, we see that the failure probability is at most $\delta$.
\end{proof}
\begin{lemma}[Efficiency of $\textsc{Reconstructor}$]\label{lem:reconstruct-efficiency}
For any $f: \bits^n \to \bits$, $s \in \N$, $\eps \in (0, \frac{1}{2})$, $\delta \in (0,1)$, particular input $z \in \bits^n$, and
\begin{align*}
q = O\left(\frac{(\log s)^9\cdot (\log n) \cdot \log(1/\delta)}{\eps^9} \right),
\end{align*}
upon receiving $z$ as input, $\textsc{Reconstructor}(f,s,\eps,\delta)$ uses $O(q)$ queries and $O(qn)$ time to return an output.
\end{lemma}
\begin{proof}
On each input, the estimator from \Cref{thm:score-estimator} is used up to $d$ times. Each uses
\begin{align*}
q_{\mathrm{inner}} \coloneqq O\left(\frac{\log n + \log(2^d/\delta)}{\tau^2} \right) = O\left(\frac{\log n + d + \log(1/\delta)}{\tau^2} \right)
\end{align*}
queries and $O(q_\mathrm{inner}n)$ time. By Hoeffding's inequality, it is sufficient to take
\begin{align*}
q_{\mathrm{leaf}} \coloneqq O\left(\frac{\log(2^d/\delta)}{\eps^2} \right) = O\left(\frac{d + \log(1/\delta)}{\eps^2} \right)
\end{align*}
random samples in \Cref{step:leaf-estimate}. Therefore, the total number of queries used is
\begin{align*}
q &= q_{\mathrm{inner}} + q_{\mathrm{leaf}} \\
&= O\left(\frac{\log n + d + \log(1/\delta)}{\tau^2} \right) + O\left(\frac{d + \log(1/\delta)}{\eps^2} \right) \\
&= O\left(\frac{\log n + ((\log s)^3/\eps^3) + \log(1/\delta)}{\eps^6 /( \log s)^6} + \frac{((\log s)^3/\eps^2) + \log(1/\delta)}{\eps^2} \right)\\
&= O\left(\frac{(\log s)^9\cdot (\log n) \cdot \log(1/\delta)}{\eps^9} \right).
\end{align*}
The time to prepare all queries is $O(q n)$, and all other computation is asymptotically faster.
\end{proof}
\begin{remark}[Local reconstruction]
\label{remark: local reconstruction}
We remark that our reconstruction algorithm can be made local in the sense of \cite{SS10}. They define a reconstruction algorithm, $\mathcal{A}$, to be local, if the output of $\mathcal{A}$ on some input $z$ is a \emph{deterministic} and easy to compute function of $z$ and some small random string $\rho$. This allows queries to the reconstructor to be answered in parallel, as long as the random string $\rho$ is shared. To make our reconstructor local, we note that the only place randomness is used is in generating samples consistent with some restriction $\alpha$. We can set $\rho$ to be $n$ bits per sample the constructor might wish to generate. Since the total number of samples the reconstructor needs per input is $\poly(\log s, 1/\eps, \log(1/\delta))\cdot \log n$, we have
\begin{align*}
|\rho| = \poly(\log s, 1/\eps, \log(1/\delta))\cdot n\log n
\end{align*}
On a particular input, the local reconstructor starts with $T^\circ$ being the empty tree. Whenever it wishes to produce a random sample consistent with $\alpha$, it sets the $\bx \in \bits^n$ to be next $n$ bits of $\rho$ and then uses $\bx_{\alpha}$ for the sample. It's easy to see that this algorithm will keep $T^\circ$ consistent between different runs because it will always compute the same variable as having the highest score given some restriction. Furthermore, the analysis goes through without issue. The only difference between this analysis and one where fresh random bits are used to for each sample is that the queries of different paths may be correlated. In our proof of \Cref{lemma:reconstructor-correctness}, we use a union bound to ensure all estimates obtained through sampling are accurate, and that union bound holds regardless of whether those estimates are independent.
\end{remark}
\subsection{Proof of~\Cref{cor:main}}
\label{sec:proof-of-tester}
In this section we derive \Cref{cor:main} as a simple consequence of~\Cref{thm:reconstruct}. The connection between reconstruction and tolerant testing has been noted in other works (see e.g.~\cite{CGR13, B08}); we provide a proof here for completeness.
\begin{theorem}[\Cref{cor:main} restated]
There is an algorithm which, given query access to $f : \bn \to\bits$ and parameters $s\in\N$ and $\eps, \delta \in (0,1)$, runs in $\poly(\log s,1/\eps)\cdot n\log n\cdot \log(1/\delta)$ time, makes $\poly(\log s,1/\eps) \cdot \log n \cdot \log(1/\delta)$ queries to $f$, and
\begin{itemize}
\item[$\circ$] Accepts w.p.~at least $1 - \delta$ if $f$ is $\eps$-close to a size-$s$ decision tree;
\item[$\circ$] Rejects w.p.~at least $1 - \delta$ if $f$ is $\Omega(\eps)$-far from size-$s^{O((\log s)^2/\eps^3)}$ decision trees
\end{itemize}
\end{theorem}
\begin{proof}
The algorithm chooses $m$ uniform random inputs, $\bx^{(1)}, \ldots, \bx^{(m)} \sim \bits^n$ where $m = O(\log(1/\delta)/\eps^2)$. Let $\bb^{(1)}, \ldots, \bb^{(m)} \in \bits$ be the output of $\textsc{Reconstructor}(f, s, \eps, \delta)$. The tester rejects if $\E_{\bi \in [m]}[f(\bx^{(\bi)}) \neq \bb^{(\bi)}] > \Omega(\eps)$ and accepts otherwise.
First, we consider the case where $f$ is $\eps$-close to a size-$s$ decision tree (i.e.~$\opt_s \le \eps$). By \Cref{lemma:reconstructor-correctness}, with probability at least $1 - \delta$ the outputs of $\textsc{Reconstructor}$ are consistent with a tree, $T$, satisfying $\dist(T, f) \leq O(\eps)$. By Hoeffding's inequality,
\[
\Prx_{\bx^{(1)}, \ldots, \bx^{(m)}}\left[\Ex_{\bi \in [m]}\big[f(\bx^{(\bi)}) \neq \bb^{(\bi)}\big] > \Omega(\eps)\right] \leq \exp(-2 m \eps^2) \le \delta.\]
By a union bound, the tester rejects with probability at most $\delta + \delta = 2\delta$.
We next consider the case where $f$ is $\Omega(\eps)$-far from size-$s^{O((\log s)^2/\eps^3)}$ decision trees. By \Cref{lemma:reconstructor-correctness} it is guaranteed to be consistent. A similar argument to the first case shows that the probability of acceptance is at most $\delta + \exp(-2 m \eps^2) = 2\delta$. Finally, the efficiency of this tester is a consequence of \Cref{lem:reconstruct-efficiency} and our choice of $m = O(\log(1/\delta)/\eps^2).$
\end{proof}
\Crefname{enumi}{Item}{Item}
\section{placeholder}
blah blah blah
\section{Proof of \Cref{cor:reconstruct-degree}
We first restate~\Cref{thm:reconstruct} with decision tree {\sl depth} instead of {\sl size} as the complexity measure:
\begin{theorem}[\Cref{thm:reconstruct} in terms of decision tree depth]
\label{thm:reconstruct-depth}
There is a randomized algorithm which,
given query access to $f : \bn\to\bits$ and parameters $d\in \N$ and $\eps\in (0,1)$, provides query access to a fixed decision tree $T$ where
\begin{itemize}
\item[$\circ$] $T$ has depth $O(d^3/\eps^2)$,
\item[$\circ$] $\dist(T,f) \le O(\opt_d) + \eps$ w.h.p., where $\opt_d$ denotes the distance of $f$ to the closest depth-$d$ decision tree.
\end{itemize}
Every query to $T$ is answered in $\poly(d,1/\eps)\cdot n\log n$ time and with $\poly(d,1/\eps)\cdot \log n$ queries to~$f$.
\end{theorem}
To see that our proof of~\Cref{thm:reconstruct} also establishes~\Cref{thm:reconstruct-depth}, we use the fact that every depth-$d$ decision tree has size $\le 2^d$, and recall that the tree $T$ that the algorithm of~\Cref{thm:reconstruct} provides query access to is a complete tree and hence has depth logarithmic in its size.
Decision tree depth and Fourier degree of boolean functions are known to be polynomially related:
\begin{fact}[Decision tree depth vs.~Fourier degree~\cite{Mid04,Tal13}]
\label{fact:depth-vs-degree}
For $g : \bn\to\bits$ let $\deg(g)$ denote $g$'s Fourier degree and $D(g)$ denote the depth of the shallowest decision tree that computes~$g$. Then $\deg(g) \le D(g)$ and $D(g) \le \deg(g)^3$.
\end{fact}
We first observe \Cref{thm:reconstruct-depth} and~\Cref{fact:depth-vs-degree} already gives a quantitatively weaker version of~\Cref{cor:reconstruct-degree} where $g$ has degree $O(d^{\,9}/\eps^2)$. To see this $f : \bn\to\bits$ be $\opt_d$-close to a degree-$d$ function $h : \bn\to\bits$. By~\Cref{fact:depth-vs-degree}, $D(h) \le \deg(h)^3$, and so the algorithm of~\Cref{thm:reconstruct-depth} provides query access to a decision tree $T : \bn\to\bits$ that is $(O(\opt_d) + \eps)$-close to $f$ and where the depth of $T$ is $O(D(h)^3/\eps^2) = O(\deg(h)^9/\eps^2)$. Applying~\Cref{fact:depth-vs-degree} again, we conclude that $\deg(T) \le D(T) \le O(\deg(h)^9/\eps^2)$.
To obtain the sharper bound of $O(\deg(h)^7/\eps^2)$, we observe that the proof of~\Cref{lemma:BGLT} in fact bounds the depth of $T$ by $O(D(h)^2\, \Inf(h)/\eps^2)$.
Influence and degree of boolean functions are related via the following basic fact (see e.g.~\cite[Theorem 37]{ODbook}):
\begin{fact}
For all $h : \bn\to\bits$, we have $\Inf(h) \le \deg(h)$.
\end{fact}
Therefore, we can bound the degree of $T$ by $O(D(h)^2\,\Inf(h)/\eps^2) \le O(\deg(h)^7/\eps^2)$.
Guarantees for the other measures listed in~\Cref{table:other-measures} follow from similar calculations and known quantitative relationships between these measures and decision tree complexity; the current best bounds are summarized in Table 1 of~\cite{ABDKRT20}.
\section{Introduction}
\violet{We study the problem of {\sl reconstructing} decision trees: given queries to a function $f$ that is close to a size-$s$ decision tree, provide fast query access to a decision tree, ideally one of size not much larger than $s$, that is close to~$f$. This can be viewed as an ``on the fly" variant of the problem of properly and agnostically learning decision trees, where the goal there is to output the entire decision tree hypothesis. More broadly, reconstruction algorithms, introduced by Ailon, Chazelle, Comandur, and Liu~\cite{ACCL08}, can be viewed as sublinear algorithms that restore structure---in our case, that of a decision tree---in a function that has been lost due to noise.}
Decision trees have long been a popular and effective model in machine learning, and relatedly, they are among the most intensively studied concept classes in learning theory. The literature on learning decision trees is vast, spanning three decades and studying the problem in a variety of models and from a variety of perspectives~\cite{EH89,Riv87,Blu92,Han93,Bsh93,BFJKMR94,HJLT96,MR02,JS06,OS07,KS06,KST09,HKY18,CM19,BDM20,BLT-ITCS,BLT-ICML,BGLT-NeurIPS1,BLT21icalp,BLQT21}.
\violet{In contrast, the problem of reconstructing decision trees has thus far been surprisingly understudied.}
\subsection{Our contributions}
\violet{We give the first reconstruction algorithm for decision trees. Our algorithm achieves a {\sl polylogarithmic} dependence on $s$ in its query and time complexities, exponentially smaller than the information-theoretic minimum required to learn.}
\begin{theorem}[Main result]
\label{thm:reconstruct}
There is a randomized algorithm which,
given queries to $f : \bn\to\bits$ and parameters $s\in \N$ and $\eps\in (0,1)$, provides query access to a fixed decision tree~$T$ where
\begin{itemize}
\item[$\circ$] $T$ has size $s^{O((\log s)^2/\eps^3)}$
\item[$\circ$] $\dist(T,f) \le O(\opt_s) + \eps$ w.h.p., where $\opt_s$ denotes the distance of $f$ to the closest size-$s$ decision tree;
\item[$\circ$] Every query to $T$ is answered with $\poly((\log s)/\eps)\cdot \log n$ queries to~$f$ and in $\poly((\log s)/\eps)\cdot n\log n$ time.
\end{itemize}
\end{theorem}
\violet{Notably, in the standard setting where $s = \poly(n)$, the query and time complexities of our algorithm are $\polylog(n)$ and $\tilde{O}(n)$ respectively. Previously, the only known approach was to simply properly and agnostically learn $f$; the current fastest such algorithm has query and time complexities $n^{O(\log\log n)}$~\cite{BLQT21}.}
Our reconstruction algorithm is furthermore {\sl local} in the sense of Saks and Seshadhri~\cite{SS10}, allowing queries to be answered in parallel assuming a shared random string. In particular, once $f, s, \eps$ and the random string are fixed, all queries are answered consistently with a single decision tree.
\subsubsection{\violet{Implications of~\Cref{thm:reconstruct} and further results}}
By a standard reduction,~\Cref{thm:reconstruct} gives a {\sl tolerant tester} for decision trees:
\begin{corollary}[Tolerant testing of decision trees]
\label{cor:main}
There is a randomized algorithm which, given queries to $f : \bn \to\bits$ and parameters $s\in\N$ and $\eps \in (0,1)$,
\begin{itemize}
\item[$\circ$] Makes $\poly((\log s)/\eps) \cdot \log n$ queries to $f$, runs in $\poly((\log s)/\eps)\cdot n\log n$ time, and
\item[$\circ$] Accepts w.h.p.~if $f$ is $\eps$-close to a size-$s$ decision tree;
\item[$\circ$] Rejects w.h.p.~if $f$ is $\Omega(\eps)$-far from size-$s^{O((\log s)^2/\eps^3)}$ decision trees
\end{itemize}
\end{corollary}
This adds to a long line of work on testing decision trees~\cite{KR00,DLMORSW07,CGM11,BBM12,Bsh20}. \violet{We give an overview of prior testers in~\Cref{sec:prior}, mentioning for now that they all have (at least) an exponentially larger dependence on $s$ in their query and time complexities.}
\paragraph{A new connection between tolerant testing and learning.} It would be preferable if our tester can be improved to reject all $f$'s that are far
from size-$s$ decision trees---or more strongly, if our reconstructor can be improved to provide query access to a size-$s$ decision tree.
We show that such a tester, even one that is considerably less efficient than ours, would yield the first polynomial-time algorithm for properly learning decision trees:
\begin{theorem}[Tolerant testing $\Longrightarrow$ Proper learning]
\label{thm:lower-bound-intro}
Suppose there is an algorithm which, given query access to $f : \bn \to \bits$ and parameters $s\in \N$ and $\eps \in (0,1)$,
\begin{itemize}
\item[$\circ$] Makes $\poly(s,n,1/\eps)$ queries to $f$, runs in $\poly(s,n,1/\eps)$ time, and
\item[$\circ$] Accepts w.h.p.~if $f$ is $\eps$-close to a size-$s$ decision tree;
\item[$\circ$] Rejects w.h.p.~if $f$ is $\Omega(\eps)$-far from size-$s$ decision trees.
\end{itemize}
Then there is a $\poly(s,n,1/\eps)$-time membership query algorithm for properly learning size-$s$ decision trees with respect to the uniform distribution.
\end{theorem}
This would represent a breakthrough on a central open problem in learning theory. Recent work of Blanc, Lange, Qiao, and Tan~\cite{BLQT21} gives a $\poly(n)\cdot s^{O(\log\log s)}$ time algorithm, improving on the prior state of the art of $n^{O(\log s)}$~\cite{EH89}. Neither~\cite{EH89}'s nor~\cite{BLQT21}'s algorithm goes through testing.
It is well known and easy to see that proper learning algorithms yield comparably efficient testers~\cite{GGR98}. \Cref{thm:lower-bound-intro} provides an example of a converse; \violet{we find the existence of such a converse surprising, and are not aware of any previous examples.}
\paragraph{Reconstructors and testers for other properties.}
Decision tree complexity is quantitatively related to numerous other complexity measures of boolean functions: Fourier degree, approximate degree, randomized and quantum query complexities, certificate complexity, block sensitivity, sensitivity, etc. Our results therefore immediately yield new reconstructors and tolerant testers for these properties. For example, we have the following:
\begin{corollary}[Reconstruction of low Fourier degree functions]
\label{cor:reconstruct-degree}
There is a randomized algorithm which,
given queries to $f : \bn\to\bits$ and parameters $d\in \N$ and $\eps\in (0,1)$, provides query access to a fixed function $g : \bn\to\bits$ where
\begin{itemize}
\item[$\circ$] $g$ has Fourier degree $O(d^7/\eps^2)$,
\item[$\circ$] $\dist(f,g) \le O(\opt_d) + \eps$ w.h.p., where $\opt_d$ denotes the distance of $f$ to closest $h : \bn\to\bits$ of Fourier degree $d$.
\item[$\circ$] Every query to $g$ is answered in $\poly(d,1/\eps)\cdot n\log n$ time and with $\poly(d,1/\eps)\cdot \log n$ queries to~$f$.
\end{itemize}
\end{corollary}
This in turn yields a tolerant tester for Fourier degree. As in the case for decision trees, all prior testers for low Fourier degree~\cite{DLMORSW07,CGM11,CGM11-soda,BBM12,BH13,Bsh20} have an exponential dependence on $d$ in their query and time complexities.
\Cref{table:other-measures} lists examples of measures for which we obtain new reconstruction algorithms, each of which in turn give new tolerant testers
\vspace{5pt}
\begin{table}[h!]
\begin{adjustwidth}{-1in}{-1in}
\renewcommand{\arraystretch}{1.8}
\centering
\begin{tabular}{|c|c|c|}
\hline
\multirow{3}{*}{\sl Complexity measure} & {\sl Assumption} & {\sl Guarantee} \\ [-.5em]
&~~Query access to $f$ that is~~&~~Query access to $g$ that is~~\\ [-.9em]
&~~$\opt_d$-close to $h$ where:~~&~~$O(\opt_d+\eps)$-close to $f$ where:~~\\ [.2em] \hline \hline
Fourier degree & $\deg(h) \le d$ &~~$\deg(g) \le O(d^7/\eps^2)$~~ \\ [.2em] \hline
Approximate degree &~~$\wt{\deg}(h) \le d$~~&~~$\wt{\deg}(g) \le O(d^{9}/\eps^2)$~~ \\ [.2em] \hline
~~Randomized query complexity~~&~~$\mathrm{R}(h) \le d$~~&~~$\mathrm{R}(g) \le O(d^7/\eps^2)$~~ \\ [.2em] \hline
~~Quantum query complexity~~&~~$\mathrm{Q}(h) \le d$~~&~~$\mathrm{Q}(g) \le {O(d^{10}/\eps^2)}$~~ \\ [.2em] \hline
~~Certificate complexity~~&~~$\mathrm{C}(h) \le d$~~&~~$\mathrm{C}(g) \le O(d^5/\eps^2)$~~ \\ [.2em] \hline
~~Block sensitivity~~&~~$\mathrm{bs}(h) \le d$~~&~~$\mathrm{bs}(g) \le O(d^8/\eps^2)$~~ \\ [.2em] \hline
~~Sensitivity~~&~~$\mathrm{s}(h) \le d$~~&~~$\mathrm{s}(g) \le O(d^{13}/\eps^{2})$~~ \\ [.2em] \hline \end{tabular}
\end{adjustwidth}
\captionsetup{width=.9\linewidth}
\caption{Performance guarantees of our reconstruction algorithms for various complexity measures. In each row, $\opt_d$ denotes the distance from $f$ to the closest function $h$ such that the complexity measure of that row for $h$ is bounded by $d$. In all cases, every query to $g$ is answered in $\poly(d,1/\eps)\cdot n\log n$ time with $\poly(d,1/\eps)\cdot \log n$ queries to $f$. }
\label{table:other-measures}
\vspace{-5pt}
\end{table}
\subsection{Background and comparison with prior work}
\label{sec:prior}
\violet{As already mentioned,~\Cref{thm:reconstruct} gives the first reconstruction algorithm for decision trees. The problem of testing decision trees, on the other hand, has been intensively studied. }
\violet{\pparagraph{Testing decision trees.}} Recent work of Bshouty~\cite{Bsh20} gives an algorithm, running in $\poly(s^s,1/\eps)\cdot n$ time and using $O((s\log s)/\eps)$ queries, that distinguishes between size-$s$ decision trees from functions that are $\eps$-far from size-$s$ decision trees. Prior to~\cite{Bsh20}, Chakraborty, Garc{\'\i}a-Soriano, and Matsliah~\cite{CGM11} gave an $O((s\log s)/\eps^2)$-query algorithm, and before that Diakonikolas, Lee, Matulef, Onak, Rubinfeld, Servedio, and Wan~\cite{DLMORSW07} gave an $\tilde{O}(s^4/\eps^2)$-query algorithm. Like \cite{Bsh20}'s algorithm, the algorithms of~\cite{CGM11,DLMORSW07} also run in $\poly(s^s,1/\eps)\cdot n$ time.\footnote{\violet{All these testers enjoy a weak form of tolerance: they are in fact able to distinguish between functions that are $O(\poly(\eps/s))$-close to size-$s$ decision trees from those that are $\eps$-far from size-$s$ decision trees. (Briefly, this is because their queries, while correlated, are each uniformly distributed.)}}
Compared to these algorithms, our algorithm in~\Cref{cor:main} solves an incomparable problem with efficiency parameters that compare rather favorably with theirs. Notably, our time and query complexities both depend {\sl polylogarithmically} on $s$ instead of exponentially and super-linearly respectively
Turning to the parameterized setting, Kearns and Ron~\cite{KR00} gave a tester with time and query complexities $\poly(n^n, (\log s)^n)$ that distinguishes size-$s$ decision trees over $[0,1]^n$ from functions that are $(\frac1{2}-n^{-\Theta(n)})$-far from size-$\poly(2^n,s)$ decision trees. The parameters of this result are such that one should think of the dimension `$n$' as being a constant rather than an asymptotic parameter.
\paragraph{Property reconstruction.}
Property reconstruction was introduced by Ailon, Chazelle, Comandur, and Liu~\cite{ACCL08}. (See also the work of Austin and Tao~\cite{AT10}, who termed such algorithms~``repair algorithms".) Reconstruction has since been studied for a number of properties, including monotone functions~\cite{ACCL08,SS10,BGJJRW12}, hypergraph properties~\cite{AT10}, convexity~\cite{CS06}, expanders~\cite{KPS13}, Lipschitz functions~\cite{JR13}, graph connectivity and diameter~\cite{CGR13}, and error correcting codes~\cite{CFM14}. Property reconstruction falls within the {\sl local computation algorithms} framework of Rubinfeld, Tamir, Vardi, and Xie~\cite{RTVX11}.
The paper of Blanc, Gupta, Lange, and Tan~\cite{BGLT-NeurIPS2}
designs a decision tree learning algorithm that is amenable to {\sl learnability estimation}~\cite{KV18,BH18}: given a training set $S$ of {\sl unlabeled} examples, the performance of this algorithm $\mathcal{A}$ trained on $S$---that is, the generalization error of the hypothesis that $\mathcal{A}$ would construct if we were to label all of $S$ and train $\mathcal{A}$ on it---can be accurately estimated by labeling only a small number of the examples in $S$. Their techniques can be used to derive a reconstruction algorithm that achieves guarantees similar to those in~\Cref{thm:reconstruct}, but only for {\sl monotone} functions~$f$. This limitation is inherent: as noted in~\cite{BGLT-NeurIPS2}, their algorithm is fails for non-monotone functions.
\subsubsection{The work of Bhsouty and Haddad-Zaknoon}
\label{sec:Bshouty}
Subsequent to the posting of our work to the ArXiv, Bshouty and Haddad-Zaknoon~\cite{BHZ21} have given a tester that is closely related, but incomparable, to~\Cref{cor:main}. Their tester:
\begin{itemize}
\item[$\circ$] Makes $\poly(s,1/\eps)$ queries to $f$, runs in $\poly(n,1/\eps)$ time, and
\item[$\circ$] Accepts w.h.p.~if $f$ is exactly a size-$s$ decision tree;
\item[$\circ$] Rejects w.h.p.~if $f$ is $\eps$-far from size-$(s/\eps)^{O(\log(s/\eps))}$ decision trees
\end{itemize}
Comparing~\cite{BHZ21}'s tester to ours, their query complexity is independent of~$n$ (whereas ours has a $\log n$ dependence), and the size of decision trees in their reject condition is only $(s/\eps)^{O(\log(s/\eps))}$ (whereas we require $s^{O((\log s)^2/\eps^3)}$).
On the other hand, our tester is tolerant and has query complexity that achieves a polylogarithmic instead of polynomial dependence on $s$. Furthermore,~\cite{BHZ21} does not give a reconstruction algorithm, \violet{while that is the main contribution of our work.}
\subsection{Future directions}
\label{sec:future}
We list a few concrete avenues for future work suggested by our results:
\begin{itemize
\item[$\circ$] {\sl Tighter connections between testing and learning:} Our tester rejects functions that are $\Omega(\eps)$-far from $\mathrm{quasipoly}(s)$ decision trees, and~\Cref{thm:lower-bound-intro} shows that a tester that rejects functions that are $\Omega(\eps)$-far from size-$s$ decision trees would yield a comparably efficient algorithm for properly learning decision trees. A concrete avenue for future work is to narrow this gap between $\mathrm{quasipoly}(s)$ and $s$, with the ultimate goal of getting them to match.
There are also other ways in which \Cref{thm:lower-bound-intro} could be strengthened: Do {\sl non-tolerant} testers for decision trees yield proper learning algorithms? Do tolerant testers yield proper learning algorithms with {\sl agnostic} guarantees?
\item[$\circ$] {\sl Improved reconstruction algorithms and testers for other properties:} The reconstruction algorithms that we obtain for the properties listed in~\Cref{table:other-measures} follow by combining~\Cref{thm:reconstruct} with known relationships between these measures and decision tree complexity. It would be interesting to obtain improved parameters by designing reconstruction algorithms that are tailored to each of these properties, without going through decision trees.
The same questions can be asked of property testers, and about properties that are not known to be quantitatively related to decision tree size. Can we achieve similar exponential improvements in the time and query complexities of non-parameterized testers by relaxing to the parameterized setting? \Cref{thm:lower-bound-intro} could be viewed as suggesting that for certain properties, efficient algorithms may only be possible in the parameterized setting.
\end{itemize}
Finally, we mention that there remains a large gap in the known bounds on the query complexity of non-tolerant testing of decision trees in the non-parameterized setting: the current best upper bound is~$\tilde{O}(s)$~\cite{Bsh20,CGM11} whereas the current best lower bound is $\Omega(\log s)$~\cite{DLMORSW07,BBM12}. It would be interesting to explore whether our techniques could be useful in closing this exponential gap.
\paragraph{Notation.} All probabilities and expectations are with respect to the uniform distribution unless otherwise stated; we use boldface (e.g.~$\bx$) to denote random variables. For two functions $f,g : \bn\to\bits$, we write $\dist(f,g)$ to denote the quantity $\Pr[f(\bx)\ne g(\bx)]$. We say that $f$ and $g$ are {\sl $\eps$-close} if $\Pr[f(\bx)\ne g(\bx)] \le \eps$, and {\sl $\eps$-far} otherwise.
For a function $f : \bn\to\bits$, a decision tree $T$ over the same variables as $f$, and a node $v$ in $T$, we write $f_v$ to denote the subfunction of $f$ obtained by restricting $f$ according to the root-to-$v$ path in $T$. We write $|v|$ to denote the depth of $v$ within $T$, and so the probability that a uniform random $\bx\sim \bn$ reaches $v$ is $2^{-|v|}$.
\section{Proofs of~\Cref{thm:reconstruct} and~\Cref{cor:main}}
Our proof of~\Cref{thm:reconstruct} has two main components:
\begin{itemize}
\item[$\circ$] A structural lemma about functions $f$ that are $\opt_s$-close to a size-$s$ decision tree $T^\star$. While we have no information about the structure of this tree $T^\star$ that $f$ is $\opt_s$-close to, we will show that $f$ is $O(\opt_s + \eps)$-close to a tree $T^\diamond$ of size $S = S(s,\eps)$ with a very specific structure.
\item[$\circ$] An algorithmic component that leverages this specific structure of $T^\diamond$ to show that for any input $x \in\bn$, the value of $T^\diamond(x)$ can be computed with only $\log S \cdot \log n$ queries to $f$.
\end{itemize}
\Cref{sec:structural} will be devoted to the structural lemma and~\Cref{sec:algorithmic} to the algorithmic component. We prove~\Cref{thm:reconstruct} in~\Cref{sec:proof-of-reconstruct}, and we derive~\Cref{cor:main} as a simple consequence of~\Cref{thm:reconstruct} in~\Cref{sec:proof-of-tester}.
\subsection{Structural component of~\Cref{thm:reconstruct}}
\label{sec:structural}
\begin{definition}[Noise sensitivity]
The {\sl noise sensitivity of $f : \bn\to\bits$ at noise rate $p$} is the quantity
\[ \NS_p(f) \coloneqq \Pr[f(\bx) \ne f(\by)],\]
where $\bx\sim \bn$ is uniform random and $\by\sim_p\bx$ is a {\sl $p$-noisy copy of $\bx$}, obtained from $\bx$ by independently rerandomizing each coordinate with probability $p$.
\end{definition}
We assign each coordinate $i\in [n]$ of a function $f$ a {\sl score}, which measures the expected decrease in the noise sensitivity of $f$ if $x_i$ is queried:
\begin{definition}[Score of a variable]
\label{def:score}
Given a function $f:\bn \to \bits$, noise rate $p \in (0,1)$, and coordinate $i \in [n]$, the {\sl score} of $x_i$ is defined as
\[ \mathrm{Score}_i(f, p) = \NS_{p}(f) - \Ex_{\bb\in\bits}\big[\NS_{p}(f_{x_i = \bb})\big].
\]
\end{definition}
(Our notion of score is equivalent, up to scaling factors depending on $p$, to the notion of ``noisy influence" as in defined in O'Donnell's monograph~\cite{ODbook}. We use our definition of score as it simplifies our presentation.) We are now ready to define the tree $T^\diamond$ described at the beginning of this section and state our structural lemma.
\begin{definition}
\label{def:top-down-tree}
For a function $f : \bn\to\bits$, parameters $d\in \N$ and $p \in (0,1)$, we write $T^{d,p}_f$ to denote the complete decision tree of depth $d$ defined as follows:
\begin{itemize}
\item[$\circ$] At every internal node $v$, query $x_i$ where $i\in [n]$ maximizes $\mathrm{Score}_i(f_v,p)$.\footnote{Ties are arbitrarily broken; our results hold regardless of how ties are broken.}
\item[$\circ$] Label every leaf $\ell$ with $\sign(\E[f_\ell])$.
\end{itemize}
\end{definition}
\begin{lemma}[Structural lemma]
\label{lemma:BGLT}
Let $f : \bn\to\bits$ be $\opt_s$-close to a size-$s$ decision tree. Then for $d = O((\log s)^3/\eps^3)$ and $p = \eps/(\log s)$, we have $\dist(f,T^{d,p}_f) \le O(\opt_s)+\eps$.
\end{lemma}
\section{Proof of~\Cref{lemma:BGLT}}
\label{appendix}
\paragraph{Noise-sensitivity-based potential function.} First, we introduce the potential function that will facilitate our proof of~\Cref{lemma:BGLT}. Every decision tree $T$ naturally induces a distribution over its leaves where each leaf $\ell$ receives weight $2^{-|\ell|}$. We write $\bell\sim T$ to denote a draw of a leaf of $T$ according to this distribution.
\begin{definition}[Noise sensitivity of $f$ with respect to a tree $T$]
\label{def:NS-wrt-T}
Let $f :\bn\to\bits$ be a function, $p\in (0,1)$, and $T$ be a decision tree. The {\sl noise sensitivity of $f$ at noise rate $p$ with respect to $T$} is the quantity
\[ \NS_p(f,T) \coloneqq \Ex_{\bell\sim T} \big[ \NS_p(f_{\bell})\big].\]
\end{definition}
Note that if $T$ is the empty tree, then $\NS_p(f,T)$ is simply $\NS_p(f)$, the noise sensitivity of $f$ at noise rate $p$. The following proposition is a bound on $\NS_p(f)$ that takes into account its distance from a small decision tree:
\begin{proposition}[Noise sensitivity of $f$]
\label{prop:NS-of-f}
Let $f : \bn\to\bits$ be $\opt_s$-close to a size-$s$ decision tree $T$. For all $p\in (0,1)$, we have $\NS_p(f) \le p\log s + 2\,\opt_s.$
\end{proposition}
\begin{proof}
Let $\bx\sim\bn$ be uniform random, $\by\sim_p \bx$ be a $p$-noisy copy of $\bx$, and $\bx^{\oplus i}$ denote $\bx$ with its $i$-th coordinate flipped. We first observe that
\begin{align*}
\NS_p(f) &= \Pr[f(\bx)\ne f(\by)] \\
&\le \Pr[f(\bx)\ne T(\bx)] + \Pr[T(\bx) \ne T(\by)] + \Pr[T(\by) \ne f(\by)] \\
&= \NS_p(T) + 2\,\opt_s.
\end{align*}
To bound $\NS_p(T)$, we use the inequality $\NS_p(T) \le p\cdot \Inf(T)$ where $\Inf(T) \coloneqq \sum_{i=1}^n \Pr[f(\bx) \ne f(\bx^{\oplus i})]$ is the total influence of $T$~\cite[Exercise 2.42]{ODbook}, along with the bound $\Inf(T) \le \log s$ (see e.g.~\cite{OS07}).
\end{proof}
We prove~\Cref{lemma:BGLT} by quantifying the difference between $\NS_p(f,T^{j+1,p}_f)$ and $\NS_p(f,T^{j,p}_f)$: we show that for every $j\in\N$, either $\dist(f,T^{j,p}_f)\le O(\opt_s + \eps)$ or it must be the case that $\NS_p(f,T^{j+1,p}_f)$ is significantly smaller than $\NS_p(f,T^{j,p}_f)$. Since $\NS_p(f,T) \ge 0$ for all trees~$T$, the second case can only happen so many times before we fall into the first case.
We will need the following result from~\cite{OSSS05}:
\begin{theorem}[Theorem 3.2 of~\cite{OSSS05}]
\label{thm:OSSS-non-buggy}
Let $T : \bn \to \bits$ be a decision tree. For all functions $g : \bn\to \R$, writing $\bx,\bx'\sim\bn$ to denote uniform random and independent inputs and $\bx^{\sim i}$ to denote $\bx$ with its $i$-th coordinate rerandomized,
\[ \mathrm{CoVr}(T,g) \le \sum_{i=1}^n \lambda_i(T) \cdot \Ex_{\bx} \big[|g(\bx)-g(\bx^{\sim i})|\big],\]
where
\begin{align*}
\mathrm{CoVr}(T,f) &\coloneqq \Ex_{\bx,\bx'}\big[|T(\bx)-g(\bx')|\big] - \Ex_{\bx}\big[|T(\bx)-g(\bx)|\big], \\
\lambda_i(T) &\coloneqq \Pr[\,\text{$T$ queries $\bx_i$}\,].
\end{align*}
\end{theorem}
By applying~\Cref{thm:OSSS-non-buggy} to a suitably smoothened version of $f$, we are able to derive a lower bound on the score of the highest-scoring variable of $f$.
\begin{definition}[$p$-smoothed version of $f$]
For $f : \bn\to\bits$ and $p\in (0,1)$, the {\sl $p$-smoothed version of $f$} is the function $\tilde{f}^{(p)} : \bn\to [-1,1]$,
\[ \tilde{f}^{(p)}(x) = \Ex_{\by\sim_p x}[f(\by)] = \sum_{S\sse [n]} (1-p)^{|S|} \wh{f}(S)\prod_{i\in S}x_i,\]
where the $\wh{f}(S)$ is the $S$-th Fourier coefficients of $f$. When $p$ is clear from context, we write $\tilde{f}$.
\end{definition}
\begin{lemma}[Score of the highest-scoring variable]
\label{lem:noisy-OSSS-truncated-agnostic-fixed}
Let $f : \bits^n \to \bits$ be a function, $p\in (0,1)$, and $\tilde{f} = \tilde{f}^{(p)}$ be its $p$-smoothed version. For all size-$s$ decision trees $T : \bn\to \bits$,
\[ \max_{i\in [n]}\big\{\sqrt{\mathrm{Score}_i(f,p)}\big\} \ge \frac{\sqrt{p}}{\log s} \cdot \Big(\lfrac1{2}\Var(\tilde{f}) - \E\big[|T(\bx)-\tilde{f}(\bx)|\big]\Big).\]
\end{lemma}
\begin{proof}
Applying~\Cref{thm:OSSS-non-buggy} with `$g$' being the $p$-smoothed version $\tilde{f}$ of $f$, we have
\begin{equation}
\label{eq:OSSS-non-buggy-for-us} \mathrm{CoVr}(T,\tilde{f}) \le \sum_{i=1}^n \lambda_i(T) \cdot \E[|\tilde{f}(\bx)-\tilde{f}(\bx^{\sim i})|].\end{equation}
We first lowerbound the LHS of~\Cref{eq:OSSS-non-buggy-for-us}. For $\bx,\bx'\sim \bn$ uniform and independent,
\begin{align}
\mathrm{CoVr}(T,\tilde{f}) &= \E\big[|T(\bx)-\tilde{f}(\bx')|\big] - \E\big[|T(\bx)-\tilde{f}(\bx)|\big] \tag*{(Definition of $\mathrm{CoVr}$)} \nonumber \\
&\ge \E\big[|\tilde{f}(\bx)-\tilde{f}(\bx')|\big] - 2\E\big[|T(\bx)-\tilde{f}(\bx)|\big] \tag*{(Triangle inequality)}\nonumber \\
&\ge \lfrac1{2}\E\big[(\tilde{f}(\bx)-\tilde{f}(\bx'))^2\big] - 2\E\big[|T(\bx)-\tilde{f}(\bx)|\big] \tag*{($\tilde{f}$ is $[-1,1]$-valued)}\nonumber \\
&\ge \Var(\tilde{f}) - 2\E\big[|T(\bx)-\tilde{f}(\bx)|\big]. \label{eq:cov-lb-fixed}
\end{align}
For a function $g : \bn\to\R$, its {\sl $i$-th discrete derivative} is the function
\[ (D_ig)(x) \coloneqq \lfrac1{2}\big(g(x^{i=1})-g(x^{i=-1})\big) = \ds\sum_{S\ni i} \wh{g}(S)\prod_{j \in S \setminus \{i\}} x_j,\]
where $x^{i=b}$ denotes $x$ with its $i$-th coordinate set to $b$. With this definition in hand, we now analyze the expectation on the RHS of~\Cref{eq:OSSS-non-buggy-for-us}. By Jensen's inequality, \[ \E[|\tilde{f}(\bx)-\tilde{f}(\bx^{\sim i})|]^2 \le \Ex_{\bx}\big[ (\tilde{f}(\bx)-\tilde{f}(\bx^{\sim i}))^2\big] = \lfrac1{2}\ds\Ex_{\bx}\big[ (\tilde{f}(\bx^{i=1})-\tilde{f}(\bx^{i=-1}))^2\big] = 2\Ex_{\bx}\big[D_i\wt{f}(\bx)^2\big]. \]
Applying Plancherel's identity twice,
\[ \Ex_{\bx}\big[ D_i\tilde{f}(\bx)^2\big] = \sum_{S \ni i} (1-p)^{2|S|}\wh{f}(S)^2 \le \sum_{S\ni i} (1-p)^{|S|}\wh{f}(S)^2 = \Ex_{\bx,\by}\big[ D_if(\bx)D_if(\by)\big] \]
where $\by\sim_p\bx$ is a $p$-noisy copy of $\bx$. It follows from a straightforward calculation~\cite[Lemma 3.2]{BGLT-NeurIPS1} that
\[ \Ex_{\bx,\by}\big[ D_if(\bx)D_if(\by)\big] = \frac{2\cdot \mathrm{Score}_i(f,p)}{p}.\]
Therefore, combining the three equations above we have shown that
\begin{equation} \label{eq:score-non-negative} \Ex_{\bx}\big[ |\tilde{f}(\bx)-\tilde{f}(\bx^{\sim i})|\big] \le \sqrt{\frac{4\cdot \mathrm{Score}_i(f,p)}{p}}.
\end{equation}
Plugging this inequality into the RHS of~\Cref{eq:OSSS-non-buggy-for-us},
\begin{align}
\sum_{i=1}^n \lambda_i(T) \cdot \ds\Ex_{\bx}\big[|\tilde{f}(\bx)-\tilde{f}(\bx^{\sim i})|\big] &\le \sum_{i=1}^n \lambda_i(T) \cdot \sqrt{\frac{4\cdot \mathrm{Score}_i(f,p)}{p}} \nonumber \\
&\le \ds \max_{i\in [n]} \big\{ \sqrt{\mathrm{Score}_i(f,p)}\big\} \cdot \frac{2}{\sqrt{p}} \cdot \sum_{i=1}^n \lambda_i(T) \nonumber \\
&\le \max_{i\in [n]} \big\{ \sqrt{\mathrm{Score}_i(f,p)}\big\} \cdot \frac{2\log s}{\sqrt{p}}, \label{eq:RHS-ub-fixed}
\end{align}
where the final inequality holds because
\[ \sum_{i=1}^n \lambda_i(T) = \sum_{i=1}^n \Pr[\,\text{$T$ queries $\bx_i$}\,] = \Ex_{\bell\sim T}\big[|\bell|\big] \le \log s. \]
The lemma follows by combining~\Cref{eq:OSSS-non-buggy-for-us,eq:cov-lb-fixed,eq:RHS-ub-fixed}.
\end{proof}
\subsubsection{Proof of~\Cref{lemma:BGLT}}
Let $T^\star$ be the size-$s$ decision tree that $f$ is $\opt_s$-close to.
Fix $j\in \N$ and consider the tree $T^{j,p}_f$. We have that:
\begin{align*}
\NS_p(f, T^{j+1,p}_f) &\le \NS_p(f, T^{j,p}_f) - \Ex_{\bell\sim T^{j,p}_f}\Big[\max_{i\in [n]}\big\{\mathrm{Score}_i(f_{\bell}, p)\big\}\Big] \tag*{(\Cref{def:score})} \\
&\le \NS_p(f, T^{j,p}_f) - \bigg(\Ex_{\bell\sim T^{j,p}_f}\Big[\max_{i\in [n]}\big\{\sqrt{\mathrm{Score}_i(f_{\bell}, p)}\big\}\Big]\bigg)^2. \tag*{(Jensen's inequality)}
\end{align*}
Recall that we write $\bell\sim T$ to denote a draw of a leaf of $T$ where each leaf $\ell$ receives weight $2^{-|\ell|}$.
We consider two cases:\medskip
\noindent {\bf{Case 1:}} $\Ex_{\bell \sim T^{j,p}_f}[\Var(\wt{f_{\bell}})] \ge 2\,(\Ex_{\bell,\bx}\big[|{T}^\star(\bx)-\wt{f_{\bell}}(\bx)|\big] + \eps)$.\medskip
In this case we apply~\Cref{lem:noisy-OSSS-truncated-agnostic-fixed} to each leaf $\ell$ of $T^{j,p}_f$ to get that
\begin{align*} \Ex_{\bell\sim T^{j,p}_f}\Big[\max_{i\in [n]}\big\{\sqrt{\mathrm{Score}_i(f_{\bell}, p)}\big\}\Big] &\ge \frac{\sqrt{p}}{\log s} \cdot \Ex_{\bell\sim T^{j,p}_f}\Big[\lfrac1{2}\Var(\tilde{f_{\bell}}) - \E\big[|T^\star(\bx)-\tilde{f_{\bell}}(\bx)|\big]\Big] \\
&\ge \frac{\eps\sqrt{p}}{\log s},
\end{align*}
and hence
\begin{align*} \NS_p(f,T^{j+1,p}_f) &\le \NS_p(f,T^{j,p}_f) - \frac{\eps^2 p}{(\log s)^2}. \\
&= \NS_p(f,T^{j,p}_f)-\frac{\eps^3}{(\log s)^3}. \tag*{(Our choice of $p = \eps/(\log s)$)}
\end{align*}
\medskip
\noindent {\bf{Case 2:} $\Ex_{\bell \sim T^{j,p}_f}[\Var(\wt{f_{\bell}})] < 2\,(\Ex_{\bell,\bx}\big[|{T}^\star (\bx)-\wt{f_{\bell}}(\bx)|\big] + \eps$).}\medskip
In this case we claim that $\dist(f,T^{j,p}_f) \le O(\opt_s + \eps)$. We will need a couple of simple propositions:
\begin{proposition}
\label{prop:NS-of-leaves}
$\Ex_{\bell,\bx}[(\wt{f_{\bell}}(\bx)-f_\ell(\bx))^2] \le 4\,\NS_p(f)$.
\end{proposition}
\begin{proof}
Since $f_\ell$ and $\wt{f_\ell}$ are $[-1,1]$-valued, we have that
\begin{align*}
\Ex_{\bell,\bx}\big[(\wt{f_{\bell}}(\bx)-f_{\bell}(\bx))^2\big] &\le 2 \Ex_{\bell,\bx}\big[|\wt{f_{\bell}}(\bx)-f_{\bell}(\bx)|\big] \\
&= 2 \Ex_{\bell} \Bigg[ \mathop{\Ex_{\bx}}_{\by\sim_p\bx} \big[ | f_{\bell}(\by) - f_{\bell}(\bx) | \big]\Bigg] \\
&= 2 \Ex_{\bell} \Bigg[ \,2\mathop{\Prx_{\bx}}_{\by\sim_p\bx} \big[ f_{\bell}(\by) \ne f_{\bell}(\bx)\big] \,\Bigg] \\
&= 4 \E_{\bell} \big[ \NS_p(f_{\bell})\big] \\ &= 4\,\NS_p(f,T^{j, p}_f) \le 4\,\NS_p(f),
\end{align*}
where the final inequality is a consequence of the fact that score is a nonnegative quantity (\Cref{eq:score-non-negative}).
\end{proof}
\begin{proposition}
\label{prop:rounding-error}
For any function $g : \bn\to\bits$ and constant $c\in \R$,
\[ \E\big[(g(\bx) - \sign(\E[g]))^2\big] \le 2\E[(g(\bx)-c)^2]. \]
\end{proposition}
\begin{proof}
Let $a \coloneqq \Pr[g(\bx) = 1]$ and assume without loss of generality that $a \geq \frac1{2}$. On one hand, we have that $\E\big[(g(\bx) - \sign(\E[g])^2\big] = \E\big[(g(\bx) - 1)^2\big] = 4(1-a)$. On the other hand, since
\[ \E[(g(\bx)-c)^2] = a(1-c)^2+(1-a)(1+c)^2 \]
this quantity is minimized for $c = 2a-1$ and attains value $4a(1-a)$ at this minimum. Therefore indeed
\[ \min_{c\in \R} \big\{ \E[(g(\bx)-c)^2]\big\} = 4a(1-a) \ge 2(1-a) = \lfrac{1}{2}\E\big[(g(\bx) - \sign(\E[g]))^2\big] \]
and the proposition follows.
\end{proof}
With~\Cref{prop:NS-of-leaves,prop:rounding-error} in hand, we now bound $\dist(f,T^{j,p}_f)$:
\begin{align*}
\dist(f,T^{j,p}_f) &= \Ex_{\bell\sim T^{j,p}_f}\big[ \dist(f_{\bell},\sign(\E[f_{\bell}]))\big] \\
&= \lfrac1{4} \ds \Ex_{\bell,\bx}\big[ (f_{\bell}(\bx) - \sign(\E[f_{\bell}]))^2 \big] \\
&\le \lfrac1{2} \ds \Ex_{\bell,\bx} \big[ (f_{\bell}(\bx) - \E[\wt{f_{\bell}}])^2\big] \tag*{(\Cref{prop:rounding-error})} \\
&\le \Ex_{\bell,\bx}\big[ (f_{\bell}(\bx)-\wt{f_{\bell}}(\bx))^2 \big] + \Ex_{\bell,\bx}\big[ (\wt{f_{\bell}}(\bx) - \E[\wt{f_{\bell}}])^2 \big] \tag*{(``almost-triangle" inequality)} \\
&\le 4\,\NS_p(f) + \Ex_{\bell}[\Var(\wt{f_{\bell}})]. \tag*{(\Cref{prop:NS-of-leaves})}
\end{align*}
By the assumption that we are in Case 2,
\begin{align*}
\Ex_{\bell}[\Var(\wt{f_{\bell}})] &< 2\Ex_{\bell,\bx}\big[|{T}^\star (\bx)-\wt{f_{\bell}}(\bx)|\big] + 2\,\eps \\
&\le 2\Big(\Ex_{\bell,\bx}\big[|{T}^\star(\bx) - f_{\bell}(\bx)|\big] + \Ex_{\bell,\bx}\big[ |f_{\bell}(\bx) - \wt{f_{\bell}}(\bx)| \big] \Big) + 2\,\eps \tag*{(Triangle inequality)} \\
&\le O\big(\Ex_{\bell}[\dist(f_{\bell},{T}^\star)] +\NS_p(f) \big) + 2\,\eps \tag*{(\Cref{prop:NS-of-leaves})} \\
&\le O \big(\opt_s + \eps + \NS_p(f)\big) \tag*{($\dist(f,{T}^\star)= \opt_s$)} \\
&\le O(\opt_s + p\log s + \eps) \tag*{(\Cref{prop:NS-of-f})} \\
&= O(\opt_s + \eps). \tag*{(Our choice of $p = \eps/\log s$)}
\end{align*}
Summarizing what we have shown through Cases 1 and 2, for all $j\in \N$, we either have
\[ \NS_p(f,T^{j+1,p}_f) \le \NS_p(f,T^{j,p}_f)-\frac{\eps^3}{(\log s)^3} \]
or it must be the case that $\dist(f,T^{j,p}_f) \le O(\opt_s + \eps)$. Since $\NS_p(f,T) \in [0,1]$ for all decision trees~$T$, we must fall into the latter case for some $j \le O((\log s)^3/\eps^3)$. Finally, since $\dist(f, T^{j+1,p}_f) \le \dist(f, T^{j,p}_f)$ for all $j\in \N$, we conclude that $\dist(f,T^{d,p}_f) \le O(\opt_s + \eps)$ for our choice of $d = O((\log s)^3/\eps^3)$, and~\Cref{lemma:BGLT} follows.
\violet{ \begin{remark}
\label{rem:differences}
A similar structural lemma \violet{was claimed in~\cite{BGLT-NeurIPS1}}. Their proof, however, relies crucially on a sophisticated result (the ``two function OSSS inequality for semi-metrics" from~\cite{OSSS05}) that was subsequently shown to be false~\cite{Qia21,OD21}. Our proof of~\Cref{lemma:BGLT} does not use this erroneous result.
There are also important differences between our setting and that of~\cite{BGLT-NeurIPS1}'s. \cite{BGLT-NeurIPS1} analyzes a tree, call it~$\Upsilon$, that is analogous to our $T^{d,p}_f$. Unlike $T^{d,p}_f$, their tree $\Upsilon$ is not necessarily complete: it is iteratively constructed in a top-down manner, where in each iteration the size of the tree grows by one. In each iteration, the leaf $\ell$ in the current tree with the highest ``value" is replaced with a query the variable of $f_\ell$ with the highest score, where the ``value" of a leaf is defined to be the score of the highest-scoring variable of $f_\ell$ normalized by $\ell$'s depth in the current tree.~\cite{BGLT-NeurIPS1}'s definition of score differs from ours: for their intended application, it was important that the score of a variable can be efficiently estimated to high accuracy from random labeled examples $(\bx,f(\bx))$ where $\bx \sim \bn$ is uniform random;~\Cref{def:score} does not lend itself to such an estimation procedure.
\end{remark} }
\begin{remark}
\label{remark:BGLT-approx}
\Cref{lemma:BGLT} concerns the tree $T^{d,p}_f$ as defined in~\Cref{def:top-down-tree}, where each internal node~$v$ of $T^{d, p}_f$ is a query the variable $x_i$ that maximizes $\mathrm{Score}_i(f_v,p)$. For the algorithmic component of~\Cref{thm:reconstruct}, we will need a robust version of~\Cref{lemma:BGLT}. An inspection of its proof shows that the same statement holds for any tree where each internal node $v$ is a query to a variable of {\sl approximately} maximal score, within $\tau \coloneqq O(\eps^3/(\log s)^3)$ of $\max_{j\in [n]} \mathrm{Score}_j(f_v,p)$. Indeed, the only change to the proof will be that for all $j\in \N$, we either have that
\[ \NS_p(f,T^{j+1,p}_f) \le \NS_p(f,T^{j,p}_f)-\frac{\eps^3}{(\log s)^3} + \tau \]
or it must be the case that $\dist(f,T^{j,p}_f) \le O(\opt_s + \eps)$. Therefore, as long as $\tau \le O(\eps^3/(\log s)^3)$ the conclusion is unaffected. Similarly, instead of labeling every leaf $\ell$ with $\sign(\E[f_\ell])$, the same conclusion holds if we only require this for leaves $\ell$ such that $|\E[f_\ell]| > \eps$.
\begin{lemma}[Robust version of~\Cref{lemma:BGLT}]
\label{lem:BGLT-approx}
Let $f : \bn\to\bits$ be $\opt_s$-close to a size-$s$ decision tree. For $d = O((\log s)^3/\eps^3)$, $p = \eps/(\log s)$, and $\tau = O(\eps^3/(\log s)^3)$, let~$T$ be any complete decision tree of depth $d$ satisfying:
\begin{itemize}
\item[$\circ$] At every internal node $v$, the variable $x_i$ that is queried at this node satisfies:
\[ \mathrm{Score}_{i}(f_v,p) \geq \max_{j \in [n]}\ \{\mathrm{Score}_j(f_v,p)\} - \tau.\]
\item[$\circ$] Every leaf $\ell$ such that $|\E[f_\ell]|> \eps$ is labeled $\sign(\E[f_\ell])$.
\end{itemize}
Then $\dist(f,T) \le O(\opt_s +\eps)$.
\end{lemma}
\end{remark}
\section*{Acknowledgements}
We thank the anonymous reviewers for their thoughtful comments and feedback.
Guy and Li-Yang are supported by NSF CAREER Award 1942123. Jane is supported by NSF Award CCF-2006664.
| 1e2117cf87d0ea9751c3a5d4565e5c91e39c430a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Contribution from the $e_g$-orbitals}
We quickly discuss the nonnegligible contribution from
the $e_g$-orbitals. In this section, we ignore a spin index
for simplicity. In the basis set for the $d$-orbitals with
\begin{equation}
\bm{d}^\dagger = \left(d_{yz}^\dagger, d_{xz}^\dagger, d_{xy}^\dagger, d_{3z^2-r^2}^\dagger, d_{x^2-y^2}^\dagger \right),
\end{equation}
the angular momentum operator $\bm{L}$ can be written as
follows:
\begin{equation}
L^x = \bm{d}^\dagger \begin{pmatrix}
0 & 0 & 0 & -\sqrt{3}i & -i \\
0 & 0 & i & 0 & 0 \\
0 & -i & 0 & 0 & 0 \\
\sqrt{3}i & 0 & 0 & 0 & 0 \\
i & 0 & 0 & 0 & 0
\end{pmatrix}\bm{d},\quad
L^y = \bm{d}^\dagger \begin{pmatrix}
0 & 0 & -i & 0 & 0 \\
0 & 0 & 0 & \sqrt{3}i & -i \\
i & 0 & 0 & 0 & 0 \\
0 & -\sqrt{3}i & 0 & 0 & 0 \\
0 & i & 0 & 0 & 0
\end{pmatrix}\bm{d},\quad
L^z = \bm{d}^\dagger \begin{pmatrix}
0 & i & 0 & 0 & 0 \\
-i & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 2i \\
0 & 0 & 0 & 0 & 0 \\
0 & 0 & -2i & 0 & 0
\end{pmatrix}\bm{d}.
\end{equation}
Though the up-left $3 \times 3$ part coincides with $-\bm{l}_\textrm{eff}$,
there are offdiagonal elements in $\bm{L}$~\cite{Stamokostas2018}.
Thus, the product of the two has a nontrivial contribution from
the $e_g$-orbitals.
\section{Computation of the anisotropy parameter}
In a nematic Kitaev spin liquid (NKSL),
the anisotropy parameter $\eta$ can be computed as follows.
First, we consider the case without a magnetic field.
In this case, the Hamiltonian is $\mathcal{H}_\textrm{NKSL}$
in the main text and the ground state is exactly solvable
by introducing Majorana fermions~\cite{Kitaev2006}.
\begin{equation}
\ket{\Psi_\textrm{GS}}=P_\textrm{phys}\ket{\Psi},
\end{equation}
where $P_\textrm{phys}$ is a projection onto the physical
subspace, and $\ket{\Psi}$ is the ground state of the itinerant
Majorana fermions $c_i$. Thus,
\begin{equation}
\langle ic_i c_j \rangle = \braket{\Psi_\textrm{GS}|ic_i c_j|\Psi_\textrm{GS}} = \braket{\Psi|P_\textrm{phys} ic_i c_j P_\textrm{phys}|\Psi}
\end{equation}
Meanwhile, $\mathcal{H}_\textrm{NKSL}$ obeys
\begin{equation}
\braket{\Psi_\textrm{GS}|\mathcal{H}_\textrm{NKSL}|\Psi_\textrm{GS}}
=\sum_{\langle ij\rangle \in \gamma} \frac{K^\gamma}{4} \braket{\Psi|P_\textrm{phys} ic_i c_j P_\textrm{phys}|\Psi},
\end{equation}
assuming that the ground state is in the 0-flux sector.
Thus, from the Hellmann-Feynman theorem $\langle ic_i c_j \rangle$
for the $\gamma$-direction can be computed by
\begin{equation}
\langle ic_i c_j \rangle_\gamma = \frac{4}{N_\textrm{unit}} \frac{\partial}{\partial K^\gamma} \braket{\Psi_\textrm{GS}|\mathcal{H}_\textrm{NKSL}|\Psi_\textrm{GS}},
\end{equation}
where $N_\textrm{unit}$ is the number of unit cells~\cite{Li2018}.
We here note that in the thermodynamic limit physical states
and unphysical states have the same energy and the same expectation
value for physical observables~\cite{Zschocke2015}. Thus, we can
ignore $P_\textrm{phys}$ and use $\ket{\Psi}$ directly to calculate
observables, as long as the thermodynamic limit can be taken
analytically.
The direct computation gives
\begin{equation}
E = \braket{\Psi|\mathcal{H}_\textrm{NKSL}|\Psi} = -\frac{1}{16\pi^2}\int_0^{2\pi}dk_1 \int_0^{2\pi}dk_2 \varepsilon_k,
\end{equation}
where $\varepsilon_k = |K^x e^{ik_1} + K^y e^{ik_2} + K^z|$.
Thus,
\begin{align}
\langle ic_i c_j \rangle_x &= -\frac{1}{4\pi^2}\int_\textrm{BZ} d^2k \frac{(K^x \cos k_1 + K^y \cos k_2 + K^z)\cos k_1 + (K^x \sin k_1 + K^y \sin k_2)\sin k_1}{\varepsilon_k},\\
\langle ic_i c_j \rangle_y &= -\frac{1}{4\pi^2}\int_\textrm{BZ} d^2k \frac{(K^x \cos k_1 + K^y \cos k_2 + K^z)\cos k_2 + (K^x \sin k_1 + K^y \sin k_2)\sin k_2}{\varepsilon_k},\\
\langle ic_i c_j \rangle_z &= -\frac{1}{4\pi^2}\int_\textrm{BZ} d^2k \frac{K^x \cos k_1 + K^y \cos k_2 + K^z}{\varepsilon_k},
\end{align}
where the numerical integration is done in the whole
Brillouin zone ($\textrm{BZ}$), $0 \leq k_1 \leq 2\pi$ and
$0 \leq k_2 \leq 2\pi$. The expectation values of $ic_i c_j$
are plotted in Fig.~\ref{icicj}(a). We also plot the derivative
of $\eta$ in the same strategy as shown in Fig.~\ref{derivative}.
\begin{figure}
\centering
\includegraphics[width=8.6cm]{figs1.pdf}
\caption{(a) $\langle ic_i c_j \rangle$ for each bond direction.
The definition of $\theta$ is included in the main text.
A cyan shaded region represents Kitaev's $A$ phase.
(b) Honeycomb lattice with a direction of the NNN bonds
shown in black dashed arrows.}
\label{icicj}
\end{figure}
The same strategy applies to the case without a time-reversal
symmetry. For simplicity, we only consider the case where
$K^x=K^y=K^z=K$. The Hamiltonian considered here is
\begin{equation}
\mathcal{H}_\textrm{mag} = -K\sum_{\langle ij\rangle \in \gamma} S_i^\gamma S_j^\gamma-\kappa \sum_{\langle ijk \rangle_{\alpha\beta\gamma}} S_i^\alpha S_j^\beta S_k^\gamma,
\end{equation}
where $\kappa$ is a real time-reversal breaking parameter.
$\langle ijk \rangle_{\alpha\beta\gamma}$ means nearest-neighbor (NN)
pairs, where $\langle ij \rangle$ and $\langle jk \rangle$ are on the
$\alpha$- and $\gamma$-directions, respectively, and $\beta$ is the
component which is neither $\alpha$ nor $\gamma$.
As is well-known, the $\kappa$ term is nothing but a
next-nearest-neighbor (NNN) hopping term for Majorana fermions.
Thus, we can compute the expectation value of NNN hopping terms
from the derivative about $\kappa$. The final expression becomes
\begin{align}
\langle ic_i c_j \rangle_\textrm{NN} &= -\frac{1}{12\pi^2}\int_\textrm{BZ} d^2k \frac{K|e^{ik_1}+e^{ik_2}+1|^2}{\epsilon_k}, \\
\langle ic_i c_j \rangle_\textrm{NNN} &= -\frac{1}{12\pi^2}\int_\textrm{BZ} d^2k \frac{\kappa[\sin k_1-\sin k_2 + \sin(k_2-k_1)]^2}{\epsilon_k},
\end{align}
where $\langle ic_i c_j \rangle_\textrm{NN}$ is a bond expectation value
for NN bonds, $\langle ic_i c_j \rangle_\textrm{NNN}$ is a bond
expectation value for NNN bonds, and
$\epsilon_k = \sqrt{K^2|e^{ik_1}+e^{ik_2}+1|^2 + \kappa^2[\sin k_1-\sin k_2 + \sin(k_2-k_1)]^2}$.
As for NNN bonds, the bond direction is defined as illustrated in
Fig.~\ref{icicj}(b), where the arrow points from $j$ to $i$.
These formulae will be used in the next section.
\begin{figure}
\centering
\includegraphics[width=8.6cm]{figs2new.pdf}
\caption{Derivative of $\eta$. A cyan shaded region represents
Kitaev's $A$ phase. As clearly seen, $d\eta/d\theta$ has
a cusp at the transition point $\theta = \theta_\textrm{c}$
between the $B$ phase and the $A$ phase.}
\label{derivative}
\end{figure}
\section{Magnetic field effect}.
In the main text, we have assumed the existence of a time-reversal
symmetry, but in real NMR or MS experiments an external magnetic field
is usually applied. First, we modify the ansatz Hamiltonian as follows:
\begin{equation}
\mathcal{H}_\textrm{Zeeman}=-\sum_{\langle ij\rangle \in \gamma} K^\gamma S_i^\gamma S_j^\gamma-\bm{h}\cdot\sum_i\bm{S}_i,
\end{equation}
where $\bm{h}=(h^x, h^y, h^z)$ represents the normalized external field.
In the first-order in $\bm{h}$, the wavefunction can be approximated as
\begin{equation}
\ket{\Psi^{(1)}}\sim\ket{\Psi_\textrm{GS}}-\frac{1-P_0}{-\Delta_\textrm{flux}}\bm{h}\cdot\sum_i\bm{S}_i\ket{\Psi_\textrm{GS}},
\end{equation}
where $\Delta_\textrm{flux}\sim 0.26K/4$ is a flux gap energy for
two neighboring vortices. In the first order, only the
contribution that creates or annihilates two neighboring
vortices survives. This condition holds for $S_i^x S_{i_z}^z$
and $S_i^y S_{i_z}^z$ in Eq.~(12) in the main text, for example,
which results in the following NNN hopping operators,
after summing up three surrounding bonds around the $i$th site:
\begin{align}
&P_0 q_{i} P_0 = \nonumber \\
&\begin{pmatrix}
\frac{3it_2^2}{(U + \frac{3}{2}\lambda)^2}c_i c_{i_x} & -\frac{4\Delta}{3\lambda} + \frac{4it_2^2 h^z}{3(U + \frac{3}{2}\lambda)^2\Delta_\textrm{flux}}(c_{i_z}c_{i_x} + c_{i_y}c_{i_z}) & -\frac{4\Delta}{3\lambda} + \frac{4it_2^2 h^y}{3(U + \frac{3}{2}\lambda)^2\Delta_\textrm{flux}}(c_{i_y}c_{i_z} + c_{i_x}c_{i_y}) \\
-\frac{4\Delta}{3\lambda} + \frac{4it_2^2 h^z}{3(U + \frac{3}{2}\lambda)^2\Delta_\textrm{flux}}(c_{i_z}c_{i_x} + c_{i_y}c_{i_z}) & \frac{3it_2^2}{(U + \frac{3}{2}\lambda)^2}c_i c_{i_y} & -\frac{4\Delta}{3\lambda} + \frac{4it_2^2 h^x}{3(U + \frac{3}{2}\lambda)^2\Delta_\textrm{flux}}(c_{i_z}c_{i_x} + c_{i_x}c_{i_y}) \\
-\frac{4\Delta}{3\lambda} + \frac{4it_2^2 h^y}{3(U + \frac{3}{2}\lambda)^2\Delta_\textrm{flux}}(c_{i_y}c_{i_z} + c_{i_x}c_{i_y}) & -\frac{4\Delta}{3\lambda} + \frac{4it_2^2 h^x}{3(U + \frac{3}{2}\lambda)^2\Delta_\textrm{flux}}(c_{i_z}c_{i_x} + c_{i_x}c_{i_y}) & \frac{3it_2^2}{(U + \frac{3}{2}\lambda)^2}c_i c_{i_z}
\end{pmatrix} \nonumber \\
&-\frac{t_2^2}{(U + \frac{3}{2}\lambda)^2}(ic_i c_{i_x}+ic_i c_{i_y}+ic_i c_{i_z})\openone_3.
\end{align}
All new terms are NNN hopping operators breaking the time-reversal
symmetry, so their contribution vanishes in the computation of
$\braket{\Psi_\textrm{GS}|q_{i}|\Psi_\textrm{GS}}=\braket{\Psi_\textrm{GS}|P_0 q_{i} P_0|\Psi_\textrm{GS}}$,
as long as $\ket{\Psi_\textrm{GS}}$ is time-reversal symmetric.
Thus, the magnetic field effect is negligible in the first order.
Though we have shown that the time-reversal breaking effect is
negligible in the first-order perturbation, there could be
another effect which produces a nonzero expectation value for
NNN bonds. For example, if the $\Gamma^\prime$ term coming
from the trigonal distortion is included in the magnetic Hamiltonian,
the gap opening by the time-reversal breaking can occur in
the first-order in $\bm{h}$~\cite{Takikawa2019,Takikawa2020}.
This type of effects can be included by adding a so-called
$\kappa$ term to the Hamiltonian by hand.
\begin{equation}
\mathcal{H}_\textrm{eff}=-\sum_{\langle ij\rangle \in \gamma} K^\gamma S_i^\gamma S_j^\gamma-\kappa \sum_{\langle ijk \rangle_{\alpha\beta\gamma}} S_i^\alpha S_j^\beta S_k^\gamma -\bm{h}\cdot\sum_i\bm{S}_i.
\end{equation}
The $\kappa$ term explicitly breaks
the time-reversal symmetry and $\langle ic_i c_j \rangle$ for NNN
bonds no longer vanish~\footnote{It is not trivial to show that
$\kappa$ does not depend on the bond direction even if $K^\gamma$
are not equal. See \textit{e.g.} Eq.~(9) in Ref.~\cite{Yamada2020}.}.
Again we use the first-order perturbation in $\bm{h}$ and now the
magnetic field effect becomes nontrivial. Even in the case where
$K^x=K^y=K^z$ without a nematic order, $\eta$ does not always vanish
unless the magnetic field is perpendicular to the (111) plane.
A careful treatment is necessary to detect a nematic order through
$\eta$.
\begin{figure}
\centering
\includegraphics[width=8.6cm]{figs3.pdf}
\caption{Field angle dependence of $\eta$. The magnetic field
is applied in the (111) plane with an angle $\phi$. A typical
ansatz of $\kappa/K=0.1$ and $|\bm{h}|/K=0.1$ is used.
$\Delta = 10$ meV, $\lambda = 150$ meV, $t_2 = 200$ meV,
and $U = 1.5$ eV are used.}
\label{phi}
\end{figure}
Fortunately, however, this effect is very small when we apply an
in-plane magnetic field. This is because the first-order contribution
of $\bm{h}$ for $\kappa(h^x, h^y, h^z)$ vanishes when $h^x+h^y+h^z=0$,
assuming the complete $C_3$ symmetry of the original electronic
Hamiltonian. $\kappa$ is effectively expanded from the third order
in $\bm{h}$, and we can assume that $\kappa$ to be $O(h^3)$, which
is consistent with experiments~\cite{Tanaka2020}.
We plot the angle dependence of $\eta$ with respect to an azimuth
angle $\phi$ of the applied magnetic field in Fig.~\ref{phi} in
the isotropic case $K^x=K^y=K^z=K$. As is clearly seen,
a small but nonzero $\eta$ appears due to the applied magnetic field.
It has a tiny angle dependence with a period of $120\degree$.
Thus, we can say that experiments should be done by comparing
$\phi = 0\degree$ with $120\degree$ to avoid this field effect.
To sum up, in reality $\eta$ is not exactly 0 even in the isotropic
case, so we should be careful about the actual experimental setup
to determine the critical point of the nematic transition.
Experiments should be combined with a measurement of an
anisotropy of the relaxation rate $1/T_1 T$, which is compared
between $\phi = 0\degree$ and $120\degree$, for example.
| 2750374dbae9154eab245de8776fcc4aea1b32b5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The $U_A(1)$ subgroup of approximate $U_L(3)\times U_R(3)$ chiral symmetry is anomalously broken in quantum chromodynamics (QCD). QCD is a strongly coupled theory, and as such, most accurate results can be expected to emerge through lattice simulations. These, however, lack the ability to simulate the system at finite density due to the notorius sign problem. To tackle this issue, it is common to build effective models upon chiral symmetry, which are expected to capture essential features of QCD in the low energy regime. Even at zero density, these models are known to provide reasonable results for temperatures below that of the chiral transition \cite{fukushima10}.
In effective theories, such as the Nambu--Jona--Lasinio or linear sigma models this is taken into account by `t Hooft's determinant term. Coefficients of operators in the Lagrangian of field theories, including the aforementioned determinant term, are considered to be (coupling) constants, without any field or environment dependence. In the quantum version of the action, however, fluctuations introduce temperature ($T$), baryochemical potential ($\mu_B$) and also field dependence as they become coefficient {\it functions}. When talking about field dependence of a given coupling one has in mind the resummation of higher dimensional operators that can reappear when Taylor expanding the coefficient functions in terms of the field variable(s) around a conveniently chosen expansion point.
In QCD, it is well established that the anomalous breaking of $U_A(1)$ symmetry should gradually disappear beyond the critical temperature, as at high $T$ the instanton density causing the anomaly exponentially vanishes \cite{thooft76,gross81}. At lower temperatures, however, the situation is far from being understood in a satisfactory fashion. One has also great interest in gaining results regarding the anomaly evolution at finite $\mu_B$ due to the sign problem, as mentioned earlier.
The finite temperature and/or density behavior of the $U_A(1)$ anomaly represents an active direction of research. More conservative results usually argue that the evaporation of the anomaly should follow that of the chiral condensate and thus the $U_A(1)$ symmetry restores around the critical temperature ($T_C$) of the chiral transition\cite{mitter14,cossu15,ishii16,gomez-nicola18,bottaro20}. There are also several arguments and results that indicate that it is visible even beyond $T_C$ \cite{dick15,sharma16,fejos16,li20,braun20}. For example, earlier renormalization group studies indicate that when considering a field dependent anomaly coefficient, it decreases as a function of the chiral condensate \cite{fejos16,fejos18}, and this profile function can also depend explicitly on the tempereture (though the former effect is more dominant). Effective restoration of the anomaly has, e.g., a consequence regarding the order of the chiral transition \cite{pelissetto13}, the axion mass \cite{horvatic19}, and the fate of $\eta'$ meson, whose mass if substantially drops \cite{nagahiro06,sakai13,sakai17,jido19,horvatic19b,rai20}, in a nuclear medium could lead to the formation of an $\eta'$-nucleon bound state.
The goal of this paper is to calculate the first field dependent correction to the coupling of the `t Hooft determinant in the effective action perturbatively, i.e., we determine the first Taylor coefficient of the anomaly function. Determining this coefficient (at least qualitatively) allows for obtaining the behavior of the anomaly strength as the function of the chiral condensate. This allows for providing answer for the question whether the aforementioned results on anomaly strengthening \cite{fejos16,fejos18} can be reproduced within a simple perturbative renormalizatrion group setting of the linear sigma model, or determining a full functional dependence of the effective action is necessary. Fluctuations will be included using the functional variant of the renormalization group (FRG) \cite{kopietz10}, in the so-called local potential approximation (LPA). Even though renormalizable, we think of the model as an effective field theory, therefore, an ultra violet (UV) cutoff is inherently part of the system, which we set to $\Lambda = 1 \GeV$ (we expect the linear sigma model to emerge from QCD around this scale). Our task is to integrate out all fluctuations below $\Lambda$.
The paper is organized as follows. In Section 2, we introduce the model and the corresponding method of the FRG. Section 3 is devoted for calculating the effective action and discussing the problem of the expansion point of the Taylor series. After appropriate parametrization of the model, in Section 4, as an application, we show how the anomaly strengthens at the zero temperature nuclear liquid--gas phase transition. Section 5 contains the summary.
\section{Model and method}
The model we are working with in this paper is the three flavor linear sigma model, which is defined via the following Euclidean Lagrangian:
\bea
\label{Eq:Lag}
{\cal L}&=&\Tr(\partial_i M^\dagger \partial_i M) + m^2 \Tr (M^\dagger M) + g_1 \left(\Tr(M^\dagger M)\right)^2 + g_2\Tr (M^\dagger M - \Tr (M^\dagger M)\cdot\mathbb{1})^2\nonumber\\
&&+a ( \det M + \det M^\dagger) - (h_0s_0+h_8s_8),
\eea
where $M$ contains the meson fields, $M=(s_a+i\pi_a)T_a$ [$T_a=\lambda_a/2$ are generators of the $U(3)$ group with $\lambda_a$ being the Gell-Mann matrices, $a=0,...,8$], $m^2$ is the mass parameter and $g_1$, $g_2$ refer to independent quartic couplings. As discussed in the previous section, the determinant term and the corresponding $a$ parameter is responsible for the $U_A(1)$ anomaly. We also have explicit symmetry breaking terms containing $h_0$ and $h_8$, which represent finite quark masses.
Our main goal is to calculate the quantum effective action, $\Gamma$, built upon the theory defined via (\ref{Eq:Lag}). As announced in the introduction, we think of (\ref{Eq:Lag}) as an inherently effective model, which is only valid up to the scale $\Lambda = 1 \GeV$, therefore, one needs to take into account fluctuations with a cutoff $\Lambda$. The scale dependent quantum effective action, $\Gamma_k$, which includes fluctuations with momenta larger than $k$ (i.e., they are integrated out) is defined as
\bea
Z_k[J]&=&\int {\cal D} M {\cal D}M^\dagger \exp \Big\{-\int {\cal L} - \int (J M + h.c.) - \int \int M^\dagger R_k M \Big\}, \nonumber\\
\Gamma_k[M]&=&-\log Z_k[J]-\int (J M + h.c.) - \int \int M^\dagger R_k M,
\eea
where we omitted matrix indices, $J$ is the conjugate source variable to $M$, and $R_k$ is an appropriately chosen (bilocal) regulator function freezing fluctuations with momenta smaller than $k$. We note that the integrals can be considered either in direct or Fourier spaces.
The $\Gamma_k$ functional, for homogeneous field configurations, obeys the following, so-called flow equation \cite{kopietz10}:
\bea
\label{Eq:floweq}
\partial_k \Gamma_k [M] = \frac12 \Tr \int_x \int_{q} (\Gamma_k^{(2)}+R_k)^{-1}(q)\partial_k R_k (q),
\eea
where $\Gamma_k^{(2)}$ is the second functional derivative of $\Gamma_k$ in a homogeneous background field $M$, thus the $x$ integral merely gives a spacetime volume factor. We also assumed that the regulator is diagonal in momentum space.
Our aim is to calculate the scale dependent effective action, $\Gamma_k$, in an approximation that takes into account the evolution of the anomaly at the next-to-leading order, i.e., we wish to determine in $\Gamma_k$ the coefficient of the operator $\Tr(M^\dagger M)\cdot (\det M+\det M^\dagger)$. Our ansatz for $\Gamma_k$ is as follows:
\bea
\label{Eq:ansatz}
{\Gamma}_k&=&\int_x \Big[\Tr(\partial_i M^\dagger \partial_i M) + m^2_k \Tr (M^\dagger M) + g_{1,k} \left(\Tr(M^\dagger M)\right)^2 + g_{2,k}\Tr (M^\dagger M - \Tr (M^\dagger M)\cdot\mathbb{1})^2\nonumber\\
&&+a_k ( \det M + \det M^\dagger) + a_{1,k} \Tr(M^\dagger M)\cdot (\det M + \det M^\dagger)- (h_0s_0+h_8s_8)\Big].
\eea
This is sometimes called the Local Potential Approximation (LPA), where momentum dependence is only introduced into the two point function, via the standard kinetic term in ({\ref{Eq:ansatz}). Note that the LPA can be considered as the leading order of the derivative expansion, and there is substantial evidence that these kind of series do converge \cite{kopietz10}. As seen in (\ref{Eq:ansatz}), instead of working with a completely general field dependent potential, we are employing perturbation theory in terms of the small parameter $1/\Lambda$. That is, by gradually including higher dimensional operators, since their coefficients scale with inverse powers of the scale, the ansatz (\ref{Eq:ansatz}) in the UV can be thought of as a power series in $1/\Lambda$. We choose (\ref{Eq:ansatz}) to be compatible with (\ref{Eq:Lag}), but all couplings come with $k$--dependence. The only exceptions are $h_0$ and $h_8$, as one point couplings do not flow with respect to the scale. Furthermore, notice the new term proportional to $a_{1,k}$, which is key for our purposes to determine the anomaly behavior at $k=0$. First, our task is to calculate $\Gamma_k^{(2)}$ from (\ref{Eq:ansatz}), then plug it into (\ref{Eq:floweq}), and identify the individual differential equations for $m_k^2$, $g_{1,k}$, $g_{2,k}$, $a_k$, and $a_{1,k}$. Finally, these equations need to be integrated from $k=1 \GeV$ to $k=0$ to obtain $\Gamma\equiv \Gamma_{k=0}$. In the ansatz (\ref{Eq:ansatz}), obviously the actual strength of the anomaly is not described by the parameter $a$, but rather $a+a_1\cdot \Tr(M^\dagger M)|_{\min}$, where we need to evaluate the chiral condensates in the minimum point of the effective action. Therefore, what we are basically after is the relative sign of $a_1$ to $a$ at $k=0$ to decide whether the anomaly strengthens or weakens as the chiral condensate gradually evaporates.
We finally note that there are various choices for the regulator function, $R_k$. In this paper we will stick to $R_k(q,p)\equiv R_k(q)(2\pi)^3 \delta(q+p)=(k^2-q^2)\Theta(k^2-q^2)(2\pi)^4 \delta(q+p)$, where $\Theta(x)$ is the step function. This variant has been shown to be the optimal choice for the LPA \cite{litim00}, maximizing the radius of convergence of an amplitude expansion.
\section{Calculation of the effective action}
The first step is to calculate $\Gamma_k^{(2)}$. In principle it is a $18\times 18$ matrix in the $s^a-\pi^a$ space, and there is not much hope that one can invert such a complicated expression analytically. Luckily, it is not necessary at all, as in (\ref{Eq:ansatz}) we kept the field dependence up to the order of ${\cal O}(M^5)$. By working with a restricted background, $\Gamma_k^{(2)}$ is easily invertable and by expanding the rhs of (\ref{Eq:floweq}) in terms of the field variables it still allows for identifying each operator that are being kept in (\ref{Eq:ansatz}).
A convenient choice is to work with $M=s_0T_0+s_8T_8$. In such a background, the operators that need to be identified are as follows:
\begin{subequations}
\label{Eq:operators}
\bea
\rho&:=&\Tr (M^\dagger M)=\frac12 (s_0^2+s_8^2), \\
\tau&:=&\Tr (M^\dagger M - \Tr (M^\dagger M)\cdot\mathbb{1})^2=\frac{1}{24} s_8^2 (8 s_0^2 - 4 \sqrt{2} s_0 s_8 + s_8^2), \\
\Delta&:=&\det M+\det M^\dagger=\frac{1}{36} (2 \sqrt{6} s_0^3 - 3 \sqrt{6} s_0 s_8^2 - 2 \sqrt{3} s_8^3).
\eea
\end{subequations}
The $\Gamma_k^{(2)}$ matrix elements in the scalar sector read
\bea
\Gamma_{k,s_0s_0}^{(2)}&\!\!\!=\!\!\!&q^2+m_k^2+g_{1,k}(3s_0^2+s_8^2)+\frac23g_{2,k}s_8^2+\sqrt{\frac23}a_ks_0+a_{k,1}\Big(\frac53\sqrt{\frac23}s_0^3-\frac{1}{2\sqrt6}s_0s_8^2-\frac{1}{6\sqrt3}s_8^3\Big), \\
\Gamma_{k,s_0s_8}^{(2)}&\!\!\!=\!\!\!&2g_{1,k} s_0s_8+g_{2,k}\Big(\frac43s_0s_8-\frac{1}{\sqrt2}s_8^2\Big)-\frac{a_k}{\sqrt6}s_8-a_{1,k}\Big(\frac{1}{2\sqrt6}s_0^2s_8+\frac{1}{2\sqrt3}s_0s_8^2+\frac{1}{\sqrt6}s_8^3\Big),\\
\Gamma_{k,s_8s_8}^{(2)}&\!\!\!=\!\!\!&q^2+m_k^2+g_{1,k}(s_0^2+3s_8^2)+g_{2,k}\Big(\frac23s_0^2-\sqrt2 s_0s_8+\frac12s_8^2\Big)-a_k\Big(\frac{1}{\sqrt6}s_0+\frac{1}{\sqrt3}s_8\Big)\nonumber \\
&&-a_{1,k}\Big(\frac{1}{6\sqrt6}s_0^3+\frac{1}{2\sqrt3}s_0^2s_8+\sqrt{\frac32}s_0s_8^2+\frac{5}{3\sqrt3}s_8^3\Big),\\
\Gamma_{k,s_1s_1}^{(2)}&\!\!\!=\!\!\!&\Gamma_{k,s_2s_2}^{(2)}=\Gamma_{k,s_3s_3}^{(2)}=q^2+m_k^2+g_{1,k}(s_0^2+s_8^2)+g_{2,k}\Big(\frac23s_0^2+\sqrt2 s_0s_8+\frac16s_8^2\Big)\nonumber\\
&&\hspace{2.6cm}+a_k\Big(-\frac{1}{\sqrt6}s_0+\frac{1}{\sqrt3}s_8\Big)+a_{1,k}\Big(-\frac{1}{6\sqrt6}s_0^3+\frac{1}{2\sqrt3}s_0^2s_8-\frac{1}{\sqrt6}s_0s_8^2+\frac{1}{3\sqrt3}s_8^3\Big), \nonumber\\
\eea
\bea
\Gamma_{k,s_4s_4}^{(2)}&\!\!\!=\!\!\!&\Gamma_{k,s_5s_5}^{(2)}=\Gamma_{k,s_6s_6}^{(2)}=\Gamma_{k,s_7s_7}^{(2)}=q^2+m_k^2+g_{1,k}(s_0^2+s_8^2)+g_{2,k}\Big(\frac23s_0^2- \frac{1}{\sqrt2}s_0s_8+\frac16s_8^2\Big)\nonumber\\
&&\hspace{4.1cm}-a_k\Big(\frac{1}{\sqrt6}s_0+\frac{1}{2\sqrt3}s_8\Big)\nonumber\\
&&\hspace{4.1cm}-a_{1,k}\Big(\frac{1}{6\sqrt6}s_0^3+\frac{1}{4\sqrt3}s_0^2s_8+\frac{1}{\sqrt6}s_0s_8^2+\frac{5}{12\sqrt3}s_8^3\Big),
\eea
while the pseudoscalar components are
\bea
\Gamma_{k,\pi_0\pi_0}^{(2)}&\!\!\!=\!\!\!&q^2+m_k^2+g_{1,k}(s_0^2+s_8^2)-\sqrt{\frac23}a_k s_0-a_{1,k}\Big(\frac13\sqrt{\frac{2}{3}}s_0^3+\frac12\sqrt{\frac{3}{2}}s_0s_8^2+\frac{1}{6\sqrt3}s_8^3\Big), \\
\Gamma_{k,\pi_0\pi_8}^{(2)}&\!\!\!=\!\!\!&g_{2,k}\Big(\frac23 s_0s_8-\frac{1}{3\sqrt2}s_8^2\Big)+\frac{a_k}{\sqrt6}s_8+a_{1,k}\Big(\frac{1}{2\sqrt6}s_0^2s_8+\frac{1}{2\sqrt6}s_8^3\Big), \\
\Gamma_{k,\pi_8\pi_8}^{(2)}&\!\!\!=\!\!\!&q^2+m_k^2+g_{1,k}(s_0^2+s_8^2)+g_{2,k}\Big(-\frac{\sqrt2}{3}s_0s_8+\frac16 s_8^2\Big)+a_k\Big(\frac{1}{\sqrt6}s_0+\frac{1}{\sqrt3}s_8\Big)\nonumber\\
&&+a_{1,k}\Big(\frac{5}{6}\sqrt{\frac16}s_0^3+\frac{1}{2\sqrt3}s_0^2s_8+\frac{1}{3\sqrt3}s_8^3\Big), \\
\Gamma_{k,\pi_1\pi_1}^{(2)}&\!\!\!=\!\!\!&\Gamma_{k,\pi_2\pi_2}^{(2)}=\Gamma_{k,\pi_3\pi_3}^{(2)}=q^2+m_k^2+g_{1,k}(s_0^2+s_8^2)+g_{2,k}\Big(\frac{\sqrt2}{3}s_0s_8-\frac16 s_8^2\Big)\nonumber\\
&&\hspace{2.8cm}+a_k\Big(\frac{1}{\sqrt6}s_0-\frac{1}{\sqrt3}s_8\Big)+a_{1,k}\Big(\frac{5}{6}\sqrt{\frac16}s_0^3-\frac{1}{2\sqrt3}s_0^2s_8-\frac{2}{3\sqrt3}s_8^3\Big), \\
\Gamma_{k,\pi_4\pi_4}^{(2)}&\!\!\!=\!\!\!&\Gamma_{k,\pi_5\pi_5}^{(2)}=\Gamma_{k,\pi_6\pi_6}^{(2)}=\Gamma_{k,\pi_7\pi_7}^{(2)}=q^2+m_k^2+g_{1,k}(s_0^2+s_8^2)+g_{2,k}\Big(-\frac{1}{3\sqrt2}s_0s_8+\frac56 s_8^2\Big)\nonumber\\
&&\hspace{4.3cm}+a_k\Big(\frac{1}{\sqrt6}s_0+\frac{1}{2\sqrt3}s_8\Big)+a_{1,k}\Big(\frac{5}{6}\sqrt{\frac16}s_0^3+\frac{1}{4\sqrt3}s_0^2s_8+\frac{1}{12\sqrt3}s_8^3\Big). \nonumber\\
\eea
Using that $\partial_k R_k(q)=2k\Theta(k^2-q^2)$ and $\Gamma^{(2)}_k(q)+R_k(q)=\Gamma^{(2)}_k(k)$ for $q<k$, from (\ref{Eq:floweq}) we get
\bea
\label{Eq:floweq2}
\partial_k \Gamma_k = \int_x \frac{k^5}{32\pi^2} \Tr \big(\Gamma_k^{(2)}(k)\big)^{-1}.
\eea
Plugging in the matrix elements calculated in the $M=s_0T_0+s_8T_8$ background, we can expand the rhs of (\ref{Eq:floweq2}) in terms of $s_0$ and $s_8$. After this step, using (\ref{Eq:operators}) we identify the $\rho$, $\tau$ and $\Delta$ operators as
\bea
\label{Eq:floweqansatz}
\partial_k \Gamma_k = \int_x \Big[ \partial_k m_k^2 \cdot \rho + \partial_k g_{1,k} \cdot \rho^2 + \partial_k g_{2,k} \cdot \tau + \partial_k a_k \cdot \Delta + \partial_k a_{1,k} \cdot \rho \Delta+...\Big],
\eea
where
\bea
\label{Eq:flowcoup0}
\partial_k m_k^2&\!\!\!=\!\!\!& - \frac{k^5}{32\pi^2} \frac{8(15g_{1,k}+4g_{2,k})}{3(k^2+m_k^2)^2}, \quad \partial_k g_{1,k} = \frac{k^5}{32\pi^2} \frac{8(117g_{1,k}^2+48g_{1,k}g_{2,k}+16g_{2,k}^2)}{9(k^2+m_k^2)^3}, \nonumber\\
\partial_k g_{2,k} &\!\!\!=\!\!\!& \frac{k^5}{32\pi^2} \frac{48g_{1,k}g_{2,k}+32g_{2,k}^2}{(k^2+m_k^2)^3}, \quad \partial_k a_k = \frac{k^5}{32\pi^2}\Big(\frac{8a_k(3g_{1,k}-4g_{2,k})}{(k^2+m_k^2)^3}-\frac{24a_{1,k}}{(k^2+m_k^2)^2}\Big), \nonumber\\
\partial_k a_{1,k} &\!\!\!=\!\!\!& \frac{k^5}{32\pi^2}\bigg(\frac{32a_k(-9g_{1,k}^2+6g_{1,k}g_{2,k}+2g_{2,k}^2)}{(k^2+m_k^2)^4}+\frac{16a_{1,k}(33g_{1,k}-2g_{2,k})}{3(k^2+m_k^2)^3}\bigg).
\eea
Note that we treated the anomaly as perturbation and dropped every term beyond ${\cal O}(a_k,a_{1,k})$. Introducing scale independent variables, from (\ref{Eq:flowcoup0}) one easily reproduces the well known 1-loop $\beta$ functions of the couplings in the linear sigma model \cite{pisarski84}. Our task now is to solve Eqs. (\ref{Eq:flowcoup0}) starting from $k=\Lambda\equiv 1\GeV$ to $k=0$.
Solving (\ref{Eq:flowcoup0}) all the way down to $k=0$ would require $m_k^2>0$ throughout the renormalization group flow. Since we wish to obtain phenomenologically reasonable results, the potential has to show spontaneous symmetry breaking. That is to say, when all fluctuations are integrated out, $m^2_{k=0}$ has to be negative. But then there exists a critical scale $k_{\crit}>0$, for which all denominators in (\ref{Eq:flowcoup0}) blow up and the flow equations lose their meaning. The way out is to realize is that one actually has the choice to determine the flow equations in the minimum point of the effective action, $\Gamma_k|_{s_{0,\min},s_{8,\min}}$, rather than evaluating it in a vanishing background. That is, all renormalization group flows are to be extracted at $s_{0,\min}$, $s_{8,\min}$. This way one always has a positive definite denominators and the flow equation is valid for any $k$.
One, therefore, repeats the calculations starting from (\ref{Eq:floweq2}), but this time expands only in terms of $s_8$ so that the $\rho$ dependence of the parameters can be traced via $s_0$. A long but straightforward calculation leads once again to the possibility of identifying the invariants appear in (\ref{Eq:floweqansatz}), whose coefficients now read as
\bea
\partial_k m_k^2 &\!\!\!=\!\!\!& -\frac{k^5}{32\pi^2} \Bigg[ \frac{18g_{1,k}}{(k^2+m_k^2+2g_{1,k}\rho_0)^2}+\frac{6g_{1,k}}{(k^2+m_k^2+6g_{1,k}\rho_0)^2}+\frac{16(3g_{1,k}+2g_{2,k})}{3(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^2}\nonumber\\
&&\hspace{1cm}+\frac{72g^2_{1,k}\rho_0}{(k^2+m_k^2+2g_{1,k}\rho_0)^3}
+\frac{72g^2_{1,k}\rho_0}{(k^2+m_k^2+6g_{1,k}\rho_0)^3}+\frac{64(3g_{1,k}+2g_{2,k})^2\rho_0}{9(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^3} \Bigg], \nonumber\\
\partial_k g_{1,k}&\!\!\!=\!\!\!&\frac{k^5}{32\pi^2}\Bigg[ \frac{36g_{1,k}^2}{(k^2+m_k^2+2g_{1,k}\rho_0)^3}+\frac{36g_{1,k}^2}{(k^2+m_k^2+6g_{1,k}\rho_0)^3}+\frac{32(3g_{1,k}+2g_{2,k})^2}{9(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^3}\Bigg], \nonumber\\
\partial_k g_{2,k}&\!\!\!=\!\!\!&\frac{k^5}{32\pi^2}\Bigg[ \frac{6g_{2,k}^2}{(k^2+m_k^2+2g_{1,k}\rho_0)^3}-\frac{9g_{2,k}/2}{\rho_0(k^2+m_k^2+2g_{1,k}\rho_0)^2}+\frac{3g_{2,k}(6g_{1,k}+g_{2,k})}{\rho_0(g_{2,k}-3g_{1,k})(k^2+m_k^2+6g_{1,k}\rho_0)^2}\nonumber\\
&&\hspace{1cm}\frac{30g_{2,k}^2}{(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^3}+\frac{3g_{2,k}(g_{2,k}-21g_{1,k})/2}{\rho_0(g_{2,k}-3g_{1,k})(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^2} \Bigg],\nonumber\\
\partial_k a_k &\!\!\!=\!\!\!&\frac{k^5}{32\pi^2} \Bigg[-\frac{36g_{1,k}(a_k+2a_{1,k}\rho_0)}{(k^2+m_k^2+2g_{1,k}\rho_0)^3}-\frac{18(a_k+a_{1,k}\rho_0)}{\rho_0(k^2+m_k^2+2g_{1,k}\rho_0)^2}-\frac{12g_{1,k}(3a_k+10a_{1,k}\rho_0)}{(k^2+m_k^2+6g_{1,k}\rho_0)^3}\nonumber\\
&&\hspace{1cm}-\frac{6a_k+10a_{1,k}\rho_0}{\rho_0(k^2+m_k^2+6g_{1,k}\rho_0)^2}+\frac{16(3g_{1,k}+2g_{2,k})(3a_k+a_{1,k}\rho_0)}{3(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^3}\nonumber\\
&&\hspace{1cm}+\frac{4(6a_k+a_{1,k}\rho_0)}{\rho_0(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^2}\Bigg], \nonumber
\eea
\bea
\label{Eq:flowcoup}
\partial_k a_{1,k} &\!\!\!=\!\!\!&\frac{k^5}{32\pi^2} \Bigg[\frac{36g_{1,k}(a_k+2a_{1,k}\rho_0)}{\rho_0(k^2+m_k^2+2g_{1,k}\rho_0)^3}+\frac{9a_k}{\rho_0^2(k^2+m_k^2+2g_{1,k}\rho_0)^2}+\frac{12g_{1,k}(3a_k+10a_{1,k}\rho_0)}{\rho_0(k^2+m_k^2+6g_{1,k}\rho_0)^3} \nonumber\\
&&\hspace{1cm}-\frac{16(3g_{1,k}+2g_{2,k})(3a_k+a_{1,k}\rho_0)}{3\rho_0(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^3}-\frac{12a_k}{\rho_0^2(k^2+m_k^2+2g_{1,k}\rho_0+4g_{2,k}\rho_0/3)^2}\Bigg],
\eea
where we have denoted the expansion point by $\rho_0$, which is to be set to the value of $\rho$ corresponding to the minimum point of the effective action [note that $\rho=\Tr (M^\dagger M)/2$]. As a side remark, one easily checks that choosing $\rho_0=0$ (\ref{Eq:flowcoup}) would lead back to the earlier results, (\ref{Eq:flowcoup0}). Our task is to integrate the system of equations (\ref{Eq:flowcoup}) from $k=\Lambda\equiv 1 \GeV$ down to $k=0$ with the boundary conditions $m_\Lambda^2=m^2$, $g_{1,\Lambda}=g_1$, $g_{2,\Lambda}=g_2$, $a_\Lambda=a$, $a_{1,\Lambda}=0$, where $m^2, g_1, g_2, a$ are such constants that reproduce as accurately as possible the mesonic spectrum in the infrared. Here we used that at the UV scale the coefficient of the operator $\rho\Delta$ can be set to zero due to perturbative renormalizability. This might be questionable if the UV scale was not high enough, as being a dimension 5 operator, dimensional analyis suggests that its coefficient, $a_1$, is of ${\cal O}(1/\Lambda)$. Obviously if the linear sigma model was not an effective theory, and $\Lambda$ could be sent to infinity, the term in question would not be present. But, in principle the $a_1$ coupling could be included already in the UV action. Investigation of such a scenario is beyond the scope of this paper.
Before solving the coupled system of equations (\ref{Eq:flowcoup}), we need to fix the explicit symmetry breaking terms, i.e., the values for $h_0$, $h_8$. Instead of $h_0$ and $h_8$, we will work in the nonstrange--strange basis, i.e., $h_{\ns}=\sqrt{\frac23}h_0+\frac{1}{\sqrt3}h_8$, $h_s=\frac{1}{\sqrt3}h_0-\sqrt{\frac{2}{3}}h_8$. The partially conserved axialvector current (PCAC) relations give
\bea
m_{\pi}^2f_{\pi}=h_{\ns}, \quad m_K^2f_K=\frac{h_{\ns}}{2}+\frac{h_s}{\sqrt2},
\eea
where $m_{\pi}^2=\delta^2 \Gamma/\delta \pi_i^2(q=0)$ $[i=1,2,3]$ and $m_K^2=\delta^2 \Gamma/\delta \pi_i^2(q=0)$ $[j=4,5,6,7]$. Using physical pion and kaon masses, $\sim 140 \MeV$, $\sim 494 \MeV$, respectively, and decay constants, $f_{\pi}=93 \MeV$, $f_K=113 \MeV$, one gets
\bea
\label{Eq:h}
h_{\ns}=m_{\pi}^2f_{\pi} \approx (122 \MeV)^3, \quad h_{\s}=\frac{1}{\sqrt2}(2m_K^2f_K-m^2_{\pi}f_{\pi}) \approx (335 \MeV)^3,
\eea
which is equivalent to
\bea
h_0&=&\sqrt{\frac23}\big(m_{\pi}^2f_{\pi}/2+m_K^2f_K)\approx (285 \MeV)^3, \quad h_8=\frac{2}{\sqrt3}\big(m_{\pi}^2f_{\pi}-m_K^2f_K) \approx -(310 \MeV)^3.
\eea
Now we use that Ward identities of chiral symmetry lead to
\bea
\frac{\delta \Gamma}{\delta s_{\ns}}(q=0)=m_{\pi}^2s_{\ns}-h_{\ns}, \quad \frac{\delta \Gamma}{\delta s_{\s}}(q=0)=\frac{m_K^2-m_{\pi}^2}{\sqrt2}s_{\ns}+m_K^2s_{\s}-h_{\s}.
\eea
Combined with (\ref{Eq:h}), this shows that no matter how we choose the remaining parameters $m^2, g_1, g_2, a$, in the minimum point of the effective action
\bea
s_{\ns,\min}=f_{\pi}, \quad s_{\s,\min}=\sqrt2(f_K-f_{\pi}/2).
\eea
That leads to $\rho_0=(s_{\ns,\min}^2+s_{\s,\min}^2)/2$, and thus we are ready to fix the aforementioned parameters. Solving (\ref{Eq:flowcoup}), the values $\{m^2, g_1, g_2, a\}\approx \{0.6835 \GeV^2, 29.7, 91.5, -4.4 \GeV\}$ lead to the masses of the pion, kaon, $\eta$, $\eta'$ as $m_{\pi} \approx 133 \MeV$, $m_K \approx 494 \MeV$, $m_{\eta} \approx 537 \MeV$, $m_{\eta'}\approx 957 \MeV$, respectively. Note that the strength of the axial anomaly, $a$, is negative, and it remains so throughout the renormalization group flow. However, after solving (\ref{Eq:flowcoup}), one concludes that at $k=0$ the coefficient $a_{1,k=0}$ is positive. That is to say, since the actual strength of the determinant term is $A:=a_{k=0}+a_{1,k=0}\cdot \rho_0$, when the chiral condensate evaporates, the absolute value of $A$ becomes larger. This shows that at low temperatures $T$, where the $T$ dependence of the anomaly parameters is negligible, the anomaly is actually strengthening as chiral symmetry gradually restores. That is one of the main results of the paper.
In what follows, we provide a rough estimate how the anomaly behaves at the zero temperature nuclear liquid--gas transition.
\section{Anomaly strengthening at the nuclear liquid--gas transition}
In this section we apply our results to the zero temperature nuclear liquid--gas transition. We assume that the nucleon field couples to the mesons via Yukawa interaction, ${\cal L}_{\inter}=g\bar{\psi}M_5\psi$, $\psi^T=(p, n)$, $M_5 =\sum_{a=\ns,1,2,3} (s^a+i\pi^a\gamma_5)T^a$, where the nonstrange generator is $T^{\ns}=\sqrt{2/3}T^0+1/\sqrt3 T^8$, while $\gamma_5$ is the fifth Dirac matrix. In principle one would also need to include the dynamics of an $\omega$ vector particle into the system \cite{floerchinger12,drews16} that models the repulsive interaction between nucleons, but as we will see in a moment, for our purposes it plays no role.
First, we exploit some of the zero temperature properties of nuclear matter. Note that, in the current model, the nucleon mass entirely originates from the spontaneous breaking of chiral symmetry,
\bea
m_{N}(s_{\ns})=g_Ys_{\ns}/2,
\eea
and since $m_{N}(f_{\pi})\approx 939 \MeV$ in the vacuum, we arrive at $g_Y\approx 20.19$. Normal nuclear density, $n_N \approx 0.17 \fm^{-3} \approx (109.131 \MeV)^3$ leads to the Fermi momentum, $p_F$, of the nucleons, since at the mean field level, for $T=0$ we have
\bea
n_N=4\int_p n_F\Big(\sqrt{p^2+m_N^2}-p_F\Big)\Big|_{T=0}\equiv \frac{2}{3\pi^2}p_F^3,
\eea
therefore, $p_F \approx 267.9 \MeV \approx 1.36 \fm^{-1}$. This leads to the nonstrange condensate in the liquid phase, $s_{\ns\!,\liq}$, because the Landau mass, which is defined as
\bea
M_L=\sqrt{p_F^2+m_N^2(s_{\ns\!,\liq})}
\eea
is known to be $M_L \approx 0.8m_N(f_{\pi}) \approx 751.2 \MeV$, and thus $s_{\ns\!,\!\liq} \approx 69.52 \MeV$ \cite{floerchinger12,drews16}. This shows that as we increase the chemical potential, the nonstrange chiral condensate, $s_{\ns}$, jumps: $f_{\pi}\rightarrow s_{\ns\!,\liq}$. This will definitely be accompanied by a jump in the strange condensate, but it has been shown to be significantly smaller \cite{fejos18}. Neglecting the change in $s_{\s}$, the $\rho$ chiral invariant jumps as $(f_\pi^2+s_{\s\!,\min}^2)/2 \rightarrow (s_{\ns\!,\liq}^2+s_{\s,\min}^2)/2$. As discussed in the previous section, the anomaly strength is $A=a_{k=0}+a_{1,k=0}\cdot \rho$, which also jumps accordingly, and the change in $A$ becomes
\bea
\Delta A = a_{1,k=0}\cdot\Delta \rho,
\eea
where $\Delta \rho=(s_{\ns,\liq}^2-f_{\pi}^2)/2$. Solving (\ref{Eq:flowcoup}) one gets $a_{k=0}\approx -9.05\GeV$ and $a_{1,k=0} \approx 494.5\GeV^{-1}$, therefore, the relative change in the anomaly at the liquid--gas transition is
\bea
\frac{\Delta A}{A}=\frac{a_{1,k=0}\cdot\Delta \rho}{a_{k=0}+a_{1,k=0}\cdot\rho_0} \approx 0.2 = 20\%,
\eea
which is in the ballpark of the result of \cite{fejos18}. One can now check how robust this result is with respect to changing the cutoff $\Lambda$. A thorough investigation reveals that in a cutoff interval of $0.8$--$1.5 \GeV$, while the mass spectrum can be maintained within a few percent error after reparametrization, the $\Delta A/A$ ratio is less stable. One finds that the latter is is a monotonically decreasing function of the cutoff and varies roughly between $15$--$40\%$ in the above interval. Results show that when going below $1 \GeV$ the cutoff dependence gets stronger, which is understandable, since non-renormalizable operators are absent at the UV scale. That is, if the latter is chosen to be too small, the model cannot provide robust results (more operators would be needed). Going beyond $1.5$--$2$ GeV, in turn, would be physically inappropriate as at those scales quark degrees of freedom would definitely play a crucial role. From these findings it is safe to say is that the relative change of the anomaly strength is of ${\cal O}( 10\%)$ at the transition point.
At this point we once again wish to emphasize that we have neglected the drop in the strange condensate, and also, the present analysis is based on perturbation theory. In principle higher order operators that break the $U_A(1)$ subgroup should also be resummed, e.g., terms such as $\sim \big(\Tr (M^\dagger M)\big)^n (\det M+ \det M^\dagger)$ could be of huge importance. The lesson we wish to point out here is that the present, rather simple perturbative calculation can also capture the phenomenon of strengthening anomaly as the chiral condensate (partially) evaporates.
\section{Conclusions}
In this paper we investigated how the $U_A(1)$ anomaly behaves as a function of the chiral condensate. We worked with the three flavor linear sigma model, and calculated the leading correction in a $1/\Lambda$ expansion to the conventional anomaly term caused by quantum fluctuations. We have found that the coefficient of the aforementioned operator, $\sim \Tr (M^\dagger M)\cdot (\det M + \det M^\dagger)$, causes the actual strength of the anomaly to become larger once the chiral condensate evaporates. For the sake of an example, we demonstrated that at the zero temperature nuclear liquid--gas transition, where (on top of a jump in the nuclear density) the chiral condensate partially restores, the actual strength of the anomaly increases. This could also happen toward the full restoration of chiral symmetry, where quark dynamics also play a significant role. Note that our findings are based solely on calculating mesonic fluctuations, and no instanton effects have been taken into account.
The linear sigma model, being an effective field theory, cannot accommodate instantons as the fundamental model of QCD. Still, there are at least two directions worth exploring in the effective model framework. Recently it has been shown \cite{pisarski20} that $3Q$-point interactions are generated by instantons with $Q$ topological charge, which can be embedded into the linear sigma model via $\sim [(\det M^\dagger)^Q + (\det M)^Q]$ operators. Another important direction could be to assign environment dependence even to the bare anomaly coefficient(s) from QCD data and see how these compete against thermal effects caused by mesonic fluctuations.
Finally, we wish to point out that our study calls for an extension via a non-perturbative treatment, where fluctuations are taken into account beyond the ${\cal O}(a)$ order, and the coefficient function of the determinant term is obtained in a functional fashion, rather than at the lowest order of its Taylor series. The aforementioned directions are under progress and will be reported in a separate study.
\section*{Acknowledgements}
The author thanks A. Patk\'os and Zs. Sz\'ep for discussions on related topics. This research was supported by the National Research, Development and Innovation Fund under Project No. PD127982, by the János Bolyai Research Scholarship of the Hungarian Academy of Sciences, and by the ÚNKP-20-5 New National Excellence Program of the Ministry for Innovation and Technology from the source of the National Research, Development and Innovation Fund.
\makeatletter
\@addtoreset{equation}{section}
\makeatother
| 6b191557552e085eb6106e55f2663daaf96eb3e4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Many important real-world problems involve optimizing a submodular function. Such problems include maximum coverage, maximum cut \cite{MaxcutSDP}, maximum influence \cite{MaxSpreadInfluence}, sensor placement problem \cite{NearSensorPlacement,SubInfoGather}, as well as many problems in the machine learning domain \cite{FeatureSelection,DataSubsetSelection,DocSum,DocSumBudget,EfficientMinDecompSubmodular}. Much work has been done in the area of submodular optimization under static constraints.
A particularly well-studied class of algorithms in this line of research is greedy algorithms, which have been shown to be efficient in exploiting submodularity \cite{LocationBankOptFloat,MaxSubmodularMatroid,GreedyMaxBoundCurvPartMatroid,GreedNonmod}.
Important recent results on the use of evolutionary algorithms for submodular optimization are summarized in~\cite{DBLP:books/sp/ZhouYQ19}.
Real-world problems are seldom solved once, but rather many times over some period of time, during which they change. Such changes demand adapting the solutions that would otherwise become poor or infeasible. The dynamic nature of these problems presents many interesting optimization challenges, which have long been embraced by many researchers. A lot of research in the evolutionary computation literature has addressed these types of problems from an applied perspective.
Theoretical investigations have been carried out for evolutionary algorithms on some example functions and classical combinatorial optimization problems such as shortest paths, but in general the theoretical understand on complex dynamic problems is rather limited~\cite{DBLP:journals/corr/abs-1806-08547}.
In this paper, we follow the approach of carrying out theoretical runtime analysis of evolutionary algorithms with respect to their runtime and approximation behavior. This well established area of research has significantly increased the theoretical understanding of evolutionary computation methods~\cite{DBLP:books/daglib/0025643,BookDoeNeu}.
Many recent studies on submodular and near-submodular optimization have investigated Pareto optimization approaches based on evolutionary computation. \citet{POMC} derive an approximation guarantee for the POMC algorithm for maximizing monotone function under a monotone constraint. They show that POMC achieves an $(\alpha_f/2)(1-1/e^{\alpha_f})$-approximation within at most cubic expected run time. The recent study of \citet{MaxMonoApproxSubmodMulti} extends the results to a variant of the GSEMO algorithm (which inspired POMC) to the problem of maximizing general submodular functions, but under a cardinality constraint. It results reveal that non-monotonicity in objective functions worsens approximation guarantees.
In our work, we extend existing results for POMC \cite{ParetoSubsetDynCon} to partition matroid constraints with dynamic thresholds. We show that the proven adaptation efficiency facilitated by maintaining dominating populations can be extended for multiple constraints with appropriately defined dominance relations. In particular, we prove that POMC can achieve new approximation guarantees quickly whether the constraints thresholds are tightened or relaxed. Additionally, we study POMC experimentally on the dynamic max cut problem and compare its results against the results of greedy algorithms for underlying static problems. Our study evaluates the efficiency in change adaptation, thus assuming immaculate change detection. Our results show that POMC is competitive to GREEDY during unfavorable changes, and outperforming GREEDY otherwise.
In the next section, we formulate the problem and introduce the Pareto optimization approach that is subject to our investigations. Then we analyze the algorithm in terms of runtime and approximation behaviour when dealing with dynamic changes. Finally, we present the results of our experimental investigations and finish with some conclusions.
\section{Problem Formulation and Algorithm}
In this study, we consider optimization problems where the objective functions are either submodular or monotone. We use the following definition of submodularity \cite{Submod1978}.
\begin{definition}\label{submodular_def}
Given a finite set $V$, a function $f:2^V\to\mathbb{R}^+$ is submodular if it satisfies for all $X\subseteq Y\subseteq V$ and $v\in V\setminus Y$,
\[f(Y\cup\{v\})-f(Y)\leq f(X\cup\{v\})-f(X).\]
\end{definition}
As in many relevant works, we are interested in the submodularity ratio which quantifies how close a function is to being modular. In particular, we use a simplified version of the definition in \cite{spectral}.
\begin{definition}\label{submodratio_def}
For a monotone function $f:2^V\to\mathbb{R}^+$, its submodularity ratio with respect to two parameters $i$, $j\geq1$ is
\[\gamma_{i,j}=\min_{|X|<i,|L|\leq j,X\cap L=\emptyset}\frac{\sum_{v\in L}[f(X\cup\{v\})-f(X)]}{f(X\cup L)-f(X)},\]
for $i>0$ and $\gamma_{0,j}=\gamma_{1,j}$.
\end{definition}
It can be seen that $\gamma_{i,j}$ is non-negative, non-increasing with increasing $i$ and $j$, and $f$ is submodular iff $\gamma_{i,j}\geq1$ for all $(i,j)$. This ratio also indicates the intensity of the function's diminishing return effect. Additionally, non-monotonicity is also known to affect worst-case performance of algorithms \cite{GreedyMaxBoundCurvPartMatroid,MaxMonoApproxSubmodMulti}. As such, we also use the objective function's monotonicity approximation term defined similarly to \cite{NearSensorPlacement}, but only for subsets of a certain size.
\begin{definition}\label{mono_approx}
For a function $f:2^V\to\mathbb{R}^+$, its monotonicity approximation term with respect to a parameter $j$ is
\[\epsilon_j=\max_{X,v:|X|<j}\{f(X\setminus\{v\})-f(X)\},\]
for $j>0$ and $\epsilon_0=0$.
\end{definition}
It is the case that $\epsilon_j$ is non-negative, non-decreasing with increasing $j$, and $f$ is monotone iff $\epsilon_{n+1}=0$. We find that adding the size parameter can provide extra insight into the analysis results.
Consider the static optimization problem with partition matroid constraints.
\begin{definition}\label{def_static}
Given a set function $f:2^V\to\mathbb{R}^+$, a partitioning $B=\{B_i\}_{i=1}^k$ of $V$, and a set of integer thresholds $D=\{d_i\}_{i=1}^k$, the problem is
\begin{align*}
\underset{X\subseteq V}{\text{maximize }}f(X),
\text{ s.t. }|X\cap B_i|\leq d_i,\quad\forall i=1,\dots,k.
\end{align*}
\end{definition}
We define notations $d=\sum_{i=1}^kd_i$, $\bar{d}=\min_i\{d_i\}$, and $OPT\subseteq V$ the feasible optimal solution. A solution $X\subseteq V$ is feasible iff it satisfies all constraints. It can be shown that $\bar{d}\leq d/k$, $|OPT|\leq d$, and any solution $X$ where $|X|\leq\bar{d}$ is feasible. Each instance is then uniquely defined by the triplet $(f,B,D)$. Without loss of generality, we assume $1\leq d_i\leq |B_i|$ for all $i$.
We study the dynamic version of the problem in Definition \ref{def_static}. This dynamic problem demands adapting the solutions to changing constraints whenever such changes occur.
\begin{definition}\label{def_dynamic}
Given the problem in Definition \ref{def_static}, a dynamic problem instance is defined by a sequence of changes where the current $D$ in each change is replaced by $D^*=\{d^*_i\}_{i=1}^k$ such that $d^*_i\in[1,|B_i|]$ for $i=1,\dots,k$. The problem is to generate a solution $X$ that maximizes $f(X)$ for each newly given $D^*$ such that
\[|X\cap B_i|\leq d^*_i,\quad\forall i=1,\dots,k.\]
\end{definition}
Such problems involve changing constraint thresholds over time. Using the oracle model, we assume time progresses whenever a solution is evaluated. We define notations $d^*=\sum_{i=1}^kd^*_i$, $\bar{d}^*=\min_i\{d^*_i\}$, and the new optimal solution $OPT^*$. Similarly, we assume $1\leq d^*_i\leq |B^*_i|$ for all $i$. Lastly, while restarting from scratch for each new thresholds is a viable tactic for any static problems solver, we focus on the capability of the algorithm to adapt to such changes.
\subsection{POMC algorithm}
The POMC algorithm \cite{POMC} is a Pareto Optimization approach for constrained optimization. It is also known as GSEMO algorithm in the evolutionary computation literature~\cite{SEMO,GSEMO,GSEMO2015}. As with many other evolutionary algorithms, the binary representation of a set solutions is used. For this algorithm, we reformulate the problem as a bi-objective optimization problem given as
\[\text{maximize }(f_1(X),f_2(X)),\]
where
\[f_1(X)=\begin{cases*}
f(X),&if $X$ is feasible\\
-\infty,&otherwise
\end{cases*},\text{ }f_2(x)=-|X|.\]
POMC optimizes two objectives simultaneously, using the dominance relation between solutions, which is common in Pareto optimization approaches. Recall that solution $X_1$ dominates $X_2$ ($X_1\succeq X_2$) iff $f_1(X_1)\geq f_1(X_2)$ and $f_2(X_1)\geq f_2(X_2)$. The dominance relation is strict ($X_1\succ X_2$) iff $X_1\succeq X_2$ and $f_i(X_1)>f_i(X_2)$ for at least one $i \in \{1,2\}$. Intuitively, dominance relation formalizes the notion of ``better'' solution in multi-objective contexts. Solutions that don't dominate any other present a trade-off between objectives to be optimized.
The second objective in POMC is typically formulated to promote solutions that are ``further'' from being infeasible. The intuition is that for those solutions, there is more room for feasible modification, thus having more potential of becoming very good solutions. For the problem of interest, one way of measuring ``distance to infeasibility'' for some solution $X$ is counting the number of elements in $V\setminus X$ that can be added to $X$ before it is infeasible. The value then would be $d-|X|$, which is the same as $f_2(X)$ in practice. Another way is counting the minimum number of elements in $V\setminus X$ that need to be added to $X$ before it is infeasible. The value would then be $\min_i\{d_i-|B_i\cap X|\}$. The former approach is chosen for simplicity and viability under weaker assumptions about the considered problem.
On the other hand, the first objective aims to present the canonical evolutionary pressure based on objective values. Additionally, $f_1$ also discourages all infeasible solutions, which is different from the formulation in \cite{POMC} that allows some degree of infeasibility. This is because for $k>1$, there can be some infeasible solution $Y$ where $|Y|\leq d$. If $f_1(Y)$ is very high, it can dominate many good feasible solutions, and may prevent acceptance of global optimal solutions into the population. Furthermore, restricting to only feasible solutions decreases the maximum population size, which can improve convergence performance. As a consequence, the population size of POMC for our formulation is at most $d+1$. Our formulation of the two objective functions is identical to the ones in \cite{GSEMO2015} when $k=1$.
\begin{algorithm}[t]
\begin{algorithmic}[1]
\STATEx \textbf{Input:} a problem instance: $(f,B,D)$
\STATEx \textbf{Parameter:} the number of iterations $T\geq0$
\STATEx \textbf{Output:} a feasible solution $x\in\{0,1\}^n$
\STATE $x\gets0^n$, $P\gets\{x\}$
\WHILE{$t<T$}
\IF{Change is detected}
\STATE $P\gets P\setminus\{x\in P|\exists y\in P,y\neq x\wedge y\succeq x\}$
\ENDIF
\STATE Randomly sample a solution $y$ from $P$
\STATE $y'\gets y$ after flipping each bit with probability $1/n$
\IF{$\nexists x\in P,x\succ y'$}
\STATE $P\gets(P\setminus\{x\in P|y'\succeq x\})\cup\{y'\}$
\ENDIF
\ENDWHILE
\STATE \textbf{return} $\argmax_{x\in P}f_1(x)$
\end{algorithmic}
\caption{POMC algorithm for dynamic problems}
\label{alg:gsemo}
\end{algorithm}
POMC (see Algorithm~\ref{alg:gsemo}) starts with initial population consisting of the search point $0^n$ which represents the empty set. In each iteration, a new solution is generated by random parent selection and bit flip mutation. Then the elitist survivor selection mechanism removes dominated solutions from the population, effectively maintaining a set of trade-off solutions for the given objectives. The algorithm terminates when the number of iteration reaches some predetermined limit. We choose empty set as the initial solution, similar to \cite{POMC} and different from \cite{GSEMO2015}, to simplify the analysis and stabilize theoretical performance. Note that POMC calls the oracle once per iteration to evaluate a new solution, so its run time is identical to the number of iterations.
We assume that changes are made known to the algorithm as they occur, and that feasibility can be checked efficiently. The reason for this is that infeasibility induced by changes in multiple thresholds has nontrivial impact on the algorithm's behaviour. For single constraint problems, this impact is limited to increases in population size \cite{ParetoSubsetDynCon}, since a solution's degree of feasibility entirely correlates with its second objective value. However, this is no longer the case for multiple constraints, as the second objective aggregates all constraints. While it reduces the population size as the result, it also allows for possibilities where solutions of small size (high in second objective) become infeasible after a change and thus dominate other feasible solutions of greater cardinality without updating evaluations. This can be circumvented by assuming that the changes' directions are the same every time, i.e., $(d^*_i-d_i)(d^*_j-d_j)\geq0$ for every $(i,j)$ pair. Instead of imposing assumptions on the problems, we only assume scenarios where POMC successfully detects and responds to changes. This allows us to focus our analysis entirely on the algorithm's adaptation efficiency under arbitrary constraint threshold change scenarios.
\section{Runtime Analysis}
For the runtime analysis of POMC for static problems, we refer to the results by~\citet{Do2020}. In short, its worst-case approximation ratios on static problems are comparable to those of the classical GREEDY algorithm, assuming submodularity and weak monotonicity in the objective function \cite{GreedyMaxBoundCurvPartMatroid}. On the other hand, a direct comparison in other cases is not straightforward as the bounds involve different sets of parameters.
The strength of POMC in dynamic constraints handling lies in the fact that it stores a good solution for each cardinality level up to $\bar{d}$. In this way, when $\bar{d}$ changes, the population will contain good solutions for re-optimization. We use the concept of greedy addition $v^*_X$ to a solution $X$.
\[v^*_X=\argmax_{v\in V\setminus X}f_1(X\cup\{v\}).\]
It can be shown that for any $X$ where $|X|<\bar{d}$, the corresponding greedy addition $v^*_X$ w.r.t. $(f,B,D)$ is the same as the one w.r.t. $(f,B,D^*)$ if $\bar{d}^*\geq\bar{d}$, since $X\cup\{v\}$ is still feasible for all $v\in V\setminus X$. Thus, we can derive the following result from Lemma 2 in \cite{MaxMonoApproxSubmodMulti}, and Lemma 1 in \cite{Do2020}.
\begin{lemma}\label{greed_prog_submodular_dyn}
Let $f$ be a submodular function, $\bar{d}^*\geq\bar{d}$, and $\epsilon_d$ be defined in Definition \ref{mono_approx}, for all $X\subseteq V$ such that $|X|=j<\bar{d}$, we have both
\begin{align*}
&f(X\cup\{v^*_X\})-f(X)\geq\frac{1}{d}\left[f(OPT)-f(X)-j\epsilon_{d+j+1}\right],\\
&f(X\cup\{v^*_X\})-f(X)\geq\frac{1}{d^*}\left[f(OPT^*)-f(X)-j\epsilon_{d^*+j+1}\right].
\end{align*}
\end{lemma}
\begin{proof}
The first part is from Lemma 1 in \cite{Do2020}, while the second follows since if $|X|<\bar{d}\leq\bar{d}^*$, then the element contributing the greedy marginal gain to $f_1(X)$ is unchanged.
\end{proof}
The result carries over due to satisfied assumption $|X|<\bar{d}^*$. Using these inequalities, we can construct a proof, following a similar strategy as the one for Theorem 5 in \cite{ParetoSubsetDynCon} which leads to the following theorem.
\begin{theorem}\label{gsemo_submodular_dyn}
For the problem of maximizing a submodular function under partition matroid constraints, assuming $\bar{d}^*\geq\bar{d}$, POMC generates a population $P$ in expected run time $\mathcal{O}(d^2n/k)$ such that
\begin{align}\label{eq:gsemo_submodular_dyn}
\begin{split}
&\forall m\in[0,\bar{d}],\exists X\in P,|X|\leq m\\&\wedge f(X)\geq\left[1-\left(1-\frac{1}{d}\right)^{m}\right][f(OPT)-(m-1)\epsilon_{d+m}]\\&\wedge f(X)\geq\left[1-\left(1-\frac{1}{d^*}\right)^{m}\right][f(OPT^*)-(m-1)\epsilon_{d^*+m}].
\end{split}
\end{align}
\end{theorem}
\begin{proof}
Let $S(X,j)$ be the expression
\begin{align*}
&|X|\leq j\\&\wedge f(X)\geq\left[1-\left(1-\frac{1}{d}\right)^{j}\right][f(OPT)-(j-1)\epsilon_{d+j}]\\&\wedge f(X)\geq\left[1-\left(1-\frac{1}{d^*}\right)^{j}\right][f(OPT^*)-(j-1)\epsilon_{d^*+j}],
\end{align*}
and $Q(i)$ be the expression $\forall j\in[0,i],\exists X\in P_t,S(X,j)$. We have that $Q(0)$ holds. For each $h\in[0,\bar{d}-1]$, assume $Q(h)$ holds and $Q(h+1)$ does not hold at some iteration $t$. Let $X\in P_t$ be the solution such that $S(X,h)$ holds, $S(Y,h)$ and $Q(h)$ holds at iteration $t+1$ for any solution $Y\in P_{t+1}$ such that $Y\succeq X$. This means once $Q(h)$ holds, it must hold in all subsequent iterations of POMC. Let $X^*=X\cup\{v^*_X\}$, Lemma \ref{greed_prog_submodular_dyn} implies
\begin{align*}
f(X^*)&\geq\left[1-\left(1-\frac{1}{d}\right)^{h+1}\right][f(OPT)-h\epsilon_{d+h+1}],\\
\text{and}&\\
f(X^*)&\geq\frac{1}{d^*}f(OPT^*)-\frac{h}{d^*}\epsilon_{d^*+h+1}\\&+\left(1-\frac{1}{d^*}\right)\left[1-\left(1-\frac{1}{d^*}\right)^h\right]\\&[f(OPT^*)-(h-1)\epsilon_{d^*+h}]\\&
\geq\left[1-\left(1-\frac{1}{d^*}\right)^{h+1}\right][f(OPT^*)-h\epsilon_{d^*+h+1}].
\end{align*}
The second inequality uses $0\leq\epsilon_{d^*+h}\leq\epsilon_{d^*+h+1}$. This means that $S(X^*,h+1)$ holds. Therefore, if $X^*$ is generated, then $Q(h+1)$ holds, regardless of whether $X^*$ is dominated afterwards. According to the bit flip procedure, $X^*$ is generated with probability at least $1/(en(d+1))$ which implies
\[\Pr[Q(h+1)|Q(h)]\geq\frac{1}{en(d+1)},\text{ and }\Pr[\neg Q(h)|Q(h)]=0.\]
Using the additive drift theorem \cite{DriftAnalysis}, the expected number of iterations until $Q(\bar{d})$ holds is at most $e\bar{d}n(d+1)=\mathcal{O}(d^2n/k)$. This completes the proof.
\end{proof}
The statement \eqref{eq:gsemo_submodular_dyn} implies the following results.
\begin{align*}
&\forall m\in[0,\bar{d}],\exists X\in P,|X|\leq m\\&\wedge f(X)\geq\left(1-e^{m/d}\right)\left[f(OPT)-(m-1)\epsilon_{d+m}\right]\\&\wedge f(X)\geq\left(1-e^{m/d^*}\right)\left[f(OPT^*)-(m-1)\epsilon_{d^*+m}\right].
\end{align*}
We did not put this more elegant form directly in Theorem \ref{gsemo_submodular_dyn} since it cannot be used in the subsequent proof; only Expression \eqref{eq:gsemo_submodular_dyn} is applicable. Note that the result also holds for $\bar{d}^*\leq\bar{d}$ if we change the quantifier to $\forall m\in[0,\bar{d}^*]$. It implies that when a change such that $\bar{d}^*\leq\bar{d}$ occurs after cubic run time, POMC is likely to instantly satisfy the new approximation ratio bound, which would have taken it extra cubic run time to achieve if restarted. Therefore, it adapts well in such cases, assuming sufficient run time is allowed between changes. On the other hand, if $\bar{d}^*>\bar{d}$, the magnitude of the increase affects the difficulty with which the ratio can be maintained. The result also states a ratio bound w.r.t. the new optimum corresponding to the new constraint thresholds. As we will show using this statement, by keeping the current population (while discarding infeasible solutions), POMC can adapt to the new optimum quicker than it can with the restart strategy.
\begin{theorem}\label{gsemo_submodular_dyn2}
Assuming POMC achieves a population satisfying \eqref{eq:gsemo_submodular_dyn}, after the change where $\bar{d}^*>\bar{d}$, POMC generates in expected time $\mathcal{O}((\bar{d}^*-\bar{d})d^*n)$ a solution $X$ such that
\[f(X)\geq\left(1-e^{\bar{d}^*/d^*}\right)\left[f(OPT^*)-(\bar{d}^*-1)\epsilon_{d^*+\bar{d}^*}\right].\]
\end{theorem}
\begin{proof}
Let $S(X,i)$ be the expression
\begin{align*}
&|X|\leq j\\&\wedge f(X)\geq\left[1-\left(1-\frac{1}{d^*}\right)^{j}\right][f(OPT^*)-(j-1)\epsilon_{d^*+j}].
\end{align*}
Assuming \eqref{eq:gsemo_submodular_dyn} holds for some iteration $t$, let $\bar{X}\in P_t$ be a solution such that $S(\bar{X},i)$ holds for some $i\in[\bar{d},\bar{d}^*)$, and $v^*_{\bar{X}}=\argmax_{v\in V\setminus\bar{X}}\{f_1(\bar{X}\cup\{v\})-f_1(\bar{X})\}$ w.r.t. $(f,B,D^*)$ for any $\bar{X}\subset V$, and $\bar{X}'=\bar{X}\cup\{v^*_{\bar{X}}\}$, Lemma \ref{greed_prog_submodular_dyn} implies
\begin{align*}
f(\bar{X}')&\geq\left[1-\left(1-\frac{1}{d^*}\right)^{i+1}\right][f(OPT^*)-i\epsilon_{d^*+i+1}].
\end{align*}
Hence, $S(\bar{X}',i+1)$ holds. It is shown that such a solution is generated by POMC with probability at least $1/(en(d^*+1))$. Also, for any solution $X$ satisfying $S(X,i)$, another solution $Y\succeq X$ must also satisfy $S(Y,i)$. Therefore, the Additive Drift Theorem implies that given $S(X,\bar{d})$ holds for some $X$ in the population, POMC generates a solution $Y$ satisfying $S(Y,\bar{d}^*)$ in expected time at most $(\bar{d}^*-\bar{d})en(d^*+1)=\mathcal{O}((\bar{d}^*-\bar{d})d^*n)$. Such a solution satisfies the inequality in the theorem.
\end{proof}
The degree with which $\bar{d}$ increases only contributes linearly to the run time bound, which shows efficient adaptation to the new search space.
For the monotone objective functions cases, without loss of generality, we assume that $f$ is normalized ($f(\emptyset)=0$). We make use of the following inequalities, derived from Lemma 1 in \cite{PPOSS}, and Lemma 2 in \cite{Do2020}, using the same insight as before.
\begin{lemma}\label{greed_prog_monotone_dyn}
Let $f$ be a monotone function, $\bar{d}^*\geq\bar{d}$, and $\gamma_{i,j}$ be defined in Definition \ref{submodratio_def}, for all $X\subseteq V$ such that $|X|=j<\bar{d}$, we have both
\begin{align*}
&f(X\cup\{v^*_X\})-f(X)\geq\frac{\gamma_{j+1,d}}{d}[f(OPT)-f(X)],\\
&f(X\cup\{v^*_X\})-f(X)\geq\frac{\gamma_{j+1,d^*}}{d^*}[f(OPT^*)-f(X)].
\end{align*}
\end{lemma}
\begin{proof}
The first part is from Lemma 2 in \cite{Do2020}, while the second follows since if $|X|<\bar{d}\leq\bar{d}^*$, then the element contributing the greedy marginal gain to $f_1(X)$ is unchanged.
\end{proof}
This leads to the following result which show POMC's capability in adapting to changes where $\bar{d}^*<\bar{d}$.
\begin{theorem}\label{gsemo_monotone_dyn}
For the problem of maximizing a monotone function under partition matroid constraints, assuming $\bar{d}^*\geq\bar{d}$, POMC generates a population $P$ in expected run time $\mathcal{O}(d^2n/k)$ such that
\begin{align}\label{eq:gsemo_monotone_dyn}
\begin{split}
&\forall m\in[0,\bar{d}],\exists X\in P,|X|\leq m\\&\wedge f(X)\geq\left[1-\left(1-\frac{\gamma_{m,d}}{d}\right)^{m}\right]f(OPT)\\&\wedge f(X)\geq\left[1-\left(1-\frac{\gamma_{m,d^*}}{d^*}\right)^{m}\right]f(OPT^*).
\end{split}
\end{align}
\end{theorem}
\begin{proof}
Let $S(X,j)$ be the expression
\begin{align*}
|X|\leq j&\wedge f(X)\geq\left[1-\left(1-\frac{\gamma_{j,d}}{d}\right)^{j}\right]f(OPT)\\&\wedge f(X)\geq\left[1-\left(1-\frac{\gamma_{j,d^*}}{d^*}\right)^{j}\right]f(OPT^*),
\end{align*}
and $Q(i)$ be the expression $\forall j\in[0,i],\exists X\in P_t,S(X,j)$. We have that $Q(0)$ holds. Similar to the proof for Theorem \ref{gsemo_submodular_dyn}, we get, using Lemma \ref{greed_prog_monotone_dyn}, that the expected number of iterations of POMC until $Q(\bar{d})$ holds is at most $\mathcal{O}(d^2n/k)$.
\ignore{
For each $h\in[0,\bar{d}-1]$, assume $Q(h)$ holds and $Q(h+1)$ does not hold at some iteration $t$, let $X\in P_t$ be the solution such that $S(X,h)$ holds, $S(Y,h)$ and $Q(h)$ holds at iteration $t+1$ for any solution $Y\in P_{t+1}$ such that $Y\succeq X$. This means once $Q(h)$ holds, it must hold in all subsequent iterations of POMC. Let $X^*=X\cup\{v^*_X\}$, Lemma \ref{greed_prog_monotone_dyn} implies
\begin{align*}
f(X^*)&\geq\left[1-\left(1-\frac{\gamma_{h,d}}{d}\right)^{h+1}\right]f(OPT),\\
\text{and}&\\
f(X^*)&\geq\frac{\gamma_{h,d^*}}{d^*}f(OPT^*)\\&+\left(1-\frac{\gamma_{h+1,d^*}}{d^*}\right)\left[1-\left(1-\frac{\gamma_{h,d^*}}{d^*}\right)^h\right]f(OPT^*)\\&
\geq\left[1-\left(1-\frac{\gamma_{h+1,d^*}}{d^*}\right)^{h+1}\right]f(OPT^*).
\end{align*}
The second inequality uses $\gamma_{h,d^*}\geq\gamma_{h+1,d^*}$. This means $S(X^*,h+1)$ holds. Therefore, if $X^*$ is generated, then $Q(h+1)$ holds, regardless of whether $X^*$ is dominated afterwards. According to the bit flip procedure, $X^*$ is generated with probability at least $1/en(d+1)$, so
\[\Pr[Q(h+1)|Q(h)]\geq\frac{1}{en(d+1)},\quad\text{and}\quad\Pr[\neg Q(h)|Q(h)]=0.\]
The Additive Drift Theorem implies that assuming $Q(0)$, the expected number of iterations before $Q(\bar{d})$ holds is at most $e\bar{d}n(d+1)=\mathcal{O}(d^2n/k)$. This proves the theorem.}
\end{proof}
Similar to Expression~\eqref{eq:gsemo_submodular_dyn}, Expression~\eqref{eq:gsemo_monotone_dyn} gives a more elegant form.
\begin{align*}
&\forall m\in[0,\bar{d}],\exists X\in P,|X|\leq m\\&\wedge f(X)\geq\left(1-e^{-\gamma_{m,d}m/d}\right)f(OPT)\\&\wedge f(X)\geq\left(1-e^{-\gamma_{m,d^*}m/d^*}\right)f(OPT^*).
\end{align*}
Just like Theorem \ref{gsemo_submodular_dyn}, this result holds for $\bar{d}^*\leq\bar{d}$ when the quantifier is $\forall m\in[0,\bar{d}^*]$. It is implied that if such a population is made, the new approximation ratio bound is immediately satisfied when the thresholds decrease. Once again, we can derive the result on POMC's adapting performance to $\bar{d}^*\geq\bar{d}$ by using Theorem \ref{gsemo_monotone_dyn}.
\begin{theorem}\label{gsemo_monotone_dyn2}
Assuming POMC achieves a population satisfying \eqref{eq:gsemo_monotone_dyn}, after the change where $\bar{d}^*>\bar{d}$, POMC generates in expected time $\mathcal{O}((\bar{d}^*-\bar{d})d^*n)$ a solution $X$ such that
\[f(X)\geq\left(1-e^{-\gamma_{\bar{d}^*,d^*}\bar{d}^*/d^*}\right)f(OPT^*).\]
\end{theorem}
\begin{proof}
Let $S(X,i)$ be a expression
\[|X|\leq j\wedge f(X)\geq\left[1-\left(1-\frac{\gamma_{\bar{d}^*,d^*}}{d^*}\right)^{j}\right]f(OPT^*).\]
Assuming \eqref{eq:gsemo_monotone_dyn} holds for some iteration $t$, let $\bar{X}\in P_t$ be a solution such that $S(\bar{X},i)$ holds for some $i\in[\bar{d},\bar{d}^*)$, and $v^*_{\bar{X}}=\argmax_{v\in V\setminus\bar{X}}\{f_1(\bar{X}\cup\{v\})-f_1(\bar{X})\}$ w.r.t. $(f,B,D^*)$ for any $\bar{X}\subset V$, and $\bar{X}'=\bar{X}\cup\{v^*_{\bar{X}}\}$, Lemma \ref{greed_prog_monotone_dyn} implies
\begin{align*}
f(\bar{X}')&\geq\left[1-\left(1-\frac{\gamma_{i+1,d^*}}{d^*}\right)^{i+1}\right]f(OPT^*).
\end{align*}
So $S(\bar{X}',i+1)$ holds. Following the same reasoning in the proof for Theorem \ref{gsemo_submodular_dyn2}, we get that POMC, given a solution $X$ in the population satisfying $S(X,\bar{d})$, generates another solution $Y$ satisfying $S(Y,\bar{d}^*)$ in expected run time at most $\mathcal{O}((\bar{d}^*-\bar{d})d^*n)$.
\end{proof}
Theorem \ref{gsemo_submodular_dyn2} and Theorem \ref{gsemo_monotone_dyn2} state that there is at least one solution in the population that satisfies the respective approximation ratio bounds in each case, after POMC is run for at least some expected number of iterations. However, the proofs for Theorem \ref{gsemo_submodular_dyn} and Theorem \ref{gsemo_monotone_dyn} also imply that the same process generates, for each cardinality level from $0$ to $\bar{d}^*$, a solution that also satisfies the respective approximation bound ratio, adjusted to the appropriate cardinality term. These results imply that instead of restarting, maintaining the non-dominated population provides a better starting point to recover relative approximation quality in any case. This suggests that the inherent diversity in Pareto-optimal fronts is suitable as preparation for changes in constraint thresholds.
\section{Experimental Investigations}
We compare the POMC algorithm against GREEDY \cite{GreedyMaxBoundCurvPartMatroid} with the restart approach, on undirected max cut problems with random graphs. Given a graph $G=(V,E,c)$ where $c$ is a non-negative edge weight function, the goal is to find, while subjected to changing partition matroid constraints,
\[X^*=\argmax_X\left\lbrace\sum_{a\in X,b\in V\setminus X}c(a,b)\right\rbrace.\]
Weighted graphs are generated for the experiments based on two parameters: number of vertices ($n$) and edge density. Here, we consider $n=200$, and 5 density values: 0.01, 0.02, 0.05, 0.1, 0.2. For each $n$-$density$ pair, a different weighted graph is randomly generated with the following procedure:
\begin{enumerate}
\item Randomly sample $E$ from $V\times V$ without replacement, until $|E|=\lfloor density\times n^2\rfloor$.
\item Assign to $c(a,b)$ a uniformly random value in $[0,1]$ for each $(a,b)\in E$.
\item Assign $c(a,b)=0$ for all $(a,b)\notin E$.
\end{enumerate}
To limit the variable dimensions so that the results are easier to interpret, we only use one randomly chosen graph for each $n$-density pair. We also consider different numbers of partitions $k$: 1, 2, 5, 10. For $k>1$, each element is assigned to a partition randomly. Also, the sizes of the partitions are equal, as with the corresponding constraint thresholds.
\begin{figure}[t!]
\centering
\includegraphics[width=.8\linewidth]{b.eps}
\caption{Threshold level over time for dynamic problems}
\label{fig:thres_changes}
\end{figure}
The dynamic component is the thresholds $d_i$. We use the approach outlined in \cite{DynKnapsack} to generate threshold changes in the form of a sequence of $m$ values $\{b_i\}_{i=1}^m$. In particular, these values range in $[0,1]$ and are applied to each instance: $d_i=\max\{b_j|B_i|,1\}$ at the $j^{th}$ change, rounded to the nearest integer. The generating formulas are as follow:
\[b_1=\mathcal{U}(0,1),\quad b_{i+1}=\max\{\min\{b_i+\mathcal{N}(0,0.05^2),1\},0\}.\]
The values $b_i$ used in the experiments are displayed in Figure \ref{fig:thres_changes} where the number of changes is $m=200$.
For each change, the output of GREEDY is obtained by running without restriction on evaluations. These values are used as baselines to compare against POMC's outputs . POMC is run with three different settings for the number of evaluations between changes: $5000$, $10000$, $20000$. Smaller numbers imply a higher change frequency leading to a higher difficulty in adapting to changes. Furthermore, POMC is run 30 times for each setting, and means and standard deviations of results are shown in POMC$_{5000}$, POMC$_{10000}$, POMC$_{20000}$, corresponding to different change frequency settings.
\ignore{
\begin{figure*}[ht!]
\centering
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_01_1.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_01_2.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_01_5.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_01_10.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_05_1.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_05_2.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_05_5.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_05_10.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_2_1.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_2_2.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_2_5.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_100_2_10.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_01_1.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_01_2.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_01_5.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_01_10.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_05_1.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_05_2.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_05_5.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_05_10.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_2_1.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_2_2.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_2_5.eps}
\end{subfigure}
\begin{subfigure}{0.246\textwidth}
\centering
\includegraphics[width=\textwidth]{t_200_2_10.eps}
\end{subfigure}
\caption{POMC's mean outputs against GREEDY's over time in 6 graph settings. Standard deviations in POMC's are minimal.
\frank{Figure doesn't really show differences. Better to do a table with statistical results?}
}.
\label{fig:timeplot}
\end{figure*}
}
\begin{table*}[t!]
\begin{small}
\renewcommand{\arraystretch}{.76}
\caption{Experimental results for dynamic max cut with $n=200$. Outputs are shown in batches of 50 threshold changes. L--W--T are numbers of losses, wins, ties POMC has over GREEDY, determined by U-tests on data in each change.}
\label{table:experimentsd}
\begin{center}
\begin{tabular}{lll@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c@{\hspace{6pt}}c}
\toprule
\multirow{2}{*}{$k$}&\multirow{2}{*}{density}&\multirow{2}{*}{Changes}&\multicolumn{2}{c}{GREEDY}&\multicolumn{3}{c}{POMC$_{5000}$}&\multicolumn{3}{c}{POMC$_{10000}$}&\multicolumn{3}{c}{POMC$_{20000}$}\\\cmidrule(l{2pt}r{2pt}){4-5}\cmidrule(l{2pt}r{2pt}){6-8}\cmidrule(l{2pt}r{2pt}){9-11}\cmidrule(l{2pt}r{2pt}){12-14}&&&mean&std&mean&std&L--W--T&mean&std&L--W--T&mean&std&L--W--T\\\cmidrule(l{2pt}r{2pt}){1-14}
\multirow{12}{*}{1}&\multirow{4}{*}{0.01}&1--50&82.93&27.5&81.26&27.01&26--17--7&82.85&27.49&17--25--8&83.68&27.81&7--31--12\\&
&51--100&57.17&24.63&56.11&23.79&39--2--9&56.88&24.26&26--7--17&57.29&24.57&13--16--21\\&
&101--150&100.7&5.001&100.9&5.535&12--32--6&101.7&5.255&5--43--2&102.3&5.11&1--48--1\\&
&151--200&103.1&1.436e-14&105.1&0.2395&0--50--0&105.1&0.2814&0--50--0&105.2&0.2764&0--50--0\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.05}&1--50&289.3&106.2&283.7&103.4&30--12--8&288.3&105.3&20--20--10&290.5&106.3&7--29--14\\&
&51--100&186.4&87.89&183.5&85.18&43--0--7&185.3&86.25&33--1--16&186.2&86.92&19--10--21\\&
&101--150&359.4&23.64&358.2&24.7&23--23--4&361.9&25.16&9--39--2&363.5&24.95&2--47--1\\&
&151--200&373.2&2.297e-13&379.5&1.618&0--50--0&380.6&1.817&0--50--0&379.5&1.344&0--50--0\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.2}&1--50&890.3&345.2&877.6&336.8&31--13--6&886.8&341&23--21--6&892&343.8&12--28--10\\&
&51--100&551.7&274.3&545.1&266.1&41--2--7&548.7&268.9&34--5--11&550.6&270.5&28--9--13\\&
&101--150&1121&79.61&1116&81.18&33--16--1&1123&81.52&18--25--7&1126&81.89&8--41--1\\&
&151--200&1167&4.594e-13&1179&1.773&0--50--0&1180&1.592&0--50--0&1181&1.734&0--50--0\\\cmidrule(l{2pt}r{2pt}){1-14}
\multirow{12}{*}{2}&\multirow{4}{*}{0.01}&1--50&83&26.99&81.06&26.38&29--14--7&82.77&26.93&20--20--10&83.65&27.27&7--30--13\\&
&51--100&56.89&24.54&55.62&23.61&42--0--8&56.45&24.08&32--3--15&56.91&24.39&17--13--20\\&
&101--150&100.7&4.921&100.5&5.562&18--27--5&101.5&5.351&5--39--6&102.2&5.136&1--48--1\\&
&151--200&103.1&1.436e-14&105.2&0.2744&0--50--0&105.1&0.3222&0--50--0&105.2&0.3047&0--50--0\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.05}&1--50&289.4&104.8&283.2&101.9&37--9--4&288&103.9&26--18--6&290.7&105.1&13--30--7\\&
&51--100&185.5&87.7&181.9&84.58&44--0--6&184&85.82&36--1--13&185.1&86.55&25--9--16\\&
&101--150&359.6&23.24&357.8&24.65&27--21--2&361.4&24.79&12--35--3&363.3&24.71&2--44--4\\&
&151--200&373.2&2.297e-13&379.5&1.547&0--50--0&380.2&1.603&0--50--0&379.6&1.72&0--50--0\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.2}&1--50&890.5&341.2&876.4&332.2&33--8--9&886.5&337&24--19--7&891.8&339.6&15--26--9\\&
&51--100&548.6&274&540.8&265&42--1--7&544.9&268.2&40--5--5&547.2&270.1&28--7--15\\&
&101--150&1121&78.79&1115&80.14&32--15--3&1122&80.65&19--25--6&1126&80.84&6--41--3\\&
&151--200&1167&4.594e-13&1178&2.186&0--50--0&1180&1.777&0--50--0&1181&1.365&0--50--0\\\cmidrule(l{2pt}r{2pt}){1-14}
\multirow{12}{*}{5}&\multirow{4}{*}{0.01}&1--50&82.89&26.89&80.18&26.47&35--10--5&82.03&26.82&28--17--5&83.19&27.09&21--22--7\\&
&51--100&57.45&23.4&55.25&22.07&46--0--4&56.51&22.63&44--0--6&57.13&22.94&35--2--13\\&
&101--150&100.3&5.205&98.99&6.439&31--15--4&100.4&6.181&17--26--7&101.3&5.851&7--38--5\\&
&151--200&103.1&1.436e-14&105.1&0.299&0--50--0&105.2&0.3053&0--50--0&105.3&0.2958&0--50--0\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.05}&1--50&288.1&104.2&276&99.2&45--2--3&282&101&40--7--3&286.2&102.6&35--10--5\\&
&51--100&185.9&83.57&179.9&79.17&44--3--3&183&80.87&41--5--4&184.9&81.91&31--13--6\\&
&101--150&357.1&25.23&347.7&27.29&41--2--7&353.3&26.81&35--13--2&357.4&26.77&23--20--7\\&
&151--200&373.1&0.1956&376&4.098&4--38--8&377.8&3.594&2--44--4&379.5&2.772&0--48--2\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.2}&1--50&891.6&337.8&868.9&324.9&42--2--6&881.1&329.8&40--7--3&888.7&333.4&27--10--13\\&
&51--100&556.6&262.2&546.6&251.8&45--0--5&552.3&256&38--1--11&555.2&258.4&26--5--19\\&
&101--150&1117&82.61&1101&84.57&37--10--3&1110&83.79&33--14--3&1117&83.62&25--20--5\\&
&151--200&1167&4.594e-13&1175&4.725&0--48--2&1178&4.139&0--50--0&1180&2.777&0--50--0\\\cmidrule(l{2pt}r{2pt}){1-14}
\multirow{12}{*}{10}&\multirow{4}{*}{0.01}&1--50&80.91&27.29&77.03&26.71&41--3--6&79.3&27.08&32--10--8&80.69&27.32&20--17--13\\&
&51--100&54.9&21.54&51.59&19.49&50--0--0&53.23&20.3&47--0--3&54.17&20.78&38--1--11\\&
&101--150&99.72&5.231&97.23&7.459&33--11--6&99.15&6.985&20--24--6&100.3&6.261&12--34--4\\&
&151--200&103.1&0.09412&105&0.3504&0--50--0&105.1&0.3296&0--50--0&105.2&0.2794&0--50--0\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.05}&1--50&282.7&105.5&268.5&99.61&50--0--0&275.1&101.5&43--2--5&279.9&103.2&32--9--9\\&
&51--100&179.9&77.93&172.3&72.08&49--0--1&176.4&74.58&43--2--5&178.9&76.24&27--10--13\\&
&101--150&356.2&23.7&344.8&28.83&44--1--5&351.2&27.92&37--11--2&355.8&26.49&21--21--8\\&
&151--200&372.6&0.7456&374.1&4.464&7--36--7&376.3&3.334&2--45--3&377.5&2.62&1--49--0\\\cmidrule(l{2pt}r{2pt}){2-14}
&\multirow{4}{*}{0.2}&1--50&877.9&340.8&850.1&324.3&45--4--1&865.3&330.6&36--8--6&873.5&334.4&24--17--9\\&
&51--100&544.6&247&532&233.3&41--4--5&539.2&238.8&30--12--8&543.5&242.3&17--21--12\\&
&101--150&1116&77.15&1093&82.69&45--2--3&1106&81.16&36--10--4&1115&79.89&22--21--7\\&
&151--200&1166&1.281&1167&9.098&8--32--10&1173&6.941&2--44--4&1176&5.929&1--48--1\\\cmidrule(l{2pt}r{2pt}){1-14}
\end{tabular}
\end{center}
\end{small
\end{table*}
We show the aggregated results in Table \ref{table:experimentsd} where a U-test \cite{Corder09} with 95\% confidence interval is used to determine statistical significance in each change.
The results show that outputs from both algorithms are very closely matched most of the time, with the greatest differences observed in low graph densities. Furthermore, we expect that GREEDY fairs better against POMC when the search space is small (low threshold levels). While this is observable, the opposite phenomenon when the threshold level is high can be seen more easily.
We see that during consecutive periods of high constraint thresholds, POMC's outputs initially fall behind GREEDY's, only to overtake them at later changes. This suggests that POMC rarely compromises its best solutions during those periods as the consequence of the symmetric submodular objective function. It also implies that restarting POMC from scratch upon a change would have resulted in significantly poorer results. On the other hand, POMC's best solutions follow GREEDY's closely during low constraint thresholds periods. This indicates that by maintaining feasible solutions upon changes, POMC keeps up with GREEDY in best objectives well within quadratic run time.
Comparing outputs from POMC with different interval settings, we see that those from runs with higher number of evaluations between changes are always better. However, the differences are minimal during the low constraint thresholds periods. This aligns with our theoretical results in the sense that the expected number of evaluations needed to guarantee good approximations depends on the constraint thresholds. As such, additional evaluations won't yield significant improvements within such small feasible spaces.
Comparing between different $k$ values, POMC seems to be at a disadvantage against GREEDY's best at increased $k$. This is expected since more partitions leads to more restrictive feasible search spaces, given everything else is unchanged, and small feasible spaces amplify the benefit of each greedy step. Nevertheless, POMC does not seem to fall behind GREEDY significantly for any long period, even when given few resources.
\section{Conclusions}
In this study, we have considered combinatorial problems with dynamic constraint thresholds, particularly the important classes of problems where the objective functions are submodular or monotone. We have contributed to the theoretical run time analysis of a Pareto optimization approach on such problems. Our results indicate POMC's capability of maintaining populations to efficiently adapt to changes and preserve good approximations. In our experiments, we have shown that POMC is able to maintain at least the greedy level of quality and often even obtains better solutions.
\section{Acknowledgements}
This work has been supported by the Australian Research Council through grants DP160102401 and DP190103894.
| 5cd292aa28780daacf444e58097c881456c9ced1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In recent years, much attention has been paid to studies of low-dimensional structures
in which new physical phenomena that have no analogues in bulk materials have been
discovered. Due to these new functional properties, these materials are considered
as very promising for their use in electronics. Ferroelectric superlattices
(SLs)---quasi-two-dimensional artificial periodic structures whose properties can
be tailored to obtain the necessary functionality---belong to this interesting
class of materials.
Most of the earlier experimental~\cite{JApplPhys.72.2840,JapJApplPhys.33.5192,
ApplPhysLett.65.1970,ApplPhysLett.70.321,ApplPhysLett.72.1394,PhysRevLett.80.4317,
ApplPhysLett.77.1520,ApplPhysLett.77.3257,JapplPhys.91.2290,PhysRevLett.88.097601,
ApplPhysLett.80.3581,ApplPhysLett.82.2118,JApplPhys.93.1180,PhysRevLett.95.177601,
ApplPhysLett.89.092905,Science.313.1614,AdvMater.19.4153,PhysRevLett.104.187601,
PhysSolidState.54.1026,PhysRevLett.109.167601,PhysSolidState.57.486,Nature.568.368,
SciRep.9.18948} and theoretical~\cite{PhysRevB.59.12771,PhysRevLett.84.5636,
ApplPhysLett.82.1586,PhysRevB.69.184101,Nature.433.395,PhysRevB.71.100103,
ApplPhysLett.87.052903,ApplPhysLett.87.102906,HandbookChap134,PhysRevB.76.020102,
ApplPhysLett.91.112914,Nature.452.732,PhysRevLett.101.087601,PhysRevB.79.024101,
PhysSolidState.51.2324,PhysSolidState.52.1448,PhysRevB.83.020104,PhysRevLett.107.217601,
PhysStatusSolidiB.249.789,PhysRevB.85.184105,PhysSolidState.55.1198,
ComputMaterSci.91.310,Nature.534.524,NanoLett.17.2246,ActaMater.152.155}
studies of ferroelectric superlattices with the perovskite
structure were performed on II-IV/II-IV or I-V/I-V type superlattices (here, the
numbers indicate the valence of atoms entering the $A$ and $B$~sites of the $AB$O$_3$
perovskite structure). In such superlattices, there are no double electric layers
at the interface between two dielectric materials, and therefore both the bulk
and interfaces of these superlattices remain macroscopically electrically neutral.
Studies of the SrTiO$_3$/LaAlO$_3$ heterostructures have revealed new interesting
phenomena that appear in these structures as a result of so-called polar
discontinuity---of a polarization jump induced by effective $\pm e/2$ charges
per unit cell which appear at the interface between SrTiO$_3$ and LaAlO$_3$
because the charged character of LaO$^+$ and AlO$_2^-$ layers~\cite{Nature.427.423}.
These phenomena include the appearance of a conducting layer near the interface
(a two-dimensional electron gas), its magnetism, and even
superconductivity~\cite{Nature.427.423, NatureMater.6.493, Science.317.1196}. These
phenomena can be controlled using an external electric field~\cite{Nature.456.624}.
The divergence of the electrostatic potential in such heterostructures, which
can result in the appearance of conductive layers at the interface between two
dielectrics, is called a polar catastrophe. The possibility of the appearance of
the two-dimensional electron gas at the interface between a ferroelectric and a
nonpolar dielectric in perovskites was systematically studied in \cite{SciRep.6.34667}.
Later it was realized that conductive layers can also be obtained in ferroelectric
structures without the polar discontinuity~\cite{PhysRevB.92.115406}. This made
it possible to create new types of electronic devices---ferroelectric structures
with switchable giant tunneling conductivity \cite{ApplPhysLett.107.232902,
PhysRevB.94.155420}; the idea of such devices was proposed earlier in
\cite{Science.313.181, NanoLett.9.427}. Note that the appearance of similar
phenomena can also be anticipated in epitaxial films of II-IV and I-V perovskites,
which are grown on DyScO$_3$, GdScO$_3$, NdScO$_3$, NdGaO$_3$, and LaGaO$_3$
substrates to create a biaxial strain in them, since the polar discontinuity
is possible in these structures~\cite{ChinPhysB.28.047101}.
Until now, theoretical studies of superlattices with the polar discontinuity have
focused on studying the distribution of polarization and electric field in these
structures and on finding the conditions for the appearance of a two-dimensional
electron gas at the interface~\cite{PhysRevB.79.100102,PhysRevLett.103.016804,
PhysRevB.80.045425,PhysRevB.80.165130,PhysRevB.80.241103,PhysRevB.81.235112,
PhysRevB.85.235109,PhysRevB.87.085305,SciRep.8.467,RCSAdv.9.35499,ChinPhysB.28.047101,
PhysChemChemPhys.21.8046}. The questions of the stability of the high-symmetry
polar structure in such SLs, possible phase transitions in them, and the physical
properties of low-symmetry phases were not analyzed. At the same time, the
ferroelectric and antiferrodistortive (AFD) instabilities characteristic of many
perovskites can result in strong distortions of the structure of SLs. This means
that the earlier predictions of their physical properties obtained without taking
these distortions into account may be incorrect.
One of the questions that remains insufficiently studied is the question about
the piezoelectric properties of superlattices with the polar discontinuity. The
calculations of these properties are limited to the calculations for the
high-symmetry tetragonal $P4mm$~phase of such SLs~\cite{JApplPhys.109.066107,
ChinPhysB.27.027701}. It is known that the record-high piezoelectric properties
of PbTiO$_3$-based solid solutions near the morphotropic boundary~\cite{JApplPhys.82.1804}
are due to the ease of inclination of the polarization vector from the
[100] direction to the [111] one under the influence of the electric
field~\cite{PhysRevLett.84.5423} or mechanical stress. The polar discontinuity
in SLs allows, under certain conditions, to obtain high values of irreversible
polarization in their structures. This is why it was interesting to check
whether it is possible to get high piezoelectric coefficients in superlattices
with the polar discontinuity using the strain-induced inclination of the
polarization vector. In Ref.~\cite{JApplPhys.109.066107}, SLs with the polar
discontinuity were already considered as a way to obtain stable, weakly
temperature-dependent piezoelectric properties.
\section{Calculation technique}
\begin{figure}
\centering
\includegraphics{fig1.ps}
\caption{Geometry of a typical superlattice.}
\label{fig1}
\end{figure}
In this work, the properties of seventeen free-standing [001]-oriented
short-period II-IV/I-V, II-IV/III-III, and I-V/III-III superlattices with the
polar discontinuity and a thickness of individual layers of one unit cell are
studied from first principles. For comparison, similar calculations are
performed for several superlattices with broken symmetry. The latter SLs have
no the polar discontinuity, but their high-symmetry polar structure is formed
by a sequence of layers with no mirror symmetry. In addition, the obtained
results are compared with the results for ordinary superlattices with the
mirror symmetry of the layer sequence. The geometry of the superlattices is
shown in Fig.~\ref{fig1}.
The choice of free-standing superlattices for our study stems from the fact that,
when any coherently strained short-period superlattice is grown on a substrate,
a transition layer containing misfit dislocations and other defects with a
thickness of $\sim$100~{\AA} appears at the interface as a result of a mismatch
between the equilibrium in-plane lattice parameter of the SL and that of the
substrate. In the defect-free part of the superlattice, the in-plane lattice
parameter relaxes to the lattice parameter of the free-standing superlattice.
This is why in the majority of rather thick experimentally grown superlattices,
the in-plane lattice parameter is close to that of free-standing superlattices.
First-principles calculations were performed within the density functional theory
using the \texttt{ABINIT} program~\cite{abinit} and norm-conserving pseudopotentials
constructed according to the RRKJ scheme~\cite{PhysRevB.41.1227} in the local
density approximation (LDA), like in~\cite{PhysSolidState.51.362}. The cutoff
energy was 30~Ha (816~eV) with the exception of Ta-containing systems in which
it was 40~Ha (1088~eV). Integration over the Brillouin zone was carried out using
a 8$\times$8$\times$4 Monkhorst--Pack mesh for high-symmetry structures and
meshes with equivalent density of ${\bf k}$-points for low-symmetry phases.
The equilibrium lattice parameters and atomic positions were calculated by
relaxing forces acting on the atoms to values less than $2 \cdot 10^{-6}$~Ha/Bohr
(0.1~meV/{\AA}). The phonon spectra, the tensors of piezoelectric stress
coefficients $e_{i\mu}$, and the elastic compliance tensors $S_{\mu\nu}$ were
calculated using the density-functional perturbation theory. The $e_{i \mu}$
values were then converted to the piezoelectric strain coefficients $d_{i\nu}$
using the formula $d_{i\nu} = e_{i\mu} S_{\mu\nu}$, and for monoclinic structures
the tensor components were transformed to the standard setting of the monoclinic
unit cell, in which the polarization vector lies in the $xz$~plane.
\section{Results and discussion}
\subsubsection{Search for the ground-state structure}
It is known that the electrical conductivity of materials is usually a factor
that prevents the practical use of their ferroelectric properties. Since the
experiments on structures with the polar discontinuity often revealed metallic
conductivity at the interface, it was necessary first to ensure that the
superlattices under study are insulating. The calculations confirmed that
in all superlattices studied in the work, the conduction band is separated
from the valence band by a sufficiently large energy gap (see Table~S1
in the Supplementary data), and therefore all studied SLs are dielectrics.
Since the layer sequence in considered superlattices does not admit the reversal
of $z \to -z$ (Fig.~\ref{fig1}), the superlattices are always polar and their
high-symmetry phase is described by the $P4mm$~space group. However, this
structure can exhibit various instabilities characteristic of crystals with
the perovskite structure: either the ferroelectric instability or the
antiferrodistortive one, or both of them simultaneously. In this work, the
ground state of
superlattices was searched in the traditional way~\cite{PhysSolidState.51.2324,
PhysSolidState.52.1448}. First, the structures resulting from the condensation
of all unstable phonons found in the phonon spectrum of the $P4mm$~phase were
calculated taking into account a possible degeneracy of phonon modes. Then, by
calculating the phonon frequencies at all high-symmetry points of the Brillouin
zone and the elastic tensor for these distorted structures, the stability of
the obtained solutions was checked. In the case when an instability was found
in any of these structures, the search for the ground state was continued until
a structure whose phonon spectrum has no unstable modes and whose matrix composed
of the elastic tensor components in the Voigt notation is positive definite is
found. In this case, the conclusion that a \emph{stable} phase which can be a
possible ground-state structure can be made. The problem here is that in chains
of phases generated by different octahedra rotations, several stable states can
be found, as was recently demonstrated for SrTiO$_3$~\cite{PhysSolidState.58.300}.
In this case, the ground state is the stable phase with the lowest total energy;
other stable phases should be considered as metastable. If the energy of such
metastable solutions differs little from the energy of the ground state, then
these solutions should be considered as phases that can be observed
in experiment. For them, as well as for the ground state, an analysis of their
physical properties should also be carried out.
Calculations of the phonon spectra show that in most studied superlattices
with the polar discontinuity (except for BaTiO$_3$/LaAlO$_3$, SrTiO$_3$/LaAlO$_3$,
PbTiO$_3$/KTaO$_3$, PbTiO$_3$/LaAlO$_3$, and KTaO$_3$/LaAlO$_3$) the high-symmetry
$P4mm$~phase exhibits the ferroelectric instability with respect to the in-plane
distortion of the structure or, in other words, to the inclination of the polarization
vector. The phonon spectrum of one of such superlattices, KNbO$_3$/BaTiO$_3$, is
shown in Fig.~\ref{fig2}. It is seen that in addition to the instability at the
$\Gamma$ point, the instabilities also appear at the $X$, $R$, and $Z$~points of
the Brillouin zone. We have already encountered a similar situation in
KNbO$_3$/KTaO$_3$~\cite{PhysStatusSolidiB.249.789} and
BaTiO$_3$/BaZrO$_3$~\cite{PhysSolidState.55.1198} SLs.
\begin{figure}
\centering
\includegraphics{fig2.ps}
\caption{\label{fig2}Phonon spectrum of the KNbO$_3$/BaTiO$_3$ superlattice
(a) in the high-symmetry $P4mm$~phase and (b) in the ground-state $Cm$~phase.
The numbers near the curves indicate the symmetry of unstable modes.}
\end{figure}
\begin{figure}
\centering
\includegraphics{fig3.ps}
\caption{\label{fig3}Phonon spectrum of the BaTiO$_3$/LaAlO$_3$ superlattice
(a) in the high-symmetry $P4mm$~phase and (b) in the ground-state $P4bm$~phase.
The numbers near the curves indicate the symmetry of unstable modes.}
\end{figure}
\begin{figure}
\centering
\includegraphics{fig4.ps}
\caption{\label{fig4}Phonon spectrum of the KNbO$_3$/SrZrO$_3$ superlattice
(a) in the high-symmetry $P4mm$~phase and (b) in the ground-state $Pc$~phase.
The numbers near the curves indicate the symmetry of unstable modes.}
\end{figure}
An instability zone, which is observed as a band of imaginary phonon frequencies
on the $\Gamma$--$Z$--$R$--$X$--$\Gamma$ line (imaginary frequencies are
represented in the figure by negative values), is a consequence of the
ferroelectric instability in the ...--Ti--O--... chains propagating in the plane
of the SL---the so-called chain instability~\cite{PhysRevLett.74.4067}. Indeed,
an analysis of the eigenvectors of these modes shows that at all the
above-mentioned points of the Brillouin zone the out-of-phase, transverse,
$xy$-polarized displacements of Ti and O atoms in chains propagating in the [100]
and [010] directions dominate in the atomic motion. In
Ba(Sr)TiO$_3$/KNb(Ta)O$_3$ superlattices, the displacements of the Nb(Ta)
atoms are small, whereas in superlattices containing no Ti atoms as well as
in the PbTiO$_3$/KNbO$_3$ SL the chain instability is observed mainly in
...--Nb--O--... chains. In BiScO$_3$/LaAlO$_3$ and PbTiO$_3$/BiScO$_3$
superlattices, additional out-of-phase, $xy$-polarized displacements of Bi and
oxygen atoms lying in the same $xy$~plane are stronger; this displacement
pattern is characteristic of BiScO$_3$. A similar displacement pattern is also
typical for unstable modes at the $Z$~point in these superlattices. At the
$\Gamma$~point, the described displacement pattern corresponds to a doubly
degenerate ferroelectric $E$~mode ($\Gamma_5$).%
\footnote{When numbering irreducible representations in this paper, we follow
their classification accepted at Bilbao Crystallographic Server~\cite{Bilbao}.}
\begin{table*}
\caption{\label{table1}Energies (per 10-atom supercell) of different low-symmetry
phases resulting from the condensation of unstable phonons in the high-symmetry
$P4mm$~phase of short-period KNbO$_3$/BaTiO$_3$ and KNbO$_3$/SrZrO$_3$ superlattices.}
\begin{ruledtabular}
\begin{tabular}{cccccccc}
Phase & Unstable & Energy & Phase & Unstable & Energy \\
& phonon & (meV) & & phonon & (meV) \\
\hline
\multicolumn{6}{c}{KNbO$_3$/BaTiO$_3$ superlattice} \\
$P4mm$ & --- & 0 & $Pm$ & $\Gamma_5~(\eta,0)$ & $-$27.8 \\
$Abm2$ & $R_4$ & $-$13.8 & $Cmc2_1$ & $Z_5~(\eta,\eta)$ & $-$31.0 \\
$Pma2$ & $X_4$ & $-$14.8 & $Cm$ & $\Gamma_5~(\eta,\eta)$ & $-$38.6 \\
$Pmc2_1$ & $Z_5~(\eta,0)$ & $-$21.3 & \\
\multicolumn{6}{c}{KNbO$_3$/SrZrO$_3$ superlattice} \\
$P4mm$ & --- & 0 & $Pc$ & $Z_5(\eta,0)+\Gamma_5(0,\xi)$ & $-$19.2 \\
$I4cm$ & $A_4$ & $-$1.3 & $Pma2$ & $M_5~(\eta,0)$ & $-$19.9 \\
$P4bm$ & $M_4$ & $-$2.1 & $Cm$ & $\Gamma_5~(\eta,\eta)$ & $-$20.0 \\
$Abm2$ & $R_4$ & $-$4.7 & $P2$ & $X_3+X_4$ & $-$20.7 \\
$Pma2$ & $X_4$ & $-$6.7 & $Cc$ & $A_4+\Gamma_5~(\eta,\eta)$ & $-$24.1 \\
$Pmm2$ & $X_3$ & $-$12.4 & $Pm$ & $X_3+\Gamma_5~(0,\eta)$ & $-$24.9 \\
$Pmc2_1$ & $Z_5~(\eta,0)$ & $-$12.9 & $Cmm2$ & $M_5~(\eta,\eta)$ & $-$26.9 \\
$Pm$ & $\Gamma_5~(\eta,0)$ & $-$13.7 & $Pm$ & $X_3+\Gamma_5~(\eta,0)$ & $-$27.6 \\
$Pc$ & $R_4+\Gamma_5~(\eta,0)$ & $-$15.4 & $Pm$ & $M_5~(\eta,0)+\Gamma_5~(\xi,\xi)$ & $-$36.9 \\
$Cm$ & $A_4+\Gamma_5(\eta,0)$ & $-$16.0 & $Cm$ & $M_4+\Gamma_5~(\eta,0)$ & $-$39.4 \\
$Pc$ & $X_4+\Gamma_5~(\eta,0)$ & $-$17.0 & $Cm$ & $M_5~(\eta,\eta)+\Gamma_5~(0,\xi)$ & $-$40.2 \\
$Cmc2_1$ & $Z_5~(\eta,\eta)$ & $-$18.9 & $Pc$ & $M_4+\Gamma_5~(\eta,\eta)$ & $-$51.9 \\
\end{tabular}
\end{ruledtabular}
\end{table*}
Of two possible polar phases resulting from the condensation of the unstable
ferroelectric $E$~mode, the $Cm$~phase with atomic displacements along
the [110] direction has the lowest energy in all systems except for
KNbO$_3$/LaAlO$_3$, PbTiO$_3$/LaGaO$_3$, and SrTiO$_3$/BiScO$_3$. The
predominant [111] displacements correlate
with the fact that at normal conditions, bulk BaTiO$_3$ and KNbO$_3$ are
polarized along the [111] direction. The same polarization direction is also
characteristic of BiScO$_3$ when the octahedral rotations are turned off.
As an example, the energies of different low-symmetry phases for the
KNbO$_3$/BaTiO$_3$ superlattice are given in Table~\ref{table1}. It is seen that
the structures obtained from the condensation of $Z_5$, $R_4$, and $X_4$~unstable
phonons have a higher energy as compared to the energy of the $Cm$~phase.
The lattice parameters and atomic positions in ground-state structures of this
and other superlattices studied in this work are given in Tables~S2--S26
of the Supplementary data.
The absence of the ferroelectric instability in the $P4mm$~phase of
PbTiO$_3$/KTaO$_3$ and PbTiO$_3$/LaAlO$_3$ SLs can be explained by a tendency
of bulk PbTiO$_3$ to be polarized along the [001] axis. This instability
was also absent in SrTiO$_3$/LaAlO$_3$ and KTaO$_3$/LaAlO$_3$ superlattices,
in which both constituents are nonpolar. In addition, the ferroelectric
instability did not appear in the BaTiO$_3$/LaAlO$_3$ superlattice, in which
its absence is a result of the strong (by 2.3\%) in-plane compression of
the BaTiO$_3$ layers. Our calculations of the influence of strain on the
ground-state structure of BaTiO$_3$ showed that the biaxial compression
of 1\% is already sufficient for the $P4mm$~phase to become the most stable
polar phase in this material, in agreement with~\cite{PhysRevLett.80.1988,
PhysRevB.69.212101}.
Along with the ferroelectric instability, in the high-symmetry $P4mm$~phase
of twelve superlattices with the polar discontinuity (except for
KNbO$_3$/PbTiO$_3$, KNbO$_3$/BaTiO$_3$, KNbO$_3$/BaZrO$_3$, BaTiO$_3$/KTaO$_3$,
and KTaO$_3$/PbTiO$_3$), an antiferrodistortive instability with octahedra
rotations around the $z$~axis is observed. This is clearly seen by the
appearance of unstable phonons on the $M$--$A$ line (Fig.~\ref{fig3}).
The appearance of the AFD instability in a superlattice clearly correlates
with its existence in one or both components of the SL (SrTiO$_3$, SrZrO$_3$,
LaAlO$_3$, BiScO$_3$). A comparison of the energies of phases resulting from
the condensation of phonons at $M$ and $A$~points of the Brillouin zone shows
that of two phases resulting from the $M_4$~phonon condensation (space
group $P4bm$) and from the $A_4$~phonon condensation (space group $I4cm$),
the $P4bm$~phase was energetically more favorable in all cases except for
the PbTiO$_3$/BiScO$_3$ superlattice.
The most complex picture is observed in superlattices, in which both the
ferroelectric and AFD instabilities are simultaneously present in the
$P4mm$~phase. Since the energy difference between the $P4bm$ and $I4cm$~phases
is small and both phases can exhibit the ferroelectric instability, it was
necessary to consider all polar subgroups of these phases when searching
for the ground state. The KNbO$_3$/SrZrO$_3$ superlattice is an example of
such a system. Its phonon spectrum is shown in Fig.~\ref{fig4}, and the
energies of different phases are given in Table~\ref{table1}.
The calculations show that the ferroelectric instability of the $P4bm$ and
$I4cm$~phases is characteristic of KNbO$_3$/SrTiO$_3$, KNbO$_3$/SrZrO$_3$,
SrTiO$_3$/KTaO$_3$ superlattices and all SLs containing BiScO$_3$. Of two phases,
$Cm(2)$ and $Pc$, into which the $P4bm$ structure can transform upon the polar
distortion, the $Pc$~phase with polarization along the [110] direction of the
pseudocubic cell had lower energy in all SLs except for PbTiO$_3$/LaGaO$_3$.
Of two phases, $Cc$ and one more phase of the $Cm$~symmetry, $Cm(3)$, into which
the $I4cm$ structure can transform upon the polar distortion, the $Cc$~phase whose
polarization is also directed along the [110] direction of the pseudocubic cell
usually had a lower energy. In BiScO$_3$/LaAlO$_3$ and KNbO$_3$/BiScO$_3$
superlattices, it was the $Cm(3)$~phase which had a lower energy, but this
phase has never become the ground state. In KNbO$_3$/SrTiO$_3$ and
SrTiO$_3$/KTaO$_3$ superlattices, the energy difference between the $Cc$ and
$Pc$~phases was only 0.5--0.7~meV, whereas in other systems it was within
8--80~meV. Since the calculations proved the stability of both $Cc$ and
$Pc$~phases
in the two superlattices, the $Cc$~phase should be considered as a metastable
one. As the energy difference between the two phases is very small, there is
a real possibility that both structures can occur in an experiment simultaneously.
That is why the properties of these superlattices were calculated below for both
phases, $Pc$ and $Cc$.
The PbTiO$_3$/LaGaO$_3$ superlattice is the only one in which the $Cm(2)$~phase
is the ground state. It originates from the $P4bm$~phase and has the polarization
directed along the $[x0z]$ direction of the pseudocubic cell.
In addition to the unstable $M_4$~mode discussed above, in phonon spectra of
the $P4mm$~phase of nine SLs (KNbO$_3$/SrTiO$_3$, KNbO$_3$/SrZrO$_3$,
SrTiO$_3$/LaAlO$_3$, PbTiO$_3$/LaAlO$_3$, PbTiO$_3$/LaGaO$_3$, and all SLs
containing BiScO$_3$) one more doubly degenerate unstable $M_5$~mode is observed
(Fig.~\ref{fig4}). This mode describes the octahedra rotations around one or
both of the $x$ and $y$~axes. The distortions described by this mode with
the order parameters of ($\eta$, 0) and ($\eta$, $\eta$) result in the $Pma2$
and $Cmm2$ phases whose energies are usually higher than that of the $P4bm$~phase
(except for PbTiO$_3$/LaGaO$_3$, SrTiO$_3$/BiScO$_3$,
KNbO$_3$/BiScO$_3$, and KNbO$_3$/SrZrO$_3$ SLs). Both the $Pma2$ and $Cmm2$
phases are characterized by the ferroelectric instability. An analysis of the
polar subgroups of these phases (with $Pc$, $Cm$, and $Pm$ space groups) shows
that the $Pc$~phase has the lowest energy among them. It should be noted,
however, that in the BiScO$_3$/LaAlO$_3$ and PbTiO$_3$/LaGaO$_3$ superlattices,
the structure of this phase differs from that of the $Pc$~phase, which
originates from the $P4bm$~phase. In the BiScO$_3$/LaAlO$_3$ superlattice,
this new $Pc(2)$ phase becomes the ground-state structure because its energy is
lower than that of the $Pc$~phase originating from the $P4bm$~phase. In the
PbTiO$_3$/LaGaO$_3$ superlattice, the energy of the $Pc(2)$~phase is intermediate
between those of the $Pc$~phase and ground-state $Cm(2)$~structure. As the
$Pc(2)$~phase demonstrates its stability with respect to all acoustic and
optical distortions, we conclude that this is a metastable phase.
In two latter superlattices, BiScO$_3$/LaAlO$_3$ and PbTiO$_3$/LaGaO$_3$,
we encounter a situation when the metastable solutions appear. The cause of
this metastability is that the relaxation paths of the $P4bm$ and $Pma2$
($Cmm2$) phases, in which the rotations are described by the $M_4$ and $M_5$
irreducible representations, are separated by a potential barrier. In other
superlattices, these barriers are absent, and the structures, regardless
of the initial rotation pattern, relax to the same ground-state $Pc$~structure.
It may seem strange why the same phase appears as a result of condensation of
distortions described by two different irreducible representations. However,
it should be taken into account that the vertical axis of the octahedra in
the $Pc$~phase is slightly inclined, which means that, in fact, this phase is
described by two non-zero rotations around the coordinate axes. This explains
why two structures, in which rotational distortions are described by different
irreducible representations, relax to the same phase when the polar distortions
are turned on.
In SrTiO$_3$/LaAlO$_3$ and PbTiO$_3$/LaAlO$_3$ superlattices, both the $Pma2$
and $Cmm2$ phases relax to the $P4bm$~structure when the polar distortions
are turned on. Here, as in the case of the KNbO$_3$/LaAlO$_3$ SL, the strong
AFD instability of LaAlO$_3$ results in a complete suppression of the
ferroelectric instability which was present in the high-symmetry $P4mm$~phase.
In all BiScO$_3$-containing superlattices with the polar discontinuity, one
more unstable mode, $A_5$, appears in the phonon spectrum at the $A$~point.
An analysis of chains of phases induced by the corresponding distortions
showed that all obtained phases have the energy much higher than that of the
ground-state structure.
\begin{table}
\caption{\label{table2}Calculated polarization and energy gain resulting from
the in-plane ferroelectric distortion for all studied short-period superlattices.}
\begin{ruledtabular}
\begin{tabular}{cccccc}
Superlattice & Space & $P_x$ & $P_y$ & $P_z$ & $\Delta E$ \\
& group & \multicolumn{3}{c}{(C/m$^2$)} & (meV) \\
\hline
KNbO$_3$/PbTiO$_3$ & $Cm$ & 0.181 & 0 & 0.567 & 3.3 \\
KNbO$_3$/BaTiO$_3$ & $Cm$ & 0.338 & 0 & $-$0.033 & 38.6 \\
KNbO$_3$/BaZrO$_3$ & $Cm$ & 0.229 & 0 & 0.181 & 33.8 \\
KNbO$_3$/SrTiO$_3$ & $Pc$ & 0.281 & 0 & 0.066 & 53.3;~9.0\footnotemark[1] \\
KNbO$_3$/SrZrO$_3$ & $Pc$ & 0.268 & 0 & 0.187 & 51.9;~49.8\footnotemark[1] \\
\hline
BaTiO$_3$/KTaO$_3$ & $Cm$ & 0.180 & 0 & 0.130 & 9.7 \\
BaTiO$_3$/LaAlO$_3$ & $P4bm$ & 0 & 0 & 0.057 & --- \\
\hline
SrTiO$_3$/KTaO$_3$ & $Pc$ & 0.105 & 0 & 0.141 & 11.9;~0.8\footnotemark[1] \\
SrTiO$_3$/LaAlO$_3$ & $P4bm$ & 0 & 0 & 0.012 & --- \\
SrTiO$_3$/BiScO$_3$ & $Pc$ & 0.501 & 0 & 0.329 & 671.1;~331.1\footnotemark[1] \\
\hline
PbTiO$_3$/KTaO$_3$ & $P4mm$ & 0 & 0 & 0.454 & --- \\
PbTiO$_3$/LaAlO$_3$ & $P4bm$ & 0 & 0 & 0.119 & --- \\
PbTiO$_3$/LaGaO$_3$ & $Cm(2)$ & 0.073 & 0 & $-$0.192 & 24.0\footnotemark[1] \\
PbTiO$_3$/BiScO$_3$ & $Pc$ & 0.710 & 0 & 0.330 & 865.5;~349.3\footnotemark[1] \\
\hline
KNbO$_3$/LaAlO$_3$ & $P4bm$ & 0 & 0 & 0.061 & --- \\
KTaO$_3$/LaAlO$_3$ & $P4bm$ & 0 & 0 & 0.109 & --- \\
KNbO$_3$/BiScO$_3$ & $Pc$ & 0.558 & 0 & 0.241 & 381.3;~281.5\footnotemark[1] \\
\hline
KNbO$_3$/NaTaO$_3$ & $Pc$ & 0.302 & 0 & 0.119 & 26.5;~26.4\footnotemark[1] \\
BaTiO$_3$/SrZrO$_3$ & $Pc$ & 0.234 & 0 & 0.072 & 186.8 \\
BaTiO$_3$/SrSnO$_3$ & $Pc$ & 0.189 & 0 & 0.072 & 275.9 \\
BiScO$_3$/LaAlO$_3$ & $Pc(2)$ & 0.493 & 0 & $-$0.116 & 535.1;~189.3\footnotemark[1] \\
\hline
BaTiO$_3$/SrTiO$_3$ & $Cm$ & 0.233 & 0 & 0.061 & 4.8 \\
KNbO$_3$/KTaO$_3$ & $Cm$ & 0.223 & 0 & 0.119 & 11.7 \\
PbTiO$_3$/PbZrO$_3$ & $Pc$ & 0.586 & 0 & 0.314 & 488.2;~147.4\footnotemark[1] \\
KNbO$_3$/NaNbO$_3$ & $Pmc2_1$ & 0.502 & 0 & 0 & 104.5;~98.6\footnotemark[2] \\
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{As compared to the non-polar $P4bm$~phase.}
\footnotetext[2]{As compared to the non-polar $P4/mbm$~phase.}
\end{table}
The space groups of the energetically most favorable phases obtained for all
studied superlattices are given in Table~\ref{table2}. It is seen that in
superlattices exhibiting only the ferroelectric instability in the high-symmetry
$P4mm$~phase, the $Cm$~phase is the ground state. In superlattices exhibiting
only the AFD instability, the $P4bm$~phase is the ground state. And, finally,
in superlattices in which both instabilities are present, the $Pc$~phase is
the ground state, with the exception of the KNbO$_3$/LaAlO$_3$ and PbTiO$_3$/LaGaO$_3$
superlattices. In the first of these exceptions, the AFD instability suppresses
the ferroelectric one, and the $P4bm$~phase becomes the ground state. In the second
of them, the ground-state structure is $Cm(2)$. The discrepancy between our
results and those obtained for the KNbO$_3$/BaTiO$_3$ SL in \cite{PhysRevB.87.085305}
stems from the fact that the calculations in Ref.~\cite{PhysRevB.87.085305} were
performed for the superlattice clamped on the SrTiO$_3$ substrate; at these
boundary conditions, the $P4mm$~phase is indeed the ground state.
Table~\ref{table2} also presents the energy gain $\Delta E$ per 10-atom formula
unit, which results from the in-plane ferroelectric distortion. The obtained
values show that, at room temperature, the predicted ground-state structures
are likely to be observed in the KNbO$_3$/BaTiO$_3$, KNbO$_3$/BaZrO$_3$,
KNbO$_3$/SrZrO$_3$, SrTiO$_3$/BiScO$_3$, PbTiO$_3$/BiScO$_3$, and
KNbO$_3$/BiScO$_3$ superlattices. For the ground-state structures of all
superlattices, the calculated spontaneous polarization and components of the
piezoelectric tensor are presented in the following sections.
\subsubsection{Calculation of polarization}
Classical electrostatics of an electrically neutral interface between two
dielectrics requires that the electric displacement field components normal to
this interface are equal in two materials. If the materials are ferroelectrics
and have different spontaneous polarizations, a bound electric charge appears
at the interface, and the electric field generated by it makes the electric
displacement fields equal in two materials.
In superlattices with the polar discontinuity, a violation of the order of
charged $A$O and $B$O$_2$ planes generates an additional electrostatic perturbation
at the interfaces and creates a polarization jump of $\Delta P = e / 2a^2$ in
superlattices with a change in ionic charges of components by~$e$ and
$\Delta P = e / a^2$ in superlattices with a change in the ionic charges by
$2e$ (here $a$ is the in-plane lattice parameter of the
superlattice)~\cite{PhysRevB.79.100102}. An elegant solution to this problem
within the modern theory of polarization, which is valid in the general case,
was proposed in~\cite{PhysRevB.80.241103}. An application of this approach to
our superlattices enables us to calculate the electric displacement field in
them and to use it to determine the \emph{average} polarization in the SLs. In
this work, we are interested precisely in this property. The polarization values
in individual layers of a SL can be calculated by correcting the obtained average
polarization taking into account the jump in the ionic contribution to the Berry
phase at the interface and dielectric constants of individual components.%
\footnote{The periodicity of a superlattice assumes that the electric field
strengths ${\cal E}_1$ and ${\cal E}_2$ in its layers satisfy the condition
${\cal E}_1 x_1 + {\cal E}_2 x_2 = 0$, where $x_1$ and $x_2$ are the
thicknesses of individual layers. This condition, combined with the equation
$(P_1 + \epsilon_1 {\cal E}_1) - (P_2 + \epsilon_2 {\cal E}_2) = \Delta P$,
where $P_1$, $P_2$, $\epsilon_1$, and $\epsilon_2$ are spontaneous
polarizations and dielectric constants of two layers, gives a solution to
this problem in the linear approximation. Solving a nonlinear problem
requires more complex calculations.}
When calculating the polarization by the Berry phase method, it should be borne
in mind that the ionic contributions to the Berry phase are different in
nonpolar $Pm{\bar 3}m$ phases of I-V, II-IV, and III-III
perovskites~\cite{PhysicsofFerroelectrics2}. That is why for each SL with the
polar discontinuity it is necessary first to find the Berry phase for the nonpolar
structure before calculating the polarization. Unfortunately, in our case,
the calculation of the $z$~component of the Berry phase becomes a problem
because it is impossible to reverse the polarization or construct a non-polar
structure for the high-symmetry $P4mm$~phase, which does not have $\sigma_z$
mirror plane. To \emph{estimate} $P_z$, we considered unrelaxed structures
with ideal atomic positions corresponding to the cubic perovskite structure in
both layers. In this structure, the electron contribution to the Berry phase
is nonzero because of the redistribution of the electron density between the
layers, and the ionic contribution reflects the difference in the Berry
phases of individual perovskites. The average polarization in a superlattice
was calculated using standard formulas from a change in the Berry phase upon
the transition from the above-described unrelaxed structure to the ground-state one.
To correctly determine the polarization from the change in the Berry phase,
which is well-defined modulo $2\pi$, for all superlattices the calculations
were also performed for at least one intermediate point at which the atoms
are halfway between the ground-state and unrelaxed structures.
The calculated polarizations are given in Table~\ref{table2}. The components
of the polarization vector are given relative to the axes of the standard
crystallographic settings for tetragonal and monoclinic unit cells (their axes
are often rotated in the $xy$~plane relative to each other by 45$^\circ$).
A comparison of polarizations calculated for structures with and without
octahedral rotations shows that in SLs with the $Pc$ ground-state structure,
the neglect of these rotations can result in an error in determining $P_z$
up to 30\% and, in some cases, even in an error in the sign of this quantity.
For the $P4mm$~phase, our values reasonably agree with the published values
of $P_z = 0.532$~C/m$^2$ for PbTiO$_3$/KNbO$_3$~\cite{ChinPhysB.27.027701},
$P_z = 0.202$~C/m$^2$ for PbTiO$_3$/LaAlO$_3$~\cite{ChinPhysB.27.027701}, and
$P_z = 0.38$~C/m$^2$ for PbTiO$_3$/KTaO$_3$~\cite{ChinPhysLett.33.026302} SLs.
\subsubsection{Piezoelectric properties}
The reason for our interest to ferroelectric instability in superlattices
with the polar discontinuity is that in such SLs it is possible to obtain rather
high piezoelectric coefficients resulting from the in-plane ferroelectric
phase transitions appearing in them. The literature data on the piezoelectric
properties of such superlattices are limited to calculations for the
high-symmetry $P4mm$~phase
of PbTiO$_3$/LaAlO$_3$ and KNbO$_3$/PbTiO$_3$ SLs~\cite{JApplPhys.109.066107,
ChinPhysB.27.027701} and of the PbTiO$_3$/KTaO$_3$ one~\cite{ChinPhysLett.33.026302}.
As the ground-state structure of the first two superlattices differs from $P4mm$,
more correct calculations for these SLs are needed. For other superlattices
considered in this work, data on their piezoelectric properties are absent.
\begin{table*}
\caption{\label{table3}Nonzero components of the piezoelectric tensor $d_{i\nu}$
(in~pC/N) for the ground-state structures of all studied short-period superlattices.}
\begin{ruledtabular}
\begin{tabular}{cccccccccccc}
Superlattice & $d_{11}$ & $d_{12}$ & $d_{13}$ & $d_{15}$ & $d_{24}$ & $d_{26}$ & $d_{31}$ & $d_{32}$ & $d_{33}$ & $d_{35}$ \\
\hline
KNbO$_3$/PbTiO$_3$ & 13.4 & 3.7 & $-$13.9 & 158.3 & 273.3 & 137.3 & $-$3.0 & $-$2.9 & $-$10.9 & $-$0.8 \\
KNbO$_3$/BaTiO$_3$ & 23.5 & 9.6 & $-$12.7 & 3.8 & $-$2.8 & 106.4 & 10.7 & 7.9 & $-$16.3 & 17.4 \\
KNbO$_3$/BaZrO$_3$ & 16.3 & 4.7 & $-$11.7 & 5.6 & 7.9 & 20.8 & $-$8.9 & $-$7.3 & 16.7 & 1.8 \\
KNbO$_3$/SrTiO$_3$ & 35.9 & 15.3 & $-$19.1 & 8.3 & 41.7 & 151.0 & $-$22.0 & $-$12.1 & 16.1 & 9.4 \\
KNbO$_3$/SrZrO$_3$ & 30.8 & 10.1 & $-$18.8 & 9.1 & 11.0 & 85.5 & $-$18.1 & $-$9.8 & 17.7 & 12.9 \\
\hline
BaTiO$_3$/KTaO$_3$ & 19.2 & 9.5 & $-$18.5 & 2.5 & 7.9 & 25.3 & $-$18.1 & $-$11.6 & 43.4 & 10.1 \\
BaTiO$_3$/LaAlO$_3$ & --- & --- & --- & 1.4 & 1.4 & --- & 3.0 & 3.0 & $-$4.6 & --- \\
\hline
SrTiO$_3$/KTaO$_3$ & 39.7 & 22.8 & $-$27.6 & 28.2 & 46.9 & 41.2 & $-$18.4 & $-$13.0 & 26.1 & $-$4.2 \\
SrTiO$_3$/LaAlO$_3$ & --- & --- & --- & 2.0 & 2.0 & --- & 1.0 & 1.0 & $-$1.6 & --- \\
SrTiO$_3$/BiScO$_3$ & 23.6 & 5.4 & $-$6.0 & 14.3 & 17.9 & 24.6 & $-$4.7 & 0.6 & 11.8 & 12.2 \\
\hline
PbTiO$_3$/KTaO$_3$ & --- & --- & --- & 91.4 & 91.4 & --- & $-$8.8 & $-$8.8 & 41.2 & --- \\
PbTiO$_3$/LaAlO$_3$ & --- & --- & --- & 158.6 & 158.6 & --- & $-$4.3 & $-$4.3 & $-$2.8 & --- \\
PbTiO$_3$/LaGaO$_3$ & 19.1 & $-$5.8 & $-$3.0 & $-$29.6 & 15.1 & $-$35.1 & 7.8 & $-$2.9 & $-$18.5 & 4.3 \\
PbTiO$_3$/BiScO$_3$ & 31.1 & 2.7 & $-$7.4 & 5.1 & 1.5 & 47.0 & $-$2.3 & $-$4.0 & 18.5 & 13.6 \\
\hline
KNbO$_3$/LaAlO$_3$ & --- & --- & --- & 5.7 & 5.7 & --- & 1.1 & 1.1 & $-$0.8 & --- \\
KTaO$_3$/LaAlO$_3$ & --- & --- & --- & 5.3 & 5.3 & --- & $-$0.8 & $-$0.8 & 3.7 & --- \\
KNbO$_3$/BiScO$_3$ & 19.2 & 9.2 & $-$5.9 & 10.0 & 23.2 & 51.2 & $-$0.5 & 3.9 & 2.1 & 6.9 \\
\hline
KNbO$_3$/NaTaO$_3$ & 90.2 & 19.8 & $-$97.1 & 24.3 & $-$16.0 & 70.2 & 154.6 & 43.1 & $-$209.2 & 73.5 \\
BaTiO$_3$/SrZrO$_3$ & 23.3 & 4.9 & $-$12.8 & 1.2 & 2.6 & 19.3 & 3.9 & 4.0 & $-$7.8 & 5.1 \\
BaTiO$_3$/SrSnO$_3$ & 23.9 & 6.6 & $-$12.7 & $-$7.1 & $-$6.4 & 23.2 & $-$1.0 & $-$2.2 & 3.2 & 2.1 \\
BiScO$_3$/LaAlO$_3$ & 30.2 & $-$1.1 & $-$5.9 & $-$5.7 & 8.3 & 23.1 & 0.3 & 1.3 & $-$3.5 & 10.5 \\
\hline
BaTiO$_3$/SrTiO$_3$ & 103.5 & 37.4 & $-$117.1 & $-$59.5 & 21.1 & 187.9 & $-$287.3 & $-$131.5 & 460.5 & 288.1 \\
KNbO$_3$/KTaO$_3$ & 108.3 & 60.4 & $-$157.1 & $-$15.9 & 18.7 & 123.9 & $-$181.6 & $-$111.3 & 315.5 & 45.4 \\
PbTiO$_3$/PbZrO$_3$ & 62.1 & 12.5 & $-$38.8 & 3.2 & 29.9 & 132.7 & $-$45.8 & $-$19.5 & 70.4 & 26.6 \\
KNbO$_3$/NaNbO$_3$ & 17.9 & 5.3 & $-$9.2 & --- & --- & 153.0 & --- & --- & --- & 60.1 \\
\end{tabular}
\end{ruledtabular}
\end{table*}
In superlattices with a tetragonal ground-state structure, in which the
polarization is directed along the $z$~axis, five components of the piezoelectric
tensor $d_{i\nu}$ are nonzero. Among them, the highest values of $d_{i\nu}$ are
obtained for PbTiO$_3$/KTaO$_3$ and PbTiO$_3$/LaAlO$_3$ SLs (Table~\ref{table3}).
Interestingly, among these coefficients, the $d_{15}$ values turned out to be
the largest. This coefficient characterizes the polarization $P_x$ that appears
as a result of the $xz$ shear strain of the unit cell, that is, as a result of
the inclination of the polarization vector. However, no clear correlation between
$d_{15}$ and $P_z$ was observed in tetragonal SLs.
Moreover, in the related system, BaTiO$_3$/LaAlO$_3$,
the piezoelectric coefficients were unexpectedly low (Table~\ref{table3}). This
means that the inclination of the polarization vector is not an effective way
for obtaining high piezoelectric coefficients.
In superlattices with a monoclinic ground-state structure, in which the
polarization vector lies in the $xz$~plane, the piezoelectric tensor is characterized
by ten nonzero components. In these SLs, the highest values of $d_{i\nu}$
are the $d_{24}$ and $d_{26}$ coefficients, which describe the appearance of
polarization in the $y$~direction normal to the $xz$~plane under the $yz$ and
$xy$ shear strain. An analysis of the obtained data also does not find a clear
correlation between the piezoelectric coefficients and the average polarization.
Stretching the unit cells of these superlattices in the $xz$~plane changes the
polarization, but the corresponding piezoelectric coefficients ($d_{11}$ and
$d_{33}$ in Table~\ref{table3}) are not the largest.
\begin{table*}
\caption{\label{table4}The values of the $\partial u_{1}^i / \partial \sigma_{15}$
component of the $\partial u_{\alpha}^i / \partial \sigma_{\mu\nu}$ tensor (in~{\AA})
for all atoms in the ground-state structure of PbTiO$_3$/LaAlO$_3$ and BaTiO$_3$/LaAlO$_3$
superlattices ($P4bm$~phase) and in ordinary superlattices, PbTiO$_3$/PbZrO$_3$
($Pc$~phase) and KNbO$_3$/NaNbO$_3$ ($Pmc2_1$~phase), of systems known for their
piezoelectric properties. Atoms 1--10 are numbered according to Fig.~\ref{fig1};
atoms 11--20 are located in the second part of the doubled unit cell of the
high-temperature phase and are shifted from atoms 1--10 along the $x$~axis.
For atomic positions of all atoms see tables in the Supplementary data.}
\begin{ruledtabular}
\begin{tabular}{ccccc}
Atom~$i$ & PbTiO$_3$/LaAlO$_3$ & BaTiO$_3$/LaAlO$_3$ & PbTiO$_3$/PbZrO$_3$ & KNbO$_3$/NaNbO$_3$ \\
\hline
1 & +5.18 & +0.09 & +0.35 & +0.00 \\
2 & +2.37 & $-$0.08 & +0.03 & $-$0.58 \\
3 & $-$1.19 & +0.24 & +0.14 & $-$0.31 \\
4 & $-$5.61 & +0.10 & +0.17 & $-$0.51 \\
5 & $-$1.02 & $-$0.11 & +0.49 & +0.00 \\
6 & +4.56 & +0.73 & +0.44 & +0.00 \\
7 & +0.76 & +0.02 & $-$0.23 & +0.58 \\
8 & $-$0.21 & $-$0.31 & $-$0.31 & +0.31 \\
9 & $-$0.22 & $-$0.15 & $-$0.83 & +0.51 \\
10 & $-$0.84 & $-$0.10 & $-$0.42 & +0.00 \\
11 & +1.71 & $-$0.38 & +0.01 & +0.00 \\
12 & +2.37 & $-$0.08 & +0.08 & $-$0.41 \\
13 & $-$1.19 & +0.24 & $-$0.11 & $-$0.51 \\
14 & $-$5.61 & +0.10 & $-$0.43 & $-$1.15 \\
15 & $-$1.02 & $-$0.11 & +0.47 & +0.00 \\
16 & +0.48 & +0.37 & +0.76 & +0.00 \\
17 & +0.76 & +0.02 & $-$0.19 & +0.41 \\
18 & $-$0.21 & $-$0.31 & $-$0.06 & +0.51 \\
19 & $-$0.22 & $-$0.15 & $-$0.26 & +1.15 \\
20 & $-$0.84 & $-$0.10 & $-$0.12 & +0.00 \\
\end{tabular}
\end{ruledtabular}
\end{table*}
To understand the mechanism of the appearance of high piezoelectric coefficients
in some superlattices with the polar discontinuity, we analyzed the third-rank
tensors $\partial u_{\alpha}^i / \partial \sigma_{\mu\nu}$. This tensor
characterizes the displacement of the $i$th atom in the unit cell in the
$\alpha$~direction produced by strain $\sigma_{\mu\nu}$. It turned out that
in SLs exhibiting strong piezoelectricity, the values of some components
of these tensors for some atoms reach 10--15~{\AA} (that is, the unit cell strain
of 1\% generates atomic displacements exceeding 0.1~{\AA}). For example, in the
PbTiO$_3$/LaAlO$_3$ superlattice, such atoms are Pb(1), La(6), and two oxygen atoms
O(4) and O(14) located in the TiO$_2$ layer (Table~\ref{table4}); the contributions
of O(8) and O(9) oxygen atoms located in the AlO$_2$ layer are 25~times smaller.
The fact that strong distortions in the PbTiO$_3$/LaAlO$_3$ SL are observed in
the TiO$_2$ layer suggests that the Ti(2) atom also actively participates
in the appearance of polarization. Indeed, the displacement pattern
of the Pb, Ti, and O atoms in the PbO and TiO$_2$ layers resembles that of the
polar mode in PbTiO$_3$. A more detailed analysis of atomic displacements in
monoclinically strained SL, whose deformation is described by a nonzero component
$\sigma_{xz} = 0.005$, simultaneously finds two types of distortions: the rotations
of corner-linked TiO$_6$ and AlO$_6$ octahedra around the $x$~axis, resembling
rotations described by an unstable phonon at the $R$~point of the cubic perovskite
structure, and out-of-phase
(polar) displacements of Pb(1), La(6), Ti(2), O(4), and O(14) atoms along the
$x$~axis, like in the polar mode in PbTiO$_3$. The observed displacement pattern
is very close (with the correlation coefficient of 0.902) to the eigenvector
of the soft $E$~mode with a surprisingly low frequency of 27~cm$^{-1}$ in the
phonon spectrum of the $P4bm$~phase of this SL. Under the influence of strain,
the irreducible representation~$E$ reduces to a sum of $A' + A''$ irreducible
representations of the low-symmetry $Cm$~phase, and the structure distorts
according to the eigenvector of the full-symmetry $A'$~mode. In other SLs with
lower piezoelectric properties, the frequencies of the corresponding modes were
higher (40--120~cm$^{-1}$), but we didn't find simple correlations between
the frequencies and effective charges of these modes and the piezoelectric
properties. This question needs more detailed investigation.
\begin{table*}
\caption{\label{table5}Nonzero components of the piezoelectric tensor $d_{i\nu}$
(in~pC/N) in the ground and metastable states of two short-period superlattices
with the polar discontinuity.}
\begin{ruledtabular}
\begin{tabular}{cccccccccccc}
Phase & $d_{11}$ & $d_{12}$ & $d_{13}$ & $d_{15}$ & $d_{24}$ & $d_{26}$ & $d_{31}$ & $d_{32}$ & $d_{33}$ & $d_{35}$ \\
\hline
\multicolumn{11}{c}{KNbO$_3$/SrTiO$_3$ superlattice} \\
$Pc$ & 35.9 & 15.3 & $-$19.1 & 8.3 & 41.7 & 151.0 & $-$22.0 & $-$12.1 & 16.1 & 9.4 \\
$Cc$ & 35.0 & 14.8 & $-$18.3 & 7.7 & 41.4 & 154.0 & $-$21.4 & $-$11.6 & 15.7 & 11.1 \\
\hline
\multicolumn{11}{c}{SrTiO$_3$/KTaO$_3$ superlattice} \\
$Pc$ & 39.7 & 22.8 & $-$27.6 & 28.2 & 46.9 & 41.2 & $-$18.4 & $-$13.0 & 26.1 & $-$4.2 \\
$Cc$ & 39.2 & 22.3 & $-$27.4 & 27.1 & 45.7 & 41.1 & $-$18.6 & $-$13.1 & 26.3 & $-$3.7 \\
\end{tabular}
\end{ruledtabular}
\end{table*}
A comparison of the piezoelectric properties of the metastable $Cc$~phase and
the ground-state $Pc$~phase in KNbO$_3$/SrTiO$_3$ and SrTiO$_3$/KTaO$_3$
superlattices, in which the energy difference between the two structures
is very small, shows that their piezoelectric tensors are quite close to each
other (Table~\ref{table5}). This means that a change in the character of the
octahedra rotations weakly affects the piezoelectric properties of SLs.
According to our calculations, the piezoelectric coefficients in studied
superlattices can reach 150--270~pC/N. A comparison of the obtained results with
the published data in some cases finds their good agreement. For example, for
the $P4mm$~phase of PbTiO$_3$/LaAlO$_3$ SL our result $e_{33} = -3.52$~C/m$^2$
is quite close to the value $e_ {33} = -2.85$~C/m$^2$ obtained in \cite{JApplPhys.109.066107}.
However, our value of $d_{33} = -18.9$~pC/N for the same phase of the same SL
differs from the value $d_{33} = 13.9$~pC/N obtained in \cite{ChinPhysB.27.027701}:
these values are close in magnitude, but have a different sign. An additional
reason for the stronger discrepancies here may be a neglect of the difference in
the displacements and effective charges of the O atoms in Ref.~\cite{ChinPhysB.27.027701}
(according to our data, the effective charges varies from $-$2.07 to $-$5.84).
As for the PbTiO$_3$/KTaO$_3$ SL studied in \cite{ChinPhysLett.33.026302}, the
$e_{33}$ and $e_{15}$ piezoelectric coefficients obtained there are almost
100~times less than our data, and the $e_{31}$~value is close to our result in
magnitude, but differs in sign.
In conclusion, it was interesting to compare the piezoelectric properties of
superlattices with the polar discontinuity with those of ``isoelectronic''
superlattices: (1)~SLs without the polar discontinuity, in which the broken
symmetry and high-symmetry $P4mm$~structure are formed, by analogy with
\cite{PhysRevLett.84.5636}, as a result of the absence of the mirror symmetry
in the sequence of layers (BaTiO$_3$/SrZrO$_3$, BaTiO$_3$/SrSnO$_3$,
KNbO$_3$/NaTaO$_3 $, and BiScO$_3$/LaAlO$_3$), and (2)~ordinary SLs with
one substituted
atom (BaTiO$_3$/SrTiO$_3$, KNbO$_3$/KTaO$_3$, and two famous systems known for
their high piezoelectric properties, PbTiO$_3$/PbZrO$_3$ and ``lead-free''
KNbO$_3$/NaNbO$_3$). The piezoelectric properties of these superlattices are
given in the last eight lines of Table~\ref{table3}. It is seen that in
spite of a large scatter of the data, superlattices with the polar discontinuity
are not much different in their piezoelectric properties from other
superlattices (high piezoelectric properties of
BaTiO$_3$/SrTiO$_3$ and KNbO$_3$/KTaO$_3$ SLs, as follows from the corresponding
values of $\Delta E$ (Table~\ref{table2}), can be realized only at low
temperatures). Thus, the only advantage of superlattices with the polar
discontinuity as well as of SLs with broken symmetry is that one component of
polarization in them is not switchable.
\section{Conclusions}
First-principles calculations have been used to study the stability of a
high-symmetry $P4mm$~polar phase in seventeen ferroelectric perovskite
superlattices with the polar discontinuity. In most superlattices, this phase
exhibits either the ferroelectric instability or the antiferrodistortive one,
or both of them simultaneously. For each superlattice, the ground-state
structure, the structure of possible metastable phases, spontaneous polarization,
and piezoelectric properties were calculated. A~comparison of the piezoelectric
properties of superlattices with the polar discontinuity and those without
the polar discontinuity (superlattices with broken symmetry and ordinary
superlattices) showed that these properties are not much different. It was
demonstrated that high piezoelectric coefficients (up to 150--270~pC/N) in some
superlattices with the polar discontinuity are due to the appearance of strong
lattice distortions, whose symmetry follows that of a low-lying polar phonon
mode of the ground-state structure under the influence of external strain.
\bigskip
\textbf{Data availability statement}
\bigskip
All data included in this study are available upon request from the author.
\bigskip
\textbf{Declaration of competing interests}
\bigskip
The author declares that he has no known competing financial interests or
personal relationships that could have appeared to influence the work reported
in this paper.
\mbox{ }
\begin{acknowledgments}
This work was partially supported by the Russian Foundation for Basic Research
under Grant 17-02-01068.
\end{acknowledgments}
| cd8a3170ad7dfb6fa50a9bf8e00a0a5f666c805e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Converting heat into useful work, heat engine lies at the core of
thermodynamics, both in classical and quantum regime \citep{Carnoteff,tolman1948irreversible,kosloff2014quantum,Binder2018}.
Absorbing heat from a hot thermal bath with the temperature $T_{\mathrm{h}}$,
the engine outputs work and releases part of the heat to the cold
bath with the temperature $T_{\mathrm{c}}$. The upper limit of the
heat engine working between two heat baths is given by the Carnot
efficiency $\eta_{\mathrm{C}}=1-T_{\mathrm{c}}/T_{\mathrm{h}}$ \citep{Carnoteff}.
Due to the limitation of the quasi-static cycle with infinite-long
operation time, the heat engine with Carnot efficiency generally has
vanishing output power and in turn is of no practical use. To design
the heat engine cycles operating in finite time, several practical
heat engine models have been proposed \citep{andresen1984thermodynamics,wu1999recent,ZCTuCPB},
such as the endo-reversible model \citep{reitlinger1929utilisation,yvon1955saclay,chambadal1957recuperation,novikov1958efficiency,CA},
the linear irreversible model \citep{BroeckPRL2005,Wang2012,izumida2014work},
the stochastic model \citep{schmiedl2008efficiency,StochasticThermodynamics},
and the low-dissipation model \citep{EspositoPRL2010,2013EfficiencyBroeck,2015Efficiency,Constraintrelationyhma,2020LDmodel,2020Optimal}.
The efficiency at maximum power (EMP), is proposed as an important
parameter to evaluate the performance of these heat engines in the
finite-time cycles.
The utilization of the low-dissipation model\citep{EspositoPRL2010,2013EfficiencyBroeck,2015Efficiency,Constraintrelationyhma,yhmaoptimalcontrol}
simplifies the optimization of the finite-time Carnot-like heat engines.
As the model assumption, the heat transfer between the engine and
the bath in the finite-time quasi-isothermal process is divided into
two parts as follow
\begin{equation}
Q_{\mathrm{h,c}}(\tau_{\mathrm{h}})=T_{\mathrm{h,c}}(\Delta S_{\mathrm{h,c}}-S_{\mathrm{h,c}}^{(\mathrm{ir})}),\label{eq:Qhc}
\end{equation}
where $\Delta S_{\mathrm{h}}=-\Delta S_{\mathrm{c}}=\Delta S$ is
the reversible entropy change of the working substance and $S_{\mathrm{h,c}}^{(\mathrm{ir})}=\Sigma_{\mathrm{h,c}}/\tau_{\mathrm{h,c}}$
is the irreversible entropy generation which is inversely proportional
to the process time $\tau_{\mathrm{\alpha}}$. Optimizing the output
power $P(\tau_{\mathrm{h}},\tau_{\mathrm{c}})=[Q_{\mathrm{h}}(\tau_{\mathrm{h}})+Q_{\mathrm{c}}(\tau_{\mathrm{c}})]/(\tau_{\mathrm{h}}+\tau_{\mathrm{c}})$
with respect to the operation time $\tau_{\mathrm{h}}$ and $\tau_{\mathrm{c}}$,
one gets the optimal operation times \citep{EspositoPRL2010} as
\begin{align}
\tau_{\mathrm{h}}^{*} & =\frac{2T_{\mathrm{h}}\Sigma_{\mathrm{h}}}{(T_{\mathrm{h}}-T_{\mathrm{c}})\Delta S}\left(1+\sqrt{\frac{T_{\mathrm{c}}\Sigma_{\mathrm{c}}}{T_{\mathrm{h}}\Sigma_{\mathrm{h}}}}\right),\label{eq:th}\\
\tau_{\mathrm{c}}^{*} & =\frac{2T_{\mathrm{c}}\Sigma_{\mathrm{c}}}{(T_{\mathrm{h}}-T_{\mathrm{c}})\Delta S}\left(1+\sqrt{\frac{T_{\mathrm{h}}\Sigma_{\mathrm{h}}}{T_{\mathrm{c}}\Sigma_{\mathrm{c}}}}\right),\label{eq:tc}
\end{align}
and the efficiency at the maximum power $\eta^{*}$ bounded by the
following inequality as \citep{EspositoPRL2010,ZCTuCPB}
\begin{equation}
\eta_{\mathrm{-}}\equiv\frac{\eta_{\mathrm{C}}}{2}\leq\eta^{*}\leq\frac{\eta_{\mathrm{C}}}{2-\eta_{\mathrm{C}}}\equiv\eta_{\mathrm{+}}.\label{eq:inequality}
\end{equation}
Due to the simplicity of the model assumption and the universality
of the obtained EMP, the low-dissipation model becomes one of the
most studied finite-time heat engine models in recent years \citep{2013EfficiencyBroeck,2015Efficiency,Constraintrelationyhma,2020LDmodel,2020Optimal,2020-finite-size}.
It is currently cleared that \citep{TradeoffrelationShiraishi,CavinaPRLtradeoffrelation,Constraintrelationyhma,yhmaoptimalcontrol,2020IEGyhma}
the low-dissipation assumption is valid in the long-time regime of
$\tau_{\mathrm{h(c)}}/t_{\mathrm{r}}\gg1$, where $t_{\mathrm{r}}$
is the relaxation time for the work substance to reach its equilibrium
with the heat bath. And the dissipation coefficient $\Sigma_{\mathrm{h(c)}}$
of the $1/\tau$-scaling is determined by both the coupling strength
$\gamma_{\mathrm{h(c)}}\sim1/t_{\mathrm{r}}$ to the bath\citep{Constraintrelationyhma,2020IEGyhma}
and the control scheme \citep{yhmaoptimalcontrol,2020IEGyhma}. Such
a relation implies that the condition $\tau_{\mathrm{h(c)}}^{*}/t_{\mathrm{r}}\gg1$
is not fulfilled simply and should be justified to reveal the regime
of validity. We check the consistency of the obtained EMP with a minimal
heat engine model consisting of a single two-level system. In Sec.
\ref{sec:Self-consistency-of-low-dissipat}, we analytically obtain
the regime, where the optimal operation time to achieve EMP is consistent
with the low-dissipation assumption. And we further show the possibility
of the exact EMP of the engine to surpass the upper bound of EMP,
i.e., $\eta_{\mathrm{+}}$, obtained with the low-dissipation model
in the large-$\eta_{\mathrm{C}}$ regime in Sec. \ref{sec:Efficiency-at-maximum}.
\section{\label{sec:Self-consistency-of-low-dissipat}Self-consistency of
the low-dissipation model in deriving efficiency at maximum power}
The two-level atomic heat engine is the simplest quantum engine to
demonstrate the relevant physical mechanisms \citep{Geva1992,QTquanQH,Su2018,Constraintrelationyhma,yhmaoptimalcontrol}.
The energy spacing of the excited state $\left|e\right\rangle $ and
ground state $\left|g\right\rangle $ is tuned by an outside agent
to extract work with the Hamiltonian $H=\frac{1}{2}\hbar\omega\left(t\right)\sigma_{z},$
where $\sigma_{z}=\left|e\right\rangle \left\langle e\right|-\left|g\right\rangle \left\langle g\right|$
is the Pauli matrix in the z-direction. The Planck's constant is taken
as $\hbar=1$ in the following discussion for convenience. For the
finite-time quasi-isothermal process with the duration $\tau$ of
the two-level system, the low-dissipation assumption of the $1/\tau$
scaling is valid in the regime $\widetilde{\gamma}\tau\gg1$ \citep{Constraintrelationyhma},
where $\widetilde{\gamma}=2\gamma T/\omega_{0}$. Here $\gamma$ is
the coupling strength between the system and the bath with the temperature
$T$ and $\omega_{0}$ is the initial energy spacing of the system
during the process.
\begin{figure}
\centering{}\includegraphics[width=8.5cm]{cycleillustrationV3}\caption{\label{fig:Schematic-diagram-of}Schematic diagram of the finite-time
Carnot-like cycle for a two-level atomic heat engine. The horizontal
axis and the vertical axis represent respectively the energy spacing
$\omega$ and excited state population $p_{\mathrm{e}}$ of the two-level
atom. The red (blue) solid curve represents the high (low) temperature
finite-time quasi-isothermal process, and the black solid lines represent
the adiabatic processes.}
\end{figure}
The finite-time Carnot-like cycle for the two-level atomic heat engine
of interest consists of four strokes, two isothermal and two adiabatic
processes. The schematic diagram of the cycle is shown in Fig. \ref{fig:Schematic-diagram-of}.
In the figure, $\omega_{\mathrm{h}}^{\mathrm{i}}$ and $\omega_{\mathrm{h}}^{\mathrm{f}}$
($\omega_{\mathrm{c}}^{\mathrm{i}}$ and $\omega_{\mathrm{c}}^{\mathrm{f}}$
) are respectively the initial and finial energy spacing of the working
substance in the high (low) temperature finite-time quasi-isothermal
process with duration $\tau_{\mathrm{h}}$ ($\tau_{\mathrm{c}}$),
which is shown with the red (blue) solid curve. The total operating
time per cycle is $t=\tau_{\mathrm{h}}+\tau_{\mathrm{c}}$. The interval
of the adiabatic processes (the black solid lines) are ignored in
comparison with $\tau_{\mathrm{h}}$ and $\tau_{\mathrm{c}}$ \citep{EspositoPRL2010,Constraintrelationyhma}.
We assume the two-level system has no energy level crossing during
the whole cycle to ensure no coherence of the system is induced by
non-adiabatic transition\citep{2012Quantum-non-adiabatic,2018Time-non-adiabatic}.
The quasi-isothermal process retains the normal isothermal process
at the quasi-static limit of $\tau_{\mathrm{h(c)}}\rightarrow\infty$.
For simplicity, we focus on the high-temperature regime, where the
reversible entropy change $\Delta S_{\alpha}$ and the irreversible
entropy generation coefficient $\Sigma_{\alpha}$ in Eq. (\ref{eq:Qhc})
are analytically written as \citep{Constraintrelationyhma}
\begin{equation}
\Delta S=\frac{\left[\left(\omega_{\mathrm{\alpha}}^{\mathrm{i}}\right)^{2}-\left(\omega_{\mathrm{\alpha}}^{\mathrm{f}}\right)^{2}\right]}{8T_{\mathrm{\alpha}}^{2}},\Sigma_{\alpha}=\frac{2\Delta S}{\widetilde{\gamma}_{\alpha}},\label{eq:deltaS-sigma}
\end{equation}
with $\alpha=\mathrm{h,c}$, and $\widetilde{\gamma}_{\alpha}=2\gamma_{\alpha}T_{\alpha}/\omega_{\mathrm{\alpha}}^{\mathrm{i}}$.
Here and after, the Boltzmann's constant $k_{\mathrm{B}}=1$ is chosen.
To obtain the above equations, the relations $\omega_{\mathrm{h}}^{\mathrm{i}}/T_{\mathrm{h}}=\omega_{\mathrm{c}}^{\mathrm{f}}/T_{\mathrm{c}}$,
$\omega_{\mathrm{h}}^{\mathrm{f}}/T_{\mathrm{h}}=\omega_{\mathrm{c}}^{\mathrm{i}}/T_{\mathrm{c}}$
have been used in the quantum adiabatic processes\citep{QTquanQH,Constraintrelationyhma}.
Substituting Eq. (\ref{eq:deltaS-sigma}) into Eqs. (\ref{eq:th})
and (\ref{eq:tc}), we obtain the corresponding optimal operation
time $\tau_{\mathrm{\alpha}}^{*}$ for achieving the maximum power
with the dimensionless time $\widetilde{\tau}_{\alpha}^{*}\equiv\tau_{\mathrm{\alpha}}^{*}\widetilde{\gamma}_{\alpha}$
as
\begin{equation}
\widetilde{\tau}_{\mathrm{h}}^{*}=\frac{2}{\eta_{\mathrm{C}}}\frac{\omega_{\mathrm{h}}^{\mathrm{i}}-\omega_{\mathrm{h}}^{\mathrm{f}}}{\omega_{\mathrm{h}}^{\mathrm{i}}+\omega_{\mathrm{h}}^{\mathrm{f}}}\left[1+\sqrt{\left(1-\eta_{\mathrm{C}}\right)\frac{\gamma_{\mathrm{h}}}{\gamma_{\mathrm{c}}}}\right],\label{eq:th-1}
\end{equation}
\begin{equation}
\widetilde{\tau}_{\mathrm{c}}^{*}=\frac{2}{\eta_{\mathrm{C}}}\frac{\omega_{\mathrm{h}}^{\mathrm{i}}-\omega_{\mathrm{h}}^{\mathrm{f}}}{\omega_{\mathrm{h}}^{\mathrm{i}}+\omega_{\mathrm{h}}^{\mathrm{f}}}\left[\sqrt{\left(1-\eta_{\mathrm{C}}\right)\frac{\gamma_{\mathrm{c}}}{\gamma_{\mathrm{h}}}}+1-\eta_{\mathrm{C}}\right].\label{eq:tc-1}
\end{equation}
The low-dissipation assumption is valid in the regime $\widetilde{\tau}_{\mathrm{h}}^{*}\gg1$
and $\widetilde{\tau}_{\mathrm{c}}^{*}\gg1$, namely,
\begin{equation}
1+\sqrt{\left(1-\eta_{\mathrm{C}}\right)\frac{\gamma_{\mathrm{h}}}{\gamma_{\mathrm{c}}}}\gg\frac{\eta_{\mathrm{C}}}{2}\frac{\omega_{\mathrm{h}}^{\mathrm{i}}+\omega_{\mathrm{h}}^{\mathrm{f}}}{\omega_{\mathrm{h}}^{\mathrm{i}}-\omega_{\mathrm{h}}^{\mathrm{f}}}\label{eq:condition1}
\end{equation}
\begin{equation}
\sqrt{\left(1-\eta_{\mathrm{C}}\right)\frac{\gamma_{\mathrm{c}}}{\gamma_{\mathrm{h}}}}+1-\eta_{\mathrm{C}}\gg\frac{\eta_{\mathrm{C}}}{2}\frac{\omega_{\mathrm{h}}^{\mathrm{i}}+\omega_{\mathrm{h}}^{\mathrm{f}}}{\omega_{\mathrm{h}}^{\mathrm{i}}-\omega_{\mathrm{h}}^{\mathrm{f}}}\label{eq:condition2}
\end{equation}
The above two inequalities are fulfilled when
\begin{equation}
\eta_{\mathrm{C}}\ll2\frac{1-\delta}{1+\delta},\label{eq:regime-1}
\end{equation}
where $\delta\equiv\omega_{\mathrm{h}}^{\mathrm{f}}/\omega_{\mathrm{h}}^{\mathrm{i}}$
is the compression ratio of the heat engine cycle in the quasi-isothermal
process. The above relation is one of the main results of the current
work and reveals the range of $\eta_{\mathrm{C}}$ in which the low-dissipation
model is applicable for finding EMP. The bound for EMP obtained in
the low-dissipation regime, as given by Eq. (\ref{eq:inequality}),
thus may be not unconditionally applicable to such two-level atomic
engine. Indeed, we will show the EMP out of the regime is larger than
the upper bound $\eta_{+}$ predicted by the low-dissipation model
in the next section.
\section{\label{sec:Efficiency-at-maximum}Efficiency at maximum power: beyond
the low dissipation model}
With the analytical discussion above, we find the EMP obtain with
the low-dissipation model is only consistent with the assumption of
the low-dissipation model in the low-$\eta_{\mathrm{C}}$ regime for
the two-level system. The question is whether the bound provided by
the low-dissipation model, i.e. $\eta_{+}$, is still the upper bound
for the achievable efficiency of the system out of the low-$\eta_{\mathrm{C}}$
regime. Unfortunately, the answer is no. In this section, we will
focus the efficiency at the maximum power in the regime with large
$\eta_{\mathrm{C}}$.
By numerically simulating the dynamics of the two-level system engine
with different cycle time, we obtain the exact power and efficiency
to find the EMP. The results in the large-$\eta_{\mathrm{C}}$ regime
show that: (i) the optimal operation time corresponding to the maximum
power of the heat engine does not meet the low-dissipation assumption;
(ii) the EMP surpass the upper bound obtained with the low-dissipation
model, namely, $\eta_{\mathrm{MP}}>\eta_{+}$.
The dynamics of the two-level atom in the finite-time quasi-isothermal
process is given by the master equation as follow \citep{Constraintrelationyhma}
\begin{equation}
\frac{dp_{\mathrm{e}}(t)}{dt}=-\kappa(t)p_{\mathrm{e}}(t)+C(t),\label{eq:master equation-1}
\end{equation}
where $p_{\mathrm{e}}(t)$ is the excited state population and $C(t)=\gamma n[\omega(t)]$.
$\kappa(t)=\gamma\left(2n[\omega(t)]+1\right)$ is the effective dissipation
rate with the mean occupation number $n[\omega(t)]=1/\left(\exp[\beta\omega(t)]-1\right)$
for the bath mode $\omega(t)$. The dissipation rate $\gamma$ equals
to $\gamma_{\mathrm{h}}$ ($\gamma_{\mathrm{c}}$) in the high (low)
temperature quasi-isothermal process with the inverse temperature
$\beta=1/(k_{\mathrm{B}}T_{\mathrm{h}})$ ($\beta=1/(k_{\mathrm{B}}T_{\mathrm{c}})$).
The energy spacing of the two-level atom is tuned linearly as $\omega(t)=\omega_{\mathrm{h}}^{\mathrm{i}}+(\omega_{\mathrm{h}}^{\mathrm{f}}-\omega_{\mathrm{h}}^{\mathrm{i}})t/\tau_{\mathrm{h}},\:t\in[0,\tau_{\mathrm{h}}]$
in the high-temperature finite-time quasi-isothermal process and as
$\omega(t)=\omega_{\mathrm{c}}^{\mathrm{i}}+(\omega_{\mathrm{c}}^{\mathrm{f}}-\omega_{\mathrm{c}}^{\mathrm{i}})t/\tau_{\mathrm{c}},\:t\in[\tau_{\mathrm{c}},\tau_{\mathrm{c}}+\tau_{\mathrm{h}}]$
in the low-temperature finite-time quasi-isothermal process. The population
of the two-level system keeps unchanged during the adiabatic processes
whose operation time is ignored in comparison with $\tau_{\mathrm{h}}$
and $\tau_{\mathrm{c}}$.
In the following simulation, we set $\gamma_{\mathrm{h}}=1$ and focus
on the regime of $\gamma_{\mathrm{c}}/\gamma_{\mathrm{h}}\rightarrow\infty$,
i.e., $\Sigma_{\mathrm{c}}/\Sigma_{\mathrm{h}}\rightarrow0$, where
the upper bound $\eta_{+}=\eta_{\mathrm{C}}/(2-\eta_{\mathrm{C}})$
of EMP of the engine is achieved according to the prediction with
the low-dissipation model\citep{EspositoPRL2010}. In this regime,
the low-temperature quasi-isothermal process approaches the isothermal
process fastly enough that the operation time $\tau_{\mathrm{c}}$
is further ignored for the optimization of the cycle's output power.
The optimization is simplified as a single parameter optimization
problem: find the maximum value $P_{\mathrm{max}}$ of the cycle's
output power with respect to $\tau_{\mathrm{h}}$, and obtain the
EMP of the engine, $\eta_{\mathrm{MP}}\equiv\eta(P=P_{\mathrm{max}})$.
\begin{figure}
\begin{centering}
\includegraphics[width=8.5cm]{cycle-th=2-10-200tr-TH=10-gamma=1-wh0=1-whf=0\lyxdot 9-etac=0\lyxdot 1}
\par\end{centering}
\centering{}\caption{\label{fig:The-finite-time-Carnot-like}The finite-time Carnot-like
cycles of for two-level atomic heat engine with different operation
time $\tau_{\mathrm{h}}$. The red curves represent the high-temperature
finite-time quasi-isothermal process with the duration $\tau_{\mathrm{h}}$,
while the blue curves represent the low-temperature isothermal process
. The adiabatic processes are plotted with the black lines. The outermost
dash-dotted curves relate to the quasi-static cycle with $\tau_{\mathrm{h}}=200t_{\mathrm{r}}$,
while the middle dashed cycle and inner solid cycle are obtained with
$\tau_{\mathrm{h}}=10t_{\mathrm{r}}$ and $\tau_{\mathrm{h}}=2t_{\mathrm{r}}$,
respectively. In this example, we choose $\omega_{\mathrm{h}}^{\mathrm{i}}=1$,
$\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$, $\gamma_{\mathrm{h}}=1$,
$T_{\mathrm{h}}=10$, and $T_{\mathrm{c}}=9$. $t_{\mathrm{r}}=\omega_{\mathrm{h}}^{\mathrm{i}}/\left(2\gamma_{\mathrm{h}}T_{\mathrm{h}}\right)=0.05$
is the relaxation time related to the high-temperature finite-time
quasi-isothermal process.}
\end{figure}
The cycles with different $\tau_{\mathrm{h}}$ are illustrated in
Fig. \ref{fig:The-finite-time-Carnot-like}, where $\omega_{\mathrm{h}}^{\mathrm{i}}=1$
and $\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$ are fixed. The temperatures
for the hot and cold bath are chosen as $T_{\mathrm{h}}=10$ and $T_{\mathrm{c}}=9$
as an example. The relaxation time is $t_{\mathrm{r}}=\omega_{\mathrm{h}}^{\mathrm{i}}/\left(2\gamma_{\mathrm{h}}T_{\mathrm{h}}\right)=0.05$.
The quasi-static cycles with $\tau_{\mathrm{h}}=200t_{\mathrm{r}}$,
$10t_{\mathrm{r}}$ and $2t_{\mathrm{r}}$ are represented by the
dash-dotted line, dashed line, and solid line, respectively. The figure
shows that the output work represented by the cycle area decreases
with $\tau_{\mathrm{h}}$.
\begin{figure}
\centering{}\includegraphics[width=8.5cm]{P-t2}\caption{\label{fig:The-normalized-power}The normalized power of the engine
$\widetilde{P}=P/P_{\mathrm{max}}$ as the function of $\tau_{\mathrm{h}}/t_{\mathrm{r}}$.
The blue solid line, the orange dash-dotted line, and the purple dashed
line are respectively obtained with $\eta_{\mathrm{C}}=0.1$, $\eta_{\mathrm{C}}=0.12$,
and $\eta_{\mathrm{C}}=0.15$. In this example, we choose $\omega_{\mathrm{h}}^{i}=1$,
$\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$, $\gamma_{\mathrm{h}}=1$,
and $T_{\mathrm{h}}=10$ with changing $T_{\mathrm{c}}=9$, $8.8$
and $8.5$. The relaxation time is $t_{\mathrm{r}}=\omega_{\mathrm{h}}^{\mathrm{i}}/\left(2\gamma_{\mathrm{h}}T_{\mathrm{h}}\right)=0.05$.}
\end{figure}
In Fig. \ref{fig:The-normalized-power}, we show the normalized power
of the engine $\widetilde{P}\equiv P/P_{\mathrm{max}}$ as the function
of $\tau_{\mathrm{h}}/t_{\mathrm{r}}$ with $\eta_{\mathrm{C}}=0.1$
(blue solid line), $\eta_{\mathrm{C}}=0.12$ (orange dash-dotted line),
and $\eta_{\mathrm{C}}=0.15$ (purple dashed line). In the simulation,
the parameters are set as $\omega_{\mathrm{h}}^{i}=1$, $\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$,
and $T_{\mathrm{h}}=10$ with changing $T_{\mathrm{c}}=9$, $8.8$
and $8.5$. The relaxation time is $t_{\mathrm{r}}=0.05$. The maximum
output power $P_{\mathrm{max}}$ is obtained numerically for different
$\eta_{\mathrm{C}}$. It is observed from the figure that the dependence
of $\widetilde{P}$ on operation time $\tau_{\mathrm{h}}$ changes
with $\eta_{\mathrm{C}}$. In the figure, the optimal $\tau_{\mathrm{h}}^{*}$
decreases with $\eta_{\mathrm{C}}$ and is away from the low-dissipation
regime of $\tau_{\mathrm{h}}/t_{\mathrm{r}}\gg1$, illustrated with
the orange dash-dotted line ($\eta_{\mathrm{C}}=0.12$, $\tau_{\mathrm{h}}^{*}/t_{\mathrm{r}}\approx0.5$)
and the blue solid line ($\eta_{\mathrm{C}}=0.1$, $\tau_{\mathrm{h}}^{*}/t_{\mathrm{r}}\approx1$).
As shown clearly by the purple dashed line with $\eta_{\mathrm{C}}=0.15$,
the maximum power $\widetilde{P}=1$ is achieved in the short-time
regime of $\tau_{\mathrm{h}}/t_{\mathrm{r}}\ll1$, where the $1/\tau$-scaling
of irreversible entropy generation is invalid \citep{Constraintrelationyhma,2020IEGyhma}.
\begin{figure}
\begin{raggedright}
(a)
\par\end{raggedright}
\begin{centering}
\includegraphics[width=8.5cm]{wh0=1-whf=0\lyxdot 6-Th=10-gamma=1-V2}
\par\end{centering}
\begin{raggedright}
(b)
\par\end{raggedright}
\begin{centering}
\includegraphics[width=8.5cm]{wh0=1-whf=0\lyxdot 9-Th=10-gamma=1-V2}
\par\end{centering}
\begin{raggedright}
(c)
\par\end{raggedright}
\begin{centering}
\includegraphics[width=8.5cm]{t-wh0=1-whf=0\lyxdot 9-0\lyxdot 6-Th=10-gamma=1}
\par\end{centering}
\raggedright{}\caption{\label{fig:Efficiency-at-maximum}Efficiency at the maximum power
$\eta_{\mathrm{MP}}$ (orange solid line) of the heat engine as the
function of the Carnot efficiency $\eta_{\mathrm{C}}$ for different
final energy spacing of the two level system(a) $\omega_{\mathrm{h}}^{\mathrm{f}}=0.6$
and (b) $\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$. The black dashed
line (black dash-dotted line) represents the upper bound $\eta_{\mathrm{+}}$
(lower bound $\eta_{\mathrm{-}}$) of EMP obtained with the low-dissipation
model {[}Eq. (\ref{eq:inequality}){]}, and the Carnot efficiency
$\eta_{\mathrm{C}}$ is plotted with the black dotted line. The gray
area represents the low-dissipation regime predicted by Eq. (\ref{eq:regime-1}).
(c) Optimal operation time $\tau_{\mathrm{h}}^{*}$ at the maximum
power as the function of $\eta_{\mathrm{C}}$. The blue solid curve
is obtained with $\omega_{\mathrm{h}}^{\mathrm{f}}=0.6$ ($\delta=0.6$)
while the red dash-dotted curve is obtained with $\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$
($\delta=0.9$). The other parameters in this figure are chosen as
$\omega_{\mathrm{h}}^{\mathrm{i}}=1$, $\gamma_{\mathrm{h}}=1$, and
$T_{\mathrm{h}}=10$. The relaxation time is $t_{\mathrm{r}}=\omega_{\mathrm{h}}^{\mathrm{i}}/\left(2\gamma_{\mathrm{h}}T_{\mathrm{h}}\right)=0.05$.}
\end{figure}
We show the obtained efficiency $\eta_{\mathrm{MP}}$ at the maximum
power of the engine as the function of $\eta_{\mathrm{C}}$ in Fig.
\ref{fig:Efficiency-at-maximum}(a) and (b), and plot the corresponding
optimal operation time $\tau_{\mathrm{h}}^{*}$ in Fig. \ref{fig:Efficiency-at-maximum}(c).
We choose the final energy spacing of the two level system as $\omega_{\mathrm{h}}^{\mathrm{f}}=0.6$
and $\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$ respectively for (a) and
(b), and other parameters are set as $\omega_{\mathrm{h}}^{\mathrm{i}}=1$,
$\gamma_{\mathrm{h}}=1$, $T_{\mathrm{h}}=10$. As shown in Fig. \ref{fig:Efficiency-at-maximum}(a)
and (b), the EMP of the engine $\eta_{\mathrm{MP}}$(orange solid
line) in the large-$\eta_{\mathrm{C}}$ regime surpasses the upper
bound of EMP, $\eta_{+}=\eta_{\mathrm{C}}/(2-\eta_{\mathrm{C}})$
(black dashed line) obtained with the low-dissipation model. The lower
bound of EMP, $\eta_{-}=\eta_{\mathrm{C}}/2$, obtained with the low-dissipation
model is plotted with the black dash-dotted line. The gray area represents
the consistent regime as demonstrated by Eq. (\ref{eq:regime-1}).The
figure shows that $\eta_{\mathrm{MP}}$ is bounded by $\eta_{+}$
and $\eta_{-}$ of Eq. (\ref{eq:inequality}) in the gray area with
relatively small $\eta_{\mathrm{C}}$. Additionally, by comparing
(b) and (a) of Fig. \ref{fig:Efficiency-at-maximum}, with the larger
the compression rate $\delta=\omega_{\mathrm{h}}^{\mathrm{f}}/\omega_{\mathrm{h}}^{\mathrm{i}}$
($\delta=0.9$ for (a) and $\delta=0.6$ for (b)), we illustrate the
narrower the range of $\eta_{\mathrm{C}}$ in which $\eta_{\mathrm{MP}}$
is bounded by $\eta_{+}$. With the increasing of the compression
ratio $\delta$, the valid regime of optimization of the engine with
the low-dissipation model becomes smaller. And it is consistent with
the theoretical analysis of Eq. (\ref{eq:regime-1}).
In Fig. \ref{fig:Efficiency-at-maximum}(c), the optimal operation
time $\tau_{\mathrm{h}}^{*}$ at the maximum power (blue solid curve
for $\omega_{\mathrm{h}}^{\mathrm{f}}=0.6$ and red dash-dotted curve
for $\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$) decreases monotonically
with increasing $\eta_{\mathrm{C}}$. The operation time at maximum
power $\tau_{\mathrm{h}}^{*}$ of the engine is not in the low-dissipation
regime of $\tau_{\mathrm{h}}/t_{\mathrm{r}}\gg1$ for the relatively
large $\eta_{\mathrm{C}}$. This explains why $\eta_{\mathrm{MP}}$
is no longer satisfies the bound provided by the low-dissipation model
in large-$\eta_{\mathrm{C}}$ regime, and verifies our analytical
analysis in Sec. \ref{sec:Self-consistency-of-low-dissipat}. In addition,
one can find in Fig. \ref{fig:Efficiency-at-maximum}(c) that the
red dash-dotted curve is lower than the blue solid curve. This leads
to a narrower parameter range of $\eta_{\mathrm{C}}$, in which the
optimal operation time $\tau_{\mathrm{h}}^{*}$ satisfies the low-dissipation
assumption, for the heat engine with $\omega_{\mathrm{h}}^{\mathrm{f}}=0.9$
than that with $\omega_{\mathrm{h}}^{\mathrm{f}}=0.6$. Therefore,
the phenomenon that the gray area in Fig. \ref{fig:Efficiency-at-maximum}(a)
is wider than that in Fig. \ref{fig:Efficiency-at-maximum}(b) is
explained from the perspective of the operation time.
\section{Conclusions and discussion}
In summary, we checked whether the optimal operation time for achieving
the maximum power is consistent with the requirement of the low-dissipation
model for the finite-time Carnot-like heat engines in this paper.
The low-dissipation model, widely used in the finite-time thermodynamics
to study EMP, relies on the assumption that the irreversible entropy
generation in the finite-time quasi-isothermal process of duration
$\tau$ follows the $1/\tau$ scaling in the long-time regime. The
operation time for the maximum power obtained from the model should
fulfill the requirement of the low-dissipation model assumption. Due
to the unknown coefficient of the $1/\tau$ scaling, the consistency
of the model in optimizing finite-time Carnot engines had not been
tested before.
In this paper, we proved that the optimal operation time for a two-level
finite-time Carnot engine achieving EMP satisfy the low-dissipation
assumption only in the low Carnot efficiency regime of $\eta_{\mathrm{C}}\ll1$.
This observation motivated us to check the EMP in the regime with
large $\eta_{\mathrm{C}}$. We calculated the EMP of the two-level
atomic heat engine in the full parameter space of $\eta_{\mathrm{C}}$.
It is found that, in the large-$\eta_{\mathrm{C}}$ regime, the true
EMP of the heat engine can surpass the upper bound for EMP, i.e.,
$\eta_{+}=\eta_{\mathrm{C}}/(2-\eta_{\mathrm{C}})$ obtained with
the low-dissipation model.
Our study on EMP in the large-$\eta_{\mathrm{C}}$ regime shall provide
a new insight for designing heat engines with better performance working
between two heat baths with large temperature difference. In addition
to affecting the EMP of the heat engine, the short-time effects caused
by fast driving may also influence the trade-off between power and
efficiency \citep{2015Efficiency,TradeoffrelationShiraishi,CavinaPRLtradeoffrelation,Constraintrelationyhma},
which needs further exploration. The predictions of this paper can
be tested on some experimental platforms \citep{BrownianHENatPhys2015,Rossnagel2016,shortcutHEexperimentSciAdv2018,2019Thermodynamic-shortcut-isothermal,2020IEGyhma,atomic-Endoreversible2020}
in the short-time regime.
\begin{acknowledgments}
This work is supported by the National Natural Science Foundation
of China (NSFC) (Grants No. 12088101, No. 11534002, No. 11875049, No. U1530402, and No. U1930403), and the National Basic Research Program of China (Grants No. 2016YFA0301201).
\end{acknowledgments}
\bibliographystyle{apsrev}
| e9c25569db78247f0d60ab56667b7dda251fac82 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In recent years, there are more and more big data based deep learning tasks, such as ImageNet classification, natural language processing, and face recognition. The datasets often include tens of millions of images, texts, voices or other data, and the related neural networks (NNs) also contain multiple millions of parameters. For such complex tasks, how to improve the training efficiency has attracted mounting attention. It is natural to use multiple Graphics Processing Unit (GPU) as the parallel computing power, which can feed a very large batch size of data into the network at each training iteration. Unfortunately, this large batch size training deteriorates the model performance in respect of accuracy, convergence time, and more. This generalization difficulty has been reported previously \cite{lecun2012efficient,diamos2016persistent,keskar2017on,goyal2017accurate,jastrzebski2018three}, and widely known as a ‘generalization gap’ \cite{keskar2017on}. We reimplement this problem using AlexNet on Cifar10 dataset (Figure \ref{fig_1}).
Many efforts have been devoted to understanding the ‘generalization gap’ \cite{breuel2015the,keuper2016distributed,neyshabur2017exploring,li2017batch,dai2018towards}. Keskar et al. \yrcite{keskar2017on} and Jastrzebski et al. \yrcite{jastrzebski2018finding} found that large batch size training is more likely to converge to sharp minima compared to small batch size training. In this vein, Wen et al. \yrcite{wen2018smoothout:} and Haruki et al. \yrcite{haruki2019gradient} proposed noise injecting methods to eliminate sharp minima. Jastrzebski et al. \yrcite{jastrzebski2018three} reported that higher values of the ratio of learning rate (LR) to batch size can lead to flatter minima.
Hoffer et al. \yrcite{hoffer2017train} put forward a new hypothesis, ‘random walk on a random landscape’, and experimentally demonstrated that the ‘generalization gap’ can be primarily attributed to the reduced number of parameter updates. This hypothesis suggests that longer training time helps to achieve better generalization.
LR scheduling is a class of methods that could alleviate the ‘generalization gap’. Goyal et al. \yrcite{goyal2017accurate} tested a ‘learning rate warm-up’ scheme, in which a small ‘safe’ LR was initialized and linearly scaled to a target ‘base’ LR. Using this method, they successfully compress the training time of ImageNet to less than 1 hour. MegDet, an object detector, obtained the state of the art performance using a combined policy: warmup LR and Cross-GPU batch normalization \cite{peng2017megdet:}. You et al. \yrcite{you2019large-batch} proposed an improved approach of warm-up, linear-epoch gradual-warmup (LEGW), which gains better generalization. Smith \& Topin \yrcite{smith2017super-convergence:} described a phenomenon named ‘super-convergence’ with periodic LR modulation. Li et al. \yrcite{li2019towards} demonstrated that NN models with small LR first generalize to low-noise and hard-to-fit patterns, and Jastrzebski et al. \yrcite{jastrzebski2018three} explained the ratio of LR to batch size is a key determinant of statistical gradient descent (SGD) dynamics.
Batch size scheduling is regarded as the counterpart of LR scheduling. The largest useful batch size can be predicted by the gradient noise scale \cite{mccandlish2018an}. Increasing batch size helps find flatter minima \cite{devarakonda2017adabatch:,smith2018don't}. Mikami \& Suganuma \yrcite{Mikami2018imageNet} have trained ImageNet in 224 seconds using a predetermined batch size scheduling.
The discovery of the unbalance distribution of both parameters and gradients across layers in a network is an important breakthrough for large batch size training. The generalization is markedly enhanced in case of setting a layer-wise LR, which is determined by some form of statistical values of the inlayer parameters and gradients. You et al. \yrcite{you2017large} firstly built an algorithm, ‘Layer-wise Adaptive Rate Scaling (LARS)’, to extend ImageNet training batch size to 32 K. In their sequential work, they showed that using LARS, the ImageNet can be even trained in a few minutes \cite{you2017scaling,you2018imagenet}. Two relatives of LARS, ‘PercentDelta’ and ‘LAMB’, have achieved better generalization \cite{abuelhaija2017proportionate,you2019reducing}. Using ‘LAMB’ optimizer, You et al. \yrcite{you2019reducing} have reduced the pre-training time of BERT to 76 minutes.
In parallel with the above-mentioned approach, new methods such as adding noise to gradients \cite{wen2019interplay}, applying second-order information \cite{martens2015optimizing,yao2018large,adya2018nonlinear}, using local SGD instead of large batch size \cite{lin2018don't}, and batch augmentation \cite{hoffer2017train} have also attracted a lot of interest.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_001.eps}}
\vskip \figtocaption{}
\caption{Impact of batch size on AlexNet accuracy using Cifar10 dataset}
\label{fig_1}
\end{center}
\vskip \figrear{}
\end{figure}
In this study, we explore the evolution of the basic geometry properties of NNs with varying batch sizes. Such geometrical properties include parameter, gradient, and loss. Our main work include:
\begin{itemize}
\item First, we make a theoretical and experimental analysis on the geometry limit of large batch size training.
\item Based on the work above, the authors propose a guidance to improve ‘generalization gap’.
\item Then we design a parameter update algorithm based on the approximated loss curve curvature.
\item In the last part, we provide theoretical backbones for the major part of the existing research results of ‘generalization gap’, and investigate the geometry picture of large batch size training.
\end{itemize}
\section{Theoretical and Experimental Analysis}
Figure \ref{fig_1} raises a highly concerned problem that large batch size deteriorates the training performance. To avoid subjective bias and search for an in-depth insight, we focus on inspecting of some fundamental network properties, including gradient, parameter update step length and loss update step length with varying batch sizes. If there’s no special specifications, all the experiments in this paper are carried out on Pytorch, using AlexNet and Cifar10 dataset, with SGD as the default optimizer.
\subsection{Gradient}
We probe the impact of batch size on gradient by combining theory and experiments. The curvature of loss determines the parameter update amplitude in terms of LR, and in a update step, the related parameter will inch closer to a nearby minima of the loss curve, with a update step length given by the product of gradient and LR. We find that curvature across layers in a network varies remarkably (Figure \ref{fig_2}), and thus it is reasonable to treat gradient in a manner of layer independent. Theoretically, the gradient of each parameter is the average value of the samples in a whole batch size, and we suppose here that the distribution of gradient on a layer is Gaussian especially in the initial training epochs:
\vskip -15pt
\begin{equation}
\label{eqn_1}
g^k_i\sim N(\mu,\sigma)
\end{equation}
\vskip \eqnlowerspace
where \(g_i^k\) is the \(i\)th gradient corresponding to the \(k\)th sample in a given layer; \(\mu\) and \(\sigma\) denote the mean and standard deviation of the distribution, respectively. \(\mu=0\) is another assumption considering the parameters evenly distributed near the minima. \(\sigma\) is treated as a constant, which relates to both the average curvature of the layer and the structure of the network. The mean gradient over a batch of sample is
\vskip \eqnupspace
\begin{equation}
\label{eqn_2}
g_i=\frac{1}{n}\sum_{k=0}^n g^k_i\sim N(0,\frac{\sigma}{\sqrt{n}})
\end{equation}
\vskip \eqnlowerspace
where \(n\) is the batch size. The probability density function is
\vskip \eqnupspace
\begin{equation}
\label{eqn_3}
f(g_i)=\frac{\sqrt{n}}{\sigma \sqrt{2\pi}}e^{-\frac{1}{2}(\frac{g_i}{\sigma \sqrt{n}})^2}
\end{equation}
\vskip \eqnlowerspace
The expectation of the absolute value of \(g_i\) on all gradients in a layer is
\vskip \eqnupspace
\begin{equation}
\label{eqn_4}
E(\left|g_i \right|)=\int_{-\infty}^\infty f(g_i)\left|g_i \right|d g_i=\frac{2\sigma}{\sqrt{\pi}}\frac{1}{\sqrt{n}}
\end{equation}
\vskip \eqnlowerspace
This function predicts that the mean gradient will decrease when batch size increases. To verify this prediction, in the first update step, we compute \(E(\left|g_i \right|)\) of the first fully connected layer of AlexNet on Cifar10 dataset from batch size 32 to 60000, and Figure \ref{fig_3} shows the expected result.
The consistence between theoretical and experimental results indicates that gradients falling to zero with large batch size is immutable, and it is doomed very large batch size training will result in a poor model performance, if there is no adaption on the network architecture.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_002.eps}}
\vskip \figtocaption{}
\caption{Curvature variations across the layers of AlexNet. The average curvature radius of Bias is artificially enlarged by10 times.}
\label{fig_2}
\end{center}
\vskip \figrear{}
\end{figure}
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_003.eps}}
\vskip \figtocaption{}
\caption{Impact of batch size on the average gradient. The average gradient is the L1 norm of the gradients divided by the number of the parameters in the first fully connected layer of AlexNet.}
\label{fig_3}
\end{center}
\vskip \figrear{}
\end{figure}
\subsection{Parameter}
The gradient vanishing effect of large batch size will definitely lower the parameter update step length. The common parameter update rule is
\vskip -15pt
\begin{equation}
\label{eqn_5}
\Delta w_i=lr(n) g_i
\end{equation}
\vskip \eqnlowerspace
where \(\Delta w_i\) and \(lr(n)\) are the parameter update step length and the LR with respect to batch size \(n\) , respectively. Combining eqn. \ref{eqn_4}, the average parameter update step length across a specific layer is
\vskip \eqnupspace
\begin{equation}
\label{eqn_6}
E(\left| \Delta w_i \right|)=lr(n) E(\left| g_i \right|)=\frac{2\sigma}{\sqrt{\pi}}\frac{lr(n)}{\sqrt{n}}
\end{equation}
\vskip \eqnlowerspace
To validate this finding, we compute the first epoch value of the ‘Normalized parameter stride’, \( E(\left| \Delta w_i \right|) \verb|/| lr(n) \), with varying batch sizes. As shown in Figure \ref{fig_4}, we observe a significant shrink of this value with large batch size.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_004.eps}}
\vskip \figtocaption{}
\caption{Impact of batch size on the normalized parameter update step length in the first epoch.}
\label{fig_4}
\end{center}
\vskip \figrear{}
\end{figure}
The parameter searching space accumulated from each update step is a key point for network optimization. It is obvious that the probability to find a better minima is bigger with larger searching space. We record the ‘Parameter update stride’, \( E(\left| \Delta w_i \right|) \), of each epoch during multiple AlexNet training tasks (Cifar10 dataset). As shown in Figure \ref{fig_5}, the area under a single result curve represents the total parameter searching space for a specific batch size. For small batch sizes, the parameter searching space is big enough to obtain an optimal minima for each parameter, while for large batch sizes, it is more likely a small-area minima will be found and thus less likely a good-performance model will be trained.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_005.eps}}
\vskip \figtocaption{}
\caption{Parameter update stride from batch-size 32 to batch-size 49152. The area below a curve represents the parameter searching space.}
\label{fig_5}
\end{center}
\vskip \figrear{}
\end{figure}
\subsection{Loss}
The vanishing of both gradient and parameter update step length will definitely affect the loss update process. We commit a brief theoretical derivation to study the rule change.
Referring the geometry relations of \( \Delta L \) and \( \Delta w_i \) in Figure \ref{fig_6}, the loss decrement induced by a single parameter update step is
\vskip -15pt
\begin{equation}
\label{eqn_7}
\Delta L_i=\Delta w_i g_i=lr(n) (g_i)^2
\end{equation}
\vskip \eqnlowerspace
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=7cm]{fig_006.pdf}}
\vskip -0.1in
\caption{A conceptual sketch of minima, gradient, parameter update step length, and loss update step length.}
\label{fig_6}
\end{center}
\vskip \figrear{}
\end{figure}
Then the average loss update step length across the batch of samples is
\vskip \eqnupspace
\begin{equation}
\label{eqn_8}
E(\Delta L_i )=lr(n)E((g_i)^2 )=\sigma^2 \frac{lr(n)}{n}
\end{equation}
\vskip \eqnlowerspace
This equation gives the theoretical results of the average loss decrement in a single update step. To check the impact of large batch size training on loss changing, we compute the ‘First epoch loss stride’, \( E(\Delta L_i ) / lr(n) \), in the first epoch of training process, and the results shown in Figure \ref{fig_7} are consistent with our theoretical predictions.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_007.eps}}
\vskip \figtocaption{}
\caption{Impact of batch size on the normalized first epoch loss update length}
\label{fig_7}
\end{center}
\vskip \figrear{}
\end{figure}
In addition to this experiment, we also compare the whole training process with varying batch sizes in Figure \ref{fig_8}. Besides the loss curves changing as expected, we observe a phenomenon that for both small and large batch sizes, the fluctuation amplitude of the loss curves are big, while for medium batch size, the amplitude are relatively small. We preliminary judge that there are two factors that can lead to network instability: the small parameter update step length in large batch-size training and the big randomness in small batch-size training.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_008.eps}}
\vskip \figtocaption{}
\caption{Loss curves from batch-size 32 to batch-size 49152}
\label{fig_8}
\end{center}
\vskip \figrear{}
\end{figure}
\section{Methods to Enlarge the Gradient}
In Section 2, we demonstrate the disadvantages of large batch size training originate from gradient vanishing effect. Thus, on the other hand, it also provides a clue to improve the performance of NN training. We have tried a lot of methods to enlarge the gradient, including discarding small-loss samples, real time tuning of batch size in a training process, discarding small value gradients, setting larger dropout ratio, and selecting sharper loss function. Two methods will be described in the following.
\subsection{Discarding Small-loss Samples}
We conjecture that the input samples with higher losses will produce larger gradients, and we design an experiment to compute the average gradients of the second fully connected layer of AlexNet.
In the first step, a set of tests are finished, with each one discards p\% of small-loss samples. The loss defined here is computed using a single input sample in a batch size, and the discarding ratio, p\%, changes equally from 10\% to 90\%. Figure \ref{fig_9} shows the average gradients with varying discarding ratios, which illustrates discarding small-loss samples will enlarge the gradient.
The next step is to test the actual model training. We compare the baseline to our model for both batch-size 2000 and 8000. The training parameters in our model are the same as the baseline except discarding 30\% small-loss samples in the first 100 epochs. The result shows an observable improvements for both batch-size 2000 and 8000.
A point to note here is that for batch size larger than 8000, the effect of discarding small-loss samples is not significant, which needs further research.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_009.eps}}
\vskip \figtocaption{}
\caption{Impact of loss discarding ration on average gradient. The small-loss samples are discarded in each experiment. The tests are carried out in AlexNet on Cifar10 dataset with batch-size 8192 and update-step 1.}
\label{fig_9}
\end{center}
\vskip \figrear{}
\end{figure}
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_010.eps}}
\vskip \figtocaption{}
\caption{Training accuracy of baseline and the model with 30\% lower loss discarded. The vertical lines indicate the standard deviation. The baseline model is AlexNet on Cifar10 dataset using SGD optimizer.}
\label{fig_10}
\end{center}
\vskip \figrear{}
\end{figure}
\subsection{Real Time Tuning of Batch Size in a Training Process}
As shown in Figure \ref{fig_3}, a smaller batch size produces larger gradients, so it is worth trying to introduce a different batch size in some steps of a training process. Mikami \& Suganuma \yrcite{Mikami2018imageNet} already used this method to achieve better generalization. In this part, we show some simple comparative experiments as a preliminary exploration.
We use the same training parameters as baseline except in the first epoch, the batch-size and LR are tuned down to 512 and 0.005, respectively, which will be reset to 8192 and 0.05 from the second epoch.
We have recorded a group of test results for both baseline model and our model (Figure \ref{fig_11}, \ref{fig_12}, \ref{fig_13}, \ref{fig_14}). For the baseline model, the small update step makes the training process more stumble. The big fluctuation in the first 80 epochs of both loss and accuracy curves hinder the subsequent training and results in models with big performance dispersion. The intrinsic reason is related to the small parameter update step length.
While for our model, the training performance is clearly improved, as shown in Figure \ref{fig_13} and \ref{fig_14}. By decreasing the batch size in the first epoch, the update step length is amplified, which enables searching a broader space for a better loss curve minima.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_011.eps}}
\vskip \figtocaption{}
\caption{Multiple loss curves of baseline model. The baseline model is AlexNet on Cifar10 dataset with batch-size 8192 and use SGD as optimizer.}
\label{fig_11}
\end{center}
\vskip \figrear{}
\end{figure}
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_012.eps}}
\vskip \figtocaption{}
\caption{Multiple accuracy curves of baseline model}
\label{fig_12}
\end{center}
\vskip \figrear{}
\end{figure}
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_013.eps}}
\vskip \figtocaption{}
\caption{Multiple loss curves with first-epoch batch size reduced}
\label{fig_13}
\end{center}
\vskip \figrear{}
\end{figure}
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_014.eps}}
\vskip \figtocaption{}
\caption{Multiple accuracy curves with first-epoch batch size reduced}
\label{fig_14}
\end{center}
\vskip \figrear{}
\end{figure}
\section{Parameter Update Rules Based on Curvature of Loss Curve}
Besides the gradient vanishing effect, the non-uniform distribution of loss curve curvature across layers in a network is another important factor affecting the performance of large batch size training. As shown in Figure \ref{fig_2}, the curvature heterogeneity is significant, and needs pay more attention to. From a geometric point of view, the LR of a specific parameter should be consistent with the loss curve curvature where the parameter located. In case of small batch size, due to the large parameter searching space (Figure \ref{fig_5}), it is applicable to set a global LR, while for large batch size, the decreasing parameter update step length (Figure \ref{fig_4}) makes the training highly sensitive to LR, and thereby, a customized LR setting method is necessary.
\subsection{Curvature-based LR Setting Method and its Approximation}
The best way to update parameters is to set a specific LR for each parameter based on its curvature. The radius of curvature of parameter \( w_i \) is
\vskip \eqnupspace
\begin{equation}
\label{eqn_9}
R_i = \left| \frac{(1+(\frac{d L}{d w_i})^2)^{\frac{3}{2}}}{\frac{d^2 L}{d w_i^2}} \right|
\end{equation}
\vskip \eqnlowerspace
where \( R_i \) is the radius of curvature, \( dL/dw_i) \) is the first order gradient and \( dL/dw_i=g_i \), and \( d^2 L/d w_i^2 \) is the second order gradient. Here the LR for \( w_i \) is
\vskip -15pt
\begin{equation}
\label{eqn_10}
\eta_i=\gamma R_i
\end{equation}
\vskip \eqnlowerspace
where \(\eta_i \) is the LR, and \( \gamma \) is a global hyper parameter. Then follow the rule below to update the parameter,
\vskip \eqnupspace
\begin{equation}
\label{eqn_11}
\Delta w_i=\eta_i\frac{d L}{d w_i}
\end{equation}
\vskip \eqnlowerspace
The above steps are the backbone of curvature-based parameter update rule. Though it is a vanilla method, the huge computation work needed to obtain curvature radius in a big NN limits its applications. The main difficulty results from the lack of high efficiency methods to compute second order gradient in almost all deep learning platforms. In this section we develop a method to approximate the curvature radius.
Referring to Figure \ref{fig_15}, Morse theory states that in the vicinity of a local critical point of a smooth curve, the geometry is equivalent to a second-order curve, such that
\vskip -15pt
\begin{equation}
\label{eqn_12}
L=a_i(w_i-b_i)^2+c_i
\end{equation}
\vskip \eqnlowerspace
where \( a_i \), \( b_i \), and \( c_i \) are the characteristic coefficients of the specific parabola.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=6.5cm]{fig_015.pdf}}
\vskip -0.1in
\caption{A conceptual sketch of minima. The local geometry equivalents to a parabola.}
\label{fig_15}
\end{center}
\vskip \figrear{}
\end{figure}
From this equation we derive the first order gradient:
\vskip \eqnupspace
\begin{equation}
\label{eqn_13}
\frac{d L}{d w_i}=2a_i(w_i-b_i)
\end{equation}
\vskip \eqnlowerspace
and second order gradient:
\vskip \eqnupspace
\begin{equation}
\label{eqn_14}
\frac{d^2 L}{d w_i^2}=2a_i
\end{equation}
\vskip \eqnlowerspace
Substituting eqn. \ref{eqn_13} into eqn. \ref{eqn_14}, we get
\vskip \eqnupspace
\begin{equation}
\label{eqn_15}
\frac{d^2 L}{d w_i^2}=\frac{\frac{d L}{d w_i}}{(w_i-b_i)}
\end{equation}
\vskip \eqnlowerspace
Using this gradient expression, the curvature radius of parameter \( w_i \) becomes
\vskip \eqnupspace
\begin{equation}
\label{eqn_16}
R_i = \left| \frac{(w_i-b_i)(1+(\frac{d L}{d w_i})^2)^{\frac{3}{2}}}{\frac{d L}{d w_i}} \right|
\end{equation}
\vskip \eqnlowerspace
where \( w_i \) and \( d L/d w_i \) are detectable quantities, and \( b_i \) is an undetectable coefficients. We propose a simplification here that \( b_i \) can be set to 0, which is reasonable in statistical meanings, especially when weight decay method is used. The \( (d L/d w_i )^2 \) item is also eliminated considering that its value is far less than 1. Then we get
\vskip \eqnupspace
\begin{equation}
\label{eqn_17}
R_i\approx \left| \frac{w_i}{\frac{d L}{d w_i}} \right|
\end{equation}
\vskip \eqnlowerspace
This equation is a rough estimate of curvature radius, and it will totally fail in either below cases:
\vskip -15pt
\begin{equation}
\label{eqn_18}
w_i\rightarrow 0
\end{equation}
\vskip \eqnlowerspace
and
\vskip \eqnupspace
\begin{equation}
\label{eqn_19}
\frac{d L}{d w_i}\rightarrow 0
\end{equation}
\vskip \eqnlowerspace
Though the majority of \( R_i \) are not accurate enough for parameter updating, we estimate some kinds of statistics in a certain group of \( R_i \) are still usable, for example some forms of layer-wise average curvature radius can be used to set the LR.
\subsection{Parameter Optimization Rule Based on the Median of Curvature Radius}
We propose a simple statistical approximation of eqn. \ref{eqn_17} to compute the median value of a group of \( R_i \)
\vskip \eqnupspace
\begin{equation}
\label{eqn_20}
R_m\approx \left| \frac{w_m}{\frac{d L}{d w_m}} \right|
\end{equation}
\vskip \eqnlowerspace
where \( w_m \), \( d L/d w_m \) and \( R_m \) are the median values of parameter, gradient and curvature radius in a group (a layer in a NN for example), respectively. Now the LR can be set as
\vskip -15pt
\begin{equation}
\label{eqn_21}
\eta_g = \gamma R_m
\end{equation}
\vskip \eqnlowerspace
where \( g \) is the group index. If weight decay is considered, instead of eqn. \ref{eqn_20}, we use
\vskip \eqnupspace
\begin{equation}
\label{eqn_22}
R_m\approx \left| \frac{w_m}{\frac{d L}{d w_m}+\beta w_m} \right|
\end{equation}
\vskip \eqnlowerspace
where \( \beta \) is the weight decay coefficient.
Following the above parameter optimization rule, we have trained ResNet18 on Cifar10 dataset for a series of batch sizes, and find it works well for batch size less than 8192. Figure \ref{fig_16} illustrates the comparation of our method to LARS, and the discrepancy is negligibly small.
We also notice that the training performance falls faster than LARS for large batch size. The reason is the decreasing average gradient (Figure \ref{fig_3}) degrades the validation of \( R_m \) (eqn. \ref{eqn_18}, \ref{eqn_19}, \ref{eqn_20}). Thus, this method can be regarded as a verification of curvature-based parameter update principle, and we do not recommend extending it to large batch size training.
\begin{figure}[ht]
\vskip \fighead{}
\begin{center}
\centerline{\includegraphics[width=\figwidth]{fig_016.eps}}
\vskip \figtocaption{}
\caption{Training accuracy of median-curvature and LARS with batch-size 1024}
\label{fig_16}
\end{center}
\vskip \figrear{}
\end{figure}
\subsection{Interpretation of Layer-wise Parameter Update Algorithms}
In order to solve large batch size training problem, researchers have proposed one specific class of algorithms, the common point of which is setting a layer-wise LR \cite{you2017large,you2017scaling,abuelhaija2017proportionate,you2018imagenet,you2019reducing}. These algorithms are based on the experimental results, without convincing theoretical foundations. Three typical methods are the so-called LARS \cite{you2017large}, PercentDelt \cite{abuelhaija2017proportionate} and LAMB \cite{you2019reducing}. The definition of the former two algorithms are
LARS,
\vskip \eqnupspace
\begin{equation}
\label{ean_23}
\eta_g=\gamma \frac{\|w_i\|_2}{\|\frac{d L}{d w_i}\|_2}
\end{equation}
\vskip \eqnlowerspace
and PercentDelta,
\vskip \eqnupspace
\begin{equation}
\label{eqn_24}
\eta_g=\gamma \frac{size(w_i)}{\| \frac{d L}{d w_i}/w_i \|_1}
\end{equation}
\vskip \eqnlowerspace
where \( size(w_i) \) is the number of parameters in \( g \)th layer. These two algorithms are conjectural solutions to the curvature heterogeneity problem (Figure \ref{fig_2}). Comparing both eqn. \ref{ean_23} and \ref{eqn_24} to eqn. \ref{eqn_17}, it is clear that the main part of LARS and PercentDelta are two transformers of curvature radius (eqn. \ref{eqn_17}). By calculating the norms of the numerator and denominator, two specific statistical mean curvatures are obtained, which can avoid the minimum and maximum problems in the original curvature formula (eqn. \ref{eqn_20}). What the researchers usually ignore is that the algorithms have the same failure conditions, eqn. \ref{eqn_18} and \ref{eqn_19}, which need to be dealt with by careful parameter initialization.
\section{Discussion}
\subsection{Where Does Parameter Stay in the Loss Curve?}
In this section, we will discuss the geometry meanings of gradient vanishing induced by large batch size training, and will provide a novel insight into how large batch size training degrades the model performance.
Referring to Figure \ref{fig_15}, the distance of \( w_i \) from its nearest minima, \( b_i \), is
\vskip -15pt
\begin{equation}
\label{eqn_25}
d_i=w_i-b_i
\end{equation}
\vskip \eqnlowerspace
Considering eqn. \ref{eqn_13}, we get
\vskip \eqnupspace
\begin{equation}
\label{eqn_26}
d_i=\frac{1}{2a_i} \frac{d L}{d w_i} = \frac{1}{2a_i} g_i
\end{equation}
\vskip \eqnlowerspace
For simplification, we suppose \( a_i \) to be a constant, such that \( a_i=a \). Then, we get
\vskip \eqnupspace
\begin{equation}
\label{eqn_27}
d_i\sim N(0,\frac{\sigma}{2a\sqrt{n}})
\end{equation}
\vskip \eqnlowerspace
The expectation of the absolute value of \( d_i \) on all parameters in a layer is
\vskip \eqnupspace
\begin{equation}
\label{eqn_28}
E(\left| d_i \right|)=\frac{\sigma}{a\sqrt{\pi}}\frac{1}{\sqrt{n}}
\end{equation}
\vskip \eqnlowerspace
If batch size, \( n \), inflates to an infinity value, this equation predicts that the average distance between a parameter and its nearest minima, \( \left| d_i \right| \), will move to 0. In other words, all the parameters, \( w_i \), in the network will center on a local minimum point of the loss curve, a consequence that makes the parameters optimizing stop and the network training stuck. Making things worse, it happens at the beginning of the training, and the random initialized loss curve strays far from the perfect model that we expect.
One more thing to note is that the vanishing of \( \left| d_i \right| \) also illustrates adopting Morse theory in Subsection 4.1 is valid.
\subsection{How Does Current Methods Work?}
In Introduction section, the main research on ‘generalization gap’ are discussed, and in this section, we try to give a simpler understanding to the main part of these work.
Method 1: LR scheduling \cite{goyal2017accurate,smith2017super-convergence:,peng2017megdet:,you2019large-batch}. This class of methods alleviates the uneven distribution of curvature in different layers (Figure \ref{fig_1}). Taking warm-up training for example, the parameters in a sharp minima (small curvature radius) are given the priority to update, and will find a flatter minima. After warm-up, the curvatures in different layers incline to converge, which is very helpful for the follow-up training.
Method 2: Batch size scheduling \cite{devarakonda2017adabatch:,smith2018don't,Mikami2018imageNet}. Following the conclusion of Subsection 3.2, Figure \ref{fig_4}, and eqn. \ref{eqn_6}, the nature of batch size scheduling is to adjust the parameter update step length, which is an effective way to jump out of sharp minima or noise.
Method 3: Training longer \cite{hoffer2017train}. According to Figure \ref{fig_5} and eqn. \ref{eqn_6}, increasing the iteration steps will enlarge the parameter searching space, which results in a higher probability of finding a better minima.
Method 4: eliminating sharp minima \cite{keskar2017on,jastrzebski2018finding,wen2018smoothout:,haruki2019gradient}. Parameter update step length and curvature of minima are the ‘two faces of the same coin’. For large batch size training, with the vanishing of parameter update step length, big curvature of minima (sharp minima) will play a greater role. Hence, it is evident that eliminating sharp minima will improve the generalization.
\section{Conclusion}
In this work, we have derived the theoretical limits of large batch size training, and predict that ‘generalization gap’ will be worsen with larger batch size. We reveal that the vanishing of gradient and parameter update step length are the key reasons to result in generalization difficulty.
We propose a curvature-based optimizer, a second order LR rule to better fit the curvature variations across layers in a network. We also show that one approximation of this optimizer, median curvature, is almost as good as LARS in small batch size training.
According to our theoretical findings, we suggest a new guidance to reduce ‘generalization gap’ by increasing the gradients in a network. Two algorithms, discarding small-loss samples and scheduling batch size, are provided for example. Our results demonstrate that the generalization is effectively enhanced.
The recent research work on ‘generalization gap’ mentioned in Introduction section have made great experiment advances, but unified theories to rule all the discrete explanations are still inadequate. Our work offers simple theories to support the major parts of current researches. The widely-known LARS and PercentDelta optimizers are proved to be approximations of our curvature-based algorithm. Many other methods are well explained based on the relation between parameter update step length and batch size.
For larger batch size, we find theoretically that the parameters are more likely to center on the bottom of minima, which makes parameter update step length reduced, and finally makes the training to stuck by small minima or noise at a higher probability.
| bfeadfa30d67a38c46ca502d04899b748549f9b0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The phenomenon of information diffusion, dubbed as information cascade, is ubiquitous in our daily lives, \textit{e.g.}, the spread of breaking news or virus. Information diffusion prediction is an important and challenging task, which aims at forecasting the future properties or behaviors of an information cascade, such as the eventual size~\cite{deepcas-macro,cascn-macro,macro-coupledGnn} or the next infected user~\cite{NDM,DCE-micro,infVAE-micro}. Current studies of information diffusion prediction have been successfully adopted for many real-world scenarios including epidemiology~\cite{wallinga2004different}, viral marketing~\cite{leskovec2007dynamics}, media advertising~\cite{li2013popularity} and the spread of news and memes~\cite{leskovec2009meme,vosoughi2018spread}.
During the last decade, deep learning techniques have shown their effectiveness in computer vision~\cite{krizhevsky2017imagenet} and natural language processing areas~\cite{devlin2018bert}. Recently, methods~\cite{deepcas-macro,CYAN-RNN,TopoLSTM,SNIDSA} based on recurrent neural networks (RNNs) also achieved promising results in information cascade modeling by treating influenced users as sequential data ranked by their infection timestamps. In RNNs, the entire information diffusion history is encoded into a real-valued vector as the hidden state. However, representing all previously infected users by a single vector could fail to encode all necessary information for future predictions due to the mode collapse~\cite{che2016mode} problem. As illustrated in Fig.\ref{fig_intro}, an information item is spreading through two communities and the cascade sequence has infected $5$ users ($4$ from community A and $1$ from community B). An information cascade model need to encode these users. Learned representations of conventional methods can encode the most important factor (community A) for future predictions, but fail to remember others (community B).
\begin{figure}[htb]
\centering
\includegraphics[scale=0.4]{pic/intro.pdf}
\caption{An illustration of mode collapse problem in information cascade modeling. }
\label{fig_intro}
\end{figure}
To address this problem, we propose to employ the idea of disentangled representation learning, which aims to extract multiple latent factors representing different aspects of the data, for modeling the diffusion history of an information cascade. Though disentangled representation learning was originally proposed in controllable image generation~\cite{chen2016infogan}, it has been successfully adopted for other areas such as text generation~\cite{stylicSPG-disentangled} as well. To the best of our knowledge, we are the first to use disentangled representation learning for information diffusion prediction.
To be more specific, we propose a novel Sequential Information Diffusion model with Disentangled Attention (SIDDA) by applying a sequential attention module and a disentangled attention module on RNNs. The former module can help better aggregate the history information and the latter one will disentangle the hidden state vector into multiple representations. We further employ Gumbel-Softmax~\cite{gumbel} technique to encourage bigger differences between the disentangled representations. Consequently, we are able to learn multiple hidden state vectors characterizing different latent factors for future predictions at each step of cascade modeling. Different hidden state vectors can complement each other and provide more comprehensive information to predict the next infected users.
We conduct experiments on three public real-world datasets and employ the task of next infected user prediction for evaluation. Experimental results show that the proposed model SIDDA significantly outperforms state-of-the-art baseline methods by up to $14\%$ in terms of $hits@50$ metric.
Our contributions are as follows:
\begin{itemize}
\item To the best of our knowledge, we are the first work to adopt the idea of disentangled representation learning for information cascade modeling.
\item We propose SIDDA, a novel method which can learn disentangled representations encoding different factors for information diffusion prediction.
\item Experiments on three real-world datasets demonstrate that our proposed method outperforms state-of-the-art baselines significantly.
\end{itemize}
\section{Related Work}
\subsection{Information Diffusion Prediction}
There are two kinds of tasks associated with information diffusion prediction: (1) Studying the growth of cascades and (2) Predicting the next activated node in the cascade. In some recent works~\cite{cascn-macro,NDM,FOREST,zhou2020survey}, these tasks are classified as macroscopic and microscopic predictions.
\subsubsection{Macroscopic Information Diffusion Prediction}
Macroscopic information diffusion prediction, also known as popularity prediction, aims at predicting cascade sizes in the future~\cite{zhao2015seismic}. Previously, related works are based on feature-based approaches~\cite{cancascadebepredicted} and generative approaches~\cite{generative-hawkes}. In recent years, with the success of deep learning, methods based on Recurrent Neural Networks (RNNs) are proposed, \textit{e.g.}, DeepCas~\cite{deepcas-macro} and Deephawkes~\cite{deephawkes-macro}. Some researchers also introduce Graph Neural Networks (GNNs) to model the underlying social graph and diffusion paths, \textit{e.g.}, CoupledGNN~\cite{macro-coupledGnn} and HDGNN~\cite{heterogeneous4-HDGNN-macro}.
\subsubsection{Microscopic Information Diffusion Prediction}
Our work focuses on microscopic information diffusion prediction, which aims to forecast the next activated node given previously infected users in an information cascade.
With the development of deep learning, various kinds of neural networks have been adopted for microscopic information diffusion prediction. Generally, RNN, GNN and attention mechanism are widely used, and give promising results. Topo-LSTM~\cite{TopoLSTM} proposes a novel LSTM to model tree-structured cascades. CYAN-RNN~\cite{CYAN-RNN} and Deep-Diffuse~\cite{deepdiffuse} take timestamps into consideration with the help of temporal point processes. DyHGCN~\cite{heterogeneous2-DyHGCN} proposes a heterogeneous graph convolutional network to model the social graph and dynamic diffusion graph jointly. HDD~\cite{heterogeneous1} exploits meta-path in the diffusion graph and learns heterogeneous network representations using GNN. There are also some models fully based on attention mechanisms, \textit{e.g.}, DAN~\cite{DAN}, Hi-DAN~\cite{Hi-DAN} and NDM~\cite{NDM}.
Besides, some works target on both macroscopic and microscopic information diffusion predictions. FOREST~\cite{FOREST} makes use of reinforcement learning to incorporate the ability of macroscopic prediction. DMT-LIC~\cite{dmt-lic-related} is a multi-task learning framework with a shared-representation layer and two different task layers for predictions.
However, existing methods represent previously infected users by a single vector and could encounter the mode collapse problem, which would fail to encode all necessary information for future predictions.
\subsection{Disentangled Representation Learning}
Disentangled representation learning which aims to learn representations of different latent factors hidden in the observed data~\cite{representationlearning,chen2016infogan}, has been successfully used in various fields. In the field of recommendation systems, a few works are proposed to learn disentangled item representations~\cite{ma2019learningcosinedot}. \cite{ma2020disentangled} performs self-supervision in the latent space, getting disentangled intentions from item sequences for product recommendations. In the field of natural language processing,
SPG~\cite{stylicSPG-disentangled} proposes an unsupervised way to generate stylistic poetry by maximizing the mutual information between representations and generated text. GNUD~\cite{GNUD-disentangled} explores user interest disentanglement in news recommendation by making use of DisenGCN~\cite{madisentangledgnn-DisenGCN}. Besides, disentangled representation learning is also successfully used in the modeling of graph-structured data~\cite{multi-aspect,madisentangledgnn-DisenGCN, disentangledGNN-IPGDN}.
As far as we know, we are the first to adopt the idea of disentangled representation learning for information diffusion prediction.
\section{Method}
In this section, we will formalize the problem of information diffusion prediction and introduce the notations. Then we will propose an RNN-based information diffusion prediction model and a novel attention mechanism to learn disentangled representations. Finally, we will predict the next infected node using the disentangled representations.
\subsection{Problem Definition}
Given a node (or user) set $V$ and a cascade set $C$, we have $V=\{v_1,v_2,...,v_N\}$ and $C=\{c_1,c_2,...,c_M\}$. Here $N$ is the size of the node set, and $M$ is the number of cascades. Each cascade $c_i \in C$ spreading among users is a sequence of nodes $\{v^i_1,v^i_2,...,v^i_{\left | c_i \right |}\}$, ordered by their activation timestamps. Following the same setting in previous works~\cite{NDM, TopoLSTM, SNIDSA, FOREST}, we only keep the order of nodes and ignore the exact timestamps. A detailed modeling of timestamp information will be left for future work.
Information diffusion prediction can be formulated as: given the cascade sequence of previously activated nodes $\{v^i_1,v^i_2,...,v^i_t\}$ in cascade $c_i$ for $t=1,2,...,\left | c_i \right |-1$, predicting the next activated node $v^i_{t+1}$. In other words, our goal is to build a model that is able to learn the conditional probability function $p(v^i_{t+1}|\{v^i_1,v^i_2,...,v^i_t\})$ for each cascade $c_i$ where $1\leq i \leq M$. We will focus on the modeling of a single cascade and omit the superscript for simplification in the rest of the paper.
\subsection{Model Architecture}
The framework of our method is shown in Fig.\ref{model_overall}. Firstly, we encode every node into node embedding by an embedding-lookup layer. Then we employ a Gated Recurrent Unit (GRU) as the basis to model the node sequence. Given the infected node sequence, the GRU model can output a hidden state representing the sequential history at each time step. Then a sequential attention module and a disentangled attention module are applied to better aggregate the history information and disentangle the hidden state vector into multiple representations. Finally, disentangled representations will be used for predicting the next infected node.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{pic/overall.pdf}
\caption{An illustration of our proposed model SIDDA. The model will derive $K$ disentangled representations $\mathbf{y_t^{(k)}}(1 \leq k \leq K)$ at time $t$, with the help of a disentangled attention module.}
\label{model_overall}
\end{figure}
\subsubsection{Gated Recurrent Unit}
Recurrent neural networks(RNNs) have shown great potentials in modeling sequence data. Previous works~\cite{TopoLSTM,SNIDSA,FOREST,CYAN-RNN,deepdiffuse,lstm-cnn-hin-kbs} used RNNs as their basis to model information diffusion cascades. Firstly, we represent each node as a vector $\mathbf{x}\in \mathbb{R}^D$. By feeding the activated node embedding $\mathbf{x_t}$ into an RNN at every timestamp, we can have a hidden state $\mathbf{h_t}$ capturing the history information of all previously activated nodes as the output.
Gated Recurrent Unit (GRU) can alleviate the vanishing gradient problem compared with a standard RNN. It can be considered as a variation of the LSTM but is computationally cheaper. Therefore, we select GRU as our model basis.
\begin{equation}
\mathbf{h_t} = GRU(\mathbf{h_{t-1}}, \mathbf{x_t})
\label{gru}
\end{equation}
\subsubsection{Sequential Attention Module}
Note that every hidden state $\mathbf{h_t}$ in GRU contains information about the input node sequence with a focus around position $t$. To aggregate the sequential history information of all positions automatically, we propose to employ attention mechanism, which was originally proposed in neural machine translation~\cite{attentionpaper}, and learn to assign different attention weights to previous hidden states $\mathbf{h_i}$ $(i\leq t)$. Formally, the sequential attention weight $\alpha_{it}$ of the $i$-th hidden state is computed by:
\begin{gather}
\alpha_{it} = \frac{exp(\frac{1}{\sqrt{D}} e_{it} )}{\sum_{j=1}^{t}exp(\frac{1}{\sqrt{D}}e_{jt})} \\
e_{it} = d(\mathbf{h_i},\mathbf{h_t})\notag
\end{gather}
where $\frac{1}{\sqrt{D}}$ is a scaling coefficient and $h_i\in \mathbb{R}^D$ is the $i$-th hidden state. $d(\cdot,\cdot)$ is a distance function, and we use dot product in this work.
\subsubsection{Disentangled Attention Module}
Inspired by the recent advances in disentangled representation learning~\cite{madisentangledgnn-DisenGCN,ma2020disentangled}, we propose our disentangled attention module to learn multiple representations characterizing different latent factors for future predictions at each step of cascade modeling.
We assume that there are $K$ main latent factors that affect the diffusion of an information item. Take Twitter as an example, an information item may spread through $K$ different communities or users are infected by tweets from $K$ topics. We want to disentangle these diverse latent factors from the hidden state $\mathbf{h_t}$ to get the disentangled representations.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{pic/pki.pdf}
\caption{An illustration of the disentangled attention module.}
\label{model_pki}
\end{figure}
To be more specific, we set $K$ prototype embeddings $\mathbf{p_k}$ $(1\leq k\leq K)$ to represent the $K$ potential factors. We expect these prototype embeddings can capture diverse properties after training. Then for each hidden state $h_i$, we will compute its similarity with $K$ prototypes by the distance between $\mathbf{h_i}$ and $\mathbf{P}$. The equations are as follows:
\begin{equation}
\begin{gathered}
\beta_{ik} = \frac{exp(\frac{1}{\sqrt{D}}d(\mathbf{h_{i}},\mathbf{p_k}))}{\sum_{{k}'=1}^{K}exp(\frac{1}{\sqrt{D}}d(\mathbf{h_{i}},\mathbf{p_{{k}'}}))} \\
d(\mathbf{h_i},\mathbf{p_k}) = \frac{\mathbf{h_i}\cdot \mathbf{p_k}}{\|{\mathbf{h_i}}\| \|{\mathbf{p_k}}\|}
\label{pki}
\end{gathered}
\end{equation}
where $k=1,2,...,K$ and $i=1,2,...,t$. $\{\mathbf{p_k}\in \mathbb{R}^D:1\leq k\leq K\}$ are prototype embeddings of $K$ latent factors. $d(\cdot,\cdot)$ is the distance function, and we select cosine-similarity instead of dot product here because the latter is more vulnerable when it comes to mode collapse~\cite{ma2019learningcosinedot}.
Then we apply a Softmax function to compute the attention score $\beta_{ik}$ for further usage. It is worth noting that by using Softmax, all of the $\beta_{ik}$ $(1\leq k\leq K)$ tend to have a similar result. The hidden state $h_i$ will be disentangled to $K$ latent factors equally, which would harm the performance of disentanglement. We further employ Gumbel-Softmax to alleviate this drawback and enhance the disentanglement effect.
Finally, we can combine the sequence attention and disentanglement attention together, using $\alpha_{ij}$ and $\beta_{ik}$ to draw the weighted sum result $y_i^{(k)}$ to represent the hidden state under the $k$-th factor:
\begin{equation}
\mathbf{y_t^{(k)}} = LayerNorm(\sum_{i=1}^{t}\alpha_{it}\cdot \beta_{ik}\cdot \mathbf{h_i})
\label{pkipi}
\end{equation}
where $k=1,2,...,K$. When $K=1$, our method degenerates into a GRU model with an attention mechanism. We name our model as Sequential Information Diffusion model with Disentangled Attention (SIDDA). We further apply Gumbel-Softmax trick~\cite{gumbel} to encourage the disentanglement.
\subsection{Optimization Objective}
We can get $K$ disentangled representations at each timestamp $t$ using Eq. (\ref{pkipi}). Different hidden state vectors can complement each other and provide more comprehensive information to predict the next infected users. Therefore, we compute the similarity, \textit{i.e.}, dot product, between these $K$ disentangled representations $\mathbf{y_t^{(k)}}$ and node embeddings $\mathbf{X}$ to predict the next activated node. Hence, the loss function can be defined as follows:
\begin{equation}
\begin{aligned}
L(\mathbf{\theta},t) &= -\log p_{\mathbb{\theta}}(\mathbf{x_{t+1}}|\mathbf{y_t}) \\
&= -\log \frac{\max_{k\in \{1,2,...,K\}}exp(\frac{1}{\sqrt{D}} \mathbf{x_{t+1}}\cdot \mathbf{y_t^{(k)}} )}
{\sum_{{v}' \in V}\max_{k\in \{1,2,...,K\}}exp(\frac{1}{\sqrt{D}} \mathbf{x_{{v}'}}\cdot \mathbf{y_t^{(k)}})}
\label{loss}
\end{aligned}
\end{equation}
It is noteworthy that in Eq. (\ref{loss}), the Softmax function is applied to all of the $K$ disentangled factors. The model is forced to preserve different information in $\mathbf{y_t^{(k)}}(1\leq k \leq K)$. Therefore, we do not need to use regularization terms to restrict the distribution of $\mathbf{y_t^{(k)}}$ in our model.
\section{Experiments}
In this section, we will introduce the datasets used in our experiments, and the state-of-the-art baselines which will be compared with our proposed method. We further introduce the evaluation metrics used to evaluate the performance of the baselines and our method. We will design comparative experiments and ablation experiments to show the superiority of our method.
\subsection{Datasets}
\textbf{Twitter}~\cite{twitterdata} dataset collects tweets published in October 2010, containing URLs in the message body. Each URL is considered to be a unique marker of information, spreading among users.
\textbf{Douban}~\cite{doubandata} is a Chinese social networking service network where users can share content about books. In our experiment, we consider each book as an information item, and a user is activated if the user reads the book.
\textbf{Memetracker}~\cite{memetrackerdata} contains millions of online mainstream social media activity, tracking the most frequent phrases, \textit{i.e.}, memes. Memes were information items being shared among websites.
\begin{table}[htbp]
\centering
\caption{Statistics of datasets.}
\begin{tabular}{ccccc}
\toprule
Dataset&\# Cascades&\# Nodes&\# Links &Average Length \\
\midrule
Twitter&3,442&12,627&309,631&32.60\\
Douban&10,602&23,123&348,280&27.14\\
Memetracker&12,661&4,709&-&16.24\\
\bottomrule
\end{tabular}
\label{dataset
\end{table}
We follow the experiment setting in~\cite{FOREST}, randomly selecting 80\% of cascades for training, 10\% for
validation, and 10\% for test. The statistics of datasets
are listed in Table \ref{dataset}. Twitter and Douban datasets have an underlying social graph, which will be used by some of the baseline models.
\subsection{Baselines and Experimental Settings}
We compared our method with several state-of-the-art baselines.
\textbf{DeepDiffuse}~\cite{deepdiffuse} is an LSTM based model with an attention mechanism, utilizing node sequence and their corresponding activated time. We replace timestamps with integer activated steps in the cascade.
\textbf{Topo-LSTM}~\cite{TopoLSTM} is a topological recurrent network, which can model diffusion topology as dynamic directed acyclic graphs (DAGs). In our experiment, the cascade structure is extracted from the underlying social graph.
\textbf{NDM}~\cite{NDM} employs convolutional network and attention mechanism for cascade modeling. It makes relax assumptions on the datasets and doesn't need a diffusion graph.
\textbf{SNIDSA}~\cite{SNIDSA} is a novel sequential neural network with structure attention. It utilizes both the sequential nature of a cascade and structural characteristics of the underlying social graph with the help of a gating mechanism.
\textbf{FOREST}~\cite{FOREST} is a novel GRU-based information diffusion model. It extracts underlying social graph information and uses reinforcement learning to incorporate macroscopic prediction ability into itself.
\subsection{Evaluation Metrics and Experiment Setting}
As pointed out by~\cite{TopoLSTM}, there can be an arbitrary number of potential candidates, information diffusion prediction can be seen as a retrieval task. We use two popular information retrieval evaluation method $\textbf{hits}@N$ and $\textbf{map}@N$. The same evaluation metrics are also used~\cite{FOREST,CYAN-RNN,deepdiffuse, Hi-DAN}. We set $N=10,50,100$ for evaluation. Since SNIDSA and TopoLSTM need an underlying social graph in the dataset, we exclude them for Memetracker.
We implement our method in PyTorch. We use Adam optimizer for mini-batch gradient descent and use dynamic learning rate. The dropout rate is set to $1e-4$. The temperature parameter of Gumbel-Softmax $\tau=1.0$.
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/1.pdf}
\vspace*{-12mm}
\caption{$hits@N$ of Twitter}
\label{overall1}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/2.pdf}
\vspace*{-12mm}
\caption{$map@N$ of Twitter}
\label{overall2}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/3.pdf}
\vspace*{-12mm}
\caption{$hits@N$ of Douban}
\label{overall3}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/4.pdf}
\vspace*{-12mm}
\caption{$map@N$ of Douban}
\label{overall4}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/5.pdf}
\vspace*{-12mm}
\caption{$hits@N$ of Memetracker}
\label{overall5}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/6.pdf}
\vspace*{-12mm}
\caption{$map@N$ of Memetracker}
\label{overall6}
\end{minipage}
\caption{The overall results, scores are the higher the better.}
\label{overall_result}
\end{figure}
\subsection{Experimental Results}
We design the comparative experiments, ablation experiments, and analyze the parameter sensitivity in this section.
\subsubsection{Overall Evaluation}
Fig. \ref{overall_result} shows the performance of different methods on three datasets. We select the results when $K=4$ for our method. For the analysis of parameters, please refer to Section \ref{Parameteranalysissec}. We find that:
SIDDA outperforms all the state-of-the-art baseline methods consistently and significantly on $hits@N$ and $map@N$. The results show that our method can predict the next activated user successfully.
\subsubsection{Ablation Study}
\label{albationstudysec}
\label{Parameteranalysissec}
\paragraph{\textbf{The effect of the number of disentangled factors.}} We study how the number of disentangled factors $K$ can affect the performance of our method. As shown in Fig.\ref{figdifferentk}, our method has the worst performance when $K=1$. Because when $K=1$, our model is a GRU model with an attention mechanism, the latent information is not disentangled thoroughly. The performance on Douban and Memetracker dataset increases with $K$ getting larger, and start to decrease when $K>4$. While performance on the Twitter dataset is much better when $K>4$, showing an increasing trend. This is because Twitter has more diverse topics and may have more factors to disentangle.
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{pic/diffk.pdf}
\caption{Performance with different numbers of disentangled factors $K$.}
\label{figdifferentk}
\end{figure}
\paragraph{\textbf{The effect of the number of model dimension.}} We also study how the dimension of node representations $D$ can affect the performance. We evaluate our method when $K=4$ and $D\in\{16,32,64,128\}$. As illustrated in Fig. \ref{figdifferentd}, on Twitter and Memetracker the performance doesn't converge until $D=128$, possibly because they have larger training sets. Douban dataset converges when $D=64$ and shows a decreasing trend when $D$ gets larger.
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{pic/diffd.pdf}
\caption{Performance with different dimensions of node representations $D$.}
\label{figdifferentd}
\end{figure}
\section{Conclusion}
In this paper, we propose a novel diffusion prediction model, SIDDA, which can extract latent factors by learning multiple hidden states at each timestamp. Specifically, we propose sequential attention to capture sequential relations of information cascades. Then we adopt a disentangled attention to learn multiple hidden states representing different latent factors. Additionally, we employ Gumbel-Softmax to further enhance the disentangled effect. The experiments on three real-world datasets show the superiority of our method when compared with state-of-the-art diffusion prediction models.
\section{Introduction}
The phenomenon of information diffusion, dubbed as information cascade, is ubiquitous in our daily lives, \textit{e.g.}, the spread of breaking news or virus. Information diffusion prediction is an important and challenging task, which aims at forecasting the future properties or behaviors of an information cascade, such as the eventual size~\cite{deepcas-macro,cascn-macro,macro-coupledGnn} or the next infected user~\cite{NDM,DCE-micro,infVAE-micro}. Current studies of information diffusion prediction have been successfully adopted for many real-world scenarios including epidemiology~\cite{wallinga2004different}, viral marketing~\cite{leskovec2007dynamics}, media advertising~\cite{li2013popularity} and the spread of news and memes~\cite{leskovec2009meme,vosoughi2018spread}.
During the last decade, deep learning techniques have shown their effectiveness in computer vision~\cite{krizhevsky2017imagenet} and natural language processing areas~\cite{devlin2018bert}. Recently, methods~\cite{deepcas-macro,CYAN-RNN,TopoLSTM,SNIDSA} based on recurrent neural networks (RNNs) also achieved promising results in information cascade modeling by treating influenced users as sequential data ranked by their infection timestamps. In RNNs, the entire information diffusion history is encoded into a real-valued vector as the hidden state. However, representing all previously infected users by a single vector could fail to encode all necessary information for future predictions due to the mode collapse~\cite{che2016mode} problem. As illustrated in Fig.\ref{fig_intro}, an information item is spreading through two communities and the cascade sequence has infected $5$ users ($4$ from community A and $1$ from community B). An information cascade model need to encode these users. Learned representations of conventional methods can encode the most important factor (community A) for future predictions, but fail to remember others (community B).
\begin{figure}[htb]
\centering
\includegraphics[scale=0.4]{pic/intro.pdf}
\caption{An illustration of mode collapse problem in information cascade modeling. }
\label{fig_intro}
\end{figure}
To address this problem, we propose to employ the idea of disentangled representation learning, which aims to extract multiple latent factors representing different aspects of the data, for modeling the diffusion history of an information cascade. Though disentangled representation learning was originally proposed in controllable image generation~\cite{chen2016infogan}, it has been successfully adopted for other areas such as text generation~\cite{stylicSPG-disentangled} as well. To the best of our knowledge, we are the first to use disentangled representation learning for information diffusion prediction.
To be more specific, we propose a novel Sequential Information Diffusion model with Disentangled Attention (SIDDA) by applying a sequential attention module and a disentangled attention module on RNNs. The former module can help better aggregate the history information and the latter one will disentangle the hidden state vector into multiple representations. We further employ Gumbel-Softmax~\cite{gumbel} technique to encourage bigger differences between the disentangled representations. Consequently, we are able to learn multiple hidden state vectors characterizing different latent factors for future predictions at each step of cascade modeling. Different hidden state vectors can complement each other and provide more comprehensive information to predict the next infected users.
We conduct experiments on three public real-world datasets and employ the task of next infected user prediction for evaluation. Experimental results show that the proposed model SIDDA significantly outperforms state-of-the-art baseline methods by up to $14\%$ in terms of $hits@50$ metric.
Our contributions are as follows:
\begin{itemize}
\item To the best of our knowledge, we are the first work to adopt the idea of disentangled representation learning for information cascade modeling.
\item We propose SIDDA, a novel method which can learn disentangled representations encoding different factors for information diffusion prediction.
\item Experiments on three real-world datasets demonstrate that our proposed method outperforms state-of-the-art baselines significantly.
\end{itemize}
\section{Related Work}
\subsection{Information Diffusion Prediction}
There are two kinds of tasks associated with information diffusion prediction: (1) Studying the growth of cascades and (2) Predicting the next activated node in the cascade. In some recent works~\cite{cascn-macro,NDM,FOREST,zhou2020survey}, these tasks are classified as macroscopic and microscopic predictions.
\subsubsection{Macroscopic Information Diffusion Prediction}
Macroscopic information diffusion prediction, also known as popularity prediction, aims at predicting cascade sizes in the future~\cite{zhao2015seismic}. Previously, related works are based on feature-based approaches~\cite{cancascadebepredicted} and generative approaches~\cite{generative-hawkes}. In recent years, with the success of deep learning, methods based on Recurrent Neural Networks (RNNs) are proposed, \textit{e.g.}, DeepCas~\cite{deepcas-macro} and Deephawkes~\cite{deephawkes-macro}. Some researchers also introduce Graph Neural Networks (GNNs) to model the underlying social graph and diffusion paths, \textit{e.g.}, CoupledGNN~\cite{macro-coupledGnn} and HDGNN~\cite{heterogeneous4-HDGNN-macro}.
\subsubsection{Microscopic Information Diffusion Prediction}
Our work focuses on microscopic information diffusion prediction, which aims to forecast the next activated node given previously infected users in an information cascade.
With the development of deep learning, various kinds of neural networks have been adopted for microscopic information diffusion prediction. Generally, RNN, GNN and attention mechanism are widely used, and give promising results. Topo-LSTM~\cite{TopoLSTM} proposes a novel LSTM to model tree-structured cascades. CYAN-RNN~\cite{CYAN-RNN} and Deep-Diffuse~\cite{deepdiffuse} take timestamps into consideration with the help of temporal point processes. DyHGCN~\cite{heterogeneous2-DyHGCN} proposes a heterogeneous graph convolutional network to model the social graph and dynamic diffusion graph jointly. HDD~\cite{heterogeneous1} exploits meta-path in the diffusion graph and learns heterogeneous network representations using GNN. There are also some models fully based on attention mechanisms, \textit{e.g.}, DAN~\cite{DAN}, Hi-DAN~\cite{Hi-DAN} and NDM~\cite{NDM}.
Besides, some works target on both macroscopic and microscopic information diffusion predictions. FOREST~\cite{FOREST} makes use of reinforcement learning to incorporate the ability of macroscopic prediction. DMT-LIC~\cite{dmt-lic-related} is a multi-task learning framework with a shared-representation layer and two different task layers for predictions.
However, existing methods represent previously infected users by a single vector and could encounter the mode collapse problem, which would fail to encode all necessary information for future predictions.
\subsection{Disentangled Representation Learning}
Disentangled representation learning which aims to learn representations of different latent factors hidden in the observed data~\cite{representationlearning,chen2016infogan}, has been successfully used in various fields. In the field of recommendation systems, a few works are proposed to learn disentangled item representations~\cite{ma2019learningcosinedot}. \cite{ma2020disentangled} performs self-supervision in the latent space, getting disentangled intentions from item sequences for product recommendations. In the field of natural language processing,
SPG~\cite{stylicSPG-disentangled} proposes an unsupervised way to generate stylistic poetry by maximizing the mutual information between representations and generated text. GNUD~\cite{GNUD-disentangled} explores user interest disentanglement in news recommendation by making use of DisenGCN~\cite{madisentangledgnn-DisenGCN}. Besides, disentangled representation learning is also successfully used in the modeling of graph-structured data~\cite{multi-aspect,madisentangledgnn-DisenGCN, disentangledGNN-IPGDN}.
As far as we know, we are the first to adopt the idea of disentangled representation learning for information diffusion prediction.
\section{Method}
In this section, we will formalize the problem of information diffusion prediction and introduce the notations. Then we will propose an RNN-based information diffusion prediction model and a novel attention mechanism to learn disentangled representations. Finally, we will predict the next infected node using the disentangled representations.
\subsection{Problem Definition}
Given a node (or user) set $V$ and a cascade set $C$, we have $V=\{v_1,v_2,...,v_N\}$ and $C=\{c_1,c_2,...,c_M\}$. Here $N$ is the size of the node set, and $M$ is the number of cascades. Each cascade $c_i \in C$ spreading among users is a sequence of nodes $\{v^i_1,v^i_2,...,v^i_{\left | c_i \right |}\}$, ordered by their activation timestamps. Following the same setting in previous works~\cite{NDM, TopoLSTM, SNIDSA, FOREST}, we only keep the order of nodes and ignore the exact timestamps. A detailed modeling of timestamp information will be left for future work.
Information diffusion prediction can be formulated as: given the cascade sequence of previously activated nodes $\{v^i_1,v^i_2,...,v^i_t\}$ in cascade $c_i$ for $t=1,2,...,\left | c_i \right |-1$, predicting the next activated node $v^i_{t+1}$. In other words, our goal is to build a model that is able to learn the conditional probability function $p(v^i_{t+1}|\{v^i_1,v^i_2,...,v^i_t\})$ for each cascade $c_i$ where $1\leq i \leq M$. We will focus on the modeling of a single cascade and omit the superscript for simplification in the rest of the paper.
\subsection{Model Architecture}
The framework of our method is shown in Fig.\ref{model_overall}. Firstly, we encode every node into node embedding by an embedding-lookup layer. Then we employ a Gated Recurrent Unit (GRU) as the basis to model the node sequence. Given the infected node sequence, the GRU model can output a hidden state representing the sequential history at each time step. Then a sequential attention module and a disentangled attention module are applied to better aggregate the history information and disentangle the hidden state vector into multiple representations. Finally, disentangled representations will be used for predicting the next infected node.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{pic/overall.pdf}
\caption{An illustration of our proposed model SIDDA. The model will derive $K$ disentangled representations $\mathbf{y_t^{(k)}}(1 \leq k \leq K)$ at time $t$, with the help of a disentangled attention module.}
\label{model_overall}
\end{figure}
\subsubsection{Gated Recurrent Unit}
Recurrent neural networks(RNNs) have shown great potentials in modeling sequence data. Previous works~\cite{TopoLSTM,SNIDSA,FOREST,CYAN-RNN,deepdiffuse,lstm-cnn-hin-kbs} used RNNs as their basis to model information diffusion cascades. Firstly, we represent each node as a vector $\mathbf{x}\in \mathbb{R}^D$. By feeding the activated node embedding $\mathbf{x_t}$ into an RNN at every timestamp, we can have a hidden state $\mathbf{h_t}$ capturing the history information of all previously activated nodes as the output.
Gated Recurrent Unit (GRU) can alleviate the vanishing gradient problem compared with a standard RNN. It can be considered as a variation of the LSTM but is computationally cheaper. Therefore, we select GRU as our model basis.
\begin{equation}
\mathbf{h_t} = GRU(\mathbf{h_{t-1}}, \mathbf{x_t})
\label{gru}
\end{equation}
\subsubsection{Sequential Attention Module}
Note that every hidden state $\mathbf{h_t}$ in GRU contains information about the input node sequence with a focus around position $t$. To aggregate the sequential history information of all positions automatically, we propose to employ attention mechanism, which was originally proposed in neural machine translation~\cite{attentionpaper}, and learn to assign different attention weights to previous hidden states $\mathbf{h_i}$ $(i\leq t)$. Formally, the sequential attention weight $\alpha_{it}$ of the $i$-th hidden state is computed by:
\begin{gather}
\alpha_{it} = \frac{exp(\frac{1}{\sqrt{D}} e_{it} )}{\sum_{j=1}^{t}exp(\frac{1}{\sqrt{D}}e_{jt})} \\
e_{it} = d(\mathbf{h_i},\mathbf{h_t})\notag
\end{gather}
where $\frac{1}{\sqrt{D}}$ is a scaling coefficient and $h_i\in \mathbb{R}^D$ is the $i$-th hidden state. $d(\cdot,\cdot)$ is a distance function, and we use dot product in this work.
\subsubsection{Disentangled Attention Module}
Inspired by the recent advances in disentangled representation learning~\cite{madisentangledgnn-DisenGCN,ma2020disentangled}, we propose our disentangled attention module to learn multiple representations characterizing different latent factors for future predictions at each step of cascade modeling.
We assume that there are $K$ main latent factors that affect the diffusion of an information item. Take Twitter as an example, an information item may spread through $K$ different communities or users are infected by tweets from $K$ topics. We want to disentangle these diverse latent factors from the hidden state $\mathbf{h_t}$ to get the disentangled representations.
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{pic/pki.pdf}
\caption{An illustration of the disentangled attention module.}
\label{model_pki}
\end{figure}
To be more specific, we set $K$ prototype embeddings $\mathbf{p_k}$ $(1\leq k\leq K)$ to represent the $K$ potential factors. We expect these prototype embeddings can capture diverse properties after training. Then for each hidden state $h_i$, we will compute its similarity with $K$ prototypes by the distance between $\mathbf{h_i}$ and $\mathbf{P}$. The equations are as follows:
\begin{equation}
\begin{gathered}
\beta_{ik} = \frac{exp(\frac{1}{\sqrt{D}}d(\mathbf{h_{i}},\mathbf{p_k}))}{\sum_{{k}'=1}^{K}exp(\frac{1}{\sqrt{D}}d(\mathbf{h_{i}},\mathbf{p_{{k}'}}))} \\
d(\mathbf{h_i},\mathbf{p_k}) = \frac{\mathbf{h_i}\cdot \mathbf{p_k}}{\|{\mathbf{h_i}}\| \|{\mathbf{p_k}}\|}
\label{pki}
\end{gathered}
\end{equation}
where $k=1,2,...,K$ and $i=1,2,...,t$. $\{\mathbf{p_k}\in \mathbb{R}^D:1\leq k\leq K\}$ are prototype embeddings of $K$ latent factors. $d(\cdot,\cdot)$ is the distance function, and we select cosine-similarity instead of dot product here because the latter is more vulnerable when it comes to mode collapse~\cite{ma2019learningcosinedot}.
Then we apply a Softmax function to compute the attention score $\beta_{ik}$ for further usage. It is worth noting that by using Softmax, all of the $\beta_{ik}$ $(1\leq k\leq K)$ tend to have a similar result. The hidden state $h_i$ will be disentangled to $K$ latent factors equally, which would harm the performance of disentanglement. We further employ Gumbel-Softmax to alleviate this drawback and enhance the disentanglement effect.
Finally, we can combine the sequence attention and disentanglement attention together, using $\alpha_{ij}$ and $\beta_{ik}$ to draw the weighted sum result $y_i^{(k)}$ to represent the hidden state under the $k$-th factor:
\begin{equation}
\mathbf{y_t^{(k)}} = LayerNorm(\sum_{i=1}^{t}\alpha_{it}\cdot \beta_{ik}\cdot \mathbf{h_i})
\label{pkipi}
\end{equation}
where $k=1,2,...,K$. When $K=1$, our method degenerates into a GRU model with an attention mechanism. We name our model as Sequential Information Diffusion model with Disentangled Attention (SIDDA). We further apply Gumbel-Softmax trick~\cite{gumbel} to encourage the disentanglement.
\subsection{Optimization Objective}
We can get $K$ disentangled representations at each timestamp $t$ using Eq. (\ref{pkipi}). Different hidden state vectors can complement each other and provide more comprehensive information to predict the next infected users. Therefore, we compute the similarity, \textit{i.e.}, dot product, between these $K$ disentangled representations $\mathbf{y_t^{(k)}}$ and node embeddings $\mathbf{X}$ to predict the next activated node. Hence, the loss function can be defined as follows:
\begin{equation}
\begin{aligned}
L(\mathbf{\theta},t) &= -\log p_{\mathbb{\theta}}(\mathbf{x_{t+1}}|\mathbf{y_t}) \\
&= -\log \frac{\max_{k\in \{1,2,...,K\}}exp(\frac{1}{\sqrt{D}} \mathbf{x_{t+1}}\cdot \mathbf{y_t^{(k)}} )}
{\sum_{{v}' \in V}\max_{k\in \{1,2,...,K\}}exp(\frac{1}{\sqrt{D}} \mathbf{x_{{v}'}}\cdot \mathbf{y_t^{(k)}})}
\label{loss}
\end{aligned}
\end{equation}
It is noteworthy that in Eq. (\ref{loss}), the Softmax function is applied to all of the $K$ disentangled factors. The model is forced to preserve different information in $\mathbf{y_t^{(k)}}(1\leq k \leq K)$. Therefore, we do not need to use regularization terms to restrict the distribution of $\mathbf{y_t^{(k)}}$ in our model.
\section{Experiments}
In this section, we will introduce the datasets used in our experiments, and the state-of-the-art baselines which will be compared with our proposed method. We further introduce the evaluation metrics used to evaluate the performance of the baselines and our method. We will design comparative experiments and ablation experiments to show the superiority of our method.
\subsection{Datasets}
\textbf{Twitter}~\cite{twitterdata} dataset collects tweets published in October 2010, containing URLs in the message body. Each URL is considered to be a unique marker of information, spreading among users.
\textbf{Douban}~\cite{doubandata} is a Chinese social networking service network where users can share content about books. In our experiment, we consider each book as an information item, and a user is activated if the user reads the book.
\textbf{Memetracker}~\cite{memetrackerdata} contains millions of online mainstream social media activity, tracking the most frequent phrases, \textit{i.e.}, memes. Memes were information items being shared among websites.
\begin{table}[htbp]
\centering
\caption{Statistics of datasets.}
\begin{tabular}{ccccc}
\toprule
Dataset&\# Cascades&\# Nodes&\# Links &Average Length \\
\midrule
Twitter&3,442&12,627&309,631&32.60\\
Douban&10,602&23,123&348,280&27.14\\
Memetracker&12,661&4,709&-&16.24\\
\bottomrule
\end{tabular}
\label{dataset
\end{table}
We follow the experiment setting in~\cite{FOREST}, randomly selecting 80\% of cascades for training, 10\% for
validation, and 10\% for test. The statistics of datasets
are listed in Table \ref{dataset}. Twitter and Douban datasets have an underlying social graph, which will be used by some of the baseline models.
\subsection{Baselines and Experimental Settings}
We compared our method with several state-of-the-art baselines.
\textbf{DeepDiffuse}~\cite{deepdiffuse} is an LSTM based model with an attention mechanism, utilizing node sequence and their corresponding activated time. We replace timestamps with integer activated steps in the cascade.
\textbf{Topo-LSTM}~\cite{TopoLSTM} is a topological recurrent network, which can model diffusion topology as dynamic directed acyclic graphs (DAGs). In our experiment, the cascade structure is extracted from the underlying social graph.
\textbf{NDM}~\cite{NDM} employs convolutional network and attention mechanism for cascade modeling. It makes relax assumptions on the datasets and doesn't need a diffusion graph.
\textbf{SNIDSA}~\cite{SNIDSA} is a novel sequential neural network with structure attention. It utilizes both the sequential nature of a cascade and structural characteristics of the underlying social graph with the help of a gating mechanism.
\textbf{FOREST}~\cite{FOREST} is a novel GRU-based information diffusion model. It extracts underlying social graph information and uses reinforcement learning to incorporate macroscopic prediction ability into itself.
\subsection{Evaluation Metrics and Experiment Setting}
As pointed out by~\cite{TopoLSTM}, there can be an arbitrary number of potential candidates, information diffusion prediction can be seen as a retrieval task. We use two popular information retrieval evaluation method $\textbf{hits}@N$ and $\textbf{map}@N$. The same evaluation metrics are also used~\cite{FOREST,CYAN-RNN,deepdiffuse, Hi-DAN}. We set $N=10,50,100$ for evaluation. Since SNIDSA and TopoLSTM need an underlying social graph in the dataset, we exclude them for Memetracker.
We implement our method in PyTorch. We use Adam optimizer for mini-batch gradient descent and use dynamic learning rate. The dropout rate is set to $1e-4$. The temperature parameter of Gumbel-Softmax $\tau=1.0$.
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/1.pdf}
\vspace*{-12mm}
\caption{$hits@N$ of Twitter}
\label{overall1}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/2.pdf}
\vspace*{-12mm}
\caption{$map@N$ of Twitter}
\label{overall2}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/3.pdf}
\vspace*{-12mm}
\caption{$hits@N$ of Douban}
\label{overall3}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/4.pdf}
\vspace*{-12mm}
\caption{$map@N$ of Douban}
\label{overall4}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/5.pdf}
\vspace*{-12mm}
\caption{$hits@N$ of Memetracker}
\label{overall5}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{pic/6.pdf}
\vspace*{-12mm}
\caption{$map@N$ of Memetracker}
\label{overall6}
\end{minipage}
\caption{The overall results, scores are the higher the better.}
\label{overall_result}
\end{figure}
\subsection{Experimental Results}
We design the comparative experiments, ablation experiments, and analyze the parameter sensitivity in this section.
\subsubsection{Overall Evaluation}
Fig. \ref{overall_result} shows the performance of different methods on three datasets. We select the results when $K=4$ for our method. For the analysis of parameters, please refer to Section \ref{Parameteranalysissec}. We find that:
SIDDA outperforms all the state-of-the-art baseline methods consistently and significantly on $hits@N$ and $map@N$. The results show that our method can predict the next activated user successfully.
\subsubsection{Ablation Study}
\label{albationstudysec}
\label{Parameteranalysissec}
\paragraph{\textbf{The effect of the number of disentangled factors.}} We study how the number of disentangled factors $K$ can affect the performance of our method. As shown in Fig.\ref{figdifferentk}, our method has the worst performance when $K=1$. Because when $K=1$, our model is a GRU model with an attention mechanism, the latent information is not disentangled thoroughly. The performance on Douban and Memetracker dataset increases with $K$ getting larger, and start to decrease when $K>4$. While performance on the Twitter dataset is much better when $K>4$, showing an increasing trend. This is because Twitter has more diverse topics and may have more factors to disentangle.
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{pic/diffk.pdf}
\caption{Performance with different numbers of disentangled factors $K$.}
\label{figdifferentk}
\end{figure}
\paragraph{\textbf{The effect of the number of model dimension.}} We also study how the dimension of node representations $D$ can affect the performance. We evaluate our method when $K=4$ and $D\in\{16,32,64,128\}$. As illustrated in Fig. \ref{figdifferentd}, on Twitter and Memetracker the performance doesn't converge until $D=128$, possibly because they have larger training sets. Douban dataset converges when $D=64$ and shows a decreasing trend when $D$ gets larger.
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{pic/diffd.pdf}
\caption{Performance with different dimensions of node representations $D$.}
\label{figdifferentd}
\end{figure}
\section{Conclusion}
In this paper, we propose a novel diffusion prediction model, SIDDA, which can extract latent factors by learning multiple hidden states at each timestamp. Specifically, we propose sequential attention to capture sequential relations of information cascades. Then we adopt a disentangled attention to learn multiple hidden states representing different latent factors. Additionally, we employ Gumbel-Softmax to further enhance the disentangled effect. The experiments on three real-world datasets show the superiority of our method when compared with state-of-the-art diffusion prediction models.
| 7ddd4c15599b9c483efa7469c68d118f08c84704 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\uppercase{Introduction}}
\label{sec:introduction}
\begin{figure*}[t!]
\begin{subfigure}{\linewidth}
\centering
\includegraphics[width=\linewidth]{img/tsne.pdf}
\caption{t-SNE visualizations. Upper-Center: Original latent space. Bottom-Left: Latent space according to our approach. Bottom-Right: Latent space according to standard approach\protect\footnotemark. Both bottom spaces represent the classes of the generated images given a latent code. Therefore, they should be as similar as possible to the original latent space.}
\label{fig:tsne_cifar}
\end{subfigure}\\[2ex]
\begin{subfigure}{\linewidth}
\centering
\includegraphics[width=\linewidth]{img/tsne_images.pdf}
\caption{Random sets of generated samples trained on CIFAR10. The solid frames contain real images and the dashed the generated.}
\label{fig:tsne_cifar2}
\end{subfigure}
\caption{Given a latent space, our approach exploits the structure using its features to condition the generative model. In this way, our system eventually can produce samples on demand. The code color is consistent within the whole figure.}
\label{fig:teaser}
\end{figure*}
\noindent Generative Adversarial Networks (GANs) \cite{goodfellow2014generative} are one of the most prominent unsupervised generative models. Their framework involves training a generator and discriminator model in an adversarial game, such that the generator learns to produce samples from the data distribution. Training GANs is challenging because they require to deal with a minimax loss that needs to find a Nash equilibrium of a non-convex function in a high-dimensional parameter space. This scenario may lead to a lack of control during the training phase, exhibiting non-desired side effects such as instability, mode collapse, among others. As a result, many techniques have been proposed to improve the stability training of GANs \cite{salimans2016improved,gulrajani2017improved,miyato2018spectral,durall2019stabilizing}.
\let\thefootnote\relax\footnotetext{Accepted in VISAPP 2021}
Conditioning has risen as one of the key technique in this vein \cite{mirza2014conditional,chen2016infogan,isola2017image,choi2018stargan}, whereby the whole model has granted access to labelled data. In principle, providing supervised information to the discriminator encourages it to behave more stable at training since it is easier to learn a conditional model for each class than for a joint distribution. However, conditioning comes with a price, the necessity for annotated data. The scarcity of labelled data is a major challenge in many deep learning applications which usually suffer from high data acquisition costs.
Representation learning techniques enable models to discover underlying semantics-rich features in data and disentangle hidden factors of variation. These powerful representations can be independent of the downstream task, leaving the need of labels in the background. In fact, there are fully unsupervised representation learning methods \cite{misra2016shuffle,gidaris2018unsupervised,rao2019continual,milbich2020unsupervised} that automatically extract expressive feature representations from data without any manually labelled annotation. Due to this intrinsic capability, representation learning based on deep neural networks have been becoming a widely used technique to empower other tasks \cite{caron2018deep,oord2018representation,chen2020simple}.
Motivated by the aforementioned challenges, our goal is to show that it is possible to recover the benefits of conditioning by exploiting the advantages that representation learning can offer (see Fig. \ref{fig:teaser}). In particular, we introduce a model that is conditioned on the latent space structure. As a result, our proposed method can generate samples on demand, without access to labeled data at the GAN level. To ensure the correct behaviour, a customized loss is added to the model. Our contributions are as follows.
\begin{itemize}
\item We propose a novel generative adversarial network
conditioned on features from a latent space representation.
\item We introduce a simple yet effective new loss function which incorporates the structure of the latent space.
\item Our experimental results show a neat control on the generated samples. We test the approach on MNIST, CIFAR10 and CelebA datasets.
\end{itemize}
\footnotetext{Standard approach refers to replace the encoded labels with latent code.}
\section{\uppercase{Related Work}}
\subsection{Conditional Generative Adversarial Networks}
Generative image modelling has recently advanced dramatically. State-of-the-art methods are GAN-based models \cite{brock2018large,karras2019style,karras2020analyzing} which are capable of generating high-resolution, diverse samples from complex datasets. However, GANs are extremely sensitive to nearly every aspect of its set-up, from loss function to model architecture. Due to optimization issues and hyper-parameter sensitivity, GANs suffer from tedious instabilities during training.
Conditional GANs have witnessed outstanding progress, rising as one of the key technique to improve stability training and to remove mode collapse phenomena. As a consequence, they have become one of the most widely used approaches for generative modelling of complex datasets such as ImageNet. CGAN \cite{mirza2014conditional} was the first work to introduce conditions on GANs, shortly followed by a flurry of works ever since. There have been many different forms of conditional image generation, including class-based \cite{mirza2014conditional,odena2017conditional,brock2018large}
, image-based \cite{isola2017image,huang2018multimodal,mao2019mode}
, mask- and bounding box-based \cite{hinz2019generating,park2019semantic,durall2020local}, as well as text-based \cite{reed2016generative,xu2018attngan,hong2018inferring}. This intensive research has led to impressive development of a huge variety of techniques, paving the road towards the challenging task of generating more complex scenes.
\subsection{Unsupervised Representation Learning}
In recent years, many unsupervised representation learning methods have been introduced \cite{misra2016shuffle,gidaris2018unsupervised,rao2019continual,milbich2020unsupervised}. The main idea of these methods is to explore easily accessible information, such as temporal or spatial neighbourhood, to design a surrogate supervisory signal to empower the feature learning. Although many traditional approaches such as random projection \cite{li2006very}, manifold learning \cite{hinton2003stochastic} and auto-encoder \cite{vincent2010stacked} have significantly improved feature representations, many of them often suffer either from being computationally too costly to scale up to large or high-dimensional datasets, or from failing to capture complex class structures mostly due to its underlying data assumption.
On the other hand, a number of recent unsupervised representation learning approaches rely on new self-supervised techniques. These approaches formulate the problem as an annotation free pretext task; they have achieved remarkable results \cite{doersch2015unsupervised,oord2018representation,chen2020simple} and even on GAN-based models as well \cite{chen2019self}. Self-supervision generally involves learning from tasks designed to resemble supervised learning in some way, where labels can be created automatically from the data itself without manual intervention.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.9\linewidth]{img/main.pdf}
\caption{Overview of the processing pipeline of our approach. It contains two main blocks, a feature extraction model and a generative model, more specifically a GAN. This structure allows the generator to incorporate a condition based on the latent space, so that eventually the system can produce images on demand using the latent representation. }
\label{fig:main}
\end{figure*}
\section{\uppercase{Method}}
In this section we describe our approach in detail. First, we present our representation learning set-up together with its sampling algorithm. Then, we introduce a new loss function capable of exploiting the structural properties from the latent space. Finally, we have a look at the adversarial framework for training a model in an end-to-end fashion. Fig. \ref{fig:main} gives an overview of the main pipeline and its components.
\subsection{Representation Learning}
The goal of representation learning or feature learning is to find an appropriate representation of data in order to perform a machine learning task.\\
\noindent \textbf{Generating Latent Space.}
The latent space must contain all the important information needed to represent reliably the original data points in a simplified and compressed space. Similar to \cite{aspiras2019active}, in our work we also try to exploit the latent space. In particular, we rely on existing topologies that can capture a high level of abstraction. Hence, we mainly focus on integrating these data descriptors and on evaluating their usability and impact. For this reason, We count on several set-ups where we can sample informative features of different qualities, i.e. level of clustering in the latent spaces.
Our feature extractor block $E$ is a convolutional-based model for classification tasks. Inspired by \cite{caron2018deep}, to extract the features we do not use the classifier output logits, but the feature maps from an intermediate convolutional layer. We refer to these hidden spaces as latent space.\\
\noindent \textbf{Sampling from Latent Space.}
Assuming that the feature extractor is able to produce a structured latent space, e.g. semi-clustered features, we can start sampling observations that will be fed to our GAN afterwards. The procedure to create sampling batches is described in Algorithm \ref{alg:sampling}.
\begin{algorithm}
\caption{Creating batches for training the GAN.}
\label{alg:sampling}
\begin{algorithmic}[1]
\STATE Sample a batch of images $\mathrm{\mathbf{x}}$
\STATE Extract features from it $\mathbf{f} = E(\mathbf{x})$
\STATE Compute distance between all features $\mathbf{D} = ||\mathbf{f}||_{1}$
\FOR {$x_i$ in $\mathrm{\mathbf{x}}$}
\STATE Select $x_i$ and sort the rest according to their distance $d(x_i)$
\STATE Select nearest neighbour from $x_i$, i.e. $d_\mathrm{min}(x_i)$
\STATE Select the farthest neighbour from $x_i$, i.e. $d_\mathrm{max}(x_i)$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\subsection{Loss Function}
\noindent \textbf{Minimax Loss.}
A GAN architecture is comprised of two parts, a discriminator $D$ and a generator $G$. While the discriminator trains directly on real and generated images, the generator trains via the discriminator model. They should therefore use loss functions that reflect the distance between the distribution of the data generated $p_{\mathrm{z}}$ and the distribution of the real data $p_{\mathrm{data}}$. Minimax loss is by default the candidate to carry on with this task and it is defined as
\begin{align}
\begin{split}
\min_{G} \max_{D} \mathcal{L}(D,G) =\, \mathbb{E}_{\mathrm{\mathbf{x}} \sim p_{\mathrm{data}}} \left[ \log \left(D(x)\right) \right] + \cr \mathbb{E}_{z \sim p_{z}}[\log(1-D(G(z)))].
\end{split}
\label{formula:main}
\end{align}
\vspace{1px}
\noindent \textbf{Triple Coupled Loss.}
In the vanilla minimax loss the discriminator expects batches of individual images. This means that there is a unique mapping between input image and output, where each input is evaluated and then classified as real or fake. Despite being a functional loss term, if we hold to that closed formulation, we cannot leverage alternatives such as conditional features or combinatorial inputs i.e. input is not any longer only a single image but a few of them.
We introduce a loss function coined triple coupled loss that incorporates combinatorial inputs acting as a semi-conditional mechanism. The approach lies on the idea of exploiting similitudes and differences between images. In fact, similar approaches have been already successfully implemented in other works \cite{chongxuan2017triple,sanchez2018triple,ho2020learning}. In our implementation, the new discriminator takes couples of images as input and classify them as true or false. Unlike minimax case, now we have two degrees of freedom (two inputs) to take advantage of. Therefore, we produce different scenarios to further enhance the capabilities of our discriminator, so that it can also be conditioned in an indirect manner by the latent representation space. We can distinguish three different coupled case scenarios and their corresponding losses
\begin{align}
\begin{split}
x_{\mathrm{adv}} = [x_{\mathrm{real}\_\mathrm{correct}},x_{\mathrm{fake}}] \,\longrightarrow\, \mathcal{L}_{\mathrm{adv}}\\
\mathcal{L}_{\mathrm{adv}} = \mathbb{E}_{\mathrm{\mathbf{x}} \sim ( p_{\mathrm{data}} \,\cup\, p_{\mathrm{z}} ) } \left[ \log \left(1-D(x_{\mathrm{adv}})\right) \right]
\end{split}
\end{align}
\begin{align}
\begin{split}
x_{\mathrm{same}} = [x_{\mathrm{real}\_\mathrm{correct}},x_{\mathrm{real}\_\mathrm{correct}}] \,\longrightarrow\, \mathcal{L}_{\mathrm{same}}\\
\mathcal{L}_{\mathrm{same}} = \mathbb{E}_{\mathrm{\mathbf{x}} \sim p_{\mathrm{data}}} \left[ \log \left(D(x_{\mathrm{same}})\right) \right]
\end{split}
\end{align}
\begin{align}
\begin{split}
x_{\mathrm{diff}} = [x_{\mathrm{real}\_\mathrm{correct}},x_{\mathrm{real}\_\mathrm{wrong}}] \,\longrightarrow\, \mathcal{L}_{\mathrm{diff}}\\
\mathcal{L}_{\mathrm{diff}} = \mathbb{E}_{\mathrm{\mathbf{x}} \sim p_{\mathrm{data}}} \left[ \log \left(1-D(x_{\mathrm{diff}})\right) \right].
\end{split}
\end{align}
\vspace{1px}
We first have $x_{\mathrm{adv}}$ case which is the combination of one generated image ($x_{\mathrm{fake}}$) and one real that belongs to the target class ($x_{\mathrm{real}\_\mathrm{correct}}$). Then, we have $x_{\mathrm{same}}$ with two different "real correct" samples. Finally, the last case is $x_{\mathrm{diff}}$ which combines one "real correct" and one "real wrong". The latter term is a real image from a different class, i.e. not target class ($x_{\mathrm{real}\_\mathrm{wrong}}$).
In order to incorporate the triple coupled loss, we need to reformulated the Formula \ref{formula:main} adding the aforementioned three case scenarios. As a result, the new objective loss is rewritten as follows
\begin{align}
\begin{split}
\min_{G} \max_{D} \mathcal{L}(D,G) =\,& \lambda_{\mathrm{a}} \mathcal{L}_{\mathrm{adv}} \,+\, \lambda_{\mathrm{s}} \mathcal{L}_{\mathrm{same}} \,+\, \lambda_{\mathrm{d}} \mathcal{L}_{\mathrm{diff}}
\end{split}
\end{align}
\noindent where $\lambda$s are the weighting coefficients.
\subsection{Training on Conditioned Latent Feature Spaces}
Our approach is divided into two distinguishable elements, the feature extractor $E$ and the generative model. With the integration of these two components into an embedded system, our model can produce samples on demand without label information.\\
\noindent \textbf{Dynamics of Training.}
Given an input batch $\mathrm{\mathbf{x}}$, the feature extractor produces the latent code $\mathrm{\mathbf{f}}$. Then, we generate a vector of random noise $\mathrm{\mathbf{z}}$ (e.g. Gaussian) and we attach to it the $\mathrm{\mathbf{f}}$, creating in this way the input for our generator $\mathrm{\mathbf{n}}$ (see Fig. \ref{fig:main}).
The expected behaviour from our generator should be similar to CGAN, where the generator needs to learn a twofold task. On the one hand, it has to learn to generate realistic images by approximating the real data distribution as much as possible. On the other hand, these synthetic images need to be conditioned consistently on $\mathrm{\mathbf{f}}$, so that later can be controlled. For example, when two similar\footnote[2]{Similarity is measured by $l_1$ distance as described in Algorithm \ref{alg:sampling}.} latent codes are fed into the model, this should produce two similar output images belonging to the same class.
\begin{figure*}[t!]
\centering
\includegraphics[width=\linewidth]{img/scale.pdf}
\caption{Random generated samples of 32x32, 64x64 and 128x128 resolutions.}
\label{fig:scale}
\end{figure*}
The discriminator, however, has a remarkable difference with CGAN when it comes to training. While CGAN employs latent codes to condition directly the outcome results, our method uses a semi-conditional mechanism through the coupled inputs. As it is explained in the upper section, the discriminator employs the triplet coupled loss which enforces to respect the latent space structure and binding in this way the output with the conditional information. Algorithm \ref{alg:training} describes the training scheme.
\begin{algorithm}
\caption{Training GAN model.}
\label{alg:training}
\begin{algorithmic}[1]
\STATE Require: $n_{\mathrm{iter}}$, the number of iterations. $n$, the number of iterations of the generator per discriminator iteration. $\lambda$'s, the weighting coefficients. $\theta_{\mathrm{gen}}$, generator's parameters. $\theta_{\mathrm{disc}}$, discriminator's parameters.
\FOR {$i < n_{\mathrm{iter}}$}
\STATE Sample batch using Algorithm 1
\STATE \# Train generator $G$
\STATE $\mathcal{L}_{\mathrm{gen}} = \mathcal{L}_{\mathrm{adv}}$
\STATE $\theta_{\mathrm{gen}} \leftarrow \theta_{\mathrm{gen}} + \nabla \mathcal{L}_{\mathrm{gen}}$
\IF {$mod(i,n) = 0$}
\STATE \# Train discriminator $D$
\STATE $\mathcal{L}_{\mathrm{disc}} = \lambda_{\mathrm{a}} \mathcal{L}_{\mathrm{adv}} \,+\, \lambda_{\mathrm{s}} \mathcal{L}_{\mathrm{same}} \,+\, \lambda_{\mathrm{d}} \mathcal{L}_{\mathrm{diff}}$
\STATE $\theta_{\mathrm{disc}} \leftarrow \theta_{\mathrm{disc}} + \nabla \mathcal{L}_{\mathrm{disc}}$
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
\section{\uppercase{Experiments}}
In this section, we show results for a series of experiments evaluating the effectiveness of our approach. We first give a detailed introduction of the experimental set-up. Then, we analyse the response of our model under different scenarios and we investigate the role that plays the structure of the latent space and its robustness. Finally, we check the impact of our customized loss function though an ablation study.
\subsection{Experimental Set-up}
We conduct a set of experiments on MNIST \cite{lecun1998gradient}, CIFAR10 and CelebA \cite{liu2015faceattributes} datasets. For each one, we use an individual classifier to ensure certain structural properties on our latent space. Next, we extract feature from one intermediate layer, and feed them into our generative model.\\
\noindent \textbf{MNIST.}
The experiments carried on MNIST are fully\textbf{ unsupervised} since we do not require any label information. We choose to deploy an untrained AlexNet model \cite{krizhevsky2012imagenet} as feature extractor. As shown in \cite{caron2018deep}, AlexNet offers an out-of-the-box clustered space at certain intermediate layer without any need of training. Hence, we extract there the features and no extra processing step is involved.\\
\noindent \textbf{CIFAR10.}
Despite the fact that CIFAR10 is fairly close to MNIST in terms of amount of samples and classes (10 in both cases), it is indeed a much more complex dataset. As a result, in this case we need to train a feature extractor to achieve a structured latent space. Inspired by the unsupervised representation learning method \cite{gidaris2018unsupervised}, we build a classifier which reaches similar accuracy.\\
\noindent \textbf{CelebA.}
Different from the previous datasets, CelebA contains only "one class" of images. In particular, this datatset is an extensive collection of faces. However, each sample can potentially contain up to 20 different attributes. So, in our experiments we build different scenarios by splitting the dataset into different classes according to their attributes, e.g. \textit{man} and \textit{woman}. Moreover, we also test our approach on different resolutions, since the size of the images of CelebA is larger. Similarly to CIFAR10 case, we need to train again a feature extractor.
\begin{figure*}[t!]
\begin{subfigure}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{img/cifar_evol.pdf}
\caption{CIFAR10.}
\label{fig:cifar_curve}
\end{subfigure}
\begin{subfigure}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{img/celeba64_evol.pdf}
\caption{CelebA.}
\label{fig:cifar_curve}
\end{subfigure}
\caption{FID and accuracy learning curves on CIFAR10 and CelebA.}
\label{fig:celeba_evol}
\end{figure*}
\begin{table}[t!]
\centering
\caption{Validation results in MNIST, CIFAR10 and CelebA.}
\begin{tabular}{cccc}
\hline
\textbf{MNIST} & IS & FID & Accuracy \\
\hline
baseline & 9.63 & - & - \\
ours & 9.78 & - & 72\% \\
\hline
\textbf{CIFAR10} & IS & FID & Accuracy \\
\hline
baseline & 7.2 & 28.76 & - \\
ours & 7.0 & 29.71 & 68\% \\
\hline
\textbf{CelebA} & IS & FID & Accuracy \\
\hline
baseline & 2.3 & 18.56 & - \\
ours (32) & 2.5 & 11.10 & 90\% \\
ours (64) & 2.7 & 13.69 & 94\% \\
ours (128) & 2.65 & 37.59 & 94\% \\
\hline
\end{tabular}
\label{table:metric}
\end{table}
\subsection{Evaluation Results}
We compare the baseline model based on
Spectral Normalization for Generative Adversarial Networks (SNGAN) \cite{miyato2018spectral} to our approach that incorporates the latent code and the coupled input on top of it. The rest of the topology remains unchanged.\footnote[3]{In CelebA, we add one and two layers into the model to be able to produce samples with resolution of 64x64 and 128x128, respectively.} We do not use CGAN architecture since our framework is not conditioned on labels. Therefore, we take an unsupervised model as a baseline. In particular, we choose SNGAN because it is a simple yet stable model that allows to control the changes applied on the system. Besides, it generates appealing results having a competitive metric scores.
To evaluate generated samples, we report standard qualitative scores on the Frechet Inception Distance (FID) and the Inception Score (IS) metrics. Furthermore, we provide the accuracy scores that eventually quantize the success of the system. We compute this score using a classifier trained on the real data that guarantees that the metric correctly assesses the percentage of generated samples that coincide with the class of the latent code. For instance, if the latent code belongs to a \textit{cat}, the generator should produce a \textit{cat}.
Table \ref{table:metric} compares scores for each metric. We observe how our model performs fairly similar to the baseline independently of the scenario. Only when we ask for a 128x128 output resolution, the FID score increases substantially. We hypothesize that this break happens due to a model architecture issue since the baseline is initially designed for 32x32 images (see Fig. \ref{fig:scale}). Fig. \ref{fig:celeba_evol} plots FID and accuracy training curves on CIFAR10 and CelebA datasets, and confirms that our approach exhibits a strong correlation between the both metrics. A better FID score (low value) means always a higher accuracy score.
It is important to notice that baseline models do not have accuracy score since they cannot choose the class of the output. On the other hand, regarding our approach, the accuracy for both MNIST and CIFAR is around 70\%, and more than 90\% for CelebA. This gap is directly related with the quality of the latent space. In other words, the more clustered the latent space is, the higher accuracy our model can have. In this case, CelebA is evaluated in a scenario with only two classes \textit{man} and \textit{women}, as a consequence the latent spaces is simpler. As a rule of thumb, an increase of classes will often lead to a more tangled latent space making the problem harder. The main reason for that are those samples located on the borders. We refer to this phenomenon as border effect and it is shown in Fig. \ref{fig:tsne_celeba2}. As it is expected, we observe how the samples that lie between the two blobs have usually a higher failure rate (colored in red).
\begin{figure*}[t!]
\centering
\includegraphics[width=\linewidth]{img/tsne_celeba2.pdf}
\caption{Visualization of border effect on CelebA for the classes \textit{man} and \textit{woman}. Upper-left: t-SNE of extracted features regarding to their classes. Upper-right: t-SNE of extracted features regarding to their capacity of conditioning the output i.e. accuracy. Bottom: Random samples from different latent codes, where the solid frame belong to the real images and the dashed frames the generated. The code color is consistent within the whole figure.}
\label{fig:tsne_celeba2}
\end{figure*}
\subsection{Impact of the Latent Space Structure}
The structure of latent space plays an important role and has a direct effect in the accuracy performance. This is mainly due to the nature of the triple coupled loss. This term relies on having at least a semi-clustered feature space to sample from. Hence, those latent spaces with almost no structure will build many false couples in training time and resulting in bad performance. Notice that the generator model does not use label information directly, but through the extracted features.
\begin{table*}[t!]
\centering
\caption{Statistics of the latent space's structure for different scenarios. }
\begin{tabular}{cccccc}
\hline
\textbf{MNIST} & Classes & Accuracy & $\;1^{\mathrm{st}}$ neighbour & $\;2^{\mathrm{nd}}$ neighbour& $\;5^{\mathrm{th}}$ neighbour\\
\hline
baseline & 10 & - & 10\% & 10\% & 10\%\\
ours & 10 & 72\% & 89\% & 84\% & 78\%\\
\hline
\textbf{CIFAR10} & Classes & Accuracy & $\;1^{\mathrm{st}}$ neighbour & $\;2^{\mathrm{nd}}$ neighbour& $\;5^{\mathrm{th}}$ neighbour\\
\hline
baseline & 10 & - & 10\% & 10\% & 10\%\\
ours & 10 & 68\% & 70\% & 68\% & 65\% \\
\hline
\textbf{CelebA} & Classes & Accuracy & $\;1^{\mathrm{st}}$ neighbour & $\;2^{\mathrm{nd}}$ neighbour& $\;5^{\mathrm{th}}$ neighbour\\
\hline
baseline & 2 & - & 50\% & 50\% & 50\%\\
ours (32) & 2 & 90\% & 88\% & 88\% & 86\%\\
ours (64) & 2 & 94\% & 95\% & 94\% & 93\%\\
\hline
baseline & 5 & - & 20\% & 20\% & 20\%\\
ours (32) & 5 & 80\% & 90\% & 89\% & 87\%\\
ours (64) & 5 & 78\% & 96\% & 95\% & 92\%\\
\hline
\end{tabular}
\label{table:statistic}
\end{table*}
\begin{figure*}[t!]
\begin{subfigure}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{img/mnist_normal.pdf}
\caption{Semi-clustered latent space.}
\label{fig:mnist_normal}
\end{subfigure}
\begin{subfigure}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{img/mnist_cluster.pdf}
\caption{Full-clustered latent space.}
\label{fig:mnist_cluster}
\end{subfigure}
\caption{t-SNE visualizations from two different latent spaces on MNIST. First row displays the classes and second row the accuracy from our approach.}
\label{fig:mnist}
\end{figure*}
We run the evaluations on MNIST, CIFAR10 and CelebA datasets as in the previous section. However, in CelebA's case there are now two different set-ups. One based on gender (\textit{man} and \textit{woman}), and a second one based on hair (\textit{blond}, \textit{black}, \textit{brown}, \textit{gray} and \textit{bold}). In order to study the impact of the latent space structure, we need to determine how clustered our space is. Therefore, we compute a set of statistics (see Table \ref{table:statistic}) that are useful to estimate the initial conditions of the latent structure, and consequently find out the boundaries that our system might not overcome. For example, our model on CIFAR10 reports 70\% on $\;1^{\mathrm{st}}$ neighbour. This value indicates that if we take one random sample from our latent space, 70\% of the time its nearest neighbour will belong to the same class. Empirically, we observe the causal effect that the structure of latent space has on the accuracy results. The more clustered, i.e. higher neighbours scores, the better the accuracy. In other words, neighbourhood information helps to understand the upper-bounds fixed by the latent space.
Fig. \ref{fig:mnist} compares two identical set-ups with different latent spaces. On the one hand, we have the semi-clustered space produced by an untrained AlexNet. This scenario achieves good accuracy scores despite the border effect. On the other hand, we have an extreme case with a fully-clustered space. As expected, all the scores are dramatically improved at the cost of having a perfect space.
\subsection{Robustness of the Latent Space}
In this section, we analyse how our approach behaves when we introduce noisy labels, and we compare it to CGAN performance. This analysis allows us to quantify how robust our system is.
We start the experiments having a set-up free of noise.\footnote[4]{Notice that for this experiment we take a feature extractor and we train it from scratch each time that we change the percentage of noise.} Then, we gradually increase the amount of noise by introducing noisy labels. Fig. \ref{fig:cgan} shows the accuracy curves evolution for both cases. We observe how CGAN has almost a perfect lineal relationship between noise and accuracy. Every time that noise increases, the accuracy decreases in a similar proportion. This demonstrate the necessity of CGAN of labels to produce the desired output and the incapacity to deal with noise. Therefore, its robustness against noise very is limited. On the other hand, our approach shows a more robust behaviour. In this case, there is not lineal relationship, and the system is able to maintain the accuracy score independently of the level of noise. Only a notable decrease happens when the percentage of noisy labels surpasses the barrier of 90\%.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{img/CGAN.pdf}
\caption{Robustness evaluation on MNIST using accuracy curves.}
\label{fig:cgan}
\end{figure}
\begin{figure*}
\begin{subfigure}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{img/training.pdf}
\caption{Losses evolution.}
\label{fig:cifar_loss}
\end{subfigure}
\begin{subfigure}{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{img/cifar_fid.pdf}
\caption{FID evolution.}
\label{fig:cifar_comp}
\end{subfigure}
\caption{Comparison between the baseline and our approach on CIFAR10.}
\label{fig:cifar}
\end{figure*}
\begin{table*}
\centering
\caption{Quantitative results of the ablation study on CIFAR10.}
\begin{tabular}{ccccccc}
\hline
& $\mathcal{L}_{\mathrm{minimax}}$ & $\mathcal{L}_{\mathrm{adv}}$ & $\mathcal{L}_{\mathrm{same}}$ & $\mathcal{L}_{\mathrm{diff}}$ & Convergence & Accuracy\\
\hline
baseline & \checkmark & & & & \checkmark & 8\% \\
prototype A & & \checkmark & & & & - \\
prototype B & & \checkmark & \checkmark & & \checkmark & 58\%\\
prototype C & & \checkmark & & \checkmark & & -\\
ours & & \checkmark & \checkmark & \checkmark & \checkmark &68\%\\
\hline
\end{tabular}
\label{table:ablation}
\end{table*}
\subsection{Analysis of Triple Coupled Loss}
The triple coupled loss is designed to exploit the structure of the latent space, so that the generative model learns how to produce samples on demand, i.e. based on the extracted features. Fig. \ref{fig:cifar} shows the itemized losses and FID learning curves for minimax loss (baseline) and for triple coupled loss (ours). We can confirm a similar behaviour between our model and the baseline, having as a side-effect an increase of convergence training time. In exchange for this delay, the proposed system has control over the outputs through the extracted features.
\section{\uppercase{Ablation Study}}
In this section, we quantitatively evaluate the impact of removing or replacing parts of the triple coupled loss. We do not only illustrate the benefits of the proposed loss compared to minimax loss, but also present a detailed evaluation of our approach. Table \ref{table:ablation} presents how the loss function behaves when we modify its components. First, we check whether the sub-optimal loss leads the system to convergence, i.e. it is able to generate realistic images. And second, for those functions that have the capacity of generating, we check their accuracy score. Notice that ideally we would achieve 100$\%$ which means producing the desired output all the time.
Based on the empirical results from the previous table, we can see the importance of each term in the loss function. In particular, we observe how $\mathcal{L}_{\mathrm{same}}$ is essential to achieve convergence, and how the combination of all three terms brings the best result.
\section{\uppercase{Conclusions}}
Motivated by the desire to condition GANs without using label information, in this work, we propose an unsupervised framework that exploits the latent space structure to produce samples on demand. In order to be able to incorporate the features from the given space, we introduce a new loss function. Our experimental results show the effectiveness of the approach on different scenarios and its robustness against noisy labels.
We believe the line of this work opens new avenues for feature research, trying to combined different unsupervised set-ups with GANs. We hope this approach can pave the way towards high quality, fully unsupervised, generative models.
\bibliographystyle{apalike}
{\small
| 424770a0b0ce57b50d5c21bbfbcb18ea94a29c50 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The past decade has witnessed a surge of interest in exploiting concepts from topology as a powerful tool to engineer material behavior~\cite{kosterlitz_ordering_1973,haldane_model_1988,sanchez_spontaneous_2012,coulais_static_2017}.
In particular, an important realization was that collections of a material's normal modes are not merely independent; instead, phase relationships between modes -- quantified, for example, by Berry curvature -- can have physical implications.
This realization provides a natural framework for linking topological ideas to material systems and distinguishing topologically distinct phases of matter~\cite{laughlin_quantized_1981,thouless_quantized_1982,prodan_topological_2009}.
These connections have revealed unique phenomena governed by the topology of the band structure of the material in question.
Crucially, the resulting exotic effects, such as the appearance of chiral edge modes, are protected from disorder by their topological origin, meaning that chiral edge modes are insensitive to geometric features of the boundary or inhomogeneity in the material's bulk~\cite{hasan_colloquium:_2010,ryu_topological_2010}.
The design of mechanical metamaterials -- engineered structures assembled from composite components -- has extended topological protection to elastic structures, enabling us to control and direct phononic excitations robustly without requiring extreme precision in the metamaterial's construction~\cite{prodan_topological_2009,kane_topological_2013,susstrunk_observation_2015}.
At the same time, metamaterials offer an ideal arena for fundamental insights into topological phases.
In metamaterials, we observe topological excitations directly at a macroscopic scale, rather than through indirect transport measurements.
Tabletop metamaterials furthermore support rapid prototyping and access to the material's individual components.
This class of platforms has allowed the direct identification of topological behavior in crystalline systems as diverse as coupled pendula~\cite{susstrunk_observation_2015}, photonic waveguides~\cite{rechtsman_photonic_2013}, isostatic frames~\cite{paulose_selective_2015}, chains of gears~\cite{meeussen_geared_2016}, and lattices of acoustic resonators ~\cite{souslov_topological_2017}.
One other emerging platform of metamaterial systems is composed of interacting gyroscopes suspended from a plate~\cite{nash_topological_2015,mitchell_realization_2018,mitchell_tunable_2018,wang_topological_2015,brun_vortex-type_2012,carta_dispersion_2014,carta_flexural_2019,susstrunk_classification_2016}.
Because networks of gyroscopes break time reversal symmetry in a geometry-dependent manner, these gyroscopic tabletop metamaterials have provided a simple platform to observe topological phase transitions~\cite{mitchell_realization_2018}
and switch between multiple topological phases with multiple gaps~\cite{mitchell_tunable_2018}.
This ease of access in turn allows for discovery.
For example, constructing amorphous materials using the gyroscopic platform provided a route to discovering that topology arises not just in crystalline materials, but even in amorphous networks~\cite{mitchell_amorphous_2018}.
This last example suggests that topological phases should arise in glassy systems composed of randomly arranged inclusions, dopants, or vacancies -- a prediction that has since been identified in models of quantum matter~\cite{agarwala_topological_2017,sahlberg_topological_2020,marsal_topological_2020,poyhonen_amorphous_2018,huang_theory_2018,agarwala_higher-order_2020}.
Some instances of these topological amorphous phases have been suggested to be distinct phases of matter~\cite{bourne_non-commutative_2018,sahlberg_topological_2020} and have been placed into a topological classification scheme for mechanical systems~\cite{barlas_topological_2018}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig_symplectic6.png}
\caption{
\textbf{Gyroscopic metamaterials provide a platform for engineering exotic behavior.}
\textit{(a)} Each gyroscope is suspended from a plate and spins rapidly with a speed $\omega_0$.
\textit{(b)}
Coupled gyroscopes experience an interaction frequency, $\Omega_k$, and precess under the influence of restoring forces such as gravity.
\textit{(c)} In an experimental realization, externally-powered motors spin small masses with magnets embedded at their cores.
All gyroscopes spin in the same direction.
A weak spring suspends the gyroscope from a plate, enabling low-loss precession.
A coil beneath each gyroscope locally modulates the effective gravitational torque on each gyroscope, allowing us to introduce disorder or staggered potentials.
\textit{(d)} Chiral waves skim along the boundary when the system is shaken at a frequency in the resulting bandgap.
Each colored circle represents the measured displacement of a gyroscope in our experimental realization, with the radius proportional to the displacement amplitude and the color denoting the phase (inset shows colorwheel).
\textit{(e)} Protected chiral waves persist even if the spatial structure of the underlying network is amorphous.
Here, a simulated amorphous network of gyroscopes consists of a kagome-like structure above and a Voronoi-like structure below a boundary shaped into the letters `CHERN'.
The upper and lower regions differ in their topological phase, so that an extended wave packet travels along the boundary between the regions.
}
\label{fig:intro_gyro_symplectic}
\end{figure}
These recent discoveries place a spotlight on the importance of local, real-space physics in determining topological phenomena.
We seek to examine this interplay using gyroscopic metameterials as a concrete foundation for our discussion.
Our conclusions, however, extend beyond this particular class of systems.
After revisiting the mathematical foundations of these odd materials, we present new insights into the symmetries underlying these materials' band gaps.
We then adopt a local, mechanical viewpoint on the mechanisms generating topological gaps.
This extends the current understanding of the mesoscopic or quasi-local character of band topology that has emerged based on studies of gyroscopic materials.
Lastly, we present experimental and numerical measurements of phase transitions driven by disorder and the localization properties of gyroscopic materials' band structures.
In contrast to recent studies of critical behavior in amorphous topological insulators~\cite{ivaki_criticality_2020,sahlberg_topological_2020}, we find the scaling behavior of critical exponents consistent with the universality class with Cartan label A.
In addition, a topological Anderson insulator phase diagram shows re-entry into the topological phase with increasing disorder strength, both in experiments and simulations.
\section{An odd interaction and its symplectic dynamics}
Our system is composed of coupled spinning gyroscopes suspended from a plate shown in~\Fig{fig:intro_gyro_symplectic}.
When the spinning speed is large, the high frequency nutation becomes negligible in amplitude.
A gyroscope with a hanging orientation vector $\hat{n}$ subjected to forces acting at $\ell \hat{n}$ exhibits motion in the $xy$ plane governed by the action of torques redirecting the angular momentum axis:
\begin{align} \label{gyrox}
\frac{I \omega_0}{\ell } \partial_t{x}
& = \ell \frac{\partial U}{\partial y} \\
\label{gyroy}
\frac{I \omega_0}{\ell} \partial_t{y}
& = - \ell \frac{\partial U}{\partial x} ,
\end{align}
where $\ell$ is distance from the pivot point to the point where forces act, $x$ and $y$ are the in-plane displacements of the same point, $U$ is the potential energy of the gyroscope from pinning forces (like gravity and the springs connecting the gyroscope to its neighbors), $\omega_0$ is the spin frequency, $I$ is the moment of inertia along the spinning axis, and $x$ and $y$ are the displacements from the equilibrium position.
\Eqs{gyrox} and~\ref{gyroy} reflect the curious behavior of gyroscopes: they respond to an applied force by moving in a direction transverse to that force.
Building a material out of coupled gyroscopes translates this unusual response into new collective material excitations that are profoundly different from those of inertial materials, where we are accustomed to having displacements follow the direction of forces.
In conventional elastic materials, waves emerge from the conversion of potential energy into kinetic energy and back again.
As energy sloshes back and forth between two forms, disturbances propagate through the material.
For interacting gyroscopes, however, inertia is dominated by the rapid spin of each gyroscope, rendering the kinetic energy of the gyroscopes' center of mass motion irrelevant.
Instead, the independent degrees of freedom are the orthogonal gyroscope displacement $x$ and $y$, and the resulting energy oscillation is an oscillation of potential energy between these spatial degrees of freedom because of the transverse response of gyroscopes to applied forces.
These dynamics are analogous to those of Tkachenko waves in vortex crystals~\cite{sonin_vortex_1987}.
The relationship between motion in these two directions forms a symplectic structure.
By defining $q = \sqrt{I \omega_0} x / \ell$ and $p = \sqrt{I \omega_0} y / \ell$, we see that~\Eq{gyrox} and~\Eq{gyroy} are the same as Hamilton's equations:
\begin{equation}
\label{eq_hamilton}
\partial_t{q} = \frac{\partial U}{\partial p}
\hspace{30pt}
\partial_t{p} = - \frac{\partial U}{\partial q},
\end{equation}
which provides symplectic structure to the dynamics.
\subsection{Symplectic structure of cyclotron orbits}
\begin{figure}
\centering
\includegraphics{fig_symplectic_charged_particle_ver2.png}
\caption{
\textbf{A charged particle in a magnetic field provides a simple setting in which to understand symplectic dynamics. }
A particle of positive unit charge moves in a confining potential (blue gradient) and experiences cyclotron orbits due to an out-of-plane magnetic field of strength $B$.}
\label{fig:charged_particle}
\end{figure}
To gain some insight into the practical consequences of these conservative transverse dynamics, it is helpful to examine a simple related system, a charged particle in a magnetic field confined by an external potential (\Fig{fig:charged_particle}).
This simple example highlights subtleties in this category of problems with time-reversal-broken dynamics.
We take the energy to be given by
\begin{equation}
H = \frac{1}{2m} \left( \vec{p} - \frac{q\vec{A}}{c} \right)^2 + \frac{k}{2} \left(x^2 + y^2\right),
\end{equation}
where we adopt the gauge such that $\vec{A} = Bx \hat{y}$ for the vector potential $\vec{A}$ so that $\vec{B} = B\hat{z} = \nabla \times \vec{A}$ and $\vec{p}$ is the canonical momentum.
In order to fully describe the normal modes of the particle's motion, both realspace coordinates $(x,y)$ and the conjugate momenta ($p_x=m\partial_t{x}$, $p_y=m\partial_t{y} + q B x/c$) must be included. \
Hamilton's equations (\Eq{eq_hamilton}) become
\begin{equation}
\partial_t \begin{pmatrix}
x \\ y \\ p_x \\ p_y
\end{pmatrix}
= \begin{pmatrix}
0 & 0 & \frac{1}{m} & 0 \\
-\frac{qB}{mc} & 0 & 0 & \frac{1}{m} \\
-\frac{q^2B^2}{mc^2}-k & 0 & 0 & \frac{qB}{mc} \\
0 & -k & 0 & 0 \\
\end{pmatrix}
\begin{pmatrix}
x \\ y \\ p_x \\ p_y
\end{pmatrix} .
\end{equation}
This is simply a recasting of Hamilton's equations seen before in~\Eq{eq_hamilton}.
Similarly to a quantum mechanical Hamiltonian, the dynamical matrix is a linear operator governing the evolution of the system, but the matrix here is not Hermitian.
Instead, it is symplectic, as we will see.
The four eigenmodes of the dynamical matrix trace out elliptical orbits, taking the form of
\begin{align}
\begin{array}{l l c c c}
e_{\pm} &=\frac{1}{\sqrt{2\omega_{\pm} \left(m \alpha_{\pm}^2+\frac{1}{k}\right)}}
\bigg(\alpha_\pm,
& \frac{-i\omega_{\pm}}{k},
& i m \alpha_\pm \omega_{\pm},
& 1 \bigg)
\end{array}
\end{align}
and their complex conjugates, $\xoverline{e}_+$ and $\xoverline{e}_-$,
where the frequencies of the orbits are $\pm \omega_{+}$ and $\pm \omega_{-}$.
Above,
\begin{equation}
\alpha_{\pm} =
\frac{c}{qBk}\left( k -m \omega_{\pm}^2 \right)
\end{equation}
and the eigenmode frequencies are
\begin{align}
\omega_{\pm} &= \sqrt{\frac{k}{m} + \frac{q^2B^2}{2c^2m^2}\left(1 \pm \sqrt{\frac{4c^2km}{q^2B^2} + 1}\right)}.
\end{align}
These eigenmodes, which correspond physically to cyclotron orbits of the particle precessing in the magnetic field,
are orthogonal under the symplectic form, defined for any two modes $\psi_\alpha$ and $\psi_\beta$ as
\begin{equation} \label{inner_product}
\langle \psi_\alpha | \cdot^\perp | \psi_\beta \rangle
= i \sum_{j \in \{1,2\}}
\xoverline{x_{\alpha,j}} \, {p_{\beta, j}}
- \xoverline{p_{\alpha,j}} \, {x_{\beta,j}},
\end{equation}
where $\vec{x} = (x, y)$ and $\vec{p} = (p_x,p_y)$ and $j$ runs over the two spatial dimensions.
This highlights a difference between parity-breaking systems and more common inertial systems that do not break parity.
In the latter case, there is a direct proportionality between momentum and velocity.
For an eigenmode in an inertial system, the momentum is then simply proportional to the position offset by a phase ($ p_j = -i m_j \omega x_j$).
This simplifies the dynamical matrix considerably to produce eigenvectors that are orthogonal under both the symplectic form and the familiar Euclidean norm defined as $\sum_j m_j x_{\alpha, j} \xoverline{x_{\beta, j}}$.
This feature persists in systems of many coupled inertial objects.
In contrast, rotational coupling between motion in $x$ and $y$ prevent the eigenvectors from being orthogonal under the Euclidean inner product.
However, under the symplectic form, the eigenmodes \textit{are} orthogonal for any value of magnetic field $B$ according to
\begin{equation} \label{eq:innerProductPoisson}
\langle {\psi}_{\alpha} |\cdot^\perp | \psi_{\beta} \rangle \propto \delta_{\alpha\beta} \, \textrm{sign}(\omega_{\alpha}).
\end{equation}
In systems which break parity, the only appropriate inner product is given by~\Eq{inner_product}, and this norm relates the time-averaged energy of a mode, $\langle U \rangle$, to the oscillation frequency, $\omega$.
This relationship is given by
\begin{equation}\label{eq:energy-frequency}
\omega_{\alpha} = \frac{\langle U \rangle}
{ \langle \psi_{\alpha} | \cdot^\perp | \psi_{\alpha} \rangle },
\end{equation}
which holds only when $\langle \psi_{\alpha} | \cdot^\perp | \psi_{\alpha} \rangle$ is the symplectic form of the eigenmode given in~\Eq{inner_product}.
This simple relationship allows one to estimate oscillation frequencies by having an intuitive picture of energies involved in a mechanical deformation.
The consequences of symplecticity extend beyond a redefinition of the inner product: a symmetry between eigenvectors at positive and negative frequencies appears.
Each mode appears twice, with one eigenvector representation at frequency $\omega$ and a second, redundant eigenvector with frequency $-\omega$ and conjugated position and momentum, but each pair represents the same physical circular cyclotron orbit.
This redundancy in the system foreshadows the emergence of ``particle-hole" symmetry in networks of gyroscopes.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig_2gyro3gyro_generic2.png}
\caption{\textbf{Interacting gyroscopes highlight the interplay between clockwise and counter-clockwise polarizations in elliptical eigenmodes.}
Eigenmode displacements are represented by ellipses traced out by the gyro tips, each colored by its phase in the ellipse at a snapshot in time.
\textit{(a)} Two interacting gyroscopes support an in-phase circular motion `ground state' ($\omega=\Omega_p$) and out-of-phase elliptical motion `excited state' ($\omega > \Omega_p$).
\textit{(b)} Three interacting gyroscopes generate two excited states in addition to an in-phase circular motion `ground state' ($\omega=\Omega_p$).
One has elliptical motion with relative phases staggered by $2\pi/3$, while the highest frequency motion is again circular, with relative phases staggered by $-2\pi/3$.
In this figure, we let $\Omega_k=\Omega_p$.
}
\label{fig:2gyro}
\end{figure}
\subsection{Coupled gyroscopes yield coupled polarizations}
Gyroscopic metamaterials similarly break parity, mixing momentum and positions and supporting waves that exchange displacements in orthogonal spatial dimensions.
We consider the motion of the gyroscopes about their hanging rest configuration.
For small displacements $(x, y)$ of the centers of mass, the equation of motion for a gyroscope takes the form
\begin{equation} \label{simulation}
\partial_t
\left(
\begin{array}{c}
x_i\\
y_i
\end{array}
\right)
\approx
\Omega_p \left( \begin{array}{c}
y_i \\
- x_i
\end{array} \right)
+
\frac{\ell^2}{I \omega_0}
\sum_{j=1}^{\textrm{N}}
\left( \begin{array}{c}
-F_{ij, y} \\ F_{ij, x}
\end{array} \right),
\end{equation}
where $\Omega_p$ is the pinning frequency due to gravity, restoring forces from the method of suspension, and other site-specific restoring potentials such as a magnetic field generated by coils placed beneath each gyroscope (see~\Fig{fig:intro_gyro_symplectic}b-e and~\cite{nash_topological_2015}) and the sum is taken over all $N$ neighbors connected to site $i$.
It is useful to write the displacement of the $p^{th}$ gyroscope as $\psi_p \equiv x_p + i y_p$, so that
\begin{equation}\label{eq:psiRpsiL}
\psi_p = \psi^R_p e^{-i\omega t} + \overline{\psi^L_p} e^{i \omega t} .
\end{equation}
If the frequency of oscillation $\omega$ is positive, the first term encodes clockwise motion, while the second encodes counter-clockwise motion, and $\overline{\psi^L_p}$ is the complex conjugate of $\psi^L_p$.
How do these dynamics play out in the simplest scenario of two interacting gyroscopes?
One possible motion is uniform precession at a frequency $\Omega_p$, which does not stretch or compress the bond connecting the two gyros, shown in the lower mode of~\Fig{fig:2gyro}a.
This motion is entirely in-phase.
The other possible motion is out-of-phase, which does stretch and compress the bond and results in a higher frequency.
This out-of-phase motion is elliptical, not circular, as shown in the upper mode of~\Fig{fig:2gyro}a.
Elliptical traces reflect a superposition of both clockwise and counter-clockwise polarizations, foreshadowing that the interaction between polarizations in the dynamical matrix will play a central role in understanding gyroscopic metamaterials.
As in the case of a charged particle in a magnetic field, here each of these modes is represented twice -- with one eigenstate of frequency $\omega$ dominated by $\psi^R$ and one of frequency $-\omega$ dominated by $\psi^L$.
Physically, both $\psi^R$ at positive frequency and $\psi^L$ at negative frequency represent the same clockwise motion according to~\Eq{eq:psiRpsiL}.
By casting the equations of motion in the basis of $(\psi^R, \psi^L)$ as
\begin{multline}\label{eq:psi-eom}
i \partial_t \left(
\begin{array}{c}
\psi^R_p \\
\psi^L_p
\end{array}
\right)
= \Omega_p \left( \begin{array}{c}
\psi^R_p \\
-\psi^L_p
\end{array}
\right)
\\
+ \frac{ \Omega_k}{2}
\sum_{q \in \textrm{N}(p)} \left(
\begin{array}{c}
\psi_{p}^R -\psi_{q}^R
+ e^{i 2 \theta_{pq}}(\psi_p^L - \psi_q^L) \\
-\psi_{p}^L +\psi_{q}^L
- e^{-i 2 \theta_{pq}}(\psi_p^R - \psi_q^R )
\end{array}
\right),
\end{multline}
where $\Omega_k = k \ell^2/ (I \omega_0)$ is the characteristic spring frequency, $\theta_{pq}$ is the bond angle between gyroscopes $p$ and $q$ fixed by network geometry, and the sum over $q$ is over all of $p$'s coupled neighbors,
we see that clockwise and counter-clockwise polarizations are coupled through the geometry of bonds between gyroscopes.
Here, $k$ is the spring constant, $\ell$ is the length of the pendulum from the pivot to the point of applied forcing, $I$ is the moment of inertia along the spinning axis, $\omega_0$ is the angular spinning frequency about the long axis of the gyroscope.
Now that we have gained intuition from the two-gyroscope case, we can formulate the many-gyroscope case as an eigenvalue problem.
\Eq{eq:psi-eom} defines the entries for a system's dynamical matrix, $ \mathbf{D}$, such that
\begin{equation} \label{eq:basicD}
i\partial_t \vec{\psi} = \mathbf{D} \vec{\psi},
\end{equation}
where the components of $\vec{\psi} = (\psi^R, \psi^L)$ encode the displacements of the gyroscopes.
The matrix elements $D_{ij}$ describe the response of gyroscope $j$ to displacements of gyroscope $i$.
\subsection{Symplecticity affects eigenmodes}
The dynamical matrix is not Hermitian.
However, due to the symplectic symmetry of our dynamical matrix, all eigenvalues of $\mathbf{D}$ are nonetheless real.
The symplectic symmetry of our dynamical matrix can be stated as:
\begin{equation}
D = Q D^\dagger Q,
\end{equation}
with
\begin{equation}
Q = \begin{pmatrix}
\mathbb{I}_n & 0 \\
0 & -\mathbb{I}_n
\end{pmatrix}.
\end{equation}
Just as the real-valued frequencies of the charged particle in a magnetic field come in positive and negative pairs, so too the frequencies of gyroscopic motion arise in positive and negative pairs.
These two copies of the system's dynamical response are related by a particle-hole symmetry: each eigenmode has two redundant descriptions with frequencies $\pm \omega$ related by conjugation of the gyroscope positions and momenta.
This is analogous to particle-hole symmetry in electronic systems, wherein the behavior of an electron's excitated state above the Fermi level is tied to the behavior of a electron `hole' left behind below the Fermi level.
In later sections, we find that the symplectic form must be used in the proper definition of Berry curvature and projector operator elements to build real-space descriptions of band topology.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./trs_triangle_nonconservation_TRS_cdonly.png}
\caption{\textbf{Gyroscopic triad provides intuition for the effects of time reversal symmetry breaking.}
\textit{(a)} Time reversal operation, given by $\psi\rightarrow \bar{\psi}$ and $t\rightarrow -t$, transforms the instantaneous displacements (black bars) of one eigenmode into another for the gyroscopic triad.
Here, the instantaneous displacements differ between the two pictured eigenmodes only by $y\rightarrow -y$.
Since eigenmodes $\vec{\psi}_{\omega_a}$ (left) and $\vec{\psi}_{\omega_p}$ (right) have different frequencies, and given that the time reversal operator conjugates these displacements, we see that time reversal symmetry is broken.
\textit{(b)} Simulating the dynamics of the triad reveals strong signatures of chirality from time reversal symmetry breaking.
Because of the phase interference between the different eigenmodes, the displacement generally propagates counter-clockwise between the three gyroscopes.
For special values of $\Omega_k/\Omega_p$, the oscillation is periodically localized to one site at regular intervals, and the site of localized excitation advances in a chiral fashion, here shown for $\Omega_k / \Omega_p = 0.4$.
We note that the excitation is not fully localized at times between the snapshots shown.
}
\label{fig:3gyro_TRS}
\end{figure}
\subsection{Time reversal symmetry breaking}
An interesting transformation to consider is $\psi\rightarrow\bar{\psi}$ and $t\rightarrow -t$.
This transformation is analogous to the time reversal symmetry (TRS) operator of the Schr\"odinger equation, which resembles~\Eq{eq:psi-eom}.
The conjugation operation $\psi \rightarrow \bar{\psi}$ corresponds to reflecting the displacements $y\rightarrow -y$ of each gyroscope about its pivot point, instead of reversing momenta.
A single gyroscope system clearly respects this symmetry, in contrast to a charged particle in a magnetic field; spinning alone does not break this form of TRS.
Interestingly, a system of two interacting gyroscopes also respects TRS.
If the bond connecting the two gyroscopes is aligned with the $x$ axis, this follows trivially from the fact that $\theta_{pq} = 0$ or $\pi$.
Even if the angle $\theta_{pq}$ is nonzero, we can globally rotate the system without penalty.
Therefore, after this coordinate change, the factor $e^{i2\theta_{pq}}$ is purely real, and there is no phase associated with the coupling between polarizations $\psi^R$ and $\psi^L$.
Introducing a third gyroscope, however, allows time reversal symmetry to be broken.
Now no global rotation of the system will result in bond angles $\theta_{pq}$ such that all factors $e^{i 2 \theta_{pq}}$ are real.
The simplest example in which one can observe the role of TRS breaking is the equilateral triad of coupled gyroscopes shown in~\Fig{fig:2gyro}b.
For this triad, three distinct eigenmodes arise: one uniform in-phase precession with no spring stretching, and two `excited state' motions $\vec{e}_a$ and $\vec{e}_b$.
In the particular case of the equilateral triad, neighboring gyroscopes differ in phases by $2\pi/3$ either clockwise (for $\vec{e}_a$) or counterclockwise (for $\vec{e}_b$) around the triangle.
The action of the time reversal operation mixes eigenstates.
This can be seen intuitively by examination of the excited states shown in~\Fig{fig:3gyro_TRS}a.
In particular, the $TR$ operation, given by $\psi \rightarrow \bar{\psi}$ and $t\rightarrow -t$, mixes $\vec{e}_a$ and $\vec{e}_b$, which have different frequencies.
In this scenario, we find that TRS breaking is manifest by the difference in frequency of the two eigenmodes that are mixed by the \textit{TR} operation.
More generally, for systems of more than two gyroscopes, \textit{TR} transforms a given configuration into a linear superposition of modes with different frequencies unless there are additional symmetries that guarantee the bond energy is preserved: $TR$ is not typically a symmetry of the dynamical equations~\cite{mitchell_tunable_2018}.
This lack of TRS is manifest in the differing frequencies of these eigenmodes.
Together with the fact that the two excited states have differing relative phases between gyroscopes ($\pm 2\pi/3$ radians), the frequency difference between these modes leads to beating patterns with a chiral flow of energy.
As shown in~\Fig{fig:3gyro_TRS}b, displacing a single gyroscope leads to large amplitude excitation that sloshes from one gyroscope to another in a chiral fashion over time.
This resulting chiral conductance of energy foreshadows the emergence of chiral edge waves.
Now that we have gained intuition for the role of TRS breaking from the three-gyroscope case, we can see the effects of TRS breaking in the dynamical matrix itself.
We can envision excitation of gyroscopes `hopping' from site to site with couplings given directly by the elements of the dynamical matrix $\mathbf{D}$ in~\Eq{eq:basicD}.
In fact, we see directly from~\Eq{eq:psi-eom} that the coupling of clockwise precession and counter-clockwise precession depends explicitly on the geometry with \textit{complex} couplings, whereas gyroscopic motion with the same polarization at different sites -- $\psi^R_i$ and $\psi^R_j$, for example -- have purely real couplings.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig_2gyro3gyro_StrongPinningLimit2.png}
\caption{\textbf{Low interaction strength enables connection to tight-binding models by diminishing coupling between polarizations.}
In the limit that $\Omega_k /\Omega_p \rightarrow 0$, eigenmode frequencies of a gyroscope pair or triad approach the precession frequency, $\Omega_p$.
The slight deviations from $\Omega_p$ are proportional to $\Omega_k$, resembling a tight-binding model with inter-site coupling of $\Omega_k/2$ (\Eq{eq:tightbinding_limit_gyro}).
\textit{(a)} Two modes represent a pair of gyroscopes' bonding and anti-bonding behavior.
Decreasing the interaction strength decreases the ellipticity of the higher-frequency mode: the counter-rotating component presents only a small perturbation on the dominant clockwise motion.
In the limit of $\Omega_k / \Omega_p \rightarrow 0$, both modes are nearly circular, like in quantum mechanical models with unitary evolution.
Here, the phase of the displacement is akin to the phase of a wavefunction.
\textit{(b)} For the gyroscopic triad, the oscillation frequency is $\Omega_p$ for the uniform precession (`ground state') mode and $\Omega_p + 3\Omega_k/2$ for the phased circular mode.
The intermediate mode frequency varies as a function of $\Omega_k/\Omega_p$, interpolating between the two bounds.
Only the intermediate eigenmode changes shape, becoming increasingly circular as the coupling strength weakens.
} \label{fig:pinning_limit}
\end{figure}
\subsection{Connection to tight-binding models: the limit of weak interactions}
If we consider the dynamics when interactions between gyroscopes are weak compared to the pinning forces restoring a gryoscope toward its hanging position, the effects of counter-clockwise motion act as a perturbation on the dominant clockwise precession.
\Fig{fig:pinning_limit} highlights how the two-gyro and three-gyro cases explored thus far vary as the interaction strength is tuned.
Modes that mix polarizations become increasingly elliptical as the coupling strength grows.
For weak interactions ($\Omega_k/\Omega_p \ll 1$), gyroscopic networks' eigenmodes become nearly circular, which highlights the perturbative effect of the counter-rotating polarization on the dominant precession.
In fact, in this limit there is a natural correspondence with hopping models ubiquitous in condensed matter.
Just as precession of a gyroscope can exert torque on its neighbors to induce precessional motion, with an energy cost depending on the bond strength, so too electrons or other excitations can hop from one site to another associated energy based on the coupling between sites.
In this `tight binding limit', in which $\psi^L$ has a perturbative effect on the dominant $\psi^R$ precession, we again see that all complex terms arise from the network geometry, meaning that \textit{phase offsets} are introduced in the response of one gyroscope to another's displacement in a manner that depends on angles between adjacent bonds in the network~\cite{nash_topological_2015}.
In particular, the approximate equations of motion allow one to predict the evolution by considering only the magnitudes and phases of the dominant rotation at each site:
\begin{multline}\label{eq:tightbinding_limit_gyro}
\omega \psi^R_i = \Omega_p \psi^R_i + \frac{\Omega_k}{2} \sum_{m \in N(i)} (\psi^R_i - \psi^R_j) \\
+ \frac{\Omega_k^2}{8 \Omega_p}\left[
- \sum_{j,j'}\left(\psi^R_i - \psi^R_j\right) e^{2i \theta_{j'ij}} \right.\\
\left.
+ \sum_{j,k} (\psi^R_j-\psi^R_k)e^{2i\theta_{ijk}}
\right],
\end{multline}
where the bond angle $\theta_{ijk}$ is the opening angle between bonds $ij$ and $jk$, the first sum is over the $N$ neighbors of $i$, and the subsequent sums are over pairs of neighbors $j$ and $j'$ of $i$ and neighbors $k$ of $j$ for each of $i$'s neighbors $j$.
Note that the phase offsets of $2\theta$ arise in the term proportional to $\Omega_k^2 /8 \Omega_p$.
This complex next-nearest-neighbor coupling is reminiscent of phases from magnetic flux in common tight-binding models~\cite{haldane_model_1988}, but here the phases depend on the arrangement of the gyroscopes in space.
Furthermore, the phase shifts arise here because of the counter-rotating polarizations, $\psi^L$, which are not present in the tight-binding picture.
We stress that the complex-valued coupling encodes the fact that perturbing one gyroscope gives rise to a \textit{phase-shifted} response at a nearby site, with a phase shift dictated by the geometric arrangement of the bonds.
Thus, geometry provides phase shifts which generically break TRS except in special circumstances such as a rectangular lattice, where all angles $\theta_{ijk}$ are multiples of $\pi/2$.
As discussed below, it is these TRS-breaking terms that generate topological bandgaps in gyroscopic networks.
\section{Band gaps without full diagonalization}
\begin{figure*}[ht]
\includegraphics[width=\textwidth]{./fig_diracmodes_final_with_ABDelta_and_springs_double_ver3.png}
\caption[]{
\textbf{Gyroscope mode displacements reveal how coupled circular polarizations lead to a gap.}
\textit{(a)}
The Dirac cones for the weak-coupling band structure (light orange bands) evolve into a band gap for the finite coupling case (black band structure), here shown for $\Omega_k = \Omega_p$.
The point of closest approach lies at the wavevectors $\vec{k} = K$ or $K'$, where the normal mode displacements take on a simple character: clockwise precession of one sublattice site and small (lower) or vanishing (upper) counter-clockwise motion at the second sublattice site (left inset).
Upon examination of the spring couplings connecting a unit cell to its neighbors (right insets), we find that spring forces cancel for the upper mode but do not cancel at the sublattice site with small excitation for the lower mode.
\textit{(b)} The geometry of these eigenmodes is determined by rotational symmetry at the $K$ and $K'$ points.
Since a $120^\circ$ rotation about any gyroscope must change the pattern only by a phase, all displacements must be purely circular.
Furthermore, the state can either be invariant under rotation (upper panel) or else phase shifted with a counter-clockwise rotation at the $A$ site (lower panel).
\textit{(c)} For $\Omega_k \ll \Omega_p$, the two normal modes are asymptotically equivalent, up to an inversion of $A\leftrightarrow B$, since the smaller counter-clockwise rotation on $A$ vanishes in this limit.
As the spring coupling increases, the counter-clockwise excitation at the $A$ site grows in magnitude, decreasing the spring energy and broadening the gap. In the limit of $\Omega_k \gg \Omega_p$ (rightmost images), the counter-clockwise excitation magnitude approaches that of the clockwise excitation.
\textit{(d)}
When the pinning frequencies at the two sublattice sites differ, so that $\Omega_p^A \rightarrow \Omega_p + \Delta_{AB}$ and $\Omega_p^B \rightarrow \Omega_p - \Delta_{AB}$, a trivial gap opens in the limit of $\Omega_k \ll \Omega_p$.
Though the lower frequency state does not exhibit torque balance at the immobile sublattice site, the spring energy is negligible compared to $\Omega_p$, so pinning torques dominate and no counter-clockwise motion is observed.
As a consequence, the excitation patterns bounding the gap do not change as the frequency splitting between the $A$ and $B$ sites grows.
}
\label{fig_diracmodes}
\end{figure*}
In this section we consider the spectral gaps that arise naturally in large collections of gyroscopes in ordered as well as amorphous configurations.
We focus on the role of symmetry and local dynamics in determining their origin.
We begin by examining the mechanical motions that underlie the frequency splitting of neighboring bands, linking these motions to TRS breaking through the coupling of left- and right-circular polarizations.
We then study the evolution of the band structure as we tune the coupling between neighboring unit cells of gyroscope pairs, allowing us to contrast the topological gap origins with gaps created from bonding/anti-bonding behavior of isolated dimers.
Lastly, we provide a simple method to predict the existence of band gaps without full diagonalization of a system's dynamical matrix.
\subsection{Band gaps in lattices}
A simple honeycomb lattice of gyroscopes has gaps in the system's bulk phonon spectrum, shown in the black band structure in~\Fig{fig_diracmodes}a.
If the spring-like coupling between sites is weak, the gap nearly closes at the corners of the Brillouin zone (orange spectrum in~\Fig{fig_diracmodes}a).
As we increase the coupling, the gap broadens.
Physically, gaps arise due to the coupling of left-polarized motion with the right-polarized motion of the gyroscopes, which breaks time reversal symmetry in the dynamics (\Eq{eq:psi-eom}).
Inspecting the honeycomb lattice's normal modes bounding the gap leads to an intuitive picture of this coupling's effect.
The lattice symmetry allows us to focus on a single hexagonal plaquette to fully describe the bulk excitations bounding the bandgap.
\Fig{fig_diracmodes}a shows these modes bounding the gap, located at the corners of the Brillouin zone $K$ and $K'$.
In the higher frequency mode bounding the gap, one sublattice site precesses clockwise and the other does not move.
In the lower frequency mode, on the other hand, the excitation is not completely localized to one of the two lattice sites: the other site precess counterclockwise with a small amplitude.
The difference in this motion provides an intuitive way to explain the gap, as we detail in a multi-step argument in~\Fig{fig_diracmodes} and Appendix~\ref{appendix:dirac_splitting}.
The crux of the argument relies on the relationship between energy and frequency given earlier in~\Eq{eq:energy-frequency}.
The small counterclockwise displacement in the lower-frequency mode relaxes the cycle-averaged spring energy, and since the oscillation frequency $\omega$ is directly proportional to the average potential energy over a normal mode cycle, the frequency of this mode is smaller than the state with a truly immobile site
(see Appendix~\ref{appendix:dirac_splitting} for details).
This lower energy is evident from the force imbalance from the neighboring springs in the lower-frequency mode (\Fig{fig_diracmodes}a).
In contrast, for the higher-frequency mode, springs are stretched or compressed by the same amount around the $B$ site, so there is no net force, and the center gyroscope does not move at all.
This mechanical mechanism for gap generation is different than in the Haldane model~\cite{haldane_model_1988}, where all excitation resides on only one of the two sublattice sites for both modes bounding the gap and the energy splitting arises purely from the next-nearest neighbor coupling.
In this case, the particular lattice symmetry of the honeycomb allowed us to make strong constraints on the shape of the mode displacements to explain the band gap (see Appendix~\ref{appendix:dirac_splitting}), but this is not always possible in general.
In later sections, in discussing structures without translational order, we will provide alternative arguments for the existence of topological band gaps.
Increasing the spring coupling amplifies the frequency separation between the two states, broadening the gap and leading to larger and larger counter-clockwise rotation on the $B$ sites at $K'$, as shown in~\Fig{fig_diracmodes}c.
If, instead of increasing the spring coupling, we detune the precession frequencies at the two sublattice sites, the gap similarly broadens.
However, the material is a trivial insulator, and the motions of the two states bounding the gap remain unchanged as the inversion symmetry breaking grows, without any signature of coupling between right and left polarizations (\Fig{fig_diracmodes}d).
What have we learned?
The interplay between coupled polarizations and gyroscope phases interact in tandem to break time reversal symmetry, reminiscent of the simple three-gyroscope triad of Figures~\ref{fig:2gyro}-~\ref{fig:pinning_limit}.
The spring forces directly induce counter-rotation because of the surrounding gyroscopes' phases.
In turn, this coupling between the clockwise and counter-clockwise states opens the gap by splitting the cycle-averaged spring energies, imparting an opposite chirality of precession between the two sublattice sites.
In more general scenarios, we may not have such symmetric configurations near the gap, but the role of coupled polarizations and relative phases remains.
Importantly, if the dynamics were artificially modified to suppress the coupling between the left polarized (dominant for $\omega < 0$) and right polarized (dominant for $\omega > 0$) excitations, this splitting would not be possible, and no gap would emerge.
That is, if we adjust the strength of block off-diagonal entries of the dynamical matrix that describe the coupling between clockwise and counter-clockwise motion,
\begin{equation}
D = \left(
\begin{array}{c|c}
D_{LL} & \varepsilon D_{LR} \\
\hline
\varepsilon D_{RL}
& D_{RR}
\end{array}
\right),
\end{equation}
so that $\varepsilon \rightarrow 0$, then the topological character of the bands vanishes.
From~\Eq{eq:psi-eom}, we recognize that this modification destroys any dependence of the dynamics on the geometry of the lattice, further highlighting the geometry-dependent topological phases that emerge.
This right-left-coupling avenue for gap formation is distinct from inversion symmetry breaking, which instead adds only diagonal offsets to the dynamical matrix in $D_{LL}$ and $D_{RR}$, with no mixing of polarizations.
\subsection{Spectrum flow: from isolated fragments to a strongly-coupled network }
\begin{figure*}[ht]
\includegraphics[width=\linewidth]{./fig_spectrumflow_ver2.png}
\caption[]{
\textbf{
The transition from isolated bonding/anti-bonding behavior to Chern insulating behavior in a Haldane-like tight-binding model mimics the analogous transition in gyroscopes.
}
\textit{(a)} Pairs of bonded sites or \textit{(b)} gyroscopes are coupled to each other through bonds of varying strength (dashed lines).
When the coupling is absent ($t_1'=0$ or $\Omega_k'=0$), the material is in a trivial insulating phase with one `bonding state', in which the excitations at site pairs are in phase, and one `anti-bonding state', in which the excitations are out of phase.
These bonding and anti-bonding behaviors thicken into bands as the coupling is increased, until the two mix at the gap closing.
Beyond this value, each band becomes topologically nontrivial, as marked by their nonzero Chern numbers $C = \pm 1$ (red and blue, respectively).
The strength of complex next-nearest neighbor couplings, $t_2$, are controlled by the weakest link between next-nearest neighbors, which is $t_1'$.
Here, we set $t_2 = i t_1' / 10$.
\textit{(b)}
The gyroscopic case, governed by the equations of motion given in~\Eq{eq:psi-eom}, mimics the tight-binding picture, though there is a redundant copy of the positive frequency band pairs at negative frequencies.
}
\label{fig:spectrumflow}
\end{figure*}
One way to disentangle local and global effects in an ordered system is to artificially partition an ordered system into subsections and study what happens as the partitions are coupled.
We do this here with our gyroscopic lattice, showing that the topological gap cannot be understood as frequency splitting between bonding and anti-bonding behaviors, and we find analogous behaviour in a familiar tight binding model.
A pair of coupled oscillators will typically have a frequency splitting between bonding (in-phase) and anti-bonding (out-of-phase) behavior.
Even in the presence of coupling between many such pairs of sites -- which endows width to bands of eigenmode frequencies -- this fundamental origin of the two bands can be seen in the in-phase and out-of-phase character of eigenmodes.
This results in a local origin of band gaps in tight binding systems, with a natural extension to the case of amorphous materials~\cite{weaire_electronic_1971}.
It is tempting to think that the two bands of our topological metamaterial seen in~\Fig{fig_diracmodes} may be continuously connected to in-phase bonding and out-of-phase anti-bonding behavior of gyroscopes in the unit cell.
This, however, cannot be the case.
We show this in \Fig{fig:spectrumflow} by considering a periodic honeycomb lattice with tunable coupling between unit cells, each of which consists of a pair of sites.
We compare both the tight-binding and gyroscopic case, where the two-banded tight-binding dynamics analogous to~\Eq{eq:tightbinding_limit_gyro} is given by the Haldane model of Ref.~\cite{haldane_model_1988}:
\begin{equation}
H = -t_1 \sum_{\langle ij \rangle} c_i^\dagger c_j - t_2
\sum_{\langle\langle ij \rangle\rangle}e^{-i \phi_{ij}} c_i^\dagger c_j,
\end{equation}
where $\langle ij \rangle$ denotes nearest neighbors, $\langle\langle ij \rangle \rangle$ denotes next-nearest neighbors, $c^\dagger_i$ is the creation operator at site $i$, and $\phi_{ij}=-\phi_{ji}$ is a phase shift associated with hopping from site $i$ to site $j$.
By comparison with~\Eq{eq:tightbinding_limit_gyro}, we see that $t_1$ is akin to the coupling $\Omega_k$ between gyroscopes, $t_2$ is akin to $\Omega_k^2/8\Omega_p$, and $\phi_{ij}$ plays the role of the gyroscopic network bond angle $2\theta_{ijk}$.
To tune between bonding/anti-bonding behavior within each pair of sites and topological behavior across the bulk, we vary the direct coupling $t'_1$ between adjacent unit cells while keeping $t_1$ fixed for sites within a unit cell, and we set $t_2=i t_1'/10$.
If the bonds denoted by dashed lines are severed completely ($t_1' = 0$), the resulting network consists of isolated pairs of sites, each coupled by a bond of strength $t_1$ and uncoupled from other pairs.
Such a system is a trivial insulator, akin to isolated molecular dimers in a lattice.
As the dashed bonds are strengthened, the width of the two bands begin to broaden while remaining in the trivial insulator phase.
The system transitions to a topological phase once $t'_1 \approx 0.5$.
Inspection of the eigenstates near the topological gap for $t_1' > 0.5$ reveals that the displacement patterns do not resemble simple bonding/anti-bonding patterns (refer back to the normal modes for $\Omega_k \ll \Omega_p$ in ~\Fig{fig_diracmodes}c, which correspond to the case of $t_1'=t_1$).
How do we interpret this?
For small $t'_1$, the gap is formed by a frequency separation between bonding and anti-bonding behavior, while the gap with large $t'_1$ arises due to the phase shifts enabled by coupling between clockwise and counter-clockwise polarizations.
Furthermore, the two phases have a macroscopic difference: the topological phase supports chiral edgemodes, indicated by the nonzero Chern number of each band.
In~\Fig{fig:spectrumflow}, the colored bands exhibit nonzero topological invariants of $C=+1$ (red) and $-1$ (blue), which we will define in depth in section~\ref{section:real-space_topology}.
In short, by interpolating between the trivial and topological phases, we find that bonding and anti-bonding bands mix in order to open a topologically nontrivial gap.
The gyroscopic case is similar: trivial bonding/anti-bonding gaps are not continuously connected to the topological gaps from TRS breaking.
As shown in Figure~\Fig{fig:spectrumflow}b, the spectrum of the gyroscopic honeycomb lattice with spring interactions resembles two copies of the Haldane model spectrum, one with $\omega > 0$ and another with $\omega < 0$, with symplectic symmetry relating the two copies.
(Elsewhere in the text, we omit the redundant, negative frequency states.)
Varying $\Omega'_k$ between adjacent unit cells drives a topological phase transition, closing the gap generated by bonding/anti-bonding behavior of the coupled dimers.
Once the gap reopens, the bonding and anti-bonding behaviors are mixed together, allowing topologically non-trivial bands and chiral edge modes in the gaps.
In this strongly interacting regime, the classic argument for the local origin of bandgaps (like that of~\cite{weaire_electronic_1971}) does not apply.
This is because the physics of TRS breaking that opens the gap is distinct from the bonding/anti-bonding picture.
The question is then: can we still predict whether there is a gap using a local argument, or do we have to compute the full band structure?
\section{Understanding the gap in the strongly interacting regime}
\label{section:understanding_gap_in_regime}
The origin of band gaps can be easily understood in systems in which individual sites are weakly coupled to each other, as compared to their on-site energy.
The dynamical matrix of such a system has the general form of $D=D_{diag}+\delta$,
where $D_{diag}$ encodes the on-site energy in the electronic tight binding picture (and the precession frequency of the individual gyroscopes in the gyroscopic case).
$\delta$ is a matrix of weak (as compared to the elements of $D_{diag}$) off-diagonal couplings.
Schematically,
\begin{equation}\label{ex_staggered}
D \approx
\left(\,\,
\begin{array}{ccc}
\diagentry{\begin{array}{cc} t_0 & \\& -t_0\end{array}} & & \delta \\
&\diagentry{\xddots}\\
\delta &&\diagentry{\begin{array}{cc} t_0 & \\ & -t_0\end{array}}
\end{array} \, \,
\right).
\end{equation}
For $\delta \ll |t_0|$, the eigenvalues of $D$ are simply close to the diagonal entries, and therefore gaps reside between different values of diagonal elements.
In the example of~\Eq{ex_staggered}, the eigenvalues bunch together around $\pm t_0$, and a spectral gap lies in between.
For trivial insulators with bonding/anti-bonding behavior, the dynamical matrix is not nearly diagonal, but can be arranged to be nearly block diagonal.
The dynamical matrix of such a system appears in nearly block diagonal form, with all strong couplings contained within blocks and only small terms in the off-diagonal sectors:
\begin{equation}
D \approx
\begin{pmatrix}
\diagentry{\begin{array}{cc} 0& t\\t& 0\end{array}} & & \delta \\
&\diagentry{\xddots}\\
\delta &&\diagentry{\begin{array}{cc} 0& t\\
t& 0\end{array}}\\
\end{pmatrix}.
\end{equation}
As highlighted in the previous section, a system that can be divided into pieces, like our weakly coupled unit cells in a honeycomb lattice, generates two separated frequency bands from bonding and anti-bonding.
The weak connections between the components broaden the spectrum by only a small amount, leaving the gap intact.
This is similar to previous works in which a bond-centric description can explain why amorphous glasses have band gaps~\cite{weaire_electronic_1971}.
Note that in this case, though $D$ is not nearly diagonal, $D^2$ \textit{is} nearly diagonal.
Unlike either diagonal or block diagonal systems, a system with a topological band gap cannot be divided up into nearly independent pieces (\Fig{fig:spectrumflow}, see also~\cite{arovas_localization_1988}).
To transform a trivial gap into a topological one, couplings must be sufficiently strong to close and reopen the gap.
From the perspective of the dynamical matrix, these strong off-diagonal couplings cannot be arranged in block diagonal form~\cite{arovas_localization_1988,huo_current_1992}.
For example, the structure of such a dynamical matrix could follow the structure of~\Eq{eq:psi-eom}, which we can roughly conceptualize as
\begin{equation}
D \approx
\begin{pmatrix}
\diagentry{0} & ... & t & ... & it' \\
&&\diagentry{\xddots}& &\\
-it' & ... & t & ... &\diagentry{0}\\
\end{pmatrix},
\end{equation}
where large couplings $t$ and TRS breaking terms $i t'$ are distributed throughout.
One can certainly characterize band gaps by diagonalizing the Hamiltonian, but it would be useful to have a protocol for estimating whether a gap exists or what frequencies a gap may occupy without this costly step.
For the case of a broad class of topological systems, we note that although $D$ is highly non-local, $D^2$ is nearly diagonal, and we construct a method for estimating the gap in these systems below.
This condition is closely connected to the condition of having a flat band, as occurs in crystal materials supporting the Fractional Quantum Hall Effect~\cite{leykam_artificial_2018}.
Such a protocol lends itself particularly well for understanding gaps in amorphous systems.
\subsection{Existence of a band gap: algebraic approach}
Consider a system with a dynamical matrix $D$ such that (1) some off-diagonal elements of $D$ are large compared to on-diagonal elements -- ie, there are some couplings that are large compared to on-site pinning -- and (2) the square of the dynamical matrix is nearly diagonal such that
\begin{equation} \label{eq:interference}
D^2=\omega_0^2\mathds{1}+\epsilon,
\end{equation}
where $\omega_0$ is defined as the square-root of the diagonal entry of $D^2$ and $\epsilon$ contains all the off-diagonal entries.
We seek an argument that indicates whether a band gap exists without diagonalization of $D$.
\begin{figure}
\includegraphics[width=.45\textwidth]{ari_fig1.png}
\caption{Square lattice model whose band gap and Chern number can be calculated by a local approximation.
\textit{(a)} The model's dynamical matrix $D$ is built from couplings $\pm t$ (solid lines) with the sign drawn on the picture.
The grey diagonal lines have a coupling $i t'$ in the direction indicated by the arrows and $-it'$ in the opposite direction.
\textit{(b)} The square of the dynamical matrix, $D^2$, includes only horizontal and vertical couplings of two steps with matrix-entry $t^2-2t'^{2}$ and diagonal couplings with matrix entry $t'^2$. Because of the phases in $t$, all the other matrix elements cancel out.
\textit{(c)} Examples of pairs of paths that cancel one another in calculating $D^2$ for the matrix element connecting site 1 (red) to site 2 (blue).}
\label{fig:Hofstadter}
\end{figure}
Both conventional insulators with strong bonding/anti-bonding behavior and topological systems meet these criteria, provided one subtracts a constant frequency so that the gap is approximately centered about zero.
As an example with which to make the argument concrete, we will use one of the simplest systems with a Chern number, shown in~\Fig{fig:Hofstadter}.
This model comes from studying electrons in a square lattice in a magnetic field (a Hofstadter model) with a half unit of magnetic flux per unit cell.
The diagonal couplings of strength $t'$ break time reversal symmetry, opening the possibility of a topological band gap, while direct couplings are given a strength $t$.
We choose signs and phases of the couplings so that the phase of all the couplings going around each right triangle is $\pi / 2$.
Using this model, we find that the gap can be understood to result from interference during propagation of waves in the system, and a variational principle shows that this system has a gap for a range of values of the complex, next-nearest-neighbor coupling strength, $t'/t$.
The method presented below can also be applied to a system of gyroscopes in the weak-interaction limit as described by~\Eq{eq:tightbinding_limit_gyro}.
We first show that there are both positive and negative
eigenvalues (i.e., eigenvalues both above and below the net pinning frequency of an individual gyroscope).
The value of $v^\dagger D v$ for any normalized vector $v$ is an upper bound for the lowest eigenvalue of $D$ if $D$ is Hermitian, according to the variational principle.
Similarly, for another vector $u$, $u^\dagger D u$ is a lower bound on the highest eigenvalue.
One can choose in a variety of ways vectors $v$ and $u$ such that $v^\dagger D v$ is either positive or negative, which forces there to be at least some positive and some negative eigenvalues.
For example, let $v$ be non-zero only on two adjacent sites of the network; the expectation value is either positive or negative for a suitable choice of relative phase between the two sites.
This will be the case if there are off-diagonal elements that are larger than the diagonal elements of $D$ -- which is certainly true for the Hofstadter model, for which diagonal elements are zero.
Not only are there states above and below, but furthermore these two collections of states do not touch.
To see this, consider the case where all diagonal couplings in the dynamical matrix are equal, and note that the \textit{square} of dynamical matrix has off-diagonal couplings that are small relative to diagonal ones: $D^2=\omega_0^2\mathds{1}+\epsilon$.
If the entries of $\epsilon$ are small compared to $\omega_0^2$ the eigenvalues of $D^2$ must be close to $\omega_0^2$, so $D$'s eigenvalues must be close to $\pm \omega_0$, leaving a gap near $0$.
To make this argument more precise, we use a bound from Appendix~\ref{appendix:proof_gap}.
From~\Eq{eq:interference} it follows that the eigenvalues of $D^2$ are greater than or equal to $\omega_0^2-\max_i \sum_j|\epsilon_{ij}|$.
If this is positive, $D$ has a gap whose width $\Delta$ is at least
\begin{equation}
\Delta \ge 2\sqrt{\omega_0^2-\max_i\sum_j|\epsilon_{ij}|}.
\end{equation}
Thus, the width of the gap is bounded by a quantity that does not require the diagonalization of $D$.
Intuitively, we can see that $D^2$ is nearly diagonal because of the effects of interference between paths.
Each entry of $D^2$ is a sum $\sum_k D_{ik}D_{kj}$ over intermediate vertices $k$ that have a nonzero coupling to both vertex $i$ and vertex $j$.
For the $D$ given above, the entries of $D^2$ are shown in \Fig{fig:Hofstadter}b.
Any pair of sites that are connected by a sequence of two bonds can in principle have a nonzero matrix element, but many of the couplings are zero because they cancel out (\Fig{fig:Hofstadter}c).
The quantities in the estimate are $\omega_0=2\sqrt{t^2+t^{'2}}$ and $\max_i\sum_j|\epsilon_{ij}|=4|t^2-2t^{'2}|+4t^{'2} $ so $\Delta \ge 4\sqrt{t^2-|t^2-2t^{'2}|}$.
If $|t'|<|t|/\sqrt{2}$, this bound is $4\sqrt{2}t'$, which is proportional to the size of the time-reversal symmetry breaking term, as expected.
In this model, a gap exists for any value of $t'\ne 0$, and this estimate proves the existence of this gap at least up to $|t'|=|t|$.
\subsection{Green's function approach yields a gap}\label{Greens_func_yields_gap}
To better understand the connection of this argument to interference and locality, consider the Green's function $G(j,i;\omega)=[(\omega-D)^{-1}]_{ji}$, the amplitude of the oscillations of the $j^\mathrm{th}$ site when the $i^\mathrm{th}$ is oscillated by a unit force at a frequency $\omega$.
This shows how far energy is transported at a given frequency.
We will show that this decays exponentially over a range of frequencies, implying that this range is a gap.~\footnote{Unlike the previous argument, here we cannot prove a vanishing density of states, but for disordered or amorphous systems, the mobility gap is the physically relevant feature, given that the density of states need not vanish in the gap.}
First consider an ordinary insulator where there
are two types of sites with different diagonal spring constants.
Write $D=D_{diag}+\delta$, where $D_{diag}$ includes the diagonal pinning frequencies and $\delta$ includes only the off-diagonal ones.
The Green's function can be expanded in a geometric series as $\frac{1}{\omega-D_{diag}}\sum_{n=0}^\infty \left(\delta \frac{1}{\omega-D_{diag}}\right)^n$.
If $\omega$ is sufficiently far from the two diagonal entries, this series decays exponentially with distance.
Let us imagine ``bonds" drawn connecting any two sites with a non-zero value of $\delta$.
The first contribution to the sum for a pair of sites $i$ and $j$ appears at the $m^\mathrm{th}$ term if it requires $m$ steps to connect $i$ to $j$ along these bonds.
The magnitude of this term decays as $(\max_{pq} |\delta_{pq}|/\min_{p} |\omega-D_{pp}|)^m$, with the rate of decay given by the ratio of the maximum off-diagonal component of $D$ to the minimum on-diagonal component of $\omega-D$.
This suggests that the Green's function decays exponentially.
This argument leaves out a factor that counts the number of paths connecting a given pair of sites, but when this factor is included, we find that the Green's function does in fact decay exponentially, as shown in Appendix~\ref{appendix:green_function}.
Intuitively, with a short-range dynamical matrix, the network's energy transport is localized unless the couplings are large enough to be comparable with the difference between the applied frequency $\omega$ and the natural frequencies on the diagonal of $D$.
For our system, this argument at first appears invalid since the off-diagonal couplings are not small when gyroscopes are strongly coupled.
A slight modification of this calculation does hold, however.
For some forcing $\vec{f}$, the system responds with $\partial_t \psi_i=-i\sum_j D_{ij}\psi_j-i f_i$.
Through repeated differentiation, we can find the derivatives of $\psi_i$ in terms of the initial values of the $\psi$'s, and typically higher derivatives depend on the displacement of more distant gyroscopes.
This suggests that energy can be transported over long distances through chains of interacting sites.
But if $D^2$ is almost diagonal, $\partial_t^2 \psi_i = -\sum_{jk} D_{ij}D_{jk}\psi_k-\sum_j D_{ij}f_j-i\partial_t {f_i}$, and the coefficient of $\psi$ on the right-hand side is $D^2=\omega_0^2\mathds{1}+\epsilon$, which is nearly diagonal.
Physically, the disturbance produced by a given gyroscope mostly reflects back to itself after two steps due to interference.
The solution to the equation for forces at a frequency $\omega$ (using vector notation) is
\begin{align}\label{eq:green_series}
\psi&=\frac{1}{\omega^2-\omega_0^2-\epsilon}(\omega+D)f\nonumber\\
&=\frac{\omega+D}{\omega^2-\omega_0^2}\sum_{n=0}^\infty \left(\frac{\epsilon}{\omega^2-\omega_0^2}\right)^n f,
\end{align}
which converges if $\epsilon$'s entries are small compared to $|\omega^2-\omega_0^2|$, so the Green's function (the coefficient of $f$) decays exponentially with the distance.
Though no part of the topological system is disconnected from the remainder by weak couplings, interference can nonetheless suppress tunneling of excitation from a given region.
This inteference can arise over short paths or long length paths, depending on the details of the network.
In the Hofstadter model we have used here, short paths of length two already generate significant interference, enabling a simple test to identify the band gap.
In general, the interfering paths might be longer, requiring more and more terms in~\Eq{eq:green_series} to converge.
Though the paths may be long, the series will necessarily converge for a suitable choice of parameters (see Appendix~\ref{appendix:generalize_chern}).
If there is a gap, this method identifies its existence if~\Eq{eq:green_series} converges, even in the case of topological systems with strong off-diagonal couplings.
This method also provides insights into the interference processes within the gap by following terms within the series.
We have discussed three cases.
In trivial insulators with nearly-diagonal $D$, we can predict the presence of a gap from the diagonal elements.
Systems dominated by bonding/anti-bonding behavior are instead nearly \textit{block} diagonal, and a gap will reside between bands of bonding and anti-bonding behavior.
Third, topological systems require strong coupling between sites to drive the system into a different phase -- one that cannot be described as a composite of non-interacting pieces.
In this strongly-coupled regime, interactions and eigenstates are non-local, threatening any prediction of band gaps without full diagonalization of the dynamical matrix~\cite{arovas_localization_1988,huo_current_1992}.
Nonetheless, we constructed a simple argument if $D^2$ is nearly diagonal.
As shown in Appendix~\ref{appendix:generalize_chern}, this argument generalizes to a local method for finding locations of gaps even when $D^2$ is not nearly diagonal, since in such a case interference occurs between sufficiently long paths.
Our argument applies both to topological systems and other systems with strong couplings for which $D^2$ is nearly diagonal in the bulk.
In topological phases, note that $D^2$ will \textit{not} be nearly diagonal on the boundary, and this delocalization signals the existence of protected edge modes.
For disordered and amorphous samples, the denominators in~\Eq{eq:green_series} could sometimes be small, but the total expression may nonetheless decay in a similar manner to the Green's function of disordered trivial insulators~\cite{anderson_absence_1958}.
\section{Real-space topology}\label{section:real-space_topology}
Now that we have gained insight into how gaps arise, we turn to characterizing their topology.
In particular, we seek local, real-space probes for global topological invariants.
A natural starting point is the notion of flux pumping, which is a natural connection between a localized perturbation and a global consequence.
After observing how flux-pumping plays out in a mechanical context, we review the traditional notion of the Chern number and its real-space generalizations in the form of the Kitaev sum and the Bott index.
We compare these three methods of computing band topology and probe the extent to which we can approximate the Chern number using only the local structure of a network.
Using a variant of the Kitaev sum, we find that, surprisingly, a purely local computation can approximate the real-space generalization of the Chern number.
\begin{figure*}
\includegraphics[width=\textwidth]{fluxpumping_swapped_ver3.png}
\caption[]{
\textbf{The topological invariant in a small system can be measured by two related methods: the spectral response to twisted boundary conditions and the Kitaev sum.}
\textit{(a)} Twisting the bonds along the dashed blue line by attaching an extensible ring to each gyroscope above the dashed line drives a phenomenon akin to flux pumping from a threaded magnetic field.
The extensible ring, like a Hoberman ring, remains centered around the pivot point of the gyroscope and is free to rotate.
This modulates the interaction by an angle, $\theta_{\textrm{twist}}$.
\textit{(b)}
Upon twisting the interaction angle by $2\pi$ radians, the spectrum returns to its initial state, but the nontrivial topology ensures that one state from the central region is pumped from the lower band to the upper band, while states localized to the outer boundary (green curves) are pumped towards the lower band.
Here, bands are shaded by a gray overlay.
The red circle in the spectrum corresponds to the normal mode excitation shown in panel (a).
\textit{(c)} The Kitaev sum reveals that there is chiral conductance within the bulk of even a system $\sim 16$ gyroscopes in extent.
Here, a region of summation half the size of the system is used to evaluate the sum at a grid of locations.
The sum is proportional to the number of states at the system's core that are pumped from the conduction to the valence band.
}
\label{fig_bulkboundary}
\end{figure*}
\subsection{Flux pumping as a local probe for band topology}
Arguably at the heart of band topology~\cite{laughlin_quantized_1981} lies the notion of flux pumping.
For an electronic Chern insulator, threading magnetic flux through the core of an annulus effectively transfers occupancy of edge states from one edge to another.
On one boundary, the process pumps electronic edge states into the bulk conduction band and bulk states in the valence band onto the edge, while on another boundary, the process pumps edge states into the valence band and pumps bulk states in the conduction band onto the edge.
With the goal of building a local description, we want to understand this topological phenomenon in terms of a localized perturbation.
A localized phenomenon that is similar to pumping flux through an annulus arises when threading one quantum of magnetic flux localized in the bulk of the material.
This localized insertion gives rise to a quantized charge accumulation at the site of flux insertion~\cite{mitchell_amorphous_2018}.
As magnetic flux is inserted into the bulk of the system, an induced electromotive force winds around the site.
Given that a Chern insulator will exhibit chiral conductance, this induced field drives a current toward or away from the site of flux insertion, accumulating charge at the insertion site that is balanced at the outer boundary.
In the mechanical version of a Chern insulator, the states can similarly be transformed into each other via a mechanical version of flux pumping, wherein upper frequency states transform into lower frequency states via edge modes on the outer boundary for a topologically nontrivial network.
Since the magnetic flux is related to the phase winding of electrons, magnetic flux can be simulated by altering the interactions such that the force of one gyro on its neighbor is rotated by an angle $\theta_{\textrm{twist}}$:
\begin{equation}
F \sim \psi_i -\psi_j \rightarrow
\psi_i - \psi_j e^{i\theta_{\textrm{twist}}}.
\end{equation}
To give a concrete picture of how this could be built in an experiment, we envision attaching an extensible Hoberman ring to a small number of gyroscopes, as illustrated in Figure~\ref{fig_bulkboundary}a.
Each ring remains centered about the hanging axis of its gyroscope, so that when a gyroscope is displaced, the ring uniformly extends in all directions.
The ring freely allows the gyroscope displacement to change in magnitude by expanding or contracting without any resistance and allows change in phase by rotating without any resistance.
To twist the gyroscope $i$'s response to its neighbor $j$, we affix the spring that couples $i$ and $j$ to a position on the ring offset from the gyroscope $i$'s tip.
This introduces a phase-shifted torque on the gyroscope.
To mimic the effects of a gauge field, we are free to concentrate the modifications to the spring attachments along a radial cut of the annulus, shown as a blue dashed line in Figure~\ref{fig_bulkboundary}a,
effectively supplying a `twisted' boundary condition to the system along the cut of the annulus.
Twisting the bonds along the cut deforms the resulting band structure, but the system returns to its original state once the twist has taken the value $\theta_{\textrm{twist}} = 2\pi$.
\Fig{fig_bulkboundary}b shows that during this operation, states localized to the outer boundary decrease in frequency, while a state localized to the center, where the effective magnetic flux resides, increases in frequency.
The net result of the operation is the exchange of one state between the two bands.
\Fig{fig_bulkboundary}b highlights this action in the gap: the edge (green) states are transformed into adjacent, lower frequency states, while the state at the center of the sample (purple) pumps from the lower band to the upper band.
In contrast, a trivial insulator has no exchange between the two bands, though extended states may be shuffled within each band separately.
Traditional treatment explains this phenomenon by looking at the band structure of the periodic bulk.
We briefly review this more restrictive perspective before moving to a real-space point of view, where the lattice structure of the bulk is not needed and we are able to compute the band topology and understand flux pumping seen even for the small system shown in~\Fig{fig_bulkboundary}.
\subsection{Traditional view: Berry curvature}
The traditional point of view for computing band topology is to work entirely in reciprocal (momentum) space, which requires a crystalline material.
While more restrictive, we briefly lay out this formulation in order to introduce the projection operator and highlight the alterations required by symplectic symmetry.
Here, the Chern number for gyroscopes on a lattice is defined from the band structure of the equations of motion for small displacements.
Diagonalizing the matrix describing the network's dynamics yields $2N$ frequencies of the dispersion bands at each value of $\vec{k}$, where $N$ is the number of gyroscopes per unit cell.
For each value of $\vec{k}$, each band has a corresponding eigenvector,
\begin{equation} |\boldsymbol{\psi}_j (\vec{k} )\rangle = \left(
\psi_1^R, ..., \psi_N^R,
\psi_1^L, ..., \psi_N^L
\right)
\end{equation}
characterizing the amplitudes and phases of the $N$ gyroscopes' collective motion.
The symplectic symmetry of the dynamical matrix enables the eigenstates to be orthogonalized such that
\begin{align}\label{symplectic_eigenstates_orthogonalized}
\langle \boldsymbol{\psi}_p |\cdot^\perp| \boldsymbol{\psi}_q \rangle &=
\sum_\alpha
\xoverline{\psi^R_{p,\alpha}} \, {\psi^R_{q, \alpha}}
- \xoverline{\psi^L_{p,\alpha}} \, {\psi^L_{q, \alpha}}
\\
&= \delta_{pq} \, \textrm{sign}(\omega_q),
\end{align}
where $\alpha$ runs over each gyroscope and $\omega_q$ is the oscillation frequency of $| \boldsymbol{\psi}_q \rangle$.
In a lattice, momentum space is periodic, just as the real space lattice configuration is a tiling of the unit cell, implying that the Brillouin zone is a torus.
Information about the connection of normal modes in each band is given by the Berry curvature, defined as
\begin{equation}
\vec{\Omega}_n = \nabla_{\mathbf{k}} \times i \langle \boldsymbol{\psi} | \nabla_{\mathbf{k}} | \boldsymbol{\psi} \rangle.
\end{equation}
This Berry curvature is akin to a magnetic vector potential in momentum space, affecting the motion of wavepackets as they traverse the system.
The integral of the Berry curvature over the Brillouin zone describes the topological obstruction to a continuous connection between states in the band, encoded in the Chern number:
\begin{equation}
\int_{BZ} \mathrm{d} \mathbf{S} \cdot \boldsymbol{\Omega}(\mathbf{k}) = 2\pi \, C.
\end{equation}
The integer $C$ is the Chern number of the band~\cite{thouless_quantized_1982}.
\Fig{fig:intro_berry_curvature} shows two schematic examples of a band colored by its Berry curvature.
In~\Fig{fig:intro_berry_curvature}a, the Berry curvature is nonzero, but its integral over the Brillouin zone vanishes: this is a trivial insulator with inversion symmetry breaking.
In~\Fig{fig:intro_berry_curvature}b, however, the contributions add, and the total value is a nonzero multiple of $2\pi$.
This nonzero Chern number signals the existence of the topologically protected chiral modes that live on the boundary of a system~\cite{thouless_quantized_1982}.
Topological protection ensures that, in the absence of available counter-propagating states, an edge wave will not scatter, passing around inclusions or voids and readily changing direction along jagged boundaries (\Fig{fig:intro_chern_number}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{./intro_berry_curvature_nopaths.png}
\caption{
\textbf{The Chern number is given by the integreated Berry curvature, which describes the connection between states in a band.}
\textit{(a)} A topologically trivial band may have nonzero Berry curvature, but its integrated Berry curvature vanishes.
\textit{(b)} For a band with Chern number of $C=1$, the integrated Berry curvature sums to $2\pi$.
}
\label{fig:intro_berry_curvature}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{./intro_chern_number_ver2.pdf}
\caption{\textbf{The Chern number counts the number of modes occupying a band gap.}
\textit{(a)} In a system with a single gap and a single boundary, the number of right-moving modes --- ie, sets of possible eigenstates with positive group velocity --- minus the number of left-moving modes is determined by the Chern number of a band.
Nonzero Chern numbers give rise to topologically protected chiral edge modes, which are robust against back-scattering.
\textit{(b)} A topologically trivial insulator does not support protected edge modes on the boundary.
}
\label{fig:intro_chern_number}
\end{figure}
An alternative representation of the Chern number uses the phase-invariant formula~\cite{avron_homotopy_1983}
\begin{equation}
\label{chern_num}
C_{j} = \frac{i}{2 \pi} \int \textrm{d}^2 k\ \mathrm{Tr}[\partial_{k_\alpha} P_j P_j \partial_{k_{\beta}} P_j \varepsilon^{\alpha \beta}],
\end{equation}
where $C_j$ is the Chern number of the $j^{\textrm{th}}$ band, $\varepsilon^{\alpha\beta}$ is the antisymmetric Levi-Civita symbol, and $P_j$ is the projection matrix defined for our system as
\begin{equation} \label{projector_symplectic}
P_j = | \psi_\alpha \rangle
\langle \psi_\alpha |
Q \, \textrm{sign}(\omega_j),
\end{equation}
where
\begin{equation}
Q = \left( \begin{array}{cc}
\mathbb{I}_{n} & 0 \\
0 & - \mathbb{I}_{n}
\end{array} \right) .
\end{equation}
The factors of $Q \, \textrm{sign}(\omega_j)$ arise from the symplectic structure of gyroscopes' equations of motion and the normalization of the states in~\Eq{symplectic_eigenstates_orthogonalized}.
If the symplectic symmetry were ignored, and the projector was instead defined using a simple outer product of bands which were not orthonormal, a non-physical distribution of Berry curvature would result, though this does not affect any of the computed Chern numbers.
By using the symplectic formulation, the correct Berry curvature distributions are readily obtained.
\begin{figure*}
\includegraphics[width=\textwidth]{kitaev_vs_bott_ver3.png}
\caption[]{
\textbf{The Kitaev sum and Bott index both measure band topology in real space.}
\textit{(a)} The Kitaev sum measures the band-integrated spectral flow at a particular location in the bulk of the material.
The sum converges to a target integer value corresponding to the Chern number of the system in the limit of an infinitely large system and large region of summation that does not enclose a boundary.
The blue curve shows the convergence of $\nu$ for increasing sizes of the summation region $\delta$.
The projector used for this calculation maps all states of the amorphous system that are above $\omega_c=2.2\Omega_g$ to themselves and maps states below $\omega_c$ to a zero vector.
\textit{(b)} Centering the summation region on different locations in an amorphous network in which all sites are trivalent (top) or in which two different local geometries are present (bottom) generates a spatially-resolved measurement of the local topological character of the material.
\textit{(c)} In electronic systems, the Kitaev sum measures the net charge that accumulates near the location of magnetic flux insertion (top panel)~\cite{mitchell_amorphous_2018}.
In gyroscopic and electronic systems alike, the contribution to the sum is dominated by sites close to the site of flux insertion, represented by the large blue markers at the center of the circular sample.
The size of each site's marker is proportional to the magnitude of the sum of terms in the Kitaev sum that include that site.
Blue markers indicate a negative net contribution, while red terms indicate a positive contribution.
An equal and opposite contribution resides on the boundary sites, mimicking the opposite charge that resides at the boundary in the electronic case, represented by the red markers at the edge of the circular sample.
\textit{(d)} The Bott index measures a topological obstruction preventing band-projected position operators from commuting. The index is a global measurement of band topology and requires closed boundary conditions.
We contrast the Kitaev sum with the Bott index by studying a small gyroscopic honeycomb network subjected to inversion symmetry breaking.
The Kitaev sum here has summation window set to include 45\% of the sites in the system.
The inset below in panel \textit{(d)} shows the pinning strength at each site (ranging from blue to orange) for the samples used in the calculation with open (periodic) boundary conditions for the Kitaev sum (Bott index).
The pinning strengths $\Omega_p$ of each site are disordered by adding an amount $V$ sampled from a Gaussian distribution with a standard deviation of $0.1\, \langle{\Omega_p}\rangle$, and the gyroscopes have pinning strengths biased by their sublattice site: $\Omega_p = \langle \Omega_p \rangle(1 \pm \Delta_{AB}) + V$.
}
\label{fig_kitaev}
\end{figure*}
\subsection{Real-space generalizations of Chern number}
If disorder is present in a lattice or if a gyroscopic network has amorphous structure, translational symmetry is broken, and the Chern number previously described cannot be defined due to the absence of a Brillouin zone.
However, several real-space generalizations to the Chern number have been proposed.
We discuss two approaches here.
We use both to investigate topological phase transitions, and one will even enable us to build an approximation to the Chern number using only the local character of the network.
\subsubsection{Kitaev sum}
The Kitaev sum generalizes the notion of Chern number to a real-space measure of chirality integrated over a band in the bulk of a 2D material~\cite{kitaev_anyons_2006,mitchell_amorphous_2018,ma_hamiltonian_2017}.
By defining a projection operator, $P$, mapping states above a cutoff frequency, $\omega_c$, to themselves and states below the cutoff to zero, we obtain a real-valued number $\nu$ via
\begin{equation}
\nu(P) = 12 \pi i \sum_{j \in A} \sum_{k \in B} \sum_{l \in C}
\left(
P_{jk} P_{kl} P_{lj} - P_{jl} P_{lk} P_{kj}
\right).
\end{equation}
The projection operator could instead be defined to map states \textit{below} the cutoff $\omega_c$ to themselves, as is often done in electronic systems; in this case, $\nu \rightarrow -\nu$.
When negative frequency bands are included, it is important to include the symplectic factors in~\Eq{projector_symplectic}.
$A$, $B$, and $C$ are three non-overlapping summation regions defined in a counterclockwise fashion as shown by the red, green, and blue regions of~\Fig{fig_kitaev}a and c, and these can be of any shape.
In electronic systems, the Kitaev sum measures the charge that accumulates for a localized magnetic field flux inserted into the bulk position where the three regions $A$, $B$, and $C$ meet.
For a single magnetic flux quantum, the amount of charge is $\nu e$.
In our gyroscopic system, $\nu$ similarly relates directly to flux pumping, signifying the number of states exchanged between the boundary and the site of flux insertion shown in~\Fig{fig_bulkboundary}a~\cite{mitchell_amorphous_2018}.
One feature of the Kitaev sum is that the result is not strictly integer-valued.
In the limit that the region of summation encloses many gyroscopes without enclosing the material's boundary, however, the value of $\nu$ does converge towards an integer.
This non-integer character of $\nu$ is useful in that the rate of convergence indicates the localization length of modes at the cutoff frequency: $\nu$ approximates its target value once the region of summation is larger than the localization length.
\subsubsection{Bott Index}
An alternative approach which rests on $K-$theoretic methods is to measure what has become known as the Bott index~\cite{loring_disordered_2010}.
Like the Chern number, a nonzero Bott index signals the impossibility of finding a complete, orthonormal basis of localized functions (`Wannier states') among the states included in the projection operator.
Unlike the Kitaev index, the Bott index requires the construction of a periodic system, so translational invariance is artificially restored on some scale, though this scale can be made arbitrarily large.
Given a fully periodic sample of spatial extent $W$ in both spatial dimensions $X$ and $Y$, the positions of each gyroscope are indexed by $\Theta = 2\pi X / W$ and $\Phi = 2\pi Y / W$.
Equipped with a projection operator $P$ as before, we define band-projected position matrices
\begin{equation}
P e^{i \Theta}P \sim \left(
\begin{array}{cc}
0 & 0 \\
0 & U
\end{array}
\right)
\hspace{30pt}
Pe^{i \Phi} P \sim \left(
\begin{array}{cc}
0 & 0 \\
0 & V
\end{array}
\right),
\end{equation}
with the right hand sides expressed in the basis of eigenmodes.
If the cutoff frequency of $P$ lies in a mobility gap, the block nonzero components $U$ and $V$ are almost unitary and will almost commute~\cite{loring_disordered_2010}.
The Bott index, $B$, defined as
\begin{equation}\label{eqn_bott_index}
B = \frac{\Im \left[ \textrm{Tr} (\log(VUV^\dagger U^\dagger)) \right]}{2\pi i},
\end{equation}
indicates if $U$ and $V$ are `close' to a different pair of matrices which are exactly unitary and exactly commute.
Since the logarithm is multivalued,~\Eq{eqn_bott_index} must be defined more precisely: we evaluate the sum of the logarithm of the eigvenvalues of $VUV^\dagger U^\dagger$ and take the imaginary components to be as small as possible.
A potential advantage of the Bott index over the Kitaev sum is that it returns an exact integer, rather than converging towards an integer value, at the cost of requiring the introduction of (potentially artificial) periodicity to the sample.
A potential disadvantage is that the Bott index measurement is global: a given sample returns a single value, whereas the Kitaev sum can be performed in a spatially-resolved manner (\Fig{fig_bulkboundary}b and~\Fig{fig_kitaev}a-c).
For a system composed of multiple patches of different topological phases, such as in the lower panel of~\Fig{fig_kitaev}b, the Bott index selects only a single global readout.
\subsection{Determining the Chern number by Local Approximations}
Given that such small systems can still be used to accurately compute real-space topological invariants and exhibit bulk-boundary correspondence, it is tempting to think that we can use the local properties of a network to approximate the topological invariant, under the assumption that the region is not exceptional.
Here, we will build on Kitaev's insight to address this challenge for systems that satisfy the same conditions as we required in section~\ref{section:understanding_gap_in_regime}: namely, that (1) some off-diagonal elements of $D$ are large compared to on-diagonal elements, and (2) the square of the dynamical matrix is nearly diagonal such that $D^2 \approx \omega_0^2 \mathds{1} + \epsilon$.
The goal is to approximate the projection operator, which is short-ranged~\cite{kohn_theory_1964,hastings_lieb-schultz-mattis_2004}.
This is a critical assumption behind the classification of interacting topological phases, which -- like amorphous materials -- are studied in real-space rather than momentum space.
To begin, we use an expression we derived in~\cite{mitchell_amorphous_2018} that casts Kitaev's formula into
\begin{equation}\label{eq:kitaev_mod}
\nu =\mathrm{Im\ } 8\pi\sum_{ijk} P_{ij}P_{jk}P_{ki},
\end{equation}
where $P$ is the projection onto states below a given frequency, and the sum is over all sets of three points in the network that form a triangle surrounding a given point $q$ in a counterclockwise orientation.
This formula can be derived as the charge that accumulates when a magnetic flux quantum is threaded through $q$, and this expression translates directly into Kitaev's (and vice versa) if $P$ is the exact projector.
However, we will use an approximation to the projector, and~\Eq{eq:kitaev_mod} seems more reliable in the situation when $P$ is known only approximately.
Kitaev's formula requires a choice of subregions $A$,$B$,$C$ that can change the value of the sum if $P$ is merely approximated, but our expression requires no such choice.
\begin{figure}
\includegraphics[width=\columnwidth]{ari_fig2_fig3_ver3_convergence.png}
\caption{
\textbf{The Chern number can be approximated by a purely local, real-space calculation.}
\textit{(a)} In the Hofstadter model, there are two three-hopping triangles formed from nonzero entries in $D$ that encircle $q$, which are needed to calculate the lowest-order approximation to the Chern number.
The two triangles are 123 and 124.
\textit{(b)} The approximation to the Chern number (\Eq{eq:kitaev_mod} and \ref{eq:projector_series_expansion}) converges toward the true value of $\nu=1$ as the number of terms included from the expansion of~\Eq{eq:projector_series_expansion} grows.
The thin green curve corresponds to the first term, given by~\Eq{eq:approx_chern}, and thicker lines varying from green to blue indicate the result from including two, three, four, five, or six terms from~\Eq{eq:projector_series_expansion}.
The measurement is shown here as a function of the complex, next-nearest-neighbor hopping strength $t'/t$ (yellow arrows in panel a).
The dashed, vertical gray line indicates the value of the ratio $t'/t = {1}/\sqrt{2}$, where the approximation should be the most accurate.
Above $t'>t$, the series does not converge for our choice of $\omega_0$ in $D^2 = \omega_0^2\mathds{1} + \epsilon$.
Note that the single-term approximation (thin green curve) reaches its maximum at this coupling strength, and further terms improve the result near $t'/t = {1}/\sqrt{2}$.
\label{fig:approximate_chern}}
\end{figure}
The projection operator (or `projector') in a coordinate system where $D$ is diagonal is $\sum_{\omega_n<\omega} |n\rangle\langle n|$, where $\omega_n$ are the frequency eigenvalues and $|n\rangle$ the eigenvectors.
This is an inherently nonlocal object, but we can overcome this difficulty by writing the sum in an algebraic form as $\sum_{\omega_n} \frac{1}{2}(1- \omega_n/\sqrt{\omega_n^2})|n\rangle\langle n|$, where the square root is the positive square root.
Recast in a coordinate-independent way, $P = \frac{1}{2}(\mathds{1}- D / \sqrt{D^2} )$ which can be expanded using the assumption $D^2=\omega_0^2 \mathds{1} +\epsilon$, where $\epsilon$ is small, as
\begin{equation}\label{eq:projector_series_expansion}
P=\frac{1}{2}\mathds{1}-\frac{D}{2\omega_0}\sum_{n=0}^\infty\left(-\frac{\epsilon}{4\omega_0^2}\right)^n \binom{2n}{n}.
\end{equation}
Here, we have used the Taylor series $\frac{1}{\sqrt{1-x}}=\sum_{n=0}^\infty \binom{2n}{n} (x/4)^n$.
Substituting into~\Eq{eq:kitaev_mod}, we note first that the first term ($\frac{1}{2}\mathds{1}$) and its cross-terms with the other factors contribute only real terms to the
sum because $D$ is Hermitian and can therefore be neglected.
Only the sum remains.
If the $n=0$ term dominates, then
\begin{equation}\label{eq:local_chern}
\nu \approx\mathrm{Im} \frac{\pi}{\omega_0^3}\sum_{ijk} D_{ij}D_{jk}D_{ki}.
\end{equation}
This is a local formula for the Chern number.
Since $D_{ij}$ is short-ranged, the number of triangles $ijk$ enclosing a point that contribute nonzero terms is limited.
A simple model both provides intuition and tests the accuracy of this result.
For the Hofstadter model we considered before, which is associated with a gyroscopic network in the weakly-interacting limit $\Omega_k/ \Omega_p \ll 1$, $\nu$ is proportional to the product of the magnitudes of the couplings around the sides of a triangle times the sine of the net phase factor, normalized by $\omega_0^3$.
Intuitively, $\omega_0$ is akin to a mean value of the coupling magnitudes, so this cancels off much of the dependence on the overall magnitude of the couplings, leaving mainly the dependence on the phases.
For the square model, take $q$ to be the point shown in~\Fig{fig:approximate_chern}a.
There are only two triangles surrounding $q$, which each contribute $it^2 t'$.
They are each counted three times, since the three vertices $ijk$ can be enumerated with any cyclic permutation while retaining the proper counter-clockwise ordering, so
\begin{equation}\label{eq:approx_chern}
\nu \approx \frac{3\pi t^2 t'}{4(t^2+t^{'2})^{\frac32}}.
\end{equation}
A true topological invariant should be quantized, but clearly this expression is not.
We expect the approximation to work best when $t'$ is close to $t/\sqrt{2}$ so that the next-nearest-neighbor entries in $D^2$ cancel.
At this point, the approximation is $\pi/ (2\sqrt{3})\approx .9$, which is close to 1.
A plot of the approximate Chern number against a range of complex hopping magnitudes $t'$ in~\Fig{fig:approximate_chern} is nearly constant for a range of values of $t'$ near $t/\sqrt{2}$.
Including more terms in the sum in~\Eq{eq:local_chern} increases the accuracy of the Chern measurement in the vecinity of $t'=t/\sqrt{2}$.
As discussed in Appendix~\ref{appendix:generalize_chern}, convergence of the series requires $\omega_0$ to be chosen so that $\omega_0^2$ is at least as large as half the maximum eigenvalue of $D^2$.
The method discussed here for computing a local approximation to $\nu$ generalizes to any system with a gap, regardless of whether $D^2$ is nearly diagonal, as we show in Appendix~\ref{appendix:generalize_chern}.
This approximation to the real-space Chern invariant does not require a full diagonalization of the dynamical matrix.
This result expands our computational machinery not only to predict if a gap will appear (as derived before), but also to predict if that gap will be endowed with chiral edge modes.
\begin{figure}
\includegraphics[width=\columnwidth]{patch_coupling_homogeneity_v2.png}
\caption[]{
\textbf{In order for patches of Chern insulator to constitute a topological sample, the patches must be sufficiently coupled}.
By varying the spring constants connecting initially uncoupled regions of a gyroscopic network, we find that a threshold coupling strength is required to create a macroscopically topological sample.
Here, we used a $21\times21$-cell gyroscopic honeycomb configuration with $\Omega_g = \Omega_k$ and open boundary conditions.
On the right side, two example eigenmodes with $\Omega'_k =0.2\Omega_k$ and $0.8 \Omega_k$ show the change in spatial structure of the excitations nearest $\omega = 2.25 \Omega_p$ for patch sizes $w/\ell=7.5$, where $\ell$ is the bond length.
Below a critical patch coupling strength, edge modes span the boundaries between patches throughout the material, whereas above a critical patch coupling strength, a single edge mode resides on the outer rim of the network.
The localization length $\xi$ for a mode in the middle of the band gap is on the order of the bond length: $\xi\approx \ell$, with the precise value depending on the orientation of the boundary with respect to the lattice.
So long as the patch size is greater than $\sim 2 \xi$, the transition from trivial to topological occurs at a fixed coupling strength.
}
\label{fig_homogeneity}
\end{figure}
\subsection{Mesoscopy vs homogeneity in determining Chern number}\label{subsection:mesoscopy_vs_homogeneity}
We have found that even in relatively small systems, a topological invariant can be readily approximated by the Kitaev sum without periodic boundary conditions.
This demonstrates that band topology can be encoded on a mesoscopic scale -- far smaller of a scale than has been traditionally investigated.
If multiple topological patches are adjacent but uncoupled, then each patch will have its own edge states, but the system as a whole will not behave as a Chern insulator, since these states are distributed throughout the bulk.
Since the `edge modes' of each patch occupy a frequency range that corresponds to the gap of a fully connected system, the material as a whole does not behave as an insulator.
What degree of coupling between patches is required for these patches to behave as a single insulating system with chiral modes only at its outer boundary, rather than throughout its interior?
Here we study the transition of weakly coupled mesoscopic patches of Chern insulator into a single unit of material.
\Fig{fig_homogeneity} shows patches of gyroscopic honeycomb networks bonded together by spring coupling with variable strength (dashed bonds in~\Fig{fig_homogeneity}).
With these bonds completely floppy ($\Omega_k' \ll \Omega_k$), the composite system is a trivial insulator, with a real-space Chern number measurement (here using the Kitaev sum) of $\nu \rightarrow 0$ (gray circles in the phase diagram of~\Fig{fig_homogeneity}).
If the coupling is strengthened, however, the material transitions to a Chern insulator at $\Omega_k' \approx 0.4 \Omega_k$.
Notably, so long as the patches are sufficiently large that the edge modes localized to their boundaries do not span the entire patch, the value of $\Omega_k'$ at the transition does not vary with patch size.
\Fig{fig_homogeneity} shows the associated phase diagram: for cell sizes larger than $w \gtrsim 2 \ell$, the transition from trivial ($\nu = 0$) to topological ($\nu = -1$) occurs at a constant coupling strength.
This is an extension of our earlier observations in~\Fig{fig:spectrumflow} in which varying the coupling between unit cells led to a topological phase transition at a critical coupling strength.
There, a sufficiently strong coupling ($\Omega_k' \approx 0.4 \Omega_k$) was necessary to transform a trivial insulator of bonding/anti-bonding states into the delocalized states of a Chern insulator.
Here, in addition, we find that even if large patches individually register as topological, a significant coupling is still required for the sample as a whole to have a nonzero topological invariant.
Below a critical coupling, the conductance of this gyroscopic material is not insulating on account of interior interfaces between insulating grains.
This highlights the need for homogeneity in the coupling for a topological material to insulate throughout its bulk
and highlights one of the ways in which topological insulation can break down in the presence of heterogeneity.
In the process of establishing a local origin of topological band gaps, we encountered the importance of homogeneity in the strengths of bonds throughout the bulk.
Violating this assumption of homogeneity in couplings can change a material's topological phase.
We now turn to the other way of breaking homogeneity in the bulk: modulating pinning strengths at each site.
Consistent with our findings so far, we find that whereas inhomogeneity can break topological insulation, spatial order is utterly irrelevant to topological order.
Indeed, in the final section of this article, we find that even in their response to on-site disorder, amorphous structures behave like their crystalline counterparts and undergo localization transitions consistent with the same scalings as in ordered crystals.
\section{Topology and Anderson localization in gyroscopic metamaterials}
Topological insulators are famous for the chiral waves living at a system's edge at frequencies in the gap.
However, there are also differences in the bulk modes at frequencies in bands bounding the gap.
These modes can conduct energy in phononic metamaterials, where there is no Pauli exclusion and all modes are available excitations.
For topologically trivial materials in 2D, any disorder is sufficient to localize all modes in a large enough system.
For Chern insulators, on the other hand, some normal modes in bands will be spatially extended across the entire system~\cite{huo_current_1992}.
The study of how states localize as disorder increases or as the system size grows -- so-called `Anderson localization' -- has provided a powerful framework to understand and predict the conductivity of a wide range of materials~\cite{anderson_absence_1958,abrahams_scaling_1979,Lagendijk_Fifty_2009}.
Recent extensions incorporate topological order into this scheme~\cite{xue_quantum_2013}.
How do \textit{amorphous} Chern insulators behave when subjected to strong disorder?
In the first half of this section, we study the scaling of the topological index with system size, with particular attention to the amorphous case.
By varying the disorder, we find behavior consistent with the expected scaling for symmetry class A, including annihilation of extended states in the mobility gap at the transition.
We then use inversion symmetry breaking to add a second dimension to this phase transition and find an avenue for disorder to drive a trivial insulating phase into a topological one.
This re-entrant topological behavior, dubbed `topological Anderson insulation', is surprising, since typically disorder drives systems away from the topological phase~\cite{groth_theory_2009,stutzer_photonic_2018,meier_observation_2018}.
Using experiments and simulations, we identify this interesting feature of the interplay between disorder and topology in the mechanical context for the first time.
\begin{figure}
\includegraphics[]{fig_levitation_annihilation.png}
\caption[]{
\textbf{Increasing disorder strength drives extended states towards the gap, where they annihilate as the system passes from topological to a trivial insulator.
}
\textit{(a)} Illustration of the annihilation of extended states (yellow states) in the gap as pinning disorder strength $V$ increases, as shown in a schematic cartoon (upper panel) and in computed spectra for a collection of 20 disordered amorphous gyroscopic networks of 450 gyroscopes (lower panel) with $\Omega_k = \langle\Omega_p \rangle$.
The colors of these spectra reflect the localization of states, $\xi$, measured in units of inverse average bond length.
These extended states (yellow) carry the Hall conductance for the system, so their annihilation triggers a transition to a trivial insulating phase, in which all modes are localized (purple).
\textit{(b)} As disorder grows, the localization length in the gap grows to the system size, aligning with the change in average topological invariant of the top band from -1 to 0. The vertical dashed line marks the disorder strength at which the topological transition occurs.
}
\label{fig_levannihilation}
\end{figure}
\subsection{Anderson Insulator transition}
Strong disorder in pinning strengths drive a transition to the trivial insulating regime.
In the canonical Anderson localization picture without any topological effects, the addition of infinitesimal disorder localizes all modes in a 2D system.
In the presence of nontrivial topology, however, stronger disorder strengths are required to drive Chern insulators into the trivial insulating state.
For small disorder, each band acquires tails of localized states, and bands carrying Hall conductance are confined to a region near the center of each band (leftmost panels in~\Fig{fig_levannihilation}a).
As the disorder strength grows, these states approach and annihilate in a frequency region which was previously a mobility gap.
This process is called `annihilation and levitation' due to the annihilation of the extended states in the gap and the subsequent rise in variance in the spacing between eigenfrequencies.
In \Fig{fig_levannihilation}, we show that amorphous gyroscopic networks display this behavior as well.
\Fig{fig_levannihilation}a shows this annihilation of extended states for
amorphous networks of 450 gyroscopes.
Without pinning disorder, the two bands of states which span the system are separated by a mobility gap.
As we increase random pinning frequencies $\Omega_p$ drawn from a Gaussian distribution with width $V$, however, the extended states move toward the gap, meet in the frequency range which was previously the middle of the gap, and annihilate, shown in~\Fig{fig_levannihilation}a.
\Fig{fig_levannihilation}b shows that the localization length in the middle of the gap ($\omega \approx \Omega_p + 1.25 \Omega_k$) rises as the extended states invade.
The localization length peaks at a value near the system size, then falls as disorder dominates.
The peak in localization length coincides precisely with the change in the ensemble-averaged topological index -- whether Kitaev sum or Bott index -- from nonzero to zero.
Thus, the annihilation of the extended states mark the transition from a topological to trivial phase.
By measuring the topological invariant for systems of different sizes, we find sharper and sharper transitions to the trivial phase as the system size increases.
Intuitively, a disordered topological system whose size is comparable to the localization length will be susceptible to random variations in disorder such that different realizations of disorder may yield a trivial phase.
Near the transition, where the localization length of modes in the center of the gap is large, proportionally larger systems are required to suppress fluctuations in the resulting topological invariant.
This is the case both for amorphous structures shown in~\Fig{fig_levannihilation} and~\ref{fig_scaling}, as well as for the honeycomb lattice we studied earlier.
If the transition is perfectly sharp in the limit of an infinitely large system, then the system with disorder strength $V < V_c$ is a Chern insulator supporting topologically protected chiral edge modes, while systems with $V > V_c$ will register $\langle \nu \rangle \rightarrow 0$ and $\langle B \rangle \rightarrow 0$.
In~\Fig{fig_scaling}, we compute the topological invariants for thousands of amorphous systems with different sizes and disorder strengths.
After tracing out the average topological invariant as a function of disorder for many system sizes, we collapse these curves by rescaling the disorder strength according to
\begin{equation}\label{eq_finsize_scaling}
\tilde{V} = V_c + (V - V_c) \left( \frac{L}{L_0} \right)^{1/\tilde{\nu}}.
\end{equation}
Here $L$ is the system size, $L_0$ is an intermediate size, $V_c$ is a critical disorder strength at which all curves should in principle intersect, and $\tilde{\nu}$ is the critical scaling exponent.
\Fig{fig_scaling} shows that both the Kitaev sum (top panels) and Bott index (bottom panels) collapse under finite-size rescaling of Eqn~\ref{eq_finsize_scaling} with the same values $V_c / \langle\Omega_p \rangle = 0.63 \pm 0.04$, and scaling exponent, $\tilde{\nu} \approx 2.6$, though small samples with fewer than $450$ gyroscopes deviate visibly from this scaling due to boundary effects in the Kitaev sum.
This value is consistent with reported scaling exponents of $\tilde{\nu} = 2.58\pm 0.03$ found in previous studies of quantum tight binding models~\cite{xue_quantum_2013}.
Notably, we obtain reasonable fits for a surprisingly wide range of critical exponents $\tilde{\nu} = 2.2-3.0$, and a $\chi^2$ analysis constrains our value only within an uncertainty of $\Delta\tilde{\nu} \approx 1$.
Nonetheless, the scaling collapse suggests that the transition is, in fact, infinitely sharp in the limit of large system sizes, with finite size scaling consistent with the universality class of Cartan label A~\cite{barlas_topological_2018}.
This result is consistent with the idea that amorphous Chern insulators share similar scaling behavior with tight binding models on lattices endowed with random pinning disorder~\cite{huo_current_1992,hung_disorder_2016}.
Our system is also different from the traditional crystalline case in another way in that gyroscopic metamaterials are symplectic rather than Hermitian.
It is interesting to note that neither difference appears to break the standard universal scaling~\cite{xue_quantum_2013}.
While two recent studies have found non-universal behavior in amorphous systems~\cite{ivaki_criticality_2020,sahlberg_topological_2020}, which motivates continued study of the scaling behavior along other directions of the phase boundary, we find no inconsistency with the scaling dictated by the symmetry class of the dynamical matrix.
\begin{figure*}[ht]
\includegraphics[width=\textwidth]{fig_scaling_Chernamorphous_transposev3.png}
\caption[]{
\textbf{Scaling collapse of the Anderson localization transition in our gyroscopic Chern insulators is consistent with the universality class of Cartan label A.
}
As disorder is increased, the topological index of the upper band averaged over many samples smoothly varies from -1 to 0.
The spatial structures of networks used here (left column) are Voronoi tessellations of hyperuniform point sets generated as in Ref.~\cite{mitchell_amorphous_2018}, with a gyroscope placed at every vertex and bonds lining the polygons of the tessellation.
As the size of the system for which the index is computed grows (blue to black to red in each panel), the transition grows sharper.
Each data point represents the average Kitaev sum $\langle \nu \rangle$ or Bott index $\langle B \rangle$ for 100 amorphous gyroscopic networks (ten disorder realizations for each of ten randomly-generated amorphous structures) subjected to on-site disorder strengths modulating $\Omega_p$ at each site by an amount taken from a Gaussian distribution.
The pinning disorder strength is measured in units of $\langle\Omega_p\rangle$, and we let $\Omega_k=\langle\Omega_p\rangle$.
Rescaling the disorder values according to a power law of the system size collapses the curves with an exponent $\tilde{\nu} = 2.6$ and $V_c = 0.63 \, \langle \Omega_p \rangle$, as shown in the right panels.
}
\label{fig_scaling}
\end{figure*}
\subsection{Topological Anderson Insulator phase diagram}
We have seen that strong disorder destroys topological behavior.
By tuning an additional parameter, however, we find the remarkable possibility of transforming a trivial insulator into a Chern insulator simply by adding disorder to the system.
Recently, this re-entrant phase diagram of so-called `topological Anderson insulation' was demonstrated for the first time in photonic~\cite{stutzer_photonic_2018} and cold atom systems~\cite{meier_observation_2018}.
As shown in~\Fig{fig_TAI}, gyroscopic metamaterials support this same transition.
Here, we simultaneously vary both random disorder strength and inversion symmetry breaking in a honeycomb lattice, both in experiment and in numerical calculations.
We introduce disorder by changing the gyroscope spinning speeds, which changes both pinning frequencies and bond strengths simultaneously due to the dependence of each on spinning speed.
Meanwhile, inversion symmetry is broken by a periodic array of staggered magnetic fields that splits the on-site precession frequencies of the two sublattice sites.
\Fig{fig_TAI}c shows the resulting topological index, here shown by the Bott index averaged over 200 realizations for each value of disorder.
Increasing either inversion symmetry breaking or disorder strength drives the gyroscopic lattice into the trivial phase, but increasing both simultaneously allows for the topological phase to persist for significantly larger values of inversion symmetry breaking.
This enables a re-entrant phase transition by increasing disorder in a system with large inversion symmetry breaking.
Using the Kitaev sum in place of the Bott index gives the same result, aside from weakened convergence near the phase boundaries.
In~\Fig{fig_TAI}c, we choose $\Omega_k = 0.67 \Omega_p$ to match the relative values in our experiment, and also choose the disorder to affect both pinning and interaction strengths proportionally, as would occur in an experiment.
The qualitative features of the resulting re-entrant phase transition, however, are indifferent to modest variations of these parameters, to the shape of the disorder distribution (flat versus Gaussian), and to whether or not interaction strength disorder is varied in tandem with pinning frequencies or instead of pinning frequencies.
\Fig{fig_TAI}d shows complementary measurements made in an experimental setup of 54 gyroscopes coupled with repulsive magnetic interactions.
Note that with magnetic interactions, the equations of motion are similar to the spring case considered thus far, as shown in Appendix~\ref{appendix:magnetic_interactions}.
The edge of the sample is shaken at a slowly varying frequency that spans the band gap range of the clean ($V=0$) and inversion symmetric ($\Delta=0$) system.
Taking the Fourier transform of the tracked gyroscope displacements extracts the normal modes, from which we measure the localization length by fitting the displacement amplitudes to a decaying exponential as a function of distance from the edge.
The average localization length $\xi$ of modes in a narrow range of frequencies that lies in the gap of the clean system is reported as a fraction of the system width, $L$.
As the system approaches the boundary between topological and trivial phases, the localization length grows (gray dots in~\Fig{fig_TAI}d).
In principle, the localization length should decrease again far from the phase boundary, though this subsequent decrease is not visible in our experiment except at low disorder.
In the experiments, the inversion symmetry is broken using a staggered array of magnetic coils, and the disorder strength is controlled by setting the spinning speeds of individual gyroscopes using a chain of microcontrollers (~\Fig{fig_TAI}b).
In our experimental setup, we set all experiments to have $\omega_0=175$ Hz spinning frequency, and we vary the width of the distribution using a pulse width modulation setup.
Briefly, we first set all gyroscopes to have identical spinning frequencies by iteratively adjusting the duty cycle (i.e., fraction of time that a gyroscope is receiving power) assigned to individual gyroscopes using a pulse width modulation controller.
This step cancels out initial disorder from variations in the construction of our 3D-printed and hand-assembled gyroscopes (Stratasys Objet 350 printer).
We then generate a target spinning speed assignment for all gyroscopes so that the precession frequencies will approximately follow a flat distribution.
We first attempt to achieve this assignment by perturbing the duty cycle values according to an average calibration relating duty cycle to spinning speed.
We then iteratively measure the spinning speeds with a high-speed camera (Vision Research Phantom v12.1) and adjust the duty cycle values to approach the target spinning speed assignment.
There are several differences between the experimental realization and the idealized model.
For large disorder strengths, some gyroscopes are set to spin very slowly.
For these slowly spinning gyroscopes, nutation and nonlinear effects play a role in the dynamics, driving the system away from the fast-spinning-limit assumption.
Additionally, the magnetic interactions are long range in the experiment.
We choose to suppress both of these complications in the model shown in~\Fig{fig_TAI}c to underscore the simplicity of the phenomenon.
We use the localization length as an indication of the phase transition rather than measuring the Chern numbers of the phases themselves.
While this is an indirect measure of topology, the results capture the essential features of the expected phase diagram.
\begin{figure*}
\includegraphics[width=\textwidth]{figure_tai_idraw_202006.png}
\caption[]{
\textbf{
Gyroscopic networks exhibit a topological Anderson insulator (TAI) transition wherein a trivial insulator phase is driven into the topological phase by adding disorder.}
\textit{(a)} In our experiments, each gyroscope is constructed from an externally powered motor encased in a 3D printed housing that minimizes damping.
A magnetic coil placed beneath each site allows variation of the on-site precession frequency $\Omega_p\rightarrow \Omega_p \pm \Delta_{AB}$, with each sublattice site (orange and blue) receiving an opposite bias from the magnetic torque.
\textit{(b)} Daisy-chained microcontrollers (TLCs) convert a serial output into a site-specific spinning speed via pulse width modulation, while a separate DC power supply drives inversion symmetry breaking by passing current through each coil.
\textit{(c)} A 2D topological phase diagram shows that either breaking inversion symmetry (increasing $\Delta$) or introducing strong disorder (increasing $V$) drives a transition from topologically nontrivial ($\langle B \rangle \ne 0$) to trivial ($\langle B \rangle =0$).
Each colored region represents the average Bott index computed numerically for 200 random disorder configurations with nearest-neighbor interactions only.
The pinning frequencies are chosen from a flat distribution of disorder of width $V = [ \max(\Omega_p) - \min({\Omega_p})] / \Omega_p^0$, and magnetic interaction disorder covaries with the pinning frequency, as would occur in the experiments.
\textit{(d)} An experiment designed to realize the topological Anderson insulator transition in 54 gyroscopes measures the localization length of normal modes excited in a frequency range within the band gap of the clean system ($\Delta=0$, $V=0$).
As either disorder strength or inversion symmetry breaking is increased, the localization length grows on average, reaching or exceeding the system size, $L$.
Disorder is introduced by controlling the spinning frequency of each gyroscope individually in a spatially uncorrelated fashion, which varies both pinning and interaction disorder strength.
The orange line is a guide to the eye taken from simulations.
Though the experimental system exhibits many differences from the numerical idealization, including long range interactions and strong nutation at low spinning speeds, the qualitative features of the transition are evident.
}
\label{fig_TAI}
\end{figure*}
\section{Conclusion}
Periodic order is dispensable for generating topology in gyroscopic metamaterials.
In light of this observation, by re-examining the fundamental dynamics of coupled gyroscopes, we have identified intuitive, real-space descriptions for the origins of topological gaps.
The signatures of time reversal symmetry breaking -- an essential ingredient in endowing bands with non-trivial topology -- appear even at the three-gyroscope level through the interaction of counter-rotating polarizations.
The three-way interaction between TRS breaking, network geometry, and coupled polarizations underlies the emergence of topological gaps, and these nontrivial gaps are topologically disconnected from bonding/anti-bonding energy splitting.
We present a method to predict the existence of a band gap without full diagonalization of the dynamical matrix that generalizes to amorphous structures.
In the same spirit of simplifying the predictive machinery for topological phases irrespective of spatial order, we approximate the topological index using only local properties of a gyroscopic network, in contrast to other real-space methods in wide use.
The ability to do so reflects the ability of even mesoscopic patches of gyroscopes to register as belonging to a topologically nontrivial phase, further supporting the notion that periodicity is not needed for nontrivial topology.
We find that disorder transitions in both amorphous and crystalline gyroscopic networks have similar behavior to periodic electronic and photonic Chern insulators.
Amorphous networks display finite-size scaling that is consistent with the universal behavior from crystalline Chern insulators when subjected to pinning disorder on top of their existing structural disorder.
Finally, we demonstrate a re-entrant topological phase diagram in a mechanical context for the first time.
\section{Acknowledgements}
Lisa M Nash contributed to this work via experimental design, gyroscope fabrication, conceptual insights, and useful discussions.
We thank Emil Prodan for his extensive insight and Nikolas Claussen for a careful reading of the manuscript and helpful comments.
We also thank Anton Souslov, Alex Edelman, Emmanuel Stamou, Tom Witten, and Daniel Hexner for useful discussions.
This work was supported in part by the University of Chicago Materials Research Science and Engineering Center, which is funded by National Science Foundation under award number DMR-1420709.
This research was supported in part by the National Science Foundation under Grant No.~NSF PHY-1748958.
NPM acknowledges support from the Helen Hay Whitney Foundation.
Additional support was provided by the Packard Foundation and by NSF EFRI NewLAW grant 1741685.
| 23518c350917b21313e7dd12751ff48a81580ff1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{s:intro} Let $G$ be a finite group. An elementary abelian subgroup $E$ of
$G$ is a called a broad subgroup if every involution of $G$ is conjugate to an element of $E$. This definition is motivated in part
by the fact that (using a Theorem of R.Kn\"orr \cite{K}) an irreducible character which vanishes identically on non-identity elements of $E$ lies
in a $2$-block of defect zero. This in turn allows us to prove that when $G$ contains a broad subgroup, the number of irreducible characters of $G$
which lie in $2$-blocks of positive defect of $G$ is at most $|C_{G}(t)|$ for some (non-trivial) involution $t \in G$. Our main result (which depends upon
the classification of finite simple groups) is:
\begin{theorem} \label{t:main} Let $G$ be a finite quasi-simple group. Then $G$ contains a broad subgroup.
\end{theorem}
Notice that it is immediate that a direct product of finite groups has a broad subgroup if each direct factor has a broad subgroup.
We have the following corollary, which may be of independent interest:
\begin{corol} Let $G$ be a finite quasi-simple group. If $x$ and $y$ are involutions of
$G$, then $x$ commutes with some conjugate of $y$.
\end{corol}
The promised application to character theory is provided by:
\begin{corol} \label{c:positive}
Let $G$ be a finite group with no non-trivial normal subgroup of odd order. Then there is a non-trivial involution $t \in F^{\ast}(G)$ such that the total number of irreducible characters of $G$ which do not lie over $2$-blocks of defect zero of $F^{\ast}(G)$ is at most $|C_{G}(t)|$.
\end{corol}
This applies in particular when $G$ is a non-Abelian simple group, and since $G = {\rm SL}(2,2^n)$ (for $n \geq 2)$ has exactly $2^{n}$ irreducible characters which lie in $2$-blocks of positive defect,
while $|C_{G}(t)| = 2^{n}$ for each non-trivial involution $t \in G$, the inequality of Corollary 3 may be sharp.
If $G$ is quasisimple and the center has even order, then there is a central element of order $2$ and so the corollary
is not useful.
Note that there is no direct analog of the theorem for odd primes. For example if $p \ge 5$, let $S=\operatorname{SL}_n(p)$ with
$4 \le n \le p$. Since $Z(S)$ has order prime to $p$, it suffices to produce an example in $S$.
Let $x$ be a regular unipotent element (i.e. $x$ has a single Jordan block of size $n$). Note that
$x$ has order $p$.
Then the centralizer of $x$ is an elementary abelian $p$-group of rank $n-1$. It is straightforward
to check that the only elements $y \in C_S(x)$ which have two Jordan blocks have Jordan blocks
of sizes $(n \pm 1)/2$ if $n$ is odd and two Jordan blocks of size $n/2$ if $n$ is even. Thus, $x$ does
not commute with an element of order $p$ that has Jordan blocks of size $n-1$ and $1$.
For $p=3$, we give an example where elements of order $3$ are semisimple. Let $S=\operatorname{SL}_9(q)$
with $q \equiv 4 \mod 9$. Then $Z= Z(S) = \langle z \rangle$ has order $3$. Let $x \in S$
satisfy $x^3=z$. So in $G=S/Z$, $xZ$ is an element of order $3$. Let $y \in S$ be an element
of order $3$ with a $7$ dimensional eigenspace with eigenvalue $1$. We claim that
no conjugate of $y$ commutes with $x$ modulo $Z$. Suppose there is a conjugate $y'$ of $y$
commuting with $x$ modulo $Z$. Then replacing $y'$ by $y$, we see that $[x,y] = u \in Z$. Then
$y$ is conjugate to $yu$ whence by considering eigenvalues, $u=1$. So $x$ and $y$ commute
in $S$. Since $y$ has a $1$-dimensional eigenspace, $x$ must preserve this space. However,
$x$ has no eigenvalues in the base field, a contradiction.
The result also fails in general for almost simple groups for $p=2$.
Let $G = \OO^+_{4m}(2^a)$. There are two conjugacy classes of involutions in $\SO_{4m}^+(2^a)= \Omega_{4m}^+(2^a)$
interchanged by an involution in $G$. Clearly such an outer involution cannot commute with
any element of either of the two conjugacy classes interchanged.
Note that all three families of counterexamples above are also counterexamples to the corollary.
We also remark that it is easy to check that
the quaternion group $Q_{8}$ is the only extra-special $2$-group which has a broad subgroup, while it is clear that for $p$ odd,
a non-Abelian $p$-group of exponent $p$ never has a broad elementary $p$-subgroup.
In the next three sections, we prove Theorem \ref{t:main} for the various families of quasi-simple groups. In the last two sections,
we prove Corollary \ref{c:positive} and discuss some other character theoretic results.
\section{Alternating and Sporadic Groups}
We make a few preliminary remarks before starting the proof. We may assume that $G$ has no odd order
normal subgroups and in particular if $G$ is quasisimple, we may assume that the center is a $2$-group.
If $G$ has a unique class of noncentral involutions, then the result holds by considering the abelian
subgroup $E = \langle Z(G), x \rangle$ where $x$ is a noncentral involution. If $G$ has $3$ conjugacy classes
of involutions, $C_1, C_2$ and $C_3$ and there exist $x_i \in C_i$ with $x_1x_2x_3=1$, then
the elementary abelian subgroup $\langle x_1, x_2 \rangle$ intersects each $C_i$. If $H$ is a subgroup of $G$
and $H$ intersects every conjugacy class of involutions of $G$ and $H$ has a broad subgroup, then so does $G$.
The case of alternating groups is trivial.
\begin{lem} Let $G=A_n, n \ge 5$. Let $E$ be the elementary abelian subgroup of $S_n$
with all orbits of size $2$ if $n$ is even and all orbits but one of size $2$ if
$n$ is odd. Then every involution in $S_n$ is conjugate to an element of $E$
by an element of $A_n$.
\end{lem}
\begin{lem} Let $G=2A_n, n \ge 5$. Then $G$ contains a broad subgroup.
\end{lem}
\begin{proof} If $x \in A_n$ is an involution, then $x$ lifts to an involution in $G$ if and only
if the number of points moved by $x$ is a multiple of $8$. Now choose a partition of
$n$ into subsets $X_0, X_1, \ldots, X_d$ where $|X_i|=8$ for $i > 0$ and $|X_0| < 8$. Let $E$ be the elementary
abelian $2$-group $E_1 \times \ldots \times E_d$ where $E_i$ acts as a regular elementary
abelian group of order $8$ on $X_i$. Note that if $x \in E$, then $x$ moves $8e$ points for some
$0 \le e \le d$ and so every element in $E$ lifts to an involution in $G$, whence the lift of $E$ to $G$
is also elementary abelian.
\end{proof}
All but five of the sproadic groups have at most $2$ noncentral conjugacy classes of
involutions and so the observations above yield the result.
The remaining cases are $Co_2, Co_1, Fi_{22}, Fi_{23}$ and $BM$. The first four have three nontrivial classes
of involutions and $BM$ has four such classes. In these cases, we just use GAP \cite{B} to produce
$E$. One deals with the quasi-simple sporadic groups in a similar manner \cite{B}.
\begin{lem} Let $G$ be a quasi-simple sporadic group. Then there exists an elementary abelian
subgroup $E$ intersecting every conjugacy class of involutions in $G$.
\end{lem}
\section{Group of Lie type in characteristic $2$}
In this section $q$ is a power of $2$. We consider the finite simple groups of Lie type over $\mathbb{F_q$.
We will typically work with the simply connected group (which will have center of odd order) which
is sufficient. We note the result that centralizer of an involution is connected.
It is not true that all centralizers of unipotent elements are connected.
\begin{lem} Let $G$ be a connected reductive algebraic group over any algebraically closed field
of characteristic $2$. Let $g \in G$ be an involution. Then $C_G(g)$ is connected.
\end{lem}
\begin{proof} Clearly, we can assume that $G$ is semisimple. Then $Z(G)$ is finite of odd order
and so the result does not depend upon the isogeny class of $G$. So we may
assume that $G$ is a direct product and so it suffices to assume that $G$ is simple.
If $G$ is exceptional, we just quote \cite[Chapter 22]{LS}. So assume that $G$ is classical.
By Lang's theorem, it suffices to show that if two involutions in $G(q)$ (untwisted) are conjugate in
the algebraic group, they are already conjugate in $G(q)$ (and indeed we may assume that $q$
is large). If $G=\operatorname{SL}_{2n}(q)$, then any involution is conjugate to one in the unipotent radical $Q$
of the stabilizer of $n$-dimensional space and two elements in $Q$ are conjugate if and only
if they have the same rank (identify $Q$ with $n \times n$ matrices), whence the result. If
$G=\operatorname{SL}_n(q)$, we can reduce to the case that $n$ is even (or use a slightly different unipotent
radical). Similarly if $G = \Sp_{2n}(q)$, then any involution acts trivially on a totally singular
$n$-dimensional subspace \cite{FGS} and so is contained in the unipotent radical $Q$ of the stabilizer of
such a space \cite{FGS}. We can identify $Q$ with the space of symmetric bilinear forms on an $n$-dimensional
space and the orbits of the Levi subgroup are the congruence classes of such forms. Two forms
are congruent if and only they have the same rank and either both are alternating or not and again
the result holds.
Next consider $G = \SO^+_{2n}(q)=\SO(V)$. Let ${\bar G}$ be the corresponding algebraic group.
Let $C$ be a $G$-conjugacy class of involutions and let ${\bar C}$ be the conjugacy class
of $\bar{G}$ containing $C$.
Recall that $G \le H = \Sp_{2n}(q)$. It follows by \cite{AS} or \cite{FGS} that if $D$ is the $H$-class containing
$C$, then $D \cap G = C$ or $C \cup \tau(C)$ for $\tau$ any element of $\OO_{2n}^+(q)$ not in $\SO_{2n}^+(q)$. By
the result for symplectic groups, $D$ is precisely the set of $\mathbb{F_q$ points of the corresponding
conjugacy class in the algebraic group.
Let $x \in C$. Let $( \ , \ )$ denote the alternating
form preserved by $G$. If $(xv,v) \ne 0$ for some
$v \in V$, then $x$ preserves a nondegenerate $2$-space and so $x$ commutes with a transvection $\tau$
(over the base field). In particular, $\tau(C) = C$ and so $C = D \cap G =\bar{C} \cap H$ as required.
Suppose that $(xv,v)=0$ for all $v \in V$. It follows
just as for the symplectic case
that $x$ acts trivially on a totally singular $n$-dimensional space and so $x$ is contained in the
unipotent radical of the stabilizer of such a space and the orbits of the Levi subgroup on the unipotent
radical correspond to the orbits of $\operatorname{GL}_n$ on skew symmetric matrices. The only invariant is the rank
and so there is no fusion after field extensions. Thus $\bar{C} \cap H = C$ as required.
\end{proof}
\subsection{Linear, Unitary and Symplectic Groups}
The case of $\operatorname{SL}_n(q)$ is obvious.
\begin{lem} Let $G=\operatorname{SL}_n(q), n \ge 2$. Let $m=n/2$ if $n$ is even and let $m=(n+1)/2$ if $n$ is odd.
Let $P$ be the stabilizer of an $m$-dimensional subspace $W$. Then any involution in $G$
is conjugate to an element acting trivially on $W$. In particular, if $Q$ is the unipotent radical
of the stabilizer of $W$, then $Q$ is elementary abelian and every involution of $G$ is conjugate to
an element of $Q$.
\end{lem}
For symplectic and unitary groups, we use the following elementary observation.
\begin{lem} Let $G=\SU_{2n}(q), n \ge 2$ or $\Sp_{2n}(q), n > 2$.
Then any involution $g$ acts trivially on a totally isotropic subspace of dimension $n$
and so the elementary abelian subgroup $Q$ that is the unipotent radical of the stabilizer of
a totally singular subspace of dimension $n$ intersects every conjugacy class of involutions.
\end{lem}
\begin{proof} For $G=\Sp_{2n}(q)$, the argument is easier (and is given in \cite{FGS}).
So we assume that $G=\SU_{2n}(q)$. If $n=1$, this is clear.
Let $W$ be the fixed space of the involution of $g$.
Note that $\dim W \ge n$. If $W$ is totally singular, the result follows.
Otherwise $g$ is trivial on a nonsingular $1$-space $L$.
Then $g$ acts on $L^{\perp}$ and since
$\dim L^{\perp} = 2n-1$, the fixed space of $g$ has dimension at least $n$ and so is not totally singular.
Thus
$g$ leaves invariant some nondegenerate $1$-space orthogonal to $L$ and so
acts trivially on a nondegenerate $2$-space. The result then follows by induction.
\end{proof}
If $G=\SU_{2n+1}(q)$, then any involution is conjugate to an element of
$\SU_{2n}(q)$. So to prove the theorem it suffices to consider only even dimensional
unitary groups. Thus, the previous lemma gives the result for $\SU$ and $\Sp$.
\subsection{Orthogonal Groups}
We now consider $G=\SO_{2n}^{\epsilon}(q)=\SO(V)$ with $n \ge 4$. This case is different than
the other families of classical groups and we have to examine the conjugacy classes of involutions
more closely. We should note that (in characteristic $2$) we denote the simple group by $\SO$ (so in particular
$\SO$ is not the full orthogonal group of determinant $1$ isometries preserving the quadratic
form -- often $SO^{\epsilon}_n(q)$ is denoted by $\Omega^{\epsilon}_n(q)$).
Suppose that $n = 2m $ is even. Decompose $V = V_1 \perp \ldots \perp V_m$ as the orthogonal sum of $m$ four dimensional
spaces. If $V$ has $+$ type, choose all the summands to be of $+$ type. Otherwise choose
$V_m$ to be of $-$ type. Let $E$ be a Sylow $2$-subgroup of $\SO(V_1) \times \ldots \times \SO(V_m)$. Note
that $E$ is elementary abelian and that the normalizer of $E$ contains elements in $\OO(V)$ not in $\SO(V)$.
Thus it suffices to show that $E$ intersects every $\OO(V)$ conjugacy class of involutions. Such involutions
are described in \cite[Chap. 6]{LS}. They possible conjugacy classes are labelled
$W(2)^e \oplus W(1)^f$ or $V(2)^2 \oplus W(2)^e \oplus W(1)^f$ (see \cite[Chapter 6]{LS}).
They are distinguished by their Jordan form and whether $(gv,v)$ vanishes identically for $v \in V$ with respect
to the alternating form left invariant by $G$ (the classes involving $V(2)^2$ are the ones where this function does
not vanish everywhere).
Note that involutions in $\SO_4^+(V)$ are either trivial, $W(2)$ or $V(2)^2$. Involutions in $\SO_4^-(V)$ are
either trivial or are in the class $V(2)^2$. If $V$ has $+$ type, then it is clear
any class of involutions intersects $E$. Similarly if $V$ has $-$ type, it is clear that any involution other than $W(2)^m$
is conjugate to an element of $E$. However, this class of involutions in the algebraic group is not invariant under the graph
automorphism (i.e. there are two $\SO$ classes which are fused in $\OO$). Since the classes are stable under any
field automorphism, it follows that this class is not present in $\SO^-_n(q)$.
If $n$ is odd, decompose $V = W \perp U$ with $\dim U = 2$ and $W$ of $+$ type (and so
$U$ as the same type as $V$).
By \cite[Chap. 6]{LS}, any involution in $g$ is conjugate to an element in $\SO(W)$ and so the result
follows by the case that $n$ is even. This completes the proof of the theorem in this case.
\subsection{Exceptional Groups}
Note that for the groups ${^2}B_2(2^{2a+1})$, ${^2}F_4(2^{2a+1})'$ or $G_2(q)$, there are at most
$2$ noncentral conjugacy classes of involutions and the theorem holds -- see \cite[Chap. 22]{LS}.
Also since centralizers of involutions are connected, all classes of involutions intersect a subfield
group and so one can work over the base field (and so use the character tables). We do not use this.
In $F_4(q)$, the center of a Sylow $2$-subgroup has order $q^2$ and intersects of three of the four
nontrivial classes of involutions and so the result holds in the case as well.
It is easy to see that each conjugacy class of involutions in ${^2}E_6(q)$ or $E_6(q)$ intersects $F_4(q)$.
This can be seen by the tables in \cite{LS} -- indeed they exhibit the classes of unipotent elements
in the exceptional groups by first working in $E_8$ and then determining which classes intersect
the smaller groups and if they split. Alternatively, using the Jordan block structure of involutions of
$F_4$ acting on the Lie algebra and the $26$ dimensional module \cite{Law} and noting that the
Lie algebra of $E_6$ as an $F_4$-module is the direct sum of the Lie algebra of $F_4$ and the $26$-dimensional module, one can see there
three of the nontrivial classes in $F_4$ remain distinct in $E_6$. Thus, the result for these cases follows from
the result for $F_4$.
In $E_7$, using the standard labelling, we see that we can choose $4$ commuting simple root subgroups
and the group generate by meets all four classes (see the labelling given in \cite{AS} or \cite{Law}). Each
class of involutions in $E_8$ intersects $E_7$ (arguing as in the case of $F_4 < E_6$). Thus, the theorem
holds for $E_7$ and $E_8$.
\subsection{Exceptional Multipliers}
If $G$ is a simple group of Lie type in characteristic $2$, then almost always its Schur mulitplier has odd order
and so there is nothing more to do. There are a handful of cases where this fails. See \cite[Table 6.1.3]{GLS}.
In each of the cases, one produces the required broad subgroup using GAP \cite{B}.
\section{Groups of Lie type in odd characteristic}
In this section, we consider finite simple groups of Lie type over a field $\mathbb{F_q$ with $q$ odd. We prove
a slightly stronger result which makes the proof easier. So let $X$ be a simple algebraic group over
an algebraically closed field of odd characteristic $p$. Let $\sigma$ be a Lang-Steinberg endomorphism
of $X$ and let $H=X_{\sigma}$ the fixed points of $\sigma$. Then $H$ is a finite group of Lie type over
a finite field $\mathbb{F_q$ for some $q$ a power of $p$. Let $G = O^{p'}(H)$. Then aside from the cases
that $H={^2}G_2(3)$ or $H$ is $\operatorname{SL}_2(3)$ or $\operatorname{PGL}_2(3)$, we have $G$ is quasi-simple. Note that in the last cases $G$
is solvable and in the first case $[H,H] \cong \operatorname{PSL}_2(8)$. We will exclude these cases (clearly the main theorem holds
in these cases).
We say that a subgroup $S$ of $H$ is toral if $S$ is contained in a torus $T$ in $X$. It follows that $S$ is contained
in a maximal torus of $H$ (which is defined to be $C_T(\sigma)$ for $T$ a $\sigma$ invariant maximal torus of $X$).
Note that if $\phi: X \rightarrow Y$ is an isogeny of algebraic groups, then tori map to tori and the inverse image of a maximal
torus is a maximal torus.
We shall prove the following:
\begin{thm} \label{t:toralbroad} Let $G$ be a finite quasi-simple group of Lie type over a field of odd characteristic. Then there exists a toral
subgroup $E$ of $G$ which intersects every conjugacy class of involutions of $G$.
\end{thm}
If $G$ is as in the theorem and $Z$ is a central subgroup, then the result for $G/Z$ implies the result for $G$
(since as noted above toral subgroups lift to toral subgroups and any involution in $G$ maps to an involution in
$G$). On the other hand, if we prove that there exists a toral subgroup $S$ of $G$ such that $S$ intersects
every conjugacy class of $2$-elements $g \in G$ with $g^2 \in Z$, then $SZ/Z$ is toral and intersects each
class of involutions in $SZ/Z$. Thus, we can choose
a particular form of the group and prove the result needed for that form.
If $G$ is simply connected and split, the result is quite easy.
\begin{lem} \label{l:splitsc} Let $G$ be a finite quasi-simple simply connected group of Lie type over the field of
$q$-elements.
Let $S$ be a maximal torus of $G$ contained in a Borel subgroup of $G$. Then $S$ intersects every conjugacy class
of involutions of $G$.
\end{lem}
\begin{proof} Let $T$ be a maximal torus of $X$ containing $S$. Every involution of $X$ is conjugate to an element
of $T$ (indeed every semisimple element of $X$ is conjugate to an element of $T$). Note that $S$ is a direct
product of $r$ copies of a cyclic group of order $q-1$ and so that $S$ contains the $2$-torsion subgroup of $T$.
Let $g \in G$ be an involution, then $g$ is conjugate in $X$ to some element of $S$. Since $X$ is simply connected,
$g^X \cap G = g^G$ and so $g$ is conjugate in $G$ to an element of $S$.
\end{proof}
We can now complete the proof for the exceptional groups. The previous lemma implies the result for
the simply connected groups of type $G_2(q)$, $F_4(q)$, $E_6(q)$, $E_7(q)$ and $E_8(q)$. Aside
from the case of $E_7(q)$, the centers are either trivial or have order $3$ and so the result holds.
We note that in the simply connected group $E_7(q)$, there are maximal tori that are direct products
of seven cyclic groups of order $q-1$ and also of order $q+1$ \cite{LSS}. Arguing as above, one of these maximal
tori contain a conjugate of any element of order $4$ in $G$ and so any element whose square is central.
This leaves only the cases of ${^2}G_2(3^{2k+1}), k \ge 1$ and ${^2}E_6(q)$. In the first case, there is only
one class of involutions and the result follows since any semisimple element is contained in some maximal
torus. In the case of ${^2}E_6(q)$ (and since the center has odd order, it suffices to consider the
simply connected case), there is a maximal torus that is the direct product of six copies of the cyclic
group of order $q+1$ \cite{LSS}. This contains the $2$-torsion subgroup of a maximal torus of the algebraic group
and we argue as above.
We next consider the classical groups. We will work with the form of the group that acts on the natural module $V$.
We first consider $\operatorname{SL}_n(q)$. If $n$ is odd, the center has odd order and so the result follows by Lemma \ref{l:splitsc}.
We next handle the case that $n$ is even.
\begin{lem} Let $G = \operatorname{SL}_{2m}(q) \le GL_{2m}(q) = L $. Then there exists a toral subgroup $A$ of
$L$ such that $A$ intersects every conjugacy class of elements of $G$ containing
elements whose square is central.
\end{lem}
\begin{proof} Note that semisimple elements in $G$ which are conjugate in $L$ are already
conjugate in $G$. Let $Z$ be the Sylow $2$-subgroup of
the center of $G$ with $z$ a generator. Consider the subgroup $B = B_1 \times \ldots \times B_m$ of $L$ with each $B_i \cong \operatorname{GL}_2(q)$
acting on a direct sum of $m$ two dimensional spaces. Let $A_i$ be a cyclic subgroup of $B_i$ of order $q^2-1$
and set $A=A_1 \times \ldots \times A_m$. Note that $A$ contains an element $w$ with $w^2=z$ (by considering
the two dimensional case). Suppose that $y^2 = z^i$ with $y \in G$. If $i$ is odd, then there is an odd power of $y$ with square $z$
and so $y$ is conjugate to a power of $w$. If $i=2j$, then $yz^{-j}$ is an involution in $G$. It is easy to see that an involution
in $G$ is conjugate to an element of $A$ (since each eigenspace is even dimensional). Thus, $A$ contains a conjugate of
every element of $G$ whose square is central. Clearly $A$ is toral and so the result holds.
\end{proof}
As we noted above, this implies the result for any quotient of $\operatorname{SL}$.
The same proof holds for $G=\SU_{2m}(q), m \ge 2$ and so the main result for any quotient of $\SU_{2m}(q)$.
If $G=\SU_n(q)$ with $n$ odd, then $Z(G)$ has odd order and we see that there
is a toral subgroup that is a direct product of $n-1$ copies of a cyclic group of order $q+1$ which contains the $2$-torsion
subgroup of a maximal torus of the algebraic group and the result follows as above.
If $G = \Sp_{2n}(q)$ with $q \equiv 1 \mod 4$, then the split torus $T$ contains the $4$-torsion
subgroup of the maximal torus containing it in the algebraic group. Thus every element
of order $4$ in $G$ is conjugate to an element of $T$ whence the results holds for $\Sp_{2n}(q)$
and its central quotient. If $q \equiv 3 \mod 4$, there is a maximal torus of $G$ that is the direct
product of $n$ copies of the cyclic group of order $q+1$ which again contains the $4$-torsion subgroup
of a maximal subgroup of the algebraic group and the result follows.
Finally we consider the quasi-simple groups related to orthogonal groups in dimension at least $7$. We will prove the
result for the groups $G=\Omega^{\pm}_n(q) \le L = \SO^{\pm}_n(q)$. As noted above, this suffices
to prove the result in general.
First suppose that $n = 2m +1$ is odd.
Then $\SO_n(q)=\SO(V)$ has no center and the simple group
$\Omega_n(q)$ consists of the elements with spinor norm $1$ and has index $2$. Decompose $V = W \perp L$ with
$W$ a hyperplane such that the central element in $\SO(W)$ has spinor norm $1$. Now decompose
$W = W_1 \perp \ldots \perp W_m$ as an orthogonal sum of $m$ nondegenerate $2$-paces with at least one summand of each type.
If $g$ is an involution in $\Omega_n(q)$, then $g$ is conjugate to an element of the maximal torus $T =
\SO(W_1) \times \ldots \times \SO(W_m)$. If the multiplicity of the trivial eigenspace of $g$ is greater than $1$, this is clear since
the class of $g$ (for $g$ an involution even in $\SO(V)$) is determined by the dimension and type of its fixed space.
If the fixed space of $g$ is $1$-dimensional and $g$ has spinor norm $1$, then the $-1$ eigenspace of $g$
has the same type as $W$ and so $g$ is conjugate to the element acting as $-1$ on $W$. Thus, $T \cap \Omega_n(q)$
is a toral subgroup intersecting each conjugacy class of involutions and the result holds.
Next consider the case that $n = 4m+ 2, m \ge 2$. First consider the case that $L = \SO_n^{\epsilon}(q)= \SO(V)$
with $q \equiv \epsilon 1 \mod 4$, Note that $Z(L)=Z(G)$ has order $2$ in this case.
Decompose $V = V_1 \perp \ldots \perp V_{2m+1}$ where the $V_i$ are $2$-dimensional spaces
of the same type as $V$. Let $T = \SO(V_1) \times \ldots \times \SO(V_{2m+1})$.
Then $T$ is a maximal torus of $\SO(V)$. It is easy to see that any involution
of spinor norm $1$ is conjugate to an element of $T$. Moreover, if $x \in G$ and $x^2 = -I$, then
$x$ is conjugate to an element of $T$ (reduce to the two dimensional case).
Thus, $T \cap \Omega(V)$ intersects any conjugacy
of elements of $G$ whose square is central. This proves the theorem for $\Omega(V)$ and it simple quotient
and so the result holds.
Suppose that $q \equiv \epsilon 3 \mod 4$. In this case the nontrivial involution in $Z(L)$ has spinor
norm $-1$ and so $\Omega(V)$ has trivial center and is the simple group.
Decompose $V = V_1 \perp \ldots \perp V_{2m+1}$ so that
there are at least $2$ summands of each type. Let $T = \SO(V_1) \times \ldots \times \SO(V_{2m+1})$.
Then $T$ intersects every conjugacy class of involutions of $\SO(V)$ and the result holds.
Finally consider the case that $n=4m, m \ge 2$. Suppose that $L = \SO_n^-(q) = \SO(V)$. Note that in
this case $-I$ has spinor norm $-1$ and so $\Omega_n^{-}(q)$ is simple. Decompose $V = V_1 \perp \ldots \perp V_{2m}$
into an orthogonal
sum of $2m$ nondegenerate $2$-dimensional spaces (note both types must occur in this decomposition). It is clear
that any involution in $\Omega(V)$ is conjugate to an element of $T= \SO(V_1) \times \ldots \times SO(V_{2m})$ and
the result holds.
Finally suppose that $L=\SO_n^+(q)$. Decompose $V = V_1 \perp \ldots \perp V_{2m}$ with each $V_i$ of dimension $2$ and all
of the same type. If $q \equiv 1 \mod 4$, take the type of $V_i$ to be $+$ and $-$ otherwise. In the case the central
involution in $L$ has spinor norm $1$ and indeed the involution in $\SO(V_i)$ has spinor norm $1$. Set $T =
SO(V_1) \times \ldots \times SO(V_{2m})$, a direct product of $2m$ cyclic groups of order $q \pm 1$ (and each
cyclic group has order a multiple of $4$). If $x \in G$ and
$x^2 = -I$, then $x$ is conjugate to an element of $T$ and any involution of spinor norm $1$ is also conjugate to
an element of $T$ and the result follows.
\section{Proof of Corollary 3}
As mentioned in the introduction, a theorem of R. Kn\"orr \cite{K}, tells us that when $p$ is a prime and $G$ is a finite group of order divisible by $p$, then the irreducible character $\chi$ of $G$ lies in a $p$-block of defect zero of $G$ (hence vanishes on all elements of $G$ of order divisible by $p$) if and only if $\chi$ vanishes on all elements of order $p$ in $G$. Hence when $G$ has an elementary Abelian subgroup $E$ which meets every conjugacy class of elements of order $p$ of $G$, the irreducible character $\chi$ lies in a $p$-block of defect zero of $G$ if and only if $\chi$ vanishes identically on the non-identity elements of $E$.
We may conclude, using the results of \cite{R} that if $G$ has such an elementary Abelian $p$-subgroup $E$, then $\sum_{ a \in E^{\#}}|\chi(a)|^{2} \geq |E|-1 $ whenever $\chi$ lies in a $p$-block of positive defect of $G$.
Using the orthogonality relations, we deduce that if $G$ has $k_{+}$ irreducible characters lying in $p$-blocks of positive defect, we have $$ \sum_{a \in E^{\#}} |C_{G}(a)| \geq k_{+}(|E|-1).$$
In particular, we have $k_{+} \leq |C_{G}(a)|$ for some $a \in E^{\#}.$
More generally, let $G$ be a finite group, and $N \lhd G$. We will first prove that if $N$ has an elementary Abelian $p$-subgroup $E$ which meets every conjugacy class of elements of order $p$ of $N$,
then there an element $x$ of order $p$ in $N$ such that at most $|C_{G}(x)|$ irreducible characters of $G$ lie in $p$-blocks which do not cover $p$-blocks of defect zero of $N$.
We first need a slight extension of the above result of Kn\"orr.
\begin{lem} Let $G$ be a finite group and $N$ be a normal subgroup of $G$. Let $B$ be a $p$-block of $G$ with defect group $D$ such that $D \cap N \neq 1.$ Then for any irreducible character $\chi \in B$, there is an element $ x \in N$ of order $p$ with $\chi(x) \neq 0$.
\end{lem}
\begin{proof} Let $(R,K,F)$ be a $p$-modular system for $G$. We note that $H = N_{G}(D \cap N) \geq N_{G}(D)$ so by Brauer's First Main Theorem, there is a unique Brauer
correspondent block for $B$ in $N_{G}(D \cap N)$, say $B^{\prime}.$ Let $\sigma = \sum_ {\{n \in N: n^{p}=1 \} } n$, which is an element of $Z(RG)$.
We claim that $\omega_{\chi}(\sigma) \in J(R)$, so that $\omega_{\chi} (\sigma - 1_{G}) \not \in J(R)$, and in particular, there must be some non-identity element
$x \in N$ of order $p$ with $\chi(x) \neq 0$.
Now, using the Brauer homomorphism, there is an irreducible character $\mu \in B^{\prime}$ (which may be assumed to have $D \cap N$ in its kernel) such that
$\omega_{\chi}(\sigma) \equiv \omega_{\mu} (\sigma^{\ast} )$ ( mod $J(R)$ ), where $\sigma^{\ast} = \sum_{ \{n \in C_{G}(D \cap N) : n^{p} = 1 \}} n.$
However, let $Z = \Omega_{1}(Z(D \cap N))$, and let $Z^{+}$ be the sum of its elements in $RG$. It is clear that $\sigma^{\ast} = Z^{+}T$ for some element $T$ of $RC_{G}(D \cap N)$ which is itself a sum of certain elements of order dividing $p$. Notice that $T$ commutes with $Z^{+}.$ It follows that $\sigma^{\ast}$ has nilpotent image in $Z(FN_{G}(D \cap N))$, so that $\omega_{\mu} (\sigma^{\ast}) \in J(R)$, which suffices to complete the proof.
\end{proof}
We conclude from the Lemma (and Clifford's Theorem) that the irreducible character $\chi$ of $G$ lies over characters in $p$-blocks of positive defect of $N$ if and only if $\chi(x) \neq 0$ for some element $x$ of order $p$ in $N$.
In particular, if $E$ is the above broad elementary Abelian $p$-subgroup of $N$, then no irreducible character $\chi$ of $G$ which lies over irreducible characters in $p$-blocks of positive defect zero of $N$ can vanish identically on $E^{\#}.$ We may then conclude as above that there is an element $x$ of order $p$ in $N$ such that the total number of irreducible characters of $G$ which do not vanish identically on $p$-singular elements of $N$ is at most $|C_{G}(x)|$.
Finally, we turn our attention to the case $p = 2$, $O(G) = 1$ and $N = F^{\ast}(G)$. If $O_{2}(G) \neq 1$, let $Z = \Omega_{1}(Z(O_{2}(G)) ) \neq 1.$ Then by Clifford's Theorem, whenever $\chi$ is an irreducible character of $G$, ${\rm Res}^{G}_{Z}(\chi)$ is not a multiple of the regular character, so that $\chi$ does not vanish identically on $Z^{\#}.$ Hence, as before, we have $\sum_{ z \in Z^{\#}}|\chi(z)|^{2} \geq |Z|-1$, and we have $k(G) \leq |C_{G}(z)|$ for some $z$ of order $2$ in $Z$, where (as usual), $k(G)$ denotes the number of complex irreducible characters of $G$.
Suppose then that $O_{2}(G) = 1$. Then $F^{\ast}(G)$ is a direct product of non-Abelian simple groups, and $F^{\ast}(G)$ has a broad elementary $2$-subgroup, say $E$, by Theorem 1 and the remarks following. Now the arguments above tell us that there is some element $t \in E^{\#}$ such that the number of irreducible characters of $G$ which do not vanish identically on $2$-singular elements of $F^{\ast}(G)$ is at most $|C_{G}(t)|$, and the proof of Corollary 3 is complete.
We remark that if equality holds in Corollary 3, then every irreducible character of $G$ which does not vanish at $t$ takes value $\pm 1$ at $t$, and that each such character has odd degree, so lies in a $2$-block of full defect of $G$ (and has height zero in that block).
\section{Further character-theoretic applications}
One of our motivations for asking the question answered by Corollary 2 is the following:
\begin{lem} Let $G$ be a finite group, and let $t,u$ be non-trivial involutions of $G$ such that $t$ does not commute with any $G$-conjugate of $u$. Let $C = C_{G}(t)$ and $D = C_{G}(u)$.
Then there is a non-trivial `irreducible character $\mu$ in the principal $2$-block of $G$ such that $\mu(t)\mu(u) \neq 0$ and $\mu(1) \leq ([C:O(C)]-1)(\sqrt{[D:O(D)]-1})$.
\end{lem}
\begin{proof} We first note that $u$ is not expressible as the product of two conjugates of $t$. More generally, the product of two conjugates of $t$ never lies in the $2$-section of $u$, for if $t^{x}t^{y}$ has $2$-part $u^{z}$, then $t^{x}$ and $t^{y}$ both invert the involution $u^{z}$, contrary to hypothesis.
By the usual character-theoretic formula for the coefficient of $g$ in the product of two class sums, we see that the class function $$\sum_{ \chi \in {\rm Irr}(G)} \frac{\chi(t)^{2} \chi}{\chi(1)} $$ vanishes identically on the $2$-section of $u$ in $G$.
By a well-known consequence of Brauer's Second Main Theorem, the class function $$\sum_{ \chi \in B} \frac{\chi(t)^{2} \chi}{\chi(1)} $$ also vanishes identically on the $2$-section of $u$ in $G$, where $B$ is the principal $2$-block of $G$. It follows easily that $$\sum_{ 1 \neq \chi \in B} \frac{\chi(t)^{2}|\chi(u)|}{\chi(1)} \geq 1. $$
Choose a non-trivial irreducible character $\mu \in B$ with $\mu(t)\mu(u) \neq 0$ and with $r(u) = \frac{|\mu(u)|}{\mu(1)}$ maximal. Then $$\sum_{ 1 \neq \chi \in B} \chi(t)^{2} \geq \frac{1}{r(u)}.$$
By Brauer's Second and Third Main Theorem, we have $$\sum_{ \chi \in B} \chi(t)^{2} = \sum_{ \theta \in b} \theta(1)^{2},$$ where $b$ is the principal $2$-block of $C = C_{G}(t)$ .
Since each irreducible character $\theta$ of $b$ has $O(C)$ in its kernel, we deduce that $$\sum_{ 1 \neq \chi \in B} \chi(t)^{2} \leq [C:O(C)] - 1$$.
A similar argument with $u$ certainly allows us to conclude that $|\mu(u)| \leq \sqrt{[D:O(D)]-1}.$ Hence we may conclude that $$\mu(1) \leq ([C:O(C)]-1)( \sqrt{[D:O(D)]-1}),$$ and the result follows.
\end{proof}
In view of Corollary 2, this result may not be applied as it stands to finite simple groups, but we have seen that the hypotheses may be satisfied within almost simple groups, so we mention:
\begin{cor} Let $G$ be an almost simple group with $F^{\ast}(G) = S$, and let $t$ be an involution of $G \backslash S$ and $u$ be an involution of $S$ which commutes with no conjugate of $t$.
Then the principal $2$-block of $S$ contains a $t$-stable non-trivial irreducible character of degree at most $([C:O(C)]-1)(\sqrt{[D:O(D)]-1}),$ where $C = C_{S}(t)$ and $D = C_{S}(u)$.
\end{cor}
\begin{proof} We might as well work within $\langle t \rangle S$, so we suppose that $G = \langle t \rangle S$. We may argue in a similar fashion to the previous result. Now the principal $2$-block of $G$ contains two linear characters, $1$ and $\lambda$, say, and we also have $\lambda(t)^{2}\lambda(u) = 1.$ This time, we find that $$\sum_{ \chi \in B: \chi(1) >1} \frac{\chi(t)^{2}|\chi(u)|}{\chi(1)}
\geq 2.$$ where $B$ is again the principal $2$-block of $G$. However, the irreducible characters $\chi \in B$ with $\chi(t) \chi(u) \neq 0$ come in pairs, both members of which lie over the same $t$-stable irreducible character of the principal $2$-block of $S$ (if $\chi$ is one such, so is $\lambda \chi \neq \chi$, and note that $\chi(t)^{2} \chi(u)$ is unchanged on replacing $\chi$ by $\lambda \chi$). Also, $C_{G}(t) = \langle t \rangle \times C_{S}(t)$, so the result follows in a fashion similar to the previous Corollary.
\end{proof}
| 64b998371618145f7f75e90aadca07bcabf39fd2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Over the past few years, there has been a soaring interest for the study of multi-agent collective behavior models.
Indeed, they can be applied in many different areas: biology with the study of collective flocks and swarms \cite{MR2310046,MR2205679,MR2064375,1582249}, aviation \cite{MR1617559}, and social dynamics \cite{HK} (among many others).
In this article, we will take a particular interest in models for opinion dynamics.
The first ones were introduced in the 50's \cite{social,social2} and already stressed the highly non-linear nature of social interactions. Some early models for consensus formation have been introduced by DeGroot \cite{DG} and more recently by Hegselmann and Krause \cite{HF,HK,MR1792007}. Since then, a wealth of models have been developed in order to study one of the key features of collective dynamics: the intriguing emergence of global patterns from local interaction rules. This phenomenon is referred to as \emph{self-organization} \cite{ACMPPRT17,MR3714980,MR2343706,JabinMotsch14,MR2064375,1582249}.
Most social dynamics models have a common structure: the dynamics of the agents' opinions are described by a system of ODEs in which the agents interact pairwise, i.e.
\begin{equation}\label{eq:HKintro}
\displaystyle \frac{d}{dt} x_i(t) = \frac{1}{N} \sum_{j=1}^N a_{ij}\,(x_j(t)-x_i(t)),
\end{equation}
where $x_i$ represents the time-evolving opinion of agent $i$.
Depending on the nature of the interaction coefficients $a_{ij}$, these models can be roughly classified in two categories.
In the first one, interactions are pre-determined by a given interaction network, which represents the inherent structure of the population's interactions \cite{olfati2007,WS98}. Then each pairwise interaction coefficient $a_{ij}$ is non-zero if and only if the edge $(i,j)$ is part of the underlying graph of interactions.
The second approach only considers the state space of opinions and defines the interaction coefficients as a function of the pairwise distances: $a_{ij}:=a(\|x_i-x_j\|)$. In this case, every agent can potentially interact with any other if the distance separating them belongs to the support of the interaction function $a$: there is no underlying network (see for instance \cite{MT14}).\\
Recently, a variant of this model has introduced the concept of \emph{weights of influence} \cite{McQuadePiccoliPouradierDuteil19,MR3965293}.
In this augmented model, each agent is not only defined by its opinion $x_i$, but also by its weight $m_i$.
Then, the influence of an agent $j$ on an agent $i$'s opinion is proportional to its weight $m_j$. These weights are assumed to evolve in time via an equation which is coupled with the opinions' evolution. In other words, the evolution of each agent's opinion does not only depend on its proximity with another agent, but also on the charisma or popularity of the latter - and this charisma also evolves in time. This can be formulated as a system of $2N$ ODEs, as follows:
\begin{equation}\label{eq:syst-gen-intro}
\displaystyle \frac{d}{dt} x_i(t) = \frac{1}{N} \sum_{j=1}^N m_j(t)\,a(\|x_i-x_j\|)(x_j(t)-x_i(t)), \qquad \frac{d}{dt} m_i(t) = \psi_i(x(t),m(t)).
\end{equation}
Interestingly, although the interaction coefficients are given by a function of the pairwise distances between opinions, in this approach we can again view the opinions as nodes of an underlying network. The corresponding weighted graph is \emph{non-symmetric}
(the edge between $i$ and $j$ being weighted by $m_i$ in one direction and by $m_j$ in the other), and \emph{time-evolving} (the weights' dynamics being coupled with the dynamics of the nodes). \\
As in all models of collective dynamics, several natural questions arise.
A first one concerns the \emph{large time limit}, that is the asymptotic behavior of the system.
Many works in the literature have delved into the question of self-organization, i.e. the spontaneous emergence of well-organized group patterns such as consensus, alignment, clustering or dancing equilibrium \cite{MR3714980,MR2343706,MR3392625}. This was studied for the augmented model with time-varying weights in \cite{McQuadePiccoliPouradierDuteil19}.\\
In this paper, we will explore another natural question, the \emph{large population limit}.
When the number of agents tends to infinity, the previous system of $2N$ equations becomes unmanageable, a problem well-known as the \emph{curse of dimension}.
A common answer to this issue consists of studying the \emph{mean-field limit} of the system.
First introduced in the context of gas dynamics (see \cite{braun1977} for instance), when describing particles interacting via a force, a mean-field limit is a limit in which the number of particles $N$ is large ($N$ goes to infinity) but is such that the interaction between particles is both weak enough so that the forces applying on one particle remain finite at the limit, and strong enough so that all the particles continue to interact.
The mean-field limit process consists of representing the population by its density probability, instead of following each agent's individual trajectory.
In the case of the classical opinion dynamics \eqref{eq:HKintro}, the limit measure $\mu(t,x)$ represents the density of agents with opinion $x$ at time $t$.
The mean-field limit of this system is now a classical result,
and one can show that the limit measure satisfies a non-local transport equation \cite{Dobrushin79}.
In the context of the augmented system with time-varying weights \eqref{eq:syst-gen-intro}, the limit measure $\mu(t,x)$ represents the total weight of the agents with opinion $x$ at time $t$. It was shown in \cite{PouradierDuteil21} that it solves a non-local transport equation with non-local source. \\
However, there is a limitation to the mean-field approach.
Since it describes the population by its density, it requires all particles to be \emph{indistinguishable}. This not only entails a significant information loss, but also greatly reduces the span of models that can be studied. It is also incompatible with the graph viewpoint.
In particular, in the case of the augmented model \eqref{eq:syst-gen-intro} with time-varying weights, it requires strong assumptions on the mass dynamics $\psi_i$.
This leads us to the other approach that will be central to this paper: the graph limit method. \\
In 2014, Medvedev used techniques from the recent theory of graph limit \cite{MR2455626,MR2277152,MR2274085,LS,MR3012035} to derive rigorously the continuum limit of dynamical models on deterministic graphs \cite{Medvedev14}.
In the present paper, we extend this idea to our collective dynamics model with time-varying weights, adopting the graph point of view described above.
The central point of this approach consists of describing the infinite population by two functions $x(s)$ and $m(s)$ over the space of continuous indices $s$.
The discrete system of ODEs is then shown to converge as $N$ goes to infinity to a system of two non-local diffusive equations in the space of continuous indices. We show that this approach is more general than the mean-field one, and the Graph Limit can be derived for a much greater variety of models. \\
In the case of dynamics preserving the indistinguishability of particles, we show that both the graph limit and the mean-field limit can be derived.
In particular, we show that there is a hierarchy between the two limit equations: the mean-field limit equation can be derived from the graph limit one.
This subordination of the mean-field limit equation to the graph limit equation, pointed out in \cite{BiccariKoZuazua19}, is natural:
Indeed, the mean-field limit process eliminates all individuality from the particles by considering only the population density.
Thus, there would be no hope of recovering the graph limit equation from the mean-field one. \\
The paper is organized as follows: in Section \ref{sec:presentation}, we present the model and state the main results. In Section \ref{sec:graphlim}, we focus on the graph limit. We start by establishing the existence and uniqueness of a solution to the graph limit equation,
and then prove the convergence of the discrete system to the graph limit equation.
In Section \ref{sec:mfl}, we study the mean-field limit. We first define the key notion of indistinguishability for a particle system. We then prove the subordination of the mean-field limit to the graph one and finish by a weaker but alternative proof of the mean-field limit based on that subordination.
Lastly, we present some numerical simulations in Section \ref{sec:numeric} with concrete models to illustrate our results.
\section{Presentation of the model and main results}
\label{sec:presentation}
We study a social dynamics model with time-varying weights introduced in \cite{McQuadePiccoliPouradierDuteil19}.
From here onward, $d\in\mathbb{N}$ will represent the dimension of the space of opinions, and $N\in\mathbb{N}$ will represent the number of agents whose opinions evolve in $\mathbb{R}^d$.
More specifically, let $x^N=(x_i^{N})_{i\in\{1,\cdots,N\}}:[0,T]\rightarrow(\mathbb{R}^d)^N$ represent the \emph{opinions} (or positions) of $N$ agents, and let $m^N=(m_i^{N})_{i\in\{1,\cdots,N\}}:[0,T]\rightarrow \mathbb{R}^N$ represent their individual \emph{weights of influence}. Each opinion's time-evolution is affected by the opinion of each neighboring agent via the interaction function $\phi\in\mathrm{Lip}(\mathbb{R}^d; \mathbb{R})$, proportionally to the neighboring agent's weight of influence. In turn, the agents' weights are assumed to evolve in time and their dynamics may depend on the opinions and weights of all the other agents, via functions $\psi_i^{(N)}:(\mathbb{R}^d)^N\times\mathbb{R}^N\rightarrow\mathbb{R}$.
Given a set $(x_i^{0,N})_{i\in\{1,\cdots,N\}}$ of initial opinions and $(m_i^{0,N})_{i\in\{1,\cdots,N\}}$ of initial weights, the evolution of the opinions and weights are given by the following system:
\begin{equation}\label{eq:syst-gen}
\begin{cases}
\displaystyle \frac{d}{dt} x_i^{N}(t) = \frac{1}{N} \sum_{j=1}^N m_j^N(t)\, \phi(x_j^{N}(t)-x_i^{N}(t))\\
\displaystyle \frac{d}{dt} m_i^{N}(t) = \psi_i^{(N)}(x^{N}(t),m^{N}(t)), \qquad i\in\{1,\cdots,N\},
\end{cases}
\end{equation}
supplemented by the initial conditions
\begin{equation*
\forall i\in\{1,\cdots,N\}, \quad x_i^{N}(0) = x_i^{0,N} \quad \text{ and } \quad m_i^{N}(0) = m_i^{0,N}
\end{equation*}
such that
\begin{equation}
\label{eq:sum_M}
\sum_{i=1}^N m_i^{0,N}=N.
\end{equation}
We point out that the choice $m_i^{0,N}=1$ and $\psi_i^{(N)}\equiv 0$ for all $i\in\{1,\cdots,N\}$ brings us back to the classical Hegselmann-Krause model for opinion dynamics \cite{HK}:
\begin{equation}\label{eq:HK}
\displaystyle \frac{d}{dt} x_i^{N}(t) = \frac{1}{N} \sum_{j=1}^N \phi(x_j^{N}(t)-x_i^{N}(t)) \qquad i\in\{1,\cdots,N\}.
\end{equation}
This model has been thoroughly studied in the literature (see \cite{ACMPPRT17} for a (non-exhaustive) review) and provides a well-known example of emergence of global patterns, such as convergence to consensus or clustering, from local interaction rules. The augmented model with time-varying weights \eqref{eq:syst-gen} was also shown to exhibit richer types of long-term behavior, such as the emergence of a single (or several) leader(s) \cite{McQuadePiccoliPouradierDuteil19}.
\begin{rem}
In the literature, the interaction function can be found of the form $\phi(x):=a(\|x\|)x$ for some continuous function $a\in C(\mathbb{R}^+;\mathbb{R})$ (see \cite{BiccariKoZuazua19,HK}).
In other works, the interaction between agents takes the form of the gradient of an interaction potential $W:\mathbb{R}\rightarrow\mathbb{R}$, i.e. $\phi(x):=\nabla W(\|x\|)$ (see for instance \cite{CarrilloChoiHauray14}).
Here, we will keep the general notation $\phi$, also used in \cite{JabinMotsch14}, which can cover these various cases.
\end{rem}
From here onward, we will make the following assumptions on the interaction function:
\begin{hyp}\label{hyp:phi}
The interaction function $\phi$ satisfies $\phi(0)=0$ and $\phi\in\mathrm{Lip}(\mathbb{R}^d; \mathbb{R})$, with $\|\phi\|_\mathrm{Lip}=L_\phi$.
\end{hyp}
Notice that at this stage, we have not made any assumptions on the interaction functions $\psi_i$. Actually, unlike the position dynamics, the weight dynamics are allowed to differ for each agent $i$. In this paper, the dependence of $\psi_i$ on the opinions $x^N$ and the weights $m^N$ will take two main forms, that will be specified in the subsequent sections.
The aim of this work is to derive the continuum limit of these dynamics, that is when the number of agents goes to infinity.
We will show that using the graph limit method, we obtain the following limit equation (that we will refer to as the graph limit equation):
\begin{equation}
\left\{\begin{array}{l}
\displaystyle \partial_t x(t,s) = \int_I m(t,s_*)\phi(x(t,s_*) - x(t,s)) ds_* \\
\partial_t m(t,s) = \psi(s,x(t,\cdot),m(t,\cdot)),
\end{array}\right.
\label{eq:GraphLimit-gen}
\end{equation}
where $x\in C([0,T];L^\infty(\mathbb{R}^d))$ and $m\in C([0,T];L^\infty(\mathbb{R}))$ are {associated with} the respective continuum limits of $x^N$ and $m^N$.
Here, $s$ represents the continuous index variable taking values in $I:=[0,1]$, as introduced in \cite{Medvedev14} and \cite{BiccariKoZuazua19}, and $\psi:I\times C([0,T];L^\infty(\mathbb{R}^d))\times C([0,T];L^\infty(\mathbb{R}))\rightarrow\mathbb{R}$ will have to be specified. Notice that the dependence of $\psi_i^{(N)}$ on the index $i$ in the microscopic dynamics \eqref{eq:syst-gen} is translated by the dependence of $\psi$ on the continuous variable $s$ in the limit \eqref{eq:GraphLimit-gen}. Similarly, the dependence of $\psi_i$ on all agents' opinions $x^N(t)$ and weights $m^N(t)$ is encoded by the non-local dependence of $\psi$ on the functions $x(t,\cdot)$ and $m(t,\cdot)$.
\begin{example}\label{ex}
A simple example of mass dynamics depending non-locally on the opinions and weights can be given by functions of the form:
\begin{equation}
\label{eq:psi-part}
\displaystyle \psi(s,x(t,\cdot),m(t,\cdot))= m(t,s) \int_I m(t,\tilde{s}) S(x(t,s),x(t,\tilde{s})) d\tilde{s}
\end{equation}
where $S:\mathbb{R}^d\times\mathbb{R}^d\rightarrow\mathbb{R}$. Note that in this example, $\psi$ depends on the continuous index $s$ only through $x$ and $m$.
More specific examples of mass dynamics $\psi$ will be presented in Section \ref{sec:numeric}. The choice \eqref{eq:modelsimuGL} of Section \ref{sec:numericindisting} provides another example of mass dynamics depending only on the opinions and weights, and not on the individual indices. The choice \eqref{eq:modelsimuGL2} of Section \ref{sec:numericnotindisting} provides an example of mass dynamics depending explicitly on $s$.
\end{example}
In order to give a meaning to the limit,
we will reformulate the discrete system \eqref{eq:syst-gen} in a continuous way, using two operators $P_\mathrm{c}^N$ and $P_\mathrm{d}^N$ respectively transforming vectors into piecewise-constant functions and $L^\infty$ functions into $N$-dimensional vectors.
From here onward, subscripts {in ($x_N$,$m_N$)} will indicate functions over the continuous space $I$ while superscripts {in ($x^N$) , ($m^N$)} will indicate vectors of $(\mathbb{R}^d)^N$ or $\mathbb{R}^N$.
Given initial conditions $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R})$ for the continuous dynamics \eqref{eq:GraphLimit-gen},
satisfying
\begin{equation}
\label{eq:integral_egal_a_1}
\int_I m_0(s) ds =1,
\end{equation}
we can define initial conditions for the microscopic dynamics \eqref{eq:syst-gen}. For each $N\in\mathbb{N}$, we define
\begin{equation}
\label{eq:ICx}
\begin{cases}
\displaystyle x^{0,N} = P_\mathrm{d}^N(x_0) := \left( N \int_{\frac{i-1}{N}}^{\frac{i}{N}} x_0(s) ds \right)_{i \in \{1, \dots, N \}} \in (\mathbb{R}^d)^N \\
\displaystyle m^{0,N} = P_\mathrm{d}^N(m_0) := \left( N \int_{\frac{i-1}{N}}^{\frac{i}{N}} m_0(s) ds \right)_{i \in \{1, \dots, N \}} \in (\mathbb{R})^N.
\end{cases}
\end{equation}
An schematic illustration of the transformation $P_\mathrm{d}^N$ is provided in Figure \ref{fig:Transfo}.
Notice that condition \eqref{eq:integral_egal_a_1} implies that
\eqref{eq:sum_M} is fulfilled.\\
Now, for all $t\in [0,T]$, the solution $(x^N(t),m^N(t))$ to the microscopic system \eqref{eq:syst-gen} at time $t$ can be transformed into a pair of piecewise-constant functions $s\mapsto(x_N(t,s),m_N(t,s))$ via the following operation: for all $s\in I$,
\begin{equation}
\label{eq:xN}
\begin{cases}
\displaystyle x_N(t,s) = P_\mathrm{c}^N(x^{N}(t)):= \sum_{i=1}^N x_i^{N}(t) \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s)\\
\displaystyle m_N(t,s) = P_\mathrm{c}^N(m^{N}(t)) := \sum_{i=1}^N m_i^{N}(t) \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s).
\end{cases}
\end{equation}
The transformation $P_\mathrm{c}^N$ is also illustrated in Figure \ref{fig:Transfo}.
In turn, this transformation will allow us to define the discrete weight dynamics $\psi_i^{(N)}$ from the continuous ones. More specifically, given a functional $\psi: I \times L^2(I;\mathbb{R}^d)\times L^2(I;\mathbb{R}) \rightarrow \mathbb{R}$, we
define
$\psi_i^{(N)}$ in the following way:
\begin{equation}\label{eq:psi}
\forall i\in\{1,\cdots,N\},\qquad \psi_i^{(N)}(x^{N}(t),m^{N}(t)) = N \int_{\frac{i-1}{N}}^{\frac{i}{N}} \psi(s,x_N(t,s),m_N(t,s)) ds,
\end{equation}
where $x_N=P_\mathrm{c}^N(x^N)$ and $m_N=P_\mathrm{c}^N(m^N)$ as defined in \eqref{eq:xN}.
\begin{example}
With this definition, for $\psi$ taken as \eqref{eq:psi-part} in Example \ref{ex}, the system \eqref{eq:syst-gen} becomes
\begin{equation*}
\begin{cases}
\displaystyle \frac{d}{dt} x_i^{N}(t) = \frac{1}{N} \sum_{j=1}^N m_j^N(t) \phi(x_j^{N}(t)-x_i^{N}(t))\\
\displaystyle \frac{d}{dt} m_i^{N}(t) = \frac{1}{N} m_i^{N}(t) \sum_{j=1}^N m_j^{N}(t) S(x_i^{N}(t),x_j^{N}(t)) , \qquad i\in\{1,\cdots,N\}.
\end{cases}
\end{equation*}
\end{example}
\begin{figure}
\includegraphics[trim= 2cm 6cm 2cm 6cm, clip=true, width=\textwidth]{ContinuousToDiscrete.pdf}
\caption{Illustration of the two transformations $P_\mathrm{d}^N$ and $P_\mathrm{c}^N$ for $N=10$ and $x_0\in L^\infty(I;\mathbb{R})$. Notice that $\lim_{N\rightarrow\infty} P_\mathrm{c}^N(P_\mathrm{d}^N(x_0))= x_0$.}\label{fig:Transfo}
\end{figure}
These transformations allow us to reveal an equivalence between the microscopic system \eqref{eq:syst-gen} and the continuous one \eqref{eq:GraphLimit-gen} in the case of piecewise-constant functions. More precisely, we have the following
\begin{prop}\label{prop:equiv}
The vectors $(x^{N},m^{N})\in \mathcal{C}([0,T]; \mathbb{R}^d)^{N}\times \mathcal{C}([0,T];\mathbb{R})^{N}$ satisfy the differential system~\eqref{eq:syst-gen} with initial condition $x^{0,N}\in(\mathbb{R}^d)^N$, $m^{0,N}\in\mathbb{R}^d$,
and mass dynamics given by \eqref{eq:psi},
if and only if the piecewise constant functions $x_N\in C([0,T];L^2(I;\mathbb{R}^d))$ and $m_N\in \mathcal{C}([0,T];L^2(I;\mathbb{R}))$ defined by
\eqref{eq:xN}
satisfy the following system of integro-differential equations:
\begin{equation}\label{eq:syst-contN}
\begin{cases}
\displaystyle \partial_t x_N(t,s) = \int_I m_N(t,s_*)\, \phi(x_N(t,s_*)-x_N(t,s)) \,ds_*\\
\displaystyle \partial_t m_N(t,s) = N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \psi(s_*, x_N(t,\cdot),m_N(t,\cdot)) \,ds_*
\end{cases}
\end{equation}
with initial conditions $\displaystyle x_N(0,s) = \sum_{i=1}^N x_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s)$ and $\displaystyle m_N(0,s) = \sum_{i=1}^N m_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s)$.
\end{prop}
\begin{rem}\label{Rem:x0m0}
Notice that according to the Lebesgue differentiation theorem, under the assumptions that $x_0\in L^1(I)$ and $m^0\in L^1(I)$, it holds
$$\lim_{N\rightarrow\infty} x_N(0,s) = x_0(s) \quad \text{ and } \quad \lim_{N\rightarrow\infty}m_N(0,s) = m_0(s) $$
for almost every $s\in I$. Figure \ref{fig:Transfo} illustrates the relationship between $x_0$ and $x_N(0,\cdot)=P_\mathrm{c}^N(P_\mathrm{d}^N(x_0))$ for a finite $N$.
\end{rem}
The main interest of this proposition is to have recast the discrete system \eqref{eq:syst-gen} into the framework of $L^\infty(I)$ functions. This will allow us to stay in this framework to state the convergence to a limit function, also belonging to $L^\infty(I)$.
Using this trick, we can now state one of the main results of this paper, namely the convergence of the solution to system \eqref{eq:syst-contN} to the solution to system \eqref{eq:GraphLimit-gen}.
\begin{theorem}\label{Th:GraphLimit}
Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R})$ satisfying \eqref{eq:integral_egal_a_1}, and consider a functional $\psi:I \times L^2(I;\mathbb{R}^d)\times L^2(I;\mathbb{R}) \rightarrow \mathbb{R}$.
Suppose that the function $\phi$ satisfies Hyp. \ref{hyp:phi}, that
$x\mapsto\psi(\cdot,x,m)$ and $m\mapsto\psi(\cdot,x,m)$
are uniformly Lipschitz functions in the $L^2$ norm, and that $m\mapsto\psi(\cdot,x,m)$ is sublinear in the $L^\infty$ norm.
Let $x^{0,N}=P_\mathrm{d}^N(x_0)$ and $m^{0,N}=P_\mathrm{d}^N(m_0)$ given by \eqref{eq:ICx}.
Then the solution $(x_N,m_N)$ to \eqref{eq:syst-contN} with initial conditions
\[\displaystyle x_N(0,s) = \sum_{i=1}^N x_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s) \quad \text{ and } \quad \displaystyle m_N(0,s) = \sum_{i=1}^N m_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s)
\]
converges when $N$ tends to infinity in the $\mathcal{C}([0,T];L^2(I))$ topology.
More specifically, there exists $(x,m)\in \mathcal{C}([0,T];L^2(I,\mathbb{R}^d))\times \mathcal{C}([0,T];L^2(I,\mathbb{R}))$ such that
\begin{equation}\label{eq:convxm}
\displaystyle \|x-x_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}^d))} \xrightarrow[N\rightarrow+\infty]{} 0
\quad \text{ and } \quad
\|m-m_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}))} \xrightarrow[N\rightarrow+\infty]{} 0.
\end{equation}
Furthermore, the limit functions $x$ and $m$ are solutions to the integro-differential system \eqref{eq:GraphLimit-gen} supplemented by the initial conditions $x(0,\cdot) = x_0$ and $m(0,\cdot) = m_0$.
\end{theorem}
Secondly, we will show that if the mass dynamics satisfy an \emph{indistinguishability} property (that we will define in Section \ref{sec:indisting}), we can take this limit process further and derive the mean-field limit of the system from the continuum graph limit. The mean-field limit will be shown to be a solution to the following transport equation with source (see \cite{PiccoliRossi14, PouradierDuteil21}):
\begin{equation*}
\partial_t \mu_t(x) + \nabla\cdot ( V[\mu_t](x) \mu_t(x)) = h[\mu_t](x)
\end{equation*}
where the non-local transport vector field $V$ and source term $h$ will be respectively defined from the interaction function $\phi$ and the mass dynamics $\psi$ (see Theorem \ref{Th:mfl}).
\section{The Graph Limit}
\label{sec:graphlim}
From here onward, in all of Section \ref{sec:graphlim}, we will assume the following properties for the mass dynamics $\psi$.
\begin{hyp}\label{hyp:psi}
The function $\psi:I\times L^\infty(I;\mathbb{R}^d)\times L^\infty(I;\mathbb{R})$ is assumed to satisfy the following Lipschitz properties: there exists $L_\psi>0$ such that for all $(x_1,x_2,m_1,m_2)\in L^2(I)^4$,
\begin{equation}\label{eq:psilip2}
\begin{cases}
\|\psi(\cdot,x_1,m_1)-\psi(\cdot,x_2,m_1)\|_{L^2(I)} \, \leq \,L_\psi \|x_1-x_2\|_{L^2(I)}\\
\|\psi(\cdot,x_1,m_1)-\psi(\cdot,x_1,m_2)\|_{L^2(I)} \, \leq \, L_\psi \|m_1-m_2\|_{L^2(I)}.
\end{cases}
\end{equation}
Assume also that there exists $C_\psi>0$ such that for all $(x,m)\in L^\infty(I,\mathbb{R}^d\times\mathbb{R})$, for all $s\in I$,
\begin{equation}\label{eq:psisublin}
|\psi(s, x, m)| \leq C_\psi( 1+\|m\|_{L^\infty(I)}).
\end{equation}
\end{hyp}
Although the assumption of sublinear growth \eqref{eq:psisublin} may seem restrictive, it is necessary in order to prevent the blow-up in finite-time of the weight function $m$.
It is coherent with the framework of the Graph Limit developed in~\cite{Medvedev14} on graphs with $L^\infty$ weights. Indeed, we can view our system \eqref{eq:GraphLimit-gen} as the evolution of the opinions $x$ on a weighted non-symmetric graph with weights $W(s,s_*)=m(t,s_*)$.
\subsection{Well posedness of the graph limit model}
This paragraph is devoted to proving the existence and uniqueness of a solution to the graph limit equation \eqref{eq:GraphLimit-gen}.
We start by proving existence and uniqueness for the decoupled system in the general case, where the mass dynamics are assumed to depend on the continuous index $s$.
In order to prove the well-posedness of system \eqref{eq:GraphLimit-gen}, we start by studying a decoupled system in which the dynamics of the opinions and of the weights are independent.
\begin{lemma}\label{Lemma:WellPos-decoupled}
Let ${\tilde{x} \in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d))}$ and ${\tilde{m} \in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}))}$. Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R})$.
Let $\phi$ satisfy Hyp. \ref{hyp:phi} and $\psi$ satisfy Hyp. \ref{hyp:psi}.
Then for any $T>0$, there exists a unique solution $(x,m)\in\mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}))$ to the decoupled integro-differential system
\begin{equation*}
\left\{\begin{array}{l}
\displaystyle \partial_t x(t,s) = \int_I \tilde{m}(t,s_*)\phi(x(t,s) - x(t,s_*)) ds_*; \qquad x(\cdot,0)=x_0 \\
\partial_t m(t,s) = \psi(s,\tilde{x}(t,\cdot),m(t,\cdot)); \qquad m(\cdot,0)=m_0.
\end{array}\right.
\end{equation*}
\end{lemma}
\begin{proof}
Since the two equations are decoupled, we will treat them independently.
Let $\tilde{T}>0$ (to be specified later such that $\tilde{T} \leq T$) and $M_{x_0}$ be a metric subspace of $\mathcal{C}([0,\tilde{T}];L^\infty(I;\mathbb{R}^d))$ consisting of functions $x$ satisfying $x(\cdot,0) = x_0$.
Let $K_{x_0}$ be the operator defined by:
$$
\begin{array}{l l l }
K_{x_0}: & M_{x_0} & \rightarrow M_{x_0} \\
& x & \displaystyle \mapsto (K_{x_0} x):(t,s)\mapsto x_0(s) + \int_0^t \int_I \tilde{m}(\tau,s_*)\phi(x(\tau,s) - x(\tau,s_*)) ds_* d\tau.
\end{array}
$$
We will show that $K_{x_0}$ is contracting for the norm $\|\cdot\|_{M_{x_0}} := \sup_{[0,\tilde{T}]} \esssup_{I} \|\cdot\|$.
Let $(x_1,x_2)\in \mathcal{C}([0,\tilde{T}];L^\infty(I;\mathbb{R}^d))^2)$.
Then for all $s\in I$, for all $t\leq \tilde{T}$,
\begin{equation*}
\begin{split}
\| K_{x_0} x_1 - K_{x_0} x_2 \| (t,s) & = \left \| \int_0^t \int_I \tilde{m}(\tau,s_*)\left[ \phi(x_1(\tau,s) - x_1(\tau,s_*))-\phi(x_2(\tau,s) - x_2(\tau,s_*))\right] ds_* d\tau \right \| \\
& \leq \int_0^t \int_I |\tilde{m}(\tau,s_*)| L_\phi \left\| (x_1(\tau,s) - x_1(\tau,s_*))- (x_2(\tau,s) - x_2(\tau,s_*))\right \| ds_* d\tau \\
& \leq \int_0^t \int_I |\tilde{m}(\tau,s_*)| L_\phi \left( \| x_1(\tau,s)- x_2(\tau,s)\| + \| x_1(\tau,s_*) - x_2(\tau,s_*)\| \right) ds_* d\tau \\
\end{split}
\end{equation*}
from which we get:
$$
\| K_{x_0} x_1 - K_{x_0} x_2 \|_{M_{x_0}} \leq 2 L_\phi \tilde{T} \sup_{t\in [0,T]} \|\tilde{m}(t,\cdot)\|_{L_1(I;\mathbb{R})} \quad \|x_1-x_2\|_{M_{x_0}}.
$$
We remark that $\sup_{t\in [0,T]} \|\tilde{m}(t,\cdot)\|_{L_1(I;\mathbb{R})}$ is well defined since $\tilde{m}\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}))$ and $I=[0,1]$.
Since $\tilde{m}$ is given, choosing $\tilde{T}\leq (4 L_\phi \sup_{t\in [0,T]} \|\tilde{m}(t,\cdot)\|_{L_1(I;\mathbb{R})})^{-1}$ ensures that $K_{x_0}$ is contracting on $[0,\tilde{T}]$.
By the Banach contraction mapping principle, there exists a unique solution $x\in \mathcal{C}([0,\tilde{T}];L^\infty(I;\mathbb{R}^d))$.
We then take $x(\cdot,\tilde{T})$ as the initial condition, and the local solution can be extended to $[0, 2\tilde{T}]$, and by repeating the same argument, to $[0,T]$. Moreover, since the integrand in $K_{x_0}$ is continuous as a map $L^\infty(I) \to L^\infty(I) $, $x$ is continuously differentiable and $x$ belongs to $\mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^d))$.\\
We now show existence and uniqueness of $m\in \mathcal{C}^1([0,T];L^2(I;\mathbb{R}))$, solution to the second decoupled equation.
Let $M_{m_0}$ be a metric subspace of $\mathcal{C}([0,\tilde{T}];L^2(I;\mathbb{R}))$ consisting of functions $m$ satisfying $m(\cdot,0) = m_0$ with again $\tilde{T}>0$ to be specified later such that $\tilde{T} \leq T$.
Let $K_{m_0}$ be the operator defined by:
$$
\begin{array}{l l l }
K_{m_0}: & M_{m_0} & \rightarrow M_{m_0} \\
& m &\displaystyle \mapsto (K_{m_0} m):(t,s)\mapsto m_0(s) + \int_0^t \psi(s,\tilde{x}(\tau,\cdot),m(\tau,\cdot)) d\tau.
\end{array}
$$
We will show that $K_{m_0}$ is contracting for the norm $\|\cdot\|_{M_{m_0}} := \sup_{[0,\tilde{T}]} \|\cdot\|_{L^2(I)}$.
Let $(m_1,m_2)\in \mathcal{C}([0,\tilde{T}];L^2(I;\mathbb{R}))^2$.
Then for all $s\in I$, for all $t\leq \tilde{T}$,
\begin{equation*}
\begin{split}
\int_I | K_{m_0} m_1 - K_{m_0} m_2 |^2 (t,s) ds & = \int_I \left | \int_0^t \psi(s,\tilde{x}(\tau,\cdot),m_1(\tau,\cdot))- \psi(s,\tilde{x}(\tau,\cdot),m_2(\tau,\cdot)) d\tau \right | ^2 ds \\
& \leq t \int_0^t \int_I \left | \psi(s,\tilde{x}(\tau,\cdot),m_1(\tau,\cdot))- \psi(s,\tilde{x}(\tau,\cdot),m_2(\tau,\cdot)) \right | ^2 ds \, d\tau \\
& \leq t \int_0^t L_\psi^2 \|m_1(\tau,\cdot)-m_2(\tau,\cdot)\|_{L^2(I)}^2 \, d\tau,
\end{split}
\end{equation*}
where the first inequality is a consequence of Cauchy-Schwarz and Jensen's inequalities and Fubini's theorem, and the second inequality comes from \eqref{eq:psilip2}.
We obtain:
\begin{equation*}
\sup_{[0,\tilde{T}]} \| K_{m_0} m_1 - K_{m_0} m_2 \|_{L^2(I)}^2 \leq \tilde{T}^2 L_\psi^2 \sup_{[0,\tilde{T}]} \| m_1 - m_2 \|_{L^2(I)}^2
\end{equation*}
which implies: $\| K_{m_0} m_1 - K_{m_0} m_2 \|_{M_{m_0}} \leq \tilde{T} L_\psi \| m_1 - m_2 \|_{M_{m_0}}$.
Thus, if $\tilde{T}\leq \frac{1}{2L_\psi}$, by the Banach contraction mapping principle, there exists a unique solution $m\in \mathcal{C}^1([0,\tilde{T}];L^2(I;\mathbb{R}))$.
We then take $m(\tilde{T}),\cdot)$ as the initial condition, and the local solution can be extended to $[0, 2\tilde{T}]$, and by repeating the same argument, to $[0,T]$.
We thus showed that there exists a unique solution $m\in \mathcal{C}([0,T];L^2(I;\mathbb{R}))$. To prove that $m\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}))$, we will use the second assumption on $ \psi$ given by the sub-linearity~\eqref{eq:psisublin}. For all $(t,s)\in I\times [0,T]$,
$$
|m(t,s)| = \left | m_0(s) + \int_0^t \psi(s,x(\tau,\cdot),m(\tau,\cdot)) d\tau \right|
\leq |m_0(s)| + \int_0^t C_\psi (1+ \|m(\tau,\cdot)\|_{L^\infty(I)}) d\tau.
$$
This implies
$$
\| m(t,\cdot)\|_{L^\infty(I)} \leq (\|m_0\|_{L^\infty(I)} + C_\psi t ) + \int_0^t C_\psi \|m(\tau,\cdot)\|_{L^\infty(I)} d\tau
$$
and from Gronwall's lemma,
\begin{equation}\label{eq:mbound}
\| m(t,\cdot)\|_{L^\infty(I)} \leq (\|m_0\|_{L^\infty(I)} + C_\psi t ) e^{C_\psi t}.
\end{equation}
Hence $m\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}))$. As previously, since the integrand in $K_{m_0}$ is continuous as a map $L^\infty(I) \to L^\infty(I) $, $m$ is continuously differentiable and $m$ belongs to $\mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}))$. This concludes the proof.
\end{proof}
\begin{rem}\label{Rem:Gronwall}
We have used the following result:
If $u(t)\leq \alpha(t) + \int_0^t \beta(\tau) u(\tau) d\tau$, where $\beta$ is positive and $\alpha$ is non-decreasing,
then $u(t) \leq \alpha(t)\exp(\int_0^t \beta(\tau) d\tau )$.
\end{rem}
By the previous lemma, we have proven that the two decoupled integro-differential equations are well-posed in $\mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}))$. Using this result, we are now ready to demonstrate the well-posedness of the fully coupled system \eqref{eq:GraphLimit-gen}.
\begin{theorem}\label{Th:GL-wellpos}
Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R})$.
Let $\phi$ satisfy Hyp. \ref{hyp:phi} and $\psi$ satisfy Hyp. \ref{hyp:psi}.
Then for any $T>0$, there exists a unique solution $(x,m)\in {\mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}))}$ to the integro-differential system
\begin{equation}
\label{eq:graphlimitcoupled}
\left\{\begin{array}{l}
\displaystyle \partial_t x(t,s) = \int_I m(t,s_*)\phi(x(t,s) - x(t,s_*)) ds_*; \qquad x(\cdot,0)=x_0 \\
\partial_t m(t,s) = \psi(s,x(t,\cdot),m(t,\cdot)); \qquad m(\cdot,0)=m_0.
\end{array}\right.
\end{equation}
\end{theorem}
\begin{proof}
The proof will consist of proving the convergence of a sequence of functions $(x^n,m^n)_{n\in\mathbb{N}}$ of $\mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}))$ defined as follows:
\begin{itemize}
\item For almost every $s \in I$, for all $t \in [0,T]$, $x^0(t,s) = x_0(s)$ and $m^0(t,s) = m_0(s)$.
\item For all $n\in \mathbb{N}^*$,
\begin{equation*}
\left\{\begin{array}{l}
\displaystyle \partial_t x^n(t,s) = \int_I m^{n-1}(t,s_*)\phi(x^n(t,s) - x^n(t,s_*)) ds_*; \qquad x^n(\cdot,0)=x_0 \\
\partial_t m^n(t,s) = \psi(s,x^{n-1}(t,\cdot),m^n(t,\cdot)); \qquad m^n(\cdot,0)=m_0.
\end{array}\right.
\end{equation*}
\end{itemize}
From Lemma \ref{Lemma:WellPos-decoupled}, the sequence is well defined and each term $(x^n,m^n)$ is indeed in {$\mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}))$}.
We begin by highlighting the fact that the terms of the sequence are uniformly bounded in $L^\infty(I;\mathbb{R}^d\times\mathbb{R})$.
Indeed, from equation \eqref{eq:mbound}, we know that for every $n\in \mathbb{N}$, for all $t\in [0,T]$,
$$ \| m^n(t,\cdot)\|_{L^\infty(I)} \leq M_T := (\|m_0\|_{L^\infty(I)} + C_\psi T ) e^{C_\psi T}.$$
Moreover, we can now use this bound to estimate the growth of $x$, noticing that Hypothesis \ref{hyp:phi} implies that $\|\phi(r)\| \leq L_\phi \|r\|$, from the fact that $\phi(0)=0$. We estimate:
$$
\|x^n(t,s)\| \leq \|x^n(0,s)\| + M_T \int_0^t \int_I \| \phi(x^n(\tau,s_*)-x^n(\tau,s) ) \| ds_* d\tau \leq \|x_0(s)\| + 2 M_T L_\phi \int_0^t \|x^n(\tau,\cdot) \|_{L^\infty(I)} d\tau .
$$
Then Gronwall's lemma implies that for every $n\in \mathbb{N}$, for all $t\in [0,T]$,
\begin{equation*}
\|x^n(t,\cdot)\|_{L^\infty(I)} \leq X_T := \|x_0\|_{L^\infty(I)} e^{2 M_T L_\phi T}.
\end{equation*}
We now prove that the sequence $(x^n,m^n)_{n\in \mathbb{N}}$ is a Cauchy sequence.
For every $n\in \mathbb{N}^*$,
\begin{equation*}
\begin{split}
|x^{n+1}-x^n|(t,s) = & \bigg | \int_0^t \int_I m^{n}(\tau,s_*)\phi(x^{n+1}(\tau,s) - x^{n+1}(\tau,s_*)) ds_* d\tau \\
& - \int_0^t \int_I m^{n-1}(\tau,s_*)\phi(x^n(\tau,s) - x^n(\tau,s_*)) ds_* d\tau \bigg | \\
= & \bigg | \int_0^t \int_I (m^{n}(\tau,s_*)-m^{n-1}(\tau,s_*)) \phi(x^{n+1}(\tau,s) - x^{n+1}(\tau,s_*)) ds_* d\tau \\
& + \int_0^t \int_I m^{n-1}(\tau,s_*)\left( \phi(x^{n+1}(\tau,s) - x^{n+1}(\tau,s_*)) - \phi(x^n(\tau,s) - x^n(\tau,s_*)) \right) ds_* d\tau \bigg | \\
\leq & \int_0^t \int_I |m^{n}(\tau,s_*)-m^{n-1}(\tau,s_*)| 2 L_\phi\, X_T \, ds_* d\tau \\
& + \int_0^t \int_I M_T \, L_\phi \|(x^{n+1}(\tau,s) - x^{n+1}(\tau,s_*)) - (x^n(\tau,s) - x^n(\tau,s_*)) \| ds_* d\tau \\
\end{split}
\end{equation*}
Then for all $t\in [0,T]$, for every $n\in \mathbb{N}^*$,
\begin{equation*}
\begin{split}
|x^{n+1}-x^n|^2(t,s) \leq & 8 t L_\phi^2 \, X_T^2 \int_0^t \int_I |m^{n}(\tau,s_*)-m^{n-1}(\tau,s_*)|^2 \, ds_* d\tau \\
& + 2 L_\phi^2 M_T^2 t \int_0^t \int_I 2(\|(x^{n+1}(\tau,s) - x^n(\tau,s)\|^2 + \| x^{n+1}(\tau,s_*)- x^n(\tau,s_*) \|^2) ds_* d\tau
\end{split}
\end{equation*}
and we get
\begin{equation*}
\begin{split}
\|x^{n+1}-x^n\|_{L^2(I)}^2(t) \leq & 8 t L_\phi^2 \, X_T^2 \int_0^t \|m^{n}-m^{n-1}\|_{L^2(I)}^2 (\tau) d\tau + 8t L_\phi^2 M_T^2 \int_0^t \|x^{n+1} - x^n\|_{L^2(I)}^2 (\tau) d\tau.
\end{split}
\end{equation*}
A similar computation for $m$ gives for every $n\in\mathbb{N}$
\begin{equation*}
\begin{split}
|m^{n+1}-m^n|(t,s) = & \bigg | \int_0^t (\psi(s,x^n(\tau,\cdot),m^{n+1}(\tau,\cdot)) - \psi(s,x^{n-1}(\tau,\cdot),m^{n}(\tau,\cdot)) d\tau \bigg | \\
\leq & \int_0^t \bigg ( |\psi(s,x^n(\tau,\cdot),m^{n+1}(\tau,\cdot)) - \psi(s,x^n(\tau,\cdot),m^{n}(\tau,\cdot)) | \\
& + \psi(s,x^n(\tau,\cdot),m^{n}(\tau,\cdot)) - \psi(s,x^{n-1}(\tau,\cdot),m^{n}(\tau,\cdot)) | \bigg ) d\tau . \\
\end{split}
\end{equation*}
Squaring and integrating yields:
\begin{equation}\label{eq:ineqmn}
\begin{split}
\|m^{n+1}-m^n\|_{L^2(I)}^2(t)
\leq & 2 t \int_0^t \int_I |\psi(s,x^n(\tau,\cdot),m^{n+1}(\tau,\cdot)) - \psi(s,x^n(\tau,\cdot),m^{n}(\tau,\cdot)) |^2 ds \, d\tau \\
& + 2t \int_0^t \int_I | \psi(s,x^n(\tau,\cdot),m^{n}(\tau,\cdot)) - \psi(s,x^{n-1}(\tau,\cdot),m^{n}(\tau,\cdot)) | ^2 ds \, d\tau \\
\leq & 2 t L_\psi^2 \int_0^t \|m^{n+1}-m^{n}\|_{L^2(I)}^2(\tau) d\tau + 2 t L_\psi^2 \int_0^t \| x^n - x^{n-1}\|_{L^2(I)}^2(\tau) d\tau .
\end{split}
\end{equation}
Thus, denoting $A_T = \max(8 T L_\phi^2 \, X_T^2,\, 8 T L_\phi^2 \, M_T^2,\, 2 T L_\psi^2)$, and denoting by $(u_n)_{n\in\mathbb{N}}$ the sequence defined by
$u_n(t) := \|x^{n+1}-x^n\|_{L^2(I)}^2(t) + \|m^{n+1}-m^n\|_{L^2(I)}^2(t)$, we obtain for every $n\in \mathbb{N}^*$ and all $t\in [0,T]$:
\begin{equation*}
\begin{split}
u_{n}(t) \leq
A_T \int_0^t u_n(\tau) d\tau + A_T \int_0^t u_{n-1}(\tau) d\tau .
\end{split}
\end{equation*}
Since $t\mapsto A_T \int_0^t u_{n-1}(\tau) d\tau$ is non-decreasing, Gronwall's lemma implies (see Remark \ref{Rem:Gronwall}):
$$
u_{n}(t) \leq A_T \; e^{A_T T} \int_0^t u_{n-1}(\tau) d\tau .
$$
Denoting $U_0 := \sup_{[0,T]} u_0(t)$, one can easily show by induction that for all $t\in [0,T]$, for all $n\in \mathbb{N}$,
$$
u_n(t) \leq \frac{(A_T e^{A_T T}t)^n}{n!} U_0.
$$
This is the general term of a convergent series, hence for all $t\in [0,T]$, $\lim_{n\rightarrow \infty } u_n(t) = 0$, which implies that for all $t\in [0,T]$, $\|x^{n+1}-x^n\|_{L^2(I)}(t)$ and $\|m^{n+1}-m^n\|_{L^2(I)}(t)$ also converge to $0$ as $n$ tends to infinity.
Thus, $(x^n)_{n\in\mathbb{N}}$ and $(m^n)_{n\in\mathbb{N}}$ are Cauchy sequences in the Banach spaces $\mathcal{C}([0,T];L^2(I;\mathbb{R}^d))$ and $\mathcal{C}([0,T];L^2(I;\mathbb{R}))$.
One can easily show that their limits $(x,m)$ satisfy the integro-differential system \eqref{eq:GraphLimit-gen}. Furthermore, from the uniform bounds on $\|x^n(t,\cdot)\|_{L^\infty(I)}$ and $\|m^n(t,\cdot)\|_{L^\infty(I)}$, we deduce that $(x,m)\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}))$, with $\|x(t,\cdot)\|_{L^\infty(I)} \leq X_T$ and $\|m(t,\cdot)\|_{L^\infty(I)} \leq M_T$ for all $t\in [0,T]$. Finally, as previously, looking at the integrand in the integral formulation of \eqref{eq:graphlimitcoupled}, we deduce that $(x,m)\in \mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}))$ which concludes the proof of existence.
\\
Let us now deal with the uniqueness. Let us assume that there exist two solutions to the equation~\eqref{eq:graphlimitcoupled} denoted $(x_1,m_1)$ and $(x_2,m_2)$ with the same initial condition. Then, we have
\begin{equation*}
\left\{\begin{array}{l}
\displaystyle (x_1-x_2)(t,s) = \int_0^t \int_I m_1(\tau,s_*) \phi(x_1(\tau,s_*)-x_1(\tau,s))ds_* d\tau\\
\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ - \int_0^t \int_I m_2(\tau,s_*) \phi(x_2(\tau,s_*)-x_2(\tau,s))ds_* d\tau \\
\displaystyle (m_1-m_2)(t,s) = \int_0^t (\psi(s,x_1(\tau,\cdot), m_1(\tau,\cdot)) -\psi(s,x_2(\tau,\cdot), m_2(\tau,\cdot))) d\tau
\end{array}\right.
\end{equation*}
that we rewrite
\begin{equation*}
\left\{\begin{array}{l}
\displaystyle (x_1-x_2)(t,s) = \int_0^t \int_I (m_1-m_2)(\tau,s_*) \phi(x_1(\tau,s_*)-x_1(\tau,s))ds_* d\tau\\
\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + \int_0^t \int_I m_2(\tau,s_*) (\phi(x_1(\tau,s_*)-x_1(\tau,s)) -\phi(x_2(\tau,s_*)-x_2(\tau,s)))ds_* d\tau \\
\displaystyle (m_1-m_2)(t,s) = \int_0^t (\psi(s,x_1(\tau,\cdot), m_1(\tau,\cdot)) -\psi(s,x_1(\tau,\cdot), m_2(\tau,\cdot))) d\tau\\
\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + \int_0^t (\psi(s,x_1(\tau,\cdot), m_2(\tau,\cdot)) -\psi(s,x_2(\tau,\cdot), m_2(\tau,\cdot))) d\tau
\end{array}\right.
\end{equation*}
Thus, we have
\begin{equation*}
\begin{array}{l}
\displaystyle |x_1-x_2|(t,s) \leq \int_0^t \int_I 2 L_\phi X_T |m_1-m_2|(\tau,s_*) ds_* d\tau + \int_0^t \int_I M_TL_\phi (|x_1-x_2|(\tau,s_*) + |x_1-x_2|(\tau,s)) ds_* d\tau
\end{array}
\end{equation*}
from which we deduce
\begin{equation*}
\begin{array}{l}
\displaystyle |x_1-x_2|^2(t,s) \leq 8 L_\phi^2 X_T^2 t \int_0^t \int_I |m_1-m_2|^2(\tau,s_*) ds_* d\tau \\
\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + 2 M_T^2 L_\phi^2 t \int_0^t \int_I |x_1-x_2|^2(\tau,s_*) ds_* d\tau + 2 M_T^2 L_\phi^2 t \int_0^t |x_1-x_2|^2(\tau,s) d\tau.
\end{array}
\end{equation*}
Thus, we have,
\begin{equation*}
\begin{array}{l}
\displaystyle \|x_1-x_2\|^2_{L^2(I)}(t) \leq A_T \left( \int_0^t \|m_1-m_2\|^2_{L^2(I)}(\tau) d\tau + \int_0^t \|x_1-x_2\|^2_{L^2(I)}(\tau) \right) d\tau.
\end{array}
\end{equation*}
Similarly,
\begin{equation*}
\begin{array}{rl}
\displaystyle \|m_1-m_2\|^2_{L^2(I)}(t) &\leq 2 t \int_0^t \|\psi(\cdot,x_1(\tau),m_1(\tau)) - \psi(\cdot,x_1(\tau),m_2(\tau)) \|_{L^2(I)}^2 d\tau \\
&\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + 2 t \int_0^t \|\psi(\cdot,x_1(\tau),m_2(\tau)) - \psi(\cdot,x_2(\tau),m_2(\tau)) \|_{L^2(I)}^2 d\tau\\
&\displaystyle \leq 2t L_\psi \int_0^t \left(\|m_1-m_2\|^2_{L^2(I)}(\tau) + \|x_1-x_2\|^2_{L^2(I)}(\tau) \right) d\tau.
\end{array}
\end{equation*}
Finally,
\begin{equation*}
\|x_1-x_2\|^2_{L^2(I)}(t) + \|m_1-m_2\|^2_{L^2(I)}(t) \leq 2 A_T \int_0^t \left(\|x_1-x_2\|^2_{L^2(I)}(\tau) + \|m_1-m_2\|^2_{L^2(I)}(\tau) \right) d\tau.
\end{equation*}
By Gronwall lemma, we deduce that, for all $t \in [0,T]$,
\begin{equation*}
\|x_1-x_2\|^2_{L^2(I)}(t) + \|m_1-m_2\|^2_{L^2(I)}(t) = 0,
\end{equation*}
which concludes the proof of uniqueness.
\end{proof}
\subsection{Well posedness of the microscopic system}
In this paragraph, we state the existence and uniqueness results of a solution to the discrete system \eqref{eq:syst-gen}.
We do not provide the proof since it consists on a straightforward adaptation at the discrete level of the proof established for the continuous case, the graph limit equation, based on a use of the fixed point theorem.
\begin{theorem}
Let $(x^{0,N},m^{0,N}) \in \mathbb{R}^{dN} \times \mathbb{R}^N$.
Let $\phi$ satisfy Hyp.~\ref{hyp:phi}, $\psi$ satisfy Hyp.~\ref{hyp:psi},
and $\psi_i^{(N)}$ be defined by \eqref{eq:psi}.
Then for any $T>0$, there exists a unique solution $(x^{N},m^{N}) \in \mathcal{C}^1([0,T];\mathbb{R}^{dN} \times \mathbb{R}^N)$ to the discrete system \eqref{eq:syst-gen} with initial condition $(x^{0,N},m^{0,N}) \in \mathbb{R}^{dN} \times \mathbb{R}^N$.
Moreover, there exists constants $\overline{X}$ and $\overline{M}$ such that for all $t \in [0,T]$, for all $i \in \{1, \dots, N \}$,
\begin{equation}
\label{eq:bornex}
\| x_i^{(N)}(t)\| \leq \overline{X}
\end{equation}
and
\begin{equation}
\label{eq:bornem}
| m_i^{(N)}(t)| \leq \overline{M}.
\end{equation}
\end{theorem}
\begin{rem}
Although this theorem provides existence and uniqueness of solutions to system \eqref{eq:syst-gen} for the special class of mass dynamics given by \eqref{eq:psi}, it can easily be adapted to general mass dynamics $\psi^{(N)}_i(x^N,m^N)$ satisfying the two conditions:
there exist $L_\psi>0$ and $C_\psi>0$ such that
for all $x^N,y^N\in (\mathbb{R}^d)^N$, for all $m^N,p^N\in\mathbb{R}^N$,
\begin{equation*}
\begin{cases}
\|\psi^{(N)}(x^N,m^N)-\psi^{(N)}(y^N,m^N)\| \, \leq \,L_\psi \|x^N-y^N\| \\
\|\psi^{(N)}(x^N,m^N)-\psi^{(N)}(x^N,p^N)\| \, \leq \, L_\psi \|m^N-p^N\|,
\end{cases}
\end{equation*}
where $\|\cdot\|$ denotes the standard Euclidean norm in $(\mathbb{R}^d)^N$ or in $\mathbb{R}^N$,
and for every $i\in\{1,\cdots,N\}$
\begin{equation*}
\displaystyle |\psi_i( x^N, m^N)| \leq C_\psi( 1 + max_{i\in\{1,\cdots,N\}} |m^N_i| ).
\end{equation*}
\end{rem}
\subsection{Convergence to the graph limit equation}
Let us now prove the main result of this article, namely that under our assumptions, the solution to the discrete problem \eqref{eq:syst-gen} converges to the solution to the integro-differential equation \eqref{eq:GraphLimit-gen} when $N$ goes to infinity.
For clarity purposes, we restate the main theorem announced in Section \ref{sec:presentation}.
\begin{maintheorem
Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R})$ satisfying \eqref{eq:integral_egal_a_1}, and consider a functional $\psi:I \times L^2(I;\mathbb{R}^d)\times L^2(I;\mathbb{R}) \rightarrow \mathbb{R}$.
Suppose that the function $\phi$ satisfies Hyp.~\ref{hyp:phi}, and that $\psi$ satisfies Hyp.~\ref{hyp:psi}.
Let $x^{0,N}:=P_\mathrm{d}^N(x_0)$ and $m^{0,N}:=P_\mathrm{d}^N(m_0)$ be given by \eqref{eq:ICx}.
Then the solution $(x_N,m_N)$ to \eqref{eq:syst-contN} with initial conditions
\[\displaystyle x_N(0,s) = \sum_{i=1}^N x_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s) \quad \text{ and } \quad \displaystyle m_N(0,s) = \sum_{i=1}^N m_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s)
\]
converges when $N$ tends to infinity in the $\mathcal{C}([0,T];L^2(I))$ topology.
More specifically, there exist $(x,m)\in \mathcal{C}([0,T];L^\infty(I,\mathbb{R}^d))\times \mathcal{C}([0,T];L^\infty(I,\mathbb{R}))$ such that
\begin{equation}\label{eq:convxm}
\displaystyle \|x-x_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}^d))} \xrightarrow[N\rightarrow+\infty]{} 0
\quad \text{ and } \quad
\|m-m_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}))} \xrightarrow[N\rightarrow+\infty]{} 0.
\end{equation}
Furthermore, the limit functions $x$ and $m$ are solutions to the integro-differential system \eqref{eq:GraphLimit-gen} supplemented by the initial conditions $x(0,\cdot) = x_0$ and $m(0,\cdot) = m_0$.
\end{maintheorem}
\begin{rem}
Theorem \ref{Th:GraphLimit} proves the convergence of the solution to \eqref{eq:syst-contN} to the solution to \eqref{eq:GraphLimit-gen}. This is equivalent to proving the convergence of the solution to the discrete system \eqref{eq:syst-gen} to the solution to \eqref{eq:GraphLimit-gen}, as shown in Proposition \ref{prop:equiv}.
\end{rem}
\begin{proof}
The proof is done following the graph limit method of \cite{Medvedev14}.
Let $\xi_N:= x_N - x$ and $\zeta_N := m_N -m$.
We will also use the slight abuse of notation $y(t,\cdot):=y(t)$, with $y$ standing for $x$, $x_N$, $m$ or $m_N$.
We compute
\begin{equation*}
\begin{split}
\frac{\partial\xi_N(t,s)}{\partial t} = & \int_I m_N(t,s_*)\, \phi(x_N(t,s_*)-x_N(t,s)) \, ds_* - \int_I m(t,s_*)\, \phi(x(t,s_*)-x(t,s)) \, ds_* \\
= & \int_I m_N(t,s_*)\, \left[ \phi(x_N(t,s_*)-x_N(t,s))- \phi(x(t,s_*)-x(t,s)) \right] \, ds_* \\
& + \int_I (m_N(t,s_*)- m(t,s_*)) \, \phi(x(t,s_*)-x(t,s)) \, ds_*.
\end{split}
\end{equation*}
By multiplying by $\xi_N$ and integrating over $I$, we obtain
\begin{equation}\label{eq:dxi}
\begin{split}
\frac{1}{2}\int_I \frac{\partial\xi_N(t,s)^2}{\partial t} ds
= & \int_{I^2} m_N(t,s_*)\, \left[ \phi(x_N(t,s_*)-x_N(t,s))- \phi(x(t,s_*)-x(t,s)) \right] \xi_N(t,s) \, ds_*\, ds \\
& + \int_{I^2} \zeta_N(t,s_*) \xi_N(t,s) \, \phi(x(t,s_*)-x(t,s)) \, ds_*\, ds.
\end{split}
\end{equation}
We study the first term. Since the solution to \eqref{eq:syst-gen} satisfies \eqref{eq:bornex}-\eqref{eq:bornem}, we have
\begin{equation}
\label{eq:bornemN}
\sup_{t \in [0,T]} \esssup_{s \in I} |m_N(t,s)| \leq \overline{M}.
\end{equation}
Then, since $\phi$ is Lipschitz, there exists $L>0$ such that
\begin{equation*}
\begin{split}
& \left | \int_{I^2} m_N(t,s_*)\, \left[ \phi(x_N(t,s_*)-x_N(t,s))- \phi(x(t,s_*)-x(t,s))
\right] \xi_N(t,s)\, ds_*\, ds \right| \\ \leq \;& \overline{M} L \int_{I^2} |\xi_N(t,s)| \,\left| x_N(t,s_*)-x_N(t,s)- x(t,s_*)+x(t,s) \right| \, ds_*\, ds \\
\leq \;& \overline{M} L \int_{I^2} |\xi_N(t,s)| \,\left| \xi_N(t,s_*) - \xi_N(t,s) \right| \, ds_*\, ds \leq 2 \overline{M} L \|\xi_N(t)\|^2_{L^2(I)}
\end{split}
\end{equation*}
We now look at the second term of \eqref{eq:dxi}.
Since $x \in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d))$ and $\phi$ is continuous, there exists a constant $\displaystyle{C_\phi:={\mathrm{ess}\sup}_{(s,s_*,t)\in I^2\times [0,T]}|\phi(x(t,s_*)-x(t,s))| }$ which is finite, and we have the following bound:
\begin{equation*}
\begin{split}
& \int_{I^2} \zeta_N(t,s_*) \xi_N(t,s) \, \phi(x(t,s_*)-x(t,s)) \, ds_*\, ds
\leq \; C_\phi \; \int_{I^2} |\zeta_N(t,s_*) \xi_N(t,s)| ds_*\, ds \\
\leq & \; C_\phi \; \|\xi_N(t)\|_{L^1(I)}\, \|\zeta_N(t)\|_{L^1(I)} \; \leq \; C_\phi \; \|\xi_N(t)\|_{L^2(I)}\, \|\zeta_N(t)\|_{L^2(I)}.
\end{split}
\end{equation*}
Hence from \eqref{eq:dxi},
\begin{equation}\label{eq:dxi2}
\frac{1}{2}\frac{d}{dt}\|\xi_N(t)\|_{L^2(I)}^2 \, \leq \, 2 \overline{M} L \|\xi_N(t)\|^2_{L^2(I)} + C_\phi \; \|\xi_N(t)\|_{L^2(I)}\, \|\zeta_N(t)\|_{L^2(I)}.
\end{equation}
We now compute
\begin{equation*}
\begin{split}
\frac{\partial\zeta_N(t,s)}{\partial t} = & N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \psi(s_*,x_N(t),m_N(t)) \,ds_* - \psi(s,x(t),m(t))\\
= & N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right] \,ds_* \\
& + N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \psi(s_*,x(t),m(t) \,ds_* - \psi(s,x(t),m(t)).
\end{split}
\end{equation*}
Multiplying by $\zeta_N(t,s)$ and integrating over $I$, we get:
\begin{equation}\label{eq:dzn}
\begin{split}
\frac{1}{2}\int_I \frac{\partial\zeta_N(t,s)^2}{\partial t} \, ds = & \int_I N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right] ds_*\, \zeta_N(t,s) \, ds \\
& + \int_I\left[ N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \psi(s_*,x(t),m(t)) \,ds_* - \psi(s,x(t),m(t)) \right] \zeta_N(t,s)\, ds \\
\leq & \; \|h_N(t)\|_{L^2(I)} \|\zeta_N(t)\|_{L^2(I)} \; + \; \|g_N(t)\|_{L^2(I)} \, \|\zeta_N(t)\|_{L^2(I)} ,
\end{split}
\end{equation}
where the last inequality was obtained from the Cauchy-Schwarz inequality, and we denoted by {$h_N$ and $g_N$} the functions
$$
h_N:(t,s) \mapsto N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right] ds_*$$
and
$$
g_N:(t,s) \mapsto N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \psi(s_*,x(t),m(t)) \,ds_* - \psi(s,x(t),m(t)).
$$
We start with the first term:
\begin{equation*}
\begin{split}
\|h_N(t)\|_{L^2(I)}^2 & = \int_I \left( N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right] ds_*\, \right)^2ds \\
& = \sum_{i=1}^N\, \int_{\frac{i-1}{N}}^{\frac{i}{N}} N^2 \left( \int_{\frac{i-1}{N}}^{\frac{i}{N}} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right] ds_*\, \right)^2ds \\
& = \sum_{i=1}^N\, N \left( \int_{\frac{i-1}{N}}^{\frac{i}{N}} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right]\cdot 1 \, ds_*\, \right)^2 \\
& \leq \sum_{i=1}^N\, N \left( \left[ \int_{\frac{i-1}{N}}^{\frac{i}{N}} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right]^2 ds_*\,\right]^{1/2} \left[ \int_{\frac{i-1}{N}}^{\frac{i}{N}} 1^2 ds_*\,\right]^{1/2}\right)^2 \\
& \leq \sum_{i=1}^N\, \int_{\frac{i-1}{N}}^{\frac{i}{N}} \left[ \psi(s_*,x_N(t),m_N(t)) - \psi(s_*,x(t),m(t)) \right]^2 ds_* \\
& = \| \psi(\cdot,x_N(t),m_N(t)) - \psi(\cdot,x(t),m(t)) \|_{L^2(I)}^2,
\end{split}
\end{equation*}
using the Cauchy Schwarz inequality.
Then, from \eqref{eq:psilip2},
\begin{equation*}
\begin{split}
\|h_N(t)\|_{L^2(I)} & \leq \| \psi(\cdot,x_N(t),m_N(t)) - \psi(\cdot,x(t),m_N(t)) \|_{L^2(I)} + \| \psi(\cdot,x(t),m_N(t)) - \psi(\cdot,x(t),m(t)) \|_{L^2(I)} \\
& \leq L_\psi ( \| x_N(t)-x(t) \|_{L^2(I)} + \| m_N(t)-m(t) \|_{L^2(I)}) = L_\psi ( \| \xi_N(t) \|_{L^2(I)} + \| \zeta_N(t) \|_{L^2(I)}).
\end{split}
\end{equation*}
We now study the second term of \eqref{eq:dzn}.
According to Lebesgue's differentiation theorem, for almost every $s\in I$,
$$
\lim_{N\rightarrow+\infty} N \int_{\frac{1}{N}\floor{sN}}^{\frac{1}{N}(\floor{sN}+1)} \psi(s_*,x(t),m(t)) \,ds_* = \psi(s,x(t),m(t)).
$$
which implies
\begin{equation}\label{eq:h2}
\lim_{N\rightarrow+\infty} \|g_N(t)\|_{L^2(I)} = 0.
\end{equation}
Summing up the contributions of $h_N$ and $g_N$, from \eqref{eq:dzn} we obtain:
\begin{equation}\label{eq:dzn2}
\frac{1}{2}\frac{d}{dt}\|\zeta_N(t)\|_{L^2(I)}^2 \, \leq \, L_\psi \left( \| \xi_N(t) \|_{L^2(I)} + \| \zeta_N(t) \|_{L^2(I)} \right) \|\zeta_N(t)\|_{L^2(I)} + \|g_N(t)\|_{L^2(I)} \|\zeta_N(t)\|_{L^2(I)}.
\end{equation}
Now, for all $\epsilon >0$ and $t\in[0,T]$, let $\bar{\zeta}_N^\epsilon(t):=\sqrt{\|\zeta_N(t)\|_{L^2(I)}^2+\epsilon}$ and $\bar{\xi}_N^\epsilon(t):=\sqrt{\|\xi_N(t)\|_{L^2(I)}^2+\epsilon}$.
From \eqref{eq:dxi2} and \eqref{eq:dzn2},
\begin{equation*}
\begin{cases}
\frac{1}{2}\frac{d}{dt}(\bar{\xi}_N^\epsilon(t)^2) \, \leq \, 2 L \overline{M} \, \bar{\xi}_N^\epsilon(t)^2 + C_\phi \; \bar{\xi}_N^\epsilon(t) \, \bar{\zeta}_N^\epsilon(t)\\
\frac{1}{2}\frac{d}{dt}(\bar{\zeta}_N^\epsilon(t)^2) \, \leq \, L_\psi \left( \bar{\xi}_N^\epsilon(t) + \bar{\zeta}_N^\epsilon(t) \right) \bar{\zeta}_N^\epsilon(t) + \|g_N(t)\|_{L^2(I)} \bar{\zeta}_N^\epsilon(t),
\end{cases}
\end{equation*}
and since for all $t\in [0,T]$, $\bar{\xi}_N^\epsilon(t)>0$ and $\bar{\zeta}_N^\epsilon(t)>0$, this implies:
\begin{equation*}
\begin{cases}
\frac{d}{dt}\bar{\xi}_N^\epsilon(t) \, \leq \, 2 L \overline{M} \, \bar{\xi}_N^\epsilon(t) + C_\phi \, \bar{\zeta}_N^\epsilon(t)\\
\frac{d}{dt}\bar{\zeta}_N^\epsilon(t) \, \leq \, L_\psi \left( \bar{\xi}_N^\epsilon(t) + \bar{\zeta}_N^\epsilon(t) \right) + \|g_N(t)\|_{L^2(I)} .
\end{cases}
\end{equation*}
Summing up, we get
$$
\frac{d}{dt}(\bar{\xi}_N^\epsilon(t)+\bar{\zeta}_N^\epsilon(t)) \, \leq \, K (\bar{\xi}_N^\epsilon(t)+\bar{\zeta}_N^\epsilon(t)) + \|g_N(t)\|_{L^2(I)}
$$
where $K:=\max\{ 2 L \overline{M} ,\, C_\phi,\, L_\psi\}$. Now, from Gronwall's inequality, for all $t\in [0,T]$,
$$
\bar{\xi}_N^\epsilon(t)+\bar{\zeta}_N^\epsilon(t) \, \leq \, (\bar{\xi}_N^\epsilon(0)+\bar{\zeta}_N^\epsilon(0))e^{Kt} + \int_0^t \|g_N(\tau)\|_{L^2(I)}e^{K(t-\tau)}d\tau.
$$
Since $\epsilon$ is arbitrary, we obtain:
$$
\sup_{t\in [0,T]} \left(\|\xi_N(t)\|_{L^2(I)}+\|\zeta_N(t)\|_{L^2(I)} \right) \leq \left(\|\xi_N(0)\|_{L^2(I)}+\|\zeta_N(0)\|_{L^2(I)}+\int_0^T \|g_N(\tau)\|_{L^2(I)}d\tau \right) e^{KT}.
$$
As seen in Remark \ref{Rem:x0m0}, $\lim_{N\rightarrow +\infty} \xi_N(0,s) = \lim_{N\rightarrow +\infty} x_N(0,s)-x(0,s) = 0$ and
$\lim_{N\rightarrow +\infty} \zeta_N(0,s) = \lim_{N\rightarrow +\infty} m_N(0,s)-m(0,s) = 0$ for almost all $s\in I$, which implies
that $\lim_{N\rightarrow +\infty}\|\xi_N(0)\|_{L^2(I)}=0$ and $\lim_{N\rightarrow +\infty}\|\zeta_N(0)\|_{L^2(I)}=0$.
From \eqref{eq:h2}, for all $t$, $\lim_{N\rightarrow +\infty}\|g_N(t)\|_{L^2(I)}=0$, so using the dominated convergence theorem for the last term, we can finally deduce the convergence result \eqref{eq:convxm}.
\end{proof}
\section{Relation between Graph Limit and Mean-field Limit}
\label{sec:mfl}
In Section \ref{sec:graphlim}, we have derived the \emph{Graph Limit} of the microscopic model \eqref{eq:syst-gen} when the number of agents $N$ goes to infinity, showing that the limit functions representing the opinions and weights satisfy a system of integro-differential equations.
The aim of this section is to relate the \emph{Graph Limit} that we obtained with the \emph{Mean-Field Limit}, much more studied in the field of collective dynamics.
However, the Mean-Field Limit can only be derived for a particular form of mass dynamics that satisfy an \emph{indistinguishability} property. We begin by shedding light on this concept.
\subsection{Indistinguishability and mean-field limit}
\label{sec:indisting}
In the context of the classical Hegselmann-Krause model without weights \eqref{eq:HK}, the Mean-Field Limit process consists of representing the population by its density rather than by a collection of individual opinions.
The limit density $\nu(t,x)$ represents the (normalized) quantity of agents with opinion $x$ at time $t$ and satisfies a non-local transport equation, where the transport vector field $V$ is defined by the convolution of $\nu$ with the interaction function $\phi$. The proof of the limit lies on the fact that the empirical measure
\begin{equation*}
\nu^N(t,x) := \frac{1}{N} \sum_{i=1}^N \delta(x-x^{N}_i(t)).
\end{equation*}
satisfies the very same transport equation. It is crucial to notice that there is an irretrievable information loss in this formalism. Indeed the empirical measure keeps a count of the number of agents with opinion $x$ at time $t$, but loses track of the individual labeling of agents (i.e. the indices).
In the case of our augmented system \eqref{eq:syst-gen} with time-evolving weights, we generalize the notion of empirical measure by defining
\begin{equation}\label{eq:empmes}
\mu^N(t,x) := \frac{1}{N} \sum_{i=1}^N m^{N}_i(t) \delta(x-x^{N}_i(t)).
\end{equation}
We stress once again the information loss: the empirical measure only keeps track of the total weight of the group of agents with opinion $x$, but loses track of the individual labeling, the individual weights, and the number of agents at each point $x$.
More specifically, we draw attention to the fact that the empirical measure is invariant :
\begin{enumerate}[(i)]
\item by relabeling of the indices: for every $(x^N,m^N)\in (\mathbb{R}^d)^N\times\mathbb{R}^N$, for any $\sigma$ permutation of $\{1,\cdots,N\}$,
$$
\frac{1}{N} \sum_{i=1}^N m^{N}_i \delta(x-x^{N}_i)=\frac{1}{N} \sum_{i=1}^N m^{N}_{\sigma(i)} \delta(x-x^{N}_{\sigma(i)})
$$
\item by grouping of the agents: for every $(x^N,m^N)\in (\mathbb{R}^d)^N\times\mathbb{R}^N$, for every $I\subset\{1,\cdots,N\}$, such that $x^N_i = x_I$ for all $i\in I$,
$$
\frac{1}{N} \sum_{i=1}^N m^{N}_i \delta(x-x^{N}_i) = \frac{1}{N} \left[ \left(\sum_{i\in I} m^{N}_{i}\right) \delta(x-x_{I}) + \sum_{i\in\{1,\cdots,N\}\setminus I} m^{N}_i \delta(x-x^{N}_i) \right].
$$
\end{enumerate}
Figure \ref{fig:Indist} illustrates this invariance by comparing two microscopic systems corresponding to the same empirical measure.
This illustrates the fact that contrarily to the graph limit seen in the previous section, the mean-field limit process entails a non-reversible information loss. The empirical measure only retains the information concerning the total mass of agents at each point, and is incapable of differentiating between differently-labeled agents or between agents grouped at the same point.
\begin{figure}[h!]
\begin{center}
\begin{tikzpicture}[scale=1]
\input{Micro_xm_1}
\end{tikzpicture}%
\begin{tikzpicture}[scale=1]
\input{Micro_xm_2}
\end{tikzpicture}%
\begin{tikzpicture}[scale=1]
\input{Micro_xm_3}
\end{tikzpicture}%
\end{center}
\caption{Schematic representation of two microscopic sets of agents $(x^5,m^5)\in\mathbb{R}^5\times \mathbb{R}^5$ and $(y^5,p^5)\in\mathbb{R}^5\times \mathbb{R}^5$ corresponding to the same empirical measure $\mu^5\in\mathcal{P}(\mathbb{R})$. Left: Representation of $(x^5,m^5)$ with $x^5 = (0.5, 0.5, 1.5, 2.5, 3 )$ and $m^5 = (1.5,0.5,1.25,0.75,1)$. Center: Representation of $(y^5,p^5)$ with $y^5 = (0.5, 0.5, 3, 2.5,1.5 )$ and $p^5 = (1.25,0.75,1,0.75,1.25)$. Right: Symbolic representation of the empirical measure $\mu^5 = \frac{1}{5}( 2\,\delta_{0.5} + 1.25 \,\delta_{1.5} + 0.75 \,\delta_{2.5}+ \,\delta_{3})$.}\label{fig:Indist}
\end{figure}
Hence, in order to study the mean-field limit, we will require System \eqref{eq:syst-gen} to satisfy the following \textit{indistinguishability} property:
\begin{definition}\label{Def:indist}
We say that system \eqref{eq:syst-gen} satisfies \emph{indistinguishability} if
for all $J\subset \{1,\cdots,N\}$,
for all initial conditions $(x^0, m^0)\in \mathbb{R}^{dN}\times\mathbb{R}^N$ and $(y^0, p^0)\in \mathbb{R}^{dN}\times\mathbb{R}^N$ satisfying
\begin{equation}\label{eq:prop-indist_0}
\begin{cases}
x_i^0 = y_i^0 = x_j^0 = y_j^0 \qquad \text{ for all } (i,j)\in J^2 \\
x_i^0 = y_i^0 \qquad \text{ for all } i\in \{1,\cdots,N\} \\
m_i^0 = p_i^0 \qquad \text{ for all } i\in J^c \\
\sum_{i\in I} m_i^0 = \sum_{i\in I} p_i^0,
\end{cases}
\end{equation}
the solutions $t\mapsto (x(t),m(t))$ and $t\mapsto (y(t),p(t))$ to system \eqref{eq:syst-gen} with respective initial conditions
$(x^0, m^0)$ and $(y^0, p^0)$ satisfy for all $t\geq 0$,
\begin{equation*
\begin{cases}
x_i(t) = y_i(t) = x_j(t) = y_j(t) \qquad \text{ for all } (i,j)\in J^2 \\
x_i(t) = y_i(t) \qquad \text{ for all } i\in \{1,\cdots,N\} \\
m_i(t) = p_i(t) \qquad \text{ for all } i\in J^c \\
\sum_{i\in I} m_i(t) = \sum_{i\in I} p_i(t).
\end{cases}
\end{equation*}
\end{definition}
In the above definition and in all that follows, $J^c$ denotes the complement of the set $J$ in $\{1,\cdots,N\}$, i.e. $J^c=\{1,\cdots,N\}\setminus J$.
We begin by noticing that part of the required property is automatically satisfied by the general system \eqref{eq:syst-gen}, without having to specify further the mass dynamics. Namely, we prove that if two agents initially start at the same position, they stay with the same position at all time.
\begin{prop}\label{Prop:equal}
Let $(x^{0,N}_i)_{i\in\{1,\cdots,N\}}\in\mathbb{R}^{dN}$ and $(m^{0,N}_i)_{i\in\{1,\cdots,N\}}\in\mathbb{R}^{N}$. Let $\psi_i^{(N)}:(\mathbb{R}^d)\times\mathbb{R}\rightarrow\mathbb{R}$ and let $\phi\in\mathrm{Lip}(\mathbb{R}^d;\mathbb{R})$ such that the system of ODE \eqref{eq:syst-gen} with initial condition $\left( x^{0,N},m^{0,N}\right)$ admits a unique solution $(x^N,m^N)$.
If $x_k^{0,N}=x_l^{0,N}$ for some $(k,l)\in \{1,\cdots,N\}^2$, then it holds
$x_k^N(t) = x_l^N(t)$ for all $t\geq 0$.
\end{prop}
\begin{proof}
Let $(x^0_i)_{i\in\{1,\cdots,N\}}\in\mathbb{R}^{dN}$ and $(m^0_i)_{i\in\{1,\cdots,N\}}\in\mathbb{R}^{N}$
Without loss of generality, suppose that $x_1^0 = x_2^0$.
Now consider the slightly modified differential system
\begin{equation}\label{eq:syst-mod12}
\begin{cases}
\displaystyle \frac{d}{dt} x_i= \frac{1}{N} \sum_{j=3}^N m_j a(\|x_i-x_j\|) (x_j-x_i), \qquad x_i(0) = x_i^0, \qquad \text{ for } i=1,2 \\
\displaystyle \frac{d}{dt} x_i = \frac{1}{N} \sum_{j=1}^N m_j a(\|x_i-x_j\|) (x_j-x_i), \qquad x_i(0) = x_i^0, \qquad \text{ for all } i\in\{3,\cdots,N\} \\
\displaystyle \frac{d}{dt} m_i = \psi_i^N(x,m), \qquad m_i(0) = m_i^0, \qquad \text{ for all } i\in\{1,\cdots,N\}.
\end{cases}
\end{equation}
System \eqref{eq:syst-mod12} has a unique solution, that we denote by $(\tilde{x},\tilde{m})$.
Notice that $\tilde{x}_1$ and $\tilde{x}_2$ satisfy the same differential equation, so since $x_1^0=x_2^0$, we have $\tilde{x}_1(t) = \tilde{x}_2(t)$ for all $t\geq 0$.
Furthermore, one easily sees that $(\tilde{x}, \tilde{m})$ is also solution to \eqref{eq:syst-gen}. By uniqueness, we conclude that the unique solution $(x,m)$ to \eqref{eq:syst-gen} satisfies $x_1(t) = x_2(t)$ for all $t\geq 0$.
\end{proof}
Thus, if System \eqref{eq:syst-gen} is well-posed, part of the requirements for indistinguishability stated in Def. \ref{Def:indist} are automatically met. However, one can easily show that for general weight dynamics $\psi^{(N)}$, System \eqref{eq:syst-gen} does not satisfy indistinguishability.
For this reason, from here onward, we will focus on a particular class of weight dynamics given by
\begin{equation}
\label{eq:psi_i_gen}
\psi_i^{(N)}(x,m) = m_i(t)\frac{1}{N^k} \sum_{j_1=1}^N \cdots \sum_{j_k=1}^N m_{j_1}(t)\cdots m_{j_k}(t) S(x_i(t), x_{j_1}(t), \cdots x_{j_k}(t)),
\end{equation}
where $k\in\mathbb{N}$ and $S:(\mathbb{R}^d)^{k+1}\rightarrow \mathbb{R}$ satisfies the following assumptions:
\begin{hyp}\label{hyp:S}
$S\in C((\mathbb{R}^d)^{k+1}; \mathbb{R})$ is globally bounded and Lipschitz. More specifically, there exist $\bar{S}$, $L_S>0$ s.t.
\begin{equation*}\label{eq:psiSkbound}
\forall y\in(\mathbb{R}^d)^{k+1}, \quad |S(y)|\leq \bar{S}.
\end{equation*}
and
\begin{equation*}\label{eq:psiSklip}
\forall y\in (\mathbb{R}^d)^{k+1}, \forall z\in (\mathbb{R}^d)^{k+1}, | S(y_0,\cdots,y_k) - S(z_0,\cdots,z_k) | \leq L_S \sum_{i=0}^k |y_i-z_i|.
\end{equation*}
Furthermore, we require that $S$ satisfy the following skew-symmetry property: there exists $(i,j)\in \{0,\cdots,k\}^2$ such that for all $y\in(\mathbb{R}^d)^{k+1}$,
\begin{equation}\label{eq:condS}
S(y_0,\cdots, y_i,\cdots,y_j,\cdots,y_k) =-S(y_0,\cdots, y_j,\cdots,y_i,\cdots,y_k) .
\end{equation}
\end{hyp}
We show that with the weight dynamics given by \eqref{eq:psi_i_gen} and Hyp. \ref{hyp:S},
System \eqref{eq:syst-gen} satisfies the indistinguishability property of Def. \ref{Def:indist}.
\begin{prop}
Let $S\in C((\mathbb{R}^d)^{k+1}; \mathbb{R})$ satisfy Hyp. \ref{hyp:S}, and consider the collective dynamics system \eqref{eq:syst-gen}, with mass dynamics given by \eqref{eq:psi_i_gen}.
Then the system satisfies the indistinguishability property of Def. \ref{Def:indist}.
\end{prop}
\begin{proof}
For conciseness and clarity, we prove the statement for the case $k=1$, i.e. $S\in C(\mathbb{R}^2;\mathbb{R})$ and
\begin{equation}\label{eq:Sbis}
\psi_i^{(N)}(x,m) = \frac{1}{N} m_i \sum_{j=1}^N m_j S(x_i,x_j).
\end{equation}
The proof in the general case $k\in\mathbb{N}$ is essentially identical.
Let $J\subset \{1,\cdots,N\}$, $J^c = \{1,\cdots,N\}\setminus J$ and $N_c:=|I^c|\leq N$ denote the cardinal of $I^c$.
Let $(x^0, m^0)\in \mathbb{R}^{dN}\times\mathbb{R}^N$ and $(y^0, p^0)\in \mathbb{R}^{dN}\times\mathbb{R}^N$ satisfy \eqref{eq:prop-indist_0}.
Let $x_J^0:= x_i^0 = y_i^0$ for $i\in J$.
Let us start by considering a different system of dimension $d(N_c+1)+(N_c+1)$:
\begin{equation}\label{eq:syst-condensed}
\begin{cases}
\displaystyle \frac{d}{dt} \tilde{x}_i = \frac{1}{N} \tilde{m}_J a(\|\tilde{x}_i-\tilde{x}_J\|) (\tilde{x}_J-\tilde{x}_i) + \frac{1}{N} \sum_{j\in J^c} \tilde{m}_j a(\|\tilde{x}_i-\tilde{x}_j\|) (\tilde{x}_j-\tilde{x}_i) , \qquad \text{ for all } i\in J^c \\
\displaystyle \frac{d}{dt} \tilde{x}_J = \frac{1}{N} \sum_{j\in J^c} \tilde{m}_j a(\|\tilde{x}_J-\tilde{x}_j\|) (\tilde{x}_j-\tilde{x}_J), \\
\displaystyle \frac{d}{dt} \tilde{m}_i = \frac{1}{N} \tilde{m}_J \tilde{m}_i S(\tilde{x}_i,\tilde{x}_J) + \frac{1}{N} \tilde{m}_i \sum_{j\in J^c} \tilde{m}_j S(\tilde{x}_i,\tilde{x}_j), \qquad \text{ for all } i\in J^c \\
\displaystyle \frac{d}{dt} \tilde{m}_J = \frac{1}{N} \tilde{m}_J^2 S(\tilde{x}_J,\tilde{x}_J) + \frac{1}{N} \tilde{m}_J \sum_{j\in J^c} \tilde{m}_j S(\tilde{x}_J,\tilde{x}_j).
\end{cases}
\end{equation}
Given a set of initial conditions, the Cauchy problem associated with \eqref{eq:syst-condensed} has a unique solution.
Let us now consider the solutions $t\mapsto (x(t),m(t)$ and $t\mapsto(y(t), p(t))$ to \eqref{eq:syst-gen} with mass dynamics given by \eqref{eq:Sbis} with respective initial conditions $(x^0, m^0)$ and $(y^0, p^0)$.
First of all, from Prop. \ref{Prop:equal}, $x_i(t)=x_j(t)$ and $y_i(t) = y_j(t)$ for all $(i,j)\in J^2$ and all $t\geq 0$.
Let $t\mapsto x_J(t):=x_i(t)$ and $t\mapsto y_J(t):=y_i(t)$ for $i\in J$.
We can compute:
\begin{equation*}
\begin{cases}
\displaystyle \frac{d}{dt} x_i = \frac{1}{N} (\sum_{j\in J} m_j) a(\|x_i-x_J\|) (x_J-x_i) + \frac{1}{N} \sum_{j\in J^c} m_j a(\|x_i-x_j\|) (x_j-x_i) \qquad \text{ for all } i\in J^c,\\
\displaystyle \frac{d}{dt} x_J = \frac{1}{N} \sum_{j\in J^c} m_j a(\|x_J-x_j\|) (x_j-x_J), \\
\displaystyle \frac{d}{dt} m_i = \frac{1}{N} (\sum_{j\in J} m_j) m_i S(x_i,x_J) + \frac{1}{N} m_i \sum_{j\in J^c} m_j S(x_i,x_j) \qquad \text{ for all } i\in J^c , \\
\displaystyle \frac{d}{dt}(\sum_{j\in J} m_j) = \frac{1}{N} (\sum_{j\in J} m_j)^2 S(x_J,x_J) + \frac{1}{N} (\sum_{j\in J} m_j) \sum_{j\in J^c} m_j S(x_J,x_j).
\end{cases}
\end{equation*}
This shows that $( (x_i)_{i\in J}, x_J, (m_i)_{i\in J}, (\sum_{j\in J} m_j))$ satisfies the differential system \eqref{eq:syst-condensed}.
Similarly, we can show that $( (y_i)_{i\in J}, y_J, (p_i)_{i\in J}, (\sum_{j\in J} p_j))$ satisfies \eqref{eq:syst-condensed}.
Furthermore, from \eqref{eq:prop-indist_0},
$$( (x_i^0)_{i\in J}, x_J^0, (m_i^0)_{i\in J}, (\sum_{j\in J} m_j^0))=( (y_i^0)_{i\in J}, y_J^0, (p_i^0)_{i\in J}, (\sum_{j\in J} p_j^0)).$$
By uniqueness, for all $t\geq 0$ it holds:
$$( (x_i)_{i\in J}, x_J, (m_i)_{i\in J}, (\sum_{j\in J} m_j)) = ( (y_i)_{i\in J}, y_J, (p_i)_{i\in J}, (\sum_{j\in J} p_j)).$$
\end{proof}
\subsection{General context and results}\label{sec:MFpres}
Let $\P(\mathbb{R}^d)$ denote the set of probability measures on $\mathbb{R}^d$.
As exposed in Section \ref{sec:indisting}, the Mean-Field Limit process only makes sense for a subclass of mass dynamics that satisfy an indistinguishability property (Def. \ref{Def:indist}).
Hence, from here onward, we will focus on the indistinguishability-preserving mass dynamics given by \eqref{eq:psi_i_gen}.
In this framework, the convergence of the empirical measure $\mu^N$ to a limit measure $\mu$ was proven in \cite{PouradierDuteil21}, and can be stated more precisely as follows:
\begin{theorem}\label{Th:mfl}
Let $T>0$.
For each $N\in\mathbb{N}$, let $(x^{0,N}_i)_{i\in\{1,\cdots,N\}}\in\mathbb{R}^{dN}$ and $(m^{0,N}_i)_{i\in\{1,\cdots,N\}}\in\mathbb{R}^{N}$ satisfying \eqref{eq:sum_M}.
Let $S\in C((\mathbb{R}^d)^{k+1}; \mathbb{R})$ satisfying Hyp. \ref{hyp:S}.
For all $t\in [0,T]$, let $t\mapsto(x^N(t), m^N(t))$ be the corresponding solution to \eqref{eq:syst-gen}-\eqref{eq:psi_i_gen} with initial data $(x^{0,N},m^{0,N})$, and let $\mu^N_t:= \frac{1}{N}\sum_{i=1}^N m_i^N(t) \delta_{x_i^N(t)}\in \P(\mathbb{R}^d)$ be the corresponding empirical measures.
If there exists $\mu_0 \in \P(\mathbb{R}^d)$ such that
$$
\lim_{N\rightarrow\infty} \rho(\mu^N_0,\mu_0) = 0,
$$
then for all $t\in [0,T]$,
$$
\lim_{N\rightarrow\infty} \rho(\mu^N_t,\mu_t) = 0,
$$
where $\mu_t$ is the solution to the transport equation with source
\begin{equation}\label{eq:mfl}
\partial_t \mu_t(x) + \nabla\cdot ( V[\mu_t](x) \mu_t(x)) = h[\mu_t](x)
\end{equation}
with the non-local vector-field given by
\[
\forall \mu\in\P(\mathbb{R}^d),\quad \forall x\in\mathbb{R}^d, \quad V[\mu](x) = \int_{\mathbb{R}^d} \phi(y-x) d\mu(y),
\]
the non-local source term given by
\[
\forall \mu\in \P(\mathbb{R}^d), \quad \forall x\in\mathbb{R}^d, \quad h[\mu](x) = \left(\int_{(\mathbb{R}^d)^k} S(x,y_1,\cdots,y_k) d\mu(y_1)\cdots d\mu(y_k)\right) \mu(x).
\]
and with initial condition $\mu_{t=0} = \mu_0$.
\end{theorem}
\noindent In Theorem \ref{Th:mfl}, $\rho$ denotes the Bounded Lipschitz distance, also known as the generalized Wasserstein distance $W^{1,1}_1$ \cite{PiccoliRossi14,PouradierDuteil21}.
The aim of this section is to explore the link between the Graph Limit equation \eqref{eq:GraphLimit-gen} and the Mean-Field Limit equation \eqref{eq:mfl}.
In the spirit of \cite{BiccariKoZuazua19}, we will actually show that the Mean-Field Limit is subordinated to the Graph Limit.
Moreover, this link can be used to revisit the proof of the mean-field limit of the discrete system \eqref{eq:syst-gen}. It is important to note however that the result is weaker than Theorem \ref{Th:mfl} and its interest lies mainly in its ability to link the two approaches. We will prove the following result :
\begin{theorem}
\label{Th:mfl2}
Let $T>0$.
Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R}^{*+})$ satisfying \eqref{eq:integral_egal_a_1}.
Let $\phi$ satisfy Hyp.~\ref{hyp:phi}.
Let $(x_N,m_N)$ denote the solution to \eqref{eq:syst-contN} with mass dynamics given by \eqref{eq:psi_i_gen}, with $S$ satisfying Hyp.~\ref{hyp:S} and with initial conditions
\[\displaystyle x_N(0,s) = \sum_{i=1}^N x_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s) \quad \text{ and }\quad \displaystyle m_N(0,s) = \sum_{i=1}^N m_i^{0,N} \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s),
\]
where $x^{0,N}=P_\mathrm{d}^N(x_0)$ and $m^{0,N}=P_\mathrm{d}^N(m_0)$ are defined by \eqref{eq:ICx}.\\
Then there exist $x\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d))$ and $m\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}))$ such that
\[
\displaystyle \|x-x_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}^d))} \xrightarrow[N\rightarrow+\infty]{} 0
\quad \text{ and } \quad
\|m-m_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}))} \xrightarrow[N\rightarrow+\infty]{} 0.
\]
Moreover, $(x,m)$ are solutions to the integro-differential system \eqref{eq:GraphLimit-gen} with $\psi:=\psi_{S,k}$ defined by
\begin{equation*}
\psi_{S,k}(s,x,m) = m(s) \int_{I^k} m(s_1) \cdots m(s_k) \, S(x(s), x(s_1), \cdots, x(s_k) ) \; ds_1\cdots ds_k.
\end{equation*}
In addition, let $\tilde{\mu}\in C([0,T];\P(\mathbb{R}^d))$ be defined by
\begin{equation*
\tilde{\mu}_t(x) := \int_I m(t,s_*) \delta(x-x(t,s_*)) ds_*.
\end{equation*}
Then, the empirical measure $\mu^N$ defined in \eqref{eq:empmes} converges weakly to $\tilde{\mu}$, and $\tilde{\mu}$ is a solution to the transport equation with source term \eqref{eq:mfl}.
\end{theorem}
Theorem \eqref{Th:mfl2} contains many different results.
We show the well-posedness of the system of integro-differential equations given by \eqref{eq:GraphLimit-gen} with $\psi:=\psi_{S,k}$ in Section \ref{sec:WellPos-Indist}. The well-posedness of the microscopic system \eqref{eq:syst-gen}-\eqref{eq:psi_i_gen} will be stated in Section \ref{sec:WellPos-Indist-micro}.
Section \ref{sec:mflsubord} is devoted to proving that $\tilde{\mu}$ is a weak solution to \eqref{eq:mfl}.
\subsection{Well-posedness of the Graph Limit equation}\label{sec:WellPos-Indist}
From here onward, as mentioned in Section \ref{sec:indisting}, we will consider a particular class of mass dynamics which
preserves mass as well as \emph{indistinguishability}.
Let $\psi_{S,k}:I\times L^\infty(I;\mathbb{R}^d)\times L^\infty(I;\mathbb{R})\rightarrow \mathbb{R}$ be defined by
\begin{equation}\label{eq:psiSk}
\psi_{S,k}(s,x,m) = m(s) \int_{I^k} m(s_1) \cdots m(s_k) \, S(x(s), x(s_1), \cdots, x(s_k) ) \; ds_1\cdots ds_k
\end{equation}
where
$k\in\mathbb{N}$ and as previously, $S\in C((\mathbb{R}^d)^{k+1}; \mathbb{R})$ satisfies Hyp. \ref{hyp:S}.
\begin{rem}
The condition \eqref{eq:condS} of Hyp. \ref{hyp:S} implies that
\begin{equation}\label{eq:psiSkint}
\int_{I} \psi_{S,k}(s,x,m) ds = 0.
\end{equation}
Actually, this condition \eqref{eq:psiSkint} is sufficient to prove all the subsequent results, but for simplicity purposes, we will keep referring to the more particular (and more tangible) condition \eqref{eq:condS}.
\end{rem}
We notice that at first glance, the mass dynamics $\psi_{S,k}$ given by \eqref{eq:psiSk} do not satisfy
Hyp. \ref{hyp:psi}, which was necessary in order to prove the existence and uniqueness of the solution to equation \eqref{eq:GraphLimit-gen}.
The aim of this section is to show that nevertheless, the system of coupled integro-differential equations \eqref{eq:GraphLimit-gen} with $\psi= \psi_{S,k}$ is well-posed, as long as Hyp. \ref{hyp:S} is satisfied.
The proof of existence and uniqueness will rely on the following a priori observations:
\begin{prop}\label{Prop:psiSkprop}
Let $(x_0,m_0)\in L^\infty(I,\mathbb{R}^d\times\mathbb{R}^{*+})$. Consider a solution $(x,m)$ on an interval $[0,T]$ to the integro-differential system \eqref{eq:GraphLimit-gen} with initial condition $(x_0,m_0)$ and weight dynamics $\psi_{S,k}$ given by \eqref{eq:psiSk} and Hyp.~\ref{hyp:S}.
Then, the following properties hold:
\begin{enumerate}[i)]
\item For almost every $s\in I$ and all $t\in [0,T]$, $m(t,s)>0$
\item For all $t\in [0,T]$, $\int_I m(t,s) ds = M_0 := \int_I m_0(s) ds$
\item For almost every $s\in I$ and all $t\in [0,T]$, $m(t,s)\leq m_0(s) \exp(M_0^k \bar{S} t)$
\end{enumerate}
\end{prop}
\begin{rem}
With the assumption \eqref{eq:integral_egal_a_1}, Properties $(ii)$ and $(iii)$ simplify to
\[\int_I m(\cdot,s) ds \equiv 1 \qquad \text{ and } \qquad m(t,s)\leq m_0(s) \exp(\bar{S} t).
\]
\end{rem}
\begin{proof}
The second property is an immediate consequence of the antisymmetry property \eqref{eq:condS}.\\
We now focus on the first point. Suppose that there exists $t_- \in [0,T]$ and a non-negligeable set $\overline{\mathcal{N}}$ in I, such that for $s \in \overline{\mathcal{N}}$, we have $m(s,t_-)\leq 0$. Let
$$t^* := \inf\{t\in [0,T], \, \text{there exists a non-negligeable set } \overline{\mathcal{N}}_k \text{ such that for } s \in \overline{\mathcal{N}}_k, ~ m(t,s)\leq 0\}.$$ Since $m_0(s)>0$ for almost every $s\in I$, there exists $s^* \in \overline{\mathcal{N}}_k$ such that $m_0(s^*) >0$ and by continuity, $m(t^*,s^*)=0$. Moreover, by definition of $t^*$, for $t<t^*$ and for almost every $s\in I$, we have $m(t,s)>0$.
Using the global bound on $S$, we then compute, for all $t\leq t^*$ :
\begin{equation*}
\begin{split}
\partial_t m(t,s^*) = & m(t,s^*) \int_{I^k} m(t,s_1) \cdots m(t,s_k) \, S(x(t,s), x(t,s_1), \cdots, x(t,s_k) ) \; ds_1\cdots ds_k \\
\geq& - \bar{S} \, m(t,s^*) \int_{I^k} m(t,s_1) \cdots m(t,s_k) \, ds_1\cdots ds_k = - \bar{S} \, M_0^k \, m(t,s^*)
\end{split}
\end{equation*}
where we used the second property.
Thus, from Gronwall's lemma, we obtain that for all $t\leq t^*$, $m(t,s^*)\geq m_0(s^*) e^{-\bar{S} \, M_0^k\, t} > 0 $, which contradicts $m(t^*,s^*)=0$. We deduce that $m(t,s)>0$ for almost every $s\in I$ and all $t\in [0,T]$. \\
The third point can be proven easily by using the positivity of the weights and the boundedness of $S$ by $\bar{S}$.
\end{proof}
We now prove that the results of Lemma \ref{Lemma:WellPos-decoupled} also hold with $\psi=\psi_{S,k}$ satisfying \eqref{eq:psiSk} and Hyp. \ref{hyp:S}.
\begin{lemma}\label{Lemma:WellPos-decoupled-psiSk}
Let $\tilde{x} \in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d))$ and $\tilde{m} \in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}))$. Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R}^{*+})$.
Let $\phi$ satisfy Hyp. \ref{hyp:phi} and let $\psi_{S,k}$ satisfy \eqref{eq:psiSk} and Hyp. \ref{hyp:S}. \\
Then for any $T>0$, there exists a unique solution $(x,m)\in\mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}^{*+}))$ to the decoupled integro-differential equations
\begin{equation*}
\left\{\begin{array}{l}
\displaystyle \partial_t x(t,s) = \int_I \tilde{m}(t,s_*)\phi(x(t,s) - x(t,s_*)) ds_*; \qquad x(\cdot,0)=x_0 \\
\partial_t m(t,s) = \psi_{S,k}(s,\tilde{x}(t,\cdot),m(t,\cdot)); \qquad m(\cdot,0)=m_0.
\end{array}\right.
\label{eq:GraphLimit-decoupled-psiSk}
\end{equation*}
\end{lemma}
\begin{proof}
Since the two equations are decoupled, the proof of existence and uniqueness of the solution $x\in\mathcal{C}([0,T];L^\infty(I;\mathbb{R}^d))$ to the first equation was already done in Lemma \ref{Lemma:WellPos-decoupled}. We focus on the well-posedness of the second equation.
Let $M_0:=\int_I m_0(s) ds$.
Let $M_{m_0}'$ be a metric subspace of $\mathcal{C}([0,T];L^1(I;\mathbb{R}^{*+}))$ consisting of functions $m$ satisfying $m(\cdot,0) = m_0$ and $\int m(t,s) ds = M_0$ for all $t\in [0,T]$.
Let $K_{m_0}'$ be the operator defined by:
$$
\begin{array}{l l l }
K_{m_0}': & M_{m_0}' & \rightarrow M_{m_0}' \\
& m &\displaystyle \mapsto (K_{m_0}' m):(t,s)\mapsto m_0(s) + \int_0^t \psi(s,\tilde{x}(\tau,\cdot),m(\tau,\cdot)) d\tau.
\end{array}
$$
From Proposition \ref{Prop:psiSkprop}, it is clear that $K_{m_0}'$ maps $M_{m_0}'$ onto $M_{m_0}'$.
Let $\tilde{T}>0$. We will show that $K_{m_0}'$ is contracting for the norm $\|\cdot\|_{M_{m_0}'} := \sup_{[0,\tilde{T}]} \|\cdot\|_{L^1(I)}$.
Let $(m_1,m_2)\in M_{m_0}'^2$.
Then for all $t\leq \tilde{T}$,
\begin{equation*}
\begin{split}
\int_I & | K_{m_0}' m_1 - K_{m_0}' m_2 | (t,s) ds = \int_I \left | \int_0^t \psi_{S,k}(s,\tilde{x}(\tau,\cdot),m_1(\tau,\cdot))- \psi_{S,k}(s,\tilde{x}(\tau,\cdot),m_2(\tau,\cdot)) d\tau \right | ds \\
& \leq \int_I \int_0^t \int_{I^k} \Big | [m_1(s) m_1(s_1) \cdots m_1(s_k) - m_2(s) m_2(s_1) \cdots m_2(s_k)] \, S(\tilde{x}(s), \tilde{x}(s_1), \cdots, \tilde{x}(s_k) ) \Big | \; ds_1\cdots ds_k \, dt \, ds\\
& \leq \bar{S} \int_0^t \int_{I^{k+1}} \left | m_1(s) m_1(s_1) \cdots m_1(s_k) - m_2(s) m_2(s_1) \cdots m_2(s_k) \right | \;ds \, ds_1\cdots ds_k \, dt \\
& \leq \bar{S} \int_0^t \Big [ \int_{I^{k+1}} | m_1(s) - m_2(s) | m_1(s_1) \cdots m_1(s_k) \;ds \, ds_1\cdots ds_k \\
& \qquad + \int_{I^{k+1}} m_2(s) | m_1(s_1) - m_2(s_1) | m_1(s_2) \cdots m_1(s_k) \;ds \, ds_1\cdots ds_k \\
& \qquad + \cdots + \int_{I^{k+1}} m_2(s) m_2(s_1) \cdots m_2(s_{k-1}) | m_1(s_k) - m_2(s_k) | \;ds \, ds_1\cdots ds_k \, \Big ] dt \\
& \leq \bar{S} (k+1) \int_0^t \int_I | m_1(s) - m_2(s) | ds M_0^k \, dt,
\end{split}
\end{equation*}
where, from the second line onward we omitted the time dependence for compactness of notation.
We obtain:
\begin{equation*}
\| K_{m_0}' m_1 - K_{m_0}' m_2 \|_{M_{m_0}'} \leq \tilde{T} \bar{S} (k+1) M_0^k \| m_1 - m_2 \|_{M_{m_0}'}.
\end{equation*}
Thus, if $\tilde{T}\leq (2\bar{S} (k+1) M_0^k)^{-1}$, by the Banach contraction mapping principle, there exists a unique solution $m\in \mathcal{C}([0,\tilde{T}];L^1(I;\mathbb{R}^{*+}))$.
We then take $m(\tilde{T},\cdot)$ as the initial condition, and the local solution can be extended to $[0, 2\tilde{T}]$, and by repeating the same argument, to $[0,T]$.
We thus showed that there exists a unique solution $m\in \mathcal{C}([0,T];L^1(I;\mathbb{R}^{*+}))$.
Furthermore, the third property of Prop. \ref{Prop:psiSkprop} implies that $m\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}^{*+}))$.
Lastly, since the integrand is continuous with respect to $\tau$, $m(\cdot,s)$ is continuously differentiable for almost all $s\in I$, which proves that ${m\in \mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^{*+}))}$.
\end{proof}
The proof of Theorem \ref{Th:GL-wellpos} relied on the fact that $\psi$ satisfies \eqref{eq:psilip2}. Although $\psi_{S,k}$ does not satisfy \eqref{eq:psilip2}, we notice that a similar property does hold as long as $m$ belongs to $L^\infty$.
\begin{lemma}\label{Lemma:psiSklip2}
Let $\psi_{S,k}$ satisfy \eqref{eq:psiSk} and Hyp. \ref{hyp:S}. Let $(m_1, m_2) \in L^\infty(I)^2$ with $max(\|m_2\|_{L^\infty(I)},\|m_1\|_{L^\infty(I)}) \leq M_1$.
Then for all $(x_1,x_2)\in L^\infty(I)^2$, it holds:
\begin{equation}\label{eq:psiSklip2}
\begin{cases}
\|\psi_{S,k}(\cdot,x_1,m_1)-\psi_{S,k}(\cdot,x_2,m_1)\|_{L^2(I)} \, \leq \, (k+1) L_S M_1^{k+1} \|x_1-x_2\|_{L^2(I)}\\
\|\psi_{S,k}(\cdot,x_1,m_1)-\psi_{S,k}(\cdot,x_1,m_2)\|_{L^2(I)} \, \leq \, \sqrt{k+1} \bar{S} M_1^k \|m_1-m_2\|_{L^2(I)}.
\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
We start by computing
\begin{equation*}
\begin{split}
\|\psi_{S,k} & (\cdot,x_1,m_1)-\psi_{S,k}(\cdot,x_2,m_1)\|_{L^2(I)}^2 \\
& = \int_I \bigg | \int_{I^k} m_1(s_0)\cdots m_1(s_k) (S(x_1(s_0),\cdots,x_1(s_k))-S(x_2(s_0),\cdots,x_2(s_k))) ds_1 \cdots ds_k \bigg |^2 ds_0 \\
& \leq \int_{I^{k+1}} m_1(s_0)^2\cdots m_1(s_k)^2 \Big | S(x_1(s_0),\cdots,x_1(s_k))-S(x_2(s_0),\cdots,x_2(s_k))\Big |^2 ds_0 \cdots ds_k \\
& \leq \int_{I^{k+1}} m_1(s_0)^2\cdots m_1(s_k)^2 L_S^2 \Big (\sum_{i=0}^k \|x_1(s_i)-x_2(s_i)\|\Big )^2 ds_0 \cdots ds_k \\
& \leq M_1^{2(k+1)} L_S^2 \int_{I^{k+1}} (k+1) \sum_{i=0}^k \|x_1(s_i)-x_2(s_i)\|^2 ds_0 \cdots ds_k \\
& \leq (k+1)^2 L_S^2 M_1^{2(k+1)} \|x_1-x_2\|_{L^2(I)}^2
\end{split}
\end{equation*}
from which we get the first inequality of \eqref{eq:psiSklip2}.
For the second inequality, we compute:
\begin{equation*}
\begin{split}
\|\psi_{S,k} & (\cdot,x_1,m_1)-\psi_{S,k}(\cdot,x_1,m_2)\|_{L^2(I)}^2 \\
& = \int_I \bigg | \int_{I^k} \big ( m_1(s_0)\cdots m_1(s_k) - m_2(s_0)\cdots m_2(s_k) \big ) S(x_1(s_0),\cdots,x_1(s_k)) ds_1 \cdots ds_k \bigg |^2 ds_0 \\
& \leq \bar{S}^2 \int_{I^{k+1}} \big | m_1(s_0)\cdots m_1(s_k) - m_2(s_0)\cdots m_2(s_k) \big |^2 ds_0 \cdots ds_k \\
& \leq \bar{S}^2 \int_{I^{k+1}} \sum_{i=0}^k \Big( \prod_{j=0}^{i-1} m_2(s_j)^2\Big) | m_1(s_i)-m_2(s_i)|^2 \Big( \prod_{j=i+1}^{k} m_1(s_j)^2\Big) ds_0 \cdots ds_k \\
& \leq \bar{S}^2 M_1^{2k} \int_{I^{k+1}} \sum_{i=0}^k | m_1(s_i)-m_2(s_i)|^2 ds_0 \cdots ds_k = (k+1) \bar{S}^2 M_1^{2k} \|m_1-m_2\|_{L^2(I)}^2.
\end{split}
\end{equation*}
\end{proof}
We are now fully equipped to prove the well-posedness of the coupled system, with $\psi=\psi_{S,k}$.
\begin{theorem}\label{Th:WellPos-psiSk}
Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R}^{*+})$.
Let $\phi$ satisfy Hyp. \ref{hyp:phi} and let $\psi_{S,k}$ satisfy \eqref{eq:psiSk} and Hyp. \ref{hyp:S}. \\
Then for any $T>0$, there exists a unique solution $(x,m)\in\mathcal{C}^1([0,T];L^\infty(I;\mathbb{R}^d \times \mathbb{R}^{*+}))$ to the integro-differential system
\begin{equation*}
\left\{\begin{array}{l}
\displaystyle \partial_t x(t,s) = \int_I m(t,s_*)\phi(x(t,s) - x(t,s_*)) ds_*; \qquad x(\cdot,0)=x_0 \\
\partial_t m(t,s) = \psi_{S,k}(s,x(t,\cdot),m(t,\cdot)); \qquad m(\cdot,0)=m_0.
\end{array}\right.
\end{equation*}
\end{theorem}
\begin{proof}
The proof is almost identical to the one of Theorem \ref{Th:GL-wellpos}.
The only difference lies in the inequality \eqref{eq:ineqmn}.
We notice that from the third point of Proposition \ref{Prop:psiSkprop}, we can apply Lemma \ref{Lemma:psiSklip2} with $m_1:=m^n$ and $m_2:= m^{n+1}$,
and $\|m^n\|_{L^\infty(I)} \leq M_T:=\| m_0\|_{L^\infty(I)} \exp(M_0^k \bar{S} T)$.
We can thus replace \eqref{eq:ineqmn} by:
\begin{equation*}
\begin{split}
\|m^{n+1}-& m^n\|_{L^2(I)}^2(t) \\
\leq & 2 t \int_0^t \int_I |\psi_{S,k}(s,x^n(\tau,\cdot),m^{n+1}(\tau,\cdot)) - \psi_{S,k}(s,x^n(\tau,\cdot),m^{n}(\tau,\cdot)) |^2 ds \, d\tau \\
& + 2t \int_0^t \int_I | \psi_{S,k}(s,x^n(\tau,\cdot),m^{n}(\tau,\cdot)) - \psi_{S,k}(s,x^{n-1}(\tau,\cdot),m^{n}(\tau,\cdot)) | ^2 ds \, d\tau \\
\leq & 2 t (k+1) \bar{S}^2 M_1^{2k} \int_0^t \|m^{n+1}-m^{n}\|_{L^2(I)}^2(\tau) d\tau + 2 t (k+1)^2 L_S^2 M_1^{2(k+1)} \int_0^t \| x^n - x^{n-1}\|_{L^2(I)}^2(\tau) d\tau .
\end{split}
\end{equation*}
The rest of the proof of existence is identical, replacing the previous definition of $A_T$ by
$$A_T = \max(8 T L_\phi^2 \, X_T^2,\, 8 T L_\phi^2 \, M_T^2,\, 2 T (k+1) \bar{S}^2 M_T^{2k}, 2T (k+1)^2 L_S^2 M_T^{2(k+1)} ).$$
The proof of uniqueness can be adapted similarly.
\end{proof}
\subsection{Well-posedness of the microscopic system}\label{sec:WellPos-Indist-micro}
The existence and uniqueness for the microscopic system \eqref{eq:syst-gen} can be shown in the case of functions $\psi_{S,k}$ satisfying Hyp. \ref{hyp:S} instead of Hyp. \ref{hyp:psi}, in the same way that we adapted the proof of well-posedness for the integro-differential equations in Section \ref{sec:WellPos-Indist}. For this reason, we do not provide the proof, which would be redundant, and merely state the result.
\begin{theorem}
Let $(x^{0,N},m^{0,N}) \in \mathbb{R}^{dN} \times \mathbb{R}^N$.
Let $\phi$ satisfy Hyp. \ref{hyp:phi}, let $\psi_{S,k}$ satisfy \eqref{eq:psiSk} and Hyp. \ref{hyp:S}, and $\psi_i^{(N)}$ be defined by
\begin{equation*}
\forall i\in\{1,\cdots,N\},\qquad \psi_i^{(N)}(x^{N}(t),m^{N}(t)) = N \int_{\frac{i-1}{N}}^{\frac{i}{N}} \psi_{S,k}(s,x_N(t,\cdot),m_N(t,\cdot)) ds.
\end{equation*}\\
Then for any $T>0$, there exists a unique solution $(x^{N},m^{N}) \in \mathcal{C}^1([0,T];\mathbb{R}^{dN} \times \mathbb{R}^N)$ to the discrete system \eqref{eq:syst-gen} with initial condition $(x^{0,N},m^{0,N})$.
Furthermore, the solution to the system has the following properties :
\begin{enumerate}
\item[(i)] $\forall i \in \{1, \dots, N \}$, $\forall t \in [0,T]$, $m_i^{(N)}(t) >0$,
\item[(ii)]$\forall t \in [0,T]$, $\displaystyle{\sum_{i=1}^N m_i^{(N)}(t) = N}$,
\item[(iii)] there exist constants $\overline{X}$ and $\overline{M}$ such that for all $t \in [0,T]$, for all $i \in \{1, \dots, N \}$,
\begin{equation*}
\| x_i^{(N)}(t)\| \leq \overline{X}
\quad \text{ and } \quad
| m_i^{(N)}(t)| \leq \overline{M}.
\end{equation*}
\end{enumerate}
\end{theorem}
\begin{rem}
The theorem can also be stated for functions $\psi_i^{(N)}$ given by \eqref{eq:psi_i_gen} satisfying Hyp. \ref{hyp:S}.
\end{rem}
\subsection{Subordination of the mean-field equation to the graph limit equation}\label{sec:mflsubord}
The first part of Theorem \ref{Th:mfl2} consists of deriving the Graph Limit, as in Theorem \ref{Th:GraphLimit}, but for mass dynamics \eqref{eq:psiSk} satisfying Hyp. \ref{hyp:S} instead of Hyp. \ref{hyp:psi}.
Nevertheless, as for the proof of the existence of a solution to the graph limit equation,
the convergence proof can be adapted in a straightforward way, by using the assumptions on $\psi_{S,k}$ given by Hyp. \ref{hyp:S}.
Thus, we can show that the solution to the discrete system \eqref{eq:syst-gen}-\eqref{eq:psi_i_gen} converges to the functions $(x,m)\in \mathcal{C}([0,T]; L^2(I;\mathbb{R}^d))\times \mathcal{C}([0,T]; L^2(I;\mathbb{R}))$, solutions to the integro-differential system \eqref{eq:GraphLimit-gen} with $\psi:I \times L^2(I;\mathbb{R}^d)\times L^2(I;\mathbb{R})$ defined in \eqref{eq:psiSk}, as stated in the following:
\begin{prop}
\label{prop:conv}
Let $x_0\in L^\infty(I;\mathbb{R}^d)$ and $m_0\in L^\infty(I;\mathbb{R}^{*+})$ satisfying \eqref{eq:integral_egal_a_1}.
Let $\phi$ satisfying Hyp. \ref{hyp:phi} and $\psi:=\psi_{S,k}$ satisfying \eqref{eq:psiSk} and Hyp. \ref{hyp:S}.
Then the solution $(x_N,m_N)$ to \eqref{eq:syst-contN}
with initial conditions
$x_N(0,s)=P_\mathrm{c}^N(P_\mathrm{d}^N(x_0))$ and $m_N(0,s)=P_\mathrm{c}^N(P_\mathrm{d}^N(m_0))$ defined by \eqref{eq:ICx}-\eqref{eq:xN}
converges when $N$ tends to infinity in the $\mathcal{C}([0,T];L^2(I))$ topology, i.e. there exists $(x,m)\in \mathcal{C}([0,T];L^2(I,\mathbb{R}^d))\times \mathcal{C}([0,T];L^2(I,\mathbb{R}))$ such that
\[
\displaystyle \|x-x_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}^d))} \xrightarrow[N\rightarrow+\infty]{} 0 \quad \text{ and } \quad
\|m-m_N\|_{\mathcal{C}([0,T];L^2(I,\mathbb{R}))} \xrightarrow[N\rightarrow+\infty]{} 0.
\]
Furthermore, the limit functions $x$ and $m$ are solutions to the integro-differential system \eqref{eq:GraphLimit-gen} with $\psi=\psi_{S,k}$,
supplemented by the initial conditions $x(0,\cdot) = x_0$ and $m(0,\cdot) = m_0$.
\end{prop}
Let us now study the link between the non-local diffusive model coming from the graph limit equation~\eqref{eq:GraphLimit-gen} and the non-local transport equation with source \eqref{eq:mfl} obtained by the mean-field approach. This will give us an alternative proof of the mean-field limit.
\begin{prop}
\label{Prop:link_mfl-gl}
Let $(x,m)\in \mathcal{C}([0,T]; L^2(I;\mathbb{R}^d))\times \mathcal{C}([0,T]; L^2(I;\mathbb{R}))$such that
\begin{equation}
\label{eq:GL-TE}
\begin{cases}
\partial_t x(t,s) = \displaystyle \int_I m(t,s_*) \phi(x(t,s_*)-x(t,s)) ds_* \\
\partial_t m(t,s) = \displaystyle m(t,s) \int_{I^k} m(t,s_1) \cdots m(t,s_k) \, S(x(t,s), x(t,s_1), \cdots, x(t,s_k) ) \; ds_1\cdots ds_k.
\end{cases}
\end{equation}
Let $\tilde{\mu}\in\P(\mathbb{R}^d)$ be defined by
\begin{equation}\label{eq:def_mu}
\tilde{\mu}_t(x) := \int_I m(t,s_*) \delta(x-x(t,s_*)) ds_*.
\end{equation}
Then $\tilde{\mu}$ satisfies the transport equation with source \eqref{eq:mfl}.
\end{prop}
\begin{proof}
Given a test function $\varphi=\varphi(x)$, we consider the term
$$(\tilde{\mu}_t,\varphi) := \int_{\mathbb{R}^d} \varphi(x) d \tilde{\mu}_t(x).$$
Let us study the quantity $\displaystyle{\frac{d}{dt} (\tilde{\mu}_t,\varphi)}.$ We start by noticing that
$$
(\tilde{\mu}_t,\varphi) = \displaystyle \int_{\mathbb{R}^d} \varphi(x) d \tilde{\mu}_t(x)=\displaystyle \int_I m(t,s) \varphi(x(t,s)) ds
$$
by definition of $\tilde{\mu}_t$. Therefore, it holds
$$
\displaystyle \frac{d}{dt} (\tilde{\mu}_t,\varphi) = \displaystyle \int_I \partial_tm(t,s) \varphi(x(t,s)) ds + \int_I m(t,s) <\partial_t x(t,s), \nabla_x \varphi(x(t,s))> ds
$$
where $<\cdot,\cdot>$ denotes the inner product in $\mathbb{R}^d$. Let us deal with the first term. Using \eqref{eq:GL-TE}, we have
\[
\begin{split}
& \displaystyle \int_I \partial_t m(t,s) \varphi(x(t,s)) ds\\
=& \displaystyle \int_{I^{k+1}} m(t,s) m(t,s_1) \cdots m(t,s_k) \, S(x(t,s), x(t,s_1), \cdots, x(t,s_k) )\varphi(x(t,s)) \; ds\, ds_1\cdots ds_k\\
=& \displaystyle \int_{\mathbb{R}^d} \varphi(x) dh[\tilde{\mu}_t](x).
\end{split}
\]
Let us now deal with the second term. We start by rewriting the right-hand side of the first equation of \eqref{eq:GL-TE}.
$$
\displaystyle \int_I m(t,s_*) \phi(x(t,s_*)-x(t,s)) ds_*
= \int_{\mathbb{R}^d} \phi(x_*-x(t,s)) d\tilde{\mu}_t(x_*)
= V[\tilde{\mu}_t](x(t,s)).
$$
Thus, we obtain
\[
\begin{split}
\displaystyle \int_I m(t,s) <\partial_t x(t,s), \nabla_x \varphi(x(t,s))> ds = & \displaystyle \int_I m(t,s) <V[\tilde{\mu}_t](x(t,s)), \nabla_x \varphi(x(t,s))> ds \\
\displaystyle = & \displaystyle \int_{\mathbb{R}^d} <V[\tilde{\mu}_t](x), \nabla_x \varphi(x)> d \tilde{\mu}_t(x).
\end{split}
\]
Finally, we obtain
\begin{equation*}
\frac{d}{dt} (\tilde{\mu}_t,\varphi) = \int_{\mathbb{R}^d} <V[\tilde{\mu}_t](x), \nabla_x \varphi(x)> d \tilde{\mu}_t(x) + \int_{\mathbb{R}^d} \varphi(x) dh[\tilde{\mu}_t](x)
\end{equation*}
which is the weak version of \eqref{eq:mfl}.
\end{proof}
In order to prove Theorem \ref{Th:mfl2}, what is left is to show the convergence of $\mu^N$ to $\tilde{\mu}$, where $\mu^N$ is the empirical measure for the microscopic system defined in \eqref{eq:empmes}, and $\tilde{\mu}$ is defined from the solution to the graph limit equation by \eqref{eq:def_mu}.
The key point in order to do that is to rewrite $\mu^N$ using the functions $x_N$ and $m_N$ introduced to perform the graph limit and to pass to the limit in that expression. More precisely, we prove the following proposition:
\begin{prop}\label{prop:convmu}
Let $x_0\in L^\infty(I;\mathbb{R}^d) $ and $m_0\in L^\infty(I;\mathbb{R}^d)$.
Let $(x^{N},m^{N})\in \mathcal{C}([0,T]; \mathbb{R}^d)^{N}\times \mathcal{C}([0,T];\mathbb{R})^{N}$ satisfy the differential system~\eqref{eq:syst-gen} with initial condition
$x^{0,N}=P_\mathrm{d}^N(x_0)$ and $m^{0,N}=P_\mathrm{d}^N(m_0)$ given by \eqref{eq:ICx}
and mass dynamics given by \eqref{eq:psi}-\eqref{eq:psiSk}.
Let $\mu^N$ be the empirical measure associated with $(x^N,m^N)$, i.e. for all $t\in [0,T]$,
\[
\mu^N_t(x) := \frac{1}{N} \sum_{i=1}^N m^N_i(t) \delta_{x^N_i(t)}.
\]
Secondly,
let $(x,m)$ be the solutions to the integro-differential system \eqref{eq:GraphLimit-gen} with weight dynamics given by \eqref{eq:psiSk} and initial conditions given by $x(0,\cdot)=x_0$ and $m(0,\cdot)=m_0$.
Let
\[
\tilde{\mu}_t(x) := \int_I m(t,s) \delta(x-x(t,s)) ds.
\]
Then, for all test function $\varphi\in \mathcal{C}^\infty_c(\mathbb{R}^d)$, and all $t\in [0,T]$, it holds
\begin{equation*}
\lim_{N\rightarrow\infty} \int_{\mathbb{R}^d} \varphi(x) (d\mu^N_t(x) - d\tilde{\mu}_t(x)) =0.
\end{equation*}
\end{prop}
\begin{proof}
Let $x_N= P_\mathrm{c}^N(x^N)\in C([0,T];L^\infty(I;\mathbb{R}^d))$ and $m_N=P_\mathrm{c}^N(m^N)\in \mathcal{C}([0,T];L^\infty(I;\mathbb{R}))$ defined by
\eqref{eq:xN} be the solutions to the integro-differential system \eqref{eq:syst-contN}-\eqref{eq:psiSk} with initial conditions $x_N(0,\cdot) = P_\mathrm{c}^N(x^{0,N})$ and $m_N(0,\cdot) = P_\mathrm{c}^N(m^{0,N})$ given by \eqref{eq:ICx}.
We begin by showing that for all test function $\varphi\in \mathcal{C}^\infty_c(\mathbb{R}^d)$,
\begin{equation}\label{eq:equivmeas}
\int_{\mathbb{R}^d} \varphi(x) d\mu^N_t(x) = \int_{\mathbb{R}^d} \varphi(x) d\tilde{\mu}_t^N(x) ,
\end{equation}
where
$\tilde{\mu}^N_t\in\P(\mathbb{R}^d)$ is the measure defined by
\[
\tilde{\mu}^N_t(x) := \int_I m_N(t,s) \delta(x-x_N(t,s)) ds.
\]
Since $x_N(t,s) = \sum_{i=1}^N x_i^{N}(t) \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s)$ and $m_N(t,s) = \sum_{i=1}^N m_i^{N}(t) \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s)$, we can compute:
\[
\begin{split}
\int_{\mathbb{R}^d} \varphi(x) \int_I m_N(t,s) \delta(x-x_N(t,s)) ds = & \int_{\mathbb{R}^d} \varphi(x) \int_I \sum_{i=1}^N m_i^{N}(t) \mathbf{1}_{[\frac{i-1}{N}, \frac{i}{N})}(s) \delta(x-\sum_{i=1}^N x_i^{N}(t) ) ds\\
= & \int_{\mathbb{R}^d} \frac{1}{N} \sum_{i=1}^N m_i^{N}(t) \varphi(x_i^{N}(t)) dx
= \int_{\mathbb{R}^d} \varphi(x) d\mu^N_t(x) .
\end{split}
\]
Secondly, we prove the following weak convergence:
\begin{equation*}
\tilde{\mu}^N_t \rightharpoonup \tilde{\mu}_t \text{~when~} N \to \infty.
\end{equation*}
Using the definitions of $\tilde{\mu}^N_t$ and $\tilde{\mu}_t$, we write:
\[
\begin{split}
\bigg|\int_{\mathbb{R}^d} \varphi(x) & (d\tilde\mu^N_t(x) - d\tilde{\mu}_t(x))\bigg| = \left| \int_{\mathbb{R}^d} \varphi(x) \int_I (m_N(t,s) \delta(x-x_N(t,s)) - m(t,s) \delta(x-x(t,s))) \, ds \, dx \right| \\
= &\left| \int_{I} \varphi(x_N(t,s)) m_N(t,s) ds - \int_{I} \varphi(x(t,s)) m(t,s) ds \right| \\
= & \left| \int_{I} (\varphi(x_N(t,s))- \varphi(x(t,s)) ) m_N(t,s) ds - \int_{I} \varphi(x(t,s)) (m(t,s)- m_N(t,s) ) ds\right| \\
\leq & \left| \left( \int_I (\varphi(x_N(t,s))- \varphi(x(t,s)) )^2ds \right)^{1/2} \left(\int_I m_N(t,s)^2 ds \right)^{1/2} \right| \\
& + \left(\int_I \varphi(x(t,s))^2 ds \right)^{1/2} \left( \int_I (m_N(t,s) - m(t,s))^2 ds\right)^{1/2}\\
\leq & \left| K \left( \int_I ((x_N(t,s)- x(t,s) )^2ds \right)^{1/2} \left(\int_I m_N(t,s)^2 ds \right)^{1/2} \right| \\
& + \left(\int_I \varphi(x(t,s))^2 ds \right)^{1/2} \left( \int_I (m_N(t,s) - m(t,s))^2 ds\right)^{1/2}
\end{split}
\]
for a certain constant $K>0$, using the fact that $\varphi$ is a regular enough test function. Moreover, since the solution to \eqref{eq:syst-gen}
satisfies \eqref{eq:bornemN} and since $x \in \mathcal{C}([0;T];L^\infty(I;\mathbb{R}^d))$, $\varphi$ being regular, there exists a constant $C>0$ such that we finally have\\
\[
\begin{split}
& \left|\int_{\mathbb{R}^d} \varphi(x) (d\tilde\mu^N_t(x) - d\tilde{\mu}_t(x)) dx \right|\\
\leq & K \overline{M} \left( \int_I ((x_N(t,s)- x(t,s) )^2ds \right)^{1/2} + C \left( \int_I (m_N(t,s) - m(t,s))^2 ds\right)^{1/2}.
\end{split}
\]
Proposition \ref{prop:conv} allows us to conclude that
\[
\lim_{N\rightarrow\infty} \left|\int_{\mathbb{R}^d} \varphi(x) (d\mu^N_t(x) - d\tilde{\mu}_t(x)) dx \right| =0,
\]
and together with the equality \eqref{eq:equivmeas}, this proves the desired convergence.
\end{proof}
Theorem \ref{Th:mfl2} is thus proven by combining Propositions \ref{prop:conv}, \ref{Prop:link_mfl-gl} and \ref{prop:convmu}.
\section{Numerical simulations}
\label{sec:numeric}
\subsection{Dynamics not preserving indistinguishability}\label{sec:numericnotindisting}
In this section, we illustrate the convergence of the solutions of the microscopic model \eqref{eq:syst-gen} to the solution of the graph limit equation satisfying \eqref{eq:GraphLimit-gen}, as stated in Theorem \ref{Th:GraphLimit}.
We focus on mass dynamics that \textit{do not} preserve indistinguishability, so for which the classical mean-field limit process does not hold.
In particular, we consider a situation in which the agents are divided into $K$ groups $I_k$, with $k\in\{1,\cdots,K\}$. Each group is composed of leaders $I_k^L$ and followers $I_k^F$ so that $I_k=I_k^L\cup I_k^F$ and $I_k^L\cap I_k^F=\emptyset$. Within each group, the weight of each leader increases proportionally to itself and to the total weight of all the followers of the group. Conversely, the weight of each follower decreases proportionally to itself and to the total weight of all the leaders. More specifically,
consider the function $\psi$ given by
\begin{equation}\label{eq:modelsimuGL2}
\forall x\in L^2(I;\mathbb{R}),\quad \forall m\in L^2(I;\mathbb{R}),\qquad \psi(s,x,m) =
\begin{cases}
\beta m(s) \int_{I_k^F} m(s')ds' \quad \text{ if } s\in I_k^L \\
-\beta m(s) \int_{I_k^L} m(s')ds' \quad \text{ if } s\in I_k^F.
\end{cases}
\end{equation}
One easily checks that the total mass $\int_I m(s) ds$ is conserved, and that $\psi$ satisfies \eqref{eq:psilip2} on any time interval $[0,T]$ with $T<\infty$.
Given a total number of groups $K\in\mathbb{N}$ and a proportion $r\in (0,1)$ of leaders in each group, we define $I_k := [\frac{k-1}{K},\cdots,\frac{k}{K})$, $I_k^L = [\frac{k-1}{K},\cdots,\frac{k-1}{K}+\frac{r}{K})$ and $I_k^F = [\frac{k-1}{K}+\frac{r}{K},\cdots \frac{k}{K})$.
Provided that $n := N/K \in\mathbb{N}$ and that $rn\in\mathbb{N}$,
the corresponding microscopic dynamics can be written simply as:
\begin{equation}\label{eq:modelsimumicro2}
\forall k\in\{1,\cdots ,K\}, \quad \dot m_{(k-1)n + i} =
\begin{cases}
\displaystyle \frac{\beta}{N} m_{(k-1)n + i} \sum_{j=rn+1}^{n} m_{(k-1)n+j} \quad \text{ if } i\in\{1,\cdots , r n\}, \\
\displaystyle - \frac{\beta}{N} m_{(k-1)n + i} \sum_{j=1}^{rn} m_{(k-1)n+j} \quad \text{ if } i\in\{rn+1,\cdots , n\}.
\end{cases}
\end{equation}
We show the behavior of the model and the convergence of the microscopic dynamics to the macroscopic ones in cases in which we have one ($K=1$) and two ($K=2$) groups, with a proportion $r=10\%$ of leaders in each one. The initial conditions for the graph limit equation were taken to be $s\mapsto x_0(s) = \sin^2(4s)$ and $s\mapsto m_0(s)= 1/M_0*s \cos^2(5s)$, with $M_0:=\int_0^1 s \cos^2(5s) ds$. The corresponding initial conditions for the microscopic model were computed from \eqref{eq:ICx}. In all the following examples, the interaction function used was $y\mapsto a(y) = \frac{1}{1+y^2}$.
Figure \ref{fig:Micro_K1} shows the evolution of the opinions and weights of 20 agents divided into 2 leaders (indexed $i=1$ and $i=2$) and 18 followers (indexed $i=3... 20$). Observe that the leaders' weights quickly increase to sum up to the total weight of the group. As a consequence, consensus is achieved at a value $x=0.2$ close to the leaders' initial positions.
Figure \ref{fig:microGL2} illustrate the convergence of the microscopic dynamics to the graph limit ones by comparing the opinions and weights for $N=100$.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.18]{MicroK1_x.png}
\includegraphics[scale=0.18]{MicroK1_m.png}
\includegraphics[scale=0.5, clip=true]{SpaceI
\end{center}
\caption{ Evolution of opinions (left) and weights (right) of 20 agents. Center: color scale for the index $i$ ranging from 1 (blue) to 20 (red). Parameters: $K=1$, $r=0.1$, $\beta=5$.}\label{fig:Micro_K1}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=0.32\textwidth]{GLxN100t0p05.eps}
\includegraphics[width=0.32\textwidth]{GLxN100t1p4.eps}
\includegraphics[width=0.32\textwidth]{GLxN100t5.eps} \\
\includegraphics[width=0.32\textwidth]{GLmN100t0p05.eps}
\includegraphics[width=0.32\textwidth]{GLmN100t1p4.eps}
\includegraphics[width=0.32\textwidth]{GLmN100t5.eps} \\
\caption{\label{fig:microGL2} Representation of the solutions $s\mapsto x_N(t,s)$ (top) and $s \mapsto m_N(t,s)$ (bottom) to the microscopic dynamics \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro2} and of the solutions $s\mapsto x(t,s)$ (top) and $s\mapsto m(t,s)$ (bottom) to the Graph Limit equation \eqref{eq:GraphLimit-gen}-\eqref{eq:modelsimuGL2} for $t=0.05$, $t=1.4$ and $t=5$. Parameters: $K = 1$, $r = 0.1$, $\beta = 5$.}
\end{figure}
Figure \ref{fig:Micro_K2} shows the evolution of the opinions and weights of 20 agents divided into 2 groups, each one containing one leader (respectively indexed $i=1$ and $i=11$) and 9 followers (respectively indexed $i=2... 10$ and $i=12...20$).
Observe that the second group leaders' weights increases much faster than the first group leaders' weight, due to the fact that the total weigh of the second group is larger than the total weight of the first group. As a consequence, consensus is achieved at a value $x=0.6$ close to the second leader's initial position.
Figure \ref{fig:microGL3} illustrate the convergence of the microscopic dynamics to the graph limit ones by comparing the opinions and weights for $N=100$.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.18]{MicroK2_x.png}
\includegraphics[scale=0.18]{MicroK2_m.png}
\includegraphics[scale=0.5, clip=true]{SpaceI
\end{center}
\caption{ Evolution of opinions (left) and weights (right) of 20 agents. Center: color scale for the index $i$ ranging from 1 (blue) to 20 (red). Parameters: $K=2$, $r=0.1$, $\beta=5$.}\label{fig:Micro_K2}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=0.32\textwidth]{GL_x_N100t0p05.eps}
\includegraphics[width=0.32\textwidth]{GL_xN100t1p4.eps}
\includegraphics[width=0.32\textwidth]{GL_x_N100t5.eps} \\
\includegraphics[width=0.32\textwidth]{GL_m_N100_t0p05.eps}
\includegraphics[width=0.32\textwidth]{GL_m_N100_t1p4.eps}
\includegraphics[width=0.32\textwidth]{GL_m_N100t5.eps} \\
\caption{\label{fig:microGL3} Representation of the solutions $s\mapsto x_N(t,s)$ (top) and $s \mapsto m_N(t,s)$ (bottom) to the microscopic dynamics \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro2} and of the solutions $s\mapsto x(t,s)$ (top) and $s\mapsto m(t,s)$ (bottom) to the Graph Limit equation \eqref{eq:GraphLimit-gen}-\eqref{eq:modelsimuGL2} for $t=0.05$, $t=1.4$ and $t=5$. Parameters: $K = 2$, $r = 0.1$, $\beta = 5$.}
\end{figure}
\subsection{Dynamics preserving indistinguishability} \label{sec:numericindisting}
\subsubsection{The model}
In this first series of simulations, we will consider a particular case of mass dynamics of the form \eqref{eq:psiSk}. Recall that as shown in Section \ref{sec:indisting}, all mass dynamics of this form preserve indistinguishability, thus we can study the two limits - graph and mean-field - of the microscopic model.
More precisely, let us focus on the following model associated with the weight dynamics
\begin{equation}\label{eq:modelsimumicro}
\psi_i(x,m) = \frac{1}{N} m_i \left( \frac{1}{N} \sum_{k=1}^N \sum_{j=1}^N m_k m_j \| \phi(x_l-x_j)\| - \sum_{j=1}^N m_j \| \phi(x_i-x_j)\|\right).
\end{equation}
Let us explain its origin. We denote by $e_{j \to i}$ the influence of $j$ on $i$ and define it as $e_{j \to i} = m_j \phi(x_i-x_j)$. Let $e_i$ represent the total group influence on $i$, defined as $$e_i=\sum_{j=1}^N e_{j \to i} = \sum_{j=1}^N m_j \|\phi(x_i-x_j)\|.$$
Now denoting by $\overline{e}$ the weighted average of the total group influence $$\overline{e} = \sum_{k=1}^N \frac{m_k}{N} e_k = \sum_{k=1}^N \sum_{j=1}^N \frac{m_k}{N} m_j \|\phi(x_k-x_j)\|,$$
the mass dynamics of model \eqref{eq:modelsimumicro} can be rewritten as:
\[
\psi_i(x,m) = \frac{1}{N} m_i \left( \overline{e} - e_i \right).
\]
Thus, in our model, if the group influence on $i$ is lower than the weighted average of the group influence among the population (i.e. agent $i$ is being less influenced than average), $i$ gains weight proportionally to this difference and to its own weight $m_i$.
On the other hand, if the group influence on $i$ is higher than the weighted average among the population (i.e. agent $i$ is more influenced than average), $i$ loses weight proportionally to this difference and to its own weight $m_i$.
In other words, in this model, the less influenced agents gain weight, thus becoming the more influential.
The Graph Limit of the microscopic model \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} is given by \eqref{eq:GraphLimit-gen}, with
\begin{equation}\label{eq:modelsimuGL}
\psi(s,x,m) = m(s) \left( \int_I \int_I m(s_*) m(\tilde{s}) \| \phi(x(\tilde s)-x(s_*))\|\, ds_*\, d\tilde{s} - \int_I m(s_*) \| \phi(x(s)-x(s_*))\|\, ds_* \right).
\end{equation}
As seen in Section \ref{sec:mfl}, the mean-field limit of \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} is given by the transport equation with source \eqref{eq:mfl}, with
\begin{equation}\label{eq:modelsimumacro}
h[\mu](x) = \left(\int_{\mathbb{R}^d} \int_{\mathbb{R}^d} \|\phi(y-z)\| d\mu(z) d\mu(y) - \int_{\mathbb{R}^d} \|\phi(y-x)\| d\mu(y)\right) \mu(x).
\end{equation}
\subsubsection{Numerical results}
For the simplicity of numerical simulations, we take $d=1$ and
we choose an interaction function compactly supported on $[0,R]$. Let $a\in \mathcal{C}(\mathbb{R}^+, \mathbb{R})$ be defined by $a:\delta \mapsto a(\delta) = \frac{1}{\delta} \sin^2( \frac{\pi}{R}\delta ) \, \mathbbm{1}_{(0,R)}(\delta)$,
so that for all $x\in\mathbb{R}$,
\[
\phi(x) = a(|x|) x =
\begin{cases}
\frac{x}{|x|} \sin^2( \frac{\pi}{R}|x| )\quad \text{ for all } x\in (-R,R)\setminus 0 \\
0\quad \text{ otherwise}.
\end{cases}
\]
Initial conditions are given by the functions $s\mapsto x^0(s)$ and $s\mapsto m^0(s)$ defined by
\begin{equation}\label{eq:ICsimuGL}
\begin{cases}
x_0(s) = \frac{1}{\pi}\arccos(2s-1) \\
m_0(s) = \tilde{m}^0(s) (\int_I \tilde{m}_0(s_*) ds_*)^{-1}
\end{cases}
\end{equation}
where $\tilde{m}_0(s) = s^{1/4}\cos^2(5s)+ 0.2 s^2+0.5$.
Graphical representations of $x_0$ and $m_0$ can be found in the left panels of Figure \ref{fig:microGL}.
Notice that all opinions are initially in the interval $[0,1]$, thus if $R\geq 1$, all agents interact with all others, and we expect consensus. From here onward, we choose $R=0.2$.
We begin by showing numerical simulations of the microscopic model \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} for $N=30$. Initial conditions were computed from \eqref{eq:ICsimuGL}-\eqref{eq:ICx}.
The left panel of Figure \ref{fig:micro_30} shows the time evolution of the opinions $t\mapsto x_i(t)$ in which the (time-dependent) thickness of the lines is proportional to the corresponding weights $t\mapsto m_i(t)$, whose evolution is shown in the right panel.
Notice that due to the compact support of the interaction function, the population divides into three clusters separated by distances greater than $R$, the interaction radius.
Although the weights initially all start within the interval $[0.5,1.6]$, the weight dynamics spread the weights by leading the least influenced agents to gain mass. This can be observed in the evolution of $m_{30}$ (represented in light green), which feels little group influence since $x_{30}$ is at the lower edge of the group. Likewise, $m_1$ (in red) initially increases since $x_1$ is at the upper edge of the group, but after $x_1$ joins its closest neighbors, the group influence that it feels increases and $m_1$ decreases.
Observe also that the total mass is conserved, as shown by the constant evolution of the average mass (black dotted line).
\begin{figure}[!h]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{ModelMicro2_opinions1D-1.eps}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{ModelMicro2_weights1D-1.eps}
\end{subfigure}
\caption{\label{fig:micro_30} Evolution of opinions (left) and weights (right) for the microscopic model \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} with $N=30$ agents. The weighted barycenter $\bar x = \frac{1}{N}\sum_{i=1}^N m_i x_i$ and the average mass $\bar m = \frac{1}{N}\sum_{i=1}^N m_i$ are represented by black dotted lines. In the left panel, the thickness of the curve representing $x_i(t)$ is proportional to the corresponding weight $m_i(t)$.}
\end{figure}
Figure \ref{fig:microGL} shows the profile of the mean-field limit $s\mapsto x(t,s)$ and $s\mapsto m(t,s)$ solving \eqref{eq:GraphLimit-gen}-\eqref{eq:modelsimuGL} with initial conditions given by \eqref{eq:ICsimuGL} at times $t=0$, $t=0.45$ and $t=1.5$ (black line). For comparison purposes, the solutions $s\mapsto x^N(t,s)$ and $s\mapsto m^N(t,s)$ to the microscopic model \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} for $N=50$ is plotted on the same figures, using the representation via step functions given by \eqref{eq:xN} (red line).
The clustering behavior is now shown by the convergence of $x$ and $x^N$ to a step function taking three distinct values.
Notice that the weight function converges to a function with three local maxima attained at the centers of the three clusters, while the agents at each cluster's edge form local minima.
\begin{figure}[!h]
\centering
\includegraphics[width=0.3\textwidth]{MicGL_x_1.eps}
\includegraphics[width=0.3\textwidth]{MicGL_x_10.eps}
\includegraphics[width=0.3\textwidth]{MicGL_x_31.eps} \\
\includegraphics[width=0.3\textwidth]{MicGL_m_1.eps}
\includegraphics[width=0.3\textwidth]{MicGL_m_10.eps}
\includegraphics[width=0.3\textwidth]{MicGL_m_31.eps} \\
\caption{\label{fig:microGL} Representation of the solutions $s\mapsto x_N(t,s)$ (top) and $s \mapsto m_N(t,s)$ (bottom) to the microscopic dynamics \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} and of the solutions $s\mapsto x(t,s)$ (top) and $s\mapsto m(t,s)$ (bottom) to the Graph Limit equation \eqref{eq:GraphLimit-gen}-\eqref{eq:modelsimuGL} for $t=0$, $t=0.45$ and $t=1.5$.}
\end{figure}
Figure \ref{fig:microGLmacro} represents the solution $x\mapsto \mu_t(x)$ to the mean-field equation \eqref{eq:mfl}-\eqref{eq:modelsimumacro} with initial condition given by
\[
\mu_0(x) = \int_I m_0(s) \delta( x-x_0(s)) ds
\]
at times $t=0$, $t=0.45$ and $t=1.5$. Again we observe convergence of the population to three clusters, as the measure converges to three Dirac masses located at the centers of mass of the clusters. As in the previous two representations, convergence to the left-most cluster is slower than convergence to the center and right clusters.
For comparison, the solution to the microscopic model \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} with $N=50$ was plotted on the same figure (red line) using the step-function representation $\mu_t^{N,n}$ defined as follows:
\[
\forall j\in \{1,...n\}, \quad \forall x\in E_j, \quad \mu^{N,n}_t(x):= \sum_{i=1}^N \frac{m_i(t)}{|E_j|}\mathbbm{1}_{E_j}(x_i(t)),
\]
where for all $j\in \{1,...n\}$, $E_j:=[\frac{j-1}{n},\frac{j}{n})$, and $n=25$.
Subordination of the mean-field limit to the graph limit is shown by representing the solution to the graph limit equation \eqref{eq:GraphLimit-gen}-\eqref{eq:modelsimuGL}
\[
\forall x\in \mathbb{R}, \quad \tilde{\mu}_t(x) = \int_I m(t,s) \delta( x-x(t,s)) ds.
\]
The solution $t\mapsto (x_i(t), m_i(t))_{i\in\{1,\cdots,N\}}$ to the system of ODE was computed using the Matlab solver ode45.
The solution $(t,s)\mapsto (x(t,s),m(t,s))$ to the integro-differential equation was computed using Euler's method for time differentiation and Simpson's method for space integration.
Lastly, the solution $(t,x)\mapsto\mu(t,x)$ to the transport PDE with source was computed using a standard Lax-Wendroff
scheme.
\begin{figure}[!h]
\centering
\includegraphics[width=0.3\textwidth]{MicMac_1.eps}
\includegraphics[width=0.3\textwidth]{MicMac_10.eps}
\includegraphics[width=0.3\textwidth]{MicMac_31.eps} \\
\caption{\label{fig:microGLmacro} Representation of the solution $\mu_t^{N,n}$ (red) to the microscopic dynamics \eqref{eq:syst-gen}-\eqref{eq:modelsimumicro} for $N=50$ and $n=25$, of the solution $\tilde{\mu}_t$ (black) to the graph limit equation \eqref{eq:GraphLimit-gen}-\eqref{eq:modelsimuGL} and of the solution $\mu_t$ (blue) to the transport equation with source \eqref{eq:mfl}-\eqref{eq:modelsimumacro} at times $t=0$, $t=0.45$ and $t=1.5$.}
\end{figure}
\newpage
| 1d9f0b0ac56e085c07357f25fbc3f2049676f270 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The problem of recovering an unknown low-rank tensor from a small
fraction of its entries is known as the tensor completion problem,
and comes up in a wide range of applications,
e.g., image processing \cite{imbiriba2018low, ng2017adaptive, zhang2021low},
computer vision \cite{liu2013, zhang2019corrected},
and machine learning \cite{romera2013multilinear, signoretto2014learning}.
The goal of low-rank tensor
completion is to recover a tensor with the lowest rank
based on observable entries of a given tensor.
Given a third-order tensor $\mathcal{M}\in\mathbb{C}^{n_1\times n_2\times n_3}$, the low rank tensor completion problem
can be expressed as follows:
\begin{equation} \label{n1}
\begin{split}
\min_{\mathcal{Z}}& ~ \operatorname{rank}(\mathcal{Z})\\
\text{s.t.}&~ \mathcal{P}_{\Omega}(\mathcal{Z})=\mathcal{P}_{\Omega}(\mathcal{M}),
\end{split}
\end{equation}
where $\textup{rank}(\mathcal{Z})$ denotes the rank of the tensor $\mathcal{Z}$,
$\Omega$ is a subset of $\{1,\ldots, n_1\}\times\{1,\dots,n_2\}\times\{1,\dots,n_3\}$,
and $\mathcal{P}_\Omega$ is the projection operator
such that the entries in $\Omega$ are given while the remaining entries are missing, i.e.,
\begin{align*}\label{p1}
(\mathcal{P}_\Omega(\mathcal{Z}))_{ijk}= \left\{
\begin{array}{ll}
\mathcal{Z}_{ijk}, & \mbox{if}\ (i,j,k)\in\Omega, \\
0, &\mbox{otherwise}.
\end{array}
\right.
\end{align*}
In particular, if $n_{3}=1$,
the tensor completion problem in (\ref{n1}) reduces to
the well-known matrix completion problem,
which has received a considerable amount of attention in the past decades, see, e.g.,
\cite{candes2010matrix,candes2009exact,candes2010the,gross2010,recht2010guaranteed}
and references therein.
In the matrix completion problem, a given incoherent $n\times n$ matrix could be recovered with high probability
if the uniformly random sample size is of order $O(r n \log (n)),$
where $r$ is the rank of the given matrix.
This bound has been shown to be optimal, see \cite{candes2010the} for detailed discussions.
The main aim of this paper is to study the incoherence conditions
of low-rank tensors based on the transformed tensor singular value decomposition (SVD)
and provide a lower bound on the number of random sample entries required for exact tensor recovery.
Different kinds of tensor ranks setting in model \eqref{n1} lead to different convex relaxation models
and different sample sizes required to exactly recover the original tensor.
When the rank in the model \eqref{n1} is chosen as the Tucker rank \cite{Tucker66},
Liu et al. \cite{liu2013}
proposed to use the sum of the nuclear norms (SNN) of unfolding matrices
of a tensor to recover a low Tucker rank tensor. Within the SNN framework,
Tomioka et al. \cite{tomioka2011statistical}
proved that a given $d$-order tensor $\mathcal{X}\in\mathbb{R}^{n\times\cdots\times n}$
with Tucker rank $(r,\ldots,r)$ can be exactly recovered with high probability
if the Gaussian measurements size is of order $O(rn^{d-1})$.
Afterwards, Mu et al. \cite{mu2014square} showed that $O(rn^{d-1})$ Gaussian measurements
are necessary for a reliable recovery by the SNN method.
In fact, the degree of freedoms of a tensor
$\mathcal{X}\in \mathbb{R}^{n \times\cdots\times n}$ with
Tucker rank $(r_{1},\ldots,r_{d})$ is $\prod_{i=1}^dr_i+\sum_{i=1}^{d}(r_{i}n -r_{i}^{2})$,
which is much smaller than $O(rn^{d-1})$.
Recently, Yuan et al. \cite{yuan2016tensor,yuan2017incoherent}
showed that an $n\times n\times n$ tensor with Tucker rank $(r,r,r)$ can be exactly recovered with high probability
by $O((r^{\frac{1}{2}}n^{\frac{3}{2}}+r^2n)\log^2(n))$ entries,
which have a great improvement compared with the
number of sampled sizes required in \cite{mu2014square} when $n$ is relatively large.
Later, based on a gradient descent algorithm designed on some product smooth manifolds,
Xia et al. \cite{Xia2019On} showed that an $n\times n\times n$ tensor with multi-linear rank
$(r,r,r)$ can be reconstructed with high probability by $O(r^{\frac{7}{2}}n^{\frac{3}{2}}\log^{\frac{3}{2}}(n)+r^7n\log^6 (n))$ entries. When the rank in the model \eqref{n1} is chosen as the CANDECOMP/PARAFAC (CP) rank
\cite{Kolda2009},
Mu et.al \cite{mu2014square} introduced a square deal method
which only uses an individual nuclear norm of a balanced matrix instead of
using a combination of all $d$ nuclear norms of unfolding matrices of the tensor. Moreover,
they showed that $O(r^{\lfloor\frac{d}{2}\rfloor}n^{\lceil \frac{d}{2}\rceil })$ samples
are sufficient to recover a CP rank $r$ tensor with high probability.
Besides, some tensor estimation and recovery problems for the observations
with Gaussian measurements were proposed
and studied in the literature
\cite{ahmed2020tensor, cai2020provable, luo2021low, rauhut2017low, tong2021scaling}.
For example, Ahmed et al. \cite{ahmed2020tensor} proposed
and studied the tensor regression problem by using low-rank and sparse Tucker decomposition,
where a tensor variant of projected gradient descent was proposed
and the sample complexity of this algorithm
for a $d$-order tensor $n_1\times \cdots\times n_d$ with Tucker rank $(r_1,\ldots, r_d)$
is $O(\bar{r}^d+\bar{s}\bar{r}d\log^2(3\bar{n}d))$
under the restricted isometry property for sub-Gaussian linear maps.
Here $\bar{r}=\max\{r_1,\ldots,r_d\}$ and $\bar{n}=\max\{n_1,\ldots, n_d\}$
and $\bar{s}=\max\{s_1,\ldots, s_d\}$,
where $s_i$ is the upper bound of the number of nonzero entries of each column of the $i$-th factor matrix in the Tucker decomposition.
Moreover, Cai et al. \cite{cai2020provable} showed that
the Riemannian gradient algorithm can reconstruct
a $d$-order tensor of size $n\times \cdots \times n$ and Tucker rank $(r, \ldots, r)$ with high
probability from only $O(nr^2 + r^{d+1})$ measurements under the tensor restricted isometry property for Gaussian measurements,
where one step of iterative hard
thresholding was used for the initialization.
The tubal rank of a third-order tensor was first proposed by Kilmer et al. \cite{kilmer,Kilmer2011},
which is based on tensor-tensor product
(t-product). Within the associated algebraic framework of t-product,
the tensor SVD was proposed and studied similarly to matrix SVD \cite{Kilmer2011}.
For tensor tubal rank minimization, Zhang et al. \cite{Zhang2017}
proved that the tensor tubal nuclear norm (TNN) can be used as
a convex relaxation of the tensor tubal rank.
Then they showed that an $n\times n\times n$ tensor with tubal rank $r$
can be exactly recovered by $O(rn^{2}\log(n^2))$ uniformly sampled entries.
However, the TNN is not the convex envelope of the tubal rank of a tensor,
which may lead to more sample entries needed to exactly recover the original tensor.
In Table \ref{table1}, we summarize existing results and our contribution for the $n\times n\times n$ tensor completion problem.
It is interesting that there are other
factors that affect sample sizes requirement such as sampling methods and incoherence conditions.
For detailed discussions, the interested readers are referred to
\cite{barak2016noisy,jain2014provable,krishnamurthy2013low,montanari2018spectral}.
\begin{table}[h!]
\small
\caption{Sampling sizes and sampling methods for third-order tensor completion.} \label{table1}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Rank & Sampling & Incoherent and Other & Requirement \\
Assumption& Method &Conditions&Sampling Sizes\\\hline
CP rank $r$ \cite{mu2014square}& Gaussian& N/A &$O(rn^{2})$ \\\hline
\multirow{2}{*}{CP rank $r$ \cite{jain2014provable}}&Uniformly & Incoherent condition of & \multirow{2}{*}{$O(n^{\frac{3}{2}}r^{5}\log^4(n))$}\\
& Random&symmetric tensor & \\\hline
Tucker rank \cite{yuan2017incoherent}&Uniformly& Matrix incoherent condition & \multirow{2}{*}{$O(rn^{\frac{3}{2}}+r^2n)\log^2(n)$} \\
$(r,r,r)$& Random &on model-$n$ unfolding & \\\hline
Tucker rank \cite{huang2014provable}& \multirow{2}{*}{Random} & Matrix incoherent condition & \multirow{2}{*}{$O(rn^{2}\log^2(n))$}\\
$(r,r,r)$& &on model-$n$ unfolding & \\\hline
Tucker rank \cite{mu2014square}& \multirow{2}{*}{Gaussian} & \multirow{2}{*}{N/A} & \multirow{2}{*}{$O(rn^{2})$} \\
$(r,r,r)$ & & & \\\hline
Tucker rank \cite{Xia2019On}&Uniformly& Matrix incoherent condition & \multirow{2}{*}{$O(r^{\frac{7}{2}}n^{\frac{3}{2}}\log^{\frac{3}{2}}(n)+r^7n\log^6 (n))$} \\
$(r,r,r)$& Random &on model-$n$ unfolding & \\\hline
\multirow{2}{*}{Tubal rank $r$ \cite{Zhang2017}}& Uniformly & \multirow{2}{*}{Tensor incoherent condition} & \multirow{2}{*}{$O(rn^{2}\log(n^2))$} \\
& Random & & \\\hline
multi-rank $(r_{1},\ldots,r_{n})$ & Uniformly & \multirow{2}{*}{Tensor incoherent condition} & \multirow{2}{*}{$O(\sum_{i=1}^{n} r_{i}n\log(n^2))$} \\
(in this paper)& Random & & \\
\hline
\end{tabular}
\end{center}
\end{table}
In this paper, we mainly study the $n_1 \times n_2 \times n_3$
third-order tensor completion problem based on transformed tensor
SVD and transformed tensor nuclear norm (TTNN) \cite{song2019robust}. We show that
such low-rank tensors can be exactly recovered with high probability when the number of
randomly observed entries is of order $O( \sum_{i=1}^{n_3}r_i \max \{ n_1, n_2 \} \log ( \max \{ n_1, n_2 \} n_3))$,
where $r_i$ is the $i$-th element of the transformed multi-rank of a tensor.
The rest of this paper is organized as follows.
In Section \ref{Sect2}, the transformed tensor
SVD related to an arbitrary unitary transformation is reviewed.
In Section \ref{mainresults}, we provide the bound on the
number of sample entries for tensor completion via any
unitary transformation. In Section \ref{Sect4},
several synthetic data and imaging data sets
are performed to demonstrate
that our theoretical result is valid and the performance of the proposed method
is better than the existing methods in terms of sample sizes requirement.
Some concluding remarks are given in Section \ref{Sect5}.
Finally, the proofs of auxiliary lemmas supporting our main
theorem are provided in the appendix.
\section{Transformed Tensor Singular Value Decomposition}\label{Sect2}
First, some notations used throughout this paper are introduced.
$\mathbb{N}_{\geq 0}^{n}$ denotes the nonnegative $n$-dimensional integers space.
We use ${\bf \Phi}$ to denote an arbitrary unitary matrix, i.e.,
${\bf \Phi} {\bf \Phi}^H = {\bf \Phi}^H {\bf \Phi} = I$,
where ${\bf \Phi}^H$ is the conjugate transpose of ${\bf \Phi}$ and $I$ is the identity matrix whose dimension
should be clear from the context.
Tensors are represented by capital Euler script letters,
e.g., $\mathcal{A}$.
A tube of a third-order tensor is defined by fixing the first two indices and varying the third \cite{Kilmer2011}.
Let $\mathcal{A}\in\mathbb{C}^{n_1\times n_2\times n_3}$ be a third-order tensor.
$\mathcal{A}_{ijk}$ denotes the $(i,j,k)$-th entry of $\mathcal{A}$.
We use $\hat{\mathcal{A}}_{{\bf \Phi}}$ to denote a
third-order tensor obtained as follows:
\begin{equation*}
\textrm{vec}\left(\hat{\mathcal{A}}_{{\bf \Phi}}(i,j,:)\right)={{\bf
\Phi}} \left(\textrm{vec}(\mathcal{A}(i,j,:))\right),
\end{equation*}
where $\textrm{vec}(\cdot)$ is the vectorization operator from
$\mathbb{C}^{1\times1\times n_3}$ to $\mathbb{C}^{n_3}$ and $\mathcal{A}(i,j,:)$ denotes the $(i,j)$-th tube of $\mathcal{A}$.
For simplicity, we denote $\hat{\mathcal{A}}_{{\bf
\Phi}}={\bf \Phi}[\mathcal{A}]$.
In the same fashion, one can also compute $\mathcal{A}$
from $\hat{\mathcal{A}}_{{\bf \Phi}}$,
i.e., ${\cal A} = {\bf\Phi}^H[\hat{\mathcal{A}}_{{\bf \Phi}}]$.
A block diagonal matrix can be derived by the frontal slices of ${\cal A}$ using the ``blockdiag'' operator:
$$
\textrm{blockdiag}(\mathcal{\mathcal{A}}):=
\left(
\begin{array}{cccc}
\mathcal{A}^{(1)} & & & \\
& \mathcal{A}^{(2)}& & \\
& & \ddots & \\
& & &\mathcal{A}^{(n_{3})}\\
\end{array}
\right),
$$
where $\mathcal{A}^{(i)}$ is the $i$-th frontal
slice of ${\cal A},~i=1,\ldots,n_{3}$.
Conversely, the block diagonal matrix $\textrm{blockdiag}(\mathcal{\mathcal{A}})$ can be converted
into a tensor via the following ``fold'' operator:
$$
\textrm{fold}(\textrm{blockdiag}(\mathcal{\mathcal{A}})) := \mathcal{\mathcal{A}}.
$$
After introducing the tensor notation and terminology,
we give the basic definitions about the tensor product,
the conjugate transpose of a tensor, the identity tensor and the unitary tensor with respect to the unitary transformation matrix ${\bf \Phi}$, respectively.
\begin{definition}\cite[Definition 1]{song2019robust}
\label{def1} The ${\bf \Phi}$-product of $\mathcal{A} \in \mathbb{C}^{n_1
\times n_2 \times n_3}$ and $\mathcal{B} \in \mathbb{C}^{n_2 \times n_4
\times n_3}$ is a tensor $\mathcal{C} \in \mathbb{C}^{n_1 \times n_4 \times
n_3}$, which is given by
\begin{equation*}
\mathcal{C} =\mathcal{A} \diamond_{\bf \Phi} \mathcal{B}={\bf
\Phi}^{H}\left[ \emph{fold} \left(
\emph{blockdiag}(\mathcal{\hat{A}}_{\bf \Phi}) \cdot
\emph{blockdiag}(\mathcal{\hat{B}}_{\bf \Phi})\right) \right].
\end{equation*}
\end{definition}
\begin{remark}
Kernfeld et al. \cite{kernfel} defined
the tensor product between two tensors by using frontal slices
products in the transformed domain based on an arbitrary
invertible linear transformation.
In this paper, we mainly focus on the tensor product
based on unitary transformations.
Moreover, the relation between ${\bf \Phi}$-product
and tensor product by using fast Fourier transform (FFT) \cite{Kilmer2011} is shown in \cite{song2019robust}.
\end{remark}
\begin{definition}\cite[Definition 2]{song2019robust}
\label{def2.5} For any $\mathcal{A}\in
\mathbb{C}^{n_{1}\times n_{2}\times n_{3}}$, its conjugate transpose with respect to ${\bf \Phi}$,
denoted by $\mathcal{A}^{H}\in \mathbb{C}^{n_{2}\times n_{1}\times n_{3}}$, is defined as
\begin{equation*}
\mathcal{A}^{H} = {\bf \Phi}^{H} \left [ \emph{fold} \left
(\emph{blockdiag}(\mathcal{\hat{A}}_{\bf \Phi})^{H}\right
)\right].
\end{equation*}
\end{definition}
\begin{definition} \cite[Proposition 4.1]{kernfel}\label{idd}
The identity tensor $\mathcal{I}_{\bf \Phi} \in \mathbb{C}^{n \times n
\times n_3}$ (with respect to ${\bf \Phi}$) is defined to be a
tensor such that ${\cal I}_{\bf \Phi} = {\bf \Phi}^H [ {\cal T}]$,
where each frontal
slice of ${\cal T} \in \mathbb{R}^{n \times n \times n_3}$ is the $n \times
n$ identity matrix.
\end{definition}
\begin{definition} \cite[Definition 5.1]{kernfel}
A tensor $\mathcal{Q} \in \mathbb{C}^{n \times n \times n_3}$ is
unitary with respect to ${\bf \Phi}$-product
if it satisfies
\begin{equation*}
\mathcal{Q}^{H} \diamond_{{\bf \Phi}} \mathcal{Q} = \mathcal{Q}
\diamond_{\bf \Phi} \mathcal{Q}^{H} =\mathcal{I}_{\bf \Phi}.
\label{eq7}
\end{equation*}
\end{definition}
In addition, $\mathcal{A}$ is a diagonal tensor if and only if each frontal
slice $\mathcal{A}^{(i)}$ of $\mathcal{A}$ is a diagonal matrix. By above definitions, the
transformed tensor SVD with
respect to ${\bf \Phi}$ can be given as follows.
\begin{theorem} \cite[Theorem 5.1]{kernfel}\label{them1}
Suppose that $\mathcal{A}\in \mathbb{C}^{n_{1}\times n_{2}\times n_{3}}$.
Then $\mathcal{A}$ can be factorized as
\begin{equation} \label{equ12}
\mathcal{A}= \mathcal{U} \diamond_{\bf \Phi} \mathcal{S}
\diamond_{\bf \Phi} \mathcal{V}^{H},
\end{equation}
where $\mathcal{U}\in \mathbb{C}^{n_{1}\times n_{1}\times n_{3}},$
$\mathcal{V}\in \mathbb{C}^{n_{2}\times n_{2}\times n_{3}}$ are unitary
tensors with respect to ${\bf \Phi}$-product, and $\mathcal{S}\in
\mathbb{C}^{n_{1}\times n_{2}\times n_{3}}$ is a diagonal tensor.
\end{theorem}
Based on the transformed tensor SVD given in Theorem \ref{them1}, the transformed multi-rank and tubal rank of a tensor can be defined as follows.
\begin{definition}\cite[Definition 6]{song2019robust}
(i) The transformed multi-rank of $\mathcal{A} \in
\mathbb{C}^{n_1 \times n_2 \times n_3}$, denoted by
$\operatorname{rank}_{t}(\mathcal{A})$, is a vector $\textup{\bf r} \in \mathbb{N}_{\geq 0}^{n_3}$
with its $i$-th entry being the rank of the $i$-th frontal slice of
$\hat{\mathcal{A}}_{\bf \Phi}$, i.e.,
\begin{equation*}
\operatorname{rank}_{t}(\mathcal{A}_{\bf \Phi})=\textup{\bf r}~~ \textup{with}~~r_i = \operatorname{rank}(\hat{\mathcal{A}}_{\bf \Phi}^{(i)}),~~i=1,\ldots,n_{3}.
\end{equation*}
(ii) The transformed
tubal rank of $\mathcal{A} \in
\mathbb{C}^{n_1 \times n_2 \times n_3}$, denoted by $\operatorname{rank}_{tt}(\mathcal{A})$, is defined as the
number of nonzero singular tubes of $\mathcal{S}$, where
$\mathcal{S}$ comes from the transformed tensor SVD of $\mathcal{A}$,
i.e.,
\begin{equation}
\operatorname{rank}_{tt}(\mathcal{A}) = \#\{i: \mathcal{S}(i, i, :) \neq \vct{0}\}
= \max_{i} r_i. \label{eq9}
\end{equation}
\end{definition}
\begin{remark}
For computational improvement,
we will use the skinny transformed tensor SVD throughout this paper unless otherwise stated, which is defined as follows:
The skinny transformed tensor SVD of $\mathcal{A}\in \mathbb{C}^{n_{1}\times n_{2}\times n_{3}}$ with $\emph{rank}_{tt}(\mathcal{A})=r$ is given as $\mathcal{A}= \mathcal{U} \diamond_{\bf \Phi} \mathcal{S}
\diamond_{\bf \Phi} \mathcal{V}^{H},$ where $\mathcal{U}\in \mathbb{C}^{n_{1}\times r\times n_{3}}$ and
$\mathcal{V}\in \mathbb{C}^{n_{2}\times r\times n_{3}}$ satisfying $\mathcal{U}^H\diamond_{\bf \Phi}\mathcal{U}=\mathcal{I}_{\bf \Phi}, \mathcal{V}^H\diamond_{\bf \Phi}\mathcal{V}=\mathcal{I}_{\bf \Phi}$, and $\mathcal{S}\in
\mathbb{C}^{r\times r\times n_{3}}$ is a diagonal tensor. Here $\mathcal{I}_{\bf \Phi}\in\mathbb{C}^{r\times r\times n_3}$ is the identity tensor.
\end{remark}
The inner product of two
tensors $\mathcal{A},\mathcal{B}\in\mathbb{C}^{n_1\times n_2\times n_3}$
related to the unitary transformation ${\bf \Phi}$ is defined as
\begin{equation} \label{tprod}
\langle\mathcal{A}, \mathcal{B} \rangle = \sum_{i=1}^{n_{3}}\langle
\mathcal{A}^{(i)}, \mathcal{B}^{(i)} \rangle =\langle
\overline{\mathcal{A}}_{\bf \Phi}, \overline{\mathcal{B}}_{\bf \Phi}
\rangle,
\end{equation}
where $\langle
\mathcal{A}^{(i)}, \mathcal{B}^{(i)} \rangle$ is the usual inner product of two matrices
and $\overline{\mathcal{A}}_{\bf \Phi}=\textrm{blockdiag}(\mathcal{A}_{\bf \Phi})$.
The following fact will be used throughout the paper:
For any tensor $\mathcal{A}\in \mathbb{C}^{n_{1}\times n_{2}\times n_3}$
and $\mathcal{B}\in \mathbb{C}^{n_{2}\times n_{4}\times n_3},$ one can get that
$\mathcal{A}\diamond_{\bf \Phi}\mathcal{B}=\mathcal{C}\Leftrightarrow
\overline{\mathcal{A}}_{\bf \Phi}\cdot\overline{\mathcal{B}}_{\bf \Phi}=\overline{\mathcal{C}}_{\bf \Phi}.$
The tensor spectral norm of an arbitrary tensor
$\mathcal{A}\in \mathbb{C}^{n_1 \times n_2 \times n_3}$ related to ${\bf
\Phi}$, denoted by $\|\mathcal{A}\|$, can be defined as
$\|\mathcal{A}\| = \|\overline{\mathcal{A}}_{\bf \Phi}\|$ \cite{song2019robust},
i.e., the spectral norm of its block
diagonal matrix $\overline{\mathcal{A}}_{\bf \Phi}$ in the transformed
domain. Suppose that ${\tt L}$
is a tensor operator, its operator norm is defined as
$\|{\tt L}\|_{\textup{op}} = \sup_{\|\mathcal{A}\|_F \leq 1} \|{\tt
L}(\mathcal{A})\|_F,$
where the tensor Frobenius norm of $\mathcal{A}$ is defined as $\|\mathcal{A}\|_{F} =
\sqrt{\sum_{i,j,k}|\mathcal{A}_{ijk}|^2}$.
Specifically, if the operator norm can be
represented as a tensor $\mathcal{L}$ via ${\bf \Phi}$-product with
$\mathcal{A}$, we have $\| {\tt L}\|_{\textup{op}} = \|\mathcal
{L}\|$. The tensor infinity norm and
the tensor $l_{\infty,2}$ and are defined as
\[
\begin{split}
&\|\mathcal{A}\|_{\infty} = \max_{i,j,k}|\mathcal{A}_{ijk}| ~ \textup{and} ~ \|\mathcal{A}\|_{\infty,2}=\max\left\{\max_{i}\sqrt{\sum_{b,k}|\mathcal{A}_{ibk}|^2},\max_{j}\sqrt{\sum_{a,k}|\mathcal{A}_{ajk}|^2} \right\}.
\end{split}
\]
Moreover, the weighted tensor $l_{\infty,w}$
norm with respect to a weighted vector $w=(\alpha_1,\ldots,\alpha_{n_3})^H\in\mathbb{R}^{n_3}$ is defined as
\begin{equation}\label{norm1}
\|\mathcal{A}\|_{\infty, w}=\max\left\{\max_{i}\sqrt{\sum_{b,k}\alpha^2_{k}|\mathcal{A}_{ibk}|^2},\max_{j}\sqrt{\sum_{a,k} \alpha^2_{k}|\mathcal{A}_{ajk}|^2} \right\},
\end{equation}
where $\sum_{k=1}^{n_3}\alpha^2_{k}=1$.
The aim of this paper is to recover a low transformed
multi-rank tensor, which motivates us to introduce the following definition
of TTNN.
\begin{definition}\cite[Definition 7]{song2019robust}
The transformed tensor nuclear norm of $\mathcal{A} \in
\mathbb{C}^{n_1 \times n_2 \times n_3}$, denoted by
$\|\mathcal{A}\|_{\textup{TTNN}}$, is the sum of nuclear norms of
all frontal slices of $\hat{\mathcal{A}}_{\bf \Phi}$, i.e.,
$\|\mathcal{A}\|_{\textup{TTNN}} =
\sum_{i=1}^{n_3}
\|{\hat{\mathcal{A}}}^{(i)}_{\bf \Phi}\|_{\ast}$. \label{def9}
\end{definition}
Recently, Song et al. \cite{song2019robust} showed that the
TTNN of a tensor is the convex envelope of the sum of the
entries of the transformed multi-rank of a tensor, which is stated in the following.
\begin{lemma}\cite[Lemma 1]{song2019robust}\label{lemm1}
For any tensor $\mathcal{A} \in \mathbb{C}^{n_1 \times n_2 \times n_3}$,
let $\operatorname{rank}_{sum}(\mathcal{A})=\sum_{i=1}^{n_3}
\operatorname{rank}(\hat{\mathcal{A}}^{(i)}_{\bf \Phi})$.
Then $\|\mathcal{A}\|_{\textup{TTNN}}$ is
the convex envelope of $\operatorname{rank}_{sum}(\mathcal{A})$ on
the set $\{ \mathcal{A} \ | \ \| \mathcal{A}\| \leq 1\}$.
\end{lemma}
Lemma \ref{lemm1} shows that the
TTNN is the tightest convex relaxation of the sum of the entries of the transformed
multi-rank of the tensor over a
unit ball of the tensor spectral norm.
That is why the TTNN is effective in
studying the tensor recovery theory with transformed multi-rank minimization.
Next we introduce two kinds of tensor basis which will be exploited to
derive our main result.
\begin{definition} \label{defn}
(i) The column basis, denoted as
$\tc{e}_{ik}$, is a tensor of size $n_1 \times 1 \times n_3$ with the
$(i,1,k)$-th element equaling to $1$ and the others equaling to 0.
(ii) Denote $\tub{e}_{k}$ as a tensor of
size $1 \times 1 \times n_3$ with the $(1,1,k)$-th element equaling to $1$ and
the remaining elements equaling to 0, $({\bf \Phi}[\tub{e}_{k}])_{j}$
as the $(1,1,j)$-th element of ${\bf \Phi}[\tub{e}_{k}]$, $j=1,\ldots,n_{3}$.
(iii) The transformed tube
basis, denoted as $\ddot{\textbf{e}}_k$,
is a tensor of
size $1 \times 1 \times n_3$ with the $(1,1,j)$-th element of $\ddot{\textbf{e}}_k$
equaling to $(({\bf \Phi}[\tub{e}_{k}])_{j})^{-1},$ if $({\bf \Phi}[\tub{e}_{k}])_{j}\neq0,$ and $0$, otherwise, $j=1,\ldots, n_3$.
\end{definition}
\begin{remark}
The transformed tube basis $\ddot{\textbf{e}}_k$ is determined by
the unitary transformation matrix and the original tube $\tub{e}_{k}$,
whose detailed formulation
can be obtained for a given unitary transformation ${\bf \Phi}$.
\end{remark}
\section{Main Results} \label{mainresults}
In this section, we consider the third-order tensor completion problem,
which aims to recover a low transformed multi-rank tensor under some limited observations. Mathematically, the problem can be described as follows:
\begin{equation*} \label{nj1}
\begin{split}
\min_{\mathcal{Z}}& ~ \sum_{i=1}^{n_3}\operatorname{rank}(\hat{\mathcal{Z}}_{\bf \Phi}^{(i)}) \\
\text{s.t.}&~ \mathcal{P}_{\Omega}(\mathcal{Z})=\mathcal{P}_{\Omega}(\mathcal{M}),
\end{split}
\end{equation*}
where $\operatorname{rank}(\hat{\mathcal{Z}}_{\bf \Phi}^{(i)})$ is
the rank of $\hat{\mathcal{Z}}_{\bf \Phi}^{(i)}$, i.e.,
the $i$-th element of the transformed multi-rank of $\mathcal{Z}, i=1,\ldots, n_3$,
$\Omega$ and $\mathcal{P}_{\Omega}(\mathcal{Z})$ are defined in model \eqref{n1}.
Note that the rank minimization problem is NP-hard
and the TTNN is the convex envelope of the
sum of the entries of the transformed multi-rank of a tensor \cite[Lemma 1]{song2019robust}.
Therefore, we propose to utilize the TTNN as a convex
relaxation of the sum of the entries of the transformed multi-rank of a tensor.
More precisely speaking, the convex
relaxation model is given by
\begin{equation}\label{ObjFT}
\begin{split}
\min_{\mathcal{Z}} & \
\|\mathcal{Z}\|_{\textup{TTNN}} \\
\textup{s.t.} & \
\mathcal{P}_\Omega(\mathcal{Z}) = \mathcal{P}_\Omega(\mathcal{M}).
\end{split}
\end{equation}
In the following, we need to introduce the tensor incoherence conditions between the underlying tensor $\mathcal{Z}$ and the column basis given in Definition \ref{defn}.
\begin{definition} Let $\mathcal{Z} \in \mathbb{C}^{n_{1}\times n_{2}\times n_{3}}$ with $\operatorname{rank}_{t}(\mathcal{Z})= \textup{\bf r}$, where $\textup{\bf r}=(r_1,\ldots, r_{n_3})$.
Assume its skinny
transformed tensor SVD is $\mathcal{Z} = \mathcal{U} \diamond_{\bf \Phi} \mathcal{S} \diamond_{\bf \Phi}
\mathcal{V}^{H}$. Then $\mathcal{Z}$ is said to satisfy the
tensor incoherence conditions,
if there exists a parameter $\mu\geq 1$, such that
\begin{align}
\max_{i=1, \ldots, n_1}\max_{k=1, \ldots, n_3} \|\mathcal{U}^{H} \diamond_{\bf \Phi} \tc{e}_{ik}\|_F \leq \sqrt{\frac{\mu \sum_{i=1}^{n_3} {r_{i}} }{n_1n_{3}}}, \label{eq16}\\
\max_{j=1, \ldots, n_2} \max_{k=1, \ldots, n_3}\|\mathcal{V}^{H} \diamond_{\bf \Phi} \tc{e}_{jk}\|_F\leq
\sqrt{\frac{\mu \sum_{i=1}^{n_3}{r_{i}}}{n_2n_{3}}}, \label{eq17}
\end{align}\label{def13}
where $\tc{e}_{ik}\in\mathbb{R}^{n_1\times 1\times n_3}$ and $\tc{e}_{jk}\in\mathbb{R}^{n_2\times 1\times n_3}$ are the column
basis.
\end{definition}
Denote by $T$ the linear space of tensors
\begin{equation}
T = \big\{\mathcal{U}\diamond_{\bf \Phi}\mathcal{Y}^{H} + \mathcal{W}\diamond_{\bf \Phi}
\mathcal{V}^{H} ~|~ \mathcal{Y} \in \mathbb{C}^{n_2 \times r \times
n_3},~\mathcal{W} \in \mathbb{C}^{n_1 \times r \times n_3} \big\},
\label{e1}
\end{equation}
and by $T^{\perp}$ its orthogonal complement,
where
$\mathcal{U}\in \mathbb{C}^{n_{1}\times r\times n_{3}}$ and
$\mathcal{V}\in \mathbb{C}^{n_{2}\times r\times n_{3}}$ are column unitary tensors,
respectively, i.e., $\mathcal{U}^H\diamond_{\bf \Phi}\mathcal{U}=\mathcal{I}_{\bf \Phi}, \mathcal{V}^H\diamond_{\bf \Phi}\mathcal{V}=\mathcal{I}_{\bf \Phi}$.
In the light of \cite[Proposition B.1]{zhang2019corrected}, for any $\mathcal{Z}\in\mathbb{C}^{n_1\times n_2\times n_3}$, the
orthogonal projections onto $T$ and its complementary are given as follows:
\[
\begin{split}
&\mathcal{P}_{T}(\mathcal{Z}) = \mathcal{U}\diamond_{\bf \Phi} \mathcal{U}^{H}\diamond_{\bf \Phi}
\mathcal{Z} +\mathcal{Z}\diamond_{\bf \Phi} \mathcal{V}\diamond_{\bf \Phi} \mathcal{V}^{H} -
\mathcal{U}\diamond_{\bf \Phi} \mathcal{U}^{H}\diamond_{\bf \Phi} \mathcal{Z}\diamond_{\bf \Phi}
\mathcal{V}\diamond_{\bf \Phi} \mathcal{V}^{H}, \\
&\mathcal{P}_{T^{\perp}}(\mathcal{Z}) = (\mathcal{I}_{\bf \Phi} - \mathcal{U}\diamond_{\bf \Phi}
\mathcal{U}^{H})\diamond_{\bf \Phi} \mathcal{Z}\diamond_{\bf \Phi} (\mathcal{I}_{\bf \Phi} - \mathcal{V}\diamond_{\bf \Phi}
\mathcal{V}^{H}).
\end{split}
\]
Denote $n_{(1)}=\max\{n_{1},n_{2}\},n_{(2)}=\min\{n_{1},n_{2}\}.$
We can improve the low bound on the number of
sampling sizes for tensor completion
by using transformed multi-rank instead of using tubal rank, which is stated in the following theorem.
\begin{theorem}\label{Theorem1}
Suppose that $\mathcal{Z} \in \mathbb{C}^{n_1 \times n_2 \times n_3}$ with $\operatorname{rank}_{t}(\mathcal{Z})= \textup{\bf r}$ and its skinny
transformed tensor SVD is $\mathcal{Z} = \mathcal{U} \diamond_{\bf \Phi} \mathcal{S} \diamond_{\bf \Phi}
\mathcal{V}^{H}$, where $\textup{\bf r}=(r_1,\ldots, r_{n_3})$, $\mathcal{U}\in \mathbb{C}^{n_{1}\times r\times n_{3}},$ $\mathcal{S}\in \mathbb{C}^{r\times r\times n_{3}}$ and
$\mathcal{V}\in \mathbb{C}^{n_{2}\times r\times n_{3}}$ with $\operatorname{rank}_{tt}(\mathcal{Z})=r$.
Suppose that $\mathcal{Z}$
satisfies the tensor incoherence conditions \eqref{eq16}-\eqref{eq17} and the observation set $\Omega$ with $|\Omega|=m$ is
uniformly distributed among all sets of cardinality, then there
exist universal constants $c_0, c_1, c_2 > 0$ such that if
\begin{align}\label{new1}
m\geq c_0 \mu\sum_{i=1}^{n_{3}}{r_{i}}n_{(1)}\log(n_{(1)}n_3),
\end{align}
$\mathcal{Z}$ is the unique minimizer to $\eqref{ObjFT}$ with
probability at least $1- c_1(n_{(1)}n_3)^{-c_2} $.
\end{theorem}
Next we compare the number of sample sizes requirement for exact recovery in \cite{Zhang2017}.
Note that the tensor incoherence conditions in \cite{Zhang2017} are given by
\begin{align*}
&\max_{i=1, \dots, n_1} \|\mathcal{U}^{H} \diamond_{\bf \Phi} \tc{e}_{i1}\|_F \leq \sqrt{\frac{\mu_{old} r}{n_1}},\\
& \max_{j=1, \dots, n_2} \|\mathcal{V}^{H} \diamond_{\bf \Phi}
\tc{e}_{j1}\|_F\leq \sqrt{\frac{\mu_{old} r}{n_2}},
\end{align*}
where $r$ is the tubal rank of the underlying tensor,
$\mu_{old}>0$ is a parameter, ${\bf \Phi}$ is FFT,
$\tc{e}_{i1}$ and $\tc{e}_{j1}$ are defined in Definition \ref{defn}.
When the number of samples is larger than or equal to $\widetilde{c}rn_{(1)}n_3\log(n_{(1)}n_3)$, the underlying tensor
can be recovered exactly \cite{Zhang2017}, where $\widetilde{c}>0$ is a given constant.
Neglecting the constants, we know that the bound of sample sizes
requirement in Theorem \ref{Theorem1} is smaller than that of \cite{Zhang2017}
since $\sum_{i=1}^{n_3}r_i$ is smaller than $rn_3$ in general.
Especially,
when the the vector of the transformed multi-rank ${\bf r}$ is sparse,
the number of sample sizes requirement given in \eqref{new1} is much smaller than that in \cite{Zhang2017} for exact recovery.
Moreover, the exact recovery theory in Theorem \ref{Theorem1}
not only holds for FFT but
also for any unitary transformation, which is very
meaningful in practical applications.
Based on Theorem \ref{Theorem1}, for a given data tensor,
we can choose a suitable unitary transformation such that
the sum of elements of
the transformed multi-rank of the underlying tensor is small \cite{song2019robust, zhang2021low},
which can guarantee to derive better results than that by using FFT directly.
To facilitate our proof of the main theorem, we will consider
the independent and identically distributed \textit{(i.i.d.)
Bernoulli-Rademacher model}. More precisely, we assume $\Omega =
\{(i, j, k) \ | \ \delta_{ijk} = 1\}$, where $\delta_{ijk}$ are
i.i.d. Bernoulli variables taking value one with probability $\rho = \frac{m}{n_1n_2n_3}$
and zero with probability $1-\rho$. Such a Bernoulli sampling is
denoted by $\Omega \sim \textup{Ber}(\rho)$ for short. As a proxy
for uniform sampling, the probability of failure under Bernoulli
sampling closely approximates the
probability of failure under uniform sampling.
Recall the definitions of tensor Frobenius norm and the tensor
incoherence conditions given in (\ref{eq16})-(\ref{eq17}), we can get the following result easily.
\begin{proposition} \label{pro1}
Let $\mathcal{Z} \in \mathbb{C}^{n_1 \times n_2 \times n_3}$ be an arbitrary
tensor with $\operatorname{rank}_{t}(\mathcal{Z})= \textup{\bf r}$, and $T$ be given as (\ref{e1}). Suppose that the
tensor incoherence conditions (\ref{eq16})-(\ref{eq17}) are
satisfied, then
$$\|\mathcal{P}_{T}( \tc{e}_{ik}\diamond_{\bf \Phi}
\ddot{\textbf{e}}_k \diamond_{\bf \Phi} \tc{e}_{jk}^H)\|^{2}_{\textrm{F}}\leq \frac{2\mu\sum_{i=1}^{n_3} {r_{i}}}{n_{(2)}n_{3}}.$$
\end{proposition}
Proposition \ref{pro1} plays an important role in the proofs of Lemmas
\ref{le1}, \ref{lemma2} and \ref{le3}, which can be found in the Appendix.
\begin{lemma} \label{le1}
Suppose that $\Omega \sim \textup{Ber}(\rho),$ where $\Omega$ with $|\Omega|=m$ is a set of indices sampled
independently and uniformly without replacement, $\rho=\frac{m}{n_{1}n_{2}n_{3}}$ and $T$ is given as
(\ref{e1}). Then with high probability,
\begin{align} \label{11}
\|\rho^{-1}\mathcal{P}_{T}\mathcal{P}_{\Omega}\mathcal{P}_{T}-\mathcal{P}_{T}\|_\textup{op}\leq
\epsilon,
\end{align}
provided that $m\geq C_0\epsilon^{-2} \mu \sum_{i=1}^{n_3}{r_{i}}n_{(1)}
\log(n_{(1)}n_3) $ for
some numerical constant $C_0
> 0$.
\end{lemma}
\begin{lemma}\label{lemma2}
Suppose that $\mathcal{Z} \in \mathbb{C}^{n_1 \times n_2 \times n_3}$ is a
tensor with $\operatorname{rank}_{t}(\mathcal{Z})= \textup{\bf r}$, $\Omega, \rho$ and $m$ are defined in Lemma \ref{le1}.
Then for all $c>1$ and $C_{0}>0$,
\begin{equation*}
\|(\rho^{-1}\mathcal{P}_{\Omega}-\mathcal{I})\mathcal{Z}\| \leq
c\left(\frac{\log (n_{(1)}n_{3})}{\rho}\|\mathcal{Z}\|_{\infty}+\sqrt{\frac{\log (n_{(1)}n_{3})}{\rho}}\|\mathcal{Z}\|_{\infty,w}\right)
\label{eq25}
\end{equation*}
holds with high probability
provided that
$m\geq C_0 \epsilon^{-2} \mu \sum_{i=1}^{n_3}{r_{i}}n_{(1)}\log(n_{(1)}n_3),$
where $\mathcal{I}$ denotes the identity operator.
\end{lemma}
\begin{lemma}\label{le3}
Suppose that $\mathcal{Z} \in \mathbb{C}^{n_1 \times n_2 \times n_3}$ is a
tensor with $\operatorname{rank}_{t}(\mathcal{Z})= \textup{\bf r}$, $\Omega, \rho$ and $m$ are defined in Lemma \ref{le1}. Then for some sufficiently large $C_{0}$,
\begin{equation}
\|(\rho^{-1}\mathcal{P}_{T}\mathcal{P}_{\Omega}-\mathcal{P}_{T})\mathcal{Z}\|_{\infty,w}\leq\frac{1}{2}\sqrt{\frac{n_{(1)}n_{3}}{\mu\sum_{i=1}^{n_{3}}r_{i}}}\|\mathcal{Z}\|_{\infty}
+\frac{1}{2}\|\mathcal{Z}\|_{\infty,w}
\end{equation}
holds with high probability provided that $m\geq C_0 \epsilon^{-2} \mu \sum_{i=1}^{n_3}{r_{i}}n_{(1)}\log(n_{(1)}n_3).$
\end{lemma}
The proofs of the three lemmas are left to the appendix.
Lemma \ref{lemma2} establishes an upper bound of the tensor
spectral norm of $(\rho^{-1}\mathcal{P}_{\Omega}-\mathcal{I})\mathcal{Z}$
in terms of $\|\mathcal{Z}\|_{\infty}$ and $\|\mathcal{Z}\|_{\infty,w},$
which is tighter than the existing result given by $\|\mathcal{Z}\|_{\infty}$
and $\|\mathcal{Z}\|_{\infty,2}$. The weights in $\|\mathcal{Z}\|_{\infty,w}$
are determined by the unitary transformation, which
guarantee the upper bound of $\|(\rho^{-1}\mathcal{P}_{\Omega}-\mathcal{I})\mathcal{Z}\|$ to be smaller.
Lemma \ref{le3} shows that the $l_{{\infty,w}}$ norm
of $(\rho^{-1}\mathcal{P}_{T}\mathcal{P}_{\Omega}-\mathcal{P}_{T})\mathcal{Z}$ can be
dominated by $\|\mathcal{Z}\|_{\infty,w}$ and $\|\mathcal{Z}\|_{\infty}$.
Moreover, Lemmas C1 and C2 in \cite{Zhang2017} can be seen as special cases of Lemmas \ref{lemma2} and \ref{le3}, respectively, if the transformation is chosen as FFT.
\begin{lemma}\cite[Lemma 4.1]{Zhang2017}\label{4.7}
Suppose that $\|\rho^{-1}{\cal P}_T{\cal P}_{\Omega}{\cal P}_T - {\cal P}_T\|_\textup{op}\leq \frac{1}{2}.$ Then for any $\mathcal{Z}\in\mathbb{C}^{n_1\times n_2\times n_3 }$ such that $\mathcal{P}_{\Omega}(\mathcal{Z})=0,$ the following inequality
\begin{equation*}
\frac{1}{2}\|\mathcal{P}_{T^{\bot}}(\mathcal{Z})\|_{\emph{TTNN}}>\frac{1}{4n_{(1)}n_{3}}\|\mathcal{P}_{T}(\mathcal{Z})\|_{F}
\end{equation*}
holds with high probability.
\end{lemma}
With the tools in hand we can list the proof of Theorem \ref{Theorem1} in detail.
\vspace{2mm}
\noindent
{\bf Proof of Theorem \ref{Theorem1}}.
The high level road map of the proof is a standard one just as shown in \cite{candes2011robust}: by convex analysis, to show $\mathcal{Z}$ is the unique optimal solution to the problem $\eqref{ObjFT}$, it is sufficient to find a dual certificate $\mathcal{Y}$ satisfying several subgradient type conditions. In our case, we need to find a tensor $\mathcal{Y}=\mathcal{P}_{\Omega}(\mathcal{Y})$
such that
\begin{align}
& \|\mathcal{P}_{T}(\mathcal{Y})-\mathcal{U}\diamond_{\bf \Phi}\mathcal{V}^{H}\|_{F}\leq \frac{1}{4n_{(1)}n_{3}^{2}}, \label{cd1}\\
& \|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|\leq\frac{1}{2}.\label{cd2}
\end{align}
It follows from \eqref{cd2} that we need to estimate the
tensor spectral norm $\|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|$.
Similar to matrix cases \cite{candes2011robust}, we can use the tensor infinite norm to establish the upper bound of $\|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|$. However, if $\|\mathcal{Y}\|_{\infty}$ is applied, then it will ultimately link to
$\|\mathcal{U}\diamond_{\bf \Phi}\mathcal{V}^H\|_{\infty}$
and lead to the joint incoherence condition.
In order to avoid applying the joint incoherence condition,
$\|\mathcal{Y}\|_{\infty,2}$ is used in \cite{chen2015incoherence} and \cite{Zhang2017}
to compute the upper bound of $\|\mathcal{Y}\|$ for the matrix and tensor cases, respectively.
It follows from \cite{chen2015incoherence} and \cite{Zhang2017}
that a lower upper bound can be derived by using $\|\mathcal{Y}\|_{\infty,2}$.
However, the upper bound derived by $\|\mathcal{Y}\|_{\infty,2}$
can be relaxed further for an arbitrary unitary transformation.
Here, we derive a new upper bound of $\|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|$
by using the $l_{\infty,w}$ norm $\|\mathcal{Y}\|_{\infty,w}$
as defined in \eqref{norm1}. Note that $\|\mathcal{Y}\|_{\infty,w}$
is not larger than $\|\mathcal{Y}\|_{\infty,2}$
for any tensor $\mathcal{Y},$ which leads to a tighter upper bound of $\|\mathcal{Y}\|$.
We now return to the proof Theorem \ref{Theorem1} in detail.
We apply the Golfing Scheme method introduced by Gross
\cite{gross2010} and modified by Cand\`{e}s et al.
\cite{candes2011robust} to construct a dual tensor $\mathcal{Y}$
supported by $\Omega^c$ iteratively. Similar to the proof of
\cite[Theorem 3.1]{Zhang2017}, we consider the set $\Omega^c\sim
\textup{Ber}(1-\rho)$ as a union of sets of support $\Omega_j$, i.e.,
$\Omega^c = \bigcup_{j=1}^p \Omega_j$, where $\Omega_j \sim
\textup{Ber}(q),$
which implies $q \geq C_0\rho/\log(n_{(1)}n_3)$.
Hence we have
$ \rho = (1 - q)^{p},$
where $p = \lfloor 5\log(n_{(1)}n_3) + 1\rfloor$. Denote
\begin{equation}\label{s1}
\mathcal{Y}= \sum_{j=1}^p
\frac{1}{q}\mathcal{P}_{\Omega_j}(\mathcal{Z}_{j-1}), ~\text{with}~\mathcal{Z}_{j} = \Big(\mathcal{P}_{T} - \frac{1}{q}\mathcal{P}_{T}
\mathcal{P}_{\Omega_j}\mathcal{P}_{T}\Big)\mathcal{Z}_{j-1},~ \mathcal{Z}_0 =
\mathcal{P}_{T}(\mathcal{U} \diamond_{\bf \Phi} \mathcal{V}^{H} ).
\end{equation}
In the following we will show that $\mathcal{Y}$ defined in \eqref{s1} satisfies the conditions \eqref{cd1} and \eqref{cd2}.
For \eqref{cd1}. Set $\mathcal{D}_{k}:=\mathcal{U}\diamond_{\bf \Phi}\mathcal{V}^{H}-\mathcal{P}_{T}(\mathcal{Z}_{k})$
for $k=0,\ldots,p$. By the definition
of $\mathcal{Z}_{k}$, we have $\mathcal{D}_{0}=\mathcal{U}\diamond_{\bf \Phi}\mathcal{V}^{H}$ and
\begin{equation}\label{eqn1}
\mathcal{D}_{k}=(\mathcal{P}_{T}-\mathcal{P}_{T}\mathcal{P}_{\Omega}\mathcal{P}_{T})\mathcal{D}_{k-1}, \ k=1,\ldots, p.
\end{equation}
Note that
$\Omega_{k}$ is independent of $\mathcal{D}_{k-1}$ and $q\geq c_{0}\mu\sum r_{i}\log(n_{(1)}n_{3})/(n_{(1)}n_{3}).$
For each $k,$ replacing $\Omega$ by $\Omega_{k},$ then by Lemma \ref{le1}, we have
\begin{equation*}
\|\mathcal{D}_{k}\|_{{F}}\leq \|\mathcal{P}_{T}-\mathcal{P}_{T}\mathcal{P}_{\Omega_{k}}\mathcal{P}_{T}\|\|\mathcal{D}_{k-1}\|_{{F}}\leq \frac{1}{2}\|\mathcal{D}_{k-1}\|_{F}.
\end{equation*}
As a consequence, one can obtain that
\begin{align*}
\|\mathcal{P}_{T}(\mathcal{Y})-\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{{F}}=\|\mathcal{D}_{p}\|_{{F}}\leq \left(\frac{1}{2}\right) ^{p}\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{{F}}\leq\frac{1}{4(n_{(1)}n_{3})^{2}}\sqrt{r}
\leq \frac{1}{4n_{(1)}n_{3}^2}.
\end{align*}
For \eqref{cd2}. Note that $\mathcal{Y}=\sum_{k=1}^{p}\mathcal{P}_{\Omega_{k}}\mathcal{P}_{T}(\mathcal{D}_{k-1}),$ thus
\begin{align}\label{17py}
\|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|\leq \sum_{k=1}^{p}\|\mathcal{P}_{T^{\bot}}(\mathcal{P}_{\Omega_{k}}\mathcal{P}_{T}-\mathcal{P}_{T})(\mathcal{D}_{k-1})\|
\leq \sum_{k=1}^{p}\|(\mathcal{P}_{\Omega_{k}}-\mathcal{I})\mathcal{P}_{T}(\mathcal{D}_{k-1})\|.
\end{align}
Applying Lemma \ref{lemma2} with $\Omega$ replaced by $\Omega_{k}$ to each summand of (\ref{17py}) yields
\begin{align}
\|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|&\leq c \sum_{k=1}^{p}\Biggl(\frac{\log (n_{(1)}n_{3})}{q}\|\mathcal{D}_{k-1}\|_{\infty}+\sqrt{\frac{\log (n_{(1)}n_{3})}{q}}\|\mathcal{D}_{k-1}\|_{\infty,w}\Biggl)\nonumber\\
&\leq \frac{c}{\sqrt{c_{0}}} \sum_{k=1}^{p}\Biggl(\frac{ n_{(1)}n_{3}}{\mu\sum r_{i}}\|\mathcal{D}_{k-1}\|_{\infty}+\sqrt{\frac{ n_{(1)}n_{3}}{\mu\sum r_{i}}}\|\mathcal{D}_{k-1}\|_{\infty,w}\Biggl).\label{eq4}
\end{align}
Using \eqref{eqn1}, and applying Lemma \ref{le1} with $\Omega$ replaced by $\Omega_{k},$ we can get
\begin{align*}
\|\mathcal{D}_{k-1}\|_{\infty}=\|(\mathcal{P}_{T^{\bot}}(\mathcal{P}_{\Omega_{k-1}}\mathcal{P}_{T}-\mathcal{P}_{T})\cdots \mathcal{P}_{T^{\bot}}(\mathcal{P}_{\Omega_{1}}\mathcal{P}_{T}-\mathcal{P}_{T}))\mathcal{D}_{0}\|_{\infty}\leq
\frac{1}{2^{k}}\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty}.
\end{align*}
It follows from Lemma \ref{le3} that
\begin{align*}
\|\mathcal{D}_{k-1}\|_{\infty,w}=\|(\mathcal{P}_{T}-\mathcal{P}_{T}\mathcal{P}_{\Omega_{k-1}}\mathcal{P}_{T})\mathcal{D}_{k-2}\|_{\infty,w}
\leq\frac{1}{2}\sqrt{\frac{n_{(1)}n_{3}}{\mu\sum r_{i}}}\|\mathcal{D}_{k-2}\|_{\infty}+\frac{1}{2}\|D_{k-2}\|_{\infty,w}
\end{align*}
holds with high probability. Moreover, it follows from \eqref{eqn1} that
\begin{align*}
\|\mathcal{D}_{k-1}\|_{\infty,w}\leq\frac{ k}{2^{k-1}}\sqrt{\frac{n_{(1)}n_{3}}{\mu\sum r_{i}}}\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty}+\frac{1}{2^{k-1}}\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty,w}.
\end{align*}
Taking them back to \eqref{eq4} yields
\begin{align}
&~\|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|\nonumber\\
\leq &~\frac{c}{\sqrt{c_{0}}}\frac{n_{(1)}n_{3}}{\mu\sum r_{i}}\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty}\sum_{k=1}^{p}(k+1)\left(\frac{1}{2}\right)^{k-1} \\
&~~ +\frac{c}{\sqrt{c_{0}}}\sqrt{\frac{n_{(1)}n_{3}}{\mu\sum r_{i}}}\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty,w}\sum_{k=1}^{p}\left(\frac{1}{2}\right)^{k-1}\nonumber\\
\leq &~\frac{6c}{\sqrt{c_{0}}}\frac{n_{(1)}n_{3}}{\mu\sum r_{i}}\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty}+ \frac{2c}{\sqrt{c_{0}}}\sqrt{\frac{n_{(1)}n_{3}}{\mu\sum r_{i}}}\|\mathcal{U}\diamond_{\bf \Phi}\mathcal{V}^{H}\|_{\infty,w}.\label{e7}
\end{align}
By the incoherence conditions given in \eqref{eq16}-\eqref{eq17}, we can get
\begin{align}
&\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty}\leq \max_{i,j,k}\|\mathcal{U}\diamond_{\bf \Phi}\tc{e}_{ik}\|_{F}\|\mathcal{V}^{H}\diamond_{\bf \Phi}\tc{e}_{ik}\|_{{F}}\leq \frac{\mu\sum r_{i}}{n_{(1)}n_{3}},\label{e10}\\
&\|\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{\infty,w}\leq \max\left\{\max_{i,k}\| \mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\diamond_{\bf \Phi}\tc{e}_{ik}\|_{{F}},\max_{j,k}\| \tc{e}_{jk}^{H}\diamond_{\bf \Phi}\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{{F}}\right\}\leq \sqrt{\frac{\mu\sum r_{i}}{n_{(1)}n_{3}}}.\label{e11}
\end{align}
Plugging \eqref{e10} and \eqref{e11} into \eqref{e7}, we obtain that
\begin{align*}
\|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|\leq\frac{6c}{\sqrt{c_{0}}}+\frac{2c}{\sqrt{c_{0}}}\leq \frac{1}{2}
\end{align*}
provided $c_{0}$ is sufficiently large.
Moreover, for any tensor $\mathcal{W}\in \{\mathcal{W} \in \mathbb{C}^{n_{1}\times
n_{1}\times n_{3}}|\mathcal{P}_{\Omega}(\mathcal{W})=0\},$ denote the skinny
transformed tensor SVD of $\mathcal{P}_{T^\bot}(\mathcal{W})$ by
$$
\mathcal{P}_{T^\bot}(\mathcal{W}) =
\mathcal{U}_\perp \diamond_{\bf \Phi}\mathcal{S}_\perp \diamond_{\bf \Phi}
\mathcal{V}^{H}_\perp.
$$
Since
$\mathcal{\overline{U}}_{\bf \Phi}^{H}\cdot(\mathcal{\overline{U}}_{\perp})_{\bf \Phi}=\mathbf{0}$ and
$\mathcal{\overline{V}}_{\bf \Phi}^{H}\cdot(\mathcal{\overline{V}}_{\perp})_{\bf \Phi}=\mathbf{0},$ we
have
$$
\|\mathcal{U} \diamond_{\bf \Phi} \mathcal{V}^{H} + \mathcal{U}_\perp
\diamond_{\bf \Phi} \mathcal{V}^{H}_\perp\| = \|\mathcal{\overline{U}}_{\bf \Phi}\cdot
\mathcal{\overline{V}}_{\bf \Phi}^{H} + (\mathcal{\overline{U}}_\perp)_{\bf \Phi}\cdot
((\mathcal{\overline{V}}_\perp)_{\bf \Phi})^{H}\|=1.
$$
Thus, we get that
\begin{align}
\|\mathcal{Z} + \mathcal{W}\|_{\textup{TTNN}}
\geq & ~ \langle \mathcal{U}
\diamond_{\bf \Phi} \mathcal{V}^{H}
+ \mathcal{U}_\perp \diamond_{\bf \Phi} \mathcal{V}^{H}, \mathcal{Z} + \mathcal{W}\rangle \nonumber \\
= &~ \langle\mathcal{U} \diamond_{\bf \Phi}
\mathcal{V}^{H},\mathcal{Z}\rangle + \langle\mathcal{U}_\perp
\diamond_{\bf \Phi} \mathcal{V}^{H}_\perp,
\mathcal{P}_{T^{\bot}}(\mathcal{W})\rangle +
\langle\mathcal{U} \diamond_{\bf \Phi} \mathcal{V}^{H}, \mathcal{W}\rangle \nonumber\\
= &~ \|\mathcal{Z}\|_{\textup{TTNN}} +
\|\mathcal{P}_{T^{\bot}}(\mathcal{W})\|_{\textup{TTNN}} +\langle \mathcal{U}
\diamond_{\bf \Phi} \mathcal{V}^{H}, \mathcal{W}\rangle \nonumber \\
\geq &~
\|\mathcal{Z}\|_{\textup{TTNN}}+\mathcal{P}_{T^{\bot}}(\mathcal{W})\|_{\textup{TTNN}}
-|\langle\mathcal{Y}-\mathcal{U}\diamond_{\bf \Phi} \mathcal{V}^{H}, \mathcal{W}\rangle
- \langle\mathcal{Y}, \mathcal{W}\rangle| \nonumber\\
\geq & ~ \|\mathcal{Z}\|_{\textup{TTNN}}+
\|\mathcal{P}_{T^{\bot}}(\mathcal{W})\|_{\textup{TTNN}}- \|\mathcal{P}_{T^{\bot}}(\mathcal{Y})\|
\|\mathcal{P}_{T^{\bot}}(\mathcal{W})\|_{\textup{TTNN}}
\nonumber\\
&~-\|\mathcal{P}_{T}(\mathcal{Y}) - \mathcal{U}
\diamond_{\bf \Phi} \mathcal{V}^{H}\|_{{F}} \|\mathcal{P}_{T}(\mathcal{W})\|_{{F}} \nonumber\\
\geq &~ \|\mathcal{Z}\|_{\textup{TTNN}}+
\frac{1}{2}\|\mathcal{P}_{T^{\bot}}(\mathcal{W})\|_{\textup{TTNN}} -
\frac{1}{4n_{(1)}n_3} \|\mathcal{P}_T(\mathcal{W})\|_{{F}}. \label{eq21}
\end{align}
Thus, it follows from Lemma \ref{4.7} that $ \|\mathcal{Z} + \mathcal{W}\|_{\textup{TTNN}}>\|\mathcal{Z}\|_{\textup{TTNN}}$ holds
for any $\mathcal{W}$ with $\mathcal{P}_{\Omega}(\mathcal{W}) = 0.$
As a consequence, $\mathcal{Z}$ is the unique minimizer to $\eqref{ObjFT}.$
This completes the proof. \qed
In the next section, we demonstrate that the theoretical results can be obtained under
valid incoherence conditions and the tensor completion performance of the proposed
method is better than that of other testing methods.
\section{Experimental Results}\label{Sect4}
In this section, numerical examples are presented to demonstrate the effectiveness of the proposed model.
All numerical experiments are obtained from a desktop computer running on 64-bit Windows
Operating System having 8 cores with Intel(R) Core(TM) i7-6700 CPU at 3.40GHz and 20 GB memory.
Firstly, we employ an alternating direction method of multipliers
(ADMM) \cite{fazel2013hankel, Glowinski1976Approximations} to solve problem (\ref{ObjFT}).
Let $\mathcal{Z}=\mathcal{Y}$. Then problem (\ref{ObjFT}) can be rewritten as
\begin{equation}\label{ObjLg}
\begin{split}
\min_{\mathcal{Z}} & \
\|\mathcal{Z}\|_{\textup{TTNN}} \\
\textup{s.t.} & \ \mathcal{Z}=\mathcal{Y}, \
\mathcal{P}_\Omega(\mathcal{Y}) = \mathcal{P}_\Omega(\mathcal{M}).
\end{split}
\end{equation}
The augmented Lagrangian function associated with (\ref{ObjLg}) is defined as
$$
L(\mathcal{Z},\mathcal{Y},\mathcal{X}):=\|\mathcal{Z}\|_{\textup{TTNN}}
-\langle \mathcal{X},\mathcal{Z}-\mathcal{Y}\rangle+\frac{\beta}{2}\|\mathcal{Z}-\mathcal{Y}\|_F^2,
$$
where $\mathcal{X}\in \mathbb{C}^{n_{1}\times n_{2}\times n_3}$ is the Lagrangian multiplier and $\beta>0$ is the penalty parameter.
The ADMM iteration system is given as follows:
\begin{eqnarray}\label{u1211}
&&\mathcal{Z}^{k+1}=\arg\min_{\mathcal{Z}}\Big\{L(\mathcal{Z},\mathcal{Y}^k,\mathcal{X}^k)\Big\},\\\label{w1}
&&\mathcal{Y}^{k+1}=\arg\min_{\mathcal{Y}}\Big\{L(\mathcal{Z}^{k+1},\mathcal{Y},\mathcal{X}^k): \mathcal{P}_\Omega(\mathcal{Y}) = \mathcal{P}_\Omega(\mathcal{M})\Big\},\\ \label{u1}
&&\mathcal{X}^{k+1}=\mathcal{X}^k-\gamma\beta\left(\mathcal{Z}^{k+1}-\mathcal{Y}^{k+1}\right),
\end{eqnarray}
where $\gamma\in(0,\frac{1+\sqrt{5}}{2})$ is the dual steplength.
It follows from \cite[Theorem 3]{song2019robust} that
the optimal solution with respect to $\mathcal{Z}$ in (\ref{u1211}) is given by
\begin{equation}\label{XS}
\mathcal{Z}^{k+1}=\mathcal{U} \diamond_{\bf \Phi} \mathcal{S}_{\beta}
\diamond_{\bf \Phi} \mathcal{V}^{H},
\end{equation}
where $\mathcal{Y}^k+\frac{1}{\beta}\mathcal{X}^k=\mathcal{U} \diamond_{\bf \Phi} \mathcal{S}
\diamond_{\bf \Phi} \mathcal{V}^{H}$,
$\mathcal{S}_{\beta}=\mathbf{\Phi}^{H}[\hat{\mathcal{S}}_{\beta}]$,
and $\hat{\mathcal{S}}_{\beta}=\max\{\hat{\mathcal{S}}_{\bf \Phi}-\frac{1}{\beta},0\}$.
The optimal solution with respect to $\mathcal{Y}$ in (\ref{w1}) is given by
\begin{equation}\label{YS}
\mathcal{Y}^{k+1}=\mathcal{P}_{\overline{\Omega}}\left(\mathcal{Z}^{k+1}-\frac{1}{\beta}\mathcal{X}^k\right)+\mathcal{P}_\Omega(\mathcal{M}),
\end{equation}
where $\overline{\Omega}$ denotes the complementary set of $\Omega$
on $\{1,\ldots,n_1\}\times\{1,\ldots,n_2\}\times\{1,\ldots,n_3\}$.
The detailed description of ADMM for solving (\ref{ObjLg}) is given in Algorithm \ref{alg1}.
\begin{algorithm}[h]
\caption{Alternating direction method of multipliers for solving
(\ref{ObjLg})} \label{alg1}
\textbf{Step 0.} Let $\tau\in(0,(1+\sqrt{5})/2), \beta>0$ be given constants.
Given $\mathcal{Y}^0, \mathcal{X}^0$. For $k=0,1,2,\ldots,$ perform the following steps: \\
\textbf{Step 1.} Compute $\mathcal{Z}^{k+1}$ by (\ref{XS}). \\
\textbf{Step 2.} Compute $\mathcal{Y}^{k+1}$ via (\ref{YS}). \\
\textbf{Step 3.} Compute $\mathcal{X}^{k+1}$ by (\ref{u1}).
\end{algorithm}
The convergence of a two-block ADMM for solving convex optimization problems has been established in \cite[Theorem B.1]{fazel2013hankel}
and the convergence of Algorithm \ref{alg1} can be derived from this theorem easily.
We omit the details here for the sake of brevity.
The Karush-Kuhn-Tucker (KKT) conditions associated with problem (\ref{ObjLg}) are given as follows:
\begin{equation}\label{KKT}
\left\{\begin{array}{lll}
0\in\partial \|\mathcal{Z}\|_{\textup{TTNN}}-\mathcal{X}, \\
\mathcal{Z} =\mathcal{Y}, \
\mathcal{P}_\Omega(\mathcal{Y}) = \mathcal{P}_\Omega(\mathcal{M}),
\end{array}\right.
\end{equation}
where $\partial \|\mathcal{Z}\|_{\textup{TTNN}}$ denotes the subdifferential of TTNN at $\mathcal{Z}$.
Based on the KKT conditions in (\ref{KKT}), we adopt the following relative residual to
measure the accuracy of a computed solution in the numerical experiments:
$$
\eta:=\max\{\eta_x,\eta_y\},
$$
where
$$
\eta_x=\frac{\|\mathcal{Z}-\mbox{Prox}_{\|\cdot\|_{\textup{TTNN}}}(\mathcal{X}+\mathcal{Z})\|_F}{1+\|\mathcal{Z}\|_F+\|\mathcal{X}\|_F}, \ \
\eta_y=\frac{\|\mathcal{Z}-\mathcal{Y}\|_F}{1+\|\mathcal{Z}\|_F+\|\mathcal{Y}\|_F}.
$$
Here $\mbox{Prox}_f(y):=\arg\min_x\{f(x)+\frac{1}{2}\|x-y\|^2\}$.
In the practical implementation,
Algorithm \ref{alg1} will be terminated
if $\eta\leq 10^{-3}$ or the maximum number of iterations exceeds $600$.
We set $\gamma=1.618$ for the convergence of ADMM \cite{fazel2013hankel} in all experiments.
Since the penalty parameter $\beta$ is not too sensitive to the recovered results,
we set $\beta=0.05$ in the following experiments.
The relative error (Rel) is defined by
$$
\mbox{Rel}:=\frac{\|\mathcal{Z}_{est}-\mathcal{Z}\|_F}{\|\mathcal{Z}\|_F},
$$
where $\mathcal{Z}_{est}$ is the estimated tensor and $\mathcal{Z}$ is the ground-truth tensor.
To evaluate the performance of the proposed method for real-world tensors,
the peak signal-to-noise ratio (PSNR) is used to
measure the quality of the estimated tensor, which is defined as follows:
$$
\mbox{PSNR}:=10\log_{10}
\frac{n_1n_2n_3({\mathcal{Z}}_{\max}-{\mathcal{Z}}_{\min})^2}{\|\mathcal{Z}_{est}-{\mathcal{Z}}\|_F^2},
$$
where ${\mathcal{Z}}_{\max}$ and ${\mathcal{Z}}_{\min}$ denote
the maximum and minimum entries of $\mathcal{Z}$, respectively.
The structural similarity (SSIM) index \cite{wang2004image} is used to
measure the quality of the recovered images:
$$
\mbox{SSIM}:=\frac{(2\mu_x\mu_y+c_1)(2\sigma_{xy}+c_2)}{(\mu_x^2+\mu_y^2+c_1)(\sigma_x^2+\sigma_y^2+c_2)},
$$
where $\mu_x, \sigma_x$ are the mean intensities and standard deviation of the original image, respectively,
$\mu_y, \sigma_y$ denote the mean intensities and standard deviation of the recovered images, respectively,
$\sigma_{xy}$ denotes the covariance of the original and recovered images,
and $c_1,c_2>0$ are constants.
For the real-world tensor data, the SSIM is used to denote the average SSIM values of all images.
\subsection{Transformations of tensor SVD}\label{trans3}
In this subsection, we use three kinds of transformations in the ${\bf \Phi}$-product and transformed tensor SVD.
The first two transformations are FFT (t-SVD (FFT)) and
discrete cosine transform (t-SVD (DCT)).
The third one is based on given data to construct a
unitary transform matrix \cite{qiu2021nonlocal, song2019robust, zhang2021low}.
Note that we unfold $\mathcal{Z}$ into a matrix $Z$ along the third-dimension (called
t-SVD (data)) and take the SVD of the unfolding matrix $Z=U {\Sigma}{V}^H$.
Suppose that $\text{rank}(Z)=r$.
It is interesting to observe that $U^H$ is the optimal transformation to obtain a low rank approximation of $Z$:
$$
\min_{\mathbf{\Phi},B}\ \|\mathbf{\Phi} Z-B\|_F^2 \quad \mbox{s.t.} \quad \text{rank}(B)=r, \ \mathbf{\Phi}^H\mathbf{\Phi} = \mathbf{\Phi}\mathbf{\Phi}^H=I.
$$
It has been demonstrated that the chosen unitary transformation $U^H$ is very effective
for the tensor completion problems in the literature, e.g., \cite{qiu2021nonlocal, song2019robust, zhang2021low}.
In practice, the estimator of $\mathcal{Z}$
obtained by t-SVD (DCT) can be used to generate $\mathbf{\Phi}$ for tensor completion.
Now we give the computational cost of TTNN based on the three transformations for any $n_1 \times n_2 \times n_3$ tensor,
which is the main cost of Algorithm \ref{alg1}. Suppose that $n_2 \leq n_1$.
The computational cost of TTNN is given as follows:
\begin{itemize}
\item The application of FFT or DCT to a tube ($n_3$-vector) is of $O(n_3 \log(n_3))$ operations.
There are $n_1n_2$ tubes in an $n_1 \times n_2 \times n_3$ tensor. In the transformed
tensor SVD based on FFT or DCT, we need to compute $n_3$ $n_1$-by-$n_2$ SVDs in the transformed domain
and then the cost is $O(n_1n_2^2n_3)$ for these matrices.
Hence, the total cost of TTNN based on FFT or DCT is of
$O(n_1n_2n_3 \log(n_3)+n_1n_2^2n_3)$ operations.
\item The application
of a unitary transformation ($n_3$-by-$n_3$) to an $n_3$-vector is of $O(n_3^2)$ operations.
And there are still $n_3$ $n_1$-by-$n_2$ SVDs to be calculated in the transformed domain.
Therefore, the total cost of computing the TTNN based on the given data
is of $O(n_1n_2n_3^2+n_1n_2^2n_3)$ operations.
\end{itemize}
\subsection{Recovery Results}
In this subsection, we show the recovery results to demonstrate
the performance of our analysis for synthetic data and real imaging data sets.
\subsubsection{Synthetic Data}
For the synthetic data,
the random tensors are generated as follows:
$\mathcal{Z}=\mathcal{A} \diamond_{\bf \Phi} \mathcal{B}\in\mathbb{C}^{n_1\times n_2\times n_3}$
with different transformed multi-rank $\mathbf{r}$,
where $\hat{\mathcal{A}}_{{\bf\Phi}}^{(i)}$ and $\hat{\mathcal{B}}_{{\bf\Phi}}^{(i)}$
are generated by MATLAB commands $\mbox{randn}(n_1,r_i)$ and $\mbox{randn}(n_2,r_i)$,
and $r_i$ is the $i$-th element of the transformed multi-rank $\mathbf{r}$.
Here ${\bf \Phi}$ denotes FFT, DCT, and an orthogonal matrix generated
by the SVD of the unfolding matrix of ${\cal Z}$ along the third-dimension, see
Section \ref{trans3}.
In Figures \ref{fixsum200} and \ref{fixsum300}, we show the actual number of sample
sizes for exact recovery and the theoretical bounds of sample sizes requirements in Theorem \ref{Theorem1}:
\begin{equation} \label{con1}
m \ge \ {\rm constant}_1 \ \sum_{i=1}^{n}r_i n \log(n^2)
\end{equation}
and
the results in \cite{Zhang2017}:
\begin{equation} \label{con2}
m \ge \ {\rm constant}_2 \ r n^2 \log(n^2)
\end{equation}
for the $n\times n\times n$
tensor with the fixed sum of the transformed multi-rank and fixed transformed
tubal rank by using
t-SVD (FFT), t-SVD (DCT), and t-SVD (data).
In the randomly generated tensor, we set
(i) $\max_{1 \le i \le n} r_i =10$
(the transformed tubal rank is 10)
and $\sum_{i=1}^nr_i=200$; and
(ii) $\max_{1 \le i \le n} r_i =20$
(the transformed tubal rank is 20)
and $\sum_{i=1}^nr_i=300$.
\begin{figure}[!t]
\centering
\subfigure[t-SVD (FFT)]{
\begin{minipage}[b]{0.99\textwidth}
\centerline{\scriptsize } \vspace{1.5pt}
\includegraphics[width=5.5in,height=1.8in]{Fig1.png}
\end{minipage}
}
\caption{\small The number of samples of exact recovery, the sample sizes required by the right hand sides of (\ref{con1}) and (\ref{con2})
for different values of $n$ with $\sum_{i=1}^{n}r_i=200$.}
\label{fixsum200}
\vspace{2mm}
\centering
\subfigure{
\begin{minipage}[b]{0.99\textwidth}
\centerline{\scriptsize } \vspace{1.5pt}
\includegraphics[width=5.5in,height=1.8in]{Fig2.png}
\end{minipage}
}
\caption{\small The number of samples of exact recovery, the sample sizes required by the right hand sides of (\ref{con1}) and (\ref{con2})
for different values of $n$ with $\sum_{i=1}^{n}r_i=300$.}\label{fixsum300}
\end{figure}
In Figures \ref{fixsum200} and \ref{fixsum300}, we test different values of $n$
(from 60 to 200 with the increment size being 20).
The exact recovery means that five trials are tested
and all of the relative errors are less than or equal to $10^{-2}$ in the experiments.
We test constant$_1$, constant$_2$ = 0.5 or 1 in the right hand sides of
(\ref{con1}) and (\ref{con2}) respectively to check
how the theoretical bounds of sample sizes
match with the actual number of samples for different values
of $n$.
It can be seen from Figures \ref{fixsum200} and \ref{fixsum300} that
the curve $\sum_{i=1}^{n}r_in\log(n^2)$ based on the proposed
bound with constant$_1=1$ is close to the curve for the number of samples
required by using t-SVD (FFT), and the curves
$0.5 \sum_{i=1}^{n}r_i n\log(n^2)$ based on the proposed bound
with constant$_1=0.5$ is
close to the curves for the number of samples required
by using t-SVD (DCT) and t-SVD (data).
Note that the corresponding slope of these lines in Figures \ref{fixsum200} and \ref{fixsum300} is equal to 1 derived by the $n$ term.
In contrast, the curves
constant$_2 r n^2 \log(n^2)$ based on
the results in \cite{Zhang2017} do not fit
the curves for the number of samples required
by using t-SVD (FFT), t-SVD (DCT) and t-SVD (data), see
the curves with constant$_2 = 0.5, 1$ in Figures \ref{fixsum200} and \ref{fixsum300}.
The main reason is that the corresponding slope of the lines
in Figures \ref{fixsum200} and \ref{fixsum300} is equal to 2 derived by the $n^2$ term.
According to Figures \ref{fixsum200} and \ref{fixsum300}, we find that
the theoretical bounds of sample sizes requirements in Theorem \ref{Theorem1} match with
the actual number of sample
sizes for recovery.
\subsubsection{Hyperspectral Images}\label{HPDS}
\begin{figure}[!t]
\centering
\subfigure{
\begin{minipage}[b]{0.99\textwidth}
\centerline{\scriptsize } \vspace{1.5pt}
\includegraphics[width=5.5in,height=2.8in]{Fig3}
\end{minipage}
}
\caption{\small The distribution of transformed multi-ranks of the hyperspectral data sets with different truncations in each band.
First row: Samson data set.
Second row: Japser Ridge data set.
(a) $\varpi=70\%$.
(b) $\varpi=80\%$.
(c) $\varpi=90\%$.
(d) $\varpi=95\%$.}\label{rankhyp}
\end{figure}
In this subsection, two hyperspectral data sets
(Samson ($95\times95\times156$) and Japser Ridge ($100\times100\times198$) \cite{zhu2014spectral})
are used to
demonstrate
the required number of samples for tensor recovery by the proposed bound.
For Samson data, $n_1$ and $n_2$ are equal to 95 and $n_3$ is equal to 156.
For Japser Ridge data, $n_1$ and $n_2$ are equal to 100 and $n_3$ is equal to 198.
Here we compare our method with
the low-rank tensor completion method using the sum of nuclear norms of
unfolding matrices of a tensor (LRTC)\footnote{\footnotesize http://www.cs.rochester.edu/$\sim$jliu/} \cite{huang2014provable, liu2013},
tensor factorization method (TF)\footnote{\footnotesize https://homes.cs.washington.edu/$\sim$sewoong/papers.html} \cite{jain2014provable}, Square Deal \cite{mu2014square},
gradient descent algorithm on Grassmannians (GoG) \cite{Xia2019On}.
These testing hyperspectral data is normalized on $[0, 1]$.
Their theoretical estimation of samples required are presented in
Table \ref{table1}.
We remark that these hyperspectral images are not exactly low multi-rank tensors,
the multi-rank ($\sum_{i=1}^{n_3} r_i$) is not available.
A truncated tensor is used to compute the transformed multi-rank and tubal rank by using the threshold
$\varpi$.
Here for a given tensor $\mathcal{Z}$ with transformed tensor SVD in (\ref{equ12}) of Theorem \ref{them1}
and $\varpi$,
we determine the smallest value $k$ such that
$$
\frac{ \sum_{i=1}^{k} \varrho_i }
{ \sum_{i=1}^{n_{(2)} n_3} \varrho_i } \ge \varpi,
$$
where $\{ \varrho_i \}$ is the sorted value in ascending order of the numbers
$\{ (\hat{{\cal S}}_{\bf \Phi})_{jj\ell} \}_{1 \le j \le n_{(2)},1 \le \ell \le n_3}$
appearing in the diagonal tensor ${\cal S}$ in (\ref{equ12}). The ratio is used to determine
the significant numbers are kept in the truncated tensor based on
the threshold $\varpi$. Now we can define the transformed multi-rank
$\mathbf{r}(\varpi)$ of the truncated tensor as follows:
$$
\mathbf{r}(\varpi):=(r_1(\varpi),\ldots, r_{n_{3}}(\varpi)), \
\mbox{with}\ r_{\ell}(\varpi):=\# \{ (\hat{{\cal S}}_{\bf \Phi})_{jj\ell} \geq \varrho_k, 1 \le j \le n_{(2)} \},
\ \ell=1,\ldots, n_{3},
$$
Accordingly, the transformed tubal rank of the truncated tensor is defined as
$r(\varpi):=\max\{r_1(\varpi),\ldots, r_{n_{3}}(\varpi)\}$.
The distributions of the transformed multi-ranks of the two hyperspectral
data sets with different $\varpi$ are shown in Figure \ref{rankhyp}.
It can be seen that the $\sum_{i=1}^{n_3} r_i(\varpi)$ obtained by t-SVD (data) is much smaller than
that obtained by t-SVD (FFT) and t-SVD (DCT) for different truncations $\varpi=70\%, 80\%, 90\%, 95\%$,
which implies that the t-SVD (data) needs lower number of samples for successful recovery
than t-SVD (FFT) and t-SVD (DCT).
The distributions of the transformed multi-ranks of the truncated tensor obtained by t-SVD (FFT) are symmetric
due to symmetry of FFT.
\begin{table}[!t]
\scriptsize
\begin{center}
\setlength{\abovecaptionskip}{-1pt}
\setlength{\belowcaptionskip}{-1pt}
\caption{\small The PSNR, SSIM values, and CPU time (in seconds) of different methods for the hyperspectral images.}\label{tab4}
\medskip
\mbox{\hspace{-0.5cm}}\begin{tabular}{|c| c| c c c c c c c | } \hline
& &\multicolumn{7}{c|}{Samson} \\ \cline{3-9}
& const$_1$ & 20 & 30 & 40 & 50 & 60 & 70 & 80 \\
\cline{2-9}
& sampling & \multirow{2}{*}{0.013} & \multirow{2}{*}{0.019} & \multirow{2}{*}{0.026} & \multirow{2}{*}{0.032} & \multirow{2}{*}{0.039} & \multirow{2}{*}{0.045} & \multirow{2}{*}{0.052} \\
& rate & & & & & & & \\ \hline \hline
\multirow{3}{*}{LRTC} & PSNR & 13.08 & 16.00 & 18.34 & 19.82 & 20.99 & 22.12 & 22.98 \\
& SSIM & 0.3254 & 0.5720 & 0.6223 & 0.6643 & 0.6964 & 0.7287 & 0.7548 \\
& CPU & 26.44 & 29.81 & 29.50 & 27.13 & 25.50 & 26.35 & 28.66 \\ \hline
\multirow{3}{*}{TF} & PSNR & 4.05 & 7.45 & 20.26 & 29.78 & 32.86 & 35.54 & 36.39 \\
& SSIM & 0.3990 & 0.4852 & 0.7798 & 0.8627 & 0.8856 & 0.9334 & 0.9440 \\
& CPU & 207.55 & 207.20 & 206.90 & 206.80 & 216.99 & 210.21 & 207.86 \\ \hline
\multirow{3}{*}{Square Deal}
& PSNR & 16.35 & 18.55 & 20.38 & 22.15 & 23.77 & 25.57 & 26.77 \\
& SSIM & 0.4502 & 0.5727 & 0.6696 & 0.7269 & 0.7916 & 0.8310 & 0.8667 \\
& CPU & 37.46 & 34.93 & 33.05 & 30.96 & 29.48 & 29.63 & 34.05 \\ \hline
\multirow{3}{*}{GoG} & PSNR & 16.72 & 26.02 & 27.13 & 30.01 & 30.45 & 30.87 & 31.43 \\
& SSIM & 0.4851 & 0.6978 & 0.7560 & 0.8372 & 0.8366 & 0.8398 & 0.8665 \\
& CPU & 1.05e4 & 1.40e4 & 3.41e4 & 3.87e4 & 4.61e4 & 3.46e4 & 4.15e4 \\ \hline
\multirow{3}{*}{t-SVD (FFT)}
& PSNR & 22.09 & 23.88 & 25.21 & 26.27 & 27.09 & 27.68 & 28.52 \\
& SSIM & 0.5743 & 0.6498 & 0.6996 & 0.7442 & 0.7689 & 0.7912 & 0.8107 \\
& CPU & 30.24 & 26.30 & 25.11 & 24.07 & 22.45 & 23.62 & 26.58 \\ \hline
\multirow{3}{*}{t-SVD (DCT)}
& PSNR & 26.35 & 28.72 & 30.42 & 31.92 & 33.03 & 33.96 & 35.04 \\
& SSIM & 0.7355 & 0.8098 & 0.8552 & 0.8875 & 0.9096 & 0.9215 & 0.9343 \\
& CPU & 172.41 & 167.56 & 177.44 & 185.69 & 193.36 & 187.29 & 190.13 \\ \hline
\multirow{3}{*}{t-SVD (data)}
& PSNR&{\bf28.24}&{\bf31.37}&{\bf33.39}&{\bf35.25}&{\bf36.53} &{\bf 37.09} & {\bf 38.62} \\
& SSIM & 0.8324 & 0.8971 & 0.9311 & 0.9539 & 0.9643 & 0.9658 & 0.9746 \\
& CPU & 241.48 & 258.86 & 270.17 & 282.82 & 284.68 & 263.78 & 291.26 \\ \hline\hline
& &\multicolumn{7}{c|}{Japser Ridge} \\ \cline{3-9}
& const$_1$ & 20 & 30 & 40 & 50 & 60 & 70 & 80 \\
\cline{2-9}
& sampling & \multirow{2}{*}{0.010} & \multirow{2}{*}{0.015} & \multirow{2}{*}{0.020} & \multirow{2}{*}{0.025} & \multirow{2}{*}{0.030} & \multirow{2}{*}{0.035} & \multirow{2}{*}{0.040} \\
& rate & & & & & & & \\ \hline \hline
\multirow{3}{*}{LRTC} & PSNR & 12.48 & 14.80 & 16.78 & 18.17 & 18.97 & 19.61 & 20.15 \\
& SSIM & 0.3245 & 0.4104 & 0.4404 & 0.4611 & 0.4793 & 0.5011 & 0.5223 \\
& CPU & 39.54 & 39.98 & 42.73 & 39.81 & 36.51 & 40.21 & 42.36 \\ \hline
\multirow{3}{*}{TF} & PSNR & 7.12 & 11.41 & 12.09 & 15.39 & 27.03 & 27.83 & 28.07 \\
& SSIM & 0.2076 & 0.3728 & 0.6120 & 0.6418 & 0.7246 & 0.7603 & 0.7693 \\
& CPU & 400.27 & 399.79 & 398.94 & 398.46 & 286.83 & 362.15 & 384.52 \\ \hline
\multirow{3}{*}{Square Deal}
& PSNR & 15.38 & 17.50 & 20.47 & 22.48 & 24.25 & 26.13 & 27.12 \\
& SSIM & 0.3890 & 0.5168 & 0.6199 & 0.7021 & 0.7596 & 0.8046 & 0.8378 \\
& CPU & 43.33 & 41.89 & 41.82 & 41.99 & 40.65 & 42.59 & 40.99 \\ \hline
\multirow{3}{*}{GoG} & PSNR & 15.32 & 22.38 & 23.54 & 24.35 & 25.44 & 25.90 & 26.35 \\
& SSIM & 0.2643 & 0.5430 & 0.5624 & 0.5835 & 0.6550 & 0.6808 & 0.7193 \\
& CPU & 3.49e4 & 4.13e4 & 5.33e4 & 5.50e4 & 5.58e4 & 5.39e4 & 5.41e4 \\ \hline
\multirow{3}{*}{t-SVD (FFT)}
& PSNR & 21.04 & 22.95 & 24.09 & 25.17 & 26.06 & 26.74 & 27.25 \\
& SSIM & 0.4877 & 0.5802 & 0.6353 & 0.6891 & 0.7270 & 0.7578 & 0.7756 \\
& CPU & 50.54 & 42.01 & 37.62 & 35.38 & 34.36 & 42.69 & 35.99 \\ \hline
\multirow{3}{*}{t-SVD (DCT)}
& PSNR & 21.93 & 23.40 & 24.48 & 25.49 & 26.36 & 27.06 & 27.65 \\
& SSIM & 0.5206 & 0.5945 & 0.6481 & 0.7015 & 0.7378 & 0.7647 & 0.7892 \\
& CPU & 200.90 & 171.96 & 164.15 & 169.65 & 172.05 & 184.23 & 193.06 \\ \hline
\multirow{3}{*}{t-SVD (data)}
& PSNR &{\bf23.54}&{\bf25.63}&{\bf27.47}&{\bf28.74}&{\bf30.09}&{\bf 30.78} &{\bf 31.33} \\
& SSIM & 0.6004 & 0.7041 & 0.7816 & 0.8323 & 0.8675 & 0.8777 & 0.8962 \\
& CPU & 235.74 & 217.07 & 217.13 & 230.64 & 235.28 & 225.95 & 231.87 \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[htbp]
\centering
\subfigure[15th band]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{Ori1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{Ob1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{LRTC1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{TF1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{SD1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{GoG1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{FFT1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{DCT1560.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{U1560.png}
\end{minipage}
}
\subfigure[60th band]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{Ori6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{Ob6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{LRTC6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{TF6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{SD6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{GoG6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{FFT6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{DCT6060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{U6060.png}
\end{minipage}
}
\subfigure[90th band]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{Ori9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{Ob9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{LRTC9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{TF9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{SD9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{GoG9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{FFT9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{DCT9060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{U9060.png}
\end{minipage}
}
\subfigure[130th band]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{Ori13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{Ob13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{LRTC13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{TF13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{SD13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{GoG13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{FFT13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{DCT13060.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{U13060.png}
\end{minipage}
}
\caption{\small Recovery results by different methods for the Samson data with const$_1=60$.
First row: Original images.
Second row: Observed images.
Third row: Recovered images by LRTC.
Fourth row: Recovered images by TF.
Fifth row: Recovered images by Square Deal.
Sixth row: Recovered images by GoG.
Seventh row: Recovered images by t-SVD (FFT).
Eighth row: Recovered images by t-SVD (DCT).
Ninth row: Recovered images by t-SVD (data).}\label{HPVisc}
\end{figure}
\begin{figure}[!t]
\centering
\subfigure{
\begin{minipage}[b]{0.99\textwidth}
\centerline{\scriptsize } \vspace{1.5pt}
\includegraphics[width=5.5in,height=2.8in]{Fig5.png}
\end{minipage}
}
\caption{\small The distribution of transformed multi-ranks of the video data sets with different truncations in each frame.
First row: Carphone data set.
Second row: Announcer data set.
(a) $\varpi=70\%$.
(b) $\varpi=80\%$.
(c) $\varpi=90\%$.
(d) $\varpi=95\%$.}\label{rankAK}
\end{figure}
In Table \ref{tab4}, we present the number of samples (const$_1 n_{(1)} \log (n_{(1)} n_3 )$) for several values
of const$_1$ and their corresponding PSNR and SSIM values
of the recovered tensors by different methods, where const$_1$ varies from
20 to 80 with step-size $10$.
Here we can regard const$_1$ as $c_0\mu\sum_{i=1}^{n_3}r_i(\varpi)$ in (\ref{new1}) of Theorem \ref{Theorem1}.
For $\varpi = 70\%$, the values $\sum_{i=1}^{n_3}r_i(\varpi)$ of Samson hyperspectral
image are $346$, $129$, $40$
for t-SVD (FFT), t-SVD (DCT), and t-SVD (data), respectively.
Thus, the chosen const$_1$ is proportional to $\sum_{i=1}^{n_3}r_i(\varpi)$
for the testing tensor of hyperspectral image with approximately low transformed multi-rank.
We can see from the table that for
different values of const$_1$,
the PSNR and SSIM values obtained by t-SVD (data) and t-SVD (DCT)
are higher than those obtained by LRTC, TF, Square Deal, GoG, and t-SVD (FFT).
Moreover, when const$_1 =60$, the recovery performance of the t-SVD (FFT),
t-SVD (DCT), and t-SVD (data) is good enough in terms of PSNR and SSIM values,
which implies the number of sizes is enough for successful recovery by Theorem \ref{Theorem1}.
When const$_1$ is larger, e.g., const$_1 =70,80$, the improvements of PSNR values of the t-SVD (FFT), t-SVD (DCT), and t-SVD (data)
are smaller than those of other small const$_1$.
For Samson data, the performance of GoG is better than that of t-SVD (FFT) for const$_1 \ge 30$. But
the computational time required by GoG is significantly more than that required by the other methods.
For Japser Ridge data, the PSNR and SSIM values obtained by all three t-SVD methods
are almost higher than those obtained by LRTC, TF, Square Deal, and GoG. Moreover, the computational time
required by the t-SVD methods are quite efficient compared with the other methods.
Figure \ref{HPVisc} shows the visual comparisons of different bands obtained by LRTC, TF, Square Deal,
GoG, and three t-SVD methods for the Samson data, where const$_1=60$.
We can see that the t-SVD (data) and t-SVD (DCT) outperform LRTC, TF, Square Deal, and t-SVD (FFT) in terms of visual quality.
Moreover, the images recovered by t-SVD (data) keep
more details than those recovered by LRTC, TF, Square Deal, t-SVD (FFT), and t-SVD (DCT).
\subsubsection{Video Data}
In this subsection, we test two video data sets (length $\times$ width $\times$ frames)
to show the performance of the proposed method,
where the testing videos include
Carphone ($144\times176\times180$) and
Announcer ($144\times176\times200$)\footnote{\footnotesize https://media.xiph.org/video/derf/},
and we just use the first channels of all frames
in the original data. Moreover, the first $180$ and $200$
frames for the two videos are chosen to improve the computational time.
The intensity range of the video images is scaled into $[0,1]$ in the experiments.
\begin{table}[!t]
\tiny
\begin{center}
\setlength{\abovecaptionskip}{-1pt}
\setlength{\belowcaptionskip}{-1pt}
\caption{\small The PSNR, SSIM values, and CPU time (in seconds) of different methods for the video data sets.}\label{tab5}
\medskip
\mbox{\hspace{-1.3cm}}\begin{tabular}{|c|c| c| c c c c c c c c c c c | } \hline
& & const$_1$ & 20 & 30 & 40 & 50 & 60 & 70 & 80 & 90 & 100 & 110 & 120 \\
\cline{3-14}
& & sampling & \multirow{2}{*}{0.008} & \multirow{2}{*}{0.012} & \multirow{2}{*}{0.016} & \multirow{2}{*}{0.020} & \multirow{2}{*}{0.024} & \multirow{2}{*}{0.028} & \multirow{2}{*}{0.032}
& \multirow{2}{*}{0.036} & \multirow{2}{*}{0.040} & \multirow{2}{*}{0.044} & \multirow{2}{*}{0.048} \\
& & rate & & & & & & &
& & & & \\ \hline \hline
\multirow{21}{*}{Carphone}&
\multirow{3}{*}{LRTC} & PSNR & 10.81 & 11.65 & 12.49 & 13.19 & 13.97 & 14.75 & 15.40 & 16.01 & 16.53 & 17.05 & 17.52 \\
& & SSIM & 0.3498 & 0.3571 & 0.3687 & 0.3870 & 0.4085 & 0.4319 & 0.4496 & 0.4701 & 0.4862 & 0.5038 & 0.5206 \\
& & CPU & 90.78 & 85.06 & 85.92 & 85.12 & 89.76 & 92.28 & 91.70 & 90.47 & 92.91 & 92.06 & 95.21 \\ \cline{2-14}
& \multirow{3}{*}{TF} & PSNR & 4.96 & 4.64 & 13.41 & 12.98 & 20.82 & 22.57 & 22.66 & 22.74 & 22.75 & 22.85 & 23.04 \\
& & SSIM & 0.2855 & 0.4507 & 0.5081 & 0.5499 & 0.5837 & 0.6461 & 0.6531 & 0.6607 & 0.6622 & 0.6705 & 0.6746 \\
& & CPU & 780.55 & 779.63 & 771.43 & 782.11 & 762.25 & 770.57 & 771.03 & 765.42 & 760.32 & 782.45 & 795.41 \\ \cline{2-14}
& \multirow{3}{*}{Square Deal}
& PSNR & 10.04 & 12.99 & 13.55 & 15.00 & 16.98 & 17.15 & 18.10 & 19.15 & 20.42 & 20.47 & 21.17 \\
& & SSIM & 0.1796 & 0.2675 & 0.3304 & 0.3988 & 0.4497 & 0.4930 & 0.5348 & 0.5719 & 0.6113 & 0.6178 & 0.6396 \\
& & CPU & 76.98 & 80.23 & 76.06 & 77.03 & 76.76 & 76.36 & 74.78 & 80.00 & 78.59 & 77.21 & 79.75 \\ \cline{2-14}
& \multirow{3}{*}{GoG} & PSNR & 18.59 & 20.18 & 20.77 & 20.46 & 21.22 & 21.35 & 21.48 & 22.11 & 22.47 & 22.86 & 23.15 \\
& & SSIM & 0.4010 & 0.4653 & 0.5229 & 0.5023 & 0.5449 & 0.5483 & 0.5547 & 0.5768 & 0.5874 & 0.5936 & 0.6058 \\
& & CPU & 1.09e5 & 2.27e5 & 1.30e5 & 1.40e5 & 2.47e5 & 6.27e5 & 3.76e5 & 6.71e5 & 6.32e5 & 6.52e5 & 6.48e5 \\ \cline{2-14}
& \multirow{3}{*}{t-SVD (FFT)}
& PSNR & 9.79 & 11.55 & 15.06 & 18.53 & 22.04 & 22.59 & 23.07 & 23.38 & 23.72 & 24.01 & 24.24 \\
& & SSIM & 0.1814 & 0.2463 & 0.3759 & 0.4899 & 0.6121 & 0.6367 & 0.6566 & 0.6690 & 0.6816 & 0.6934 & 0.7017 \\
& & CPU & 53.78 & 58.69 & 77.43 & 88.78 & 107.26 & 103.15 & 98.25 & 93.03 & 88.88 & 90.36 & 92.81 \\ \cline{2-14}
& \multirow{3}{*}{t-SVD (DCT)}
& PSNR & 20.24 & 21.18 & 21.93 & 22.42 & 22.91 & 23.24 & 23.60 & 23.87 & 24.15 & 24.44 & 24.65 \\
& & SSIM & 0.5230 & 0.5720 & 0.6044 & 0.6286 & 0.6506 & 0.6642 & 0.6772 & 0.6890 & 0.6484 & 0.7106 & 0.7185 \\
& & CPU & 323.11 & 291.24 & 271.55 & 251.17 & 237.77 & 226.54 & 214.85 & 205.35 & 198.58 & 200.15 & 197.32 \\ \cline{2-14}
& \multirow{3}{*}{t-SVD (data)}
& PSNR &{\bf20.39}&{\bf21.43}&{\bf22.30}&{\bf22.87}&{\bf23.45}&{\bf23.75}&{\bf24.23}&{\bf24.56}&{\bf24.85}&{\bf 24.94}&{\bf 25.17} \\
& & SSIM & 0.5294 & 0.5841 & 0.6236 & 0.6499 & 0.6747 & 0.6877 & 0.7063 & 0.7200 & 0.7298 & 0.7329 & 0.7426 \\
& & CPU & 366.20 & 337.21 & 314.40 & 293.42 & 278.06 & 261.55 & 253.14 & 242.91 & 235.86 & 261.02 & 248.59 \\ \hline
& & const$_1$ & 20 & 30 & 40 & 50 & 60 & 70 & 80 & 90 & 100 & 110 & 120 \\
\cline{3-14}
& & sampling & \multirow{2}{*}{0.007} & \multirow{2}{*}{0.011} & \multirow{2}{*}{0.015} & \multirow{2}{*}{0.018} & \multirow{2}{*}{0.022} & \multirow{2}{*}{0.025} & \multirow{2}{*}{0.029}
& \multirow{2}{*}{0.033} & \multirow{2}{*}{0.036} & \multirow{2}{*}{0.040} & \multirow{2}{*}{0.044} \\
& & rate & & & & & & &
& & & & \\ \hline \hline
\multirow{21}{*}{Announcer}&
\multirow{3}{*}{LRTC} & PSNR & 12.72 & 14.03 & 15.28 & 16.42 & 17.76 & 18.83 & 19.88 & 20.74 & 21.46 & 22.06 & 22.58 \\
& & SSIM & 0.4727 & 0.4900 & 0.5087 & 0.5336 & 0.5601 & 0.5869 & 0.6120 & 0.6341 & 0.6520 & 0.6710 & 0.6871 \\
& & CPU & 100.75 & 96.25 & 100.67 & 103.05 & 104.02 & 105.00 & 104.68 & 103.02 & 100.29 & 101.23 &100.96 \\ \cline{2-14}
& \multirow{3}{*}{TF} & PSNR & 4.39 & 9.99 & 11.36 & 15.48 & 24.12 & 25.65 & 26.22 & 26.28 & 27.99 & 28.29 & 28.34 \\
& & SSIM & 0.4229 & 0.6176 & 0.6599 & 0.6743 & 0.7076 & 0.7363 & 0.7534 & 0.7544 & 0.8169 & 0.8265 & 0.8300 \\
& & CPU & 887.20 & 862.11 & 872.41 & 874.64 & 869.01 & 864.04 & 822.11 & 835.96 & 858.93 & 862.62 & 839.15 \\ \cline{2-14}
& \multirow{3}{*}{Square Deal}
& PSNR & 11.50 & 14.61 & 16.71 & 19.36 & 20.86 & 23.34 & 24.15 & 24.27 & 26.76 & 28.24 & 28.87 \\
& & SSIM & 0.1692 & 0.2702 & 0.3780 & 0.5004 & 0.6207 & 0.6950 & 0.7488 & 0.7928 & 0.8394 & 0.8743 & 0.8918 \\
& & CPU & 86.39 & 87.33 & 80.39 & 78.26 & 77.23 & 77.46 & 80.36 & 80.28 & 81.06 & 80.25 & 80.38 \\ \cline{2-14}
& \multirow{3}{*}{GoG} & PSNR & 21.85 & 23.59 & 23.72 & 24.80 & 24.96 & 25.42 & 26.20 & 26.87 & 27.14 & 27.43 & 27.79 \\
& & SSIM & 0.5113 & 0.5854 & 0.6144 & 0.6691 & 0.6775 & 0.6976 & 0.7183 & 0.7566 & 0.7977 & 0.8236 & 0.8395 \\
& & CPU & 1.94e5 & 2.31e5 & 2.60e5 & 3.54e5 & 6.09e5 & 6.36e5 & 5.49e5 & 5.98e5 & 5.90e5 & 5.89e5 & 6.01e5 \\ \cline{2-14}
& \multirow{3}{*}{t-SVD (FFT)}
& PSNR & 11.01 & 16.64 & 27.03 & 27.98 & 28.63 & 29.11 & 29.46 & 29.83 & 30.12 & 30.35 & 30.62 \\
& & SSIM & 0.2606 & 0.5339 & 0.8125 & 0.8380 & 0.8564 & 0.8686 & 0.8776 & 0.8865 & 0.8922 & 0.8987 & 0.9022 \\
& & CPU & 61.87 & 97.82 & 153.63 & 144.82 & 132.18 & 125.67 & 118.75 & 106.63 & 103.44 & 113.25 & 108.63 \\ \cline{2-14}
& \multirow{3}{*}{t-SVD (DCT)}
& PSNR & 26.06 & 27.12 & 27.92 & 28.58 & 29.16 & 29.58 & 29.95 & 30.32 & 30.69 & 30.93 & 31.17 \\
& & SSIM & 0.7645 & 0.8078 & 0.8353 & 0.8875 & 0.8703 & 0.8798 & 0.8893 & 0.8985 & 0.9046 & 0.9099 & 0.9140 \\
& & CPU & 537.80 & 466.05 & 432.45 & 407.57 & 377.23 & 350.59 & 329.40 & 298.93 & 287.79 & 298.46 & 284.32 \\ \cline{2-14}
& \multirow{3}{*}{t-SVD (data)}
& PSNR &{\bf26.08}&{\bf27.25}&{\bf28.12}&{\bf28.76}&{\bf29.41}&{\bf29.88}&{\bf30.24}&{\bf30.71}&{\bf31.14}&{\bf 31.29}&{\bf 31.58} \\
& & SSIM & 0.7649 & 0.8113 & 0.8399 & 0.8581 & 0.8741 & 0.8838 & 0.8941 & 0.9037 & 0.9106 & 0.9150 & 0.9198 \\
& & CPU & 608.38 & 531.39 & 472.35 & 460.30 & 425.59 & 395.42 & 353.94 & 341.43 & 330.40 & 365.25 & 341.29 \\ \hline
\end{tabular}
\end{center}
\end{table}
Similar to Section \ref{HPDS}, the video data sets are not exactly low-rank.
First, we show the distributions of the transformed multi-ranks with different transformations
and truncations $\varpi$ in Figure \ref{rankAK}.
It can be observed that the $\sum_{i=1}^{n_3} r_i(\varpi)$ obtained by t-SVD (data)
is much smaller than that obtained by t-SVD (FFT) and t-SVD (DCT) for different $\varpi$.
Therefore, the number of samples required by t-SVD (data) would be smaller than that
required by t-SVD (FFT) and t-SVD (DCT) for the same performance.
The distributions of transformed multi-ranks obtained by t-SVD (FFT)
are symmetric for different $\varpi$ since the FFT has symmetric property.
\begin{figure}[htbp]
\centering
\subfigure[20th frame]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{AKOri20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKObs80-20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKLRTC80-20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKTF80-20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKSD80-20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKGoG80-20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKFFT80-20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKDCT80-20.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKU80-20.png}
\end{minipage}
}
\subfigure[80th frame]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{AKOri80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKObs80-80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKLRTC80-80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKTF80-80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKSD80-80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKGoG80-80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKFFT80-80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKDCT80-80.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKU80-80.png}
\end{minipage}
}
\subfigure[120th frame]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{AKOri120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKObs80-120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKLRTC80-120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKTF80-120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKSD80-120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKGoG80-120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKFFT80-120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKDCT80-120.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKU80-120.png}
\end{minipage}
}
\subfigure[180th frame]{
\begin{minipage}[b]{0.2\textwidth}
\centerline{\scriptsize } \vspace{1pt}
\includegraphics[width=1.2in,height=0.76in]{AKOri180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKObs80-180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKLRTC80-180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKTF80-180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKSD80-180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKGoG80-180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKFFT80-180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKDCT80-180.png}
\centerline{\scriptsize }\\ \vfill \vspace{-23pt}
\includegraphics[width=1.2in,height=0.76in]{AKU80-180.png}
\end{minipage}
}
\caption{\small Recovery results by different methods for the Announcer data with const$_1=80$.
First row: Original images.
Second row: Observed images.
Third row: Recovered images by LRTC.
Fourth row: Recovered images by TF.
Fifth row: Recovered images by Square Deal.
Sixth row: Recovered images by GoG.
Seventh row: Recovered images by t-SVD (FFT).
Eighth row: Recovered images by t-SVD (DCT).
Ninth row: Recovered images by t-SVD (data).}\label{VideoVisc}
\end{figure}
We use const$_1 n_{(1)}\log(n_{(1)}n_3)$
number of samples based on Theorem \ref{Theorem1}. The
range const$_1$ is from 20 to 120 with increment
step-size being 10.
For example, when $\varpi=70\%$,
the $\sum_{i=1}^{n_3}r_i(\varpi)$ of Announcer obtained by t-SVD (FFT), t-SVD (DCT), t-SVD (data)
are $1886$, $1649$, $1502$, respectively. When const$_1$ is equal to 100,
the required sample size would be enough for successful recovery by Theorem \ref{Theorem1}.
When const$_1 >100$, e.g., const$_1= 110,120$, the improvements of PSNR values by the three t-SVD methods are very small.
In Table \ref{tab5}, we show the PSNR, SSIM values, and CPU time (in seconds)
of different methods for the testing video data sets with different const$_1$.
It can be seen that the performance of t-SVD (data) is better
than that of LRTC, TF, Square Deal, GoG, t-SVD (FFT), and t-SVD (DCT) in terms of PSNR and SSIM values.
The PSNR and SSIM values obtained by t-SVD (DCT) are higher than those obtained by t-SVD (FFT).
Hence, the number of samples required by t-SVD (DCT) and t-SVD (data) is
smaller than that required by LRTC, TF, Square Deal, GoG, and t-SVD (FFT) for the same recovery performance,
which demonstrates the conclusion of Theorem \ref{Theorem1}.
Moreover, the CPU time (in seconds) required by GoG is much more than that required by other methods.
Figure \ref{VideoVisc} shows the visual quality of the 20th, 80th, 120th, 180th frames of the recovered images by LRTC, TF, Square Deal, GoG, and three t-SVD methods, where const$_1 = 80$.
It can be seen that the three t-SVD methods outperform LRTC, TF, Square Deal,
and GoG for different frames in terms of visual quality,
where the recovered images by the t-SVD methods are more clear.
\section{Concluding Remarks}\label{Sect5}
In this paper, we have established the sample size requirement
for exact recovery in the tensor completion problem
by using transformed tensor SVD.
We have shown that for any $\mathcal{Z}\in\mathbb{C}^{n_1\times n_2\times n_3}$
with transformed multi-rank $(r_1,r_2,\ldots, r_{n_3})$,
one can recover the tensor exactly with high probability under some incoherence conditions
if the sample size of observations is of the order
$O(\sum_{i=1}^{n_3}r_i\max \{n_1, n_2 \} \log ( \max \{ n_1, n_2 \} n_3))$
under uniformly sampled entries.
The sample size requirement of our theory for exact recovery is smaller
than that of existing methods for tensor completion.
Moreover, several numerical experiments on both synthetic data and real-world data sets are
presented to show the superior performance of our methods in comparison with other state-of-the-art methods.
In further work, it would be of great interest to extend the transformed tensor SVD and
tensor completion results to higher-order tensors (cf. \cite{martin2013order}).
It would be also of great interest to extend the result of the unitary transformation
to any invertible linear transformation for tensor completion (cf. \cite{kernfel}).
\section*{Acknowledgments}
The authors are grateful to Dr. Cun Mu at Walmart Labs
and Dr. Dong Xia at The Hong Kong University of Science and Technology
for sharing the codes of the Square Deal \cite{mu2014square} and GoG methods \cite{Xia2019On}, respectively.
The authors are also grateful to the anonymous referees for
their constructive suggestions and comments to improve the presentation of the paper.
\section*{Appendix A.} \label{Attechment}
We first list the following lemma, which is the main tool to prove our conclusions.
\begin{lemma}(\cite[Theorem 4]{Recht2011})\label{lem5}
Let $X_{1},\ldots,X_{L}\in \mathbb{R}^{n\times n}$ be independent zero mean random matrices of dimension $d_{1}\times d_{2}$. Suppose
$\rho_{k}^2=\max\left\{\|\mathbb{E} [X_{k}X_{k}^{T}]\|,\|\mathbb{E} [X_{k}^{T}X_{k}]\|\right\}$
and $\|X_{k}\|\leq B$ almost surely for all $k=1,\ldots,L.$ Then for any $\tau>0,$
\begin{equation*}
\mathbb{P}\left[\Biggl\|\sum_{k=1}^{L}X_{k}\Biggl\|>\tau\right]\leq(d_{1}+d_{2})\exp\left(\frac{-\tau^2/2}{\sum_{k=1}^{L}\rho_{k}^2+B\tau/3}\right).
\end{equation*}
Moreover, if
$
\max\left\{\big\|\sum_{k=1}^{L}X_{k}X_{k}^{T}\big\|,\big\|\sum_{k=1}^{L}X_{k}^{T}X_{k}\big\|\right\}\leq \sigma^2,
$
then for any $c>1,$ we have
\begin{align*}
\left\|\sum_{k=1}^{L}X_{k}\right\|\leq \sqrt{4c\sigma^2\log(d_{1}+d_{2})}+cB\log(d_{1}+d_{2})
\end{align*}
holds with probability at least $1-(d_{1}+d_{2})^{-(c-1)}.$
\end{lemma}
\vspace{2mm}
\noindent
\subsection*{Proof of Lemma \ref{le1}}
Let $\mathcal{E}_{ijk}$ be a unit tensor whose $(i,j,k)$-th
entry is 1 and others are 0. Then for an arbitrary tensor $ \mathcal{Z}\in \mathbb{C}^{n_1 \times n_2 \times n_3}$, we have $\mathcal{Z} = \sum_{i,j,k} \<\mathcal{E}_{ijk},
\mathcal{Z}\>\mathcal{E}_{ijk}.$ Recall Definition \ref{defn}, $\mathcal{E}_{ijk}$ can be expressed as $\mathcal{E}_{ijk}= \tc{e}_{ik}\diamond_{\bf \Phi}
\ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}_{jk}^H,$ then ${\cal P}_T(\mathcal{Z})$ can also be decomposed as
\begin{align*}
{\cal P}_T(\mathcal{Z}) = \sum_{i,j,k} \<{\cal P}_T(\mathcal{Z}), \tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk}\> \tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk},
\end{align*}
where $T$ is defined as \eqref{e1}. Similarly,
\begin{align}
\rho^{-1}{\cal P}_T{\cal P}_{\Omega}{\cal P}_T(\mathcal{Z}) = \sum_{i,j,k}
\rho^{-1}\delta_{ijk}\<\mathcal{Z}, {\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\>{\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk}). \nonumber
\end{align}
Define the operator $\mathcal{T}^{ijk}$ as:
\begin{align*}
\mathcal{T}^{ijk}(\mathcal{Z})=\rho^{-1}\delta_{ijk}\<\mathcal{Z}, {\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\>{\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk}).
\end{align*}
Then by the definition of tensor operator norm, we can get
\begin{align*}
\|\mathcal{T}^{ijk}\|_{\textup{op}} =\frac{1}{\rho}\|{\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\|^2_F, ~\text{and}~\|{\cal P}_T\|_{\textup{op}} \leq 1.
\end{align*}
Note the fact that for any two positive semidefinite matrices ${A},{B}\in\mathbb{C}^{n\times n}$, we have $\|{A} -
{B}\| \leq \max\{\|{A}\|, \|{B}\|\}.$
Therefore, by the inequality given in Proposition \ref{pro1}, we have
\begin{equation}
\Big\|\mathcal{T}^{ijk} -
\frac{1}{n_1n_2n_3}{\cal P}_T\Big\|_{\textup{op}} \leq
\max\left\{\frac{1}{\rho}\|{\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\|^2_F,
\frac{1}{n_1n_2n_3}\right\} \leq \frac{2\mu \sum{r_{i}}}{n_{(2)}n_{3}\rho}.
\nonumber
\end{equation}
In addition, one has
\begin{align}
&~\Big\|\mathbb{E}\Big(\mathcal{T}^{ijk}- \frac{1}{n_1 n_2 n_3}{\cal P}_T\Big)^2\Big\|_\textup{op}\nonumber\\
\leq &~\left\|\mathbb{E}\left(\frac{1}{\rho}\|{\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\|^2_F\mathcal{T}^{ijk}\right) - \frac{2}{n_1 n_2 n_3}{\cal P}_T\mathbb{E}(\mathcal{T}^{ijk}) + \frac{1}{n^2_1n^2_2n^2_3}{\cal P}_T\right\| \nonumber\\
=& ~ \left\|\frac{1}{\rho}\|{\cal P}_T(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\|^2_F\frac{1}{n_1n_2n_3}{\cal P}_T - \frac{1}{n^2_1 n^2_2 n^2_3}{\cal P}_T\right\|
\leq \frac{2\mu \sum{r_{i}}}{n_{(1)}n^2_{(2)}n^2_3\rho}. \nonumber
\end{align}
Setting $\tau = \sqrt{\frac{14 \mu \beta \sum{r_{i}}
\log(n_{(1)}n_3)}{3n_{(2)}n_{3}\rho}} \leq
\frac{1}{2}$ with any $\beta>1$ and using Lemma~\ref{lem5}, we have
\begin{align}
&\mathbb{P}[\|\rho^{-1}{\cal P}_T{\cal P}_{\Omega}{\cal P}_T - {\cal P}_T\|_\textup{op} > \tau]
= \mathbb{P}\left[\left\|\sum_{i,j,k} \left(\mathcal{T}^{ijk} - \frac{1}{n_1n_2n_3}{\cal P}_T\right)\right\|_{\textup{op}}> \tau \right] \nonumber\\
\leq~ & 2n_{(1)}n_3\exp\Bigg(\frac{\frac{-7 \mu \beta\sum{r_{i}}
\log(n_{(1)}n_3)}{3n_{(2)}n_{3}\rho}}{\frac{2\mu \sum{r_{i}}}{n_{(2)}n_{3}\rho} + \frac{2\mu \sum{r_{i}}}{6n_{(2)}n_{3}\rho}}\Bigg)
= 2n_{(1)}n_3\exp({-\beta \log n_{(1)}n_{3}})= 2(n_{(1)}n_3)^{1-\beta},
\nonumber
\end{align}
which implies that
\begin{align}
\mathbb{P}\left[\|\rho^{-1}{\cal P}_T{\cal P}_{\Omega}{\cal P}_T - {\cal P}_T\|_\textup{op} \leq \epsilon\right] \geq 1 - 2(n_{(1)}
n_3)^{1-\beta}.\nonumber
\end{align}
This completes the proof.
\vspace{2mm}
\noindent
\section*{Appendix B. Proof of Lemma \ref{lemma2}}
Denote
$$
\rho^{-1}\mathcal{P}_{\Omega}(\mathcal{Z})-\mathcal{Z}= \sum_{i,j,k}\mathcal{G}^{ijk}=\sum_{i,j,k}\Big(\frac{1}{\rho}\delta_{ijk}-1\Big)\mathcal{Z}_{ijk}\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk}.
$$
Then by the independence of $\delta_{ijk}$, we have $\mathbb{E}[\mathcal{G}^{ijk}]=\bf{0}$ and $\|\mathcal{G}^{ijk}\|\leq \frac{1}{\rho}\|\mathcal{Z}\|_{\infty}.$
Moreover,
\begin{align*}
\left\|\mathbb{E}\left[\sum_{i,j,k}(\mathcal{G}^{ijk})^H\diamond_{\bf \Phi}\mathcal{G}^{ijk}\right]\right\|
& =\left\|\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\tc{e}_{jk}\diamond_{\bf \Phi}\tc{e}_{jk}^H\mathbb{E}\left(\frac{1}{\rho}\delta_{ijk}-1\right)^2\right\| \\
& =\left\|\frac{1-\rho}{\rho}\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\tc{e}_{ik}\diamond_{\bf \Phi}\tc{e}_{ik}^H\right\|.
\end{align*}
Recall the definition of tensor basis, we can get that
${\bf \Phi}[\tc{e}_{jk}\diamond_{\bf \Phi}\tc{e}_{jk}^H]$ is a tensor except the $(j,j,t)$-th tube entries equaling to $({\bf \Phi}[\tub{e}_{k}])_{t}^2=\alpha^2_{t}, t=1,\ldots,n_3 $ with $\sum_{t=1}^{n_{3}}\alpha_{t}^2=1,$
and 0 otherwise. Hence, we get that
\begin{align*}
\left\|\mathbb{E}\left[\sum_{i,j,k}(\mathcal{G}^{ijk})^H\diamond_{\bf \Phi}\mathcal{G}^{ijk}\right]\right\|
&=\frac{1-\rho}{\rho}\left\|\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\tc{e}_{jk}\diamond_{\bf \Phi}\tc{e}_{jk}^H\right\| \\
& =\frac{1-\rho}{\rho}\max_{j}\left\|\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\tc{e}_{jk}\diamond_{\bf \Phi}\tc{e}_{jk}^H\right\|\\
&=\frac{1-\rho}{\rho}\max_{j}\left\|\sum_{i,k}|\mathcal{Z}_{ijk}|^2(\overline{\tc{e}}_{jk})_{\bf \Phi}\cdot(\overline{\tc{e}}_{jk})_{\bf \Phi}^H\right\|
\leq \frac{1}{\rho}\|\mathcal{Z}\|^2_{\infty,w}.
\end{align*}
Moreover, $\left\|\mathbb{E}\left[\sum_{i,j,k}\mathcal{G}^{ijk}\diamond_{\bf \Phi}(\mathcal{G}^{ijk})^H\right]\right\|$ can be also bounded similarly.
Then by Lemma \ref{lem5}, we can get that
\begin{align*}
\|\rho^{-1}\mathcal{P}_{\Omega}(\mathcal{Z})-\mathcal{Z}\|_\textup{op}\leq
c\left(\frac{\log(n_{(1)}n_{3})}{\rho}\|\mathcal{Z}\|_{\infty}+\sqrt{\frac{\log (n_{(1)}n_{3})}{\rho}}\|\mathcal{Z}\|_{\infty,w}\right)
\end{align*}
holds with high probability provided that
$m\geq C_0 \epsilon^{-2} \mu \sum{r_{i}}n_{(1)}\log(n_{(1)}n_3).$
\vspace{2mm}
\noindent
\section*{Appendix C. Proof of Lemma \ref{le3}}
Denote the weighted $b$-th lateral slice of $(\rho^{-1}\mathcal{P}_{T}\mathcal{P}_{\Omega}-\mathcal{P}_{T})\mathcal{Z}$ as
\begin{align*}
\sum_{i,j,k}\mathcal{F}^{ijk}:&=(\rho^{-1}\mathcal{P}_{T}\mathcal{P}_{\Omega}-\mathcal{P}_{T})\mathcal{Z}\diamond_{\bf \Phi}\tc{e}_{bk} \\
&=\sum_{i,j,k}\Big(\frac{1}{\rho}\delta_{ijk}-1\Big)\mathcal{Z}_{ijk}\mathcal{P}_{T}(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\diamond_{\bf \Phi}\tc{e}_{bk},
\end{align*}
where $\mathcal{F}^{ijk}\in \mathbb{C}^{n_1 \times 1 \times n_3}$ are zero-mean independent lateral slices. By the incoherence conditions given in Proposition \ref{pro1}, we have
\begin{align*}
\|\mathcal{F}^{ijk}\|_{{F}}=\left\|\left(\frac{1}{\rho}\delta_{ijk}-1\right)\mathcal{Z}_{ijk}\mathcal{P}_{T}(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\diamond_{\bf \Phi}\tc{e}_{bk}\right\|_{{F}}
\leq \frac{1}{\rho}\sqrt{\frac{2\mu \sum_{i=1}^{n_{3}}r_{i}}{n_{1}n_3}}\|\mathcal{Z}\|_{\infty}.
\end{align*}
Furthermore,
\begin{align*}
\left\|\mathbb{E}\left[\sum_{i,j,k}(\mathcal{F}^{ijk})^H\diamond_{\bf\Phi}\mathcal{F}^{ijk}\right]\right\|_{{F}}
=\frac{1-\rho}{\rho}\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\|\mathcal{P}_{T}(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\diamond_{\bf\Phi}\tc{e}_{bk}\|^2_{{F}}.
\end{align*}
Then by the definition of $\mathcal{P}_{T}$, we can get
\begin{align*}
&~\|\mathcal{P}_{T}(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\diamond_{\bf \Phi}\tc{e}_{bk}\|^2_{F}\\
=&~\|\mathcal{U}\diamond_{\bf \Phi}\mathcal{U}^H\diamond_{\bf\Phi}\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk}\diamond_{\bf \Phi}\tc{e}_{bk} \\
&~~~~+(\mathcal{I}_{\bf \Phi}-\mathcal{U}\diamond_{\bf\Phi}\mathcal{U}^H)\diamond_{\bf\Phi}\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk}\diamond_{\bf\Phi}\mathcal{V}\diamond_{\bf\Phi}
\mathcal{V}^H\diamond_{\bf \Phi}\tc{e}_{bk}\|^2_{{F}}\\
\leq&~\frac{\mu \sum_{i=1}^{n_{3}}r_{i}}{n_{1}n_3}\|\ddot{\bm{e}}_k\diamond_{\bf \Phi}\tc{e}_{jk}\diamond_{\bf \Phi}\tc{e}_{bk}\|^2_{{F}}
+\|\tc{e}_{jk}^H\diamond_{\bf\Phi}\mathcal{V}\diamond_{\bf\Phi}
\mathcal{V}^H\diamond_{\bf \Phi}\tc{e}_{bk}\|^2_{{F}}.
\end{align*}
Therefore, we obtain
\begin{align*}
~& \left\|\mathbb{E}\left[\sum_{i,j,k}(\mathcal{F}^{ijk})^H\diamond_{\bf\Phi}\mathcal{F}^{ijk}\right]\right\|_{F} \\
\leq ~&
\frac{1}{\rho}\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\|\mathcal{P}_{T}(\tc{e}_{ik} \diamond_{\bf \Phi} \ddot{\bm{e}}_k \diamond_{\bf \Phi} \tc{e}^{H}_{jk})\diamond_{\bf\Phi}\tc{e}_{bk}\|^2_{{F}}\\
\leq ~& \frac{1}{\rho}\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\frac{\mu \sum_{i=1}^{n_{3}}r_{i}}{n_{1}n_3}\|\ddot{\bm{e}}_k\diamond_{\bf \Phi}\tc{e}_{jk}\diamond_{\bf \Phi}\tc{e}_{bk}\|^2_{{F}}
+\frac{1}{\rho}\sum_{i,j,k}|\mathcal{Z}_{ijk}|^2\|\tc{e}_{jk}^H\diamond_{\bf\Phi}\mathcal{V}\diamond_{\bf\Phi}
\mathcal{V}^H\diamond_{\bf \Phi}\tc{e}_{bk}\|^2_{{F}}\\
\leq ~& \frac{\mu \sum_{i=1}^{n_{3}}r_{i}}{\rho n_{1}n_3}\|\mathcal{Z}\|^2_{\infty,w}+\frac{1}{\rho}\sum_{i,k}|\mathcal{Z}_{ijk}|^2\|\tc{e}_{jk}^H\diamond_{\bf\Phi}\mathcal{V}\diamond_{\bf\Phi}
\mathcal{V}^H\diamond_{\bf \Phi}\tc{e}_{jk}\|^2_{{F}}\\
\leq ~& \frac{2\mu \sum_{i=1}^{n_{3}}r_{i}}{\rho n_{(2)}n_3}\|\mathcal{Z}\|_{\infty,w}^2,
\end{align*}
where the third inequality can be derived by $\tc{e}_{jk}^H\diamond_{\bf\Phi}\tc{e}_{bk}=\bf{0}$ if $j\neq b.$ By the same argument, $\left\|\mathbb{E}[\sum_{ijk}\mathcal{F}^{ijk}\diamond_{\bf\Phi}(\mathcal{F}^{ijk})^H]\right\|_{F}$ can be bounded by the same quantity. Therefore, by Lemma \ref{lem5}, we get that
\begin{align*}
\|(\rho^{-1}\mathcal{P}_{T}\mathcal{P}_{\Omega}-\mathcal{P}_{T})\mathcal{Z}\diamond_{\bf \Phi}\tc{e}_{bk}\|_{{F}}\leq \frac{1}{2}\|\mathcal{Z}\|_{\infty,w}+\frac{1}{2}\sqrt{\frac{n_{(1)}n_{3}}{\mu \sum_{i=1}^{n_{3}}r_{i}}}\|\mathcal{Z}\|_{\infty}
\end{align*}
holds with high probability.
We can also get the same results with respect to $\tc{e}^H_{ak}\diamond_{\bf \Phi}(\rho^{-1}\mathcal{P}_{T}\mathcal{P}_{\Omega}-\mathcal{P}_{T})\mathcal{Z}.$
Then Lemma \ref{le3} follows from using a union bound over
all the tensor columns and rows, and the desired results hold with high probability.
\bibliographystyle{abbrv}
| 34f32ddd70e7a7a52517d47b24db8cc6ef946cef | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The Schur functions $s_\lambda$ are an important family of symmetric functions.
They form a linear basis of the space of symmetric functions and have
connections to different areas of mathematics including combinatorics and
representation theory. Schur functions have been extensively studied and there
are numerous generalizations and variations of them in the literature. In
particular, Macdonald \cite{Macdonald_Schur} introduced nine variations of Schur
functions. Macdonald's 9th variation of Schur functions generalize Schur
functions and many of their variations. The main focus of this paper is to find
determinant identities for Macdonald's 9th variation of Schur functions.
We briefly review several known determinant formulas for Schur functions. The
classical Jacobi--Trudi formula and its dual formula express a Schur function
$s_\lambda$ as a determinant in terms of complete homogeneous symmetric functions
$h_k$ and elementary symmetric functions $e_k$:
\[
s_\lambda = \det \left( h_{\lambda_i+j-i} \right)_{i,j=1}^{\ell(\lambda)}, \qquad
s_\lambda = \det \left( e_{\lambda'_i+j-i} \right)_{i,j=1}^{\ell(\lambda')},
\]
where $\ell(\lambda)$ is the number of parts in the partition $\lambda$ and
$\lambda'$ is the transpose of $\lambda$. Observe that $h_k$ and $e_k$ are also
Schur functions whose shapes are partitions with one row and one column,
respectively. The Giambelli formula \cite{Giambelli} and the Lascoux--Pragacz
formula \cite{Lascoux1988} express a Schur function as a determinant of Schur
functions whose shapes are hooks and border strips, respectively. Using
so-called outside decompositions, Hamel and Goulden \cite{Hamel_1995} found a
determinant formula for a Schur function, which generalizes all of the
aforementioned formulas. Chen, Yan, and Yang \cite{Chen_2005} restated the
Hamel--Goulden formula using certain border strips called cutting strips, and
showed that all these formulas are equivalent up to simple matrix operations.
Recently, Jin \cite{Jin_2018} generalized the Hamel--Goulden formula by
considering more general border strips called thickened border strips.
\medskip
In this paper, inspired by the original proof of the Lascoux--Pragacz formula in
\cite[Section~2]{Lascoux1988}, we find determinant identities for Macdonald's
9th variation of Schur functions using the Bazin identity \cite{Bazin}. Our
results restricted to Schur functions are also new and include the
Hamel--Goulden formula \cite{Hamel_1995} and its generalization due to Jin
\cite{Jin_2018}. As an application we prove a determinant identity (after some
correction) for factorial Schur functions conjectured by Morales, Pak, and
Panova \cite{MPP2}. Another special case of our results is the Hamel--Goulden
formula for Macdonald's 9th variation of Schur functions. We note that recently
Bachmann and Charlton \cite{Bachmann_2020}, and Foley and King \cite{Foley20}
also proved this using the Lindstr\"om--Gessel--Viennot lemma.
To illustrate, we compare one of our results with the Lascoux--Pragacz formula
\cite{Lascoux1988}.
\begin{thm}[Lascoux--Pragacz formula]\label{thm:LP}
Suppose that $\mu\subseteq \lambda$ and $\theta=(\theta_1,\dots,\theta_k)$ is
the Lascoux--Pragacz decomposition of $\lm$. Then
\[
s_{\lambda/\mu} = \det \left( s_{\lambda^0[p_j,q_i]}\right)_{i,j=1}^k,
\]
where $p_i$'s and $q_i$'s are the contents of the starting and ending cells of
$\theta_i$, respectively.
\end{thm}
In the theorem above, the shape $\lambda^0[p_j,q_i]$ in each entry is a border
strip. See Sections~\ref{sec:definitions} and \ref{sec:lasc-prag-kreim} for the
undefined terms. The following theorem will be proved in
Section~\ref{sec:lasc-prag-kreim}.
\begin{thm}\label{thm:main_LP_intro}
Let $\lambda$, $\mu$, and $\nu$ be partitions. Suppose that $\mu\subseteq \lambda$
and $\theta=(\theta_1,\dots,\theta_k)$ is the Lascoux--Pragacz decomposition
of $\lm$. Then we have
\begin{align}
\label{eq:main1intro}
\ts_{\lambda/\nu}^{k-1} \ts_{\mu/\nu}
&= \det \left( (-1)^{\chi(p_j>q_i)} \ts_{\lambda(q_i,p_j-1)/\nu}\right)_{i,j=1}^k, \\
\label{eq:main2intro}
\ts_{\nu/\lambda}^{k-1} \ts_{\nu/\mu}
&= \det \left( (-1)^{\chi(p_j>q_i)} \ts_{\nu/\lambda(q_i,p_j-1)}\right)_{i,j=1}^k.
\end{align}
\end{thm}
Here $\ts_{\lm}$ is Macdonald's 9th variation of Schur functions and
$\lambda(q_i,p_j-1)$ is a partition obtained from $\lambda$ by either adding or
deleting a border strip. See Section~\ref{sec:definitions} for their precise
definitions. If $\nu=\lambda$ in \eqref{eq:main2intro} we obtain the
Lascoux--Pragacz formula for $\ts_{\lm}$, thus Theorem~\ref{thm:LP} follows. If
$\nu=\emptyset$ and Macdonald's 9th variation is specialized to the factorial
Schur functions in \eqref{eq:main1intro}, then we obtain a corrected version of
the conjecture of Morales, Pak, and Panova \cite{MPP2}.
Now we recall the reformulation of the Hamel--Goulden formula \cite{Hamel_1995}
due to Chen, Yan, and Yang \cite{Chen_2005}.
\begin{thm}[Hamel--Goulden formula]\label{thm:HG}
Let $\lambda$ and $\mu$ be partitions and let $\gamma$ be a border strip.
Suppose that $\mu\subseteq \lambda$ and $\theta=(\theta_1,\dots,\theta_k)$ is
the decomposition of $\lm$ determined by the cutting strip $\gamma$. Then we
have
\[
s_{\lambda/\mu} = \det \left( s_{\gamma[p_j,q_i]}\right)_{i,j=1}^k,
\]
where $p_i$ and $q_i$ are the contents of the starting and ending cells of
$\theta_i$, respectively.
\end{thm}
In Theorem~\ref{thm:HG}, $\gamma[p_j,q_i]$ is the set of cells in $\gamma$ whose
contents are in the closed interval $[p_j,q_i]$. See
Sections~\ref{sec:definitions} and \ref{sec:lasc-prag-kreim} for the undefined
terms. Observe that in Theorem~\ref{thm:HG} the $p_i$'s and $q_i$'s are
particular integers determined by $\lambda$, $\mu$, and $\gamma$. A natural
question is whether the determinant in this theorem can represent a skew Schur
function for arbitrary $p_i$'s and $q_i$'s. We show that this is indeed true up
to sign. This may be considered as a converse of the Hamel--Goulden theorem.
More generally, we can take $\gamma$ to be any connected skew shape, not
necessarily a border strip.
\begin{thm}\label{thm:HG3}
Let $\alpha$ be any connected skew shape. Suppose that $(a_1,\dots,a_k)$ and
$(b_1,\dots,b_k)$ are sequences of integers such that $\alpha[a_j,b_i]$ is a
skew shape for all $i$ and $j$. Then either $\det \left(
s_{\alpha[a_j,b_i]}\right)_{i,j=1}^k = 0$ or there exists a skew shape
$\rho$ such that
\[
\det \left( s_{\alpha[a_j,b_i]}\right)_{i,j=1}^k = \pm s_{\rho}.
\]
\end{thm}
\medskip
The rest of this paper is organized as follows. In Section~\ref{sec:definitions}
we give basic definitions. In Section~\ref{sec:bazin-sylv-ident} we derive a
determinant identity involving $\ts_{\lm}$ using the Bazin identity.
In Section~\ref{sec:lasc-prag-kreim} we restate the result in the previous
section using border strip decompositions. As an application we prove a
conjecture of Morales, Pak, and Panova \cite{MPP2}. In
Section~\ref{sec:gener-hamel-gould} we prove a generalization of the
Hamel--Goulden formula. In Section~\ref{sec:conv-hamel-gould} we prove Theorem
\ref{thm:HG3}.
\section{Definitions}
\label{sec:definitions}
In this section we give basic definitions which will be used throughout this paper.
A \emph{partition} is a weakly decreasing sequence
$\lambda=(\lambda_1,\dots,\lambda_k)$ of positive integers. Each $\lambda_i>0$
is called a \emph{part} of $\lambda$. The \emph{length} $\ell(\lambda)$ of
$\lambda$ is the number of parts in $\lambda$. For integers $r>\ell(\lambda)$,
we use the convention $\lambda_r=0$. The set of partitions with at most $n$
parts is denoted by $\Par_n$. By appending zeros at the end if necessary we will
write each element $\lambda\in \Par_n$ as $\lambda=(\lambda_1,\dots,\lambda_n)$.
A pair $(i,j)$ of integers is called a \emph{cell}. The \emph{Young diagram} of
a partition $\lambda$ is the set of cells $(i,j)$ with $1\le i\le \ell(\lambda)$
and $1\le j\le \lambda_i$. We will often identity a partition $\lambda$ with its
Young diagram. The Young diagram $\lambda$ is visualized as an array of squares
so that there is a square in row $i$ and column $j$ for each $\lambda\in (i,j)$,
according to the matrix coordinates. See Figure~\ref{fig:yd}.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.45]
\draw[thick] (0,0)--(1,0)--(1,1)--(3,1)--(3,2)--(4,2)--(4,3)--(0,3)--(0,0)--cycle;
\draw[thick] (0,1)--(1,1)
(0,2)--(3,2)
(1,1)--(1,3)
(2,1)--(2,3)
(3,2)--(3,3);
\node[] at (0,-.5) {\phantom{1}};
\end{tikzpicture}\qquad\qquad
\begin{tikzpicture}[scale=.45]
\draw[thick] (0,0)--(2,0)--(2,2)--(1,2)--(1,1)--(0,1)--(0,0)--cycle
(4,3)--(5,3)--(5,4)--(4,4)--(4,3)--cycle
(1,0)--(1,1)--(2,1);
\draw[dashed] (0,1)--(0,4)
(1,2)--(1,4)
(2,2)--(2,4)
(3,2)--(3,4)
(0,4)--(4,4)
(0,3)--(4,3)
(0,2)--(1,2)
(2,2)--(4,2)--(4,3);
\end{tikzpicture}\qquad\qquad
\begin{tikzpicture}[scale=.45]
\draw[thick] (0,0)--(2,0)--(2,2)--(1,2)--(1,1)--(0,1)--(0,0)--cycle
(4,3)--(5,3)--(5,4)--(4,4)--(4,3)--cycle
(1,0)--(1,1)--(2,1);
\draw[dashed] (0,1)--(0,4)
(1,2)--(1,4)
(2,2)--(2,4)
(0,4)--(4,4)
(0,3)--(4,3)
(0,2)--(1,2)
(2,2)--(3,2)--(3,3)--(4,3)
(3,3)--(3,4);
\end{tikzpicture}
\caption{The Young diagram of the partition $(4,3,1)$ (left), the skew shape
$(5,4,2,2)/(4,4,1)$ (middle), and the skew shape $(5,3,2,2)/(4,3,1)$
(right).}
\label{fig:yd}
\end{figure}
For two partitions $\lambda$ and $\mu$, we write $\mu\subseteq\lambda$ to mean
that the Young diagram of $\mu$ is contained in that of $\lambda$. A \emph{skew
shape}, denoted by $\lm$, is a pair $(\lambda,\mu)$ of partitions satisfying
$\mu\subseteq\lambda$. We also consider the skew shape $\lm$ as the
set-theoretic difference $\lambda\setminus\mu$ of their Young diagrams. Note,
however, that when we consider a skew shape $\lm$ we have the information on the
partitions $\lambda$ and $\mu$ as well as the difference $\lambda-\mu$ of their
Young diagrams. For example, the two skew shapes in Figure~\ref{fig:yd} have the
same set of cells but are considered as different skew shapes.
For a cell $x=(i,j)$, the \emph{content} $c(x)$ of $x$ is defined by $c(x)=j-i$.
For a skew shape $\alpha$, we define
\[
\Cont(\alpha) = \{c(x):x\in \alpha\}.
\]
For a skew shape $\alpha$ we define $\alpha+(r,s)$ to be the skew shape obtained
by shifting $\alpha$ by $(r,s)$, i.e.,
\[
\alpha + (r,s) = \{x+(r,s): x\in \alpha\}.
\]
In this paper ``connected'' means edgewise connected. A \emph{connected
component} of a skew shape $\alpha$ is a maximal connected subset of $\alpha$,
see Figure~\ref{fig:connected}.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.45]
\draw[thick] (0,0)--(0,2)--(2,2)--(2,3)--(3,3)--(3,4)--(6,4)--(6,1)--(4,1)--(4,0)--(0,0)
(3,3)--(6,3)
(2,2)--(6,2)
(0,1)--(4,1)
(1,0)--(1,2)
(2,0)--(2,2)
(3,0)--(3,3)
(4,1)--(4,4)
(5,1)--(5,4);
\draw[dashed] (0,2)--(0,4)--(3,4)
(0,3)--(2,3)
(1,2)--(1,4)
(2,3)--(2,4);
\node[] at (0,-.5) {\phantom{1}};
\end{tikzpicture}\qquad\qquad\qquad
\begin{tikzpicture}[scale=.45]
\draw[thick] (0,0)--(0,2)--(2,2)--(2,3)--(3,3)--(3,1)--(2,1)--(2,0)--(0,0)--cycle
(0,1)--(2,1)--(2,2)--(3,2)
(1,0)--(1,2)
(3,3)--(5,3)--(5,5)--(4,5)--(4,4)--(3,4)--(3,3)
(4,3)--(4,4)--(5,4);
\draw[dashed] (0,2)--(0,5)--(4,5)
(0,4)--(3,4)
(0,3)--(2,3)
(1,2)--(1,5)
(2,3)--(2,5)
(3,4)--(3,5);
\end{tikzpicture}
\caption{The skew shape on the left is connected, whereas the skew shape on
the right is disconnected. There are two connected components in the right
skew shape.}
\label{fig:connected}
\end{figure}
A \emph{border strip} is a connected skew shape that contains no $2\times 2$ block of
squares. For a border strip $\gamma$, the contents of the starting and ending
cells of $\gamma$ are denoted by $p(\gamma)$ and $q(\gamma)$, respectively:
\[
p(\gamma)=\min(\Cont(\gamma)),\qquad q(\gamma)=\max(\Cont(\gamma)).
\]
For a border strip $\gamma$ and integers $a$ and $b$, we define
\[
\gamma[a,b] =
\begin{cases}
\{x\in \gamma: a\le c(x)\le b\} & \mbox{if $a\le b$},\\
\emptyset & \mbox{if $a=b+1$}, \\
\mbox{undefined} & \mbox{if $a>b+1$}.
\end{cases}
\]
See Figure~\ref{fig:gamma}. If $\gamma[a,b]$ is undefined, then we define
$\ts_{\gamma[a,b]}$ to be $0$.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.5]
\draw[densely dashed] (0,1)--(0,4);
\draw[densely dashed] (1,1)--(1,4);
\draw[densely dashed] (2,2)--(2,4);
\draw[densely dashed] (3,3)--(3,4);
\draw[densely dashed] (4,3)--(4,4);
\draw[densely dashed] (0,4)--(4,4);
\draw[densely dashed] (0,3)--(4,3);
\draw[densely dashed] (0,2)--(2,2);
\draw[densely dashed] (0,1)--(1,1);
\draw[thick] (0,0)--(0,1)--(1,1)--(1,2)--(3,2)--(3,3)--(4,3)--(4,4)--(5,4)--(5,3)--(4,3)--(4,1)--(2,1)--(2,0)--(0,0)--cycle;
\draw[thick] (4,2)--(5,2)--(5,3)--(6,3)--(6,4)--(5,4);
\draw[thick] (1,1)--(4,1);
\draw[thick] (2,2)--(4,2);
\draw[thick] (1,0)--(1,1);
\draw[thick] (2,1)--(2,2);
\draw[thick] (3,1)--(3,3);
\node[] at (4.5, 3.5) {$4$};
\node[] at (5.5, 3.5) {$5$};
\node[] at (3.5, 2.5) {$2$};
\node[] at (4.5, 2.5) {$3$};
\node[] at (1.5, 1.5) {$-1$};
\node[] at (2.5, 1.5) {$0$};
\node[] at (3.5, 1.5) {$1$};
\node[] at (.5, .5) {$-3$};
\node[] at (1.5, .5) {$-2$};
\node[] at (-1.6,2) { $ \gamma=$};
\end{tikzpicture} \qquad
\begin{tikzpicture}[scale=.5]
\draw[thick] (0,0)-- (0,2)--(2,2)--(2,3)--(4,3)--(4,2)--(3,2)--(3,1)--(1,1)--(1,0)--(0,0)--cycle;
\draw[thick] (0,1)--(1,1)
(1,1)--(1,2)
(2,1)--(2,2)
(2,2)--(3,2)
(3,2)--(3,3);
\node[] at (.5, .5) {$-2$};
\node[] at (.5, 1.5) {$-1$};
\node[] at (1.5, 1.5) {$0$};
\node[] at (2.5, 1.5) {$1$};
\node[] at (2.5, 2.5) {$2$};
\node[] at (3.5, 2.5) {$3$};
\node[] at (-2.5, 2) {$\gamma [-2, 3] $=};
\end{tikzpicture}
\caption{The left diagram shows a border strip $\gamma$ with $p(\gamma)=-3$ and
$q(\gamma)=5$. The right diagram shows $\gamma[-2,3]$. In both diagrams the
contents of the cells are shown. We have $\gamma[3,2]=\emptyset$ and
$\gamma[2,-1]$ is undefined.}
\label{fig:gamma}
\end{figure}
Consider a skew shape $\alpha$ and a border strip $\gamma$ such that
$\Cont(\alpha)\subseteq\Cont(\gamma)$. Consider the diagonal shifts
$\gamma+(i,i)$ of $\gamma$, for $i\in\ZZ$, that cover $\alpha$. The intersection
of each diagonal shift of $\gamma$ with $\alpha$ is a union of border strips.
Let $\theta$ be the collection of the border strips obtained in this way. Then
$\theta$ is a decomposition of $\alpha$. In this case we say that $\gamma$ is
the \emph{cutting strip} of $\theta$. See Figure~\ref{fig:cutting} for an
example.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.34]
\fill[fill=yellow!30!white] (-4,3) rectangle (1,3.9)
(.1, 3.9) rectangle (1, 4)
(.1,4) rectangle (4,4.85)
(3.1,4.85) rectangle (4,9);
\fill[fill=asparagus!30!white] (-3,2) rectangle (2,2.9)
(1.1,2.9) rectangle (5,3.85)
(4.1,3.85) rectangle (5,8);
\fill[fill=cyan!30!white] (-2,1) rectangle (3,1.9)
(2.1,1.9) rectangle (6,2.85)
(5.1,2.85) rectangle (6,7);
\fill[fill=pink] (-1,0) rectangle (4,.9)
(3.1, .9) rectangle (4, 1)
(3.1,1) rectangle (7,1.85)
(6.1,1.85) rectangle (7,6);
\fill[fill=orange!35!white] (0,-1) rectangle (5,-.1)
(4.1,-.1) rectangle (8,.85)
(7.1,.85) rectangle (8,5);
\draw[thick] (0,0)--(4,0)--(4,1)--(7,1)--(7,3)--(8,3)--(8,5)--(3,5)--(3,4)--(2,4)--(2,3)--(0,3)--(0,0)--cycle;
\end{tikzpicture}\qquad\qquad\quad
\begin{tikzpicture}[scale=.34]
\draw[thick] (0,0)--(4,0)--(4,1)--(7,1)--(7,3)--(8,3)--(8,5)--(3,5)--(3,4)--(2,4)--(2,3)--(0,3)--(0,0)--cycle;
\draw[thick] (7,3)--(7,5);
\draw[thick] (0,1)--(3,1)--(3,2)--(6,2)--(6,5);
\draw[thick] (0,2)--(2,2)--(2,3)--(5,3)--(5,5);
\draw[thick] (3,4)--(4,4)--(4,5);
\end{tikzpicture}
\caption{The left diagram shows the diagonal shifts of a border strip $\gamma$
covering a skew shape $\alpha$. The right diagram shows the decomposition
$\theta$ with respect to the cutting strip $\gamma$.}
\label{fig:cutting}
\end{figure}
For a partition $\lambda$ with at most $n$ parts, let
\[
C_n(\lambda)=\{\lambda_i-i: 1\le i\le n\}.
\]
See Figure~\ref{fig:C}. Note that for any partition $\lambda\in\Par_n$, we have
$|C_n(\lambda)|=n$ and $\min(C_n(\lambda))\ge -n$. Conversely, one can easily
see that for any set $C$ of integers with $|C|=n$ and $\min(C)\ge -n$, there is
a unique partition $\lambda\in\Par_n$ with $C_n(\lambda)=C$.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.5]
\draw[thick] (0,0)--(0,6)--(6,6)--(6,4)--(4,4)--(4,3)--(2,3)--(2,2)--(0,2);
\draw[color=gray] (0,5)--(6,5)
(0,4)--(4,4)
(0,3)--(2,3)
(1,2)--(1,6)
(2,3)--(2,6)
(3,3)--(3,6)
(4,4)--(4,6)
(5,4)--(5,6);
\node[] at (5.5, 5.5) {$5$};
\node[] at (4.5, 5.5) {$4$};
\node[] at (3.5, 5.5) {$3$};
\node[] at (2.5, 5.5) {$2$};
\node[] at (1.5, 5.5) {$1$};
\node[] at (.5, 5.5) {$0$};
\node[] at (-.5, 5.5) {$-1$};
\node[] at (5.5, 4.5) {$4$};
\node[] at (4.5, 4.5) {$3$};
\node[] at (3.5, 4.5) {$2$};
\node[] at (2.5, 4.5) {$1$};
\node[] at (1.5, 4.5) {$0$};
\node[] at (.5, 4.5) {$-1$};
\node[] at (-.5, 4.5) {$-2$};
\node[] at (3.5, 3.5) {$1$};
\node[] at (2.5, 3.5) {$0$};
\node[] at (1.5, 3.5) {$-1$};
\node[] at (.5, 3.5) {$-2$};
\node[] at (-.5, 3.5) {$-3$};
\node[] at (1.5, 2.5) {$-2$};
\node[] at (.5, 2.5) {$-3$};
\node[] at (-.5, 2.5) {$-4$};
\node[] at (-.5, 1.5) {$-5$};
\node[] at (-.5, .5) {$-6$};
\draw[thick, dredcolor] (-.5, .5) circle (13pt);
\draw[thick, dredcolor] (-.5, 1.5) circle (13pt);
\draw[thick, dredcolor] (1.5, 2.5) circle (13pt);
\draw[thick, dredcolor] (3.5, 3.5) circle (13pt);
\draw[thick, dredcolor] (5.5, 4.5) circle (13pt);
\draw[thick, dredcolor] (5.5, 5.5) circle (13pt);
\end{tikzpicture}
\caption{The Young diagram of $\lambda=(6,6,4,2)$, where the contents of the
cells in $\lambda\cup\{(i,0):1\le i\le 6\}$ are shown. The circled integers
are the elements in $C_6(\lambda)$.}
\label{fig:C}
\end{figure}
\begin{defn}
The \emph{outer strip} $\lambda^0$ of a partition $\lambda$ is the set of
cells $x\in \lambda$ such that $x+(1,1)\not\in \lambda$. The \emph{extended
outer strip} $\lambda^+$ of $\lambda$ is the (infinite) set of cells $x\in
\ZZ_{>0}^2\setminus\lambda$ such that $x+(-1,-1)\not\in
\ZZ_{>0}^2\setminus\lambda$. See Figure~\ref{fig:outer2}.
\end{defn}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.48]
\draw[color=gray] (0,6)--(4,6)
(0,7)--(6,7)
(0,8)--(7,8)
(1,5)--(1,9)
(2,5)--(2,9)
(3,5)--(3,9)
(4,6)--(4,9)
(5,6)--(5,9)
(6,7)--(6,9);
\draw[thick] (0,5)--(4,5)--(4,6)--(6,6)--(6,7)--(7,7)--(7,9)--(0,9)--(0,5)--cycle;
\fill[fill=pink!40!white] (.1, 5.1) rectangle (3.9,6)
(3, 6.1) rectangle (5.9, 7)
(5, 7) rectangle (5.95, 8)
(6, 7.1) rectangle (6.9, 8.9);
\draw[very thick, color=dredcolor] (0.1,5.1)--(3.9,5.1)--(3.9,6.1)--(5.9, 6.1)--(5.9,7.1)--(6.9,7.1)--(6.9,8.9)--(6,8.9)--(6,8)--(5,8)--(5,7)--(3,7)--(3,6)--(.1, 6)--(.1, 5.1)--cycle
(1,5.1)--(1,6)
(2,5.1)--(2,6)
(3,5.1)--(3,6)
(3,6.05)--(3.9, 6.05)
(3.95, 6.1)--(3.95, 7)
(5, 6.1)--(5, 7)--(5.9, 7)
(5.95, 7.1)--(5.95, 8)--(6.9, 8);
\fill[fill=cyan!8!white] (0, .5) rectangle (1, 4.9)
(1, 4) rectangle (5, 4.9)
(4.1, 4.9) rectangle (7.1, 5.9)
(6.1, 5.9) rectangle (7.1, 6.9)
(7.1, 5.9) rectangle (8.1,9)
(8.1, 8) rectangle (11.5, 9);
\draw[very thick, color=frenchblue] (0, .5)--(0, 4.9)--(4.1,4.9)--(4.1, 5.9)--(6.1, 5.9)--(6.1, 6.9)--(7.1, 6.9)--(7.1, 9)--(11.5, 9)
(1, .5)--(1,4.9)
(0, 4)--(5,4)--(5, 5.9)
(4.1, 4.9)--(7.1,4.9)--(7.1, 6.9)
(6.1, 5.9)--(8.1, 5.9)--(8.1, 9)
(7.1, 8)--(11.5, 8)
(0,1)--(1,1)
(0,2)--(1,2)
(0,3)--(1,3)
(2,4)--(2,4.9)
(3.05,4)--(3.05,4.9)
(4.1, 4)--(4.1, 4.9)
(6.1, 4.9)--(6.1, 5.9)
(7.1, 6.95)--(8.1,6.95)
(9.1, 8)--(9.1,9)
(10.1, 8)--(10.1,9)
(11.1, 8)--(11.1,9);
\node [] at (.5, 5.5) {$-3$};
\node [] at (1.5, 5.5) {$-2$};
\node [] at (2.5, 5.5) {$-1$};
\node [] at (3.5, 5.5) {$0$};
\node [] at (3.5, 6.5) {$1$};
\node [] at (4.5, 6.5) {$2$};
\node [] at (5.5, 6.5) {$3$};
\node [] at (5.5, 7.5) {$4$};
\node [] at (6.5, 7.5) {$5$};
\node [] at (6.5, 8.5) {$6$};
\node [] at (.5, .6) {$\vdots$};
\node [] at (.5, 1.4) {$-7$};
\node [] at (.5, 2.4) {$-6$};
\node [] at (.5, 3.4) {$-5$};
\node [] at (.5, 4.4) {$-4$};
\node [] at (1.5, 4.4) {$-3$};
\node [] at (2.5, 4.4) {$-2$};
\node [] at (3.6, 4.4) {$-1$};
\node [] at (4.6, 4.4) {$0$};
\node [] at (4.6, 5.4) {$1$};
\node [] at (5.6, 5.4) {$2$};
\node [] at (6.6, 5.4) {$3$};
\node [] at (6.6, 6.4) {$4$};
\node [] at (7.6, 6.4) {$5$};
\node [] at (7.6, 7.4) {$6$};
\node [] at (7.6, 8.5) {$7$};
\node [] at (8.6, 8.5) {$8$};
\node [] at (9.6, 8.5) {$9$};
\node [] at (10.6, 8.5) {$10$};
\node [] at (11.6, 8.5) {$\cdots$};
\end{tikzpicture}\quad
\begin{tikzpicture}[scale=.5]
\fill[fill=cyan!8!white] (0,0) rectangle (1,2)
(1,1) rectangle (5,2)
(4,2) rectangle (6,3);
\draw[very thick, color=frenchblue] (0,0)--(1,0)--(1,1)--(5,1)--(5,2)--(6,2)--(6,3)--(4,3)--(4,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)
(4,1)--(4,2)--(5,2)--(5,3);
\node [] at (.5, .5) {$-5$};
\node [] at (.5, 1.5) {$-4$};
\node [] at (1.5, 1.5) {$-3$};
\node [] at (2.5, 1.5) {$-2$};
\node [] at (3.5, 1.5) {$-1$};
\node [] at (4.5, 1.5) {$0$};
\node [] at (4.5, 2.5) {$1$};
\node [] at (5.5, 2.5) {$2$};
\fill[fill=pink!40!white] (1,5) rectangle (5,6)
(4,6) rectangle (6,7);
\draw[very thick, color=dredcolor] (1,5)--(5,5)--(5,6)--(6,6)--(6,7)--(4,7)--(4,6)--(1,6)--(1,5)--cycle
(2,5)--(2,6)
(3,5)--(3,6)
(4,5)--(4,6)--(5,6)--(5,7);
\node [] at (1.5, 5.5) {$-3$};
\node [] at (2.5, 5.5) {$-2$};
\node [] at (3.5, 5.5) {$-1$};
\node [] at (4.5, 5.5) {$0$};
\node [] at (4.5, 6.5) {$1$};
\node [] at (5.5, 6.5) {$2$};
\node[] at (0,-1) {\phantom{1}};
\end{tikzpicture}
\caption{The left diagram shows the partition $\lambda=(7,7,6,4)$, where its
outer strip $\lambda^0$ and extended outer strip $\lambda^+$ are colored red
and blue, respectively. On the right are shown $\lambda^0[-5,2]$ at the top
and $\lambda^+[-5,2]$ at the bottom.}
\label{fig:outer2}
\end{figure}
Note that by definition, for $p\le q$, we always have
$\lambda^0[p,q]\subseteq\lambda$, but $\lambda^+[p,q]\not\subseteq\lambda$.
\begin{defn}
Consider a partition $\lambda\in\Par_n$ and integers $a,b\ge -n$ such that
$a\in C_n(\lambda)$ and $b\not\in C_n(\lambda)\setminus\{a\}$. We define
$\lambda(a,b)$ to be the unique partition $\mu\in\Par_n$ satisfying
\begin{equation}
\label{eq:C_n}
C_n(\mu) =(C_n(\lambda) \setminus\{a\}) \cup \{b\}.
\end{equation}
\end{defn}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.48]
\fill[fill=pink!50!white] (3, 3) rectangle (4, 5);
\fill[fill=pink!50!white] (3, 4) rectangle (6, 5);
\draw[color=gray] (0,5)--(6,5)
(0,4)--(3.9,3.9)
(0,3)--(2,3)
(1,2)--(1,6)
(2,3)--(2,6)
(3,3)--(3,6)
(4,4.1)--(4,6)
(5,4.1)--(5,6);
\node[] at (5.5, 5.5) {$5$};
\node[] at (4.5, 5.5) {$4$};
\node[] at (3.5, 5.5) {$3$};
\node[] at (2.5, 5.5) {$2$};
\node[] at (1.5, 5.5) {$1$};
\node[] at (.5, 5.5) {$0$};
\node[] at (-.6, 5.5) {$-1$};
\node[] at (5.5, 4.5) {$4$};
\node[] at (4.5, 4.5) {$3$};
\node[] at (3.5, 4.5) {$2$};
\node[] at (2.5, 4.5) {$1$};
\node[] at (1.5, 4.5) {$0$};
\node[] at (.5, 4.5) {$-1$};
\node[] at (-.6, 4.5) {$-2$};
\node[] at (3.5, 3.5) {$1$};
\node[] at (2.5, 3.5) {$0$};
\node[] at (1.5, 3.5) {$-1$};
\node[] at (.5, 3.5) {$-2$};
\node[] at (-.6, 3.5) {$-3$};
\node[] at (1.5, 2.5) {$-2$};
\node[] at (.5, 2.5) {$-3$};
\node[] at (-.6, 2.5) {$-4$};
\node[] at (-.6, 1.5) {$-5$};
\node[] at (-.6, .5) {$-6$};
\draw[thick, color=dredcolor!60!white] (3,3)--(4,3)--(4, 4)--(6, 4)--(6, 5)--(3, 5)--(3,3)--cycle;
\draw[thick] (0,0)--(0,6)--(6,6)--(6,5)--(3,5)--(3,3)--(2,3)--(2,2)--(0,2);
\draw[thick, dredcolor] (-.6, .5) circle (13pt);
\draw[thick, dredcolor] (-.6, 1.5) circle (13pt);
\draw[thick, dredcolor] (1.5, 2.5) circle (13pt);
\draw[thick, dredcolor] (2.5, 3.5) circle (13pt);
\draw[thick, dredcolor] (2.5, 4.5) circle (13pt);
\draw[thick, dredcolor] (5.5, 5.5) circle (13pt);
\node[] at (5.6, 3.5) {\small\color{red}removed};
\end{tikzpicture}\qquad\qquad\quad
\begin{tikzpicture}[scale=.48]
\fill[fill=cyan!14!white] (0,1) rectangle (3,2);
\fill[fill=cyan!14!white] (2,2) rectangle (5,3);
\fill[fill=cyan!14!white] (4,3) rectangle (5,4);
\draw[thick, cyan] (0,1)--(3,1)--(3,2)--(5,2)--(5,4)--(4,4)--(4,3)--(2,3)--(2,2)--(0,2)--(0,1)--cycle;
\draw[thick] (0,0)--(0,6)--(6,6)--(6,4)--(5,4)--(5,2)--(3,2)--(3,1)--(0,1);
\draw[color=gray] (0,5)--(6,5)
(0,4)--(5,4)
(0,3)--(5,3)
(0,2)--(3,2)
(1,1)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,4)--(5,6);
\node[] at (5.5, 5.5) {$5$};
\node[] at (4.5, 5.5) {$4$};
\node[] at (3.5, 5.5) {$3$};
\node[] at (2.5, 5.5) {$2$};
\node[] at (1.5, 5.5) {$1$};
\node[] at (.5, 5.5) {$0$};
\node[] at (-.6, 5.5) {$ -1$};
\node[] at (5.5, 4.5) {$4$};
\node[] at (4.5, 4.5) {$3$};
\node[] at (3.5, 4.5) {$2$};
\node[] at (2.5, 4.5) {$1$};
\node[] at (1.5, 4.5) {$0$};
\node[] at (.5, 4.5) {$-1$};
\node[] at (-.6, 4.5) {$-2$};
\node[] at (4.5, 3.5) {$2$};
\node[] at (3.5, 3.5) {$1$};
\node[] at (2.5, 3.5) {$0$};
\node[] at (1.5, 3.5) {$-1$};
\node[] at (.5, 3.5) {$-2$};
\node[] at (-.6, 3.5) {$-3$};
\node[] at (1.5, 2.5) {$-2$};
\node[] at (.5, 2.5) {$-3$};
\node[] at (-.6, 2.5) {$-4$};
\node[] at (-.6, 1.5) {$-5$};
\node[] at (-.6, .5) {$-6$};
\node[] at (2.5, 1.5) {$-2$};
\node[] at (1.5, 1.5) {$-3$};
\node[] at (.5, 1.5) {$-4$};
\node[] at (4.5, 2.5) {$1$};
\node[] at (3.5, 2.5) {$0$};
\node[] at (2.5, 2.5) {$-1$};
\draw[thick, dredcolor] (-.6, .5) circle (13pt);
\draw[thick, dredcolor] (2.5, 1.5) circle (13pt);
\draw[thick, dredcolor] (4.5, 2.5) circle (13pt);
\draw[thick, dredcolor] (4.5, 3.5) circle (13pt);
\draw[thick, dredcolor] (5.5, 4.5) circle (13pt);
\draw[thick, dredcolor] (5.5, 5.5) circle (13pt);
\node[] at (4.3, 1.5) {\small\color{blue}added};
\end{tikzpicture}
\caption{For the partition $\lambda=(6,6,4,2)$ in Figure~\ref{fig:C}, the left
is $\lambda(4,0)=\lambda\setminus \lambda^0[1,4]$ and the right is
$\lambda(-5,2)=\lambda\cup \lambda^+ [-4,2]$.}
\label{fig:la(a,b)}
\end{figure}
See Figure~\ref{fig:la(a,b)}. It is easy to check that
under the above conditions on $a$ and $b$,
the partition $\lambda(a,b)$ is well defined and
\begin{equation}\label{eq:la(a,b)}
\lambda(a,b) =
\begin{cases}
\lambda & \mbox{if $a=b$},\\
\lambda\setminus\lambda^0[b+1,a] & \mbox{if $a>b$}, \\
\lambda\cup\lambda^+[a+1,b] & \mbox{if $a<b$}.
\end{cases}
\end{equation}
\begin{remark}
The notations $\lambda^0[p,q]$ and $\lambda(p,q)$ will be used frequently
throughout this paper. Note that $\lambda^0[p,q]$ is a border strip and
$\lambda(p,q)$ is a partition obtained from $\lambda$ by adding or deleting a
(possibly empty) border strip. The notation $\lambda/\mu(a,b)$ always means
$\lambda/(\mu(a,b))$ and not $(\lambda/\mu)(a,b)$.
\end{remark}
Finally, we define Macdonald's 9th variation of Schur functions.
\begin{defn}
Let $h_{r,s}$, for $r, s\in\ZZ$ with $r\ge1$, be independent indeterminates.
Define $h_{0,s}=1$ and $h_{r,s}=0$ for all $r<0$ and $s\in \ZZ$. For any
partitions $\lambda$ and $\mu$ with at most $n$ parts, the \emph{Macdonald's
9th variation} $\ts_{\lm}$ of Schur functions is defined by
\[
\ts_{\lm} = \det\left(h_{\lambda_i-\mu_j-i+j,\mu_j-j+1}\right)_{i,j=1}^n.
\]
\end{defn}
If $h_{r,s}$ is set to be equal to the complete homogeneous symmetric function
$h_r$ for all $r$ and $s$, then $\ts_{\lm}$ reduces to the Schur function
$s_{\lm}$. Note that the Schur functions have the property that
$s_{\alpha}=s_{\alpha+(r,s)}$ for any skew shape $\alpha$ and integers $r,s$
such that $\alpha+(r,s)$ is a skew shape. However,
$\ts_{\alpha}\ne\ts_{\alpha+(r,s)}$ unless $r=s$.
\section{The Bazin identity}
\label{sec:bazin-sylv-ident}
In this section we recall the Bazin identity and derive
a determinant identity for Macdonald's 9th variation of Schur functions from it.
\begin{defn}
Let $M=(M_{ij})_{i\in \ZZ, 1\le j\le n}$ be a matrix whose rows and columns are indexed by $\ZZ$ and $\{1,2,\dots,n\}$ respectively.
For any sequence $\vec a=(a_1,\dots,a_n)$ of $n$ integers (not necessarily distinct), we define
\[
M[\vec a] = \det (M_{a_i,j})_{1\le i,j\le n}.
\]
Note that if $\vec a$ has repeated elements, then $[\vec a]=0$, and otherwise $[\vec a]$ is, up to sign, equal to the minor of $M$ obtained by selecting the rows of $M$ indexed by the integers in $\vec a$.
\end{defn}
For sequences $\vec a = (a_1,\dots,a_r)$ and $\vec b = (b_1,\dots,b_s)$ of
integers, let
\begin{align*}
\vec a \sqcup \vec b &= (a_1,\dots,a_r,b_1,\dots,b_s),\\
\vec a \setminus a_i &= (a_1,\dots,a_{i-1},a_{i+1},\dots,a_r).
\end{align*}
An integer $b$ is also considered as the sequence $(b)$ consisting of only one
element. For example,
\[
b \sqcup \vec a= (b)\sqcup \vec a = (b,a_1,\dots,a_r).
\]
The key lemma in this paper is the following result proved by Bazin \cite{Bazin}
in 1851, see also \cite[Lemma~2.1]{Lascoux1988}.
\begin{lem}[Bazin identity]\label{lem:BS}
Let $\vec a=(a_1,\dots,a_k), \vec b=(b_1,\dots,b_k)$ and $\vec c=(c_1,\dots,c_{n-k})$ be any sequences of integers. Then
\[ [\vec a \sqcup \vec c]^{k-1} [\vec b\sqcup\vec c] = (-1)^{\binom k2}\det
([b_j\sqcup (\vec a\setminus a_i)\sqcup \vec c])_{i,j=1}^k.
\]
\end{lem}
\begin{remark}
The original statement of the Bazin identity is
\[
[\vec a \sqcup \vec c]^{k-1} [\vec b\sqcup\vec c]
= \det ([(a_1,\dots,a_{i-1},b_j,a_{i+1},\dots,a_k) \sqcup \vec c])_{i,j=1}^k.
\]
Since
\[
[(a_1,\dots,a_{i-1},b_j,a_{i+1},\dots,a_k) \sqcup \vec c] =
(-1)^{i-1}[b_j\sqcup (\vec a\setminus a_i)\sqcup \vec c],
\]
the above identity is equivalent to the one in Lemma~\ref{lem:BS}. The Bazin
identity is also attributed to Sylvester (see \cite[p.~563]{Lascoux1988}), Reiss,
and Picquet (see \cite[p.~195]{Doubilet_1974}).
We note that Okada~\cite[Corollary~3.2]{okada17:sylvester} found a more general
determinant identity and derived the Bazin identity from it.
\end{remark}
For a permutation $\pi$ of $\{1,2,\dots,n\}$ we denote by $\inv(\pi)$ the number
of pairs $(i,j)$ of integers $1\le i<j\le n$ satisfying $\pi(i)>\pi(j)$. For two
sequences $\vec x = (x_1,\dots, x_r)$ and $\vec y = (y_1,\dots,y_s)$ of
integers, let $\inv(\vec x, \vec y)$ denote the number of pairs $(i,j)$ of
integers $1\le i\le r$ and $1\le j\le s$ satisfying $x_i>y_j$.
For a statement $p$ we define $\chi(p)=1$ if $p$ is true and $\chi(p)=0$
otherwise.
Now we derive a determinant identity for Macdonald's 9th variation of Schur
functions using the Bazin identity. In later sections we will give
some applications of this identity.
\begin{thm}\label{thm:main}
Let $\lambda$, $\mu$, and $\nu$ be partitions with at most $n$ parts. Suppose
that $\vec a=(a_1,\dots,a_k)$ and $\vec b=(b_1,\dots,b_k)$ are the sequences
defined by
\begin{align*}
C_n(\lambda)\setminus C_n(\mu) &= \{a_1>a_2>\dots>a_k\},\\
C_n(\mu)\setminus C_n(\lambda) &= \{b_1>b_2>\dots>b_k\}.
\end{align*}
Then we have
\begin{equation}
\label{eq:main1}
\ts_{\lambda/\nu}^{k-1} \ts_{\mu/\nu} =(-1)^{\inv(\vec b,\vec a)+\binom k2}\det \left( (-1)^{\chi(b_j>a_i)} \ts_{\lambda(a_i,b_j)/\nu}\right)_{i,j=1}^k,
\end{equation}
and
\begin{equation}
\label{eq:main2}
\ts_{\nu/\lambda}^{k-1} \ts_{\nu/\mu} =(-1)^{\inv(\vec b,\vec a) +\binom k2}\det \left( (-1)^{\chi(b_j>a_i)} \ts_{\nu/\lambda(a_i,b_j)}\right)_{i,j=1}^k.
\end{equation}
\end{thm}
\begin{proof}
Let
\[
M=\left(h_{i-\nu_j+j,\nu_j-j+1}\right)_{i\in\ZZ, 1\le j\le n}
\]
and $\vec c =(c_1,\dots,c_{n-k})$, where
$c_1>\cdots>c_{n-k}$ are the elements of $C_n(\lambda)\setminus \vec a = C_n(\mu)\setminus \vec b$ written in decreasing order. Then Lemma~\ref{lem:BS} says that
\begin{equation}
\label{eq:2}
[\vec a \sqcup \vec c]^{k-1} [\vec b\sqcup\vec c] = (-1)^{\binom k2}\det ([b_j\sqcup (\vec a\setminus a_i)\sqcup \vec c])_{i,j=1}^k.
\end{equation}
Since $\vec a \sqcup \vec c$ is a rearrangment of the decreasing sequence
$(\lambda_1-1,\lambda_2-2, \dots,\lambda_n-n)$, we have
\begin{equation}
\label{eq:5}
[\vec a \sqcup\vec c] = (-1)^{\inv(\vec c,\vec a)} \det(h_{\lambda_i-i - \nu_j+j,\nu_j-j+1})_{i,j=1}^n
= (-1)^{\inv(\vec c,\vec a)} \ts_{\lambda/\nu}.
\end{equation}
Similarly, we have
\begin{equation}
\label{eq:6}
[\vec b\sqcup\vec c] = (-1)^{\inv(\vec c,\vec b)} \ts_{\mu/\nu}.
\end{equation}
Now, we claim that
\begin{equation}
\label{eq:3}
[b_j\sqcup (\vec a\setminus a_i)\sqcup \vec c] =
(-1)^{\inv(\vec c, \vec a) - \inv(\vec c, a_i)+\inv(\vec c, b_j) + \inv(\vec a, b_j)+\chi(a_i>b_j)} \ts_{\lambda(a_i,b_j)/\nu}.
\end{equation}
If $b_j\in \vec a\setminus a_i$, both sides of \eqref{eq:3} are zero. Suppose
$b_j\not\in \vec a\setminus a_i$. Then $b_j\sqcup (\vec a\setminus a_i)\sqcup
\vec c$ is a rearrangment of $(\rho_1-1,\rho_2-2,\dots,\rho_n-n)$, where
$\rho=(\rho_1,\dots,\rho_n)$ is the partition $\lambda(a_i,b_j)$. Thus
\[
[b_j\sqcup (\vec a\setminus a_i)\sqcup \vec c] = (-1)^t \ts_{\lambda(a_i,b_j)/\nu},
\]
where
\begin{align*}
t &= \inv(\vec a\setminus a_i, b_j) +\inv(\vec c, b_j) + \inv(\vec c, \vec a\setminus a_i)\\
&=\inv(\vec a, b_j) - \inv(a_i, b_j) +\inv(\vec c, b_j) + \inv(\vec c, \vec a) - \inv(\vec c, a_i).
\end{align*}
Since $\inv(a_i, b_j) = \chi(a_i>b_j)$, we obtain \eqref{eq:3}.
By factoring out common factors from each row and each column using \eqref{eq:3}, we get
\begin{equation}
\label{eq:1}
\det ([b_j\sqcup (\vec a\setminus a_i)\sqcup \vec c])_{i,j=1}^k
= (-1)^{ (k-1)\inv(\vec c,\vec a) + \inv(\vec c, \vec b)+\inv(\vec a, \vec b) } \det ((-1)^{\chi(a_i>b_j)} \ts_{\lambda(a_i,b_j)/\nu}).
\end{equation}
By substituting \eqref{eq:5}, \eqref{eq:6} and \eqref{eq:1} to \eqref{eq:2}, we obtain
\begin{equation}
\label{eq:4}
\ts_{\lambda/\nu}^{k-1} \ts_{\mu/\nu} =(-1)^{\inv(\vec a,\vec b) +\binom k2}\det \left( (-1)^{\chi(a_i>b_j)} \ts_{\lambda(a_i,b_j)/\nu}\right)_{i,j=1}^k.
\end{equation}
Since $a_i\ne b_j$ for all $i,j$, we have $\inv(\vec a,\vec b) = k^2-\inv(\vec b,\vec a)$ and $\chi(a_i>b_j)=1-\chi(b_j>a_i)$. Thus \eqref{eq:4} is equivalent to \eqref{eq:main1}, which completes the proof of the first identity.
The second identity \eqref{eq:main2} is proved by the same arguments except that
in this case we use the matrix $N=\left(h_{\nu_j-j+i,i+1}\right)_{i\in\ZZ, 1\le
j\le n}$ in place of $M$.
\end{proof}
\section{Lascoux--Pragacz and Kreiman decompositions}
\label{sec:lasc-prag-kreim}
In this section we restate Theorem~\ref{thm:main} using the Lascoux--Pragacz and
Kreiman decompositions for the case $\mu\subseteq\lambda$. As a corollary we
prove (a corrected version of) a conjecture of Morales, Pak, and Panova
\cite{MPP2}.
Recall that for a border strip $\gamma$, we denote by $p(\gamma)$
(resp.~$q(\gamma)$) the content of the starting (resp.~ending) cell of $\gamma$.
\begin{defn}
A \emph{(border strip) decomposition} of a skew shape $\alpha$ is a sequence
$\theta=(\theta_1,\dots,\theta_k)$ of border strips satisfying the following
conditions.
\begin{itemize}
\item $\theta_i\cap \theta_j =\emptyset$ for all $i\ne j$,
\item $\theta_1\cup \cdots\cup \theta_k = \alpha$.
\end{itemize}
If there is no possible confusion, we will simply write $p_i=p(\theta_i)$,
$q_i=q(\theta_i)$, $\vec p=(p_1,\dots,p_k)$ and $\vec q=(q_1,\dots,q_k)$. By
convention we will always label the border strips so that $q_1\ge q_2\ge \dots \ge q_k$.
\end{defn}
Recall that the outer strip of a partition $\lambda$ is the set of cells
$x\in\lambda$ satisfying $x+(-1,-1)\not\in\lambda$. For a skew shape $\lm$, the
\emph{outer strip} of $\lm$ is defined to be the outer strip of $\lambda$. Let
$\gamma$ be the outer strip of $\lm$ and let $\rho$ be the set of cells $x\in
\lm$ with $x+(-1,-1)\not\in\lm$. We define the \emph{inner strip} of $\lm$ to be
the set
\[
\rho \cup \{x\in \gamma: c(x)\notin \Cont(\rho) \}.
\]
See Figure~\ref{fig:inner}.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\fill[fill=cyan!27!white, rounded corners] (.2, .2) rectangle (1.8, .8)
(1.2, .2) rectangle (1.8, 1.8)
(1.2, 1.2) rectangle (2.8, 1.8)
(2.2, 1.2) rectangle (2.8, 2.8)
(2.2, 2.2) rectangle (3.8, 2.8)
(3.2, 2.2) rectangle (3.8, 4.8)
(3.2, 4.2) rectangle (6.8, 4.8)
(6.2, 4.2) rectangle (6.8, 6.8)
(6.2, 6.2) rectangle (9.8, 6.8)
(9.2, 6.2) rectangle (9.8, 7.8)
(9.2, 7.2) rectangle (10.8, 7.8)
(10.2, 7.2) rectangle (10.8, 8.8);
\draw[dashed, color=gray] (0,3)--(0,9)--(11,9)--(11,8)
(4,4)--(6,4)
(7,5)--(7,6)--(8,6)
(0,8)--(8,8)
(0,7)--(8,7)
(0,6)--(7,6)
(0,5)--(6,5)
(0,4)--(1,4)
(1,4)--(1,9)
(2,4)--(2,9)
(3,4)--(3,9)
(4,4)--(4,9)
(5,4)--(5,9)
(6,5)--(6,9)
(7,6)--(7,9)
(8,8)--(8,9)
(9,8)--(9,9)
(10, 8)--(10,9);
\draw[thick] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(4,2)--(4,4)--(1,4)--(1,3)--(0,3)--(0,0)--cycle
(6,4)--(7,4)--(7,5)--(6,5)--(6,4)--cycle
(8,6)--(10,6)--(10,7)--(11,7)--(11,8)--(8,8)--(8,6)--cycle
(0,1)--(2,1)
(0,2)--(3,2)
(1,3)--(4,3)
(1,0)--(1,3)
(2,1)--(2,4)
(3,2)--(3,4)
(9,6)--(9,8)
(10,7)--(10,8)
(8,7)--(10,7);
\end{tikzpicture}\qquad\quad
\begin{tikzpicture}[scale=.46]
\fill[fill=asparagus!35!white, rounded corners] (.2, .2) rectangle (.8, 2.8)
(.2, 2.2) rectangle (1.8, 2.8)
(1.2, 2.2) rectangle (1.8, 3.8)
(1.2, 3.2) rectangle (3.8, 3.8)
(3.2, 3.2) rectangle (3.8, 4.8)
(3.2, 4.2) rectangle (6.8, 4.8)
(6.2, 4.2) rectangle (6.8, 6.8)
(6.2, 6.2) rectangle (8.8, 6.8)
(8.2, 6.2) rectangle (8.8, 7.8)
(8.2, 7.2) rectangle (10.8, 7.8)
(10.2, 7.2) rectangle (10.8, 8.8);
\draw[dashed, color=gray] (0,3)--(0,9)--(11,9)--(11,8)
(4,4)--(6,4)
(7,5)--(7,6)--(8,6)
(0,8)--(8,8)
(0,7)--(8,7)
(0,6)--(7,6)
(0,5)--(6,5)
(0,4)--(1,4)
(1,4)--(1,9)
(2,4)--(2,9)
(3,4)--(3,9)
(4,4)--(4,9)
(5,4)--(5,9)
(6,5)--(6,9)
(7,6)--(7,9)
(8,8)--(8,9)
(9,8)--(9,9)
(10, 8)--(10,9);
\draw[thick] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(4,2)--(4,4)--(1,4)--(1,3)--(0,3)--(0,0)--cycle
(6,4)--(7,4)--(7,5)--(6,5)--(6,4)--cycle
(8,6)--(10,6)--(10,7)--(11,7)--(11,8)--(8,8)--(8,6)--cycle
(0,1)--(2,1)
(0,2)--(3,2)
(1,3)--(4,3)
(1,0)--(1,3)
(2,1)--(2,4)
(3,2)--(3,4)
(9,6)--(9,8)
(10,7)--(10,8)
(8,7)--(10,7);
\end{tikzpicture}
\caption{The outer strip of $\lm$ on the left and the inner strip of $\lm$ on
the right.}
\label{fig:inner}
\end{figure}
\begin{defn}
Let $\mu$ and $\lambda$ be partitions with $\mu\subseteq\lambda$. The
\emph{Lascoux--Pragacz decomposition} (resp.~\emph{Kreiman decomposition}) of
$\lm$ is the decomposition of $\lm$ obtained by using the outer (resp.~inner)
strip of $\lambda$ as the cutting strip. See Figure~\ref{fig:LPK}.
\end{defn}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.48]
\fill[fill=orange!40!white, rounded corners] (.2, .2) rectangle (1.8, .8);
\fill[fill=orange!40!white, rounded corners] (1.2, .2) rectangle (1.8, 1.8);
\fill[fill=orange!40!white, rounded corners] (1.2, 1.2) rectangle (4.8, 1.8);
\fill[fill=orange!40!white, rounded corners] (4.2, 1.2) rectangle (4.8, 3.8);
\fill[fill=orange!40!white, rounded corners] (4.2, 3.2) rectangle (6.8, 3.8);
\fill[fill=orange!40!white, rounded corners] (6.2, 3.2) rectangle (6.8, 4.8);
\fill[fill=orange!40!white, rounded corners] (6.2, 4.2) rectangle (8.8, 4.8);
\fill[fill=orange!40!white, rounded corners] (8.2, 4.2) rectangle (8.8, 5.8);
\fill[fill=orange!40!white, rounded corners] (8.2, 5.2) rectangle (9.8, 5.8);
\fill[fill=red!40!white, rounded corners] (.2, 1.2) rectangle (.8, 2.8);
\fill[fill=red!40!white, rounded corners] (.2, 2.2) rectangle (3.8, 2.8);
\fill[fill=red!40!white, rounded corners] (3.2,2.2) rectangle (3.8, 4.8);
\fill[fill=red!40!white, rounded corners] (3.2, 4.2) rectangle (5.8, 4.8);
\fill[fill=red!40!white, rounded corners] (5.2, 4.2) rectangle (5.8, 5.8);
\fill[fill=red!40!white, rounded corners] (5.2, 5.2) rectangle (7.8, 5.8);
\fill[fill=cyan!27!white, rounded corners] (3.2, 5.2) rectangle (4.8, 5.8);
\fill[fill=pink, rounded corners] (.2, 3.2) rectangle (2.8, 3.8);
\fill[fill=pink, rounded corners] (2.2, 3.2) rectangle (2.8, 4.8);
\draw[thick] (0,0)--(2,0)--(2,1)--(5,1)--(5,3)--(7,3)--(7,4)--(9,4)--(9,5)--(10,5)--(10,6)--(3,6)--(3,5)--(2,5)--(2,4)--(0,4)--(0,0)--cycle;
\draw[densely dashed] (0,4)--(0,6)--(3,6);
\draw[densely dashed] (0,5)--(2,5);
\draw[densely dashed] (1,4)--(1,6);
\draw[densely dashed] (2,5)--(2,6);
\draw (3,5)--(9,5)
(2,4)--(7,4)
(0,3)--(5,3)
(0,2)--(5,2)
(0,1)--(2,1)
(1,0)--(1,4)
(2,1)--(2,4)
(3,1)--(3,5)
(4,1)--(4,6)
(5,3)--(5,6)
(6,3)--(6,6)
(7,4)--(7,6)
(8,4)--(8,6)
(9,5)--(9,6);
\node[] at (9.5, 6.5) {$\theta_1$};
\node[] at (7.5, 6.5) {$\theta_2$};
\node[] at (4.5, 6.5) {$\theta_3$};
\node[] at (-.5, 3.5) {$\theta_4$};
\node[] at (2, -1) {\phantom{d}};
\end{tikzpicture}\qquad\qquad
\begin{tikzpicture}[scale=.48]
\fill[fill=orange!40!white, rounded corners] (.2, .2) rectangle (.8, 3.8)
(.2, 3.2) rectangle (2.8, 3.8)
(2.2, 3.2) rectangle (2.8, 4.8)
(2.2, 4.2) rectangle (3.8, 4.8)
(3.2, 4.2) rectangle (3.8, 5.8)
(3.2, 5.2) rectangle (9.8, 5.8);
\fill[fill=red!40!white, rounded corners] (1.2, .2) rectangle (1.8, 2.8)
(1.2, 2.2) rectangle (3.8, 2.8)
(3.2, 2.2) rectangle (3.8, 3.8)
(3.2, 3.2) rectangle (4.8, 3.8)
(4.2, 3.2) rectangle (4.8, 4.8)
(4.2, 4.2) rectangle (8.8, 4.8);
\fill[fill=pink, rounded corners] (2.2, 1.2) rectangle (4.8, 1.8)
(4.2, 1.2) rectangle (4.8, 2.8);
\fill[fill=cyan!27!white, rounded corners] (5.2, 3.2) rectangle (6.8, 3.8);
\draw[thick] (0,0)--(2,0)--(2,1)--(5,1)--(5,3)--(7,3)--(7,4)--(9,4)--(9,5)--(10,5)--(10,6)--(3,6)--(3,5)--(2,5)--(2,4)--(0,4)--(0,0)--cycle;
\draw[densely dashed] (0,4)--(0,6)--(3,6);
\draw[densely dashed] (0,5)--(2,5);
\draw[densely dashed] (1,4)--(1,6);
\draw[densely dashed] (2,5)--(2,6);
\draw (3,5)--(9,5)
(2,4)--(7,4)
(0,3)--(5,3)
(0,2)--(5,2)
(0,1)--(2,1)
(1,0)--(1,4)
(2,1)--(2,4)
(3,1)--(3,5)
(4,1)--(4,6)
(5,3)--(5,6)
(6,3)--(6,6)
(7,4)--(7,6)
(8,4)--(8,6)
(9,5)--(9,6);
\node[] at (10.6, 5.4) {$\theta_1$};
\node[] at (9.6, 4.4) {$\theta_2$};
\node[] at (7.6, 3.4) {$\theta_3$};
\node[] at (5.6, 2.4) {$\theta_4$};
\node[] at (2, -1) {\phantom{d}};
\end{tikzpicture}
\caption{The Lascoux--Pragacz decompositions of a connected skew shape
(left) and the Kreiman decomposition of the same skew shape (right)}
\label{fig:LPK}
\end{figure}
\begin{lem}\label{lem:C/C}
Let $\lambda,\mu\in \Par_n$ with $\mu\subseteq\lambda$. Suppose that
$\theta=(\theta_1,\dots,\theta_k)$ is the Lascoux--Pragacz or Kreiman
decomposition of $\lm$. Then $p_i\ne p_j$, $q_i\ne q_j$, $p_i\ne q_j$, and
$p_i-1\ne q_j$ for all $1\le i\ne j\le k$, and
\begin{align*}
C_n(\lambda)\setminus C_n(\mu) &= \{q_1>q_2>\dots>q_k\},\\
C_n(\mu)\setminus C_n(\lambda) &= \{p_1-1,p_2-1,\dots,p_k-1\}.
\end{align*}
\end{lem}
\begin{proof}
We will only consider the case that $\theta$ is the Lascoux--Pragacz
decomposition of $\lm$ since it can be proved similarly for the case of the
Kreiman decomposition. We proceed by induction on $k$. If $k=0$,
then $\lambda=\mu$ and there is nothing to prove. Suppose that $k\ge1$ and the
lemma is true for $k-1$.
By the definition of the Lascoux--Pragacz decomposition, the border strip
$\theta_1$ is a connected component of $(\lm)\cap \lambda^0$. This implies
that $p_1-1,q_1+1\not\in \Cont(\lm)$ and the starting and ending cells of
$\theta_1$ are the only cells in $\lm$ whose contents are $p_1$ and $q_1$,
respectively. Therefore, if $\Cont(\theta_i)\cap\Cont(\theta_1)\ne\emptyset$
for some $i\ge2$, then $\Cont(\theta_i)\subseteq[p_1+1,q_1-1]$ because
$\Cont(\theta_i)$ is a set of consecutive integers. This shows that $p_i\ne
p_1$ $q_i\ne q_1$, $p_i\ne q_1$, $p_i-1\ne q_1$, and $p_1-1\ne q_i$ for all
$2\le i\le k$.
Let $\rho=\lambda\setminus\theta_1$. Then $\mu\subseteq\rho$ and
$(\theta_2,\theta_3,\dots,\theta_k)$ is the Lascoux--Pragacz decomposition of
$\rho/\mu$. By the induction hypothesis, we have $p_i\ne p_j$, $q_i\ne q_j$,
$p_i\ne q_j$, and $p_i-1\ne q_j$ for all $2\le i\ne j\le k$, and
\begin{align*}
C_n(\rho)\setminus C_n(\mu) &= \{q_2>q_3>\dots>q_k\},\\
C_n(\mu)\setminus C_n(\rho) &= \{p_2-1,p_3-1,\dots,p_k-1\}.
\end{align*}
It is straightforward to check that $p_1-1\in C_n(\mu)$, $q_1\not\in
C_n(\mu)$, $p_1-1\not\in C_n(\lambda)$, $q_1\in C_n(\lambda)$, and
\[
C_n(\rho) = (C_n(\lambda)\setminus\{q_1\})\cup\{p_1-1\}.
\]
Combining the above results we obtain the lemma for $k$. The induction then
completes the proof.
\end{proof}
By Lemma~\ref{lem:C/C}, if $\mu\subseteq\lambda$, then we can restate
Theorem~\ref{thm:main} using the Lascoux--Pragacz or Kreiman decompositions as
follows.
\begin{thm}[Theorem~\ref{thm:main_LP_intro}]\label{thm:main_LP}
Let $\lambda$, $\mu$, and $\nu$ be partitions and suppose that $\mu\subseteq
\lambda$ and $\theta=(\theta_1,\dots,\theta_k)$ is the Lascoux--Pragacz
decomposition of $\lm$. Then we have
\begin{align}
\label{eq:main_LP1}
\ts_{\lambda/\nu}^{k-1} \ts_{\mu/\nu}
&= \det \left( (-1)^{\chi(p_j>q_i)} \ts_{\lambda(q_i,p_j-1)/\nu}\right)_{i,j=1}^k, \\
\label{eq:main_LP2}
\ts_{\nu/\lambda}^{k-1} \ts_{\nu/\mu}
&= \det \left( (-1)^{\chi(p_j>q_i)} \ts_{\nu/\lambda(q_i,p_j-1)}\right)_{i,j=1}^k.
\end{align}
\end{thm}
\begin{proof}
Let $\vec a=(a_1,\dots,a_k)$ and $\vec b=(b_1,\dots,b_k)$ be the sequences
defined by
\begin{align*}
C_n(\lambda)\setminus C_n(\mu) &= \{a_1>a_2>\dots>a_k\},\\
C_n(\mu)\setminus C_n(\lambda) &= \{b_1>b_2>\dots>b_k\}.
\end{align*}
By \eqref{eq:main1}, we have
\[
\ts_{\lambda/\nu}^{k-1} \ts_{\mu/\nu}
=(-1)^{\inv(\vec b,\vec a)+\binom k2}\det
\left( (-1)^{\chi(b_j>a_i)} \ts_{\lambda(a_i,b_j)/\nu}\right)_{i,j=1}^k.
\]
By Lemma~\ref{lem:C/C}, we have $\vec a = \vec q$ and
$\{p_1-1,p_2-1,\dots,p_k-1\}=\{b_1,b_2,\dots,b_k\}$. Let $\vec b' =
(b_1',b_2',\dots,b_k') = (b_k,b_{k-1},\dots,b_1)$ be the increasing
rearrangement of $\vec b$. Then we can rewrite the above equation as
\begin{align}
\notag
\ts_{\lambda/\nu}^{k-1} \ts_{\mu/\nu}
&=(-1)^{\inv(\vec b,\vec q)}\det
\left( (-1)^{\chi(b'_j>q_i)} \ts_{\lambda(q_i,b'_j)/\nu}\right)_{i,j=1}^k\\
\label{eq:bq}
&=(-1)^{\inv(\vec b,\vec q)+\inv(\pi)} \det
\left( (-1)^{\chi(p_j-1>q_i)} \ts_{\lambda(q_i,p_j-1)/\nu}\right)_{i,j=1}^k,
\end{align}
where $\pi$ is the permutation of $\{1,2,\dots,k\}$ satisfying
\[
\vec p - \vec 1:=(p_1-1,p_2-1,\dots,p_k-1) = (b'_{\pi(1)}, b'_{\pi(2)},\dots,b'_{\pi(k)}).
\]
Note that $\inv(\pi)$ is the number of pairs $(i,j)$ with $i<j$ and
$\pi(i)>\pi(j)$, where $\pi(i)>\pi(j)$ is equivalent to
$b'_{\pi(i)}>b'_{\pi(j)}$, which in turn is equivalent to $p_i>p_j$. Thus
$\inv(\pi)$ is equal to the number of pairs $(i,j)$ with $i<j$ and $p_i>p_j$. By
the construction of the Lascoux--Pragacz decomposition, if $i<j$, then we must
have either $p_j<q_j<p_i<q_i$ or $p_i<p_j\le q_j<q_i$. This shows that
$\inv(\pi)$ is equal to the number of pairs $(\theta_i,\theta_j)$ of border
strips such that $p(\theta_i)>q(\theta_j)$, which automatically implies $i<j$.
Therefore $\inv(\pi)=\inv(\vec p,\vec q)$. On the other hand, by
Lemma~\ref{lem:C/C}, we have $p_i-1\ne q_j$ for all $1\le i,j\le n$. Thus
$\inv(\pi) = \inv(\vec p,\vec q)=\inv(\vec p-\vec 1,\vec q)=\inv(\vec b,\vec q)$
and $\chi(p_j-1>q_i)=\chi(p_j>q_i)$. Therefore the right hand sides of
\eqref{eq:bq} and \eqref{eq:main_LP1} are equal, which completes the proof of
\eqref{eq:main_LP1}.
The second identity \eqref{eq:main_LP2} can be proved similarly.
\end{proof}
\begin{thm}\label{thm:main_K}
Let $\lambda,\mu$, and $\nu$ be partitions. Suppose that $\mu\subseteq \lambda$
and $\theta=(\theta_1,\dots,\theta_k)$ is the Kreiman decomposition of $\lm$.
Then we have
\begin{align}
\label{eq:main_K1}
\ts_{\mu/\nu}^{k-1} \ts_{\lambda/\nu}
&= \det \left( (-1)^{\chi(p_i> q_j)} \ts_{\mu(p_i-1,q_j)/\nu}\right)_{i,j=1}^k, \\
\label{eq:main_K2}
\ts_{\nu/\lambda}^{k-1} \ts_{\nu/\mu}
&= \det \left( (-1)^{\chi(p_i> q_j)} \ts_{\nu/\mu(p_i-1,q_j)}\right)_{i,j=1}^k.
\end{align}
\end{thm}
\begin{proof}
Let $\vec a=(a_1,\dots,a_k)$ and $\vec b=(b_1,\dots,b_k)$ be the sequences
defined by
\begin{align*}
C_n(\mu)\setminus C_n(\lambda) &= \{a_1>a_2>\dots>a_k\},\\
C_n(\lambda)\setminus C_n(\mu) &= \{b_1>b_2>\dots>b_k\}.
\end{align*}
Then by \eqref{eq:main1} with the roles of $\lambda$ and $\mu$ switched, we have
\[
\ts_{\mu/\nu}^{k-1} \ts_{\lambda/\nu}
=(-1)^{\inv(\vec b,\vec a)+\binom k2}\det
\left( (-1)^{\chi(b_j>a_i)} \ts_{\mu(a_i,b_j)/\nu}\right)_{i,j=1}^k.
\]
By Lemma~\ref{lem:C/C}, we have $\vec b = \vec q$ and
$\{p_1-1,p_2-1,\dots,p_k-1\}=\{a_1,a_2,\dots,a_k\}$.
Let $\vec a' = (a_1',a_2',\dots,a_k') = (a_k,a_{k-1},\dots,a_1)$ be the
increasing rearrangement of $\vec a$. Then we can rewrite the above equation as
\begin{align}
\notag
\ts_{\mu/\nu}^{k-1} \ts_{\lambda/\nu}
&=(-1)^{\inv(\vec q,\vec a)}\det
\left( (-1)^{\chi(q_j>a'_i)} \ts_{\mu(a'_i,q_j)/\nu}\right)_{i,j=1}^k\\
\label{eq:bq2}
&=(-1)^{\inv(\vec q,\vec a)+\inv(\pi)}\det
\left( (-1)^{\chi(q_j>p_i-1)} \ts_{\mu(p_i-1,q_j)/\nu}\right)_{i,j=1}^k,
\end{align}
where $\pi$ is the permutation of $\{1,2,\dots,k\}$ satisfying
\[
\vec p-\vec 1:= (p_1-1,p_2-1,\dots,p_k-1) = (a'_{\pi(1)}, a'_{\pi(2)},\dots,a'_{\pi(k)}).
\]
By the same argument as in the proof of Theorem~\ref{thm:main_LP}, we have
\[
\inv(\pi) = \inv(\vec p,\vec q)=\inv(\vec p-\vec 1,\vec q)
=\inv(\vec a,\vec q) = k^2- \inv(\vec q,\vec a).
\]
Thus we can rewrite \eqref{eq:bq2} as
\[
\ts_{\mu/\nu}^{k-1} \ts_{\lambda/\nu}
=\det \left( (-1)^{1-\chi(q_j>p_i-1)} \ts_{\mu(p_i-1,q_j)/\nu}\right)_{i,j=1}^k.
\]
Since $q_j\ne p_i-1$ for all $1\le i,j\le k$ by Lemma~\ref{lem:C/C}, we have
$1-\chi(q_j>p_i-1)=\chi(q_j\le p_i-1)=\chi(q_j<p_i)$, which together with the
above equation shows \eqref{eq:main_K1}.
The second identity \eqref{eq:main_K2} can be proved similarly.
\end{proof}
\begin{exam}\label{exam:LSthm} Let $\lambda=(6,6,6,3,3)$, $\mu=(4,3,2)$ and
$\nu=(7,6,6,5,3,2)$. The
Lascoux--Pragacz decomposition of $\lambda/\mu$ and the skew shape $\nl$ are
shown in Figure~\ref{fig:LP}. Then \eqref{eq:main_LP2} of Theorem
\ref{thm:main_LP} implies that $\ts_{\nu/\lambda}^2\ts_{\nu/\mu}$ is equal to
the determinant shown in Figure~\ref{fig:det}.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.43]
\draw[densely dashed] (0,5)--(4,5)
(0,4)--(3,4)
(0,3)--(2,3)
(0,2)--(0,5)
(1,2)--(1,5)
(2,3)--(2,5)
(3,4)--(3,5);
\draw[color=gray] (4,5)--(6,5)
(3,4)--(6,4)
(2,3)--(6,3)
(0,2)--(6,2)
(0,1)--(3,1)
(0,0)--(3,0)
(0,0)--(0,2)
(1,0)--(1,2)
(2,0)--(2,3)
(3,0)--(3,4)
(4,2)--(4,5)
(5,2)--(5,5)
(6,2)--(6,5);
\draw[thick, line width=1.2pt] (0,1)--(2,1)--(2,2)--(0,2)--(0,1)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(5,5)--(5,3)--(2,3)--(2,1)--(0,1)--(0,0)--cycle;
\draw[thick, line width=1.2pt] (3,3)--(5,3)--(5,5)--(4,5)--(4,4)--(3,4)--(3,3)--cycle;
\node[] at (.5, .5) {$\color{blue} p_1$};
\node[] at (5.5, 4.5) {$\color{red} q_1$};
\node[] at (3.5, 3.5) {$\color{blue} p_2$};
\node[] at (4.5, 4.5) {$\color{red} q_2$};
\node[] at (.5, 1.5) {$\color{blue} p_3$};
\node[] at (1.5, 1.5) {$\color{red} q_3$};
\node[] at (5.5, 5.6) {$\theta_1$};
\node[] at (-.6, 1.5) {$\theta_3$};
\node[] at (4.5, 5.6) {$\theta_2$};
\end{tikzpicture} \qquad\qquad\quad
\begin{tikzpicture}[scale=.43]
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (1,0)--(1,1)
(4,1)--(4,3)
(3,2)--(5,2);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(0,1)--(0,0)--cycle
(3,1)--(5,1)--(5,3)--(3,3)--(3,1)--cycle
(6,5)--(7,5)--(7,6)--(6,6)--(6,5)--cycle;
\end{tikzpicture}
\caption{The Lascoux--Pragacz decomposition of $\lambda/\mu$ on the left
and the skew shape $\nl$ on the right. }
\label{fig:LP}
\end{figure}
\begin{figure}
\centering
\[
\det \begin{bmatrix}
~\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (0,1)--(3,1)--(3,3)--(6,3)--(6,6)--(5,6)--(5,4)--(2,4)--(2,2)--(0,2)--(0,1)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (5,5)--(6,5)
(5,4)--(6,4)
(2,3)--(5,3)
(2,2)--(5,2)
(0,1)--(2,1)
(1,0)--(1,2)
(2,1)--(2,2)
(3,1)--(3,4)
(4,1)--(4,4)
(5,3)--(5,4)
(6,5)--(6,6);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(3,1)--(5,1)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(5,6)--(5,4)--(2,4)--(2,2)--(0,2)--(0,0)--cycle;
\node[] at (.5, 1.5) {$\color{blue} p_1$};
\node[] at (5.5, 5.5) {$\color{red} q_1$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (4,3)--(6,3)--(6,6)--(5,6)--(5,4)--(4,4)--(4,3)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (1,0)--(1,1)
(3,2)--(5,2)
(4,1)--(4,3)
(5,3)--(5,4)
(6,5)--(6,6)
(5,5)--(6,5)
(5,4)--(6,4)
(4,3)--(5,3);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(0,1)--(0,0)--cycle
(3,1)--(5,1)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(5,6)--(5,4)--(4,4)--(4,3)--(3,3)--(3,1)--cycle;
\node[] at (4.5, 3.5) {$\color{blue} p_2$};
\node[] at (5.5, 5.5) {$\color{red} q_1$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (1,1)--(3,1)--(3,3)--(6,3)--(6,6)--(5,6)--(5,4)--(2,4)--(2,2)--(1,2)--(1,1)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (5,5)--(6,5)
(3,2)--(5,2)
(5,4)--(6,4)
(2,3)--(5,3)
(2,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,2)
(2,1)--(2,2)
(3,1)--(3,4)
(4,1)--(4,4)
(5,3)--(5,4)
(6,5)--(6,6);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(3,1)--(5,1)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(5,6)--(5,4)--(2,4)--(2,2)--(1,2)--(1,1)--(0,1)--(0,0)--cycle;
\node[] at (1.5, 1.5) {$\color{blue} p_3$};
\node[] at (5.5, 5.5) {$\color{red} q_1$};
\end{tikzpicture}~\\\\~\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (0,1)--(3,1)--(3,3)--(6,3)--(6,5)--(5,5)--(5,4)--(2,4)--(2,2)--(0,2)--(0,1)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (5,5)--(6,5)
(5,4)--(6,4)
(2,3)--(5,3)
(2,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,2)
(2,1)--(2,2)
(3,2)--(5,2)
(3,1)--(3,4)
(4,1)--(4,4)
(5,3)--(5,4)
(6,5)--(6,6);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(3,1)--(5,1)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(6,6)--(6,5)--(5,5)--(5,4)--(2,4)--(2,2)--(0,2)--(0,0)--cycle;
\node[] at (.5, 1.5) {$\color{blue} p_1$};
\node[] at (5.5, 4.5) {$\color{red} q_2$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (4,3)--(6,3)--(6,5)--(5,5)--(5,4)--(4,4)--(4,3)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,1)--(3,6)
(4,1)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (1,0)--(1,1)
(3,2)--(5,2)
(4,1)--(4,3)
(5,3)--(5,4)
(6,5)--(6,6)
(5,5)--(6,5)
(5,4)--(6,4)
(4,3)--(5,3);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(0,1)--(0,0)--cycle
(3,1)--(5,1)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(6,6)--(6,5)--(5,5)--(5,4)--(4,4)--(4,3)--(3,3)--(3,1)--cycle;
\node[] at (4.5, 3.5) {$\color{blue} p_2$};
\node[] at (5.5, 4.5) {$\color{red} q_2$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (1,1)--(3,1)--(3,3)--(6,3)--(6,5)--(5,5)--(5,4)--(2,4)--(2,2)--(1,2)--(1,1)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (5,5)--(6,5)
(5,4)--(6,4)
(2,3)--(5,3)
(2,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,2)
(2,1)--(2,2)
(3,1)--(3,4)
(3,2)--(5,2)
(4,1)--(4,4)
(5,3)--(5,4)
(6,5)--(6,6);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(3,1)--(5,1)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(6,6)--(6,5)--(5,5)--(5,4)--(2,4)--(2,2)--(1,2)--(1,1)--(0,1)--(0,0)--cycle;
\node[] at (1.5, 1.5) {$\color{blue} p_3$};
\node[] at (5.5, 4.5) {$\color{red} q_2$};
\end{tikzpicture}~\\\\~
\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (0,1)--(3,1)--(3,2)--(0,2)--(0,1)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (0,1)--(2,1)
(1,0)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)
(3,2)--(5,2)
(4,1)--(4,3);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(5,1)--(5,3)--(3,3)--(3,2)--(0,2)--(0,0)--cycle
(6,5)--(7,5)--(7,6)--(6,6)--(6,5)--cycle;
\node[] at (.5, 1.5) {$\color{blue} p_1$};
\node[] at (2.5, 1.5) {$\color{red} q_3$};
\end{tikzpicture}~&~
\begin{tikzpicture}[scale=.4]
\fill[fill=red!10!white] (3,1)--(4,1)--(4,2)--(5,2)--(5,3)--(3,3)--(3,1)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (1,0)--(1,1);
\draw[thick, color=dredcolor] (3,1)--(4,1)--(4,2)--(5,2)--(5,3)--(3,3)--(3,1)--cycle;
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(0,1)--(0,0)--cycle
(4,1)--(5,1)--(5,2)--(4,2)--(4,1)--cycle
(6,5)--(7,5)--(7,6)--(6,6)--(6,5)--cycle;
\node[] at (4.5, 3.5) {$\color{blue} p_2$};
\node[] at (2.5, 1.5) {$\color{red} q_3$};
\node[] at (-.75, 3) {$ -$};
\end{tikzpicture}~&~
\begin{tikzpicture}[scale=.4]
\fill[fill=cyan!10!white] (1,1)--(3,1)--(3,2)--(1,2)--(1,1)--cycle;
\draw[dashed, line width = .2pt, color=gray] (0,0)--(2,0)--(2,1)--(3,1)--(3,2)--(5,2)--(5,3)--(6,3)--(6,5)--(7,5)--(7,6)--(0,6)--(0,0)--cycle
(0,5)--(6,5)
(0,4)--(6,4)
(0,3)--(5,3)
(0,2)--(3,2)
(0,1)--(2,1)
(1,0)--(1,6)
(2,1)--(2,6)
(3,2)--(3,6)
(4,2)--(4,6)
(5,3)--(5,6)
(6,5)--(6,6);
\draw[color=gray] (0,1)--(2,1)
(1,0)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)
(3,2)--(5,2)
(4,1)--(4,3);
\draw[line width=1.2pt] (0,0)--(2,0)--(2,1)--(5,1)--(5,3)--(3,3)--(3,2)--(1,2)--(1,1)--(0,1)--(0,0)--cycle
(6,5)--(7,5)--(7,6)--(6,6)--(6,5)--cycle;
\node[] at (1.5, 1.5) {$\color{blue} p_3$};
\node[] at (2.5, 1.5) {$\color{red} q_3$};
\end{tikzpicture}~
\end{bmatrix}
\]
\caption{ An illustrate of the determinant $\det \left(
(-1)^{\chi(p_j>q_i)} \ts_{\nu/\lambda(q_i,p_j-1)}\right)_{i,j=1}^k$. The
$(i,j)$-entry shows the sign $(-1)^{\chi(p_j>q_i)}$ and the (not necessarily
connected) skew shape $\nu/\lambda(q_i,p_j-1)$ in thick lines. The partition
$\lambda(q_i,p_j-1)$ is obtained from $\lambda$ by removing (blue) or adding
(red) a border strip depending on the sign.}
\label{fig:det}
\end{figure}
\end{exam}
If $\nu=\lambda$, then the $(i,j)$-entry of the matrix in \eqref{eq:main_LP2} is
\[
(-1)^{\chi(p_j>q_i)} \ts_{\nu/\lambda(q_i,p_j-1)} =
\begin{cases}
\ts_{\lambda^0[p_j,q_i]} & \mbox{if $p_j\le q_i$},\\
\ts_{\lambda} & \mbox{if $p_j-1 = q_i$},\\
0 & \mbox{if $p_j-1\ge q_i$},
\end{cases}
\]
where the right hand side is exactly the same as the definition of
$\ts_{\lambda^0[p_j,q_i]}$. Therefore we obtain the following Lascoux--Pragacz
identity for $\ts_\lm$, in which the case $\mu=\emptyset$ is proved by Macdonald
\cite[(9,9)]{Macdonald_Schur}.
\begin{cor}
Suppose that $\mu\subseteq \lambda$ and $\theta=(\theta_1,\dots,\theta_k)$ is
the Lascoux--Pragacz decomposition of $\lm$. Then we have
\[
\ts_{\lambda/\mu} = \det \left(\ts_{\lambda^0[p_j,q_i]}\right)_{i,j=1}^k.
\]
\end{cor}
By setting $\nu=\emptyset$ in \eqref{eq:main_LP1} and \eqref{eq:main_K1}, we obtain the
following two corollaries.
\begin{cor}\label{cor:main_LP}
Suppose that $\mu\subseteq \lambda$ and $\theta=(\theta_1,\dots,\theta_k)$ is
the Lascoux--Pragacz decomposition of $\lm$. Then we have
\[
\ts_{\lambda}^{k-1} \ts_{\mu}
= \det \left( (-1)^{\chi(p_j>q_i)} \ts_{\lambda(q_i,p_j-1)}\right)_{i,j=1}^k.
\]
\end{cor}
\begin{cor}\label{cor:main_K}
Suppose that $\mu\subseteq \lambda$ and $\theta=(\theta_1,\dots,\theta_k)$ is
the Kreiman decomposition of $\lm$. Then we have
\[
\ts_{\mu}^{k-1} \ts_{\lambda} = \det \left( (-1)^{\chi(p_i> q_j)}
\ts_{\mu(p_i-1,q_j)} \right)_{i,j=1}^k.
\]
\end{cor}
\begin{exam}
For $\lambda/\mu=(6,6,6,3,3)/(4,3,2)$, the Lascoux--Pragacz decomposition of
$\lambda/\mu$ is shown in Figure~\ref{fig:LP}. Then
Corollary~\ref{cor:main_LP} implies that
$\ts_{\lambda}^2 \ts_{\mu}$ is equal to the determinant shown in Figure~\ref{fig:det2}.
\begin{figure}
\centering
\[
\ts_{\lambda}^2 \ts_{\mu}= \det \begin{bmatrix}
~\begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (0,0) rectangle (3,1);
\fill[fill=red!15!white] (2,1) rectangle (3,3);
\fill[fill=red!15!white] (3,2) rectangle (6,3);
\fill[fill=red!15!white] (5,3) rectangle (6,5);
\draw[thick, line width=1.2pt, pink] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(5,5)--(5,3)--(2,3)--(2,1)--(0,1)--(0,0)--cycle;
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt] (0,1)--(2,1)--(2,3)--(5,3)--(5,5)--(0,5)--(0,1)--cycle;
\node[] at (.5, .5) {$\color{blue} p_1$};
\node[] at (5.5, 4.5) {$\color{red} q_1$};
\node[] at (4.2, 2.5) {\tiny\color{red}removed};
\end{tikzpicture}~
&~ \begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (4,2) rectangle (6,3);
\fill[fill=red!15!white] (5,3) rectangle (6,5);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, pink] (4,2)--(6,2)--(6,5)--(5,5)--(5,3)--(4,3)--(4,2)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(3,0)--(3,2)--(4,2)--(4,3)--(5,3)--(5,5)--(0,5)--(0,0)--cycle;
\node[] at (4.5, 2.5) {$\color{blue} p_2$};
\node[] at (5.5, 4.5) {$\color{red} q_1$};
\end{tikzpicture}~
& ~\begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (1,0) rectangle (3,1);
\fill[fill=red!15!white] (2,1) rectangle (3,3);
\fill[fill=red!15!white] (3,2) rectangle (6,3);
\fill[fill=red!15!white] (5,3) rectangle (6,5);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, pink] (1,0)--(3,0)--(3,2)--(6,2)--(6,5)--(5,5)--(5,3)--(2,3)--(2,1)--(1,1)--(1,0)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(1,0)--(1,1)--(2,1)--(2,3)--(5,3)--(5,5)--(0,5)--(0,0)--cycle;
\node[] at (1.5, .5) {$\color{blue} p_3$};
\node[] at (5.5, 4.5) {$\color{red} q_1$};
\end{tikzpicture}~\\\\
~\begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (0,0) rectangle (3,1);
\fill[fill=red!15!white] (2,1) rectangle (3,3);
\fill[fill=red!15!white] (3,2) rectangle (6,3);
\fill[fill=red!15!white] (5,3) rectangle (6,4);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, pink] (0,0)--(3,0)--(3,2)--(6,2)--(6,4)--(5,4)--(5,3)--(2,3)--(2,1)--(0,1)--(0,0)--cycle;
\draw[thick, line width=1.2pt] (0,1)--(2,1)--(2,3)--(5,3)--(5,4)--(6,4)--(6,5)--(0,5)--(0,1)--cycle;
\node[] at (.5, .5) {$\color{blue} p_1$};
\node[] at (5.5, 3.5) {$\color{red} q_2$};
\end{tikzpicture}~
&~ \begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (4,2) rectangle (6,3);
\fill[fill=red!15!white] (5,3) rectangle (6,4);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, pink] (4,2)--(6,2)--(6,4)--(5,4)--(5,3)--(4,3)--(4,2)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(3,0)--(3,2)--(4,2)--(4,3)--(5,3)--(5,4)--(6,4)--(6,5)--(0,5)--(0,0)--cycle;
\node[] at (4.5, 2.5) {$\color{blue} p_2$};
\node[] at (5.5, 3.5) {$\color{red} q_2$};
\end{tikzpicture}~
&~ \begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (1,0) rectangle (3,1);
\fill[fill=red!15!white] (2,1) rectangle (3,3);
\fill[fill=red!15!white] (3,2) rectangle (6,3);
\fill[fill=red!15!white] (5,3) rectangle (6,4);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, pink] (1,0)--(3,0)--(3,2)--(6,2)--(6,4)--(5,4)--(5,3)--(2,3)--(2,1)--(1,1)--(1,0)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(1,0)--(1,1)--(2,1)--(2,3)--(5,3)--(5,4)--(6,4)--(6,5)--(0,5)--(0,0)--cycle;
\node[] at (1.5, .5) {$\color{blue} p_3$};
\node[] at (5.5, 3.5) {$\color{red} q_2$};
\end{tikzpicture}~\\\\
~\begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (0,0) rectangle (3,1);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, pink] (0,0)--(3,0)--(3,1)--(0,1)--(0,0)--cycle;
\draw[thick, line width=1.2pt] (0,1)--(3,1)--(3,2)--(6,2)--(6,5)--(0,5)--(0,1)--cycle;
\node[] at (.5, .5) {$\color{blue} p_1$};
\node[] at (2.5, .5) {$\color{red} q_3$};
\end{tikzpicture}~
&~ \begin{tikzpicture}[scale=.4]
\fill[fill=cyan!15!white] (3,0) rectangle (4,2);
\fill[fill=cyan!15!white] (4,1) rectangle (5,2);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(4,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,1)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, cyan] (3,0)--(3,2)--(5,2)--(5,1)--(4,1)--(4,0)--(3,0)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(4,0)--(4,1)--(5,1)--(5,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\node[] at (4.5, 2.5) {$\color{blue} p_2$};
\node[] at (2.5, .5) {$\color{red} q_3$};
\node[] at (4, 1.5) {\tiny\color{blue}added};
\node[] at (-.75, 2.5) {$ -$};
\end{tikzpicture}~
& ~\begin{tikzpicture}[scale=.4]
\fill[fill=red!15!white] (1,0) rectangle (3,1);
\draw[color=gray] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\draw[color=gray] (0,4)--(6,4)
(0,3)--(6,3)
(0,2)--(3,2)
(0,1)--(3,1)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,5)
(5,2)--(5,5);
\draw[thick, line width=1.2pt, pink] (1,0)--(3,0)--(3,1)--(1,1)--(1,0)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(1,0)--(1,1)--(3,1)--(3,2)--(6,2)--(6,5)--(0,5)--(0,0)--cycle;
\node[] at (1.5, .5) {$\color{blue} p_3$};
\node[] at (2.5, .5) {$\color{red} q_3$};
\end{tikzpicture}~
\end{bmatrix}
\]
\caption{An illustrate of the determinant $\det \left( (-1)^{\chi(p_j>q_i)}
\ts_{\lambda(q_i,p_j-1)}\right)_{i,j=1}^k$. The $(i,j)$-entry shows the sign
$(-1)^{\chi(p_j>q_i)}$ and the partition $\lambda(q_i,p_j-1)$, which is
obtained from $\lambda$ by either removing (red) or adding (blue) a border
strip.}
\label{fig:det2}
\end{figure}
\end{exam}
We now show that Corollary~\ref{cor:main_LP} implies (a corrected version of) a
conjecture of Morales, Pak, and Panova \cite{MPP2}.
Macdonald's 6th variation $s_\lambda(\vec x|\vec a)$ of Schur functions
\cite{Macdonald_Schur}, also known as \emph{factorial Schur functions}, are
defined by
\[
s_\lambda(\vec x|\vec a) = \frac{\det \left(
(x_i-a_1)(x_i-a_2)\cdots(x_i-a_{\lambda_j+d-j})\right)_{i,j=1}^d}{\prod_{1\le
i<j\le d}(x_i-x_j)},
\]
where $\vec x = (x_1,\dots,x_d)$ is a sequence of variables and $\vec a =
(a_1,a_2,\dots)$ is a sequence of parameters. Note that $s_\lambda(\vec x|\vec
a)$ is a symmetric polynomial in the variables $\vec x$ with parameters $\vec
a$. If $a_i=0$ for all $i$, then $s_\lambda(\vec x|\vec a)$ becomes the Schur
polynomial $s_\lambda(\vec x)$. The factorial Schur functions $s_\lambda(\vec
x|\vec a)$ are a special case of Macdonald's 9th variation $\ts_\lambda$ of
Schur functions.
By specializing the Macdonald's 9th variation to the factorial Schur functions
in Corollary~\ref{cor:main_LP}, we obtain the following result, which is a
corrected version of a conjecture proposed by Morales, Pak, and Panova
\cite{MPP2}.
\begin{cor}
\label{cor:mpp}
Let $\theta=(\theta_1,\dots,\theta_k)$ be the Lascoux--Pragacz decomposition
of a skew shape $\lm$ with $\ell(\lambda)\le d$. Then
\begin{equation}\label{eq:mpp2}
s_\mu(\vec x|\vec a) s_\lambda(\vec x|\vec a)^{k-1} =
\det\left((-1)^{\chi(p_j>q_i)} s_{\lambda(q_i,p_j-1)}(\vec x|\vec a)
\right)_{i,j=1}^k,
\end{equation}
where $\vec x = (x_1,\dots,x_d)$ and $\vec a = (a_1,a_2,\dots)$.
\end{cor}
\begin{remark}
The original conjecture of Morales, Pak, and Panova
\cite[Conjecture~7.9]{MPP2} states that, under the same assumption in
Corollary~\ref{cor:mpp},
\begin{equation}\label{eq:mpp}
s_\mu(\vec x|\vec a) s_\lambda(\vec x|\vec a)^{k-1} =
\det\left(s_{\lambda\setminus \lambda^0[p_j,q_i]}(\vec x|\vec a)
\right)_{i,j=1}^k.
\end{equation}
In fact \eqref{eq:mpp} is not true even for the Schur function case. As a
counterexample, if $\lambda=(2,1)$, $\mu=(1)$, $\vec x = (x_1,x_2)$, and $\vec
a=(0,0,\dots)$, then the Lascoux--Pragacz decomposition
$\theta=(\theta_1,\theta_2)$ has two border strips with $p_1=q_1=1$ and
$p_2=q_2=-1$. One can easily check using Pieri's rule that the left hand side of
\eqref{eq:mpp} is
\begin{equation}\label{eq:ce}
s_\mu(\vec x) s_\lambda(\vec x) = s_{(3,1)}(\vec x) + s_{(2,2)}(\vec x)
+s_{(2,1,1)}(\vec x) = s_{(3,1)}(\vec x) + s_{(2,2)}(\vec x).
\end{equation}
However, the right hand side of \eqref{eq:mpp} is
\begin{align*}
\det\left(s_{\lambda\setminus \lambda^0[p_j,q_i]}(\vec x) \right)_{i,j=1}^2
&= \det
\begin{pmatrix}
s_{\lambda\setminus \lambda^0[1,1]}(\vec x) & s_{\lambda\setminus \lambda^0[-1,1]}(\vec x) \\
s_{\lambda\setminus \lambda^0[1,-1]}(\vec x) & s_{\lambda\setminus \lambda^0[-1,-1]}(\vec x) \\
\end{pmatrix}\\
&= \det
\begin{pmatrix}
s_{(1,1)}(\vec x) & s_{\emptyset}(\vec x) \\
0 & s_{(2)}(\vec x) \\
\end{pmatrix}\\
&= s_{(3,1)}(\vec x) + s_{(2,1,1)}(\vec x) = s_{(3,1)}(\vec x).
\end{align*}
Note that the right hand side of \eqref{eq:mpp2} for the running example is
\begin{align*}
\det\left((-1)^{\chi(p_j>q_i)} s_{\lambda(q_i,p_j-1)}(\vec x) \right)_{i,j=1}^2
&= \det
\begin{pmatrix}
s_{(1,1)}(\vec x) & s_{\emptyset}(\vec x) \\
-s_{(2,2)}(\vec x) & s_{(2)}(\vec x) \\
\end{pmatrix}\\
&= s_{(3,1)}(\vec x) + s_{(2,1,1)}(\vec x) + s_{(2,2)}(\vec x)= s_{(3,1)}(\vec x)
+s_{(2,2)}(\vec x),
\end{align*}
which is equal to $s_\mu(\vec x) s_\lambda(\vec x)$ as shown in \eqref{eq:ce}.
\end{remark}
\section{A generalized Hamel--Goulden formula}
\label{sec:gener-hamel-gould}
In this section we give a generalization of the Hamel--Goulden formula. Our
result involves generalizations of cutting strips as well as Schur functions. As corollaries we
obtain Jin's result \cite{Jin_2018} and a generalized Giambelli formula.
We first introduce some definitions.
\begin{defn}\label{defn:compatible}
Let $\lambda$ and $\nu$ be partitions with $\lambda\subseteq\nu$. A border
strip $\gamma$ is \emph{$\nl$-compatible} if the following condition holds:
\begin{itemize}
\item $\Cont(\nl)\subseteq \Cont(\lambda)$, and
\item for any connected component $\alpha$ of $\nl$, we have
\[
\lambda^0[a-1,b+1] = \gamma[a-1,b+1],
\]
where $a=\min(\Cont(\alpha))$ and $b=\max(\Cont(\alpha))$.
\end{itemize}
See Figure~\ref{fig:compatible}.
\end{defn}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\fill[fill=cyan!10!white] (0,2) rectangle (1,4)
(1,3) rectangle (2,4)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)
(6,8)--(7,8)--(7,9)--(8,9)--(8,10)--(10,10)--(10,11)--(7,11)--(7,10)--(6,10)--(6,8);
\draw (0,2)--(1,2)--(1,3)--(2,3)--(2,5)--(4,5)--(4,7)--(7,7)--(7,9)--(8,9)--(8,10)--(11,10)--(11,13)--(12,13)--(12,14)--(0,14)--(0,2)--cycle
(1,4)--(1,6)--(3,6)--(3,5)
(1,5)--(2,5)--(2,6)
(11,11)--(10,11)--(10,14)
(10,13)--(11,13)--(11,14)
(10,12)--(11,12)--(11,13);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,3)--(2,3)--(2,4)--(0,4)--(0,2)--cycle
(0,3)--(1,3)--(1,4)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8)
(6,8)--(7,8)--(7,9)--(8,9)--(8,10)--(10,10)--(10,11)--(7,11)--(7,10)--(6,10)--(6,8)--cycle
(6,9)--(7,9)--(7,10)--(8,10)--(8,11)
(9,10)--(9,11);
\fill [fill=red!10!white] (1.1, 2)--(1.1, 2.9)--(2, 2.9)--(2,2)--(1.1,2)--cycle
(4.1, 6)--(6,6)--(6,6.9)--(4.1, 6.9)--(4.1, 6)--cycle
(7.1, 8)--(7.1, 8.9)--(8.1, 8.9)--(8.1, 9.9)--(10, 9.9)--(10, 8)--(7.1, 8)--cycle;
\draw [very thick, color=dredcolor] (1.1, 2)--(1.1, 2.9)--(2, 2.9)--(2,2)--(1.1,2)--cycle
(4.1, 6)--(6,6)--(6,6.9)--(4.1, 6.9)--(4.1, 6)--cycle
(5,6)--(5,6.9)
(8.1,8.95)--(10,8.95)
(8.1, 8)--(8.1,8.95)
(9.05, 8)--(9.05, 9.9)
(7.1, 8)--(7.1, 8.9)--(8.1, 8.9)--(8.1, 9.9)--(10, 9.9)--(10, 8)--(7.1, 8)--cycle;
\node[] at (3, 10.5) {$\lambda$};
\node[] at (9,6) {$\nu$};
\node[] at (.5, 2.5) {\tiny $-11$};
\node[] at (.5, 3.5) {\tiny $-10$};
\node[] at (1.5, 3.5) {\small $-9$};
\node[] at (1.5, 4.5) {\small $-8$};
\node[] at (1.5, 5.5) {\small $-7$};
\node[] at (2.5, 5.5) {\small $-6$};
\node[] at (3.5, 5.5) {\small $-5$};
\node[] at (3.5, 6.5) {\small $-4$};
\node[] at (3.5, 7.5) {\small $-3$};
\node[] at (4.5, 7.5) {\small $-2$};
\node[] at (5.5, 7.5) {\small $-1$};
\node[] at (6.5, 7.5) {\small $0$};
\node[] at (6.5, 8.5) {\small $1$};
\node[] at (6.5, 9.5) {\small $2$};
\node[] at (7.5, 9.5) {\small $3$};
\node[] at (7.5, 10.5) {\small $4$};
\node[] at (8.5, 10.5) {\small $5$};
\node[] at (9.5, 10.5) {\small $6$};
\node[] at (10.5, 10.5) {\small $7$};
\node[] at (10.5, 11.5) {\small $8$};
\node[] at (10.5, 12.5) {\small $9$};
\node[] at (10.5, 13.5) {\small $10$};
\node[] at (11.5, 13.5) {\small $11$};
\end{tikzpicture}\quad
\begin{tikzpicture}[scale=.46]
\fill[fill=cyan!10!white] (0,0)--(1,0)--(1,1)--(2,1)--(2,2)--(0,2)--(0,0)--cycle
(5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle
(7,5)--(8,5)--(8,6)--(9,6)--(9,7)--(11,7)--(11,8)--(8,8)--(8,7)--(7,7)--(7,5)--cycle;
\draw (2,1)--(6,1)--(6,2)--(2,2)--(2,1)--cycle
(3,1)--(3,2)
(4,1)--(4,2)
(5,1)--(5,2)
(7,4)--(7,5)
(8,4)--(8,5)
(11,7)--(12,7)--(12,8)--(11,8);
\draw[very thick, color=frenchblue] (0,0)--(1,0)--(1,1)--(2,1)--(2,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)--(1,2)
(5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle
(5,3)--(6,3)--(6,4)
(7,3)--(7,4)
(7,5)--(8,5)--(8,6)--(9,6)--(9,7)--(11,7)--(11,8)--(8,8)--(8,7)--(7,7)--(7,5)--cycle
(7,6)--(8,6)--(8,7)--(9,7)--(9,8)
(10,7)--(10,8);
\node[] at (.5, .5) {\tiny $-11$};
\node[] at (.5, 1.5) {\tiny $-10$};
\node[] at (1.5, 1.5) {\small $-9$};
\node[] at (2.5, 1.5) {\small $-8$};
\node[] at (3.5, 1.5) {\small $-7$};
\node[] at (4.5, 1.5) {\small $-6$};
\node[] at (5.5, 1.5) {\small $-5$};
\node[] at (5.5, 2.5) {\small $-4$};
\node[] at (5.5, 3.5) {\small $-3$};
\node[] at (6.5, 3.5) {\small $-2$};
\node[] at (7.5, 3.5) {\small $-1$};
\node[] at (7.5, 4.5) {\small $0$};
\node[] at (7.5, 5.5) {\small $1$};
\node[] at (7.5, 6.5) {\small $2$};
\node[] at (8.5, 6.5) {\small $3$};
\node[] at (8.5, 7.5) {\small $4$};
\node[] at (9.5, 7.5) {\small $5$};
\node[] at (10.5, 7.5) {\small $6$};
\node[] at (11.5, 7.5) {\small $7$};
\node[] at (10, 3) {$\gamma$};
\node[] at (0,-1.5) {\phantom{1}};
\end{tikzpicture}
\caption{A skew shape $\nl$ and the outer strip $\lambda^0$ with contents are
shown on the left, where the cells of $\nl$ are colored red. The border strip
$\gamma$ on the right is $\nl$-compatible. }
\label{fig:compatible}
\end{figure}
\begin{defn}\label{defn:comp2}
Let $\lambda$ and $\nu$ be partitions with $\lambda\subseteq\nu$. A partition
$\mu\subseteq\lambda$ is \emph{$\nl$-compatible} if the following condition
holds:
\begin{itemize}
\item $\Cont(\nl)\subseteq \Cont(\lambda)$, and
\item for any connected component $\alpha$ of $\nl$, we have
\[
\lambda^0[a,b] = \mu^+[a,b],
\]
where $a=\min(\Cont(\alpha))$ and $b=\max(\Cont(\alpha))$.
\end{itemize}
See Figure~\ref{fig:compatible2}.
\end{defn}
In Definitions~\ref{defn:compatible} and \ref{defn:comp2} one can use $\lambda^+$
instead of $\lambda^0$; if $\alpha$ is a connected component of $\nl$, then
since $\Cont(\alpha)\subseteq\Cont(\lambda)$, we have
\[
\lambda^0[a-1,b+1] = \lambda^+[a-1,b+1],
\]
where $a=\min(\Cont(\alpha))$ and $b=\max(\Cont(\alpha))$.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.43]
\fill[fill=cyan!10!white] (0,1) rectangle (1,2)
(2,6) rectangle ( 4,7)
(4,9) rectangle (6,10)
(5, 10) rectangle (7,11);
\draw (0,0)--(1,0)--(1,1)--(2,1)--(2,3)--(4,3)--(4,5)--(7,5)--(7,7)--(8,7)--(8,8)--(11,8)--(11,11)--(12,11)--(12,12)--(0,12)--(0,0)--cycle;
\draw [line width= 1.5pt, color=oxfordblue] (0,4)--(1,4)--(1,5)--(2,5)--(2,7)--(3,7)--(3,10)--(5,10)--(5,11)--(8,11)--(8,12)--(0,12)--(0,4)--cycle;
\draw[very thick, color=frenchblue] (0,1)--(1,1)--(1,2)--(0,2)--(0,1)--cycle
(2,6)--(4,6)--(4,7)--(2,7)--(2,6)--cycle
(3,6)--(3,7)
(4,9)--(6,9)--(6,10)--(7,10)--(7,11)--(5,11)--(5,10)--(4,10)--(4,9)--cycle
(5,9)--(5,10)--(6,10)--(6,11);
\fill [fill=red!10!white] (1.1,0)--(2,0)--(2,.9)--(1.1,.9)--(1.1,0)--cycle
(4.1,4)--(6,4)--(6,4.9)--(4.1,4.9)--(4.1,4)--cycle
(7.1,6)--(10,6)--(10,7.9)--(8.1,7.9)--(8.1,6.9)--(7.1,6.9)--(7.1,6)--cycle;
\draw [line width=1.3pt, color=dredcolor] (1.1,0)--(2,0)--(2,.9)--(1.1,.9)--(1.1,0)--cycle
(4.1,4)--(6,4)--(6,4.9)--(4.1,4.9)--(4.1,4)--cycle
(7.1,6)--(10,6)--(10,7.9)--(8.1,7.9)--(8.1,6.9)--(7.1,6.9)--(7.1,6)--cycle
(5,4)--(5,4.9)
(8.1, 6.9)--(10, 6.9)
(8.1, 6)--(8.1, 7.1)
(9.05,6)--(9.05,7.9);
\draw[thick, dashed, color=orange] (.5, 1.5)--(1.5, .5)
(2.5, 6.5)--(4.5, 4.5)
(3.5, 6.5)--(5.5, 4.5)
(6.5, 10.5)--(9.5, 7.5)
(4.5, 9.5)--(7.5, 6.5);
\filldraw[color=orange] (.5, 1.5) circle (3pt)
(1.5, .5) circle (3pt)
(2.5, 6.5) circle (3pt)
(4.5, 4.5) circle (3pt)
(3.5, 6.5) circle (3pt)
(5.5, 4.5) circle (3pt)
(4.5, 9.5) circle (3pt)
(7.5, 6.5) circle (3pt)
(6.5, 10.5) circle (3pt)
(9.5, 7.5) circle (3pt);
\node[] at (1.5,9.6) {$\mu$};
\end{tikzpicture}
\caption{A $\nl$-compatible partition $\mu$ for the skew shape $\nl$ in
Figure~\ref{fig:compatible}.}
\label{fig:compatible2}
\end{figure}
Note that if $\nu=\lambda$, then by definition every border strip $\gamma$ and
every partition $\mu\subseteq\lambda$ are $\nu/\lambda$-compatible.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.48]
\fill[fill=cyan!10!white] (0,0)--(1,0)--(1,1)--(2,1)--(2,2)--(0,2)--(0,0)--cycle
(5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle
(7,5)--(8,5)--(8,6)--(9,6)--(9,7)--(11,7)--(11,8)--(8,8)--(8,7)--(7,7)--(7,5)--cycle;
\draw (2,1)--(6,1)--(6,2)--(2,2)--(2,1)--cycle
(3,1)--(3,2)
(4,1)--(4,2)
(5,1)--(5,2)
(7,4)--(7,5)
(8,4)--(8,5)
(11,7)--(12,7)--(12,8)--(11,8);
\draw[very thick, color=frenchblue] (0,0)--(1,0)--(1,1)--(2,1)--(2,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)--(1,2)
(5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle
(5,3)--(6,3)--(6,4)
(7,3)--(7,4)
(7,5)--(8,5)--(8,6)--(9,6)--(9,7)--(11,7)--(11,8)--(8,8)--(8,7)--(7,7)--(7,5)--cycle
(7,6)--(8,6)--(8,7)--(9,7)--(9,8)
(10,7)--(10,8);
\fill [fill=red!10!white] (1.1, 0)--(1.1, .9)--(2, .9)--(2,0)--(1.1,0)--cycle
(6.1, 2)--(8,2)--(8,2.9)--(6.1, 2.9)--(6.1, 2)--cycle
(8.1, 5)--(8.1, 5.9)--(9.1, 5.9)--(9.1, 6.9)--(11, 6.9)--(11, 5)--(8.1, 5)--cycle;
\draw [line width=1.3pt, color=dredcolor] (1.1,0)--(2,0)--(2,.9)--(1.1,.9)--(1.1,0)--cycle
(6.1, 2)--(8,2)--(8,2.9)--(6.1, 2.9)--(6.1, 2)--cycle
(8.1, 5)--(8.1, 5.9)--(9.1, 5.9)--(9.1, 6.9)--(11, 6.9)--(11, 5)--(8.1, 5)--cycle
(7,2)--(7,2.9)
(9.1, 5.9)--(11, 5.9)
(9.1, 5)--(9.1, 6.1)
(10.05,5)--(10.05,6.9);
\node[] at (.5, .5) {\tiny $-11$};
\node[] at (.5, 1.5) {\tiny $-10$};
\node[] at (1.5, 1.5) {\small $-9$};
\node[] at (2.5, 1.5) {\small $-8$};
\node[] at (3.5, 1.5) {\small $-7$};
\node[] at (4.5, 1.5) {\small $-6$};
\node[] at (5.5, 1.5) {\small $-5$};
\node[] at (5.5, 2.5) {\small $-4$};
\node[] at (5.5, 3.5) {\small $-3$};
\node[] at (6.5, 3.5) {\small $-2$};
\node[] at (7.5, 3.5) {\small $-1$};
\node[] at (7.5, 4.5) {\small $0$};
\node[] at (7.5, 5.5) {\small $1$};
\node[] at (7.5, 6.5) {\small $2$};
\node[] at (8.5, 6.5) {\small $3$};
\node[] at (8.5, 7.5) {\small $4$};
\node[] at (9.5, 7.5) {\small $5$};
\node[] at (10.5, 7.5) {\small $6$};
\node[] at (11.5, 7.5) {\small $7$};
\node[] at (9.5, 0) {$\gamma\oplus\nl$};
\end{tikzpicture}
\caption{The skew shape $\gamma\oplus \nl$ for the border strip $\gamma$ and the
skew shape $\nl$ in Figure~\ref{fig:compatible}. The contents of the cells in
$\gamma$ are shown and the cells coming from $\nl$ are colored red.}
\label{fig:gamma+nl}
\end{figure}
Recall that we have defined $\gamma[a,b]=\{x\in \gamma: c(x)\in [a,b]\}$ for a
border strip $\gamma$. We extend this definition to any connected skew shape
$\alpha$, that is,
\[
\alpha[a,b]=\{x\in \alpha: c(x)\in [a,b]\}.
\]
\begin{defn}
Suppose that $\gamma$ is a $\nl$-compatible border strip. Define $\gamma\oplus
\nl$ to be the skew shape obtained from $\gamma$ by gluing each connected
component $\alpha$ of $\nl$ below $\gamma$ after shifting $\alpha$ diagonally
so that the southeast boundary of $\gamma$ and the northwest boundary of
$\alpha$ have common edges. See Figure~\ref{fig:gamma+nl}.
\end{defn}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.48]
\draw (2,1)--(4,1)--(4,2)--(2,2)
(3,1)--(3,2);
\fill[fill=cyan!10!white] (0,0)--(1,0)--(1,1)--(2,1)--(2,2)--(0,2)--(0,0)--cycle;
\draw[very thick, color=frenchblue] (0,0)--(1,0)--(1,1)--(2,1)--(2,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)--(1,2);
\fill [fill=red!15!white] (1.1, 0)--(1.1, .9)--(2, .9)--(2,0)--(1.1,0)--cycle;
\draw [line width=1.3pt, color=dredcolor] (1.1, 0)--(1.1, .9)--(2, .9)--(2,0)--(1.1,0)--cycle;
\node[] at (.5, .5) {\tiny $-11$};
\node[] at (.5, 1.5) {\tiny $-10$};
\node[] at (1.5, 1.5) {\small $-9$};
\node[] at (2.5, 1.5) {\small $-8$};
\node[] at (3.5, 1.5) {\small $-7$};
\node[] at (3, -2) {\phantom{d}};
\end{tikzpicture}\qquad\qquad
\begin{tikzpicture}[scale=.48]
\draw (5,2)--(4,2)--(4,1)--(6,1)--(6,2)
(5,1)--(5,2);
\fill[fill=cyan!10!white] (5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle;
\draw[very thick, color=frenchblue] (5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle
(5,3)--(6,3)--(6,4)
(7,3)--(7,4);
\fill[fill=red!10!white] (6.1, 2)--(8,2)--(8,2.9)--(6.1, 2.9)--(6.1, 2)--cycle;
\draw [line width=1.3pt, color=dredcolor] (6.1, 2)--(8,2)--(8,2.9)--(6.1, 2.9)--(6.1, 2)--cycle
(7.05, 2)--(7.05, 2.9);
\node[] at (4.5, 1.5) {\small $-6$};
\node[] at (5.5, 1.5) {\small $-5$};
\node[] at (5.5, 2.5) {\small $-4$};
\node[] at (5.5, 3.5) {\small $-3$};
\node[] at (6.5, 3.5) {\small $-2$};
\node[] at (7.5, 3.5) {\small $-1$};
\node[] at (6, 0) {\phantom{d}};
\end{tikzpicture}\qquad\qquad
\begin{tikzpicture}[scale=.48]
\draw (5,2)--(4,2)--(4,1)--(6,1)--(6,2)
(5,1)--(5,2)
(4,2)--(3,2)--(3,1)--(4,1)
(7,4)--(7,5)
(8,4)--(8,5)
(11,7)--(12,7)--(12,8)--(11,8);
\fill[fill=cyan!10!white] (5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle
(7,5)--(8,5)--(8,6)--(9,6)--(9,7)--(11,7)--(11,8)--(8,8)--(8,7)--(7,7)--(7,5)--cycle;
\draw[very thick, color=frenchblue] (5,2)--(6,2)--(6,3)--(8,3)--(8,4)--(5,4)--(5,2)--cycle
(5,3)--(6,3)--(6,4)
(7,3)--(7,4)
(7,5)--(8,5)--(8,6)--(9,6)--(9,7)--(11,7)--(11,8)--(8,8)--(8,7)--(7,7)--(7,5)--cycle
(7,6)--(8,6)--(8,7)--(9,7)--(9,8)
(10,7)--(10,8);
\fill[fill=red!15!white] (6.1, 2)--(8,2)--(8,2.9)--(6.1, 2.9)--(6.1, 2)--cycle
(8.1, 5)--(8.1, 5.9)--(9.1, 5.9)--(9.1, 6.9)--(11, 6.9)--(11, 5)--(8.1, 5)--cycle;
\draw [line width=1.3pt, color=dredcolor] (6.1, 2)--(8,2)--(8,2.9)--(6.1, 2.9)--(6.1, 2)--cycle
(8.1, 5)--(8.1, 5.9)--(9.1, 5.9)--(9.1, 6.9)--(11, 6.9)--(11, 5)--(8.1, 5)--cycle
(7,2)--(7,2.9)
(9.1, 5.9)--(11, 5.9)
(9.1, 5)--(9.1, 6.1)
(10.05,5)--(10.05,6.9);
\node[] at (3.5, 1.5) {\small $-7$};
\node[] at (4.5, 1.5) {\small $-6$};
\node[] at (5.5, 1.5) {\small $-5$};
\node[] at (5.5, 2.5) {\small $-4$};
\node[] at (5.5, 3.5) {\small $-3$};
\node[] at (6.5, 3.5) {\small $-2$};
\node[] at (7.5, 3.5) {\small $-1$};
\node[] at (7.5, 4.5) {\small $0$};
\node[] at (7.5, 5.5) {\small $1$};
\node[] at (7.5, 6.5) {\small $2$};
\node[] at (8.5, 6.5) {\small $3$};
\node[] at (8.5, 7.5) {\small $4$};
\node[] at (9.5, 7.5) {\small $5$};
\node[] at (10.5, 7.5) {\small $6$};
\node[] at (11.5, 7.5) {\small $7$};
\end{tikzpicture}
\caption{The diagrams $(\gamma\oplus\nl)[-11,-7]$ (left),
$(\gamma\oplus\nl)[-6,-1]$ (middle), and $(\gamma\oplus\nl)[-7,7]$
(right), for the $\gamma\oplus\nl$ in Figure~\ref{fig:gamma+nl}.}
\label{fig:gamma+nl[a,b]}
\end{figure}
Observe that if $\gamma$ is a $\nl$-compatible border strip and
\begin{equation}
\label{eq:a,b}
a,b\in (\Cont(\lambda)\setminus\Cont(\nl))
\cup\{\min(\Cont(\lambda)), \max(\Cont(\lambda))\},
\end{equation}
then $(\gamma\oplus\nl)[a,b]$ is a connected skew shape. See
Figure~\ref{fig:gamma+nl[a,b]}.
We now state the generalized Hamel--Goulden formula.
\begin{thm}\label{thm:main_HG}
Let $\mu\subseteq \lambda\subseteq\nu$ be partitions. Suppose that $\mu$ is
$\nl$-compatible and $\gamma$ is a $\nl$-compatible border strip. Let
$\theta=(\theta_1,\dots,\theta_k)$ be the decomposition of $\lm$ determined by
the cutting strip $\gamma$. If $\alpha_1,\dots,\alpha_\ell$ are the connected
components of $\nl$, we have
\[
\ts_{\nu/\mu} \prod_{s=1}^\ell \ts_{\alpha_s}^{r_s-1}
= \det \left( \ts_{(\gamma\oplus\nl)[p_j,q_i]}\right)_{i,j=1}^k,
\]
where $r_s$ is the number of strips $\theta_i$ such that
$\Cont(\alpha_s)\subseteq\Cont(\theta_i)$.
\end{thm}
Before proving Theorem~\ref{thm:main_HG} we give an example and some of its
applications.
\begin{exam}
Consider the partitions $\nu=(8,8,8,6,6,5,4)$, $\lambda=(8,8,6,6,6,3,3)$, and
$\mu=(4,3)$ as shown in Figure~\ref{fig:lmn}. The border strip $\gamma$ in
Figure~\ref{fig:gt} is $\nu/\lambda$-compatible. The decomposition $\theta$ of
$\lambda/\mu$ with cutting strip $\gamma$ is also shown in
Figure~\ref{fig:gt}. Since $C(\alpha_1)\subseteq C(\theta_i)$ for $i=1,2,4$
and $C(\alpha_2)\subseteq C(\theta_j)$ for $j=1,2$, we have $r_1=3$ and
$r_2=2$. Thus by Theorem \ref{thm:main_HG}, $\ts_{\nu/\mu} \ts_{\alpha_1}^2
\ts_{\alpha_2}$ is equal to the determinant shown in Figure~\ref{fig:det3}.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\fill[fill=red!10!white] (3,0) rectangle (5,2)
(6,4) rectangle (8,5);
\draw[color=gray] (0,1)--(4,1)
(0,2)--(3,2)
(0,3)--(6,3)
(0,4)--(6,4)
(3,5)--(6,5)
(4,6)--(8,6)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,1)--(4,6)
(5,2)--(5,7)
(6,5)--(6,7)
(7,4)--(7,7)
(4,0)--(4,1)--(5,1);
\draw[dashed, color=gray] (0,5)--(0,7)--(4,7)
(0,6)--(3,6)
(1,5)--(1,7)
(2,5)--(2,7)
(3,6)--(3,7);
\draw [line width=1.3pt, color=dredcolor] (3,0)--(5,0)--(5,2)--(3,2)--(3,0)--cycle
(6,4)--(8,4)--(8,5)--(6,5)--(6,4)--cycle;
\draw[thick, line width=1.2pt] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(8,5)--(8,7)--(4,7)--(4,6)--(3,6)--(3,5)--(0,5)--(0,0)--cycle;
\node[] at (-.7, 6) {$\mu$};
\node[] at (8.7, 6) {$\lambda$};
\node[] at (7, 2.6) {$\nu$};
\node[] at (5.8, .8) {\color{red}$\alpha_1$};
\node[] at (8.7, 4.4) {\color{red}$\alpha_2$};
\end{tikzpicture}
\caption{The skew shape $\lm$ (black) and the connected components $\alpha_1$
and $\alpha_2$ of $\nl$ (red).}
\label{fig:lmn}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (0,0)--(1,0)--(1,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6)
(6,7)--(7,7)--(7,8)--(6,8);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (.5, .5) {\small $-6$};
\node[] at (.5, 1.5) {\small $-5$};
\node[] at (.5, 2.5) {\small $-4$};
\node[] at (.5, 3.5) {\small $-3$};
\node[] at (.5, 4.5) {\small $-2$};
\node[] at (1.5, 4.5) {\small $-1$};
\node[] at (2.5, 4.5) {\small $0$};
\node[] at (2.5, 5.5) {\small $1$};
\node[] at (2.5, 6.5) {\small $2$};
\node[] at (3.5, 6.5) {\small $3$};
\node[] at (3.5, 7.5) {\small $4$};
\node[] at (4.5, 7.5) {\small $5$};
\node[] at (5.5, 7.5) {\small $6$};
\node[] at (6.5, 7.5) {\small $7$};
\node[] at (5, 3.5) {$\gamma$};
\end{tikzpicture}
\qquad\qquad
\begin{tikzpicture}[scale=.46]
\draw[color=gray] (0,1)--(3,1)
(0,2)--(3,2)
(0,3)--(6,3)
(0,4)--(6,4)
(3,5)--(6,5)
(4,6)--(8,6)
(1,0)--(1,5)
(2,0)--(2,5)
(3,2)--(3,5)
(4,2)--(4,6)
(5,2)--(5,7)
(6,5)--(6,7)
(7,5)--(7,7);
\draw[dashed, color=gray] (0,5)--(0,7)--(4,7)
(0,6)--(3,6)
(1,5)--(1,7)
(2,5)--(2,7)
(3,6)--(3,7);
\draw[thick, line width=1.2pt] (0,0)--(3,0)--(3,2)--(6,2)--(6,5)--(8,5)--(8,7)--(4,7)--(4,6)--(3,6)--(3,5)--(0,5)--(0,0)--cycle
(1,0)--(1,4)--(3,4)--(3,5)
(2,0)--(2,3)--(4,3)--(4,5)--(5,5)--(5,6)--(8,6)
(5,2)--(5,4)--(6,4);
\node[] at (1.5, .5) {$\color{blue} p_1$};
\node[] at (7.5, 6.5) {$\color{red} q_1$};
\node[] at (2.5, .5) {$\color{blue} p_2$};
\node[] at (7.5, 5.5) {$\color{red} q_2$};
\node[] at (5.5, 2.5) {$\color{blue} p_3$};
\node[] at (5.5, 3.5) {$\color{red} q_3$};
\node[] at (.5, .5) {$\color{blue} p_4$};
\node[] at (2.5, 4.5) {$\color{red} q_4$};
\node[] at (8.5, 6.5) {$\theta_1$};
\node[] at (8.5, 5.5) {$\theta_2$};
\node[] at (6.5, 3.5) {$\theta_3$};
\node[] at (-.5, .5) {$\theta_4$};
\end{tikzpicture}
\caption{A border strip $\gamma$ (left) and the decomposition $\theta$ (right)
of $\lm$ with cutting strip $\gamma$. }
\label{fig:gt}
\end{figure}
\begin{figure}
\centering
\[
\det \begin{bmatrix}
~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3)
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (0,1)--(1,1)--(1,2)--(0,2)--(0,1)--cycle
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6)
(6,7)--(7,7)--(7,8)--(6,8);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\node[] at (6.5, 7.5) {\tiny $7$};
\end{tikzpicture}~&
~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3)
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6)
(6,7)--(7,7)--(7,8)--(6,8);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (1.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\node[] at (6.5, 7.5) {\tiny $7$};
\end{tikzpicture}~&
~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white]
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor]
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white]
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6)
(6,7)--(7,7)--(7,8)--(6,8);
\draw[very thick, color=frenchblue]
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (1.5, .5) {\tiny \phantom{$-6$}};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\node[] at (6.5, 7.5) {\tiny $7$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3)
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (0,0)--(1,0)--(1,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6)
(6,7)--(7,7)--(7,8)--(6,8);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (.5, .5) {\tiny $-6$};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\node[] at (6.5, 7.5) {\tiny $7$};
\end{tikzpicture}~\\
~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3)
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (0,1)--(1,1)--(1,2)--(0,2)--(0,1)--cycle
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3)
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (1.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\end{tikzpicture}~&
~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white]
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor]
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white]
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\draw[very thick, color=frenchblue]
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (1.5, .5) {\tiny \phantom{$-6$}};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3)
(5,6)--(5,7)
(4,6)--(6,6)--(6,7)--(4,7)--(4,6)--cycle;
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle;
\draw[thick] (0,0)--(1,0)--(1,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5)
(3,6)--(4,6)--(4,7)--(6,7)--(6,8)--(3,8)--(3,6)--cycle
(3,7)--(4,7)--(4,8)
(5,7)--(5,8);
\node[] at (.5, .5) {\tiny $-6$};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\node[] at (3.5, 6.5) {\tiny $3$};
\node[] at (3.5, 7.5) {\tiny $4$};
\node[] at (4.5, 7.5) {\tiny $5$};
\node[] at (5.5, 7.5) {\tiny $6$};
\end{tikzpicture}~\\
~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3);
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle;
\draw[thick] (0,1)--(1,1)--(1,2)--(0,2)--(0,1)--cycle
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5);
\node[] at (.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3);
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle;
\draw[thick]
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5);
\node[] at (.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.38]
\draw[thick]
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\node[] at (2.5, .5) {\tiny\phantom{ $-6$}};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\end{tikzpicture}~&~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3);
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle;
\draw[thick] (0,0)--(1,0)--(1,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1)
(2,5)--(3,5)--(3,7)--(2,7)--(2,5)--cycle
(2,6)--(3,6);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5);
\node[] at (.5, .5) {\tiny $-6$};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny $1$};
\node[] at (2.5, 6.5) {\tiny $2$};
\end{tikzpicture}~\\
~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3);
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle;
\draw[thick] (0,1)--(1,1)--(1,2)--(0,2)--(0,1)--cycle
(0,1)--(1,1);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5);
\node[] at (.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\node[] at (2.5, 5.5) {\tiny \phantom{$1$}};
\end{tikzpicture}~& ~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3);
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle;
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5);
\node[] at (.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\end{tikzpicture}~&
~\begin{tikzpicture}[scale=.38]
\node[] at (.5, .5) {\tiny \phantom{$-6$}};
\node[] at (.5, 3.5) {$\emptyset$};
\end{tikzpicture}~& ~\begin{tikzpicture}[scale=.38]
\fill[fill=red!10!white] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle;
\draw [line width=.9pt, color=dredcolor] (1,2)--(3,2)--(3,4)--(1,4)--(1,2)--cycle
(2,2)--(2,4)
(1,3)--(3,3);
\fill[fill=cyan!10!white] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle;
\draw[thick] (0,0)--(1,0)--(1,2)--(0,2)--(0,0)--cycle
(0,1)--(1,1);
\draw[very thick, color=frenchblue] (0,2)--(1,2)--(1,4)--(3,4)--(3,5)--(0,5)--(0,2)--cycle
(0,3)--(1,3)
(0,4)--(1,4)--(1,5)
(2,4)--(2,5);
\node[] at (.5, .5) {\tiny $-6$};
\node[] at (.5, 1.5) {\tiny $-5$};
\node[] at (.5, 2.5) {\tiny $-4$};
\node[] at (.5, 3.5) {\tiny $-3$};
\node[] at (.5, 4.5) {\tiny $-2$};
\node[] at (1.5, 4.5) {\tiny $-1$};
\node[] at (2.5, 4.5) {\tiny $0$};
\end{tikzpicture}~
\end{bmatrix}
\]
\caption{An illustrate of the determinant $\det \left(
\ts_{(\gamma\oplus\nl)[p_j,q_i]}\right)_{i,j=1}^k$. The $(i,j)$ entry only
shows the skew shape $(\gamma\oplus\nl)[p_j,q_i]$, where the contents of the
cells coming from $\gamma$ are shown and the cells coming from $\nl$ are
colored red.}
\label{fig:det3}
\end{figure}
\end{exam}
If $\nu=\lambda$ in Theorem~\ref{thm:main_HG}, then we obtain the following
Hamel--Goulden formula for Macdonald's 9th variation of Schur functions, which
was also proved by Bachmann and Charlton \cite{Bachmann_2020}, and Foley and
King \cite{Foley20} using the Lindstr\"om--Gessel--Viennot lemma.
\begin{cor}\label{cor:HG}
Let $\lambda$ and $\mu$ be partitions with $\mu\subseteq \lambda$. Suppose
that $\gamma$ is a border strip with $\Cont(\lambda)\subseteq\Cont(\gamma)$
and $\theta=(\theta_1,\dots,\theta_k)$ is the decomposition of $\lm$ with
cutting strip $\gamma$. Then
\[
\ts_{\lambda/\mu} = \det \left( \ts_{\gamma[p_j,q_i]}\right)_{i,j=1}^k.
\]
\end{cor}
If $\nu/\lambda$ is a disjoint union of single cells and if we restrict
Theorem~\ref{thm:main_HG} to Schur functions, then we obtain Jin's result
\cite[Theorem~2]{Jin_2018} for the cases when the ``enriched diagrams'' are not necessary.
For sequences $\vec a = (a_1,\dots,a_r)$ and $\vec b = (b_1,\dots,b_r)$ with
$a_1>\dots>a_r\ge0$ and $b_1>\dots>b_r\ge0$, the \emph{Frobenius notation}
$(\vec a|\vec b)$ denotes the partition
\[
\{(i,i): 1\le i\le r\} \cup
\{(i,j): 1\le i\le r, i<j\le a_i\} \cup
\{(i,j): 1\le j\le r, j<i\le b_i\}.
\]
See Figure~\ref{fig:Frobenius_ex}.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.48]
\fill[fill=red!10!white] (1.1, 3.1) rectangle (4.9, 3.9);
\fill[fill=red!10!white] (2.1, 2.1) rectangle (3.9, 2.9);
\fill[fill=red!10!white] (3.1, 1.1) rectangle (3.9, 1.9);
\fill[fill=blue!10!white] (.1, .1) rectangle (.9, 2.9);
\fill[fill=blue!10!white] (1.1, 1.1) rectangle (1.9, 1.9);
\draw [thick] (0,0)--(0,4)--(5,4)--(5,3)--(4,3)--(4,1)--(1,1)--(1,0)--(0,0)--cycle;
\draw (0,1)--(1,1);
\draw (0,2)--(4,2);
\draw (0,3)--(4,3);
\draw (1,1)--(1,4);
\draw (2,1)--(2,4);
\draw (3,1)--(3,4);
\draw (4,3)--(4,4);
\draw[thick, color=dredcolor] (1.1,3.1)--(4.9, 3.1)--(4.9, 3.9)--(1.1, 3.9)--(1.1, 3.1)--cycle;
\draw[thick, color=dredcolor] (2.1,2.1)--(3.9, 2.1)--(3.9, 2.9)--(2.1, 2.9)--(2.1, 2.1)--cycle;
\draw[thick, color=dredcolor] (3.1,1.1)--(3.9, 1.1)--(3.9, 1.9)--(3.1, 1.9)--(3.1, 1.1)--cycle;
\draw[thick, color=dbluecolor] (.1,.1)--(.1, 2.9)--(.9, 2.9)--(.9, .1)--(.1, .1)--cycle;
\draw[thick, color=dbluecolor] (1.1,1.1)--(1.1, 1.9)--(1.9, 1.9)--(1.9, 1.1)--(1.1, 1.1)--cycle;
\draw [dashed, color=gray] (0,4)--(3,1);
\end{tikzpicture}
\caption{An illustrate of the Frobenius notation $(4,2,1|3,1,0)$ for the
partition $\lambda=(5,4,4,1)$.}\label{fig:Frobenius_ex}
\end{figure}
If we take $\nu=(\vec a\sqcup \vec c| \vec b\sqcup \vec d)$, $\lambda=(\vec a|
\vec b)$, $\mu=(\vec c|\vec d)$, and $\theta$ to be the Kreiman decomposition of
$\lm$ in Theorem~\ref{thm:main_HG}, then we obtain the following generalized
Giambelli formula.
\begin{cor}\label{cor:Giam}
Let $\vec a = (a_1,\dots,a_r)$, $\vec b = (b_1,\dots,b_r)$,
$\vec c = (c_1,\dots,c_s)$, and $\vec d = (d_1,\dots,d_s)$ be sequences of
integers such that $r,s\ge0$ and
\[
a_1>\dots>a_r>c_1>\dots>c_s\ge0, \qquad
b_1>\dots>b_r>d_1>\dots>d_s\ge0.
\]
Then
\[
\ts_{(\vec c|\vec d)}^{r-1} \ts_{(\vec a\sqcup \vec c| \vec b\sqcup \vec d)}
=\det\left( \ts_{(a_i\sqcup\vec c| b_j\sqcup\vec d)} \right) _{1\le i,j\le r}.
\]
\end{cor}
If $\vec c = \vec d=\emptyset$, then Corollary~\ref{cor:Giam} reduces to the
Giambelli formula. If $\vec c = \vec d=(0)$ in Corollary~\ref{cor:Giam}, we
obtain
\[
\ts_{(1)}^{r-1} \ts_{(\vec a\sqcup 0|\vec b\sqcup 0)}
=\det\left( \ts_{((a_i,0)| (b_j,0))} \right) _{1\le i,j\le r},
\]
where each entry has a \emph{near hook shape} $((a_i,0)| (b_j,0))$.
For the rest of this section we give a proof of Theorem~\ref{thm:main_HG}. We
first recall a known property of $\ts_{\lm}$.
Let $\alpha$ and $\beta$ be skew shapes and let $a$ and $b$ be, respectively,
the top-right corner of $\alpha$ and the bottom-left corner of $\beta$. We
define $\alpha\rightarrow \beta$ (resp.~$\alpha\uparrow \beta$) to be the skew
shape obtained from $\alpha$ by attaching $\beta$ so that $b$ is to the right
of $a$ (resp.~above $a$).
\begin{lem} \label{lem:combine}
\cite[Chapter I, \S5, Example~30 (d)]{Macdonald}
Let $\alpha$ and $\beta$ be skew shapes such that the top-right corner $a$ of $\alpha$
and the bottom-left corner $b$ of $\beta$ satisfy $c(b)=c(a)+1$. Then
\[
\ts_\alpha \ts_\beta = \ts_{\alpha\rightarrow \beta} + \ts_{\alpha\uparrow \beta}.
\]
\end{lem}
Chen, Yan, and Yang \cite{Chen_2005} showed that the Hamel--Goulden formula can
be obtained from the Lascoux--Pragacz formula using simple matrix operations.
Their proof uses only the fact that the Schur functions satisfy
\[
s_\alpha s_\beta = s_{\alpha\rightarrow \beta} + s_{\alpha\uparrow \beta}.
\]
Hence, by Lemma~\ref{lem:combine}, their proof extends to the identity in
Theorem~\ref{thm:main_HG}. This implies that it is sufficient to prove this
theorem for the case when $\theta$ is the Lascoux--Pragacz decomposition of
$\lm$. To this end we need three lemmas.
\begin{lem}\label{lem:9th}
The following properties hold.
\begin{enumerate}
\item If $\mu\not\subseteq\lambda$, then $\ts_{\lm}=0$.
\item If $\mu\subseteq\lambda$ and $\alpha_1,\dots,\alpha_s$ are the connected
components of $\lm$, then $\ts_{\lm}=\ts_{\alpha_1}\cdots \ts_{\alpha_s}$.
\end{enumerate}
\end{lem}
\begin{proof}
Both properties can be proved similarly as in the Schur function case
\cite[Chapter I, \S5, (5.7)]{Macdonald}.
\end{proof}
\begin{lem}\label{lem:LP1}
Suppose that $\mu\subseteq \lambda\subseteq\nu$ are partitions such that $\mu$
is $\nl$-compatible. Let $\alpha_1,\dots,\alpha_\ell$ be the connected
components of $\nl$ and let $\gamma$ be the outer strip of $\lambda$. Finally,
let $\theta=(\theta_1,\dots,\theta_k)$ be the Lascoux--Pragacz decomposition
of $\lm$. Then
\[
\ts_{\nu/\lambda(q_i,p_j-1)}=\ts_{(\gamma\oplus\nl)[p_j,q_i]} \prod_{s=1}^{\ell}
\ts_{\alpha_s}^{\chi(q_i<\min(\Cont(\alpha_s)))+\chi(p_j>\max(\Cont(\alpha_s)))}.
\]
\end{lem}
\begin{proof}
By Lemma~\ref{lem:C/C}, we have $p_j-1\ne q_i$ for all $i,j$. Thus there are
two cases $p_j\le q_i$ and $p_j-1>q_i$.
\textbf{Case 1:} $p_j\le q_i$. Since
$\lambda(q_i,p_j-1)=\lambda\setminus\lambda^0[p_j,q_i]$, the connected component
$\alpha_s$ of $\nl$ remains the same in $\nu/\lambda(q_i,p_j-1)$ if
$q_i<\min(\Cont(\alpha_s))$ or $p_j>\max(\Cont(\alpha_s))$. Otherwise,
$\Cont(\alpha_s)\subseteq [p_j,q_i]$ and therefore $\alpha_s$ is contained in
$(\gamma\oplus\nl)[p_j,q_i]$, which is a connected component of
$\nu/\lambda(q_i,p_j-1)$. Thus by Lemma~\ref{lem:9th},
\[
\ts_{\nu/\lambda(q_i,p_j-1)}=\ts_{(\gamma\oplus\nl)[p_j,q_i]} \prod_{s=1}^{\ell}
\ts_{\alpha_s}^{\chi(q_i<\min(\Cont(\alpha_s)) \mbox{\scriptsize{ or }} p_j>\max(\Cont(\alpha_s)))}.
\]
Since $q_i<\min(\Cont(\alpha_s))$ and $p_j>\max(\Cont(\alpha_s))$ cannot be satisfied
at the same time, the above equation is the same as the one in the lemma.
\textbf{Case 2:} $p_j-1>q_i$. Since $\ts_{(\gamma\oplus\nl)[p_j,q_i]}=0$, it
suffices to show that $\nu\not\subseteq\lambda(q_i,p_j-1)$, which implies
$\ts_{\nu/\lambda(q_i,p_j-1)}=0$ by Lemma~\ref{lem:9th}. Since
$\lambda(q_i,p_j-1)=\lambda\cup\lambda^+[q_i+1,p_j-1]$ has a cell with content
$p_j-1$, to show $\nu\not\subseteq\lambda(q_i,p_j-1)$ it is enough to show that
$p_j-1\not\in \Cont(\nl)$.
For a contradiction suppose $p_j-1\in \Cont(\nl)$. Then
$p_j-1\in\Cont(\alpha_s)$ for some $s$. Since $\mu$ is $\nl$-compatible and
$p_j-1\in\Cont(\alpha_s)$, we must have $\lambda^+[p_j-1,p_j] =
\mu^+[p_j-1,p_j]$. On the other hand, we have $p_j-1\in C_n(\mu)\setminus
C_n(\lambda)$ by Lemma~\ref{lem:C/C}. Since $p_j-1\in C_n(\mu)$, the border
strip $\mu^+[p_j-1,p_j]$ must be a vertical domino. However, since $p_j-1\not\in
C_n(\lambda)$, the border strip $\lambda^+[p_j-1,p_j]$ must be a horizontal
domino. Then $\lambda^+[p_j-1,p_j] \ne \mu^+[p_j-1,p_j]$, which is a
contradiction. Therefore we must have $p_j-1\not\in \Cont(\nl)$, which completes
the proof.
\end{proof}
\begin{lem}\label{lem:LP2}
Under the same assumptions in Lemma~\ref{lem:LP1}, we have
\[
\ts_{\nu/\mu} \prod_{s=1}^\ell \ts_{\alpha_s}^{k-1-I_s-J_s}
= \det \left( \ts_{(\gamma\oplus\nl)[p_j,q_i]}\right)_{i,j=1}^k,
\]
where $I_s$ is the number of $1\le i\le k$ such that
$q_i<\min(\Cont(\alpha_s))$
and $J_s$ is the number of $1\le j\le k$ such that
$p_j>\max(\Cont(\alpha_s))$.
\end{lem}
\begin{proof}
By \eqref{eq:main_LP2}, we have
\[
\ts_{\nu/\mu} \ts_{\nu/\lambda}^{k-1}
= \det \left( (-1)^{\chi(p_j>q_i)} \ts_{\nu/\lambda(q_i,p_j-1)}\right)_{i,j=1}^k.
\]
By Lemmas~\ref{lem:9th} and \ref{lem:LP1}, the above
equation can be written as
\begin{equation}\label{eq:ss}
\ts_{\nu/\mu} \prod_{s=1}^\ell \ts_{\alpha_s}^{k-1}
= \det \left(
\ts_{(\gamma\oplus\nl)[p_j,q_i]} \prod_{s=1}^{\ell}
\ts_{\alpha_s}^{\chi(q_i<\min(\Cont(\alpha_s)))+\chi(p_j>\max(\Cont(\alpha_s)))}
\right)_{i,j=1}^k,
\end{equation}
where the factor $(-1)^{\chi(p_j>q_i)}$ can be omitted because
$\ts_{(\gamma\oplus\nl)[p_j,q_i]}=0$ if $p_j>q_i$. By factoring out the factor
$\prod_{s=1}^{\ell} \ts_{\alpha_s}^{\chi(q_i<\min(\Cont(\alpha_s)))}$ from each
row $i$ and the factor $\prod_{s=1}^{\ell}
\ts_{\alpha_s}^{\chi(p_j>\max(\Cont(\alpha_s)))}$ from each column $j$ and
dividing both sides of \eqref{eq:ss} by these factors we obtain the desired
identity.
\end{proof}
Now we are ready to prove Theorem~\ref{thm:main_HG}.
\begin{proof}[Proof of Theorem~\ref{thm:main_HG}]
As we have already discussed, it suffices to show the theorem for the case
when $\theta$ is the Lascoux--Pragacz decomposition of $\lm$. Then, by
Lemma~\ref{lem:LP2}, it suffices to show that $k-I_s-J_s=r_s$ for all $s$. By
definition, $k-I_s-J_s$ is the number of border strips $\theta_i$ such that
$p(\theta_i)\le \max(\Cont(\alpha_s))$ and $\min(\Cont(\alpha_s))\le
q(\theta_i)$. Since $\mu$ is $\nl$-compatible and $\theta$ is the
Lascoux--Pragacz decomposition, every border strip $\theta_i$ with
$\Cont(\alpha_s)\cap\Cont(\theta_i)\ne\emptyset$ must satisfy
$\Cont(\alpha_s)\subseteq\Cont(\theta_i)$. This implies that $p(\theta_i)\le
\max(\Cont(\alpha_s))$ and $\min(\Cont(\alpha_s))\le q(\theta_i)$ if and only
if $\Cont(\alpha_s)\subseteq\Cont(\theta_i)$. Therefore the number of such
border strips $\theta_i$ is equal to $r_s$, which completes the proof.
\end{proof}
\section{A generalization of a converse of Hamel--Goulden's theorem}
\label{sec:conv-hamel-gould}
In this section we prove Theorem~\ref{thm:HG3}, which is a generalization of a
converse of Hamel--Goulden's theorem. We restate Theorem~\ref{thm:HG3} as
follows.
\begin{thm}\label{thm:HG_conv}
Let $\alpha$ be any connected skew shape. Suppose that $(a_1,\dots,a_k)$ and
$(b_1,\dots,b_k)$ are sequences of integers such that $\alpha[a_j,b_i]$ is a
skew shape for all $i,j$. Then either $\det \left(
s_{\alpha[a_j,b_i]}\right)_{i,j=1}^k = 0$ or there exists a skew shape
$\rho$ such that
\[
\det \left( s_{\alpha[a_j,b_i]}\right)_{i,j=1}^k = \pm s_{\rho}.
\]
\end{thm}
To prove this theorem we need the following lemma.
\begin{lem}\label{lem:gamma_to_la}
Let $\gamma$ be a border strip and let $(a_1<\dots<a_k)$ and $(b_1<\dots<b_k)$
be sequences of integers such that $a_i,b_i\in \Cont(\gamma)$ for all $i$, and
for any integer $a$,
\begin{equation}\label{eq:dyck}
|\{i: a_i\le a\}| \ge |\{i: b_i\le a\}|.
\end{equation}
Then there is a skew shape $\tau$ such that if $\theta=(\theta_1,\dots,\theta_r)$
is the decomposition of $\tau$ with cutting strip $\gamma$, then $r=k$ and
$\{p(\theta_1),\dots,p(\theta_k)\}=\{a_1,\dots,a_k\}$ and
$\{q(\theta_1),\dots,q(\theta_k)\}=\{b_1,\dots,b_k\}$.
\end{lem}
\begin{proof}
We proceed by induction on $k$. If $k=0$, we can take $\tau=\emptyset$. Let
$k\ge1$ and suppose that the lemma is true for $k-1$.
We consider the following two cases depending on the shape of
$\gamma[b_1,b_1+1]$.
\begin{description}
\item[Case 1] $\gamma[b_1,b_1+1]$ is a horizontal domino. Let
$\theta_0=\gamma[a_1,b_1]$. It is easy to see that the two sequences
$(a_2<\dots<a_k)$ and $(b_2<\dots<b_k)$ satisfy \eqref{eq:dyck}. Thus, by
the induction hypothesis, there is a skew shape $\sigma$ whose decomposition
with cutting strip $\gamma$ is $(\theta_1,\dots,\theta_{k-1})$ such that
$\{p(\theta_1),\dots,p(\theta_{k-1})\}=\{a_2,\dots,a_k\}$ and
$\{q(\theta_1),\dots,q(\theta_{k-1})\}=\{b_2,\dots,b_k\}$. The fact that
$a_1<a_2$ and $\gamma[b_1,b_1+1]$ is a horizontal domino guarantees that
$\tau=\sigma\cup \theta_0$, where $\theta_0$ is attached below $\sigma$
after an appropriate diagonal shift, is a skew shape satisfying the desired
properties.
\item[Case 2] $\gamma[b_1,b_1+1]$ is a vertical domino. Let
$\theta_0=\gamma[a_m,b_1]$, where $m$ is the largest integer such that
$a_m\le b_1$. It is easy to see that the two sequences
$(a_1<\dots<a_{m-1}<a_{m+1}<\dots<a_k)$ and $(b_2<\dots<b_k)$ satisfy
\eqref{eq:dyck}. Thus, by the induction hypothesis, there is a skew shape
$\sigma$ whose decomposition with cutting strip $\gamma$ is
$(\theta_1,\dots,\theta_{k-1})$ such that
$\{p(\theta_1),\dots,p(\theta_{k-1})\}=\{a_1,\dots,a_{m-1},a_{m+1},\dots,a_k\}$
and $\{q(\theta_1),\dots,q(\theta_{k-1})\}=\{b_2,\dots,b_k\}$. The fact that
$a_i<a_m$, for all $1\le i<m$, and $\gamma[b_1,b_1+1]$ is a vertical domino
guarantees that $\tau=\sigma\cup \theta_0$, where $\theta_0$ is attached
above $\sigma$ after an appropriate diagonal shift, is a skew shape
satisfying the desired properties.
\end{description}
The above two cases show that the lemma is true for $k$ and the proof is
completed by induction.
\end{proof}
For an illustration of the construction in the proof of
Lemma~\ref{lem:gamma_to_la}, consider the border strip $\gamma$ and the two
sequences $(a_1,a_2,a_3)$ and $(b_1,b_2,b_3)$ in Figure~\ref{fig:to_skew}. Since
$\gamma[b_1,b_1+1]$ is a horizontal domino, we construct the border strip
$\gamma[a_1,b_1]$ as in Figure~\ref{fig:to_skew1} (a). Since $\gamma[b_2,b_2+1]$
is a vertical domino, we construct the border strip $\gamma[a_3,b_2]$ as in
Figure~\ref{fig:to_skew1} (b). Finally we construct the border strip
$\gamma[a_2,b_3]$ as in Figure~\ref{fig:to_skew1} (c). By combining these border
strips after appropriate diagonal shifts we obtain the skew shape shown in
Figure~\ref{fig:to_skew}.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\draw[color=gray!30!white] (0,1)--(1,1)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)--(4,2)--(4,3)
(5,2)--(5,3)--(6,3)
(5,4)--(6,4)--(6,5)
(7,4)--(7,5)--(8,5)--(8,6);
\draw[thick] (0,0)--(1,0)--(1,1)--(4,1)--(4,2)--(6,2)--(6,4)--(8,4)--(8,5)--(9,5)--(9,6)--(7,6)--(7,5)--(5,5)--(5,3)--(3,3)--(3,2)--(0,2)--(0,0)--cycle;
\node[] at (.5, .5) {\color{red}$a_1$};
\node[] at (2.5, 1.5) {\color{red}$a_2$};
\node[] at (4.5, 2.5) {\color{red}$a_3$};
\node[] at (3.5, 2.5) {\color{blue}$b_1$};
\node[] at (5.5, 2.5) {\color{blue}$b_2$};
\node[] at (8.5, 5.5) {\color{blue}$b_3$};
\node[] at (4.5, -.5) {\phantom{\color{red}$a_3$}};
\end{tikzpicture}\qquad
\begin{tikzpicture}[scale=.46]
\fill[fill=pink, rounded corners] (2.2, .2) rectangle (2.8, 1.8)
(2.2, 1.2) rectangle (5.8, 1.8)
(5.2, 1.2) rectangle (5.8, 2.8);
\fill[fill=orange!30!white, rounded corners] (3.2, 2.2) rectangle (4.8, 2.8)
(4.2, 2.2) rectangle (4.8, 3.8)
(4.2, 3.2) rectangle (6.8, 3.8)
(6.2, 3.2) rectangle (6.8, 5.8)
(6.2, 5.2) rectangle (8.8, 5.8)
(8.2, 5.2) rectangle (8.8, 6.8)
(8.2, 6.2) rectangle (9.8, 6.8);
\fill[fill=cyan!20!white, rounded corners] (4.1, 4.2) rectangle (5.9, 4.8);
\draw[color=gray!30!white] (7,7)--(8,7)
(5,6)--(10,6)
(5,5)--(10,5)
(3,4)--(8,4)
(0,3)--(8,3)
(1,2)--(6,2)
(2,1)--(3,1)
(1,3)--(1,4)
(2,2)--(2,4)
(3,1)--(3,4)
(4,1)--(4,5)
(5,1)--(5,5)
(6,1)--(6,7)
(7,2)--(7,7)
(8,4)--(8,8)
(9,4)--(9,7)
(10,5)--(10,6);
\draw[ thick] (2,1)--(1,1)--(1,3)--(3,3)
(1,2)--(0,2)--(0,4)--(3,4)--(3,5)--(4,5)
(5,5)--(5,7)--(7,7)--(7,8)--(9,8)--(9,7)
(6,2)--(8,2)--(8,4)--(10,4)--(10,5)--(11,5)--(11,6)--(10,6)
(3,2)--(5,2)--(5,3)--(6,3)
(4,4)--(6,4)--(6,5);
\draw[line width=1.6pt, color=frenchblue!90!cyan] (2,0)--(3,0)--(3,1)--(6,1)--(6,3)--(7,3)--(7,5)--(9,5)--(9,6)--(10,6)--(10,7)--(8,7)--(8,6)--(6,6)--(6,5)--(4,5)--(4,3)--(3,3)--(3,2)--(2,2)--(2,0)--cycle;
\node[] at (2.5, .5) {\color{red}$a_1$};
\node[] at (3.5, 2.5) {\color{red}$a_2$};
\node[] at (4.5, 4.5) {\color{red}$a_3$};
\node[] at (5.5, 2.5) {\color{blue}$b_1$};
\node[] at (5.5, 4.5) {\color{blue}$b_2$};
\node[] at (9.5, 6.5) {\color{blue}$b_3$};
\end{tikzpicture}
\caption{A border strip $\gamma$ and the two sequences $(a_1,a_2,a_3)$ and
$(b_1,b_2,b_3)$ on the left and the corresponding skew shape and its
decomposition on the right.}
\label{fig:to_skew}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\fill[fill=pink, rounded corners] (.2, .2) rectangle (.8, 1.8)
(.2, 1.2) rectangle (3.8, 1.8)
(3.2, 1.2) rectangle (3.8, 2.8);
\draw[thick] (0,0)--(1,0)--(1,1)--(4,1)--(4,2)--(6,2)--(6,4)--(8,4)--(8,5)--(9,5)--(9,6)--(7,6)--(7,5)--(5,5)--(5,3)--(3,3)--(3,2)--(0,2)--(0,0)--cycle;
\draw[color=gray] (0,1)--(1,1)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)--(4,2)--(4,3)
(5,2)--(5,3)--(6,3)
(5,4)--(6,4)--(6,5)
(7,4)--(7,5)--(8,5)--(8,6);
\node[] at (.5, .5) {\color{red}$a_1$};
\node[] at (2.5, 1.5) {\color{red}$a_2$};
\node[] at (4.5, 2.5) {\color{red}$a_3$};
\node[] at (3.5, 2.5) {\color{blue}$b_1$};
\node[] at (5.5, 2.5) {\color{blue}$b_2$};
\node[] at (8.5, 5.5) {\color{blue}$b_3$};
\node[] at (4.5, -.5) {(a)};
\end{tikzpicture}~
\begin{tikzpicture}[scale=.46]
\fill[fill=cyan!20!white, rounded corners] (4.1, 2.2) rectangle (5.9, 2.8);
\draw[thick] (0,0)--(1,0)--(1,1)--(4,1)--(4,2)--(6,2)--(6,4)--(8,4)--(8,5)--(9,5)--(9,6)--(7,6)--(7,5)--(5,5)--(5,3)--(3,3)--(3,2)--(0,2)--(0,0)--cycle;
\draw[color=gray] (0,1)--(1,1)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)--(4,2)--(4,3)
(5,2)--(5,3)--(6,3)
(5,4)--(6,4)--(6,5)
(7,4)--(7,5)--(8,5)--(8,6);
\node[] at (2.5, 1.5) {\color{red}$a_2$};
\node[] at (4.5, 2.5) {\color{red}$a_3$};
\node[] at (5.5, 2.5) {\color{blue}$b_2$};
\node[] at (8.5, 5.5) {\color{blue}$b_3$};
\node[] at (4.5, -.5) {(b)};
\end{tikzpicture}~
\begin{tikzpicture}[scale=.46]
\fill[fill=orange!30!white, rounded corners] (2.2, 1.2) rectangle (3.8, 1.8)
(3.2, 1.2) rectangle (3.8, 2.8)
(3.2, 2.2) rectangle (5.8, 2.8)
(5.2, 2.2) rectangle (5.8, 4.8)
(5.2, 4.2) rectangle (7.8, 4.8)
(7.2, 4.2) rectangle (7.8, 5.8)
(7.2, 5.2) rectangle (8.8, 5.8);
\draw[thick] (0,0)--(1,0)--(1,1)--(4,1)--(4,2)--(6,2)--(6,4)--(8,4)--(8,5)--(9,5)--(9,6)--(7,6)--(7,5)--(5,5)--(5,3)--(3,3)--(3,2)--(0,2)--(0,0)--cycle;
\draw[color=gray] (0,1)--(1,1)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)--(4,2)--(4,3)
(5,2)--(5,3)--(6,3)
(5,4)--(6,4)--(6,5)
(7,4)--(7,5)--(8,5)--(8,6);
\node[] at (2.5, 1.5) {\color{red}$a_2$};
\node[] at (8.5, 5.5) {\color{blue}$b_3$};
\node[] at (4.5, -.5) {(c)};
\end{tikzpicture}
\caption{The decomposition in Figure~\ref{fig:to_skew} is obtained recursively
using these three border strips.}
\label{fig:to_skew1}
\end{figure}
We now prove Theorem~\ref{thm:HG_conv}.
\begin{proof}[Proof of Theorem~\ref{thm:HG_conv}]
We may assume $\det \left(s_{\alpha[a_j,b_i]}\right)_{i,j=1}^k \ne 0$ because
otherwise there is nothing to prove. Since permuting the $a_i$'s or $b_i$'s
only changes the sign of the determinant, we may assume $a_1<\dots<a_k$ and
$b_1<\dots<b_k$. We first claim that the condition \eqref{eq:dyck} holds for
all $a\in \ZZ$. For a contradiction, suppose that \eqref{eq:dyck} does not
hold for some $a\in \ZZ$. Then we have $a_1<\dots<a_r\le a < a_{r+1}<\dots<
a_k$ and $b_1<\dots<b_s\le a < b_{s+1}<\dots< b_k$ for some integers $0\le
r<s\le k$. Then $s_{\alpha[a_j,b_i]}=0$ for all $1\le i\le s$ and $r+1\le j\le
k$, which implies $\det \left(s_{\alpha[a_j,b_i]}\right)_{i,j=1}^k = 0$, which
is a contradiction to the assumption given in the beginning of the proof.
Therefore \eqref{eq:dyck} holds for all $a\in \ZZ$.
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\draw[color=gray] (0,1)--(3,1)
(3,2)--(4,2)
(5,3)--(8,3)
(5,4)--(9,4)
(7,5)--(9,5)
(1,0)--(1,2)
(2,0)--(2,2)
(3,1)--(3,2)
(4,2)--(4,3)
(5,2)--(5,3)
(6,2)--(6,5)
(7,2)--(7,5)
(8,3)--(8,6);
\draw[line width=1.2pt] (0,0)--(3,0)--(3,1)--(4,1)--(4,2)--(8,2)--(8,3)--(9,3)--(9,6)--(7,6)--(7,5)--(5,5)--(5,3)--(3,3)--(3,2)--(0,2)--(0,0)--cycle;
\node[] at (.5, .5) {\color{red}$a_1$};
\node[] at (2.5, 1.5) {\color{red}$a_2$};
\node[] at (4.5, 2.5) {\color{red}$a_3$};
\node[] at (3.5, 2.5) {\color{blue}$b_1$};
\node[] at (5.5, 2.5) {\color{blue}$b_2$};
\node[] at (8.5, 5.5) {\color{blue}$b_3$};
\end{tikzpicture}\qquad\qquad
\begin{tikzpicture}[scale=.46]
\fill[fill=red!10!white] (1,0) rectangle (3,1)
(6,2) rectangle (8,4)
(8,3) rectangle (9,5);
\draw[thick, color=dredcolor] (1,0)--(3,0)--(3,1)
(2,0)--(2,1)
(6,2)--(8,2)--(8,3)--(9,3)--(9,5)
(7,2)--(7,4)
(8,3)--(8,4)
(6,3)--(8,3)
(8,4)--(9,4);
\draw[color=gray] (0,1)--(1,1)--(1,2)
(2,1)--(2,2)
(3,1)--(3,2)--(4,2)--(4,3)
(5,2)--(5,3)--(6,3)
(5,4)--(6,4)--(6,5)
(7,4)--(7,5)--(8,5)--(8,6);
\draw[line width=1.3pt] (0,0)--(1,0)--(1,1)--(4,1)--(4,2)--(6,2)--(6,4)--(8,4)--(8,5)--(9,5)--(9,6)--(7,6)--(7,5)--(5,5)--(5,3)--(3,3)--(3,2)--(0,2)--(0,0)--cycle;
\node[] at (.5, .5) {\color{red}$a_1$};
\node[] at (2.5, 1.5) {\color{red}$a_2$};
\node[] at (4.5, 2.5) {\color{red}$a_3$};
\node[] at (3.5, 2.5) {\color{blue}$b_1$};
\node[] at (5.5, 2.5) {\color{blue}$b_2$};
\node[] at (8.5, 5.5) {\color{blue}$b_3$};
\end{tikzpicture}
\caption{A connected skew shape $\alpha$ and two sequences $a_1<a_2<a_3$ and
$b_1<b_2<b_3$ on the left. The inner strip $\gamma$ of $\alpha$ is shown on
the right with thick lines, where the cells in $\alpha\setminus\gamma$ are
colored red. }
\label{fig:inner_gamma}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}[scale=.46]
\fill[fill=pink, rounded corners] (.2, .2) rectangle (.8, 1.8)
(.2, 1.2) rectangle (3.8, 1.8)
(3.2, 1.2) rectangle (3.8, 2.8);
\fill[fill=orange!30!white, rounded corners] (1.2, 2.2) rectangle (2.8, 2.8)
(2.2, 2.2) rectangle (2.8, 3.8)
(2.2, 3.2) rectangle (4.8, 3.8)
(4.2, 3.2) rectangle (4.8, 5.8)
(4.2, 5.2) rectangle (6.8, 5.8)
(6.2, 5.2) rectangle (6.8, 6.8)
(6.2, 6.2) rectangle (7.8, 6.8);
\fill[fill=cyan!20!white, rounded corners] (2.1, 4.2) rectangle (3.9, 4.8);
\fill[fill=red!10!white] (1,0) rectangle (3,1)
(5,3) rectangle (7,5)
(7,4) rectangle (8,6);
\draw[thick, color=dredcolor] (1,0)--(3,0)--(3,1)
(2,0)--(2,1)
(5,3)--(7,3)--(7,4)--(8,4)--(8,6)
(6,3)--(6,5)
(7,4)--(7,5)
(5,4)--(7,4)
(7,5)--(8,5);
\draw[color=gray] (0,1)--(1,1)
(1,2)--(4,2)
(2,3)--(4,3)
(2,4)--(5,4)
(4,5)--(5,5)
(6,6)--(7,6)
(1,1)--(1,2)
(2,1)--(2,3)
(3,1)--(3,5)
(4,3)--(4,5)
(5,5)--(5,6)
(6,5)--(6,6)
(7,6)--(7,7);
\draw[thick] (1,2)--(3,2)--(3,3)--(4,3)
(2,4)--(4,4)--(4,5);
\draw[line width=1.6pt, color=frenchblue!90!cyan] (0,0)--(1,0)--(1,1)--(4,1)--(4,3)--(5,3)--(5,5)--(7,5)--(7,6)--(8,6)--(8,7)--(6,7)--(6,6)--(4,6)--(4,5)--(2,5)--(2,3)--(1,3)--(1,2)--(0,2)--(0,0)--cycle;
\node[] at (.5, .5) {\color{red}$p_1$};
\node[] at (1.5, 2.5) {\color{red}$p_2$};
\node[] at (2.5, 4.5) {\color{red}$p_3$};
\node[] at (3.5, 2.5) {\color{blue}$q_1$};
\node[] at (3.5, 4.5) {\color{blue}$q_2$};
\node[] at (7.5, 6.5) {\color{blue}$q_3$};
\end{tikzpicture}
\caption{The boundary of the skew shape $\lm$ is colored blue and the cells in
$\nl$ are colored red. Here, using the $\alpha$, $\gamma$, $a_i$'s and
$b_i$'s in Figure~\ref{fig:inner_gamma}, $\lm$ is the skew shape whose
decomposition with respect to the cutting strip $\gamma$ is
$\theta=(\theta_1,\theta_2,\theta_3)$ such that
$\{p(\theta_1),p(\theta_2),p(\theta_3)\}=\{a_1,a_2,a_3\}$ and
$\{q(\theta_1),q(\theta_2),q(\theta_3)\}=\{b_1,b_2,b_3\}$, and the connected
components of $\nl$ are obtained from those of $\alpha/\gamma$ by
appropriate diagonal shifts.}
\label{fig:nl}
\end{figure}
Now let $\gamma$ be the inner strip of $\alpha$ as shown in
Figure~\ref{fig:inner_gamma}. Then by Lemma~\ref{lem:gamma_to_la} there is a
skew shape $\lm$ whose decomposition with respect to the cutting strip $\gamma$
is $\theta=(\theta_1,\dots,\theta_k)$ such that
$\{p(\theta_1),\dots,p(\theta_k)\}=\{a_1,\dots,a_k\}$ and
$\{q(\theta_1),\dots,q(\theta_k)\}=\{b_1,\dots,b_k\}$. Let $\nu$ be the
partition obtained from $\lambda$ by attaching each connected component of
$\alpha\setminus\gamma$ below $\lambda$ after an appropriate diagonal shift, see
Figure~\ref{fig:nl}. The assumption that $\alpha[a_j,b_i]$ is a skew shape, for
all $i,j$, implies that $\mu$ and $\gamma$ are $\nl$-compatible. Therefore, by
Theorem~\ref{thm:main_HG} with $\ts_{\tau}$ specialized to $s_{\tau}$ for all
skew shapes $\tau$, we have
\begin{equation}\label{eq:det = s}
\det \left( s_{(\gamma\oplus\nl)[a_j,b_i]}\right)_{i,j=1}^k = \pm s_{\nu/\mu}
\prod_{s=1}^\ell s_{\alpha_s}^{r_s-1},
\end{equation}
where $\alpha_1,\dots,\alpha_\ell$ are the connected components of $\nl$, and
$r_s$ is the number of strips $\theta_i$ such that
$\Cont(\alpha_s)\subseteq\Cont(\theta_i)$.
By the construction, we have $(\gamma\oplus\nl)[a_j,b_i]=\alpha[a_j,b_i]$. Since
a product of skew Schur functions can be expressed as a single skew Schur
function, the theorem follows from \eqref{eq:det = s}.
\end{proof}
\section*{Acknowledgments}
The authors would like to thank Byung-Hak Hwang for helpful discussions
and Alejandro Morales for useful comments.
| 8d0378c21820281d14c00f57dea0c8d3a630cf81 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The use of neural networks (NNs) as universal approximators has recently attracted great attention in various problems in physical sciences.
Despite the recent enthusiasm, there are limited studies of interpretable NN methods, which have led to the characterization of neural networks as ``black-boxes" which do not provide feedback from a modeling strategy.
A decisive reason for this is the specification of modeling tasks from the deterministic regression point-of-view: $L_2$ norms are well-understood and widely utilized in linear least-squares applications. However, this limits the amount of information one can extract from a modeling effort. In contrast, Bayesian models, for e.g., variational inference \cite{blundell2015} and Gaussian process approximations \cite{damianou2019gp} offer significant flexibility, and may be utilized to assess model and data quality as well as quantify uncertainty under one umbrella.
In the present paper, we introduce the use of probabilistic neural network (PNN) that assumes a generative distribution for the target data \cite{bishop1994mixture}. The PNN is able to estimate the expected value of the targets while also predicting a confidence interval for each estimation. For our presentation, we consider the construction of reduced-order models for fluid dynamics.
\section{Methods}
\subsection{Probabilistic neural network (PNN)}
Neural networks $\cal F$, which output ${\bm y}$ from inputs $\bm x$, are generally trained to obtain optimized weights ${\bm w}$ by minimizing a cost function $E$ such that ${\bm w}={\rm argmin}_{\bm w}[E({\bm y},{\cal F}({\bm x};{\bm w}))]$.
The output obtained through this process has no information about the estimation uncertainty. To obtain the probability distribution of estimations $p({\bm y}_p|{\bm x})$, we consider the use of a PNN.
Instead of the general map for an NN, $\mathcal{F}: {\bm x} \rightarrow {\bm y}({\bm x})$, the mapping of a PNN can be written as $\mathcal{F}: {\bm x} \rightarrow (\pi_1, \mu_1, \sigma_1, \pi_2, \mu_2, \sigma_2, ..., \pi_N, \mu_N, \sigma_N)$, where $\pi_i$ is the mixing probability for each Gaussian component satisfying the condition $\sum_{i=1}^{m}\pi_i = 1$, the mean $\mu$ and standard deviation $\sigma$ parametrize a Gaussian probability distribution function $\mathcal{N}(\mu,\sigma)$.
Therefore, the distribution function in this model is a linear combination of several Gaussian components,
\begin{equation}\label{eq:GMM}
p({\bm y} | {\bm x}) = \sum_{i=1}^{m}\pi_i ({\bm x}) \mathcal{N}(\mu_i ({\bm x}),\sigma_i ({\bm x})).
\end{equation}
The value of $m$ is generally pre-specified based on the expectation of posterior distribution.
Utilizing the mixture of Gaussians in the manner above, we can expect to be able to handle a complex probability distribution for the output of PNN $p({\bm y}_p|{\bm x})$. Because our PNN attempts to output a distribution of the estimation $p({\bm y}_p|{\bm x})$ instead of a target variable of ${\bm y}$ directly, care should be taken to choose the loss function so that the full distribution of the estimation can be utilized. In the present study, the cost function $E$ for the PNN ${\mathcal F}({\bm x};{\bm w})$ is given in terms of average log-likelihood $\mathcal{L}$ such that
\begin{align}\label{eq:negloglike}
\begin{gathered}
{\bm w}={\rm argmin}_{\bm w}[\mathcal{E}],\quad \text{where} \quad \mathcal{E} \equiv -\log \mathcal{L} = -\sum_{k=1}^{K} p({y}_{k,p}|{x_k}) \log p({y}_{k,t})
\end{gathered}
\end{align}
with $k$ indicates each data point in the training data and $K$ denotes a number of training samples.
The term $p({y}_{p,k}|{x_k})$ in the error function is evaluated for each data point using the output of the network given in equation \ref{eq:GMM}. Note that the likelihood maximizing model such as the present model is equivalent to minimizing the cross entropy $H(p({\bm y}_p|{\bm x}), p({\bm y}_t) )$, which encourages us to apply the model to a wide variety of problems.
\subsection{PNN-based reduced-order model with proper orthogonal decomposition}
In the present study, we demonstrate the capability of the PNN for predicting the temporal evolution of coefficients for proper orthogonal decomposition (POD) from the state at the first instantaneous field, as shown in figure \ref{fig1}.
Using the POD basis $\bm \varphi$ and the POD coefficients, the state $\bm q$ can be decomposed as ${\bm q}={\bm q}_0+\sum_{i=1}^{M}{a_i}{\bm \varphi}_i$, where ${\bm q}_0$ is the temporal average of the state and $M$ denotes the number of POD modes.
The present PNN attempts to predict the POD coefficients over $n$ time steps ${\bm a}_t=[{\bm a}^1,{\bm a}^2,...,{\bm a}^{n}]$, where ${\bm a}^{\iota}=[a^{\iota}_1,a^{\iota}_2,...,a^{\iota}_{M}]$, from the state (e.g., sensor information) at the first instantaneous field ${\bm s}^1$,
\begin{align}
\begin{gathered}
\{\pi({\bm s}^1),\mu({\bm s}^1),\sigma({\bm s}^1)\}={\cal F}({\bm s}^1),~~~p({\bm a}_t | {\bm s}^1) = \sum_{i=1}^{m}\pi_i ({\bm s}^1) \mathcal{N}(\mu_i ({\bm s}^1),\sigma_i ({\bm s}^1)).
\end{gathered}
\end{align}
The present problem setting can be regarded as a reduced-order surrogate for high-dimensional systems since the original dynamics can be approximated in both space and time by combining the estimated POD coefficients and the POD basis.
In this study, we will assess the proposed modeling framework for the wake estimation of a NACA0012 airfoil with a Gurney flap and the flow reconstruction of the sea surface temperature (from the NOAA Optimum Interpolation dataset).
\section{Results}
\subsection{Comparison with benchmark linear method}
Let us first consider the estimation of POD coefficients ${\bm a}$ from local sensor measurements ${\bm s}$ at a given instant in time, by using the two-dimensional inviscid shallow water equations which are a prototypical system for geophysical flows.
The problem setting can be mathematically expressed as
\begin{align}
\begin{gathered}
\{\pi({[{\bm s},{q_t}]}),\mu([{\bm s},{q_t}]),\sigma([{\bm s},{q_t}])\}={\cal F}([{\bm s},{q_t}]),~~~p({\bm a} | [{\bm s},{q_t}]) = \sum_{i=1}^{m}\pi_i ([{\bm s},{q_t}]) \mathcal{N}(\mu_i ([{\bm s},{q_t}]),\sigma_i ([{\bm s},{q_t}])).
\end{gathered}
\end{align}
As shown, the time stamp, $q_t$, which indicates the progress to the final time of the evolution is also utilized as an input attribute in addition to the sensor measurements.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{./figs/fig0.jpg}
\caption{A comparison between PNN and Gappy POD for POD coefficient reconstructions. The values underneath the contours indicate the normalized $L_1$ error norms. Sensor locations (shown in red) are not provided for the validation with 3800 sensors for visual clarity.}
\label{fig0}
\end{figure}
We here compare the PNN and a well-known linear reconstruction method --- Gappy POD \cite{everson1995karhunen} for flow reconstruction, as presented in figure \ref{fig0}.
The present model outperforms Gappy POD at the same number of sensors $n_{\rm sensor}=30$. Even at $n_{\rm sensor}=3800$, the PNN exhibits a significant advantage, as observed through an $L_1$ norm comparison.
Note that the present model can provide confidence intervals for coefficient estimation as well (demonstrated in the following sections), in comparison to Gappy POD which is determinstic.
\subsection{NACA0012 airfoil with a Gurney flap}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{./figs/fig2.png}
\caption{Application of probabilistic neural network to the wake of NACA0012 airfoil with a Gurney flap. $(a)$ Estimation of POD coefficients over a half of vortex shedding period. $(b)$ Final instantaneous vorticity snapshots reconstructed by the estimated coefficients. The value inside the contour indicates the $L_2$ error norm of reconstructed field with the estimated coefficients.}
\label{fig2}
\end{figure}
Next, let us consider the wake behind a NACA0012 airfoil with a Gurney flap.
The training data set is prepared using two-dimensional direct numerical simulation at a chord based Reynolds number of ${\rm Re}_c =1000$ \cite{GMTA2018}.
We use 100 snapshots collected over approximately 4 periods for the training of PNN.
For the example of NACA0012 wake, the PNN tries to predict the temporal evolution of fifty POD coefficients for the vorticity field ${\bm a}^{\iota}=[a^{\iota}_1,a^{\iota}_2, a^{\iota}_{3},...,a^{\iota}_{50}]$ over 12 time steps ${\bm a}_t=[{\bm a}^1,{\bm a}^2,...,{\bm a}^{100}]$ corresponding to approximately a half of vortex shedding period, from the five vorticity sensor measurements on surface of an airfoil at the first instantaneous field ${\bm s}^1$.
Hence, the dimensions of input and output are 5 and 600, respectively.
Our investigation for the NACA0012 with a Gurney flap is summarized in Figure \ref{fig2}.
The PNN can accurately estimate the temporal evolution for the POD coefficients while showing its confidence interval as shown in figure \ref{fig2}$(a)$, which corresponds to the normalized $L_2$ error norm for the POD coefficients $\epsilon_c=||{\bm a}_{\rm Ref}-{\bm a}_{\rm PNN}||_2/||{\bm a}_{\rm Ref}||_2$ of 0.145.
By combining to the POD basis, the high-dimensional flow can also represented well with the normalized $L_2$ error norm for the reconstructed fields $\epsilon_f=||{\omega}_{\rm Ref}-{\omega}_{\rm PNN}||_2/||{\omega}_{\rm Ref}||_2$ of 0.0267, as presented in figure \ref{fig2}$(b)$.
\subsection{NOAA sea surface temperature}
We also consider the NOAA sea surface temperature data set for field reconstruction from limited sensor measurements.
The spatial resolution of data set is $360\times180$ based on a one degree grid and is obtained from satellite and ship observations.
Training data set is prepared from 20 years of data (1040 snapshots spanning 1981 to 2001), while the test data set is obtained from 874 snapshots from years 2001 to 2018.
The problem setting here is inspired by Callaham et al. \cite{CMB2019} who capitalized on sparse representations to reconstruct fluid flow fields from local sensors.
For the task with the sea surface temperature, the PNN predicts the temporal evolution of 4 POD coefficients ${\bm a}^{\iota}=[a^{\iota}_1,a^{\iota}_2, a^{\iota}_{3},a^{\iota}_{4}]$ over 100 weeks ${\bm a}_t=[{\bm a}^1,{\bm a}^2,...,{\bm a}^{100}]$ from the local 30 sensor measurements on the first week snapshot ${\bm s}^1$.
Following Callaham et al. \cite{CMB2019}, the input sensors are chosen randomly from the region between $50{^\circ}$ S and $50{^\circ}$ N.
\begin{figure}
\centering
\includegraphics[width=0.70\textwidth]{./figs/fig3.png}
\caption{Application of probabilistic neural network to the NOAA sea surface temperature. $(a)$ Estimation of POD coefficients over 100 weeks. $(b)$ Final instantaneous snapshots reconstructed by the estimated coefficients. The reference contour includes input sensor locations. The value underneath the contour indicates the $L_2$ error norm of reconstructed field with the estimated coefficients.}
\label{fig3}
\end{figure}
The results for NOAA sea surface temperature are presented in figure \ref{fig3}.
As shown, the temporal trend of POD coefficients can be captured well using PNN.
What is striking here is that the estimated uncertainty is larger than that of NACA0012 example, which is likely caused by the noisy nature of real-world data. This drives home an important point about the quality of training data for neural network optimization --- probabilistic methods offer insight into the nature of an experimentally generated data set as well.
Since estimation of POD coefficients works well, the reconstruction for the temperature field is also in good agreement with the reference data.
\section{Concluding remarks}
This article introduced the use of probabilistic neural networks (PNN) to quantify uncertainties within a surrogate model for fluid flow reconstruction and forecasting applications.
The advantage against the conventional linear method was first demonstrated with a shallow water equation.
We then assessed the framework on the wake estimation of a NACA0012 airfoil with a Gurney flap and the NOAA sea surface temperature data set.
The present PNN showed promising potential as a reduced-order surrogate with uncertainty estimates of the targets conditioned on the inputs and the training data.
Although not shown here, the confidence intervals obtained by the present modeling strategy may also be utilized additional sensor placements in fluid flow data recovery tasks \cite{MFRFT2020}.
These observations enable us to expect various extensions of PNN to a wide spread of engineering applications.
\small
\newpage
\bibliographystyle{unsrt}
\section{Introduction}
The use of neural networks (NNs) as universal approximators has recently attracted great attention in various problems in physical sciences.
Despite the recent enthusiasm, there are limited studies of interpretable NN methods, which have led to the characterization of neural networks as ``black-boxes" which do not provide feedback from a modeling strategy.
A decisive reason for this is the specification of modeling tasks from the deterministic regression point-of-view: $L_2$ norms are well-understood and widely utilized in linear least-squares applications. However, this limits the amount of information one can extract from a modeling effort. In contrast, Bayesian models, for e.g., variational inference \cite{blundell2015} and Gaussian process approximations \cite{damianou2019gp} offer significant flexibility, and may be utilized to assess model and data quality as well as quantify uncertainty under one umbrella.
In the present paper, we introduce the use of probabilistic neural network (PNN) that assumes a generative distribution for the target data \cite{bishop1994mixture}. The PNN is able to estimate the expected value of the targets while also predicting a confidence interval for each estimation. For our presentation, we consider the construction of reduced-order models for fluid dynamics.
\section{Methods}
\subsection{Probabilistic neural network (PNN)}
Neural networks $\cal F$, which output ${\bm y}$ from inputs $\bm x$, are generally trained to obtain optimized weights ${\bm w}$ by minimizing a cost function $E$ such that ${\bm w}={\rm argmin}_{\bm w}[E({\bm y},{\cal F}({\bm x};{\bm w}))]$.
The output obtained through this process has no information about the estimation uncertainty. To obtain the probability distribution of estimations $p({\bm y}_p|{\bm x})$, we consider the use of a PNN.
Instead of the general map for an NN, $\mathcal{F}: {\bm x} \rightarrow {\bm y}({\bm x})$, the mapping of a PNN can be written as $\mathcal{F}: {\bm x} \rightarrow (\pi_1, \mu_1, \sigma_1, \pi_2, \mu_2, \sigma_2, ..., \pi_N, \mu_N, \sigma_N)$, where $\pi_i$ is the mixing probability for each Gaussian component satisfying the condition $\sum_{i=1}^{m}\pi_i = 1$, the mean $\mu$ and standard deviation $\sigma$ parametrize a Gaussian probability distribution function $\mathcal{N}(\mu,\sigma)$.
Therefore, the distribution function in this model is a linear combination of several Gaussian components,
\begin{equation}\label{eq:GMM}
p({\bm y} | {\bm x}) = \sum_{i=1}^{m}\pi_i ({\bm x}) \mathcal{N}(\mu_i ({\bm x}),\sigma_i ({\bm x})).
\end{equation}
The value of $m$ is generally pre-specified based on the expectation of posterior distribution.
Utilizing the mixture of Gaussians in the manner above, we can expect to be able to handle a complex probability distribution for the output of PNN $p({\bm y}_p|{\bm x})$. Because our PNN attempts to output a distribution of the estimation $p({\bm y}_p|{\bm x})$ instead of a target variable of ${\bm y}$ directly, care should be taken to choose the loss function so that the full distribution of the estimation can be utilized. In the present study, the cost function $E$ for the PNN ${\mathcal F}({\bm x};{\bm w})$ is given in terms of average log-likelihood $\mathcal{L}$ such that
\begin{align}\label{eq:negloglike}
\begin{gathered}
{\bm w}={\rm argmin}_{\bm w}[\mathcal{E}],\quad \text{where} \quad \mathcal{E} \equiv -\log \mathcal{L} = -\sum_{k=1}^{K} p({y}_{k,p}|{x_k}) \log p({y}_{k,t})
\end{gathered}
\end{align}
with $k$ indicates each data point in the training data and $K$ denotes a number of training samples.
The term $p({y}_{p,k}|{x_k})$ in the error function is evaluated for each data point using the output of the network given in equation \ref{eq:GMM}. Note that the likelihood maximizing model such as the present model is equivalent to minimizing the cross entropy $H(p({\bm y}_p|{\bm x}), p({\bm y}_t) )$, which encourages us to apply the model to a wide variety of problems.
\subsection{PNN-based reduced-order model with proper orthogonal decomposition}
In the present study, we demonstrate the capability of the PNN for predicting the temporal evolution of coefficients for proper orthogonal decomposition (POD) from the state at the first instantaneous field, as shown in figure \ref{fig1}.
Using the POD basis $\bm \varphi$ and the POD coefficients, the state $\bm q$ can be decomposed as ${\bm q}={\bm q}_0+\sum_{i=1}^{M}{a_i}{\bm \varphi}_i$, where ${\bm q}_0$ is the temporal average of the state and $M$ denotes the number of POD modes.
The present PNN attempts to predict the POD coefficients over $n$ time steps ${\bm a}_t=[{\bm a}^1,{\bm a}^2,...,{\bm a}^{n}]$, where ${\bm a}^{\iota}=[a^{\iota}_1,a^{\iota}_2,...,a^{\iota}_{M}]$, from the state (e.g., sensor information) at the first instantaneous field ${\bm s}^1$,
\begin{align}
\begin{gathered}
\{\pi({\bm s}^1),\mu({\bm s}^1),\sigma({\bm s}^1)\}={\cal F}({\bm s}^1),~~~p({\bm a}_t | {\bm s}^1) = \sum_{i=1}^{m}\pi_i ({\bm s}^1) \mathcal{N}(\mu_i ({\bm s}^1),\sigma_i ({\bm s}^1)).
\end{gathered}
\end{align}
The present problem setting can be regarded as a reduced-order surrogate for high-dimensional systems since the original dynamics can be approximated in both space and time by combining the estimated POD coefficients and the POD basis.
In this study, we will assess the proposed modeling framework for the wake estimation of a NACA0012 airfoil with a Gurney flap and the flow reconstruction of the sea surface temperature (from the NOAA Optimum Interpolation dataset).
\section{Results}
\subsection{Comparison with benchmark linear method}
Let us first consider the estimation of POD coefficients ${\bm a}$ from local sensor measurements ${\bm s}$ at a given instant in time, by using the two-dimensional inviscid shallow water equations which are a prototypical system for geophysical flows.
The problem setting can be mathematically expressed as
\begin{align}
\begin{gathered}
\{\pi({[{\bm s},{q_t}]}),\mu([{\bm s},{q_t}]),\sigma([{\bm s},{q_t}])\}={\cal F}([{\bm s},{q_t}]),~~~p({\bm a} | [{\bm s},{q_t}]) = \sum_{i=1}^{m}\pi_i ([{\bm s},{q_t}]) \mathcal{N}(\mu_i ([{\bm s},{q_t}]),\sigma_i ([{\bm s},{q_t}])).
\end{gathered}
\end{align}
As shown, the time stamp, $q_t$, which indicates the progress to the final time of the evolution is also utilized as an input attribute in addition to the sensor measurements.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{./figs/fig0.jpg}
\caption{A comparison between PNN and Gappy POD for POD coefficient reconstructions. The values underneath the contours indicate the normalized $L_1$ error norms. Sensor locations (shown in red) are not provided for the validation with 3800 sensors for visual clarity.}
\label{fig0}
\end{figure}
We here compare the PNN and a well-known linear reconstruction method --- Gappy POD \cite{everson1995karhunen} for flow reconstruction, as presented in figure \ref{fig0}.
The present model outperforms Gappy POD at the same number of sensors $n_{\rm sensor}=30$. Even at $n_{\rm sensor}=3800$, the PNN exhibits a significant advantage, as observed through an $L_1$ norm comparison.
Note that the present model can provide confidence intervals for coefficient estimation as well (demonstrated in the following sections), in comparison to Gappy POD which is determinstic.
\subsection{NACA0012 airfoil with a Gurney flap}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{./figs/fig2.png}
\caption{Application of probabilistic neural network to the wake of NACA0012 airfoil with a Gurney flap. $(a)$ Estimation of POD coefficients over a half of vortex shedding period. $(b)$ Final instantaneous vorticity snapshots reconstructed by the estimated coefficients. The value inside the contour indicates the $L_2$ error norm of reconstructed field with the estimated coefficients.}
\label{fig2}
\end{figure}
Next, let us consider the wake behind a NACA0012 airfoil with a Gurney flap.
The training data set is prepared using two-dimensional direct numerical simulation at a chord based Reynolds number of ${\rm Re}_c =1000$ \cite{GMTA2018}.
We use 100 snapshots collected over approximately 4 periods for the training of PNN.
For the example of NACA0012 wake, the PNN tries to predict the temporal evolution of fifty POD coefficients for the vorticity field ${\bm a}^{\iota}=[a^{\iota}_1,a^{\iota}_2, a^{\iota}_{3},...,a^{\iota}_{50}]$ over 12 time steps ${\bm a}_t=[{\bm a}^1,{\bm a}^2,...,{\bm a}^{100}]$ corresponding to approximately a half of vortex shedding period, from the five vorticity sensor measurements on surface of an airfoil at the first instantaneous field ${\bm s}^1$.
Hence, the dimensions of input and output are 5 and 600, respectively.
Our investigation for the NACA0012 with a Gurney flap is summarized in Figure \ref{fig2}.
The PNN can accurately estimate the temporal evolution for the POD coefficients while showing its confidence interval as shown in figure \ref{fig2}$(a)$, which corresponds to the normalized $L_2$ error norm for the POD coefficients $\epsilon_c=||{\bm a}_{\rm Ref}-{\bm a}_{\rm PNN}||_2/||{\bm a}_{\rm Ref}||_2$ of 0.145.
By combining to the POD basis, the high-dimensional flow can also represented well with the normalized $L_2$ error norm for the reconstructed fields $\epsilon_f=||{\omega}_{\rm Ref}-{\omega}_{\rm PNN}||_2/||{\omega}_{\rm Ref}||_2$ of 0.0267, as presented in figure \ref{fig2}$(b)$.
\subsection{NOAA sea surface temperature}
We also consider the NOAA sea surface temperature data set for field reconstruction from limited sensor measurements.
The spatial resolution of data set is $360\times180$ based on a one degree grid and is obtained from satellite and ship observations.
Training data set is prepared from 20 years of data (1040 snapshots spanning 1981 to 2001), while the test data set is obtained from 874 snapshots from years 2001 to 2018.
The problem setting here is inspired by Callaham et al. \cite{CMB2019} who capitalized on sparse representations to reconstruct fluid flow fields from local sensors.
For the task with the sea surface temperature, the PNN predicts the temporal evolution of 4 POD coefficients ${\bm a}^{\iota}=[a^{\iota}_1,a^{\iota}_2, a^{\iota}_{3},a^{\iota}_{4}]$ over 100 weeks ${\bm a}_t=[{\bm a}^1,{\bm a}^2,...,{\bm a}^{100}]$ from the local 30 sensor measurements on the first week snapshot ${\bm s}^1$.
Following Callaham et al. \cite{CMB2019}, the input sensors are chosen randomly from the region between $50{^\circ}$ S and $50{^\circ}$ N.
\begin{figure}
\centering
\includegraphics[width=0.70\textwidth]{./figs/fig3.png}
\caption{Application of probabilistic neural network to the NOAA sea surface temperature. $(a)$ Estimation of POD coefficients over 100 weeks. $(b)$ Final instantaneous snapshots reconstructed by the estimated coefficients. The reference contour includes input sensor locations. The value underneath the contour indicates the $L_2$ error norm of reconstructed field with the estimated coefficients.}
\label{fig3}
\end{figure}
The results for NOAA sea surface temperature are presented in figure \ref{fig3}.
As shown, the temporal trend of POD coefficients can be captured well using PNN.
What is striking here is that the estimated uncertainty is larger than that of NACA0012 example, which is likely caused by the noisy nature of real-world data. This drives home an important point about the quality of training data for neural network optimization --- probabilistic methods offer insight into the nature of an experimentally generated data set as well.
Since estimation of POD coefficients works well, the reconstruction for the temperature field is also in good agreement with the reference data.
\section{Concluding remarks}
This article introduced the use of probabilistic neural networks (PNN) to quantify uncertainties within a surrogate model for fluid flow reconstruction and forecasting applications.
The advantage against the conventional linear method was first demonstrated with a shallow water equation.
We then assessed the framework on the wake estimation of a NACA0012 airfoil with a Gurney flap and the NOAA sea surface temperature data set.
The present PNN showed promising potential as a reduced-order surrogate with uncertainty estimates of the targets conditioned on the inputs and the training data.
Although not shown here, the confidence intervals obtained by the present modeling strategy may also be utilized additional sensor placements in fluid flow data recovery tasks \cite{MFRFT2020}.
These observations enable us to expect various extensions of PNN to a wide spread of engineering applications.
\small
\newpage
\bibliographystyle{unsrt}
| f6c8189b10530e4943cb390179c19611b4479430 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Dark Stars}
The Nobel prize this year has been awarded for research that
demonstrated that black holes can form in realistic astrophysical
situations, and observational evidence for a supermassive compact
object at the centre of our Galaxy.
Black holes are exotic objects.
These have been in popular imagination and these have also been used
in movies.
Black holes are thought of as objects from where nothing can escape.
The first speculation about such objects, then named as {\sl dark
stars} was the result of combining an increasing confidence in Newtonian
gravity and the speed of light as determined by Ole
R\"omer\mfnote{R\"omer used the eclipses of Jupiter's satellites to
measure the speed of light. He found that the estimated time
between successive eclipses is shorter when the Earth is moving
towards Jupiter and longer when the Earth is moving away from
Jupiter. He correctly interpreted this as a result of the finite
speed of light.}.
The speed of light is many orders of magnitude larger than the speeds
one encounters in day to day life.
In 18th century, John Michel and Pierre-Simon de Laplace speculated
about existence of stars so massive that light emitted cannot escape
to infinity.
A simple calculation shows that the mass and the radius of such stars
satisfies the relation based on the standard formula for escape velocity:
\begin{equation}
R \leq \frac{2 G M}{c^2}
\end{equation}
Here, $G$ is the universal constant of gravitation, $M$ is the mass of
the star, $c$ is the speed of light and $R$ is the radius\mfnote{For
masses that are familiar to us, the radius is very small. The
radius corresponding to the Sun is just under $3$~km, whereas the
radius of the Sun is close to $7\times 10^5$~km.}.
It is to be noted though that in this scenario, any observer at a
finite distance from the star will still see the light from the star.
Michel even speculated that {\sl the only way to discover the presence
of such a dark star may be to study motion of luminous objects in
its vicinity. }
Laplace conjectured that a significant fraction of objects in the
Universe may be dark stars.
\section{General Relativity and Black Holes}
Albert Einstein introduced the special theory of relativity in 1905
and this elevated the speed of light from a very high speed to the
maximum speed possible in nature.
General theory of relativity, introduced by Einstein in 1915,
connected the curvature of space-time with the notion of gravity while
retaining aspects of the special theory.
Karl Schwarzschild\mfnote{Schwarzschild presented this solution in
1916. Unfortunately he died shortly after this.} presented the
first exact solution for the equations of the general theory of
relativity.
He presented the solution for gravitational field of a point mass.
Two aspects of this solution are noteworthy: the gravitational field
is singular with a singularity at the position of the point mass, and,
nothing can escape from a sphere of radius given by Eqn.(1).
This sphere acts as a one way membrane and is referred to as the
horizon\mfnote{The understanding that this is a one way membrane came
much later with contributions from many, including Roger Penrose.
At the time, it was believed that there is a singularity at the
surface of this sphere but it was eventually understood that this
problem is due to the choice of coordinates in the Schwarzschild
metric.}.
The radius of this sphere is referred to as the Schwarzschild radius.
This was the first relativistic expression of what we now call a black
hole.
It was found that at large distances from the black hole, the
gravitational field and orbits deduced from Schwarzschild metric are
well approximated by Newtonian gravity with small corrections.
These corrections lead to precession of bound orbits: precession of
Mercury's orbit was one of the first verifications of the general
theory of relativity.
Close to the black hole the differences between the two are very
significant.
There are no stable bound orbits possible with an approach radius
smaller than three times the Schwarzschild radius.
Thus any massive particle approaching the black hole closer than this
distance is expected to fall into the black hole.
Photons can orbit around such a black hole with an orbital radius
equal to $1.5$ times the Schwarzschild radius.
Photons coming in from larger radii and approaching closer than this
distance fall into the black hole.
At large distances, photons get deflected from straight line by a
small amount.
This was observationally verified for the first time by Dyson,
Eddington and Davidson (1920) during the total solar eclipse on May
29, 1919.
This observational verification was critical in making general theory
of relativity as the accepted theory of gravitation and international
fame for Einstein.
Schwarzschild also presented a solution for space-time due to a star
and he showed that this is not singular and there is no horizon in
such a case.
He assumed the star to have a finite radius and uniform density.
In light of this, the point mass solution remained a curiosity for
some time.
Chandrasekhar limit for white dwarf stars\mfnote{S. Chandrasekhar
derived the limit by combining special relativity with quantum
statistics. He showed that a consequence of this combination is
that as a more massive white dwarf star has electrons that are
relativistic and the effective equation of state leads to an
instability if a white dwarf star made predominantly of Helium has a
mass more than $1.4$~M$_\odot$, where M$_\odot = 2 \times 10^{30}$~kg
is the mass of the Sun.} raised the question of what
happens if such a star goes beyond the mass limit.
In a star like the Sun, gravitational pull is finely balanced by a
combination of gas pressure and radiation pressure.
Thus one requires a time dependent study in order to address the
question raised by Chandrasekhar's computation of the mass limit for
white dwarf stars.
\section{Collapse and Singularities}
Bishveshwar Datt from Presidency College, Kolkata solved for time
dependent evolution of spherically symmetric density distributions.
His interest was in cosmological expansion of inhomogeneous regions.
He published his solutions for the cosmological scenario and an
expanding universe in 1938.
He passed away in the same year during a surgery.
His paper from 1938 was republished in 1997 as a golden oldie in the
journal {\sl General Relativity and Gravitation}.
The same solution was discovered independently by Oppenheimer and
Snyder in 1939 and applied to a collapsing star.
They showed that in absence of pressure the star collapses and
continues to collapse.
However, the collapse slows down from the perspective of a distant
observer as the radius of the star approaches the Schwarzschild
radius.
A very important link in the development of ideas about time dependent
space-times was provided by Professor Amal Kumar Raychaudhuri.
Working in Ashutosh College in Kolkata, he developed a general
equation for describing local evolution of such space-times without
imposing any restrictions of symmetry or constraining the type of
matter that drives the evolution of space-time.
The equation describes how matter moves in such a space time and how
the space time evolves with it.
The equation, named the Raychaudhuri equation in his honour describes
space-time in terms of an overall expansion or contraction, rotation,
and shear\cite{2008Res13319K}.
However, the assumption of zero pressure raised doubts about the
relevance of the solution for real stars.
We have discussed above that the Schwarzschild solution has a
singularity at the centre and it has a horizon enclosing the
singularity.
The development of this understanding came through the work of Roger
Penrose.
If we consider the view of a distant observer, gravitational
time dilation implies that it takes an infinite time for any object to
fall into the horizon.
Penrose introduced a set of coordinates that demonstrate that in the
frame of a particle falling into the black hole, fall towards the
singularity is inevitable once it crosses the horizon and it happens
in finite time.
More interestingly, if this particle is emitting light then all light
falls into the black hole once the particle crosses the horizon.
Penrose attributed the idea to Eddington and Finkelstein and these
coordinates are known as Eddington-Finkelstein coordinates.
Thus Penrose was able demonstrate that in a Schwarzschild black hole
infalling objects reach the singularity at the centre in a finite
time, and this fall is inevitable if the infalling object crosses the
horizon.
Penrose terms such surfaces that act as one way membranes as a trapped
surface.
\begin{figure}[!t]
\caption{Sir Roger Penrose is the Emeritus Rouse Ball Professor of
Mathematics at the University of Oxford, an emeritus fellow of
Wadham College, Oxford and an honorary fellow of St John's
College, Cambridge, and of University College London (UCL).
(Photo by Cirone-Musi, Festival della Scienza, CC BY-SA 2.0, https://commons.wikimedia.org/w/index.php?curid=19318743)}
\includegraphics[width=2.5truein]{RogerPenrose.jpg}
\end{figure}
Penrose proceeded further and was able to prove that if a trapped
surface forms during collapse of a star then the formation of a
singularity at the centre is inevitable.
This is the singularity theorem that he proved.
The proof required application of topological methods in general
relativity.
The implication of this theorem is that a star can collapse and form a
black hole, this astrophysical phenomena can lead to the formation of
a black hole.
Given that the known black hole solutions are stationary, this is a
significant step that connects these idealized solutions with complex
reality.
This proof was timely as there were doubts about whether black holes
can form in astrophysical processes or not.
Measurement of redshift of quasars like 3C273 already implied that
the total emission from these was very large and accretion around
black holes appeared to be the only feasible explanation at the time.
The proof also brings out an internal limitation of the general theory
of relativity.
Existence of singularities implies that the theory breaks down at some
points and hence the theory has limitations.
It is expected that as and when the quantum theory of gravity is
found, the singularities will be replaced by something more
tractable.
Roy Kerr provided solution for a rotating black hole in 1963.
This is the black hole solution that is clearly relevant in
astrophysics as all stars have a non-zero angular momentum and we
expect black holes that form due to collapse to have some angular
momentum as well.
Roger Penrose proposed a process that can lead to extraction of energy
from a rotating black hole.
This is also a very significant contribution.
The singularity theorems were generalized by Hawking and Penrose to
prove that the Universe had a singularity in the past.
However, what remains unproven is the so called {\sl cosmic censorship
conjecture}.
The conjecture states that a space-time singularity is always
surrounded by a horizon and hence cannot be seen by a distant
observer.
\section{Super Massive Black Holes}
Evidence has grown over the last five decades for presence of super
massive black holes at the centre of each galaxies.
This evidence has come from a variety of observations of velocities in
the vicinity of the black hole.
These observations are of stars or hot gas orbiting the black hole.
It is interesting that we had much better evidence for existence of a
super massive compact objects in other galaxies well before such
observations were attempted in our own galaxy.
A major obstacle in such observations in the Galaxy is scattering and
absorption by the intervening gas and dust, this effect is given the
name {\sl extinction} in astronomy.
One half of the Nobel prize for physics has been given to Reinhard
Genzel and Andrea Mia Ghez for their work that has established the
presence of a super massive compact object at the centre of the
Galaxy.
\begin{figure}[!t]
\caption{Reinhard Genzel is a German astrophysicist, co-director of
the Max Planck Institute for Extraterrestrial Physics, a professor
at LMU and an emeritus professor at the University of California,
Berkeley.
(Photo by MPE - Own work, CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=18151111)}
\includegraphics[width=2.5truein]{ReinhardGenzel.jpg}
\end{figure}
The program to observe the central region of the Galaxy was started by
Reinhard Genzel nearly three decades ago.
Andrea Ghez started her own program a few years later.
He identified a near infra-red band (K band, $\lambda \sim
2.2$~$\mu$m) as a suitable band for observations of the galactic
centre.
Extinction reduces the flux by nearly three orders of magnitudes in
the K band, the situation is much worse in the optical.
The first challenge that came up was the effect of the atmosphere.
Variations of density and temperature in the atmosphere lead to
distortion of the incoming wave front: astronomers refer to this
effect as {\sl seeing}.
We are familiar with this effect in terms of twinkling of stars: the
image of the star shifts around by a small amount and its intensity
fluctuates.
In images of stars taken over a long period, we get the sum of all the
shifted positions and hence image of each star becomes a bit blurry.
The number of stars in the direction of the galactic centre is very
large and therefore images of stars start to overlap.
The first technique used to get images without this problem was to
take many short images, align them and obtain a deep image of the
region without blurring of images.
These early studies allowed astronomers to get first estimates of the
speeds of stars in the central region and therefore constrain the mass
contained near the centre.
Shift to larger telescopes permitted astronomers to take deeper images
and observe fainter stars.
The most significant development was the use of adaptive optics.
\subsection{Active Optics and Adaptive Optics}
Active optics was developed three decades ago to overcome problems
that prevented construction of large telescopes.
Astronomers use reflector telescopes as mirrors can be supported from
the base and these do not suffer from chromatic aberrations.
As telescopes are reoriented to look at sources in different parts of
the sky, the shape of the mirror changes under its own weight.
In order to prevent this from causing distortion of images, thick
glass was used to make mirrors with the typical thickness being about
$1/5$ of the diameter of the mirror.
For large mirrors this causes two problems: the mirrors become very
heavy, and, the response of the material to changing temperature
causes distortion in shape as the temperature is not constant through
the body of the mirror.
Active optics was an attempt to work with thin mirrors of large
diameter where the shape is adjusted with the help of a computer model
that takes the orientation and temperature into account.
It is this technology that has allowed construction of telescopes with
large diameters in the last three decades and the same technology will
be used for making even larger telescopes with composite mirrors,
e.g., the thirty meter telescope (TMT) that India is also involved
in.
\begin{figure}[!t]
\caption{Andrea Mia Ghez is an American astronomer and professor in
the Department of Physics and Astronomy at the University of
California, Los Angeles. Her research focuses on the center of the
Milky Way galaxy.
(Photo by John D. - https://www.nsf.gov/discoveries/disc\_images.jsp?cntn\_id=133541\&org=NSF, Public Domain, https://commons.wikimedia.org/w/index.php?curid=94809121)}
\includegraphics[width=3truein]{AndreaMiaGhez.jpg}
\end{figure}
Adaptive optics makes use of the technology developed for active
optics to adjust the shape of the mirror for distortion of the
incoming wave front by the atmosphere.
This requires a measurement of the incoming wavefront and a rapid
adjustment of the shape of the mirror.
Several approaches have been developed for this.
If we have a bright point source in the field then we can sense the
wavefront by an array of lenses/mirrors: if the wavefront did not have
any distortions then all images will form at the same relative
location, whereas there will be small shifts if the wavefront cannot
be described as a plane wave.
The amount of shift can be used to sense the level of distortion and
an appropriate correction can be applied.
Bright stars are not present in all directions and astronomers use
laser guide stars to overcome this problem.
Adaptive optics permits astronomers to obtain diffraction limited
images and hence observe much fainter sources while improving
localization on the sky.
\subsection{Zooming in}
Observations with large telescopes: Keck telescopes in Hawaii and the
Very Large Telescopes (VLT) in Chile have allowed astronomers to
observe stars and track these in their orbits.
The orbital parameters of these stars constrain the mass contained
inside the orbit.
A key step was the discovery of a few stars with
time periods short enough for them to be tracked for more than
one complete orbit. This enables reliable determination of orbital
parameters for these stars.
The orbital parameters (semi-major axis and time period) can be used
to deduce the enclosed mass.
A very important inference from these studies has been that the mass
contained within the orbit reaches a constant value, indicating that
the contribution in the innermost regions is from a single source and
individual stars do not contribute any significant mass.
Indeed, the contribution of stars to the total mass up to a
radius that is twenty thousand times the orbit of the Earth around
the Sun is insignificant, i.e., the mass of the compact object dominates
in this region.
This, in essence is the observational evidence of the presence of an
object with a mass of about $4 \times 10^6$ times the mass of the Sun
in a region that is at most twenty times larger than the orbit of the
Earth around the Sun.
This is only about 200 times larger than the Schwarzschild
radius\cite{2020ApJ...889...61P}.
The mass remains almost constant to a scale that is a thousand times
larger than this.
\subsection{Recent Developments}
Astronomers are now starting to use interferometry for observations of
the galactic centre.
This will improve the resolving power by an order of magnitude and
improve determination of orbital parameters.
This will also permit discovery of fainter objects.
Improved sensitivity is expected to permit testing general theory of
relativity via observations of stars in orbits around the compact
object.
\section{Summary}
The Nobel prize for Physics in this year has recognized the
theoretical work that proved that it is possible for black holes to
form through astrophysical processes.
The work by Roger Penrose is built on foundations laid by a number of
scientists, most importantly Amal Kumar Raychaudhuri.
The Raychaudhuri equation was critical to the proof of the singularity
theorem given by Roger Penrose.
Recognition of the work by Reinhard Genzel and Andrea Mia Ghez in
establishing the presence of a super massive compact object at the
centre of the Galaxy stops short of calling it a black hole.
Observations using other means have established a stronger constraint
on supermassive compact objects in other galaxies.
Perhaps the observational discovery of black holes will be recognized
later in another Nobel.
Roy Kerr, who found the solution for a black hole with angular
momentum has also contributed significantly in this journey.
Without the Kerr solution, there would have been many concerns about
realistic models of black holes.
Readers can study more details at the Nobel foundation
website\footnote{https://www.nobelprize.org/uploads/2020/10/advanced-physicsprize2020.pdf}
and other reviews\cite{2020arXiv201106656S}.
I would like to thank Professor Rajaram Nityananda for useful comments
on an earlier draft of the manuscript.
| f74b80c2c7ca0bbef427e6c07fa3cbf78d051e74 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{}
\begin{abstract}
Online experimentation, also known as A/B testing, is the gold standard for measuring product impacts and making business decisions in the tech industry. The validity and utility of experiments, however, hinge on unbiasedness and sufficient power. In two-sided online marketplaces, both requirements are called into question. The Bernoulli randomized experiments are biased because treatment units interfere with control units through market competition and violates the “stable unit treatment value assumption”(SUTVA). The experimental power on at least one side of the market is often insufficient because of disparate sample sizes on the two sides. Despite the important of online marketplaces to the online economy and the crucial role experimentation plays in product improvement, there lacks an effective and practical solution to the bias and low power problems in marketplace experimentation. Our paper fills this gap by proposing an experimental design that is unbiased in any marketplace where buyers have a defined budget, which could be finite or infinite. We then provide generalizable system architecture for deploying this design to online marketplaces. Finally, we confirm our findings with empirical performance from experiments run in two real-world online marketplaces.
\end{abstract}
\keywords{A/B testing, controlled experiment, interference, online marketplaces}
\section{Introduction}
Online experimentation (also known as A/B testing) has become the gold standard for measuring product impact and making business decisions in the tech industry. Companies such as Google, Facebook, Amazon, Microsoft and LinkedIn have all embraced this paradigm and built in-house platforms to enable large-scale online experimentation \citep{10.1145/1835804.1835810, letham2019, Deng2016, Kohavi2012, 10.1145/2783258.2788602}, commercial platforms have also emerged to provide online experimentation as a service to the broader audience\cite{10.1145/3097983.3097992}. Because of the crucial role online experimentation plays in decision making and the scale at which it impacts the business, it is of utmost importance to ensure the unbiasedness and power of experiments. When an experiment is biased, the impact estimate is systematically incorrect and the quality of business decisions is compromised. This is especially detrimental if the bias is systematic. When the experiments do not have enough power, even large impacts are not statistically significant and are incorrectly neglected as "neutral", and when impacts do get detected, the throughput of experiments, therefore speed of innovation, is limited. These considerations are precisely why the majority of online experimentation literature is devoted to either designing unbiased experiments under an interference structure \citep{10.1145/3097983.3098192, DesignandAnalysisofExperimentsinNetworksReducingBiasfromInterference}, or to increasing the power of an unbiased experiment \citep{10.1145/2433396.2433413}.
The optimal strategy for designing unbiased and powerful experiments is highly dependent on the underlying interference structure. The most studied example of interference in the tech industry is that of the online social network \citep{10.1145/3097983.3098192, DesignandAnalysisofExperimentsinNetworksReducingBiasfromInterference}, where each unit is the same type of entity (for example, a member of a social network). Each unit can connect with other units and interact with her connections.
Another important interference structure in online platforms, yet received little coverage, is that of online marketplaces. Unlike in a social network, the marketplace population consists of at least two types of units, sellers who offer goods, services or attention and buyers who purchase them with some currency. Marketplaces are an important part of the online economy, and usually are the products responsible for revenue generation in a tech companies. Therefore, it is important for both tech companies and the health of the online economy that we are able to experiment in online marketplaces unbiasedly and with high power. Some examples of online marketplaces are ads marketplace, where advertisers purchase the attention of platform members in the form of ads impressions \citep{GoogleAds, FacebookAds}; hiring marketplace, where employers purchase the attention of job seekers in the form of job postings; and education platforms, where buyers purchase courses from educators that offer them.
The interference and network structure in online marketplaces present experimentation challenges to both unbiasedness and power. Bernoulli randomized experiments on either the seller-side or buyer-side violate the ``stable unit treatment value assumption”(SUTVA) \citep{doi:10.1198/016214504000001880, Johari2020ExperimentalDI} and are biased. To illustrate why SUTVA is violated, we use seller-side Bernoulli randomized experiments as an example. When a treatment increases the competitiveness of sellers, unless all buyers still have enough budget reserve to purchase more, the increased attractiveness of treatment sellers translates to relatively less attractiveness of control sellers, and the control sales decreases as treatment sales increases. Such interference in marketplaces violates the SUTVA assumption, which states a unit's potential outcome should depend only on the treatment it receives and not on treatments other units receive. Analogous argument explains why buyer-side Bernoulli randomized experiments also violate SUTVA and are biased. In the marketing sphere, this bias is termed the ``cannibalization bias", to highlight the fact that treatment sales is higher than control sales not because total sales grows, but because treatment sales ``cannibalizes" on control sales. Making a decision based on Bernoulli randomized experiments when cannibalization is present results in systematic overestimation of treatment effects. In any real-world marketplace, there are always limits to buyers' budgets and sellers' inventory, therefore both seller-side and buyer-side Bernoulli randomized experiments are subject to the cannibalization bias.
Beyond the cannibalization bias, marketplace experimentation is also plagued with insufficient experimental power. This happens because the two sides of the market usually consist of different types of entities, and are drastically disparate in numbers. For example, in an Ads marketplace, where buyers are enterprise advertisers or agencies, and sellers are consumer users, the number of buyers \citep{googleAdvertiser} can be three orders of magnitude smaller than the number of sellers \citep{googleStats}. Similarly, in a physical goods marketplace, where buyers are consumers and sellers are more likely to be small businesses, the number of buyers is 60 times larger than the number of sellers \citep{amazonStats}. This disparity in sample size naturally leads to disparity in experimental power, and the insufficient power will be further exacerbated if the side with smaller sample size also has longer tails in metric distributions. This power disparity, or relative insufficiency on one side of the marketplace is undesirable because the marketplace business is only sustainable if both sides continues to be optimized, and one side of the marketplace having significantly lower experimental power means the speed of innovation and optimization is throttled on this side.
Solutions to marketplace experimentation in existing literature roughly follow two approaches: the analysis approach which first models the interference structure and cannibalization bias, and then apply model-based adjustments to the biased Bernoulli randomized experiments \citep{Johari2020ExperimentalDI, 10.1145/2600057.2602837}; the design approach, which explore alternative designs that are unbiased under weak and verifiable assumptions on the marketplace \citep{doi:10.1080/01621459.2018.1527225, ebayMarket, DBLP:conf/aistats/BasseSL16}. The analysis approach typically requires first simplifying the marketplace to a parametric model, when the modeling assumptions themselves are often untestable. In fact, if the model could accurately capture the marketplace, experimentation is not necessary and treatment effects can be measured through simulations. Some design approaches \citep{doi:10.1080/01621459.2018.1527225, ebayMarket} are successful in achieving low bias under reasonable assumption, but the low bias comes at the cost of even lower power. Our paper contributes to the marketplace experimentation literature by proposing the budget-split design which overcomes both the bias and power challenges. The unbiasedness of the budget-split design requires one easy-to-validate assumption only, which is the buyers have a continuous budget (the budget can be finite or infinite). It is up to 30 times more powerful than alternative designs when deployed in the real-world marketplaces. The implementation and deployment of this design does not require an overhaul of the existing online marketplace infrastructure or the experimentation platform. Rather, only incremental changes to modules of the infrastructure. We note that our design shares some similarities with the budget allocation design in \cite{DBLP:conf/aistats/BasseSL16}. Although both designs split the ``budgets'' in some ways, they are fundamentally different: in \cite{DBLP:conf/aistats/BasseSL16}, an indirect, artificial “budget” (quota) is given to each buyer while the budget-split design considers the direct, natural budget that the buyers define. We also notice a recent blogpost in \cite{stitchFixBlog} that outlines an inventory-split counterpart of the budget-split design. Our formal analysis of the budget-split design can be directly applied to the inventory-split settings with little modification.
The impact of the budget-split design is significant. It has enabled marketplace experimentation free of cannibalization bias; boosted experimental power to detect treatment effects that previously would have been dominated by noise. As a result, marketplace innovation and optimizations are carried out with much higher confidence and velocity. In this paper, we use the ads marketplace as a running example to present the budget-split design. The methodology, system architecture and empirical results can be readily extend to any marketplace where the buyer has a well-defined budget, such as the hiring marketplace. The rest of the paper is organized as follows. In Section 2, the causal inference problem in marketplaces is formally defined using the potential outcome framework. Section 3 introduces the budget-split design contrasts it with alternative designs such as cluster-based randomization and alternating-day design. Section 4 highlights the key considerations in system architecture and deployment of the budget-split design to online marketplaces. The empirical results on unbiasedness and power of the budget-split design are reported in Section 5. Finally, Section 6 concludes the paper with discussions and future research.
\section{Experimentation in online ads marketplace}
\label{sec:oam}
Experiments in an online ads marketplace (OAM) falls into the family of bipartite experiments where experimentation takes place in a bipartite graph, linking two types of entities -- the buyers and the sellers\citep{pouget2019variance, doudchenko2020causal}. \textit{Sellers} are consumer members who visit the website, offer their attention and create requests for ads impressions, while \textit{buyers} are advertisers that fulfill these requests with ads impressions. Each member visit can result in multiple ads requests to ``sell'' (i.e., the webpage can have multiple slots to show ads impressions during one visit). Each advertiser can create multiple \textit{campaigns} that serve difference purposes, have different budgets, and target different member segments. Without loss of generality, we set our experimental units on the two sides as members and campaigns. Our framework can be applied to units at other granularity levels though. For example, it is common to set member sessions as the seller side units (finer) and advertisers as the buyer side units (coarser).
Suppose the marketplace contains $N$ members and $M$ campaigns, denoted by $i\in [N]=\{1,2,\ldots, N\}$ and $\cC_j, j\in [M]=\{ 1,2,\ldots, M\}$, respectively. For the $j$-th campaign $\cC_j$, let $B_j \geq 0$ be its budget and the member set $I_j \subset [N]$ be its target population. Let $\bm{B}=\{B_1,B_2,\ldots,B_M \}$ denote the budgets of all campaigns. We note that targeting is at the member level. Thus if a member is targeted by a campaign, all requests created by this member are targeted. In addition, the campaigns can have different strategy and behavior, such as the bidding strategy and the pacing strategy. We do not model these details and instead simply represent these nuisances with parameter $\bThe_j$. With these notations, we write the entire OAM as $\mathcal{M} = \{ (\cC_j, I_j, B_j, \bThe_j)^M_{j=1}\}$. Figure~\ref{fig:illustration} gives an illustration of an OAM.
There are two types of experiments in an OAM classified by the side of the market where the new feature is to be launched. When this feature acts on the seller-side, for example changing the placement of ads from top of the page to right rail of the page, a \textit{member-level experiment} (ME) or a seller-side experiment needs to be performed to test its effect where the randomization units are members. Similarly, when the feature acts on the buyer-side, such as adjusting the bid price of a campaign, we need to conduct a \textit{campaign-level experiment} (CE) or a buyer-side experiment where the randomization units are campaigns.
In this paper, we use ME to introduce the ideas. The budget-split design we propose also applies to CE. To simplify the demonstration, we defer the discussion of CE to Section~\ref{sec:bs_campaign_level} and focus on ME for now. In an ME, let $W_i\in\{0,1\}$ be the treatment received by member $i$, $\bW = (W_1,\ldots, W_N)$ the treatment assignment vector of all members. Classical designs, such as the completely randomized design, can be directly applied to the seller side:
\begin{defn}[Member-level completely randomized design]
A member-level completely randomized design is a probability distribution $\eta$ on $\{0,1\}^N$ such that for some $1\leq N_1\leq N$, $\eta(\bW) = 1$ if $\sum^N_{i=1} W_i = N_1$ and $\eta(\bW) = 0$ otherwise.
\end{defn}
Unfortunately, as we show in Section~\ref{sec:cannibalization_bias} and Section~\ref{sec:6.2}, applying this design in OAM without modification and ignoring the buyer side can severely bias the treatment estimation. Similar bias is entailed when applying classical designs on the buyer side and ignoring the seller side.
\def\mystrut{\vrule height 0.8cm depth 0.8cm width 0pt}
\def\mystrutt{\vrule height 0.4cm depth 0.4cm width 0pt}
\def\mystruttt{\vrule height 0.2cm depth 0.2cm width 0pt}
\begin{figure}[!h]
\begin{center}
\resizebox{0.55\textwidth}{!}{
\begin{tikzpicture}
\node[rectangle split, rectangle split parts=4,
draw, minimum width=3cm, font=\large, rectangle split part fill={gray!40,white, white,white},
rectangle split part align={center}] (t1)
{ {$\mathbf{Adv_1}$}
\nodepart{two}
\mystrut $campaign_{11}$
\nodepart{three}
\mystrutt $campaign_{12}$
\nodepart{four}
\mystrutt $campaign_{13}$
};
\begin{scope}[xshift=7cm, yshift=0.5cm]
\node[rectangle split, rectangle split parts=5,
draw, minimum width=2cm,font=\small, rectangle split part fill={blue!15,white, white, white},
rectangle split part align={center}] (t2)
{ {$\mathbf{Member_1}$}
\nodepart{two}
$request_{11}$
\nodepart{three}
$request_{12}$
\nodepart{four}
$request_{13}$
\nodepart{five}
$request_{14}$ };
\end{scope}
\begin{scope}[xshift=0cm, yshift=-5.5cm]
\node[rectangle split, rectangle split parts=3,
draw, minimum width=3cm,font=\large, rectangle split part fill={gray!40,white, white,white},
rectangle split part align={center}] (t3)
{ {$\mathbf{Adv_2}$}
\nodepart{two}
\mystrutt $campaign_{21}$
\nodepart{three}
\mystrut $campaign_{22}$ };
\end{scope}
\begin{scope}[xshift=7cm, yshift=-3cm]
\node[rectangle split, rectangle split parts=3,
draw, minimum width=2cm,font=\small, rectangle split part fill={blue!15,white, white, white},
rectangle split part align={center}] (t4)
{ $\mathbf{Member_2}$
\nodepart{two}
$request_{21}$
\nodepart{three}
$request_{22}$ };
\end{scope}
\begin{scope}[xshift=0cm, yshift=-9cm]
\node[rectangle split, rectangle split parts=3,
minimum width=4cm,font=\Large,
rectangle split part align={center}] (t5)
{ \textbf{$\cdots$}};
\end{scope}
\begin{scope}[xshift=7cm, yshift=-6cm]
\node[rectangle split, rectangle split parts=3, draw,
minimum width=2cm,font=\small, rectangle split part fill={blue!15,white, white, white},
rectangle split part align={center}] (t6)
{ $\mathbf{Member_3}$
\nodepart{two}
$request_{31}$
\nodepart{three}
$request_{32}$};
\end{scope}
\begin{scope}[xshift=7cm, yshift=-9cm]
\node[rectangle split, rectangle split parts=3,
minimum width=2cm,font=\Large,
rectangle split part align={center}] (t7)
{ \textbf{$\cdots$}};
\end{scope}
\draw[->] (t1.two east) to [out=0, in=180](t2.text west);
\draw[->] (t1.two east) to [out=0, in=180](t4.text west);
\draw[->] (t1.two east) to [out=0, in=180](t6.text west);
\draw[->] (t1.three east) to [out=0, in=180](t2.text west);
\draw[->] (t1.four east) to [out=0, in=180](t2.text west);
\draw[->] (t1.four east) to [out=0, in=180](t4.text west);
\draw[->] (t3.two east) to [out=0, in=180](t4.text west);
\draw[->] (t3.three east) to [out=0, in=180](t4.text west);
\draw[->] (t3.three east) to [out=0, in=180](t6.text west);
\draw[->] (t3.three east) to [out=0, in=180](t7.text west);
\end{tikzpicture}}
\vspace{-0.3cm}
\end{center}
\caption{An illustration of the online ads marketplace. The size of each campaign block represents its budget.}
\label{fig:illustration}
\end{figure}
\subsection{Potential outcomes and causal estimands}
\label{sec:po}
We
adopt the potential outcomes framework \citep{rubin1974estimating} for causal inference. For each $i\in[N]$ and $j\in[M]$, let $Y_{ij}(\bW;\bB)$ be the scalar potential outcome of member $i$ and campaign $\cC_j$ given the treatment assignment vector and the budgets. Note that the potential outcomes are specified for each $(member, campaign)$ pair, which is the finest level in our design. A similar setup is also considered in \cite{basse2016randomization}. This general setup allows outcomes measured at coarser levels to be written as aggregations. For example, for situations in \cite{doudchenko2020causal} and \cite{pouget2019variance} where the potential outcomes are specified in terms of only one side of the market, they can be defined by summing over indexes in $Y_{ij}(\bW;\bB)$ of the other side. Let $\bW_{-i}$ be the $(N-1)$ vector by removing the $i$-th element from $\bW$, we also consider an alternative parametrization of the potential outcomes as suggested by \cite{chin2018central}: $Y_{ij}(\bW;\bB)= Y_{ij}(W_i, \bW_{-i}; \bB)$.
We focus on two metrics of special interests in an OAM: the value delivered to an advertiser and the revenue received by the marketplace operator. Specifically, let $Y_{ij}(\bW;\bB) \geq 0$ be the value delivered to campaign $\cC_j$ (in fact, to the underlying advertiser) by the ads requests created by member $i$. If $i\not\in I_j$, member $i$ is not in the target population of $\cC_j$ and $Y_{ij}(\bW;\bB)=0$ by definition. If $i\in I_j$, $\cC_j$ competes for the ads slots from member $i$'s visits with all other campaigns $\cC_{j^\prime}$ such that $i\in I_{j^\prime}$ by participating in an auction, such as the firce price auction or the generalized second price auction \citep{edelman2007internet}. Note that the potential outcomes depend on the campaign budgets. In an OAM, $\cC_j$'s own budget $B_j$ is used by the system to guide its pacing and bidding strategy \citep{agarwal2014budget}, affecting its own bid underlying each $Y_{ij}$. At the same time, the resulting $Y_{ij}$ also depends on bids from other campaigns, which are themselves determined by their own budgets and actual spending. Unlike in \citep{basse2016randomization}, we directly focus on the outcomes of the auction, ignoring the underlying bids as well as the auction mechanism for generality.
Aside from the value delivered to the advertisers, another crucial metric is the corresponding revenue received by the marketplace operator, denoted as $Y^*_{ij}(\bW;\bB)$. It is worth noting that $Y^*_{ij}(\bW;\bB)$ may or may not equal $Y_{ij}(\bW;\bB)$. Due to the complex optimization procedure under the hood, $Y_{ij}$ depends on $\bm{B}$ in a ``not exact" manner, meaning that it is possible for $\sum_{i\in I_j}Y_{ij}(\bW;\bB) > B_j$, a phenomenon referred to as ``over-delivery". On the other hand, the platform cannot charge a campaign more than its budget. Therefore, we have for each $j$,
\begin{equation}
\begin{aligned}
0 \leq Y^*_{ij}(\bW;\bB) & \leq Y_{ij}(\bW;\bB) \\
\sum\limits_{i\in I_j} Y^*_{ij}(\bW;\bB) &\leq B_j.
\end{aligned}
\end{equation}
In this work, we conduct inference conditioning on the potential outcomes and take treatment assignment as the only source of randomness. This randomization based inference strategy has the advantage of not entailing any modeling assumptions on $Y_{ij}(\bW;\bB)$. Due to the complex and evolving ecosystem in an OAM, modeling these potential outcomes can be extremely challenging.
We would like to estimate the treatment effect in terms of the total delivered value and the total received revenue conditioning on the budgets:
\begin{equation}
\begin{aligned}
\tau(\bB) & =\sum\limits_{j\in[M]} \sum\limits_{i\in I_j} [Y_{ij}(\bm{1};\bB) -Y_{ij}(\bm{0};\bB)] = \sum\limits_{j\in[M]} \tau_j(\bB) \\
\tau^*(\bB) & = \sum\limits_{j\in[M]} \sum\limits_{i\in I_j} [Y^*_{ij}(\bm{1};\bB) -Y^*_{ij}(\bm{0};\bB)] = \sum\limits_{j\in[M]} \tau^*_j(\bB),
\end{aligned}
\end{equation}
where $\bm{1} = (1,1,\ldots,1)_N$ and $\bm{0} = (0,0,\ldots,0)_N$ denote the all treatment and all control treatment assignment.
The validity of vanilla A/B testing hinges critically on the stable unit treatment value assumption (SUTVA) \citep{imbens2015causal}, which states that (i) there is only a single form of each treatment variant and (ii) a unit's potential outcome is unaffected by the treatment assignment of other units. The no interference assumption, however, is usually violated in marketplace experiments due to the limited campaign budgets. One special scenario where SUTVA at can be justified is in an imaginary world without budget constraints, so that the campaigns are free to bid on any of their targeted members based on their own evaluation of that member's value. This is essentially the SUTVA for bids assumption made by \cite{basse2016randomization}. Formally, we have
\begin{assmp}[No interference under unlimited budget]
If $B_j = \infty$ for each $j\in[M]$, we have $Y_{ij}(\bW; \bm{\infty}) = Y_{ij}(W_i; \bm{\infty})$.
\label{assmp:sutva}
\end{assmp}
In this case, the following simple plug-in estimator is unbiased:
\begin{equation}
\begin{aligned}
\hat\tau_j(\bm{\infty}) &= \sum\limits_{i\in I_j} \frac{W_i\cdot Y_{ij}(\bW; \bm{\infty})}{N_1/N} - \sum\limits_{i\in I_j} \frac{(1-W_i)\cdot Y_{ij}(\bW; \bm{\infty})}{N_0/N},
\end{aligned}
\label{eq:naive}
\end{equation}
where $N_1 = \sum^{N}_{i=1}w_i$, $N_0 = N - N_1$. Moreover, since there is no budget constraint, $Y_{ij}^*(\bW) = Y_{ij}(\bW) = Y_{ij}(W_i)$ and the corresponding plug-in estimator $\hat\tau_j^*(\bm{\infty})$ for $\tau_j^*(\bm{\infty})$ is also unbiased.
\subsection{Finite budget and the cannibalization bias}
\label{sec:cannibalization_bias}
In reality, it is very unlikely that ads campaigns have unlimited budgets. As a result, Assumption \ref{assmp:sutva} no longer holds and the potential outcomes are affected. With a finite budget, each campaign now needs to take a more ``global" view over all its target members when planning its bid such that the total budget is split to different members wisely. For example, if the treatment has an edge over the control, a wise campaign should take this advantage and allocate more budget to members in the treatment group, resulting in interference between the potential outcomes. Due to this interference, the naive estimator for $\tau_j(\bB)$ and $\tau^*_j(\bB)$ as in (\ref{eq:naive}) is no longer unbiased. To further illustrate this, we consider two examples. The first example illustrate that the naive estimator $\hat\tau_j^*(\bB)$ of $\tau_j^*(\bB)$ is biased in a very common scenario that can happen in a real OAM due to the strict budget constraints. The second example considers a more fundamental case, where even $\hat\tau_j(\bB)$ is biased. From now on, we assume that $0 < B_j < \infty$ for campaign $j\in [M]$.
\begin{exam}[Full budget utilization]
Imagine an OAM that has a single buyer with a \$100 budget, willing to purchase the sellers attention at \$20 per impression; and many sellers each with a different propensity to pay attention to the buyers ads. Naturally, to better serve both the buyer's and sellers' interest, the marketplace improves the algorithm for matching the buyer with sellers who are most interested paying attention to the buyers ads. If the buyer already manages to spend the \$100 budget under sub par seller ranking ($\sum _{i\in[N]}Y_{i1}^*(\bm{0}; 100) = 100$), then improving seller ranking cannot further increase the buyer spending because there simply is no untapped budget (by definition, $\sum _{i\in[N]}Y_{i1}^*(\bm{1}; 100) \leq 100$ and $\tau^*_1(100) \leq 0$). On the other hand, had we run a member-level completely randomized experiment, with sellers in treatment being ranked by a better algorithm, we would have observed a positive treatment effect of the improved algorithm from the plug-in estimator.
\end{exam}
\begin{exam}[Universally beneficial treatment with diminishing returns]
For some campaign $j$ with $I_j = [N]$, suppose that
{\small\begin{equation}
\begin{aligned}
& \hspace{0.3cm} Y_{ij}(0,|\bW_{-i}| = N - 1;\bB) \\
&< \cdots \\
& < Y_{ij}(0,|\bW_{-i}| = 1;\bB) \\
&< Y_{ij}(0,\bm{0};\bB) < Y_{ij}(1, \bm{1};\bB) \\
& < Y_{ij}(1, |\bW_{-i}| = N-2;\bB) \\
& < \cdots \\
& <Y_{ij}(1, |\bW_{-i}| = 0;\bB),
\end{aligned}
\label{eq:dimreturn}
\end{equation}}
for any $i\in [N]$, where for $0\leq K < N$, $Y_{ij}(W_i, |\bW_{-i}| = K;\bB) = \{Y_{ij}(W_i, \bW_{-i}; \bB): \sum_{i\in I_j}\bW_{-i} = K \}$. Intuitively, in this scenario, the treatment increases the value of each member. However, members in the treatment group are more valuable when the size of the treatment group is smaller; members in the control group are less valuable when the size of the treatment group is larger. (Consider an analogy to a job market in a county. If a job applicant is the only college graduate, he/she would probably be paid very high; with more and more applicants holding a college degree, the value of such a degree diminishes while those only holding a high school degree can find it harder to even get a job.) In this case,
\begin{equation}
\begin{aligned}
& \hspace{0.55cm}\E[\hat\tau_j(\bB)] - \tau_j(\bB) \\
& = \left\{ \E\left[ \sum\limits_{i =1}^N \frac{W_i\cdot Y_{ij}(1, |\bW_{-1}| = N_{1} - 1;\bB)}{N_1/N}\right] - \sum\limits_{i=1}^NY_{ij}(1,\bm{1}; \bB)\right\} \\
& + \left\{ \sum\limits_{i=1}^NY_{ij}(0,\bm{0}; \bB) - \E\left[ \sum\limits_{i=1}^N \frac{(1-W_i)\cdot Y_{ij}(0, |\bW_{-1}| = |N_{1}| - 1;\bB)}{N_0/N}\right] \right\}.
\end{aligned}
\label{eq:canni}
\end{equation}
By (\ref{eq:dimreturn}), the two terms on the RHS are all positive. Thus $\hat\tau_j(\bB)$ has a positive bias.
\end{exam}
Note that in (\ref{eq:canni}), $\sum_{i=1}^NY_{ij}(1,\bm{1}; \bB)$ is overestimated while $\sum_{i=1}^NY_{ij}(0,\bm{0}; \bB)$ is underestimated. We thus refer to the bias of this naive estimator as the \textit{cannibalization bias} as the new version of the product appears to be having an edge over the old one in an experiment, not only because it drives extra values of members in treatment, but also because it ``cannibalises" the members' values in control.
\subsection{Switchback design}
\label{sec:alternating}
One popular strategy to get rid of this cannibalization bias is to treat the entire OAM at different times as the experimental units and adopt the switchback design \citep{bojinov2020design}. Let $t=1,2,\ldots, T$ be $T$ ordered time points. The switchback design assigns random treatment to each $\mathcal{M}_t$. Thus for each fixed $t$, all units (members or campaigns depending on whether the experiment is ME or CE) are assigned to the control or the treatment simultaneously. Since only one treatment is active at any time, there is no cannibalization bias due to cross treatment group intervention.
In practice, applying a switchback design to an OAM can be challenging. Firstly, although the design prevents any cross treatment interventions at any time point, it allows interventions across time points, often referred to as a carryover effect \citep{bojinov2020design}. In an OAM, the level of the carryover effect depends on the features being tested as well as other factors in the ecosystem. In general, it is hard to understand these carryover effects and analysis the switchback experiment properly. Secondly, a switchback experiment can be highly inefficient when it takes sometime for the effect of the feature to become measurable. For example, some features target at improving the performance of a campaign over a week. In this case, the time gap between different assignments has to be at least a week and the overall experimental period can stretch over weeks or even months. Thirdly, there is a conflict between the validity and the efficiency of a switchback experiment. On the one hand, a switchback experiment needs to run long enough to collect enough samples; on the other hand, it implicitly assumes that the OAM does not change much during the experimentation period, which is likely to be unrealistic when this period is long.
\section{Budget-split design}
In this section, we introduce a novel design that removes the cannibalization bias in both the member-level experiment and the campaign-level experiment in an OAM while bypassing the challenges in the switchback design. Our key motivation is that the cannibalization bias comes from interactions between units in different treatment groups. If these interactions are blocked, the cannibalization bias will not exist. Similar to the idea in \cite{ha2020counterfactual},
our goal is to simulate the two ``counterfactuals'' of the ads marketplace---one under the control and another under the treatment---with a design, such that a causal relationship can be established by directly comparing these two simulated marketplaces. The success of our strategy depends on two factors: (i) how similar the two simulated marketplaces are before the treatment, and (ii) how similar these simulated marketplaces are to the original marketplace before the treatment. As we shall see, our design automatically ensures (i), that is, the two marketplaces have exactly the same campaigns, very similar members, and the same underlying mechanism (i.e., auction mechanism, bidding strategy, etc.) We state (ii) formally in Assumption~\ref{assmp:average} and validate this assumption with real examples in Section~\ref{sec:6.2}.
Without loss of generality, we assume $I_j=[N]$ for all $\cC_j, j\in[M]$ in this section. Recall that the original OAM is $\mathcal{M} = \{ (\cC_j, I_j, B_j, \bThe_j)^M_{j=1}\}$. We create two independent marketplaces $\mathcal{M}^{(0)}$ and $\mathcal{M}^{(1)}$ out of $\mathcal{M}$ with the following steps:
\begin{enumerate}
\item perform a random split of the members into two buckets of size $N^{(0)}$ and $N^{(1)} = N - N^{(0)}$. Let $d_i \in \{0, 1\}$ be the bucket indicator of member $i$ and $\bd = (d_1,d_2,\ldots, d_N)$;
\item split the budget of each campaign $j$ accordingly: \begin{equation}
\begin{aligned}
B_j^{(0)} = \frac{N^{(0)}}{N}\cdot B_j &, \quad B_j^{(1)} = \frac{N^{(1)}}{N}\cdot B_j;
\end{aligned}
\end{equation}
\item create a campaign $\cC_{j_0}$ with a budget $B_j^{(0)}$ and a target population $I_{j_0} = \{i: d_i = 0\}$. Let this campaign inherit all features (summarized by $\bThe_j$) from campaign $\cC_j$. Similarly, create a campaign $\cC_{j_1}$;
\item replace $\mathcal{M}$ with two independent marketplaces:
\begin{equation}
\begin{aligned}
\mathcal{M}^{(0)} & = \{ (\cC_{j_0}, I_{j_0}, B^{(0)}_j, \bThe_j)^M_{j_0=1}\}\\
\mathcal{M}^{(1)} & = \{ (\cC_{j_1}, I_{j_1}, B^{(1)}_j, \bThe_j)^M_{j_1=1}\}.
\end{aligned}
\end{equation}
\end{enumerate}
Recall that the treatment effect of interests is
\begin{equation}
\begin{aligned}
\tau_j(\bB) = \sum\limits^{N}_{i=1} [Y_{ij}(\bm{1};\bB) - Y_{ij}(\bm{0};\bB)].
\end{aligned}
\end{equation}
Intuitively, under the budget split setup, we use one of the
$\mathcal{M}^{(l)}$'s to estimate $\sum^{N}_{i=1}Y_{ij}(\bm{1};\bB)$, and use another to estimate $\sum^{N}_{i=1}Y_{ij}(\bm{0};\bB)$. To this end, let $\tilde W_l\in\{0,1\}$ be the treatment indicator of OAM $\mathcal{M}^{l}$, where $\tilde W_1 \sim \text{Bernoulli}(p)$ and $\tilde W_0 = 1-\tilde W_1$. If $\tilde W_l = 1$, we apply the treatment to members in bucket $l$; if $\tilde W_l = 0$, these members will receive the control.
Let $\bB^{(l)} = \{ B_1^{(l)}, B_2^{(l)}, \ldots, B_M^{(l)} \}$ and $\bW^{(l)} = \{W_i: i\in I_{j_l} \}$ for $l=0,1$. Similar to the definition of $Y_{ij}(\bW;\bB)$, we define the potential outcomes associated with campaign $\cC_{j_l}$ as $Y_{ij}^{(l)}(\bW^{(l)}; \bB^{(l)}\mid \bm{d})$ for $l=0, 1$ and $i\in I_{j_l}$. Note that these potential outcomes also depend on $\bd$. (When defining these potential outcomes, we are actually making a no interference assumption, which is formally stated in Assumption~\ref{assmp:bucket}.) Given a split $\bd$, we consider the following estimator:
\begin{equation}
\begin{aligned}
\hat\tau_j^{BS}(\bB) & = \sum\limits_{l=0,1} \tilde{W}_l\sum\limits_{i=1}^N\frac{\mathbbm{1}(d_i = l)Y_{ij}^{(l)}(\bm{1}_{N^{(l)}}; \bB^{(l)} \mid \bd)}{N^{(l)}/N} \\ & - \sum\limits_{l=0,1} (1-\tilde{W}_l) \sum\limits_{i = 1}^N \frac{\mathbbm{1}(d_i = l)Y_{ij}^{(l)}(\bm{0}_{N^{(l)}}; \bB^{(l)} \mid \bd)}{N^{(l)}/N}.
\end{aligned}
\label{eq:bs}
\end{equation}
To study the properties of $\hat\tau_j^{BS}(\bB)$, we first introduce the following definition:
\begin{defn}[proportionally restricted campaign]
For $j\in[M]$, we define a \textit{proportionally restricted version} of $\mathcal{C}_j$ as $\mathcal{C}_j(K)$, for $1\leq K\leq N$, such that $\mathcal{C}_j(K)$ is only allowed to target at $K$ members completely at random with a budget of $\cB_j(K) = KB_j/N$. Moreover, suppose that all campaigns are proportionally restricted in a coupled manner, that is, they are restricted to the same members. Then for each $i\in [N]$, we can define the potential outcomes associated with $\cC(K)$ as $\cY_{ij}(\bW;\cB(K) \mid \bm{d})$, where $\cB(K)=\{\cB_1(K), \ldots, \cB_M(K) \}$, $\bm{d} = (d_1, d_2,\ldots, d_N)$, $d_i$ the indicator of whether member $i$ is in the restricted target population. For convenience, we define $\cY_{ij}(\bW;\cB(K) \mid \bm{d}) = 0$ if $d_i = 0$.
\end{defn}
In practice, this definition is relevant to the throttling procedure that can be applied before the auctions take place \citep{basse2016randomization}. For example, due to resource constraint, the system could not afford to allow each campaign to bid for all its targeting members. We next state two assumptions on the potential outcomes of proportionally restricted campaigns. Assumption~\ref{assmp:bucket} limits the interference to members within the restricted target population. Assumption~\ref{assmp:average} ensures that the restricted campaigns are on average ``similar enough" to the unrestricted ones such that they can be used to perform inference about the original campaigns.
\begin{assmp}[Limited interference]
For each $i\in[N]$, $j\in[M]$ and $\bW$, we assume $\cY_{ij}(\bW;\cB(K) \mid \bm{d}) = \cY_{ij}(\bW_{\bm{d} = 1};\cB(K) \mid \bm{d})$, where $\bW_{\bm{d} = 1}$ represents the elements in $\bW$ with $\bm{d} = 1$.
\label{assmp:bucket}
\end{assmp}
\begin{assmp}[Stable system]
For each $j$ and $1\leq K\leq N$, we assume
\begin{equation}
\begin{aligned}
\frac{1}{N}\sum\limits^{N}_{i=1}Y_{ij}(\bm{1};\bB) & = \E_{\bm{d}}\left[ \frac{1}{K}\sum\limits_{i=1}^N d_i \cY_{ij}(\bm{1}_K;\cB(K) \mid \bm{d}) \right] \\
\frac{1}{N}\sum\limits^{N}_{i=1}Y_{ij}(\bm{0};\bB) & = \E_{\bm{d}}\left[ \frac{1}{K}\sum\limits_{i=1}^N d_i \cY_{ij}(\bm{0}_K;\cB(K) \mid \bm{d}) \right].
\end{aligned}
\end{equation}
\label{assmp:average}
\end{assmp}
Assumption~\ref{assmp:average} characterizes a degree of stability of the system. It requires that a proportionally restricted campaign behaves in a similar manner to its original campaign so that the two created OAMs are similar to the original one. In practice, this assumption is more likely to hold when (i) $K$ is large and (ii) $N$ is large. When $K$ is close to $N$, the $\cC_j(K)$ is essentially $\cC_j$. When $K$ decreases, the ecosystem underlying $\cC_j(K)$ becomes less similar to that of $\cC_j$ and the reliability of this assumption decreases accordingly. In the extreme case when $K$ is close to $1$, $\cC_j(K)$ essentially loses all its flexibility in participating in auctions. As a result, the potential outcomes of $\cC_j(K)$ can be totally distorted from those of $\cC_j$. In reality, there is usually some burn-in threshold $K_0$ such that the system (i.e., the bidding strategies of the campaigns) stabilizes with $K_0$ members. In this case, if $N$ is very large such that $K > K_0$, $\cC_j(K)$ would behave very similarly to $\cC_j$. Fortunately, in OAM, $N$ is usually at least at the scale of millions, in which case $K > K_0$ can be easily satisfied with $K = O(N)$. We note that $K_0$ depends on the level of homogeneity of members. $K_0$ of a market where the members are very similar to each other is generally smaller than $K_0$ in a more heterogeneous market.
\begin{thm}\label{thm1}
Under Assumption~\ref{assmp:bucket} and Assumption~\ref{assmp:average}, if in addition $N^{(0)} = N^{(1)} = N/2$, $p = 0.5$, the budget split estimator in (\ref{eq:bs}) is unbiased: for each $j\in[M]$,
\begin{equation}
\begin{aligned}
\E_{\tilde W_1, \bd}[\hat\tau_j^{BS}(\bB) ] = \tau_j(\bB).
\end{aligned}
\end{equation}
\end{thm}
The proof is standard and can be found in Appendix B.
\subsection{Campaign-level experiments}
\label{sec:bs_campaign_level}
So far, we have been focusing on ME. Since the budget split framework creates two complete OAM with both the buyer side and the seller side, it can also be used for CE. A CE is involved when the features to be tested act on the buyer-side. Examples of such features include UI changes at the campaign management portal and adjustments to the underlying campaign bidding strategy.
Similar to the setup in Section~\ref{sec:oam}, let $W^c_j \in \{0, 1\}$ be the treatment received by campaign $\cC_j$, $j\in[M]$ and let $Y^c_{ij}(\bW^c; \bB)$ be the corresponding potential outcomes. Again, we consider the treatment effect in terms of the total delivered value:
\begin{equation}
\begin{aligned}
\tau^c(\bB) & =\sum\limits_{j\in[M]} \sum\limits_{i\in I_j} [Y^c_{ij}(\bm{1}^c;\bB) -Y^c_{ij}(\bm{0}^c;\bB)] = \sum\limits_{j\in[M]} \tau^c_j(\bB).
\end{aligned}
\end{equation}
Under the budget split framework, the setup for CE is essentially the same as for ME. In this case, $\tau^c_j(\bB)$ is estimated directly by comparing the two campaigns $\cC_{j_0}$ and $\cC_{j_1}$ with an estimator similar to (\ref{eq:bs}).
Although the same budget-split framework is used for both ME and CE, the resulting experiments have very different nature. In ME, the budget split design essentially performs a completely randomized experiment at the 50-50 split, where the treatment assignments coincide with the budget allocations. Here we adopt the budget split design to remove the cannibalization bias. In CE, the budget split design has a more direct motivation. Namely, we create two (approximately) identical replicates for each experimental units before the experiment and solve the ``fundamental problem of causal inference" \citep{holland1986statistics} by making the two potential outcomes of each unit directly observable. Instead of estimated or imputed, the potential outcomes are \textit{observed} under the budget split design. Therefore, the causal estimands are also observed directly.
For CE, both the switchback design (SB) and the budget split design (BS) aim at creating ``counterfactuals" of any experimental unit (campaign) and making the potential outcomes observable under both treatments. In SB, campaign $\cC_j$ at different time points serve as its ``counterfactuals". In the BS, the ``counterfactuals" are represented by the two half-sized campaigns $\cC_{j_0}$ and $\cC_{j_1}$ that are active simultaneously. When making causal comparisons, SB keeps the size of the campaigns unchanged while BS controls for the time factor.
\section{System Architecture}
\label{sec:sys}
Having presented the budget-split design and highlighted its advantages over member-level experiments, campaign-level experiments and switchback experiments, we now give guidelines for the system architecture changes for deploying this design onto the online ads marketplace. Note that the modules in the ads marketplace are also ubiquitous in other types of marketplaces such as hiring marketplaces and physical goods marketplaces.
To set the context, we introduce the important components in the infrastructure of an online ads marketplace:
\begin{enumerate}
\item At the core of an online ads marketplace, is an \textbf{ad server} that receives incoming ads requests and responds with ads impressions in real time. Within the ads server, a pacing/bidding module controls the spending speed of each campaign, and eventually when to stop serving impressions. The pacing/bidding module makes serving decisions based on auction signals such as bid price, targeting and auction participation rate.
\item The \textbf{tracker service} supports the ad server by keeping track of ads events such as impressions, clicks and tallying the campaign-level cumulative spending, along with other performance metrics, such as the click-through rate. Combined with the core ad server, it forms a feedback loop for ads serving, enabling the pacing/bidding module to be built as a feedback controller.
\item Finally, the marketplace interfaces with advertisers through the \textbf{campaign management system}, where advertiers create and manage their ads campaigns, as well as get billed and reported on campaign performance.
\end{enumerate}
\begin{figure}[h]
\centering
\includegraphics[width=0.75\linewidth]{arch1.png}
\caption{System architecture for the budget-split design.}
\label{fig:arch}
\end{figure}
To implement the budget-split design, the components marked as blue in figure \ref{fig:arch} are modified, in the following ways:
\begin{enumerate}
\item In the request receiving and responding tier of the ad server, all requests from a member is randomized into either treatment or control depending on the member ID. This is the member-level randomization as described in Section 3.
\item In the tracker service, treatment assignment in step 1 together with ads campaign ID are recorded in the tracking information of the request.
\item In the pacing/bidding module of the ad server, all the originally campaign-level controller inputs are now replaced with their campaign-treatment-level counterparts. As an example, budget is now set at campaign-treatment-level as described in Section 3 rather than at campaign-level. Furthermore, feedback signals are provided at per-treatment level. Consequently, the pacing/bidding controller outputs auction signals at the campaign-treatment-level.
Then, the real time auctions, along with any other serving decisions such as whether an ad should stop or resume serving, are carried out using the campaign-treatment-level signals in lieu of campaign-level ones.
\end{enumerate}
With these changes in place, the feedback data and control paths for each treatment become entirely independent of other treatments, for each campaign and across all campaigns. Effectively every treatment now owns a standalone marketplace. This is exactly an implementation of the budget-split design as defined in Section 3.
\section{Empirical Results}
The budget-split experimentation has since been deployed to two online marketplaces. Similar to what was proven in Section 3, it was shown to be much more powerful than alternative designs such as campaign-level experiments or switchback experiments; and not susceptible to cannibalization bias as campaign-level and member-level experiments are.
\subsection{Power Gain}
In the two marketplaces where budget-split was deployed, we compared the power curves of budget-split experiments versus campaign-level and switchback experiments. Switchback experiments were analyzed with paired permutation tests to boost the power; all other designs are analyzed with two-sample t-tests.
\begin{figure}[h]
\centering
\includegraphics[width=0.6\linewidth]{power_ads.png}
\caption{Power curves for marketplace 1}
\label{fig1}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.6\linewidth]{power_jobs.png}
\caption{Power curves for marketplace 2}
\label{fig2}
\end{figure}
As shown in Figure \ref{fig1} and Figure \ref{fig2}, for the effect size that budget-split has 80\% power to detect, campaign-level experiments have power 5.2\% (Marketplace 1) and 12\% (Marketplace 2), while switchback experiments have only 5.1\% (Marketplace 1) and 5.2\% (Marketplace 2) power to detect the effect.
Even though switchback experiments are expected to be unbiased when no carry-over effect exists in the marketplace, they practically have such low power that detecting treatment effects becomes challenging. It is worth noting that 5\% power is actually the lowest power possible for a 5\% level test. Campaign-level experiments are not only biased, but also just barely more powerful than switchback experiments. In practice, budget-split experiments have enabled detection of negative product impacts on the order of millions of dollars (in terms of annual revenue) that previously went undetected in a campaign-level experiment because of insufficient power.
\subsection{Robustness against Cannibalization Bias}\label{sec:6.2}
As discussed in Section 3, the budget-split design effectively removes interference between treatment units and control units through splitting the campaign budgets, and is not susceptible to the cannibalization bias even in place of market competition. In comparison, switchback experiments are unbiased when no carry-over effect is present; cluster-based randomization is unbiased when there is no interference among clusters, which often is not practically feasible\cite{10.1145/3097983.3098192}; member-level and campaign-level experiments are biased because they do not account for the interference from market competition at all. In this section, we compare the impacts measured in budget-split design versus in member-level experiments to showcase the amount of cannibalization bias. It is worth noting that budget-split design was not compared against campaign-level experiments, switchback experiments or cluster-based experiments because these designs are too insufficiently powered for comparison.
\begin{figure}[h]
\centering
\includegraphics[width=0.6\linewidth]{Cannibalization_bias.png}
\caption{Measurement of Cannibalization Bias}
\label{fig3}
\end{figure}
The presence of cannibalization bias in marketplace experiment when interference are not accounted for is confirmed in Figure \ref{fig3}. The exact amount of bias depends on the nature of the treatment, but not accounting for interference typically results in one to two times overestimation of the treatment effect. We have not yet encountered a treatment where cannibalization bias changes the sign of the treatment effect, but this is not impossible if, for example, the treatment positively impacts budget-constraint campaigns and negatively impacts non budget-constraint campaigns.
\section{Discussions}
In this paper, we presented the budget-split design as the solution to cannibalization bias and insufficient power in marketplace experimentation. Rather than relying on modeling assumptions that are often impossible to validate, the budget-split design only requires one side of the marketplace having a defined and splittable budget, and no additional assumptions. We formulated the marketplace experimentation problem under the potential outcome framework, defined the estimand, derived an unbiased estimator for the estimand under budget-split design. We showed the bias and variance reduction nature of budget-split design relative to alternative designs with empirical results.
This work, however, is only the start of experimentation in online marketplaces. Future work is need to accurately compute the variance estimators under the budget-split design; extend the budget-split design to marketplaces where the budget is discrete; and increase throughput of budget-split experiments by enabling multiple simultaneous orthogonal budget-split experiments in the marketplace.
\section*{Acknowledgements}
We would like to thank Weitao Duan, Shan Ba, Xingyao Ye and Giorgio Martini for their careful review and feedback; Wei Wei, Qing Duan, Di Luo for implementing the budget-split platform; Shahriar Shariat, Vangelis Dimopoulos, Giorgio Martini and Junyu Yang for contributing to experimental and metric designs and analyses; and everyone who utilized and helped improve the platform. We would also like to thank Ya Xu, Parvez Ahammad, Le Li and Jerry Shen for their continued support. The authors declare no conflict of interest.
| f5ebc988b7a49c505786fffbfc45fa877b87322d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\label{sec:introduction}Introduction}
The exactly solvable Kitaev honeycomb model \cite{Kitaev06} has evoked considerable research interest in recent years, as it offers a distinct route to quantum spin liquids (QSLs). Realizing the model in real materials requires specific crystal structures and magnetic interactions \cite{Takagi19}. Former studies showed that bond-dependent Ising interactions, also known as Kitaev interactions, can arise from spin-orbit entangled pseudospin-1/2 degrees of freedom of $d^5$ transition metal ions situated in edge-shared octahedral crystal fields, and that a honeycomb lattice of such ions might approximate the Kitaev model \cite{Jackeli09}, as is believed to be the case for \ch{Na_2IrO_3} \cite{ChaloupkaPRL2010} and $\alpha$-\ch{RuCl_3} \cite{PlumbPRB2014}. More recent studies showed that $d^7$ ions Co$^{2+}$ with a high-spin $t_{2g}^5e_g^2$ configuration can also provide pseudospin-1/2 degrees of freedom with Kitaev interactions \cite{Liu18, Sano18}, as long as non-octahedral crystal fields are weak enough to leave the spin-orbit entanglement intact. But since Co has weaker spin-orbit coupling than Ir and Ru, this last hypothesis requires close scrutiny. Two layered cobaltates, \ch{Na_2Co_2TeO_6} and \ch{Na_3Co_2SbO_6}, in which oxygen ligands of Co$^{2+}$ form a nearly regular octahedron with small trigonal distortions \cite{ViciuJSSC2007}, have been proposed to potentially realize the Kitaev model \cite{Liu18,LiuPRL2020}. The Co-O sublattice of \ch{Na_2Co_2TeO_6} is displayed in Fig.~\ref{fig:1}(a).
\begin{figure}[b]
\includegraphics[width=0.5\textwidth]{Fig-1.pdf}
\caption{\label{fig:1} (a) Co-O layers of \ch{Na_2Co_2TeO_6}. Co are at the centers of the edge-sharing O octahedra, forming a honeycomb lattice (lower layer, where O are not shown and Co outside the dashed unit cell are faded). (b) Upper half: momentum slices of neutron diffraction raw data obtained at $T = 5$ K. Sharp magnetic diffractions are seen at the $M$-points of the 2D structural Brillouin zone (BZ), consisting of rod-like signals running along $L$ from the 2D order, and peaks at integer $L$ (in reciprocal lattice units, used throughout this work) from the 3D order. Lower half: 2D structural (black hexagons) and magnetic BZs under the zigzag (blue rectangle, for domains characterized by the blue $M$-point) and triple-$\mathbf{q}$ (purple hexagon) schemes, see text. Colored spheres are $M$-points of the BZ. Dashed line illustrates the cut displayed in Fig.~\ref{fig:3}(a).}
\end{figure}
\begin{figure*}[!ht]
\includegraphics[width=0.85\textwidth]{Fig-2.pdf}
\caption{\label{fig:2} (a) Neutron diffraction in the ($H$, 0, $L$) plane at different $T$. Intensities are plotted in false colors on a log scale. (b) $T$ dependence of signal at two $\mathbf{q}$ positions indicated by ``x'' in (a), characterizing 3D and 2D correlations (see text). (c) RXD in the ($H$, 0, $L$) plane. The signal is primarily due to magnetic scattering at $T < 24$~K, whereas the weak remaining intensity (note the different color scales) at $T=33$~K is due to a weak superstructure in the crystal lattice that likely originates from the Na layers (Fig.~S1 in \cite{SM}), which we consider unimportant here. (d) $T$ dependence of the RXD signals. Critical scattering (triangles) was measured in the same run by regions on the area detector surrounding the Bragg peak (circles) (Fig.~S2 in \cite{SM}). The temperature reading had not been calibrated at the sample position (see text).}
\end{figure*}
In most candidate Kitaev magnets, the systems develop long-range magnetic orders at low temperatures, instead of having a QSL ground state \cite{Takagi19}. Nevertheless, for understanding the materials and assessing their likelihood of realizing QSLs, such order may be instrumental because the symmetry of the order, as well as the associated magnon spectrum, may be used to determine the magnetic interactions \cite{BanerjeeScience2017,WinterNatCommun2017,RanPRL2017}. The information can be used to estimate whether and how the ground state can be tuned towards QSLs \cite{LiuPRL2020,WinterPRL2018}, and it provides the base for understanding exotic excitations \cite{BanerjeeScience2017,DoNatPhys2017}. The most commonly found long-range order is the so-called zigzag antiferromagnetic order, which is believed to be shared by the aforementioned Ir-, Ru-, and Co-based compounds \cite{LiuPRB2011,FengYe12, Choi12,Johnson15, Sears15,Lefrancois16, Bera17,WongJSSC2016,YanPRM2019}. Historically, observation of zigzag order had been considered promising for finding a nearby QSL phase, because the two phases are adjacent in the parameter space of the Kitaev-Heisenberg model \cite{ChaloupkaPRL2013}. However, this notion only applies to antiferromagnetic Kitaev interactions ($K>0$), yet all of the above systems are considered to have $K<0$ \cite{KatukuriNJP2014,YamajiPRL2014,SizyukPRB2014,HuPRL2015,
WinterPRB2016,KimPRB2016,RanPRL2017,Liu18,Sano18}, even though the Co-based cases are less certain according to recent studies \cite{Songvilay20,Lin20,Kim20}. When additional interactions (further-neighbor ones in particular) are included to rationalize the observation of zigzag order, theoretical phase diagrams become complicated \cite{KatukuriNJP2014,RauPRL2014,SizyukPRB2014,WinterPRB2016,KimchiPRB2011}, and it depends on model details whether a QSL phase remains accessible nearby. In this regard, the Co-based systems are likely dominated by nearest-neighbor interactions because their $3d$ orbitals are more localized than $4d$ and $5d$ \cite{LiuPRL2020}. While this may help simplify theories, it also raises the question about the stability of the zigzag order, which becomes particularly important for shaping our understanding of these $3d$ systems.
Here we present a systematic study of high-quality single crystals \cite{XiaoCGD2019,Yao20} of \ch{Na_2Co_2TeO_6}, bringing insight to the above two fronts. We show that the system's first thermal transition, out of the paramagnetic phase, is towards a hitherto unnoticed two-dimensional (2D) long-range magnetic order, which precedes the formation of the 3D order that is commonly considered of zigzag nature. As the Mermin-Wagner theorem precludes spontaneous breaking of continuous symmetry at finite temperature in 2D systems, the 2D order must hence break discrete symmetry, which would be natural if the order is driven by strongly anisotropic interactions such as $K$. We furthermore find a surprising result about the 3D order: magnons deeply in the ordered state are \textit{not} described by Brillouin zones (BZs) of zigzag order; instead, they seem to arise from a triple-$\mathbf{q}$ ground state, which may be indistinguishable from zigzag order in diffraction experiments performed on powder \cite{Lefrancois16, Bera17} or multi-domain samples. We see experimental signs that the 3D order is strongly frustrated upon its initial formation, and that it affects the high-energy orbital degrees of freedom. These results indicate that spin-orbit physics and frustration are at the heart of magnetism in \ch{Na_2Co_2TeO_6}.
\begin{figure*}[!ht]
\includegraphics[width=0.85\textwidth]{Fig-3.pdf}
\caption{\label{fig:3} (a) INS measurement of low-energy magnons (intensities plotted in false colors). Data at different $L$ have been combined, given the lack of dispersion along $L$ (inset). (b) Energy distribution of INS intensities at different $T$, measured at $\mathbf{q}_{\parallel}=$ (0.5, 0) (upper panel) and (0.75, 0) (lower panel). Data below and above 5 meV were obtained with incident neutron energies of 6.1 and 10.0~meV, respectively. (c) Total energy carried by magnons (estimated by weight-integrating INS intensities over [0.5, 4.5] meV), and its $T$-derivative as an estimate of the associated heat capacity, see \cite{SM} for detail. (d) High-resolution heat capacity measurement on a single crystal. Similar results have been obtained in separate runs and for two different crystals. The lower panel zooms into the part near 31 K. (e) $^{23}$Na NMR spin-lattice relaxation rate as a function of $T$, measured with $\mu_0 H = 0.75$ T $\parallel \mathbf{a^*}$. In (c-e), the dashed and dash-dotted lines indicate $T = 30.97$~K and 26.7~K, respectively.}
\end{figure*}
Figure~\ref{fig:1}(b) displays an overview of our neutron diffraction \cite{SM} result in momentum ($\mathbf{q}$) space. Long-range magnetic correlations manifest themselves as sharp peaks in the 2D ($H$, $K$) plane, at the $M$-points of the structural BZ. In the view of zigzag order \cite{Lefrancois16, Bera17}, the six peaks come from domains in which the orientations of zigzag chains differ by 120$^\circ$, as each domain contributes signal at a pair of opposite $\mathbf{q}$. Later we will show that a more suitable view of the data is that the order is of triple-$\mathbf{q}$ nature, producing diffraction signals at all the $M$-points. Turning our attention to the 3D structure of the diffraction, it is seen that the signal has a nearly $L$-independent part (``rods'' of scattering running along $L$), plus extra peaks at integer $L$. This result is presented in detail in Fig.~\ref{fig:2}(a-b), which also shows the temperature ($T$) dependence. While the integer-$L$ signal (at $H=0.5$) is much greater in amplitude at $T=5$~K, the total intensity of the rod-like signal is weaker but within an order of magnitude comparable to it. Thus there is a coexistence of 2D and 3D magnetic orders at $T=5$~K. As $T$ increases, the 3D integer-$L$ signal disappears above 26.7 K, yet the 2D rod-like signal clearly survives up to higher $T$. At 28~K we only see the 2D order [Fig.~\ref{fig:2}(a)]. We consider the (lack of) inter-layer correlation in the 2D order not merely limited by stacking faults \cite{Johnson15} manifested by the diffractions at $H=1$, since the latter are actually more structured in $L$ and with nearly no dependence on $T$.
The 2D signal [at $\mathbf{q_\mathrm{b}}$ in Fig.~\ref{fig:2}(b)] has rich structures in its $T$ dependence: (1) a local maximum at 26.7 K, presumably due to critical correlations associated with the 3D transition; (2) a slight decrease below 18~K, possibly due to subtle changes in the magnetic structure \cite{ViciuJSSC2007,Yao20,XiaoCGD2019}, which we will revisit later; (3) a smeared transition around 30~K. The smearing could be due to the limited $\mathbf{q}$ and energy resolution of neutron scattering on our large array of crystals. Figure~\ref{fig:2}(c-d) displays our resonant X-ray diffraction (RXD) \cite{SM} results from a small crystal, which beautifully complements the neutron diffraction result. An intense signal, with resolution-limited width in $H$ and perfect rod-like shape along $L$, sharply sets in below nominal $T=30$ K (the sample's actual temperature may be slightly higher, due to limited thermal shielding). Critical scattering is observed as a peak at the same $T$ near the 2D ordering $\mathbf{q}$, which unequivocally shows that the transition is of 2nd-order nature and marks spontaneous symmetry breaking. This result is significant, because if a 2D system undergoes spontaneous symmetry breaking at finite temperature, the broken symmetry must be discrete according to the Mermin-Wagner theorem. For our magnetic system, it strongly suggests that the underlying model is not Heisenberg-like or anisotropic only with a global easy-plane \cite{XiaoCGD2019}, and would be consistent with the prominence of Kitaev interactions in driving the transition.
We now turn to presenting a few aspects about the excitation spectra, which shed light on the nature of the orders and represent another main finding of this work. Analysis of the full excitation spectra will be reported elsewhere. We begin by showing low-energy magnons measured by inelastic neutron scattering (INS) \cite{SM}, deep in the 3D ordered state at $T=5$~K, along an in-plane momentum cut connecting two neighboring structural BZ centers [Fig.~\ref{fig:3}(a), the lack of dispersion along $L$ (inset) shows negligible inter-layer coupling]. As illustrated in Fig.~\ref{fig:1}(b), the cut passes through an $M$-point at (0.5,~0). Had the magnetic order been zigzag, this $M$-point would be a magnetic BZ center for one of the domains [``blue'' in Fig.~\ref{fig:1}(b)]; meanwhile, it would be a BZ \textit{corner} for the other two domains. Since (0,~0) and (1,~0) are always BZ centers regardless of domain, and because our coaligned sample must have (statistically) equal domain population, the chosen cut should manifest \textit{two inequivalent} magnon dispersions superposed on each other, but the data in Fig.~\ref{fig:3}(a) clearly disprove this. The data instead show a single magnon branch without any sign of domain superposition. Given the previous neutron diffraction evidence for zigzag order \cite{Lefrancois16, Bera17}, we interpret our result as strong evidence for a triple-$\mathbf{q}$ order, the BZ of which [Fig.~\ref{fig:1}(b)] would be fully consistent with Fig.~\ref{fig:3}(a). Indeed, a triple-$\mathbf{q}$ order can be understood as a vector sum of three 120$^\circ$-different zigzag orders (Fig.~S3 in \cite{SM}); because each zigzag order, when existing alone, would produce diffraction signals at a different $M$-point, diffraction from the triple-$\mathbf{q}$ order would be \textit{in principle} indistinguishable from the zigzag order with equal domain population. Additional indirect evidence against interpreting the order as zigzag can be found in \cite{SM}.
In recent studies \cite{Songvilay20,Lin20,Kim20} of \ch{Na_2Co_2TeO_6} powder, INS spectra of magnons have been simulated based on an assumed zigzag ground state. In light of our results, conclusions based on those analyses might need reevaluation. Our INS spectra are consistent with those of \cite{Songvilay20,Lin20,Kim20}: we see a gap below $\sim1$ meV, a lowest magnon branch up to $\sim3$ meV, and additional branches at higher energies. Figure~\ref{fig:3}(b) presents spectra at selected $\mathbf{q}$ and their $T$ dependence. We find that the high-energy branches start to form below 26.7~K, as a pile-up of intensities above 6 meV is seen at $T = 21$ K (but not at 28 K), meanwhile the low-energy intensities are depleted. Using $\mathbf{q}$-integrated INS intensities as a measure of the thermally populated boson number, we can estimate the magnons' total energy, the $T$-derivative of which provides an estimate of their heat capacity [Fig.~\ref{fig:3}(c)]. The calculated heat capacity shows a prominent peak at 26.7 K, consistent with direct measurement results [Fig.~\ref{fig:3}(d)].
It therefore seems that the order below 26.7 K satisfies a considerable portion of the leading interactions, whereas the 2D order does not. This result, in conjunction with the lack of inter-layer coupling [Fig.~\ref{fig:3}(a) inset], indicates that the 2D order is \textit{not} simply the 3D order without inter-layer correlation, and must be something fundamentally different. This understanding is consistent with our $^{23}$Na nuclear magnetic resonant (NMR) experiment \cite{SM}, which shows a pronounced peak in the spin-lattice relaxation rate $1/T_1$ at 26.7~K [Fig.~\ref{fig:3}(e)], but no pronounced anomaly upon the formation of the 2D order. The lack of 2D order's clear signature in $1/T_1$ might be because the order is not strictly static, such that the slow dynamics wipes out its NMR signal (Fig.~S4 in \cite{SM}). We emphasize that the 2D ordering is still a genuine phase transition, as supported both by the RXD data in Fig.~\ref{fig:2}(d) and by our high-resolution heat capacity measurement \cite{SM}, which does show a tiny yet well-defined peak at $T=30.97$~K [zoom-in panel of Fig.~\ref{fig:3}(d)].
\begin{figure}[!ht]
\includegraphics[width=0.5\textwidth]{Fig-4.pdf}
\caption{\label{fig:4} (a) High-energy Raman spectra at different $T$ ($T$ in log scale), normalized to the peak at 119 meV. (b) $T$ dependence of area ratio between the two peaks at 136 and 130 meV (data and fits are presented in Fig.~S7 of \cite{SM}). (c) Summary of the thermal phases. A: paramagnetic; B: 2D order, no clear magnon branches; C: 2D and 3D orders, gapless excitations and magnons; D: 2D and 3D orders, gapped magnons.}
\end{figure}
The 3D order's impact goes well beyond the magnons. At much higher energies, probed by Raman spectroscopy \cite{SM} above 100 meV (Fig.~\ref{fig:4}), we still find a pronounced spectral change related to the order's formation. The high energies cannot be attributed to the pseudospin-1/2 magnetism, and although they could be related to multi-phonon scattering, we do not find any magnetoelastic effects across the 3D transition in the single-phonon spectra (Fig.~S5 in \cite{SM}). We therefore attribute the features to spin-orbit excitons \cite{SartePRB2019} and/or multi-phonon scattering assisted by electronic excitations. In either case, the excitations likely involve the orbital degrees of freedom, which further signifies the role of spin-orbit physics in driving the 3D order. In the opposite energy limit, in contrast, the 3D order is unable to immediately open up the 1 meV gap at the $M$-point [Fig.~\ref{fig:3}(b)], as the first clear indication of such a gap is not seen until $T=14$~K (detailed $T$ dependence is presented in Fig.~S6 of \cite{SM}). The persistence of gapless excitations, down to the subtle transition temperature 15-18 K shown in Fig.~\ref{fig:2}(b) and in Refs. \cite{ViciuJSSC2007,Yao20,XiaoCGD2019,Lin20}, suggests that the 3D order manifests substantial frustration upon its initial formation, until the system is cooled much more deeply to freeze out the nearly degenerate configurations. Our view of the thermal phases is summarized in Fig.~\ref{fig:4}(c).
On a final note, we discuss possible implications of our discovery of the 3D triple-$\mathbf{q}$ order in conjunction with the other phases. Similar to other multiple-$\mathbf{q}$ magnetic orders \cite{ChristensenPRB2015,GastiasoroPRB2015}, we expect the 3D triple-$\mathbf{q}$ order here to give rise to a charge (orbital) order (Fig.~S3 in \cite{SM}) that might exist on its own as a ``vestigial'' order \cite{FernandesPRB2016}. An intriguing possibility is that the 2D order is of this type -- the spin-orbit entangled pseudospin-1/2 degree of freedom might further enrich such orders by intermixing the magnetic and charge order parameters. Moreover, the triple-$\mathbf{q}$ order, formed by the superposition of three zigzag components, possesses non-zero pseudospin vorticity (Fig.~S3 in \cite{SM}) that is shared by previously proposed ``vortex'' \cite{ChaloupkaPRB2015} and ``cubic'' \cite{RousochatzakisPRL2017} orders, albeit the latter orders are characterized by the $K$- rather than the $M$-points of the BZ. We are not aware of previous reports of the triple-$\mathbf{q}$ order that we propose. The vorticity might leave distinct signatures in experiments sensitive to time-reversal symmetry breaking, \textit{e.g.}, optical Kerr rotation. In fact, the emergent vorticity may couple to external magnetic fields in a fashion similar to ferro- and/or ferrimagnetism, which has been observed \cite{Yao20}. The self-organization of vorticity might be related to the transition near 15-18 K, hence explaining the freezing of the gapless excitations below this temperature.
To summarize, our work calls for a revision of the magnetic ground state of \ch{Na_2Co_2TeO_6}, highlighting the importance of triple-$\mathbf{q}$ rather than zigzag order for navigating our determination of the microscopic model. We present systematic evidence that spin-orbit physics, highly anisotropic interactions, and magnetic frustrations are prominent characteristics of the model, which are in line with proposals related to realizing the Kitaev model. The system clearly manifests intriguingly rich physics.
\begin{acknowledgments}
We are grateful for discussions with Gang Chen, Rafael Fernandes, Giniyat Khaliullin, Haijun Liao, Yingying Peng, Natalia B. Perkins, Yiming Qiu, Kaiwei Sun, Fa Wang, Liusuo Wu, Tao Xiang, and Yang Zhao. We thank Weiliang Yao for his assistance with sample preparation and the neutron scattering experiment, which was performed at the MLF, J-PARC, Japan, under a user programme (proposal No. 2019B0062). Part of the research described in this work was performed at the Canadian Light Source, a National Research Facility of University of Saskatchewan, which is supported by the Canada Foundation for Innovation (CFI), the Natural Sciences and Engineering Research Council (NSERC), the National Research Council (NRC), the Canadian Institutes of Health Research (CIHR), the Government of Saskatchewan, and the University of Saskatchewan. Y.L. acknowledges support by the NSF of China (Grant No. 11888101) and the National Basic Research Program of China (Grant No. 2018YFA0305602). X.L. acknowledges support by the NSF of China (Grant No. 11921005) and the Strategic Priority Research Program of Chinese Academy of Sciences (Grant No. XDB28000000). W.Y. acknowledges support by the Ministry of Science and Technology of China (Grant No. 2016YFA0300504) and the NSF of China (Grant No.~51872328).
\end{acknowledgments}
| b26bab5b13c8aec3b0d52d87716ad9b5e482f769 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the IR traditional setting, a user expresses his information needs with the help of a query. Yet, it is sometimes difficult to match the query with the documents, for instance because of the query vocabulary may differ from the documents. Especially when the query is short, the performance of the system is usually poor, as it is difficult to detect the precise focus of the information need, and the relative importance of the query terms.
Query expansion aims at tackling these problems by transforming the short query into a larger text (or set of words) that make it easier to match documents from the collection.
The main difficulty of query expansion is obviously to add only relevant terms to the initial query. Several techniques have been proposed in the literature, based on linguistic resources (e.g. synonym lists) or based on the documents themselves (e.g. pseudo-relevance feedback).
In this paper, we explore the use of recent text generation models to expand queries. We experimentally demonstrate that the recent advances in neural generation can dramatically improve ad-hoc retrieval, even when dealing with specialized domains. More precisely, through different experiments, we show that:
\begin{enumerate}
\item texts artificially generated from the query can be used for query expansion;
\item this approach does not only provide new terms to the query, but also a better estimate of their relative weights;
\item in addition, it also provides a better estimate of the importance (i.e. weight) of original query words;
\item this approach can also be used on specialized domains.
\end{enumerate}
The paper is structured as follows. After a presentation of the related work, Section~\ref{sec:approach} details the different components of our approach. Several experiments are then detailed in Section~\ref{sec:expes}. Last, some concluding remarks are given in Section~\ref{sec:concl}.
\section{Related work}
Query expansion is a well-established technique to try to improve the performance of an IR system. Adding new terms to the query is expected to specifically improve recall, yet, since the query is, hopefully, better formulated, it may also improve the top rank results and be beneficial to precision.
One might classify the existing automatic approaches based on the origins of resource used to expand the query.
\paragraph{External resources.}
One obvious way to expand a query is to add semantically related terms to it (synonyms or sharing other semantic relations like hyponyms, quasi-synonyms, meronyms...). Existing lexical resources can be used to add, for each query term, a list of semantically related terms; yet, one has to deal with different problems: existence of lexical resources for the collection language, or for the specific domain of the collection, choice of the appropriateness of certain relations, need of sense disambiguation for polysemous words...
WordNet~\cite{miller-ijl-90} is among the best-known resources for English (general language) and have been used with mitigated results at first \cite{Voorhees94}, but later shown to be effective \cite[inter alia]{Claveau-COLING2016}.
\paragraph{Collection-based resources.}
Distributional thesauri have also been exploited to enrich queries. Since they can be built from the document collection (or from a large corpus with similar characteristics), they are suited to the domain, the vocabulary...
Traditional techniques to build these thesauri have obtained good results for query expansion \cite{Claveau-COLING2016}.
Neural approaches, that is, word embedding approaches are now widely used to build such semantic resources.
In the recent years, static embeddings (word2vec \cite{Mikolov13}, Glove \cite{pennington2014glove} or FastText \cite{bojanowski2016enriching} to name a few) were also used in IR, in particular to enrich the query. Indeed, these trainable dense representations make it easy to find new words that are semantically close to query words.
Even more recently, dynamic word representations obtained with transformer-based architectures, such as BERT \cite{devlin2019bert} or GPT \cite{Radford2019}, have been proposed. They build a representation for each word according to its context, and this ability have been exploited to obtain competitive results in IR tasks \cite[inter alia]{Dai_2019,Khattab2020}.
Bert has been also used for query expansion in the framework of a neural IR system based on reranking \cite{Padaki2020}
\paragraph{Pseudo-relevance feedback}
A last category of studies considers only a small set of document to help to expand or reformulate the query. To be automatic, they replace the user feedback by the hypothesis that the best ranked documents retrieved with the original query are relevant and may contain useful semantic information \cite{Ruthven03}. It is interesting to note that in this case, not only semantically relevant terms are extracted, but also distributional/statistical information on them and on the original query terms.
In this category, Rocchio, developed in the 60's for vector space model was among the first one popularized \cite{Manning_IR2008}. One of the current best known approach is RM3, which was developed in the framework of language model based IR systems \cite{Abdul-RM3}. It is often reported to yield the best results in ad-hoc retrieval tasks, even compared with recent neural models \cite{NeuralHype}.
Neural approaches have also been proposed to integrate pseudo-relevance feedback information \cite{Li-NPRF-EMNLP2018}, yet, as it is reported by the authors, the results are still lower than traditional models with query expansion.
~\newline
In this paper, we propose to use constrained text generation to expand queries. In this approach, the original query is used as a seed for a generative model which will output texts that are, hopefully, related to the query.
While text generation with language model is not new, the performance of neural models based on transformers \cite{Vaswani_NIPS2017} makes this task realistic. In this paper, we use the Generative Pre-Trained Transformers (GPT). They are built from stacked transformers (precisely, decoders) trained on a large corpus by auto-regression, i.e. unsupervisedly trained to predict the next token given the previous ones.
The second version, GPT-2 \cite{Radford2019}, contains 1.5 billion parameters for its largest pre-trained model, trained on more than 8M documents from Reddit (i.e. mostly English and general language such as discussion on press articles).
A newer version, GPT-3, has been released in July 2020; it is much more larger (175 billion parameters) and outperforms GPT-2 on any tested task. Yet, the experiments reported below were done before this release, thus with GPT-2.
\section{Generated query expansion}
\label{sec:approach}
\subsection{Overview of our approach}
As it was previously explained, our approach is very simple as it relies on existing tools and techniques.
From a query, multiple texts are generated by a GPT-2 model using the query as the seed. Note that the generation process is not deterministic, and thus, even with the same seed, the texts are different. In the experiments reported below, 100 texts per query are generated (unless specified otherwise).
These texts are concatenated and considered as the new query.
In our experiments, this new, very large, query is then fed to a simple BM25+ IR system, but it could obviously be used in any other IR system.
An example of a text generated from a query is presented in Fig.~\ref{fig:example_701}. As one can see, the generated text, while completely invented (note the barrel prices), is relevant for the query. It contains many terms, absent from the original query, that are more or less closely related to the information need, such as orthographic variants (United States vs. U.S.), meronyms-metonyms (barrel vs. oil), hypernyms (energy vs. oil) and more generally any paradigmatic (consumer, producer vs. industry) or syntagmatic (production for oil) relations.
It is worth noting that such texts also give a valuable information about the relative frequency of each terms (contrary to thesauri or embeddings).
\begin{figure}
\noindent\fbox{%
\parbox{\textwidth}{%
U.S. oil production has been declining steadily for decades and it is not expected to reverse. In fact, some argue that it may even get worse. The long-term trend is for oil production to decline at a rate of about 1 percent per year. With production of about 8 million barrels per day now, there is no way the United States can replace its current output.
The U.S. oil boom was a result of an energy revolution in the 1970s that led to increased oil production, and a significant change in the global oil market. The U.S. now produces about 2.3 million barrels of oil per day, the highest it has been in over 30 years.
The United States is now the world's largest oil producer and the fourth largest oil exporter.
What happened?
When oil prices peaked in the 1970s, the United States was the world's largest oil producer. But over the next several decades, the United States' oil production began to decline. The decline was most pronounced in the 1980s, when the United States began to fall behind other oil producing countries.
The oil price decline in the 1970s was not entirely voluntary. The United States was producing less oil and exporting more oil than it was consuming. The Federal Reserve controlled the amount of dollars in the Federal Reserve's reserves, so the United States was not exporting as much oil as it was producing. The decline in U.S. oil production was a result of the declining price of oil.
The price of oil had declined from \$8 per barrel in 1973 to \$2.50 per barrel in 1977. In 1979, the price of oil reached a high of \$15.75 per barrel. By 1983, the price of oil had fallen to \$4.65 per barrel. By 1986, the price of oil had fallen to \$1.86 per barrel. By the end of the 1980s, the price of oil had fallen to \$1.24 per barrel.
The decline in oil prices was a direct result of the energy revolution in the 1970s. The United States was the world's largest oil producer, but the United States was also the world's largest consumer of oil. When oil prices fell, so did the cost of producing oil. %
}}
\caption{Example of a document generated with the pre-trained GPT-2 large model from the text seed "U.S. oil industry history" (query 701 from .GOV collection)}
\label{fig:example_701}
\end{figure}
\subsection{Pre-trained models, fine-tuning and parameters}
\label{sec:fine_tuning}
GPT-2 comes with several pre-trained models, having different size in terms of parameters (from 124M to 1.5B). As it was previously said, their training data was news-oriented general language. The largest model was used for two of the tested collections (see below).
While these all-purpose models are fine for IR collections whose documents are also general language, it may not be appropriate for domain-specific IR collections. In the experiment reported in the next section, we use the \textsc{ohsumed} collection, made of medical documents.
For this collection, we have fine-tuned the GPT-2 355M model on the documents of the collection in order to adapt the language model to the specific medical syntax and vocabulary. The fine-tuning was stopped after 250,000 samples were processed (this number of sample process indirectly controls the under/over-fitting to the specialized corpus) was set to and other parameters (batch size, optimizer, learning rate...) let to their defaults. Although a larger set of medical documents could be used (from Pubmed for instance), this small fine-tuned model is expected to be more suited to generate useful documents to enrich the query.
Concerning the generation of documents, for reproducibity purposes, here are the main GPT-2 parameters used (please refer to GPT-2 documentation\footnote{\url{https://github.com/openai/gpt-2}}):
length=512, temperature=0.5, top\_p=0.95, top\_k = 40.
\subsection{IR Systems}
In the experiments reported in the next section, we use two IR models.
The first one is BM25+ \cite{Lv2011}, a variant of BM25 \cite{RWB98}. The parameters $k_1$, $k_3$, $b$ and $\delta$ were kept at their default value (resp. 1.2, 1000, 0.75, 1). It is implemented as a custom modification of the \textsc{gensim} toolkit \cite{Rehurek_GENSIM}.
The second IR model is Language modeling with Dirichlet smoothing \cite{Zhai2001} as implemented in Indri \cite{MetzlerCroft2004,StrohmanMetzlerTurtleCroft2005}. The smoothing parameter $\mu$ is set to 2\,500.
Both models are regarded as yielding state-of-the-art performance for bag-of-words representation \cite{NeuralHype}.
Their RSV function can be written:
$$RSV(q,d) =\sum_{t \in q} w_q(t) \cdot w_d(t) $$
with $w_q(t)$ the weight of term $t$ in query $q$ and $w_d(t)$ the weight in document $d$, as illustrated in Tab.~\ref{tab:weight} (from \cite{Lv2011}).
\begin{table}
\begin{center}
\begin{tabular}{l|cc}
IR model & weighting\\
\hline
BM25+ $w_d(t)$ & $\left( \frac{(k1+1)c(t,d)}{k1(1-b+b\cdot dl(d)/avdl)+c(t,d)}+\delta \right) \cdot \log \frac{N+1}{df(t)}$\\
BM25+ $w_q(t)$ & $\frac{(k3+1)c(t,q)}{k3+c(t,q)}$ \\
& with $k1, k_3, b$ and $\delta$ fixed parameters \\
& \\
LM $w_d(t)$ & $\log \left( \frac{\mu}{dl(d)+\mu} + \frac{c(t,d)}{(dl(d)+\mu)p(t|C)} \right)$\\
LM $w_q(t)$ & $c(t,q)$ \\
& $\mu>0$ a smoothing parameter\\
&\\
\end{tabular}
\end{center}
\caption{\label{tab:weight}IR models (weighting functions of terms in the query and the document) for BM25+ \protect\cite{RWB98,Lv2011} and Language modeling with Dirichlet smoothing LM \protect\cite{Zhai2001}}
\end{table}
For RM3 expansion, we also rely on the Indri implementation; the results reported in the next section corresponds to the best performing parameters tested for each collection (number of documents considered for pseudo relevance feedback and number of terms kept).
\section{Experiments}
\label{sec:expes}
\subsection{Experimental settings}
Three IR collections are used in our experiments: Tipster, GOV2 and \textsc{ohsumed}. Some basic statistics are given in Tab.~\ref{tab:collection}.
\begin{table}[tb]
\begin{center}
\begin{tabular}{l|ccc}
& Tipster & GOV2 &\textsc{ohsumed} \\
\hline
nb of documents & 170,000 & 25M & 350,000 \\
nb of queries & 50 & 150 & 106 \\
avg size of queries & 6.74 & 3.15 & 7.24\\
language & En & En & En \\
avg nb of relevant doc per query & 849 & 179 & 21 \\
\end{tabular}
\end{center}
\caption{\label{tab:collection}Statistics on the IR collections used}
\end{table}
Tipster was used in TREC-2. The documents are articles from newspaper, patents and specialized press (computer related) in English. The queries are composed of several parts, including the query itself and a narrative detailing the relevance criteria; in the experiments reported below, only the actual query part is used.
GOV2 is a large collection of Web pages crawled from the .gov domain and used in several TREC tracks. In the experiments reported below, 150 queries from TREC 2004-2006 ad-hoc retrieval tasks are used; as for Tipster, only the actual query part is used (i.e. description and narrative fields are not included in the query).
\textsc{Ohsumed} \cite{Ohsumed94} contains bibliographical notices from Medline and queries from the TREC-9 filtering task. Its interest for our experiments is that it deals with a specialized domain, hence it contains a specific vocabulary.
Performance are assessed with standard scores: Precision at different thresholds (P@x), R-precision (R-prec), \textit{Mean Average Precision} (MAP). When needed, a paired t-test with $p=0.05$ is performed to assess the statistical significance of the difference between systems.
\subsection{General language}
Tables~\ref{tab:res-tipster} and \ref{tab:res-gov2} respectively present the results for the general-language collections Tipster and GOV2.
For comparison purposes, we indicate the results of BM25+ without expansion, Indri's Language Model (LM) with and without RM3 expansion. The best performing setting for RM3 on Tipster is 100 terms from the top 20 documents, and 100 terms for the top 10 documents GOV2. The statistical significance is computed by comparing with the LM+RM3 baseline.
\begin{table}[tb]
\begin{center}
\begin{tabular}{l|ccccccc}
& MAP & R-Prec & P@5 & P@10 & P@20 & P@100 \\
\hline
BM25+ & 25.06 & 32.16 & 95.60 & 92.60 & 89.70 & 73.64 \\
LM & 24.48 & 31.48 & 92.40 & 89.00 & 85.40 & 70.70 \\
LM + RM3 & 31.01 & 36.38 & 94.40 & 93.20 & 90.60 & 81.22 \\
\hline
BM25+ and expansion & \textbf{35.22}$^*$ & \textbf{39.87}$^*$ & \textbf{99.60}$^*$ & \textbf{98.40}$^*$ & \textbf{98.20}$^*$ & \textbf{87.84}$^*$ \\
\end{tabular}
\end{center}
\caption{\label{tab:res-tipster}Performance (\%) on Tipster with query expansion; best results in bold, statistical significance over LM+RM3 noted with $*$}
\end{table}
\begin{table}[tb]
\begin{center}
\begin{tabular}{l|ccccccc}
& MAP & R-Prec & P@5 & P@10 & P@20 & P@100 \\
\hline
BM25+ & 25.66 & 31.25 & 52.92 & 49.97 & 46.52 & 34.63 \\
LM & 27.96 & 33.01 & 56.08 & 55.20 & 51.59 & 37.32 \\
LM with RM3 & 30.22 & 34.20 & 55.00 & 56.08 & 53.67 & \textbf{45.86} \\
\hline
BM25+ and expansion & \textbf{34.54}$^*$ & \textbf{37.76} & \textbf{67.91}$^*$ & \textbf{63.88}$^*$ & \textbf{57.94} & 44.30 \\
\end{tabular}
\end{center}
\caption{\label{tab:res-gov2}Performance (\%) on GOV2 with query expansion; best results in bold, statistical significance over LM+RM3 noted with $*$}
\end{table}
On both collections, and on every performance measure, expanding the queries with the generated texts brings important gains compared with the system without expansion. Also, our approach outperforms RM3 expansion in almost every situation, and with a large margin on MAP, R-prec and precision on the top-ranked documents (P@5, P@10).
\subsection{Specialized language}
The same setting is used on the \textsc{ohsumed} collection. For these medical-oriented IR dataset, we report two versions of our approach: one is using the pre-trained model as before, and one relies on a model fine-tuned on the documents of the collection. The best performing setting for RM3 is 80 terms for the top 10 documents. The results are reported in Tab.~\ref{tab:res-ohsumed}.
\begin{table}[tb]
\begin{center}
\begin{tabular}{l|ccccccc}
& MAP & R-Prec & P@5 & P@10 & P@20 & P@100 \\
\hline
BM25+ & 18.27 & 19.94 & 31.88 & 26.04 & 20.50 & 9.48 \\
LM & 17.61 & 20.35 & 29.31 & 24.06 & 19.21 & 9.28 \\
LM + RM3 & 20.80 & 22.54 & 30.89 & 26.83 & 22.18 & 10.51 \\
\hline
BM25+ and expansion (no fine-tuning) & 21.60 & 23.75 & 33.47$^*$ & 27.62 & 22.92 & 11.16\\
BM25+ and expansion (fine-tuning) & \textbf{23.07}$^*$ & \textbf{24.65}$^*$ & \textbf{34.65}$^*$ & \textbf{29.41}$^*$ & \textbf{24.31} & \textbf{11.42} \\
\end{tabular}
\end{center}
\caption{\label{tab:res-ohsumed}Performance (\%) on \textsc{ohsumed} with query expansion; best results in bold, statistical significance over LM+RM3 noted with $*$}
\end{table}
Here again, the GPT-based expansion significantly improves the results of the IR system and outperforms RM3 expansion.
Yet, the gains are lower than for the two previous collections. This difference can be explained by the following factors:
\begin{enumerate}
\item the queries are longer more complex and more specific (as can be seen in Tab.~\ref{tab:collection}, few documents are relevant);
\item the generation model is not sufficiently suited to the documents.
\end{enumerate}
Concerning this latter reason, we can indeed see the interest of fine-tuning the generation model, but better results may be obtained by using a larger set of medical documents, or adopting different fine-tuning parameters (in particular the number of epoch/samples processed, see Sect.~\ref{sec:fine_tuning}). Unfortunately, defining a priori the best parameters for our IR task is not possible and the cost of the fine-tuning process makes it impossible to test a wide range of possible values.
\subsection{Query expansion and weight information}
One of the interest of having complete texts that are generated is that we can collect information on the relative importance of words, to the contrary of expanding queries with a thesaurus.
To observe the impact of the number of occurrences in the generated texts, we evaluate the effect of keeping the $k$ most frequent terms of the generated texts and either weighting them by their frequency (as done usually by BM25) or by giving a fixed weight ($1/k$). The results for different values of $k$ are presented in Fig.~\ref{fig:exp_size}.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{MAP_vs_expansion_size.png}
\caption{MAP (\%) according to size of the expansion, terms with fixed weight or weight depending on their frequency in the generated documents; Tipster collection}
\label{fig:exp_size}
\end{figure}
On can observe that adding terms to the query with a fixed weight slightly improves the MAP, but most of the gain is indeed brought by a proper weighting based on the frequency of the term in the generated documents.
In the next experiment, we examine how the generated texts can help to re-weight the initial query terms; there is no query expansion, since only the initial query terms are kept, but their frequency in the generated texts are used in the BM25+ $w_q$. The results reported in Tab.~\ref{tab:res-reweighting} show that there is indeed a small improvement of the MAP, that is more noticeable at high DCV.
\begin{table}[tb]
\begin{center}
\begin{tabular}{l|ccccccc}
& MAP & R-Prec & P@5 & P@10 & P@20 & P@100 \\
\hline
BM25+ & 25.06 & 32.16 & 95.60 & 92.60 & 89.70 & 73.64 \\
\hline
BM25+ with re-weighting & 27.12 & 33.22 & 96.80 & 94.20 & 92.70 & 77.12 \\
\end{tabular}
\end{center}
\caption{\label{tab:res-reweighting}Performance (\%) on Tipster with re-weighting query words}
\end{table}
These two experiments demonstrate the usefulness of dealing with full texts and not only word-to-word similarity.
\subsection{Number of generated documents}
Since text generation can be costly, it is interesting to see how many generated texts are necessary. In Fig.~\ref{fig:exp_nb_doc}, the MAP obtained for up to 100 documents is presented. One can observe that a plateau is rapidly reached (at around 20 documents per query). Of course, the size of the generated documents (can be set as a parameter of the generation process) is also to be considered.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{MAP_vs_nb_doc.png}
\caption{MAP (\%) according to the number of generated documents (average over 5 runs); Tipster collection}
\label{fig:exp_nb_doc}
\end{figure}
\section{Conclusive remarks and foreseen work}
\label{sec:concl}
Neural approaches are increasingly used in IR, with mitigated results, especially when compared with "traditional" bag-of-word approaches \cite{NeuralHype,NeuralHype2}. Here, the neural part is successfully used outside of a "traditional" IR system (but note that it could be used with any IR systems, since it simply enriches the query).
The expansion approach presented in this paper is simple and easy to implement (thanks to the availability of the GPT models and code) while offering impressive gains.
Lot of parameters could be further optimized, especially on the GPT model side (to influence the "creativity" of the text generation), and the fine-tuning capabilities should also be explored more thoroughly (influence of bigger specialized corpus if available, precise mix between pre-trained and fine-tuning, etc.).
The recent availability of GPT-3\footnote{\url{https://github.com/openai/gpt-3}} makes it possible to even get greater gains thanks to the alleged high quality of its outputs.
This whole approach also offers many research avenues: in this work, we have used text generation as a way to perform data augmentation on the query side, but it could also be used to augment the representation of the documents (even if in practice, the cost is still prohibitive on large collection).
All machine-learning (neural or not) approaches based on pseudo-relevance feedback to train their model could instead use similar text generation with the advantage that they would not be limited by the number of potential relevant documents in the shortlist.
And of course, similar data-augmentation strategy could be used for other tasks than document retrieval.
More fundamentally, the recent improvements of text generation also question the relevance of the document retrieval task. Indeed, it is possible to envision systems that will be able to generate one unique document answering the user's information need, similarly to question-answering. If the generative model is trained on the document collection, the generated document will serve as a summary (which is one of the popular applications of GPT-x models) of the relevant documents.
Yet, the current limitations of the models tested in this paper make them far from being suited for this ultimate task: the generated documents do deal with the subject of the query, and thus use a relevant vocabulary, but do not provide accurate, factual information (as seen in the Example in Fig.~\ref{fig:example_701} about the price of oil barrels).
\bibliographystyle{splncs04}
| b4cd13ceeedca9aef31d5dbbc70a27a9f1c56e67 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Algebraic cobordism is supposed to be the universal oriented cohomology theory in algebraic geometry. While still largely hypothetical, it promises to be a valuable invariant in algebraic geometry: Voevodsky's original approach to proving Milnor conjecture was based on the algebraic cobordism spectrum $MGL$, Levine and Pandharipande used in \cite{levine-pandharipande} the classification of varieties up to cobordism over a field of characteristic 0 in their proof of degree 0 Donaldson--Thomas conjectures, and Sechin and Sechin--Semenov have found applications of algebraic Morava K-theories (defined using algebraic cobordism) to approximating torsion in Chow groups \cite{sechin-chern} and to arithmetic questions about algebraic groups \cite{sechin-semenov}. Moreover, finding the correct definition of algebraic cobordism is intimately connected with fundamental questions in algebraic geometry, such as the existence of a good Chow cohomology theory (and more generally a good bivariant intersection theory).
After the foundational treatise \cite{Levine:2007} by Levine and Morel, algebraic cobordism has attracted a lot of attention. The groups have gone through several redefinitions and generalizations in \cite{levine-pandharipande}, \cite{lowrey-schurg}, \cite{annala-cob}, \cite{annala-yokura}, \cite{annala-pre-and-cob}, which, starting from \cite{lowrey-schurg} make serious use of derived algebraic geometry. In the most recent work \cite{annala-pre-and-cob} it was shown that given a Noetherian ring $A$ of finite Krull dimension, one can construct bivariant $A$-cobordism groups $\Omega^*_A(X \to Y)$ for all morphisms $X \to Y$ of quasi-projective derived $A$-schemes, and that these groups generalize all the previous geometric models of algebraic cobordism on quasi-projective derived schemes. This still leaves a lot to hope for: first of all, it is not clear how these groups depend on the base ring $A$, neither is it clear how the corresponding homology theory --- algebraic bordism --- fits into this picture as the algebraic $A$-bordism groups $\Omega^A_*(X) := \Omega^{-*}_A\big(X \to \mathrm{Spec}(A) \big)$ are not expected to model algebraic bordism unless $A$ is regular. Moreover, it is conceivable that one would like to study the cobordism of schemes that are not quasi-projective over a base. The work presented in this article deals with these issues, among others.
\subsection{Summary of results}
Let us then summarize the contributions of this paper. Our first main result is the following base independence result.
\begin{thm*}[Corollary \ref{BivACobIndFromACor}]
Let $X \to Y$ be a morphism of derived schemes with $X$ and $Y$ quasi-projective over Noetherian rings $A$ and $B$ of finite Krull dimension. Then
$$\Omega^*_A(X \to Y) = \Omega^*_B(X \to Y).$$
\end{thm*}
\noindent In order to prove this result, we have to study the bivariant ideal of snc relations $\langle \mathcal{R}^\mathrm{snc}_A \rangle \subset {\underline\Omega}^*_A$ as $\Omega^*_A := {\underline\Omega}^*_A/\langle \mathcal{R}^\mathrm{snc}_A \rangle$, where ${\underline\Omega}^*_A$ is the universal precobordism theory over $A$. In Definition \ref{SNCIdealAltDef} and Theorem \ref{BivACobAltConsThm} we find explicit descriptions of the groups $\langle \mathcal{R}^\mathrm{snc}_A \rangle(X \to Y)$ and Corollary \ref{BivACobIndFromACor} is an immediate consequence of this description. Besides helping us to prove the desired base independence, and being interesting for its own sake, the simplified construction of bivariant algebraic cobordism also generalizes easily to give rise to bivariant algebraic cobordism $\Omega^\bullet$, which is bivariant theory defined on the $\infty$-category of finite dimensional Noetherian derived schemes admitting an ample family of line bundles (also called divisorial derived schemes, see Definition \ref{AmpleFamDef}). Note that the groups ${\underline\Omega}^\bullet(X \to Y)$ do not admit a grading in general, which is why we have adopted the notation ${\underline\Omega}^\bullet$ instead of ${\underline\Omega}^*$. We also state pleasant universal properties for the bivariant theories ${\underline\Omega}^\bullet$ and $\Omega^\bullet$ (Theorem \ref{UnivPropOfPCobThm} and Corollary \ref{UnivPropOfCobCor}) analogous to the universal properties of ${\underline\Omega}^*_A$ and $\Omega^*_A$ proved in \cite{annala-pre-and-cob}.
The second main result of this article is the definition of algebraic bordism in general, which is the oriented Borel--Moore homology version of algebraic cobordism, and the generalization of Levine--Morel's algebraic bordism outside the realm of characteristic 0 geometry. Since the bivariant theory $\Omega^\bullet$ is defined so generally, we can simply use the following definition.
\begin{defn*}[Definition \ref{AlgBordDef}]
Let $X$ be a divisorial and finite dimensional Noetherian derived scheme. We define the \emph{algebraic bordism} of $X$ as the Abelian group
$$\Omega_\bullet(X) := \Omega^\bullet \big(X \to \mathrm{Spec}(\mathbb{Z})\big).$$
In other words $\Omega_\bullet(X)$ classifies complete intersection derived schemes (see Section \ref{DerComplIntSubSect}) mapping projectively to $X$ ``up to cobordism''.
\end{defn*}
\noindent Of course, as hinted by the above definition, we could have defined $\Omega_\bullet(X)$ directly by itself, but realizing it as a part of a stably oriented bivariant theory automatically verifies that it has good formal properties and that it interacts well with the algebraic cobordism ring $\Omega^\bullet(X)$. Note that the above definition uses in crucial way the fact that the bivariant groups $\Omega^\bullet(X \to Y)$ are defined even when $X \to Y$ is not a morphism of finite type. If $A$ is a regular Noetherian ring and $X$ is quasi-projective over $A$, then $\Omega_\bullet(X)$ is isomorphic to $\Omega_*^A(X)$, but this not true in general. We note that it might be more correct to consider the modified bordism group $\Omega'_\bullet(X)$ (see the next paragraph) whenever $X$ fails to have an ample line bundle.
The third main result of this paper is the generalization of the results of \cite{annala-chern} to divisorial derived schemes. The first obstacle is that the old proof of nilpotence of Euler classes of vector bundles does not seem to generalize to derived schemes without an ample line bundle, leading us in Section \ref{NilpEulerClassSubSect} to define the modified bivariant theories ${\underline\Omega}'^\bullet$ and $\Omega'^\bullet$, for which the desired nilpotence can be proven. Most of Section \ref{GeneralPBFSect} is then dedicated to proving the projective bundle theorem.
\begin{thm*}[Theorem \ref{GeneralPBFThm}]
Let $\mathbb{B}^\bullet$ be a bivariant quotient theory of ${\underline\Omega}'^\bullet$, let $X \to Y$ be a morphism of finite dimensional divisorial Noetherian derived schemes, and let $E$ be a vector bundle of rank $r$ on $X$. Then the morphism
$$\mathcal{P} roj: \bigoplus_{i=0}^{r-1} \mathbb{B}^\bullet(X \to Y) \to \mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$$
given by
$$(\alpha_0, \alpha_1, ..., \alpha_r) \mapsto \sum_{i=0}^{r-1} e(\mathcal{O}(1))^i \bullet 1_{\mathbb{P}(E)/X} \bullet \alpha_i$$
is an isomorphism.
\end{thm*}
\noindent Even though the statement of the theorem is a straightforward generalization of the corresponding result in \cite{annala-chern}, the result itself does not come for free: the basic idea of the proof is still the same, but the details have to be considerably altered for derived schemes without an ample line bundle. Moreover, our modified proof has the advantage that we do not have to define and study Chern classes before proving the projective bundle formula; rather, we construct a good theory of Chern classes in ${\underline\Omega}'^\bullet$ (see Definition \ref{GeneralChernClassDefn} and Theorem \ref{GeneralChernClassThm}) using the argument of Grothendieck based on the projective bundle formula. The following result is a standard consequence of a good theory of Chern classes.
\begin{thm*}[Theorem \ref{GeneralCFThm}]
Suppose $X$ is a finite dimensional Noetherian derived scheme admitting an ample family of line bundles. Then the morphism
$$\mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}'^\bullet(X) \to K^0(X)$$
given by
$$[V \stackrel f \to X] \mapsto [Rf_* \mathcal{O}_V]$$
is an isomorphisms of rings. The theorem remains true if we replace ${\underline\Omega}'^\bullet$ by $\Omega'^\bullet$.
\end{thm*}
\noindent We also generalize the Grothendieck--Riemann--Roch theorem in Theorem \ref{GeneralGRRThm}. Since having an ample family of line bundles is a much weaker condition that being quasi-projective over an affine base (for instance, all regular schemes are divisorial), the results of Section \ref{GeneralPBFSect} of this paper considerably generalize those of \cite{annala-chern}. Note that this is not just abstract nonsense either: if $X$ is a divisorial (e.g. regular) non-quasi-projective variety over a field $k$ of positive characteristic, then none of the previous work on algebraic cobordism proves such a Conner--Floyd theorem above for $X$. We also note that divisorial derived schemes is almost as far as one could conceivably generalize the current proof of Conner--Floyd theorem: it might be possible to generalize everything to a derived scheme $X$ having the resolution property, but without this assumption, $K^0(X)$ is no longer generated by vector bundles on $X$, completely breaking down the construction of the Chern character.
Finally, in order to define our bivariant theories in this generality, we had to carefully define and study ample line bundles an quasi-projective morphisms of derived schemes (see Sections \ref{AmpleSubSect} and \ref{QProjSubSect}). The results obtained are of course very similar to the classical story, but generalizing them to derived geometry proved to be very useful for us, and hopefully will prove to be useful for others as well. Such results for derived schemes, as far as the author is aware, have not appeared anywhere else (except weaker versions in the unpublished note \cite{annala-qpnote}).
\subsection{Conventions}
We will freely use the language of $\infty$-categories and derived schemes whenever necessary. Given an additive sheaf $\mathcal{F}$ on a derived scheme $X$, we will denote its \emph{global sections} by $\Gamma(X; \mathcal{F})$: this is naturally regarded as a spectrum object in simplicial Abelian groups or equivalently as a chain complex of Abelian groups. We will denote by $\abs{\Gamma(X; \mathcal{F})}$ the underlying (connective) simplicial set of $\Gamma(X; \mathcal{F})$. The cohomology groups are defined as
$$H^i(X; \mathcal{F}) := \pi_{-i} \big( \Gamma(X; \mathcal{F})\big);$$
if $\mathcal{F}$ is a discrete sheaf, then these recover the usual cohomology groups of sheaves, while in general they are more closely related to hypercohomology. If $\mathcal{F}$ is discrete, we denote by $\Gamma_\mathrm{cl}(X; \mathcal{F})$ the classical global sections of $\mathcal{F}$.
A derived scheme $X$ has an underlying classical scheme $X_\mathrm{cl}$ which is also called the \emph{truncation}. A morphism $f: X \to Y$ of derived schemes is \emph{of finite type} (\emph{proper}) if $f_\mathrm{cl}$ is of finite type (proper) in the classical sense. Given a vector bundle $E$ on a derived scheme $X$, the projective bundle $\mathbb{P}(E) \to X$ classifies all the subbundles of $E$ of rank $1$, while, if $\mathcal{F}$ is a connective quasi-coherent sheaf on $X$, $\mathbf{P}(\mathcal{F}) := \mathcal{P} roj(\mathrm{LSym}^*_X(\mathcal{F}))$ classifies all surjections (maps inducing a surjection of sheaves on $\pi_0$) from $\mathcal{F}$ to line bundles. If $A$ is a Noetherian ring of finite Krull dimension, then a quasi-projective derived $A$-scheme $X$ is implicitly assumed to be Noetherian. All derived schemes are assumed to be separated, and to have finite Krull-dimensional Noetherian underlying space. Given a morphism $X \to Y$ of derived schemes, we will denote by $\mathbb{L}_{X/Y}$ its \emph{relative cotangent complex}. In the case where $Y \simeq \mathrm{Spec}(\mathbb{Z})$ this is denoted by $\mathbb{L}_{X}$ and called the \emph{absolute cotangent complex}.
Given an object $X$ in an $\infty$-category $\mathcal{C}$ with a distinguished final object $pt$, we will use the shorthand notation $\pi_X$ to denote the essentially unique morphism $X \to pt$. Bivariant orientations are denoted by $1_{X/Y}$ instead of $\theta(X \to Y)$ used in previous work. We will denote by $\mathbb{L}$ the Lazard ring. Given a formal group law $F(x,y)$, we will denote by $\mathrm{inv}_F(x)$ the formal inverse power series.
\subsection*{Acknowledgements}
The author is supported by Vilho, Yrj\"o and Kalle V\"ais\"al\"a Foundation of the Finnish Academy of Science and Letters.
\section{Background}
\subsection{Derived algebraic geometry}
The construction of our cohomology theories makes use of derived algebraic geometry. The main references for the basic results in derived algebraic geometry are the work of Toën--Vezzosi \cite{HAG1,HAG2} and the work of Lurie \cite{HTT,HA,SAG}, but the reader can also consult the background section of \cite{annala-yokura} where a concise summary of useful results is presented. Derived schemes are Zariski locally modeled by homotopy types of simplicial commutative algebras. Since we only consider derived schemes whose underlying topological space is Noetherian and of finite Krull dimension, we can use the following vanishing result.
\begin{lem}[Grothendieck vanishing]\label{GrothendieckVanishingLem}
Let $X$ be a Noetherian topological space with Krull dimension $n < \infty$, and let $\mathcal{F}$ be a sheaf of simplicial Abelian groups on $X$. Then, for all $j \in \mathbb{Z}$,
$$\pi_j\big(\Gamma(X; \mathcal{F})\big) = \pi_j\big(\Gamma(X; \tau_{\leq i}(\mathcal{F}))\big)$$
as soon as $i \geq n+j$.
\end{lem}
\begin{proof}
Using the fundamental cofibre sequences
$$\Gamma\big(X; \pi_{i+1}(\mathcal{F})[i+1] \big) \to \Gamma\big(X; \tau_{\leq i+1}(\mathcal{F}) \big) \to \Gamma\big(X; \tau_{\leq i}(\mathcal{F}) \big)$$
and the classical Grothendieck vanishing, we conclude that $\pi_j\big( \Gamma(X; \tau_{\leq i}(\mathcal{F})) \big)$
stabilize for $i \gg 0$, and therefore (by Milnor sequence)
$$\pi_j \big( \Gamma(X; \mathcal{F}) \big) = \lim_i \pi_j\big(\Gamma(X; \tau_{\leq i}(\mathcal{F}))\big).$$
The bound for $i$ follows from the fact that a Noetherian topological space of Krull dimension $n$ has cohomological dimension $n$.
\end{proof}
Moreover, for technical reasons it is often useful to restrict our attention to Noetherian schemes:
\begin{defn}\label{NoetherianDef}
A derived scheme $X$ is \emph{Noetherian} if $X_\mathrm{cl}$ is Noetherian as a classical scheme and all the homotopy sheaves $\pi_i(\mathcal{O}_X)$ are coherent $\mathcal{O}_{X_\mathrm{cl}}$-modules.
\end{defn}
\begin{defn}\label{CoherentDef}
A quasi-coherent sheaf $\mathcal{F}$ on a Noetherian derived scheme $X$ is called \emph{almost perfect} if the homotopy sheaves $\pi_i(\mathcal{F})$ are coherent sheaves on $X_\mathrm{cl}$, and they vanish for $i \ll 0$. An almost perfect sheaf $\mathcal{F}$ is \emph{coherent} if all but finitely many homotopy sheaves $\pi_i(\mathcal{F})$ vanish.
\end{defn}
\subsubsection{Ample line bundles and ample families}\label{AmpleSubSect}
The purpose of this section to define and to study ample line bundles and schemes admitting an ample family of line bundles in the context of derived algebraic geometry. The results in this section generalize the well known classical results as well as the (weaker) results of the unpublished note \cite{annala-qpnote}. Let us start with the fundamental definitions.
\begin{defn}\label{AmpleDef}
Let $X$ be a Noetherian derived scheme and let $\mathscr{L}$ be a line bundle on $X$. We say that $\mathscr{L}$ is \emph{ample} if for any point $x \in X$ there exists $n \geq 0$ and a section $s \in \Gamma(X; \mathscr{L}^{\otimes n})$ such that $X_s := X \backslash V(s)$ is affine and contains $x$.
\end{defn}
\begin{defn}\label{AmpleFamDef}
Let $X$ be a Noetherian derived scheme. We say that $X$ is \emph{divisorial} (or $X$ \emph{has an ample family of line bundles}) if for any point $x \in X$ there exists a line bundle $\mathscr{L}$ and a global section $s \in \Gamma(X; \mathscr{L})$ such that $X_s$ is affine and contains $x$.
\end{defn}
The following Lemma will be important when analyzing the ramifications of the above definitions.
\begin{lem}\label{LocLem}
Let $X$ be a Noetherian derived scheme, let $\mathscr{L}$ be a line bundle on $X$ and let $s$ be a global section of $\mathscr{L}$. Let $\mathcal{F}$ be a quasi-coherent sheaf on $X$, and consider the sequences
$$H^i(X; \mathcal{F}) \xrightarrow{s \cdot} H^i(X; \mathscr{L} \otimes \mathcal{F}) \xrightarrow{s \cdot} H^i(X; \mathscr{L}^{\otimes 2} \otimes \mathcal{F}) \xrightarrow{s \cdot} \cdots$$
Then there are natural isomorphisms
$$\mathrm{colim}_n H^i \big(X; \mathscr{L}^{\otimes n} \otimes \mathcal{F} \big) \xrightarrow{\cong} H^i(X_s; \mathcal{F})$$
given by sending $f \in H^i(X; \mathscr{L}^{\otimes n} \otimes \mathcal{F})$ to $f/s^n \in H^i(X_s; \mathcal{F})$.
\end{lem}
\begin{proof}
It is enough to prove this for $i=0$. Since applying the truncation $\tau_{\geq 0}$ does not change the zeroth cohomology, we can assume that $\mathcal{F}$ is connective. Finally, we can use Grothendieck vanishing (Lemma \ref{GrothendieckVanishingLem}) to reduce to the situation where the nontrivial homotopy sheaves of $\mathcal{F}$ lie between degrees $0$ and $n$, where $n$ is at most the Krull dimension of $X$. Our argument proceeds by induction on $n$. Notice that the base case $n=0$ is classical (see e.g. \cite{Stacks} Tag 09MR). To prove the general case, we consider the cofibre sequence
$$\mathcal{F}' \to \mathcal{F} \to \pi_0(\mathcal{F}).$$
Since the nontrivial homotopy sheaves of $\mathcal{F}'$ lie between degrees $1$ and $n$, we can apply the induction assumption to $\mathcal{F}'$ and $\pi_0(\mathcal{F})$, and the claim follows from 5-lemma applied to the diagram comparing the cohomology long exact sequences (note that sequential colimits preserve exact sequences).
\end{proof}
We can now prove the following useful characterization of ampleness.
\begin{prop}\label{AmpleCharProp1}
Let $X$ be a Noetherian derived scheme, and let $\mathscr{L}$ be a line bundle on $X$. Then the following are equivalent
\begin{enumerate}
\item $\mathscr{L}$ is ample;
\item for every connective almost perfect sheaf $\mathcal{F}$ on $X$, the sheaves $\mathscr{L}^{\otimes n} \otimes \mathcal{F}$ are globally generated for $n \gg 0$.
\end{enumerate}
\end{prop}
\begin{rem}
We recall that a connective quasi-coherent sheaf $\mathcal{F}$ on $X$ is \emph{globally generated} if there exists a set $I$ and a morphism
$$\mathcal{O}_X^{\oplus I} \to \mathcal{F}$$
inducing a surjection on $\pi_0$.
\end{rem}
\begin{proof}
Using Grothendieck vanishing (Lemma \ref{GrothendieckVanishingLem}) we can assume that $\mathcal{F}$ is coherent. Let us first assume that $\mathscr{L}$ is ample, and let us begin by choosing $s_i \in \Gamma(X; \mathscr{L}^{\otimes n_i})$, $i=1..r$, so that $X_{s_i}$ form an affine open cover of $X$. As
$$\pi_0\big(\Gamma(X_{s_i}; \mathcal{F})\big) \cong \Gamma\big(X_{s_i}; \pi_0(\mathcal{F}) \big),$$
we can use Lemma \ref{LocLem} to conclude that for $d_i \gg 0$ there exist global sections of $\mathscr{L}^{\otimes d_i n_i} \otimes \mathcal{F}$ generating it at the points of $X_{s_i}.$ Letting $N = n_1 \cdots n_r$, it then follows that $\mathscr{L}^{\otimes iN} \otimes \mathcal{F}$ is globally generated for $i \gg 0$. Since the last conclusion is also true for the sheaves $\mathscr{L} \otimes \mathcal{F}, ..., \mathscr{L}^{\otimes N-1} \otimes \mathcal{F}$, it follows that $\mathscr{L}^{\otimes n} \otimes \mathcal{F}$ is globally generated for $n \gg 0$.
Suppose then that for all coherent sheaves $\mathcal{F}$, $\mathscr{L}^{\otimes n} \otimes \mathcal{F}$ is globally generated for $n \gg 0$. Let $x \in X$ be a point, and let $U$ be an affine open neighbourhood of $x$ s.t. $\mathscr{L} \vert_U$ is trivial. Let $Z \hookrightarrow X$ be the complement of $U$ with the reduced scheme structure, and consider the cofibre sequence
$$\mathcal{I} \to \mathcal{O}_X \to \mathcal{O}_Z.$$
Supposing that $\mathscr{L}^{\otimes n} \otimes \mathcal{I}$ is globally generated, it is possible to find a global section $s$ so that its image in $s' \in \Gamma(X; \mathscr{L}^{\otimes n})$ does not vanish at $x$. It follows that $X_{s'} \subset U$ is an affine open subset containing $x$.
\end{proof}
Admitting an ample family of line bundles has an alternative characterization as well.
\begin{prop}\label{AmpleFamCharProp}
Let $X$ be a Noetherian derived scheme. Then the following are equivalent
\begin{enumerate}
\item $X$ is divisorial;
\item for any connective almost perfect sheaf $\mathcal{F}$ on $X$, there exists a surjection
$$\mathscr{L}_1 \oplus \mathscr{L}_2 \oplus \cdots \oplus \mathscr{L}_r \to \mathcal{F}$$
from a direct sum of line bundles.
\end{enumerate}
\end{prop}
\begin{proof}
Suppose first that $X$ admits an ample family of line bundles. Choose global sections $s_1, ... , s_r$ of line bundles $\mathscr{L}_1, ..., \mathscr{L}_r$ such that $X_{s_i}$ form an affine open covering for $X$. Given a connective almost perfect sheaf $\mathcal{F}$ on $X$, we can use Lemma \ref{LocLem} to find $n_i \geq 0$ and global sections $s_{i1},...,s_{id_i}$ which generate $\mathscr{L}^{\otimes n_i}_i \otimes \mathcal{F}$ at the points of $X_{s_i}$. Said otherwise, we find a surjective morphism
$$\big( \mathscr{L}^{\otimes - n_1}_1 \big)^{\oplus d_1} \oplus \cdots \oplus \big( \mathscr{L}^{\otimes - n_r}_r \big)^{\oplus d_r} \to \mathcal{F},$$
which is exactly what we wanted.
To prove the converse, let $U$ be an affine open neighborhood of $x$ with reduced complement $Z$, and consider the cofibre sequence
$$\mathcal{I} \to \mathcal{O}_X \to \mathcal{O}_Z$$
of almost perfect sheaves on $X$. Given a surjective morphism
$$\mathscr{L}_1 \oplus \cdots \oplus \mathscr{L}_r \to \mathcal{I},$$
then for some $1 \leq i \leq r$ the induced map $\mathscr{L}_i \to \mathcal{O}_X$ is surjective at $x$. In other words $\mathscr{L}_i^\vee$ has a global section $s$ such that $x \in X_s$. As $X_s \subset U$ and $U$ is affine, also $X_s$ is affine.
\end{proof}
Next we consider the relative version of ampleness.
\begin{defn}\label{RelAmpleDef}
Let $f: X \to Y$ be a morphism of derived schemes. We say that a line bundle $\mathscr{L}$ on $X$ is \emph{relatively ample} (or \emph{$f$-ample}) if for every affine open set $\mathrm{Spec}(B) \to Y$, $\mathscr{L}$ restricts to an ample line bundle on the inverse image $f^{-1} \mathrm{Spec}(B)$.
\end{defn}
if $Y$ is affine, then a line bundle $\mathscr{L}$ on $X$ is relatively ample over $Y$ if and only if $\mathscr{L}$ is ample. Combining this observation with the following lemma allows us to check relative ampleness on an affine cover of $Y$.
\begin{lem}\label{RelAmpleLocalLem}
Let $f: X \to Y$ be a morphism of Noetherian derived schemes, let $(U_i)_{i \in I}$ be an open cover of $Y$, and let $\mathscr{L}$ be a line bundle on $X$. Then $\mathscr{L}$ is $f$-ample if and only if for all $i \in I$, we have that $\mathscr{L}_i := \mathscr{L} \vert_{f^{-1} U_i}$ is $f_i$-ample, where $f_i: f^{-1} U_i \to U_i$ is the restriction of $f$.
\end{lem}
\begin{proof}
The only if direction is clear, so let us prove that $\mathscr{L}_i$ being $f_i$-ample implies the relative ampleness of $\mathscr{L}$. In other words we have to prove the following: if $Y = \mathrm{Spec}(B)$ is affine, $b_1, ..., b_r \in B$ are functions such that $(Y_{b_i})_{i=1}^r$ is an affine open cover of $Y$ with $\mathscr{L} \vert_{X_{b_i}}$ ample for all $i$, then $\mathscr{L}$ is ample. But proving this is easy: if $x \in X$, then $x$ lies inside one of the open sets $X_{b_i}$, and we can use Lemma \ref{LocLem} to find a global section $s$ of $\mathscr{L}^{\otimes n}$ with $X_{s} \subset X_{b_i}$ affine and containing $x$.
\end{proof}
As a consequence, we prove that relatively ample line bundles behave well in compositions and in derived pullbacks.
\begin{prop}\label{AmpleLineBundlesInCompositionsLem}
Let $f: X \to Y$ be a morphism of Noetherian derived schemes having a relatively ample line bundle $\mathscr{L}$. Then
\begin{enumerate}
\item if $g: Y \to Z$ and $\mathscr{L}$ is $(g \circ f)$-ample, then $\mathscr{L}$ is $f$-ample;
\item if $g: Y \to Z$ admits a relatively ample line bundle $\mathscr{M}$, then $\mathscr{L} \otimes f^* \mathscr{M}^{\otimes n}$ is $(g \circ f)$-ample for $n \gg 0$;
\item if
$$
\begin{tikzcd}
X' \arrow[]{r}{f'} \arrow[]{d}{h'} & Y' \arrow[]{d}{h} \\
X \arrow[]{r}{f} & Y
\end{tikzcd}
$$
is derived Cartesian, then $h'^* \mathscr{L}$ is $f'$-ample.
\end{enumerate}
\end{prop}
\begin{proof}
\begin{enumerate}
\item The proof is a trivial consequence of Lemma \ref{RelAmpleLocalLem} and the fact that ample line bundles are relatively ample and stable under restrictions to open subschemes.
\item The proof is the same as the classical proof, see \cite{Stacks} Tags 0C4K (1) and 0892.
\item Without loss of generality we can assume that both $Y$ and $Y'$ are affine. It follows that $h'$ is affine, and as ample line bundles are stable under affine pullbacks, the claim follows. \qedhere
\end{enumerate}
\end{proof}
We also record the following useful fact.
\begin{prop}\label{AmpleFamilyInQuasiProjProp}
Let $f: X \to Y$ be a morphism of Noetherian derived schemes having a relatively ample line bundle. If $Y$ has an ample family of line bundles, then so does $X$.
\end{prop}
\begin{proof}
Indeed, if $\mathscr{L}$ is a relatively ample line bundle, and $s$ is a global section of a line bundle $\mathscr{M}$ on $Y$ with $Y_s$ affine, then by Lemma \ref{LocLem} and the definition of relatively ample bundles, we conclude that we can find an affine open cover of $f^{-1} Y_s$ using global sections of $f^*(\mathscr{M})^{\otimes n} \otimes \mathscr{L}^{\otimes m}$ for $n,m \gg 0$. Therefore, if $Y$ has an ample family, then so does $X$.
\end{proof}
For completeness' sake, we also prove that relative ampleness can be checked on the truncation, at least for proper morphisms.
\begin{prop}\label{AmpleCharProp2}
Let $A$ be a Noetherian simplicial ring, and let $X \to \mathrm{Spec}(A)$ be a proper morphism of derived schemes. Then, for a line bundle $\mathscr{L}$ on $X$, the following are equivalent:
\begin{enumerate}
\item $\mathscr{L}_\mathrm{cl}$ is ample on $X_\mathrm{cl}$;
\item given an almost perfect sheaf $\mathcal{F}$ on $X$ and $i \in \mathbb{Z}$, we have that
$$\pi_i \big( \Gamma(X; \mathcal{F} \otimes \mathscr{L}^{\otimes n}) \big) \cong \Gamma_\mathrm{cl}(X; \pi_i(\mathcal{F} \otimes \mathscr{L}^{\otimes n}))$$
for $n \gg 0$;
\item $\mathscr{L}$ is ample.
\end{enumerate}
\end{prop}
\begin{proof}
Let us first assume that the truncation $\mathscr{L}_\mathrm{cl}$ is ample. By Grothendieck vanishing, it is enough to check that $2$ holds for $\mathcal{F}$ coherent. We will proceed by induction on $n$, where $n$ is the largest $i$ with $\pi_i(\mathcal{F})$ nontrivial. Notice that the base case $n=0$ is easy: then
$$\mathcal{F} \otimes_{\mathcal{O}_X} \mathscr{L}^{\otimes m} \simeq \mathcal{F} \otimes_{\mathcal{O}_{X_\mathrm{cl}}} \mathscr{L}_\mathrm{cl}^{\otimes m},$$
and the claim follows from the classical result of Serre on the vanishing of sheaf cohomology (see, e.g., \cite{Stacks} Tag 0B5U). In general we have a cofibre sequence
$$\mathcal{F}' \to \mathcal{F} \to \pi_0(\mathcal{F}).$$
Applying the induction assumption to $\mathcal{F}'$ and $\pi_0(\mathcal{F})$, the result holds for them for $m$ large enough. Then for any such $m$, it follows from the associated cohomology long exact sequence that
$$\pi_i\big(\Gamma(X; \mathcal{F} \otimes \mathscr{L}^{\otimes n})\big) \cong
\begin{cases}
\pi_i\big(\Gamma(X; \mathcal{F}' \otimes \mathscr{L}^{\otimes n})\big) \cong \Gamma_\mathrm{cl}(X; \pi_i(\mathcal{F} \otimes \mathscr{L}^{\otimes n})) & \text{for $i > 0$;} \\
\Gamma_\mathrm{cl}(X; \pi_0(\mathcal{F} \otimes \mathscr{L}^{\otimes n}))\big) & \text{for $i = 0$;} \\
0 & \text{for $i < 0$.}
\end{cases}
$$
This shows that $\mathscr{L}$ satisfies condition 2.
Suppose then that $\mathscr{L}$ satisfies condition 2. Applying this to all truncated coherent sheaves, we can immediately conclude that $\mathscr{L}_\mathrm{cl}$ is ample using the result of Serre cited above. As global sections of $\mathscr{L}^{\otimes n}_\mathrm{cl}$ lift to global sections of $\mathscr{L}^{\otimes n}$ for $n \gg 0$, the ampleness of $\mathscr{L}_\mathrm{cl}$ implies that of $\mathscr{L}$.
\end{proof}
\subsubsection{Quasi-projective morphisms}\label{QProjSubSect}
In this section, we define and study the basic properties of quasi-projective morphisms. Like in the previous section, the results generalize some of those in the unpublished note \cite{annala-qpnote}. Let us begin with explaining what we mean by a quasi-projective morphism.
\begin{defn}\label{QProjDef}
A finite type morphism $f: X \to Y$ of Noetherian derived schemes is \emph{quasi-projective} if it admits a relatively ample line bundle. A proper quasi-projective morphism is called \emph{projective}.
\end{defn}
We have chosen to use the above definition for quasi-projectivity as it has the following good properties.
\begin{prop}\label{StabilityOfQProjProp}
Quasi-projective and projective morphisms between Noetherian derived schemes are stable under compositions and derived pullbacks.
\end{prop}
\begin{proof}
This follows immediately from Proposition \ref{AmpleLineBundlesInCompositionsLem}.
\end{proof}
Besides the good properties verified by the above result, projective morphisms have at least one additional benefit over the general proper morphisms: they often admit nice global factorizations.
\begin{thm}\label{ExistenceOfFactorizationsThm}
Let $f: X \to Y$ be a projective morphism of Noetherian derived schemes.
\begin{enumerate}
\item If $Y$ has the resolution property, then there exists a vector bundle $E$ on $Y$ and a factorization of $f$ as
$$X \stackrel i \hookrightarrow \mathbb{P}(E) \stackrel \pi \to Y,$$
where $\pi$ is the canonical projection and $i$ is a closed embedding.
\item If $Y$ has an ample family of line bundles, then there exists line bundles $\mathscr{L}_1, ..., \mathscr{L}_r$ on $Y$ and a factorization of $f$ as
$$X \stackrel i \hookrightarrow \mathbb{P}(\mathscr{L}_1 \oplus \cdots \oplus \mathscr{L}_r) \stackrel \pi \to Y,$$
where $\pi$ is the canonical projection and $i$ is a closed embedding.
\item If $Y$ has an ample line bundle, then it is possible to find a similar factorization of form
$$X \stackrel i \hookrightarrow \mathbb{P}^n_Y \stackrel \pi \to Y.$$
\end{enumerate}
\end{thm}
\begin{rem}
We recall that a derived scheme $X$ is said to have the \emph{resolution property} if every connective coherent sheaf $\mathcal{F}$ admits a surjection $E \to \mathcal{F}$ (meaning that this induces a surjection on $\pi_0$) with $E$ a vector bundle on $X$. Every divisorial derived scheme has the resolution property by Proposition \ref{AmpleFamCharProp}.
\end{rem}
\begin{proof}
Let $\mathscr{L}$ be an $f$-ample line bundle on $X$. Then, by Proposition \ref{AmpleCharProp2}, we have that $f_*(\mathscr{L}^{\otimes n})$ is connective and that the natural morphism
$$f_*(\mathscr{L}^{\otimes n}) \to f_* \big( \pi_0(\mathscr{L}^{\otimes n}) \big)$$
identifies the right hand side as the $\pi_0$ of the left hand side for $n \gg 0$. In particular, the $\pi_0$ of the natural map $f^*f_* (\mathscr{L}^{\otimes n}) \to \mathscr{L}^{\otimes n}$ can be identified with the classical morphism $f_\mathrm{cl}^* f_{\mathrm{cl} *} (\mathscr{L}_\mathrm{cl}^{\otimes n}) \to \mathscr{L}_\mathrm{cl}^{\otimes n}$. On the other hand, by \cite{Stacks} Tags 01VR and 01VU, the classical map is a surjection of sheaves and the induced morphism
$$X \hookrightarrow \mathbf{P}\big(f_{*} (\mathscr{L}^{\otimes n}) \big)$$
is a closed embedding over $Y$ for $n \gg 0$ (since the truncation is a closed embedding). As $f_* \mathscr{L}$ is coherent and connective, and as $Y$ has the resolution property, we can choose a surjection $E \to f_* \mathscr{L}$ from a vector bundle $E$, and the composition
$$i: X \hookrightarrow \mathbf{P}\big(f_{*} (\mathscr{L}^{\otimes n}) \big) \hookrightarrow \mathbf{P}(E) = \mathbb{P}(E^\vee)$$
is a closed embedding of $Y$-schemes, proving the first claim.
If $Y$ has an ample family of line bundles, then we may choose $E$ to be a sum of line bundles, proving the second claim. In the special case where $Y$ has an ample line bundle $\mathscr{L}$, we can choose $E$ to be a direct sum of copies of $\mathscr{L}^{\otimes -n}$ for some $n \geq 0$, proving the third claim.
\end{proof}
\subsubsection{Derived regular embeddings and quasi-smooth morphisms}
\begin{defn}\label{RegEmbDef}
Let $i: X \hookrightarrow Y$ be a closed embedding of derived schemes. Then it is called a \emph{derived regular embedding} of \emph{codimension $r$} if, Zariski locally on the target, it is equivalent to
$$\mathrm{Spec}\big(A /\!\!/ (a_1, ..., a_r)\big) \hookrightarrow \mathrm{Spec}(A),$$
where $/\!\!/$ denotes the derived quotient ring construction (see \cite{khan-rydh} 2.3.1). In other words, $Z$ locally the derived vanishing locus of $r$ regular functions on $Y$.
\end{defn}
The following differential characterization is essentially just Proposition 2.3.8 of \cite{khan-rydh}.
\begin{prop}\label{ModifiedKhanRydh}
Let $f: X \hookrightarrow Y$ be a closed embedding of Noetherian derived schemes with $X$ connected. Then the following are equivalent
\begin{enumerate}
\item $f$ is a derived regular embedding;
\item the shifted cotangent complex $\mathbb{L}_{X/Y}[-1]$ is a vector bundle;
\item the shifted cotangent complex $\mathbb{L}_{X/Y}[-1]$ has Tor-amplitude 0.
\end{enumerate}
Henceforward we will denote $\mathcal{N}^\vee_{X/Y} := \mathbb{L}_{X/Y}[-1]$ and call it the \emph{conormal bundle}.
\end{prop}
\begin{proof}
The proof of \emph{loc. cit.} shows the equivalence of 1 and 2, but we have to pay some attention to one key step. Zariski locally on $Y$ the morphism $f$ is represented by a morphism $A \to B$ of simplicial commutative rings inducing a surjection on $\pi_0$. Let $F$ denote the homotopy fibre of this morphism considered as an $A$-module. The argument of \emph{loc. cit.} goes through once we show that $\pi_0(F)$ is finitely generated over $\pi_0(A)$ (otherwise one can not use Nakayama). But this is easy: we have an exact sequence
$$\pi_1(B) \to \pi_0(F) \to \pi_0(A) \to \pi_0(B) \to 0$$
and as $\pi_1(B)$ is a finitely generated $\pi_0(A)$-module, the claim follows.
Notice that 3 is equivalent to $\mathbb{L}_{X/Y}[-1]$ being flat. Since, we have already shown that $\pi_0(\mathbb{L}_{X/Y}[-1])$ is coherent, we conclude that 3 implies 2. Since vector bundles are flat, we are done.
\end{proof}
\begin{defn}\label{StrSmQSmDef}
Let $f: X \to Y$ be a morphism of derived schemes. Then
\begin{enumerate}
\item $f$ is called \emph{strong} if the natural maps $f^*_\mathrm{cl} \pi_i(\mathcal{O}_Y) \to \pi_i(\mathcal{O}_X)$ are isomorphisms for all $i \geq 0$;
\item $f$ is called \emph{smooth} if it is strong and the truncation $f_\mathrm{cl}$ is smooth in the classical sense;
\item $f$ is called \emph{quasi-smooth} if, Zariski locally on $X$, it admits a factorization
$$X \stackrel i \hookrightarrow P \stackrel p \to Y$$
with $i$ a derived regular embedding and $p$ a smooth morphism.
\end{enumerate}
\end{defn}
The following alternative characterization of smooth morphisms will be useful for us.
\begin{lem}\label{SmCharLem}
A morphism $f: X \to Y$ of Noetherian derived schemes is smooth if and only if $f_\mathrm{cl}$ is of finite type and $\mathbb{L}_{X/Y}$ is a vector bundle.
\end{lem}
\begin{proof}
This is well known if $X$ and $Y$ are classical (finite type morphisms of Noetherian schemes are of finite presentation). To prove the equivalence in general, consider the derived Cartesian square
$$
\begin{tikzcd}
X' \arrow[]{r}{f'} \arrow[]{d}{\iota_X} & Y_\mathrm{cl} \arrow[]{d}{\iota_Y} \\
X \arrow[]{r}{f} & Y.
\end{tikzcd}
$$
If $f$ is smooth, then it is strong and flat, and therefore $X'$ is classical and identified with $X_\mathrm{cl}$, which proves that $\iota_X^* \mathbb{L}_{X/Y}$, and hence $\mathbb{L}_{X/Y}$, is a vector bundle. To prove the converse implication, it is enough to show that if $\mathbb{L}_{X/Y}$ is a vector bundle and $Y$ is classical, then $X$ is classical as well. Locally we have
$$A \to A [x_1,...,x_n] \to B := A [x_1,...,x_n] /\!\!/ (f_1,...,f_r)$$
and the assumptions imply that we have a split injection of $\pi_0(B)$-modules
$$\langle df_1,...,df_r \rangle_{\pi_0(B)} \hookrightarrow \pi_0(B) \otimes_{A[x_1,...,x_n]} \Omega_{A[x_1,...,x_n] / A}.$$
But this implies that
$$A \to A [x_1,...,x_n] \to B := A [x_1,...,x_n] / (f_1,...,f_r)$$
is a relative global complete intersection (in the sense of \cite{Stacks} Tag 00SP) with smooth fibres (differential criterion for smoothness). Hence $f_1,..,f_r$ is a regular sequence around the points of $V(f_1,...,f_r)$ and $B \simeq A [x_1,...,x_n] / (f_1,...,f_r)$ is discrete.
\end{proof}
The following result is then an easy consequence of Proposition \ref{ModifiedKhanRydh}.
\begin{prop}\label{QSmCharProp}
Let $f: X \to Y$ be a morphism of Noetherian derived schemes. Then the following are equivalent:
\begin{enumerate}
\item $f$ is quasi-smooth;
\item the truncation $f_\mathrm{cl}$ is of finite type and the cotangent complex $\mathbb{L}_{X/Y}$ is a perfect complex of Tor-dimension 1;
\item the truncation $f_\mathrm{cl}$ is of finite type and the cotangent complex $\mathbb{L}_{X/Y}$ has Tor-dimension 1.
\end{enumerate}
\end{prop}
\begin{proof}
Since the truncation is of finite type, $f$ admits a local factorization
$$X \stackrel i \hookrightarrow P \stackrel p \to Y$$
with $p$ smooth (take $P = \mathbb{A}^n \times Y$ for $n$ large enough). The claim then follows by investigating the fundamental triangle
$$i^* \mathbb{L}_{P / Y} \to \mathbb{L}_{X/Y} \to \mathbb{L}_{X/P}$$
and using Proposition \ref{ModifiedKhanRydh} and Lemma \ref{SmCharLem}.
\end{proof}
If $f: X \to Y$ is quasi-smooth and $\mathbb{L}_{X/Y}$ has constant virtual rank $d$ at the points of $x$, then we say that $f$ has \emph{relative virtual dimension $d$}. It is then easy to prove the following basic result.
\begin{prop}
Quasi-smooth morphisms of Noetherian derived schemes are stable under compositions and derived pullbacks. Moreover, relative virtual dimension is additive under composition, and is preserved under derived pullback. \qed
\end{prop}
\subsubsection{Derived complete intersection rings}\label{DerComplIntSubSect}
The purpose of this section is to introduce an absolute version of quasi-smoothness, and verify it has desirable properties. A simplicial commutative ring is called \emph{local} if its truncation is local.
\begin{defn}
A local Noetherian simplicial ring $A$ is called \emph{derived complete intersection} if only finitely many $\pi_i(A)$ are nontrivial and the cotangent complex $\mathbb{L}_{\kappa / A}$, where $\kappa$ is the residue field of $A$, is perfect and concentrated in degrees less than or equal to 2.
A Noetherian derived scheme is called \emph{derived complete intersection} if its local rings are derived complete intersection rings. We also call such derived schemes \emph{derived regular}.
\end{defn}
\begin{ex}
If $A$ is a regular local ring and $a_1,...,a_r$ are elements of $A$, then the derived quotient ring $A /\!\!/ (a_1, ..., a_r)$ is a derived complete intersection ring.
\end{ex}
The following result explains what we want to capture with this definition.
\begin{prop}\label{AbsoluteQSmProp}
Suppose $B$ is a derived complete intersection local ring, and let $f: A \to B$ be a morphism from a regular local ring $A$ which is a surjection on $\pi_0$. Then there exist elements $a_1,..., a_r$ and an equivalence
$$B \simeq A /\!\!/ (a_1,...,a_r)$$
of $A$-algebras. In other words, a closed immersion from a derived regular scheme to a regular scheme is a derived regular embedding.
\end{prop}
\begin{proof}
Let $\kappa$ be the residue field of $A$, and consider the sequence $A \to B \to \kappa$ inducing the fundamental triangle
$$\kappa \otimes_B^L \mathbb{L}_{B/A} \to \mathbb{L}_{\kappa / A} \to \mathbb{L}_{\kappa / B}.$$
As $A$ is regular, $\mathbb{L}_{\kappa / A}$ is a $\kappa$-vector space in degree 1, and $B$ is a derived complete intersection, it follows that $\kappa \otimes_B^L \mathbb{L}_{B/A}$ is concentrated in degree 1, and the claim follows from Proposition \ref{ModifiedKhanRydh}.
\end{proof}
The following alternative characterization of derived complete intersection schemes is also useful.
\begin{prop}\label{DCompIntAltCharProp}
Let $X$ be a Noetherian derived scheme. Then the following are equivalent:
\begin{enumerate}
\item $X$ is a derived complete intersection scheme;
\item the absolute cotangent complex $\mathbb{L}_X$ has Tor-dimension 1;
\item for all maps $X \to Y$ with $Y$ a regular Noetherian scheme, $\mathbb{L}_{X/A}$ has Tor-dimension 1.
\end{enumerate}
\end{prop}
\begin{proof}
Let $B$ be a derived local ring of $X$ with residue field $\kappa$, let $A$ be a regular ring and consider the sequence $A \to B \to \kappa$ inducing the cofibre sequence
$$\kappa \otimes_B \mathbb{L}_{B/A} \to \mathbb{L}_{\kappa/A} \to \mathbb{L}_{\kappa/B}.$$
Since $\mathbb{L}_{\kappa/A}$ has Tor dimension 1, we conclude that $\mathbb{L}_{\kappa/B}$ has Tor-dimension 2 if and only if $\mathbb{L}_{B/A}$ has Tor-dimension 1, proving the equivalence of the three conditions.
\end{proof}
Finally we need to be able to talk about the dimension of a derived complete intersection scheme.
\begin{defn}\label{VDimDef}
Let $A$ be a derived complete intersection local ring with residue field $\kappa$. Then the \emph{virtual dimension} of $A$ is defined as
$$\dim(A) := \dim_\kappa \pi_1(\mathbb{L}_{\kappa / A}) - \dim_\kappa \pi_2(\mathbb{L}_{\kappa / A}).$$
If $X$ is a derived complete intersection scheme and $x \in X$ is a point, then we define the \emph{virtual dimension of $X$ at $x$}, $\dim_x(X)$, as the virtual dimension of the derived local ring $\mathcal{O}_{X,x}$. Finally, the \emph{virtual dimension} of such an $X$, $\dim(X)$, is defined as the supremum of $\dim_x(X)$, where $x$ ranges over all the points of $X$.
\end{defn}
\begin{rem}
Notice that if $A$ is a classical complete intersection ring, then the virtual dimension of $A$ agrees with the Krull dimension of $A$. Indeed, for regular local rings this is easy, and for more general complete intersection local rings it follows from the stability of cotangent complex in derived pullbacks that
$$\dim(A) = \dim(\hat A),$$
where $\hat A$ denotes the completion of $A$ with respect to its maximal ideal. On the other hand, since $\hat A$ is a quotient of a regular local ring by a regular sequence, one observes easily that the virtual dimension of $\hat A$ agrees with its Krull dimension. Since also the Krull dimensions of $A$ and $\hat A$ are the same, we conclude that the virtual dimension of $A$ agrees with its Krull dimension.
\end{rem}
\begin{rem}\label{DimProblemRem}
One would expect that, at least under suitable hypotheses, a connected derived complete intersection scheme is equicodimensional in the sense that $\dim_x(X) = \dim(X)$ for all closed points $x$ of $X$. It is not hard to see that this is true for $X$ finite type over a field $k$, but we won't try to pursue this in any greater generality (doing so would take us too far afield). This means that we don't know if virtual dimension is a good notion: for instance, consider a derived regular embedding $Z \hookrightarrow X$ of codimension $r$, and suppose $X$ is a derived complete intersection scheme of virtual dimension $d$. Although $Z$ is clearly a derived complete intersection scheme, it is not clear at all that it has virtual dimension $d-r$ (and in fact we would expect this to be false in general).
This potential for ill behavior will have ramifications later in this article. Indeed, we will construct base independent homological bordism groups $\Omega_\bullet(X)$ but we won't be able to prove that these groups admit a grading by the virtual dimension of cycles.
\end{rem}
\subsubsection{Derived blow ups}
One of the main technical tools we are going to need in this article is the construction of derived blow ups and derived deformation to normal cone from \cite{khan-rydh}. Let us recall the definitions and the results we are going to use:
\begin{defn}\label{DivisorOverZDef}
Let $Z \hookrightarrow X$ be a derived regular embedding. Then, for any $X$-scheme $S$, a \emph{virtual Cartier divisor} on $S$ \emph{lying over $Z$} is the datum of a commutative diagram
\begin{center}
\begin{tikzcd}
D \arrow[hook]{r}{i_D} \arrow[]{d}{g} & S \arrow[]{d}{} \\
Z \arrow[hook]{r}{} & X
\end{tikzcd}
\end{center}
such that
\begin{enumerate}
\item $i_D$ is a quasi-smooth closed embedding of codimension 1;
\item the truncation is a Cartesian square;
\item the canonical morphism
$$g^* \mathcal{N}^\vee_{Z/X} \to \mathcal{N}^\vee_{D/S}$$
induces a surjection on $\pi_0$.
\end{enumerate}
\end{defn}
It is now possible to define derived blow ups via a universal property.
\begin{defn}\label{DerivedBlowUpDef}
Let $Z \hookrightarrow X$ be a derived regular embedding. Then the \emph{derived blow up} $\mathrm{Bl}_Z(X)$ is the $X$-scheme representing virtual Cartier divisors lying over $Z$. In other words, given an $X$-scheme $S$, the space of $X$-morphisms
$$S \to \mathrm{Bl}_Z(X)$$
is naturally identified with the maximal sub $\infty$-groupoid of the $\infty$-category of virtual Cartier divisors of $S$ that lie over $Z$.
\end{defn}
\begin{thm}\label{BlowUpPropertiesThm}
Let $i: Z \hookrightarrow X$ be a derived regular embedding with $X$ Noetherian. Then
\begin{enumerate}
\item the derived blow up $\mathrm{Bl}_Z(X)$ exists and is unique up to contractible space of choices;
\item the structure morphism $\pi: \mathrm{Bl}_Z(X) \to X$ is projective, quasi-smooth, and induces an equivalence
$$\mathrm{Bl}_Z(X) - \mathcal{E} \to X - Z,$$
where $\mathcal{E}$ is the universal virtual Cartier divisor on $\mathrm{Bl}_Z(X)$ lying over $Z$ (also called the \emph{exceptional divisor});
\item the derived blow up $\mathrm{Bl}_Z(X) \to X$ is stable under derived base change;
\item the exceptional divisor $\mathcal{E}$ is naturally identified with $\mathbb{P}_Z(\mathcal{N}_{Z/X})$;
\item if $Z \stackrel i \hookrightarrow X \stackrel j \hookrightarrow Y$ is a sequence of quasi-smooth closed embeddings, then there is a natural derived regular embedding $\tilde j: \mathrm{Bl}_Z(X) \hookrightarrow \mathrm{Bl}_Z(Y)$ called the \emph{strict transform};
\item given derived regular embeddings $i: Z \hookrightarrow X$ and $j: Y \hookrightarrow X$, the strict transforms $\tilde i$ and $\tilde j$ do not meet in $\mathrm{Bl}_{Z \cap Y}(X)$;
\item if $Z$ and $X$ are classical schemes (so that $Z \hookrightarrow X$ is lci), there is a natural equivalence
$$\mathrm{Bl}_Z(X) \simeq \mathrm{Bl}_{Z}^\mathrm{cl}(X),$$
where the right hand side is the classical blow up.
\end{enumerate}
\end{thm}
\begin{proof}
The statements 1, 3, 4, 5 and 7 are directly from \cite{khan-rydh} Theorem 4.1.5. The second claim is otherwise from \emph{loc. cit.}, except that we claim $\pi$ to be projective and not just proper, which follows from the fact that $\mathcal{O}(-\mathcal{E})$ is $\pi$-ample. For a proof of 6, see for example \cite{annala-cob} Lemma 4.5.
\end{proof}
We will also make use of the following results.
\begin{prop}\label{BlowUpOfVectBundleSectionProp}
Let $X$ be a derived scheme, $E$ a vector bundle on $X$, $s$ a global section of $E$ and $Z$ the derived vanishing locus of $s$ on $X$. Then the natural short exact sequence
$$0 \to \mathcal{O}(-1) \to E \to Q \to 0$$
induces a natural equivalence
$$\abs{\Gamma(X; E)} \simeq \abs{\Gamma(\mathbb{P}(E); Q)}$$
of spaces of global sections sending $s$ to $\tilde s$, and the derived vanishing locus of $\tilde s$ inside $\mathbb{P}(E)$ is equivalent to $\mathrm{Bl}_Z(X)$ as an $X$-scheme. Moreover, $\mathcal{O}(-1)$ restricts to $\mathcal{O}(\mathcal{E})$ on $\mathrm{Bl}_Z(X)$.
\end{prop}
\begin{proof}
This is \cite{annala-chern} Proposition 2.7.
\end{proof}
\begin{prop}\label{BlowUpIsResolutionSchemeProp}
Let $X$ be a derived scheme, $E$ a vector bundle on $X$, $s$ a global section of $E$ and $Z$ the derived vanishing locus on $X$. Then $s: \mathcal{O} \to E$ factors through the natural injection
$$\mathcal{O}(\mathcal{E}) \hookrightarrow E$$
of vector bundles on $\mathrm{Bl}_Z(X)$.
\end{prop}
\begin{proof}
Indeed, since the lower row of the diagram
$$
\begin{tikzcd}
& \mathcal{O} \arrow[dashed]{ld}{} \arrow[]{d}{s} \arrow[]{rd}{\tilde s} & \\
\mathcal{O}(-1) \arrow[hook]{r} & E \arrow[twoheadrightarrow]{r} & Q
\end{tikzcd}
$$
is a cofibre sequence, a homotopy $\tilde s \sim 0$ induces in a natural way the dashed arrow, and therefore the result follows from Proposition \ref{BlowUpOfVectBundleSectionProp}.
\end{proof}
\subsection{Bivariant algebraic cobordism over $A$}
Let $A$ be a Noetherian ring of finite Krull dimension. The purpose of this section is to recall the bivariant algebraic $A$-cobordism from \cite{annala-pre-and-cob}. Let us recall the following definition.
\begin{defn}\label{FunctorialityDef}
A \emph{functoriality} is a tuple $\mathscr{F} = (\mathcal{C}, \mathscr{C}, \mathscr{I}, \mathscr{S})$, where
\begin{enumerate}
\item $\mathcal{C}$ is an $\infty$-category with a (distinguished) final object $pt$ and all fibre products;
\item $\mathscr{C}$ is a class of morphisms in $\mathcal{C}$ called \emph{confined morphisms} which contains all equivalences and is closed under compositions, pullbacks and homotopy equivalences of morphisms;
\item $\mathscr{I}$ is a class of Cartesian squares in $\mathcal{C}$ called \emph{independent squares} which contains all squares of form
\begin{center}
\begin{tikzcd}
X \arrow[]{r}{f} \arrow[]{d}{\mathrm{Id}_X} & Y \arrow[]{d}{\mathrm{Id}_Y} \\
X \arrow[]{r}{f} & Y
\end{tikzcd}
\ \ and \ \ \
\begin{tikzcd}
X \arrow[]{r}{\mathrm{Id}_X} \arrow[]{d}{f} & X \arrow[]{d}{f} \\
Y \arrow[]{r}{\mathrm{Id}_Y} & Y,
\end{tikzcd}
\end{center}
and is closed under vertical and horizontal compositions in the obvious sense as well as equivalences of Cartesian squares
\item $\mathscr{S}$ is a class of morphisms in $\mathcal{C}$ called \emph{specialized morphisms} which contains all equivalences and is closed under compositions and homotopy equivalences of morphisms.
\end{enumerate}
\end{defn}
Given a functoriality $\mathscr{F}$, one defines the notion of a bivariant theory with functoriality $\mathscr{F}$ as in \cite{annala-pre-and-cob} Section 2.1. Let us recall what it means for a bivariant theory to be additive.
\begin{defn}\label{AdditivityDef}
Suppose $\mathscr{F} = (\mathcal{C}, \mathscr{C}, \mathscr{I}, \mathscr{S})$ is such a functoriality that $\mathcal{C}$ has all finite coproducts and the canonical inclusions are confined. We say that a bivariant theory $\mathbb{B}^*$ with functoriality $\mathscr{F}$ is \emph{additive} if the morphisms
$$\mathbb{B}^*(X_1 \to Y) \oplus \mathbb{B}^*(X_2 \to Y) \xrightarrow{\iota_{1*} + \iota_{2*}} \mathbb{B}^*\Bigr(X_1 \coprod X_2 \to Y\Bigr)$$
are isomorphisms for all $X_1, X_2$ and $Y$, where $\iota_i$ is the canonical inclusion $X_i \to X_1 \coprod X_2$.
\end{defn}
We will denote by $\mathscr{F}_A$ the functoriality $(\mathcal{C}_A, \mathscr{C}_A, \mathscr{I}_A, \mathscr{S}_A)$, where $\mathcal{C}_A$ is the $\infty$-category of quasi-projective derived $A$-schemes, $\mathscr{C}_A$ consists of the projective morphisms, $\mathscr{I}_A$ consists of all derived Cartesian squares and $\mathscr{S}_A$ consists of quasi-smooth morphisms. Let us start by recalling the universal $A$-precobordism.
\begin{defn}\label{UnivAPrecobDef}
The \emph{universal $A$-precobordism theory} ${\underline\Omega}^*_A$ is the additive stably oriented bivariant theory with functoriality $\mathscr{F}_A$ defined as follows: given a homotopy class of morphisms $X \to Y$ in $\mathcal{C}$, the Abelian group ${\underline\Omega}^d_A(X \to Y)$ is generated by symbols $[V \to X]$ of equivalence classes of such projective morphisms $V \to X$ that the composition $V \to Y$ is quasi-smooth and of relative virtual dimension $-d$. These symbols are subjected to two sets of relations. First of all, taking disjoint unions is identified with summation (this enforces additivity). Secondly, given a projective morphism $W \to \mathbb{P}^1 \times X$ so that the composition $W \to \mathbb{P}^1 \times Y$ is quasi-smooth and of relative virtual dimension $-d$, and supposing that the fibre $W_\infty$ of $W \to \mathbb{P}^1$ over $\infty$ is a sum of virtual Cartier divisors $D_1$ and $D_2$, then
$$[W_0 \to X] = [D_1 \to X] + [D_1 \to X] - [\mathbb{P}_{D_1 \cap D_2}(\mathcal{O}(D_1) \oplus \mathcal{O}) \to X] \in {\underline\Omega}^d_A(X \to Y),$$
where $W_0$ is the fibre of $W \to \mathbb{P}^1$ over $0$, and where $\cap$ denotes the derived intersection.
Bivariant pushforward and pullback are defined in the obvious way using compositions of morphisms and derived pullbacks. The bivariant product is defined by bilinearly extending the following formula: given $[V \to X] \in {\underline\Omega}^a_A(X \to Y)$ and $[W \to Y] \in {\underline\Omega}^b_A(Y \to Z)$, form the derived Cartesian square
$$
\begin{tikzcd}
V' \arrow[]{d} \arrow[]{r} & X' \arrow[]{d} \arrow[]{r} & W \arrow[]{d} & \\
V \arrow[]{r} & X \arrow[]{r} & Y \arrow[]{r} & Z,
\end{tikzcd}
$$
and define
$$[V \to X] \bullet [W \to Y] := [V' \to X] \in {\underline\Omega}^{a + b}_A(X \to Z).$$
The orientation is given as follows: if $X \to Y$ is a quasi-smooth morphism, then $1_{X/Y} := [X \to X] \in {\underline\Omega}^*(X \to Y)$.
We will denote by ${\underline\Omega}^*_A(X)$ the universal $A$-precobordism rings ${\underline\Omega}^*_A(X \to X)$. Notice that the bivariant product has a simpler formula in this case:
$$[V \to X] \bullet [W \to X] = [V \times_X W \to X] \in {\underline\Omega}^*_A(X).$$
We will use shorthand notation for the unit $1_X := 1_{X/X} \in {\underline\Omega}^*(X)$. Let us also denote by ${\underline\Omega}^A_*(X)$ the universal $A$-prebordism groups ${\underline\Omega}^{-*}_A(X \to pt)$. Note that the bivariant product gives ${\underline\Omega}^A_*(X)$ the structure of an ${\underline\Omega}^*_A(X)$-module.
\end{defn}
Recall that given a line bundle $\mathscr{L}$ on $X \in \mathcal{C}_A$, we define the \emph{Euler class} (usually called the \emph{first Chern class}, but we refrain from using this terminology until we have constructed Chern classes in Section \ref{GenChernClassSubSect}) of $\mathscr{L}$ as
$$e(\mathscr{L}) := [Z_s \hookrightarrow X] \in {\underline\Omega}^*_A(X),$$
where $s$ is a global section of $\mathscr{L}$, and $Z_s$ is the derived vanishing locus of $s$. We recall that by \cite{annala-chern} Theorem 3.4 there is a formal group law
$$F_A(x,y) \in {\underline\Omega}^*_A(pt)[[x,y]]$$
such that for all line bundles $\mathscr{L}_1$ and $\mathscr{L}_2$ on $X \in \mathcal{C}_A$, the equality
$$e(\mathscr{L}_1 \otimes \mathscr{L}_2) = F_A\big(e(\mathscr{L}_1), e(\mathscr{L}_2)\big) \in {\underline\Omega}^*(X)$$
holds. Following \cite{annala-pre-and-cob}, we will define bivariant algebraic $A$-cobordism as a quotient of ${\underline\Omega}^*_A$. However, before doing so, we have to recall some terminology.
Suppose we are given the data of a virtual Cartier divisor $D \hookrightarrow W$, and an equivalence
$$D \simeq n_1 D_1 + \cdots + n_r D_r$$
of virtual Cartier divisors on $W$ with $n_i > 0$. Then, denoting by $+_{F_A}$ the formal addition given by the formal group law $F_A$ and by $[n]_{F_A} \cdot$ the formal multiplication (iterated formal addition), the formal power series
$$[n_1]_{F_A} \cdot x_1 +_{F_A} \cdots +_{F_A} [n_r]_{F_A} \cdot x_r$$
in $r$ variables has a unique expression of form
$$\sum_{I \subset \{1,2,...,r \}} \textbf{x}^I F^{n_1,...,n_r}_{A,I}(x_1,...,x_r),$$
where
$$\textbf{x}^I = \prod_{i \in I} x_i$$
and $F_{A, I}^{n_1,...,n_r}(x_1,...,x_r)$ contains only variables $x_i$ such that $i \in I$. Note that $F_{A, \emptyset}^{n_1,...,n_r}(x_1,...,x_r) = 0$. Using this notation, we make the following definition.
\begin{defn}\label{ZetaClassDef}
Let everything be as above. We define
$$
\zeta_{W,D,D_1,...,D_r} := \sum_{I \subset \{1,2,...,r\}} \iota^I_*\Biggl(F^{n_1,...,n_r}_{A,I} \Bigl( e\bigl(\mathcal{O}(D_1)\bigr), ...,e\bigl(\mathcal{O}(D_r)\bigr) \Bigr) \bullet 1_{D_I/pt} \Biggr) \in {\underline\Omega}^A_*(D),
$$
where $\iota^I$ is the inclusion $D_I \hookrightarrow D$ of the derived intersection $\cap_{i \in I} D_i$ inside $W$.
\end{defn}
In the most important special case, the composition is provided by the data of an $A$-snc divisor.
\begin{defn}\label{ASNCDivDef}
Let $W$ be a smooth quasi-projective $A$-scheme. Then an \emph{$A$-snc divisor} on $W$ is the data of an effective Cartier divisor $D \hookrightarrow W$, Zariski connected effective Cartier divisors $D_1,...,D_r \hookrightarrow W$ with
$$D_I := \cap_{i \in I} D_i$$
smooth and of the expected relative dimension over $A$ for all $I \subset \{1,2,...,r\}$, so that there exist positive integers $n_1,...,n_r$ such that
$$D \simeq n_1 D_1 + \cdots + n_r D_r$$
as effective Cartier divisors on $W$. Somewhat abusively, we will use the shorthand notation
$$\zeta_{W,D} := \zeta_{W,D, D_1, ..., D_r}$$
whenever $D \hookrightarrow W$ and $D_i$ form an $A$-snc divisor.
\end{defn}
We are now ready to recall the definition of bivariant algebraic cobordism over $A$.
\begin{defn}\label{BivACobDef}
\emph{Bivariant algebraic $A$-cobordism theory} is the additive stably oriented bivariant theory with functoriality $\mathscr{F}_A$ defined as the quotient
$$\Omega^*_A := {\underline\Omega}^*_A / \langle \mathcal{R}^\mathrm{snc}_A \rangle,$$
where $\mathcal{R}^\mathrm{snc}_A$ is the bivariant subset of \emph{snc relations} consisting of
$$\zeta_{W,D} - \theta(\pi_D) \in {\underline\Omega}^*_A(D \to pt),$$
where $D \hookrightarrow W$ ranges over all $A$-snc divisors, and $\langle \mathcal{R}^\mathrm{snc}_A \rangle$ is the bivariant ideal generated by $\mathcal{R}^\mathrm{snc}_A$.
\end{defn}
We recall that ${\underline\Omega}^*_A$ and $\Omega^*_A$ were proven to have many good properties in \cite{annala-cob, annala-yokura, annala-chern, annala-pre-and-cob}: if $k$ is a field of characteristic 0, the associated homology theory of $\Omega^*_k$ recovers Levine--Morel's algebraic bordism, the Grothendieck ring of vector bundles can be recovered from ${\underline\Omega}^*_A$ analogously to Conner--Floyd theorem, projective bundle formula holds for ${\underline\Omega}^*_A$, and ${\underline\Omega}^*_A$ can be used to construct a candidate for Chow-cohomology theory satisfying a generalization of Grothendieck--Riemann--Roch theorem. We will not recall the statements here, as we are going to give more general versions in Section \ref{GenChernClassSubSect}.
\section{Bivariant algebraic cobordism}\label{BivAlgCobSect}
The purpose of this section is to prove that bivariant algebraic $A$-cobordism groups are independent from the base ring $A$, and to define these groups more generally for morphisms of finite dimensional divisorial Noetherian derived schemes. In particular, we will not need to assume quasi-projectivity over a Noetherian base ring $A$ anymore.
The structure of this section is as follows: in Section \ref{BaseIndSubSect} we prove that $\Omega^*_A(X \to Y)$ is independent from the base ring $A$ by giving an alternative characterization for the bivariant ideal of ${\underline\Omega}^*_A$ generated by the snc relations. Using a slightly modified version of this alternative characterization, we define in Section \ref{BivCobConsSubsect} bivariant theories ${\underline\Omega}^\bullet$ and $\Omega^\bullet$ on the $\infty$-category of divisorial and finite dimensional Noetherian derived schemes which generalize the theories ${\underline\Omega}^*_A$ and $\Omega^*_A$ respectively. In Section \ref{NilpEulerClassSubSect} we analyze the potential failure of nilpotence for Euler classes of vector bundles on $X \in \mathcal{C}_a$, and define modified theories ${\underline\Omega}'^\bullet$ and $\Omega'^\bullet$ where Euler classes of vector bundles are provably nilpotent. Note that nilpotence of Euler classes is a fundamental ingredient in the proofs of Section \ref{GeneralPBFSect}, explaining why we go to such lengths to make sure that this property holds.
\subsection{Base independence of bivariant algebraic $A$-cobordism}\label{BaseIndSubSect}
The purpose of this section is to prove that if $X \to Y$ is a morphism of derived schemes with $X$ and $Y$ quasi-projective over Noetherian rings $A$ and $B$ of finite Krull dimension, then $\Omega^*_A(X \to Y)$ and $\Omega^*_B(X \to Y)$ are canonically isomorphic. While proving this, we also obtain an explicit construction of $\Omega^*_A(X \to Y)$ as a quotient of ${\underline\Omega}^*_A(X \to Y)$, without the need of considering either of these groups as parts of a bivariant theory.
We begin with the following trivial observation.
\begin{lem}
Suppose $X \to Y$ is a morphism of derived schemes with $X$ and $Y$ quasi-projective over Noetherian rings $A$ and $B$ of finite Krull dimension. Then the groups ${\underline\Omega}^*_A(X \to Y)$ and ${\underline\Omega}^*_B(X \to Y)$ are canonically isomorphic. \qed
\end{lem}
In order to obtain the desired result, we need to analyze the bivariant ideal $\langle \mathcal{R}^\mathrm{snc}_A \rangle$ used in Definition \ref{BivACobDef}. The main tool we are going to use are the following schemes.
\begin{cons}\label{VSchemeCons}
Let $n \geq 0$, and consider the scheme $\mathbb{P}^n \times \mathbb{P}^n \times \mathbb{P}^n \times \mathbb{P}^n$, where the projective spaces are over $\mathrm{Spec}(\mathbb{Z})$. Consider the vector bundle $E_n := \mathcal{O}(1,-1,0,0) \oplus \mathcal{O}(0,0,1,-1)$ over $(\mathbb{P}^n)^{\times 4}$, and notice that there are tautological sections
$$s_1: \mathcal{O}_{E_n} \to \mathcal{O}(1,-1,0,0)$$
and
$$s_2: \mathcal{O}_{E_n} \to \mathcal{O}(0,0,1,-1)$$
on $E_n$, so that the (derived) vanishing locus of $s_1$ is $\mathcal{O}(0,0,1,-1)$ and the vanishing locus of $s_2$ is $\mathcal{O}(1,-1,0,0)$. We define $\mathcal{V}^n$ as the vanishing locus of $s_1 s_2$ inside $E_n$. We also define $\mathcal{V}^n_i$ as the vanishing loci of $s_i$, and $\mathcal{V}^n_{12}$ as the vanishing locus of $s_1, s_2$; notice that there are (non-regular) closed embeddings $\iota^i: \mathcal{V}^n_i \hookrightarrow \mathcal{V}^n$ which give a closed cover of $\mathcal{V}^n$, and a (non-regular) closed embedding $\iota^{12}: \mathcal{V}^n_{12} \hookrightarrow \mathcal{V}^n$. We will also denote
$$\mathcal{N}_1 := \mathcal{N}_{\mathcal{V}^n_{12} / \mathcal{V}^n_1} \cong \mathcal{O}(0,0,1,-1) \vert_{\mathcal{V}^n_{12}}$$
and
$$\mathcal{N}_2 := \mathcal{N}_{\mathcal{V}^n_{12} / \mathcal{V}^n_2} \cong \mathcal{O}(1,-1,0,0) \vert_{\mathcal{V}^n_{12}}.$$
If $Y$ is a derived scheme, we define $\mathcal{V}^n_Y, \mathcal{V}^n_{Y,i}$ and $\mathcal{V}^n_{Y,12}$ as $Y \times \mathcal{V}^n$, $Y \times \mathcal{V}^n_i$ and $Y \times \mathcal{V}^n_{12}$ respectively.
\end{cons}
Before moving on, let us record the following easy observation.
\begin{lem}
Let $V$ and $Y$ be Noetherian derived schemes, and suppose $f: V \to \mathcal{V}^n_Y$ is a morphism. Then the following are equivalent
\begin{enumerate}
\item $f$ is quasi-smooth;
\item the maps $V_i := \mathcal{V}^n_{Y,i} \times_{\mathcal{V}^n_Y} V \to Y$ are quasi-smooth for $i=1,2$.
\end{enumerate}
\end{lem}
\begin{proof}
Since $\mathcal{V}^n_i$ cover $\mathcal{V}^n$, it is easy to check (say, using Proposition \ref{QSmCharProp}) that $f$ being quasi-smooth is equivalent to $V_i \to \mathcal{V}^n_{Y,i}$ being quasi-smooth for $i=1,2$. Since $\mathcal{V}^n_{Y,i} \to Y$ are smooth, the claim follows.
\end{proof}
We then consider the following bivariant ideal.
\begin{defn}\label{SNCIdealAltDef}
Let $A$ be a Noetherian ring of finite Krull dimension, and let $X \to Y$ be a morphism in $\mathcal{C}_A$. We will define a bivariant ideal $\mathcal{I}_A \subset {\underline\Omega}^*_A(X \to Y)$. Given a projective morphism $f: V \to X$ with the composition $V \to Y$ quasi-smooth and a quasi-smooth $Y$-morphism $V \to \mathcal{V}^n_Y$, we will denote by $V_i$ and $V_{12}$ the derived fibre products $\mathcal{V}^n_{Y,i} \times_{\mathcal{V}^n_Y} V$ and $\mathcal{V}^n_{Y,12} \times_{\mathcal{V}^n_Y} V$ respectively, and by $f^{12}$ the induced morphism $V_{12} \to X$. Given such data, consider the element
$$[V \to X] - [V_1 \to X] - [V_2 \to X] - f^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e(\mathcal{N}_{V_{12}/V_1})^{i-1} \bullet e(\mathcal{N}_{V_{12}/V_2})^{j-1} \bullet 1_{V_{12}/Y}\Bigg)$$
of ${\underline\Omega}^*_A(X \to Y)$, where $a_{ij}$ are the coefficients of the formal group law $F_A$ of ${\underline\Omega}^*_A$, and define $\mathcal{I}_A(X \to Y)$ as the subgroup of ${\underline\Omega}^*_A(X \to Y)$ generated by all such elements. The fact that $\mathcal{I}_A$ is a bivariant ideal follows trivially from the fact that projective and quasi-smooth morphisms are stable under pullbacks and compositions.
\end{defn}
Our goal is to prove that $\mathcal{I}_A = \langle \mathcal{R}^\mathrm{snc}_A \rangle$. The following lemma shows that $\langle \mathcal{R}^\mathrm{snc}_A \rangle \subset \mathcal{I}_A$.
\begin{lem}\label{LSRelationsContainedInILem}
Let $A$ be a Noetherian ring of finite Krull dimension. Suppose $X$ is a quasi-smooth and quasi-projective derived $A$-scheme, and let $D \hookrightarrow X$ be a virtual Cartier divisor. Then, if $D \simeq n_1 D_1 + \cdots + n_r D_r$ for some virtual Cartier divisors $D_i \hookrightarrow X$ and $n_i > 0$, the equality
$$1_{D/pt} = \zeta_{X,D,D_1,...,D_r} \in {\underline\Omega}^A_*(D) / \mathcal{I}_A(D \to pt)$$
holds, where $\zeta_{X,D,D_1,...,D_r}$ is as in Definition \ref{ZetaClassDef}.
\end{lem}
\begin{proof}
For simplicity we will denote by $x_i$ the Euler class $e(\mathcal{O}(D_i))$, by $+_{F_A}$ the formal addition and by $[n]_{F_A} \cdot$ the formal multiplication by $n$. Note that the claim is trivially true for $r=1$ and $n = n_1 = 1$. We will proceed by induction on $n$. If $D \simeq (n+1) D_1$, then $D$ is the derived vanishing locus of $s^{n+1}$, where $s \in \mathcal{O}(D_1)$. Since $X$ is quasi-projective over $A$, it is possible to find $d \geq 0$ and a morphism $X \to (\mathbb{P}^d)^{\times 4}$ such that $\mathcal{O}(1,-1,0,0)$ and $\mathcal{O}(0,0,1,-1)$ pull back to $\mathcal{O}(D_1)$ and $\mathcal{O}(n D_1)$ respectively. The global sections $s$ and $s^{n}$ then induce a morphism
$$X \to E_n.$$
so that the tautological sections $s_1$ and $s_2$ (see Construction \ref{VSchemeCons}) pull back to $s$ and $s^{n}$ respectively. This also induces a quasi-smooth morphism $D \to \mathcal{V}^d_A$ with
$$
\begin{tikzcd}
D_1 \arrow[hook]{r}{\iota^1} \arrow[]{d} & D \arrow[]{d} & \arrow[hook']{l}[swap]{\iota^2} n D_1 \arrow[]{d} \\
\mathcal{V}^d_{A,1} \arrow[hook]{r} & \mathcal{V}^d_A & \arrow[hook']{l} \mathcal{V}^d_{A,2}
\end{tikzcd}
$$
homotopy Cartesian. In particular
\begin{align*}
[D \to D] &= [D_1 \to D] + [n D_1 \to D] \\
&+ \iota^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e\big(\mathcal{O}(D_1)\big)^{i-1} \bullet e\big(\mathcal{O}(n D_1)\big)^{j-1} \bullet 1_{D_1 \cap n D_1 / pt} \Bigg)
\end{align*}
in $\Omega^*\big(D \to \mathrm{Spec}(A)\big)$, where $\iota^{12}$ is the embedding $D_1 \cap (n-1) D_1 \hookrightarrow D$. Applying the inductive assumption, we compute that
$$[n D_1 \to D] = \iota^1_*\big(F^n_{A,\{1\}} (x_1) \bullet 1_{D_1 / pt}\big)$$
and
\begin{align*}
& \iota^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e\big(\mathcal{O}(D_1)\big)^{i-1} \bullet e\big(\mathcal{O}(n D_1)\big)^{j-1} \bullet 1_{D_1 \cap n D_1 / pt} \Bigg) \\
=& \iota^{2}_* \Bigg( \sum_{i,j \geq 1} a_{ij} x_1^i \bullet ([n]_{F_A} \cdot x_1)^{j-1} \bullet 1_{n D_1 / pt} \Bigg) \\
=& \iota^{1}_* \Bigg( \sum_{i,j \geq 1} a_{ij} x_1^i \bullet ([n]_{F_A} \cdot x_1)^{j-1} \bullet F^n_{A,\{1\}} (x_1) \bullet 1_{D_1/pt} \Bigg).
\end{align*}
Combining the above, we see that
\begin{align*}
[D \to D] &= \iota^{1}_* \Bigg( \Big(1_{D_1/pt} + F^n_{A,\{1\}} (x_1) + \sum_{i,j \geq 1} a_{ij} x_1^i \bullet ([n]_{F_A} \cdot x_1)^{j-1} \bullet F^n_{A,\{1\}} (x_1)\Big) \bullet 1_{D_1/pt} \Bigg) \\
&= \iota^{1}_* \Big( F^{n+1}_{A,\{1\}} (x_1) \bullet 1_{D_1/pt} \Big),
\end{align*}
where the last equality follows from the fact that $x F^m_{A,\{1\}} (x) = [m]_{F_A} \cdot x$ for all $m > 0$.
Let us then proceed by induction on $r$. Similarly to the argument above, we can find a quasi-smooth morphism $D \to \mathcal{V}^d_A$ giving us the derived Cartesian diagram
$$
\begin{tikzcd}
n_1 D_1 \arrow[]{d} \arrow[hook]{r}{\iota^1} & D \arrow[]{d} & \arrow[hook']{l}[swap]{\iota^2} D' \arrow[]{d} \\
\mathcal{V}^d_{A,1} \arrow[hook]{r} & \mathcal{V}^d_{A} & \arrow[hook']{l} \mathcal{V}^d_{A,2},
\end{tikzcd}
$$
where $D' = n_2 D_2 + \cdots + n_r D_r$. Therefore
\begin{align*}
[D \to D] &= [n_1 D_1 \to D] + [D' \to D] \\
&+ \iota^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e\big(\mathcal{O}(n_1 D_1)\big)^{i-1} \bullet e\big(\mathcal{O}(D')\big)^{j-1} \bullet 1_{n_1 D_1 \cap D'} \Bigg)
\end{align*}
in $\Omega^*\big(D \to \mathrm{Spec}(A) \big)$, where $\iota^{12}$ is the embedding $n_1 D_1 \cap D' \hookrightarrow D$. We can then inductively compute that
\begin{align*}
[n_1 D_1 \to D] &= \iota^{\{1\}}_*\big(F^{n_1,...,n_r}_{A,\{1\}}(x_1) \bullet 1_{D_1/pt} \big) \\
[D' \to D] &= \sum_{I \subset \{2,...,r\}} \iota^I_*\big(F^{n_1,...,n_r}_{A,I}(x_2,...,x_r) \bullet 1_{D_I / pt} \big)
\end{align*}
and
\begin{align*}
&\iota^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e\big(\mathcal{O}(n_1 D_1)\big)^{i-1} \bullet e\big(\mathcal{O}(D')\big)^{j-1} \bullet 1_{n_1 D_1 \cap D' / pt} \Bigg) \\
=& \iota^{2}_* \Bigg( \sum_{i,j \geq 1} a_{ij} ([n_1]_{F_A} \cdot x_1)^{i} \bullet e\big(\mathcal{O}(D')\big)^{j-1} \bullet 1_{D' / pt} \Bigg) \\
=& \sum_{I \subset \{2,...,r\}} \iota^I_*\Bigg(\sum_{i,j \geq 1} a_{ij} ([n_1]_{F_A} \cdot x_1)^{i} \bullet e\big(\mathcal{O}(D')\big)^{j-1} \bullet F^{n_1,...,n_r}_{A,I}(x_2,...,x_r) \bullet 1_{D_I / pt} \Bigg).
\end{align*}
Combining the above lets us compute that
\begin{align*}
&[D \to D] \\
=& \iota^{\{1\}}_*\big(F^{n_1,...,n_r}_{A,\{1\}}(x_1) \bullet 1_{D_1/pt} \big) + \sum_{I \subset \{2,...,r\}} \iota^I_*\big(F^{n_1,...,n_r}_{A,I}(x_2,...,x_r) \bullet 1_{D_I / pt} \big) \\
+& \sum_{I \subset \{2,...,r\}} \iota^I_*\Bigg(\sum_{i,j \geq 1} a_{ij} ([n_1]_{F_A} \cdot x_1)^{i} \bullet e\big(\mathcal{O}(D')\big)^{j-1} \bullet F^{n_1,...,n_r}_{A,I}(x_2,...,x_r) \bullet 1_{D_I / pt} \Bigg) \\
=& \sum_{I \subset \{2,...,r\}} \iota^I_*\big(F^{n_1,...,n_r}_{A,I}(x_2,...,x_r) \bullet 1_{D_I / pt} \big) \\
+& \sum_{I \subset \{2,...,r\}} \iota^{\{1\} \cup I}_*\Big( F^{n_1,...,n_r}_{A,\{1\} \cup I}(x_1,...,x_r) \bullet 1_{D_{\{1\} \cup I} / pt} \Big),
\end{align*}
where the last equality follows from the fact that
$$e\big(\mathcal{O}(D')\big) = [n_2]_{F_A} \cdot x_2 +_{F_A} \cdots +_{F_A} [n_r]_{F_A} \cdot x_r$$
and the definition of $F_{A,I}^{n_1,...,n_r}.$ Hence
$$[D \to D] = \sum_{I \subset \{1,...,r\}} \iota^I_*\big(F^{n_1,...,n_r}_{A,I}(x_1,...,x_r) \bullet 1_{D_I / pt} \big)$$
and we are done.
\end{proof}
The following lemma shows that $\mathcal{I}_A \subset \langle \mathcal{R}^\mathrm{snc}_A \rangle$.
\begin{lem}\label{IContainedInLSRelationsLem}
Let $A$ be a Noetherian ring of finite Krull dimension. Then $\mathcal{I}_A = \langle \mathcal{R}^{\mathcal{V}}_A \rangle$, where $\mathcal{R}^{\mathcal{V}}_A$ is the bivariant subset of ${\underline\Omega}^*_A$ consisting of
$$
1_{\mathcal{V}^n_A / pt} - [\mathcal{V}^n_{A,1} \to \mathcal{V}^n_{A}] - [\mathcal{V}^n_{A,2} \to \mathcal{V}^n_{A}] - \iota^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e(\mathcal{N}_1)^{i-1} \bullet e(\mathcal{N}_2)^{j-1} \bullet 1_{\mathcal{V}^n_{A,12}/pt}\Bigg)
$$
in ${\underline\Omega}^A_*(\mathcal{V}^n_A)$.
\end{lem}
\begin{proof}
Clearly $\langle \mathcal{R}^{\mathcal{V}}_A \rangle$ contains elements
$$s_Y = 1_{\mathcal{V}^n_Y / Y} - [\mathcal{V}^n_{Y,1} \to \mathcal{V}^n_{Y}] - [\mathcal{V}^n_{Y,2} \to \mathcal{V}^n_{Y}] - \iota^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e(\mathcal{N}_1)^{i-1} \bullet e(\mathcal{N}_2)^{j-1} \bullet 1_{\mathcal{V}^n_{Y,12}/Y}\Bigg)$$
in ${\underline\Omega}^*_A(\mathcal{V}^n_Y \to Y)$, because bivariant ideals are stable under bivariant pullbacks. Therefore, if we are given a projective morphism $f: V \to X$ so that the composition $V \to Y$ is quasi-smooth, and a quasi-smooth $Y$-morphism $V \to \mathcal{V}^n_Y$, we can compute that
\begin{align*}
f_* \big(1_{V / \mathcal{V}^n_Y} \bullet s_Y\big) &= [V \to X] - [V_1 \to X] - [V_2 \to X] \\
&- f^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e(\mathcal{N}_{V_{12}/V_1})^{i-1} \bullet e(\mathcal{N}_{V_{12}/V_2})^{j-1} \bullet 1_{V_{12}/Y}\Bigg)
\end{align*}
is contained in $\langle \mathcal{R}^{\mathcal{V}}_A \rangle$, proving the claim.
\end{proof}
Combining Lemmas \ref{LSRelationsContainedInILem} and \ref{IContainedInLSRelationsLem}, we obtain the following result.
\begin{thm}\label{BivACobAltConsThm}
Let $A$ be a Noetherian ring of finite Krull dimension. Then the bivariant ideals $\mathcal{I}_A \subset {\underline\Omega}^*_A$ and $\langle \mathcal{R}^\mathrm{snc}_A \rangle \subset {\underline\Omega}^*_A$ coincide. \qed
\end{thm}
As a corollary, we get the desired independence of bivariant algebraic cobordism groups from the base ring.
\begin{cor}\label{BivACobIndFromACor}
Let $X \to Y$ be a morphism of derived schemes with $X$ and $Y$ quasi-projective over Noetherian rings $A$ and $B$ of finite Krull dimension. Then
$$\Omega^*_A(X \to Y) = \Omega^*_B(X \to Y).$$
\end{cor}
\begin{proof}
Clearly $\mathcal{I}_A(X \to Y) = \mathcal{I}_B(X \to Y)$, so the claim follows from Theorem \ref{BivACobAltConsThm}.
\end{proof}
\subsection{Algebraic cobordism for divisorial derived schemes}\label{BivCobConsSubsect}
In this section, we define bivariant algebraic cobordism in great generality. We will denote by $\mathscr{F}_a$ the functoriality $(\mathcal{C}_a, \mathscr{C}_a, \mathscr{I}_a, \mathscr{S}_a)$, where $\mathcal{C}_a$ is the full subcategory of the $\infty$-category derived schemes consisting of finite dimensional Noetherian derived schemes having an ample family of line bundles, $\mathscr{C}_a$ consists of projective morphisms in $\mathcal{C}_a$, $\mathscr{I}_a$ consists of all derived Cartesian squares in $\mathcal{C}_a$ and $\mathscr{S}_a$ consists of all morphisms in $\mathcal{C}_a$ whose relative cotangent complex has Tor-dimension 1. Consider the following universal theory defined by Yokura in \cite{yokura09} (see also \cite{annala-pre-and-cob} Section 2.2 and the beginning of Section 3).
\begin{defn}
Let us define the \emph{universal additive bivariant theory} $\mathcal{M}^\bullet_{\mathcal{F}_a, +}$ as the stably oriented bivariant theory so that $\mathcal{M}^\bullet_{\mathcal{F}_a, +}(X \to Y)$ is the group completion of the Abelian monoid on symbols
$$[V \to X]$$
on equivalence classes of projective morphisms $V \to Y$ with $\mathbb{L}_{V/Y}$ of Tor-dimension $1$, and where the monoid operation is given by disjoint union. The bivariant operations and the orientation is defined as in Definition \ref{UnivAPrecobDef}.
\end{defn}
\begin{rem}
The careful reader must have noticed that we did not define a grading on the bivariant groups $\mathcal{M}^\bullet_{\mathcal{F}_a, +}(X \to Y)$. This is because if $V \to Y$ is a general morphism of Noetherian derived schemes, then $\mathbb{L}_{V/Y}$ having Tor-dimension one does not imply that it is perfect, implying that the virtual rank of $\mathbb{L}_{V/Y}$ need not to make sense, and therefore we do not have a sensible notion of relative virtual dimension. Of course, if $V \to Y$ happens to be of finite type, then $\mathbb{L}_{V/Y}$ is perfect by \ref{QSmCharProp}, so the problem disappears. Hence, if $X \to Y$ is a finite type morphism in $\mathcal{C}_a$, we can define a grading on $\mathcal{M}^\bullet_{\mathcal{F}_a, +}(X \to Y)$
by saying that $[V \to X]$ has degree $d$ if $V \to Y$ has relative virtual dimension $-d$. We will denote the graded Abelian groups thus obtained by $\mathcal{M}_{\mathcal{F}_a, +}^*(X \to Y)$. In particular, the associated cohomology rings are graded rings. Similar remarks apply to the bivariant theories ${\underline\Omega}^\bullet$ and $\Omega^\bullet$ defined below.
\end{rem}
Let us recall the following universal property from \cite{yokura09}.
\begin{thm}
$\mathcal{M}^\bullet_{\mathcal{F}_a, +}$ is the universal stably oriented additive bivariant theory with functoriality $\mathscr{F}_a$. In other words, if $\mathbb{B}^\bullet$ is a stably oriented additive bivariant theory with functoriality $\mathscr{F}_a$, then there exists a unique orientation preserving Grothendieck transformation
$$\mathcal{M}^\bullet_{\mathcal{F}_a, +} \to \mathbb{B}^\bullet.$$
\end{thm}
We can then define the universal precobordism theory.
\begin{defn}\label{UnivPrecobDef}
We define the \emph{universal precobordism theory} ${\underline\Omega}^\bullet$ as a quotient of $\mathcal{M}^\bullet_{\mathcal{F}_a, +}$ by enforcing the derived double point cobordism relations: given a projective morphism $W \to \mathbb{P}^1 \times X$ so that $\mathbb{L}_{W / \mathbb{P}^1 \times Y}$ has Tor-dimension 1, and virtual Cartier divisors $D_1$ and $D_2$ so that the fibre $W_\infty$ of $W \to \mathbb{P}^1$ over $\infty$ is the sum of $D_1$ and $D_2$, then
$$[W_0 \to X] = [D_1 \to X] + [D_1 \to X] - [\mathbb{P}_{D_1 \cap D_2}(\mathcal{O}(D_1) \oplus \mathcal{O}) \to X] \in {\underline\Omega}^\bullet(X \to Y),$$
where $W_0$ is the fibre of $W \to \mathbb{P}^1$ over $0$, and where $\cap$ denotes the derived intersection. It is straightforward to check that these relations respect the bivariant operations.
\end{defn}
Recall that the \emph{Euler class} $e(\mathscr{L})$ of a line bundle $\mathscr{L}$ on $X \in \mathcal{C}_a$ is defined as $[Z_s \hookrightarrow X] \in {\underline\Omega}^1(X)$, where $s$ is any global section of $\mathscr{L}$ and $Z_s$ is the derived vanishing locus of $s$. Before going further, we need to make the following observation.
\begin{lem}\label{NilpChernLem}
Suppose $X \in \mathcal{C}_a$ and let $\mathscr{L}$ be a a line bundle on $X$. Then the Euler class $e(\mathscr{L}) \in {\underline\Omega}^1(X)$ is nilpotent.
\end{lem}
\begin{proof}
Choose line bundles $\mathscr{L}_1, ..., \mathscr{L}_r$ and sections $s_i \in \Gamma(X; \mathscr{L}_i)$ so that the $X_{s_i}$ give an affine open cover of $X$. Using Lemma \ref{LocLem}, for $N \gg 0$ and $n \gg 0$ it is possible to find $n$ global sections $s'_{i,1}, ..., s'_{i,n}$ of $\mathscr{L}_i^{\otimes N} \otimes \mathscr{L}$ such that the intersection of the restriction of their derived vanishing loci to $X_{s_i}$ is empty, i.e., $e(\mathscr{L}_i^{\otimes N} \otimes \mathscr{L})^{n}$ are represented by cobordism cycles supported outside $X_{s_i}$.
We can use Lemma \ref{GeomFGLLem} below to compute that
$$e(\mathscr{L}) = e(\mathscr{L}_i^{\otimes N} \otimes \mathscr{L}) + \alpha_i \bullet e(\mathscr{L}_i^{\otimes -N})$$
and
$$e(\mathscr{L}_i^{\otimes -N}) = \beta_i \bullet e(\mathscr{L}_i^{\otimes N})$$
so that
$$e(\mathscr{L}) = e(\mathscr{L}_i^{\otimes N} \otimes \mathscr{L}) + \gamma_i \bullet e(\mathscr{L}_i^{\otimes N}) \in {\underline\Omega}^*(X).$$
Hence
\begin{align*}
e(\mathscr{L})^{rn} &= \prod_{i=1}^r \big( e(\mathscr{L}_i^{\otimes N} \otimes \mathscr{L}) + \gamma_i \bullet e(\mathscr{L}_i^{\otimes N}) \big)^n \\
&= \prod_{i=1}^r \Bigg( \sum_{j=0}^n e(\mathscr{L}_i^{\otimes N} \otimes \mathscr{L})^j \bullet \gamma_i^{n-j} \bullet e(\mathscr{L}_i^{\otimes N})^{n-i} \Bigg),
\end{align*}
and as the $i^{th}$ part of the product can clearly be represented by bordism cycles supported outside $X_{s_i}$, the product must vanish for the simple reason that $X_{s_i}$ cover $X$. This proves that $e(\mathscr{L})$ is nilpotent.
\end{proof}
\begin{lem}\label{GeomFGLLem}
Let $X \in \mathcal{C}_a$ and let $\mathscr{L}_1$ and $\mathscr{L}_2$ be line bundles on $X$. Then
\begin{align*}
e(\mathscr{L}_1 \otimes \mathscr{L}_2) &= e(\mathscr{L}_1) + e(\mathscr{L}_2) - e(\mathscr{L}_1) \bullet e(\mathscr{L}_2) \bullet [\mathbb{P}_1 \to X] \\
&- e(\mathscr{L}_1) \bullet e(\mathscr{L}_2) \bullet e(\mathscr{L}_1 \otimes \mathscr{L}_2) \bullet \big( [\mathbb{P}_2 \to X] - [\mathbb{P}_3 \to X] \big)
\end{align*}
in ${\underline\Omega}^1(X)$, where
\begin{align*}
\mathbb{P}_1 &= \mathbb{P}_X(\mathscr{L}_1 \oplus \mathcal{O}); \\
\mathbb{P}_2 &= \mathbb{P}_X(\mathscr{L}_1 \oplus (\mathscr{L}_1 \otimes \mathscr{L}_2) \oplus \mathcal{O}); \\
\mathbb{P}_3 &= \mathbb{P}_{\mathbb{P}_X(\mathscr{L}_1 \oplus (\mathscr{L}_1 \otimes \mathscr{L}_2))}(\mathcal{O}(1) \oplus \mathcal{O}).
\end{align*}
\end{lem}
\begin{proof}
See \cite{levine-pandharipande} Section 5.2 for the original proof for algebraic schemes in characteristic 0 or \cite{annala-yokura} Lemma 6.2 for a proof in derived setting.
\end{proof}
Now that we know that the Euler classes of line bundles are nilpotent, we may prove the following result using the arguments of \cite{annala-chern} Section 3.
\begin{thm}\label{PCobFGLThm}
${\underline\Omega}^\bullet$ satisfies the formal group law axiom: in other words Euler classes of line bundles are nilpotent and there exists a formal group law
$$F(x,y) \in {\underline\Omega}^*\big(\mathrm{Spec}(\mathbb{Z})\big)[[x,y]]$$
so that for all $X \in \mathcal{C}_a$ and for all line bundles $\mathscr{L}_1$ and $\mathscr{L}_2$ on $X$, the equality
$$e(\mathscr{L}_1 \otimes \mathscr{L}_2) = F\big(e(\mathscr{L}_1), e(\mathscr{L}_2)\big) \in {\underline\Omega}^1(X)$$
holds. Moreover, since $\mathscr{F}_a$ is admissible and ${\underline\Omega}^\bullet$ satisfies the section axiom (see \cite{annala-pre-and-cob}), the formal group law $F$ is unique. \qed
\end{thm}
Moreover, following \cite{annala-pre-and-cob} Section 3.2 and \cite{annala-chern} Section 3.2 one proves the following universal property for ${\underline\Omega}$.
\begin{thm}\label{UnivPropOfPCobThm}
${\underline\Omega}^\bullet$ is the universal stably oriented additive bivariant theory with functoriality $\mathscr{F}_a$ satisfying the section and the formal group law axioms (as defined in \cite{annala-pre-and-cob}). In other words, if $\mathbb{B}^\bullet$ is another such a bivariant theory, then there exists a unique orientation preserving Grothendieck transformation ${\underline\Omega}^\bullet \to \mathbb{B}^{\bullet}.$ \qed
\end{thm}
Before constructing the bivariant theory $\Omega^\bullet$ as a quotient of ${\underline\Omega}^\bullet$, we consider the following stacky version of Construction \ref{VSchemeCons}.
\begin{cons}\label{VStackCons}
Consider the $\mathbb{Z}$-stack $[\mathbb{A}^1/ \mathbb{G}_m] \times [\mathbb{A}^1/ \mathbb{G}_m]$, i.e., the stack classifying the data $(\mathscr{L}_1, s_1, \mathscr{L}_2, s_2)$, where $\mathscr{L}_i$ are line bundles and $s_i$ are global sections of $\mathscr{L}_i$. Then we define the stack $\mathcal{V}$ as the derived vanishing locus of $s_1 s_2$ inside $[\mathbb{A}^1/ \mathbb{G}_m] \times [\mathbb{A}^1/ \mathbb{G}_m]$. The derived vanishing loci of $s_i$ give (non-regular) closed immersions $\iota^i: \mathcal{V}_i \hookrightarrow \mathcal{V}$, and the vanishing locus of $s_1$ and $s_2$ gives a (non-regular) closed embedding $\iota^{12}: \mathcal{V}_{12} \hookrightarrow \mathcal{V}$. We will also denote
$$\mathcal{N}_1 := \mathcal{N}_{\mathcal{V}_{12} / \mathcal{V}_1}$$
and
$$\mathcal{N}_2 := \mathcal{N}_{\mathcal{V}_{12} / \mathcal{V}_2}.$$
Given a derived scheme $Y$, we will denote by $\mathcal{V}_Y$, $\mathcal{V}_{Y,i}$ and $\mathcal{V}_{Y,12}$ the corresponding stacks over $Y$.
\end{cons}
Armed with these stacks, we can make the following definition.
\begin{defn}\label{BivCobDef}
We define \emph{bivariant algebraic cobordism} $\Omega^\bullet$ as a quotient of ${\underline\Omega}^\bullet$ by enforcing the following relations: let $X \to Y$ be a morphism in $\mathcal{C}_a$, let $f: V \to X$ be a projective morphism with $\mathbb{L}_{V/Y}$ of Tor-dimension one, and let $V \to \mathcal{V}^n_Y$ be a $Y$-morphism with $\mathbb{L}_{V / \mathcal{V}^n_Y}$ of Tor-dimension 1. We will denote by $V_i$ and $V_{12}$ the derived fibre products $\mathcal{V}_{Y,i} \times_{\mathcal{V}_Y} V$ and $\mathcal{V}_{Y,12} \times_{\mathcal{V}_Y} V$ respectively, and by $f^{12}$ the induced morphism $V_{12} \to X$. Given such data, we enforce the relation
$$[V \to X] = [V_1 \to X] + [V_2 \to X] + f^{12}_* \Bigg( \sum_{i,j \geq 1} a_{ij} e(\mathcal{N}_{V_{12}/V_1})^{i-1} \bullet e(\mathcal{N}_{V_{12}/V_2})^{j-1} \bullet 1_{V_{12}/Y}\Bigg)$$
in $\Omega^\bullet(X \to Y)$, where $a_{ij}$ are the coefficients of the formal group law $F$. It is straightforward to check that these relations respect the bivariant operations.
\end{defn}
Before continuing, we have to make sure that $\Omega^\bullet$ as defined above really extends $\Omega^*_A$ defined earlier.
\begin{lem}
Let $A$ be a Noetherian ring of finite Krull dimension, and let $X \to Y$ be a morphism of quasi-projective derived $A$-schemes. Then
$$\Omega^*(X \to Y) = \Omega^*_A(X \to Y).$$
\end{lem}
\begin{proof}
The issue is that $\Omega^*(X \to Y)$ was defined using the stack $\mathcal{V}_Y$ whereas $\Omega^*_A(X \to Y)$ was defined using the schemes $\mathcal{V}^n_Y$. However, since any line bundle on a quasi-projective derived scheme is of form $\mathscr{L}_1 \otimes \mathscr{L}_2^\vee$ with $\mathscr{L}_i$ globally generated, it follows that any morphism $X \to \mathcal{V}_Y$ factors through the canonical morphism $\mathcal{V}^n_Y \to \mathcal{V}_Y$ for $n \gg 0$, and therefore $\Omega^*(X \to Y)$ and $\Omega^*_A(X \to Y)$ have the same relations.
\end{proof}
As in \cite{annala-pre-and-cob}, the following result is an immediate consequence of Theorem \ref{UnivPropOfPCobThm}.
\begin{cor}\label{UnivPropOfCobCor}
$\Omega^\bullet$ is the universal stably oriented additive bivariant theory with functoriality $\mathscr{F}_a$ satisfying the section, the formal group law and the snc axioms (as defined in \cite{annala-pre-and-cob}). In other words, if $\mathbb{B}^\bullet$ is another such a bivariant theory, then there exists a unique orientation preserving Grothendieck transformation ${\underline\Omega}^\bullet \to \mathbb{B}^\bullet.$ \qed
\end{cor}
Bivariant algebraic cobordism in this generality allows us to define the corresponding absolute homology theory.
\begin{defn}\label{AlgBordDef}
Let $X \in \mathcal{C}_a$. We define the \emph{algebraic bordism} of $X$ as the Abelian group
$$\Omega_\bullet(X) := \Omega^\bullet \big(X \to \mathrm{Spec}(\mathbb{Z})\big).$$
We do not define a grading on $\Omega_\bullet(X)$.
\end{defn}
\begin{rem}
By Proposition \ref{DCompIntAltCharProp}, we see that algebraic bordism $\Omega_\bullet(X)$ classifies derived complete intersection schemes projective over $X$ up to ``cobordism''. A natural grading would be given by the virtual dimension of the cycles, but unfortunately this is not well defined in general, as noted in Remark \ref{DimProblemRem}. However, at least when $X$ is of finite type over a field $k$, the grading by virtual dimension makes sense. We will denote the graded Abelian groups obtained this way by $\Omega_*(X)$. If $X$ is quasi-projective over $k$, then these graded groups agree with the ones defined by Lowrey and Schürg in \cite{lowrey-schurg} by the results of \cite{annala-pre-and-cob}. In the special case where $k$ has characteristic 0, $\Omega_*(X)$ is canonically isomorphic to the algebraic bordism group $\Omega_*^\mathrm{LM}(X_\mathrm{cl})$ of Levine--Morel.
\end{rem}
\iffalse
The following results are proven exactly as in \cite{annala-chern}.
\begin{thm}\label{ChernClassThm}
Suppose $X \in \mathcal{C}_a$ admits an ample line bundle, and suppose $E$ is a vector bundle of rank $r$ on $X$. Then there exist classes
$$c_i(E) \in {\underline\Omega}^i(X)$$
for $1 \leq i \leq r$ called the \emph{Chern classes}, and the \emph{total Chern class}
$$c(E) := 1 + c_1(E) + \cdots + c_r(E),$$
which satisfy the following properties:
\begin{enumerate}
\item \emph{naturality in pullbacks:} $f^* c_i(E) = c_i(f^*E)$ whenever $f: Y \to X$ is such that $Y$ admits an ample line bundle;
\item \emph{Whitney sum formula:} given a short exact sequence
$$0 \to E' \to E \to E'' \to 0$$
of vector bundles on $X$, we have
$$c(E) = c(E') \bullet c(E'') \in {\underline\Omega}^*(X);$$
\item \emph{normalization:} the top Chern class $c_r(E)$ is equal to the \emph{Euler class}
$$e(E) := [Z_s \hookrightarrow X] \in {\underline\Omega}^r(X),$$
where $s$ is a global section of $E$ and $Z_s$ is the derived vanishing locus of $s$. \qed
\end{enumerate}
\end{thm}
Armed with a good theory of Chern classes, one can proof the Conner--Floyd theorem.
\begin{thm}\label{CohCFThm}
Let $\mathbb{Z}_m$ be the integers considered as an $\mathbb{L}$-algebra via the formal group law $x + y - xy$, and suppose $X \in \mathcal{C}_a$ admits an ample line bundle. Then there is a natural sequence of isomorphism
$$\mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X) \stackrel \cong \to \mathbb{Z}_m \otimes_\mathbb{L} \Omega^*(X) \stackrel \cong \to K^0(X),$$
where the second map is given by $[V \xrightarrow{f} X] \mapsto [f_* \mathcal{O}_V]$.
\end{thm}
\begin{proof}
The arguments from \cite{annala-chern} prove that $\mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X) \to K^0(X)$ is an isomorphism. However, since this clearly factors through $\mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X) \to \mathbb{Z}_m \otimes_\mathbb{L} \Omega^*(X)$, the claim follows.
\end{proof}
One can also prove the projective bundle formula.
\begin{thm}\label{PBFThm}
Let $X \to Y$ be a morphism in $\mathcal{C}_a$, suppose $X$ admits an ample line bundle and let $E$ be a vector bundle of rank $r$ on $X$. Then we get an isomorphism
$$\bigoplus_{i=0}^{r-1} {\underline\Omega}^\bullet(X \to Y) \to {\underline\Omega}^\bullet(\mathbb{P}(E) \to Y)$$
where the $i^{th}$ map is given by
$$e\big(\mathcal{O}(1)\big)^i \bullet [\mathbb{P}(E) \to X] \bullet -: {\underline\Omega}^\bullet(X \to Y) \to {\underline\Omega}^\bullet(\mathbb{P}(E) \to Y).$$
The result remains true for $\Omega$. \qed
\end{thm}
Let then $\mathbb{Z}_a$ be the integers considered as an $\mathbb{L}$-algebra via the additive formal group law $x + y$, and let ${\underline\Omega}^\bullet_a$ and $\Omega^\bullet_a$ be the bivariant theories defined as $\mathbb{Z}_a \otimes_\mathbb{L} {\underline\Omega}^\bullet$ and $\mathbb{Z}_a \otimes_\mathbb{L} \Omega^\bullet$ respectively. These give candidates for bivariant intersection theory, satisfying the following cohomological Grothendieck--Riemann--Roch theorem.
\begin{thm}\label{GeneralRR}
Suppose $X \in \mathcal{C}_a$ admits an ample line bundle. Then there exists Chern character morphism
$$ch_a: K^0(X) \to \mathbb{Q} \otimes {\underline\Omega}_a^*(X) \cong \mathbb{Q} \otimes \Omega_a^*(X),$$
which is a ring homomorphisms and commutes with pullbacks. Moreover, given $f: Y \to X$ quasi-smooth and projective, we have that
$$
f_! \bigl( ch_a(\alpha) \bullet \mathrm{Td}(\mathbb{L}_{Y/X}) \bigr) = ch_a(f_!(\alpha))
$$
for all $\alpha \in K^0(Y)$, where $\mathrm{Td}(\mathbb{L}_{Y/X})$ denotes the Todd class of $\mathbb{L}_{Y/X}$. Finally, the induced map
$$ch_a: \mathbb{Q} \otimes K^0(X) \to \mathbb{Q} \otimes {\underline\Omega}_a^*(X)$$
is an isomorphism.
\end{thm}
\begin{proof}
Again, the only thing to check is that $\mathbb{Q} \otimes {\underline\Omega}_a^*(X)$ and $\mathbb{Q} \otimes \Omega_a^*(X)$ are isomorphic. But using Theorem \ref{CohCFThm}, the arguments of \cite{annala-chern} show that
$$ch_a: \mathbb{Q} \otimes K^0(X) \to \mathbb{Q} \otimes \Omega_a^*(X)$$
is an isomorphism, and the claim follows.
\end{proof}
\fi
\subsection{Nilpotency of Euler classes on divisorial schemes and applications}\label{NilpEulerClassSubSect}
The purpose of this section is to analyze the potential failure of Euler classes of vector bundles to be nilpotent in ${\underline\Omega}^\bullet$, and to construct theories ${\underline\Omega}'^\bullet$ and $\Omega'^\bullet$ where the desired nilpotence is provably true. Note that nilpotence of Euler classes is of crucial importance in Section \ref{GeneralPBFSect}, and most of the proofs fail without it.
Let us start by recalling that given $X \in \mathcal{C}_a$ and a vector bundle $E$ of rank $r$ on $X$, then the \emph{Euler class} of $E$ is defined as
$$e(E) := [Z_s \hookrightarrow X] \in {\underline\Omega}^r(X),$$
where $s$ is an arbitrary global section of $E$, and $Z_s$ is the derived vanishing locus of $s$. We also recall that these classes are multiplicative in exact sequences of vector bundles by \cite{annala-chern} Lemma 4.1. Note that by Lemma \ref{NilpChernLem}, Euler classes of line bundles are nilpotent, but the proof does not seem to generalize to higher rank bundles. Moreover, the proof of the nilpotence of Euler classes given in \cite{annala-chern} Lemma 4.2 does not seem to generalize either, unless we restrict our attention to derived schemes admitting an ample line bundle. Hence we we need to come up with another argument.
Suppose $E$ is a rank $r > 1$ vector bundle. Then, as there exists an exact sequence
$$0 \to \mathcal{O}(-1) \to E \to Q \to 0$$
of vector bundles on $\mathbb{P}(E)$, the nilpotency of
$$e(E) \in {\underline\Omega}^*(\mathbb{P}(E))$$
follows from the nilpotency of Euler classes of line bundles. We would be done if we could show that the pullback ${\underline\Omega}^*(X) \to {\underline\Omega}^*(\mathbb{P}(E))$ is injective, which we will try to prove by constructing a class $\eta \in {\underline\Omega}^*(\mathbb{P}(E))$ pushing forward to $1_X$. We fail to achieve this, but instead we find a family of bivariant classes in ${\underline\Omega}^\bullet$ which have to be killed.
We will set up the following notation: $Z$ is the derived vanishing locus of a global section of $E$, $\widetilde X$ is the derived bow up $\mathrm{Bl}_Z(X)$, and $\mathcal{E} \hookrightarrow \widetilde X$ is the exceptional divisor. We begin by recalling the following formula, which is easy to prove using the double point cobordism formula.
\begin{lem}\label{BlowUpFormulaLem}
Let everything be as above. Then
$$1_X = [\widetilde X \to X] + e(E) \bullet [\mathbb{P}(E \oplus \mathcal{O}) \to X] - [\mathbb{P}_{\mathcal{E}}(\mathcal{O}(\mathcal{E}) \oplus \mathcal{O}) \to X]$$
in ${\underline\Omega}^*(X)$. \qed
\end{lem}
Since we do not know that $e(E) \in {\underline\Omega}^*(X)$ is nilpotent, the argument of \cite{annala-chern} Lemma 4.5 gives us only the following truncated version.
\begin{lem}\label{TruncatedBlowUpLiftLem}
Let everything be as above. Then, for all $n \geq 0$,
\begin{align*}
1_X &= \sum_{i=0}^n e(E)^i \bullet [\mathbb{P}(E \oplus \mathcal{O}) \to X]^i \bullet \big( [\widetilde X \to X] - [\mathbb{P}_\mathcal{E}(\mathcal{O}(\mathcal{E}) \oplus \mathcal{O}) \to X] \big) \\
&+ e(E)^{n+1} \bullet [\mathbb{P}(E \oplus \mathcal{O}) \to X]^{n+1}
\end{align*}
in ${\underline\Omega}^*(X)$. \qed
\end{lem}
Let us define
$$\eta_n := \sum_{i=0}^n e(E)^i \bullet [\mathbb{P}_{\mathbb{P}(E)}(E \oplus \mathcal{O}) \to \mathbb{P}(E)]^i \bullet \big( [\widetilde X \to \mathbb{P}(E)] - [\mathbb{P}_\mathcal{E}(\mathcal{O}(\mathcal{E}) \oplus \mathcal{O}) \to \mathbb{P}(E)] \big)$$
in ${\underline\Omega}^*(\mathbb{P}(E))$. Since $e(E) \in {\underline\Omega}^*(\mathbb{P}(E))$ is nilpotent, $\eta_n$ stabilize for $n \gg 0$, and as
$$1_X - \pi_!(\eta_n) = e(E)^{n+1} \bullet [\mathbb{P}(E \oplus \mathcal{O}) \to X]^{n+1} \in {\underline\Omega}^*(X),$$
we conclude that also the right hand side stabilizes when $n$ tends to infinity. We can therefore make the following definition.
\begin{defn}
We define
$$\epsilon_{E, X} \in {\underline\Omega}^0(X)$$
as the limit of $e(E)^n \bullet [\mathbb{P}(E \oplus \mathcal{O}) \to X]^n$ when $n \to \infty$. Define $n_{X,E} \geq 0$ as the smallest integer $n$ such that $\epsilon_{E, X} = e(E)^n \bullet [\mathbb{P}(E \oplus \mathcal{O}) \to X]^n$.
\end{defn}
\begin{defn}
We define the bivariant theory ${\underline\Omega}'^\bullet$ as the quotient of ${\underline\Omega}^\bullet$ by the bivariant ideal $\mathcal{I}_\epsilon$ generated by all the elements $\epsilon_{E,X}$, where $X \in \mathcal{C}_a$ and $E$ is a vector bundle on $X$.
\end{defn}
We record the following expression of $\mathcal{I}_\epsilon(X \to Y)$, which might be useful.
\begin{prop}
Let $X \to Y$ be a map in $\mathcal{C}_a$. Then $\mathcal{I}_\epsilon(X \to Y)$ is the subgroup of ${\underline\Omega}^\bullet(X \to Y)$ generated by
$$f_* \big( e(E)^n \bullet [\mathbb{P}(E \oplus \mathcal{O}) \to V]^n \bullet 1_{V/Y}\big)$$
where $f: V \to X$ is any projective morphism so that $\mathbb{L}_{V/Y}$ has Tor-dimension 1, $E$ is a vector bundle on $V$ and $n \geq n_{E,V}$.
\end{prop}
\begin{proof}
It is enough to show that these groups form a bivariant ideal, i.e., they are stable under pushforwards, pullbacks and left and right multiplication by ${\underline\Omega}^\bullet$. Stability under pullbacks and pushforwards is shown similarly to \cite{annala-cob} Lemma 3.8 using the fact that for $W \to V$, the inequality
$$n_{E, W} \leq n_{E, V}$$
holds.
To prove stability under left multiplication, it suffices to consider multiplication by
$$[W \stackrel h \to Z] \in \Omega^\bullet(Z \to X).$$
To investigate this, we form the derived Cartesian diagram
$$
\begin{tikzcd}
V \arrow[]{r}{f} & X \arrow[]{r} & Y \\
Z' \arrow[]{r}{f'} \arrow[]{u}[swap]{g'} & Z \arrow[]{u}[swap]{g} \\
W' \arrow[]{r}{f''} \arrow[]{u}[swap]{h'} & W \arrow[]{u}[swap]{h}
\end{tikzcd}
$$
and since a direct computation shows that
\begin{align*}
&[W \to Z] \bullet f_* \big( e(E)^n \bullet [\mathbb{P}_V(E \oplus \mathcal{O}) \to V]^n \bullet 1_{V/Y}\big) \\
=& (h \circ f'')_* \big( e(E)^n \bullet [\mathbb{P}_{W'}(E \oplus \mathcal{O}) \to W']^n \bullet 1_{W'/Y}\big),
\end{align*}
stability under left multiplication follows from the inequality $n_{E,W'} \leq n_{E,V}$. The proof that these groups are stable under right multiplication is similar.
\end{proof}
Note that by construction ${\underline\Omega}'^\bullet$ satisfies the following.
\begin{lem}\label{BlowUpLiftLem}
Suppose $X$ is a finite dimensional divisorial Noetherian derived scheme, and let $E$ be a vector bundle on $X$. Then the class
$${1_{\mathbb{P}(E)} - e(\mathcal{O}(-1)) \bullet [\mathbb{P}_{\mathbb{P}(E)}(\mathcal{O}(-1) \oplus \mathcal{O}) \to \mathbb{P}(E)] \over 1_{\mathbb{P}(E)} - e(E) \bullet [\mathbb{P}_{\mathbb{P}(E)}(E \oplus \mathcal{O}) \to \mathbb{P}(E)]} \bullet e(Q) \in {\underline\Omega}'^\bullet\big(\mathbb{P}(E)\big)$$
pushes forward to $1_X$.
\end{lem}
\begin{proof}
Indeed, this follows from Lemma \ref{TruncatedBlowUpLiftLem}, as derived vanishing loci of $Q$ correspond to derived blow ups of $X$.
\end{proof}
We constructed ${\underline\Omega}'^\bullet$ so that the following result would be true.
\begin{prop}\label{NilpEulerProp}
Let $X \in \mathcal{C}_a$, and let $E$ be a vector bundle on $X$. Then
$$e(E) \in {\underline\Omega}'^*(X)$$
is nilpotent.
\end{prop}
\begin{proof}
We already know that $e(E) \in {\underline\Omega}'^*(\mathbb{P}(E))$ is nilpotent. Moreover, it follows from Lemma \ref{BlowUpLiftLem} and the projection formula that the pullback morphism ${\underline\Omega}'^*(X) \to {\underline\Omega}'^*(\mathbb{P}(E))$ is an injection, and therefore $e(E)$ is nilpotent in ${\underline\Omega}'^*(E)$ as well.
\end{proof}
We also record the following useful observation, which also leads to universal properties for ${\underline\Omega}'^\bullet$ and $\Omega'^\bullet$.
\begin{lem}\label{EquivalentModificationsLem}
Let $\nu: {\underline\Omega}^\bullet \to \mathbb{B}^\bullet$ be a Grothendieck transformation. Then the following conditions are equivalent:
\begin{enumerate}
\item for all $X \in \mathcal{C}_a$ and all vector bundles $E$ over $X$, $\nu(e(E)) \in \mathbb{B}^*(X)$ is nilpotent;
\item for all $X \in \mathcal{C}_a$ and all vector bundles $E$ over $X$, $\nu(\epsilon_{E,X}) = 0 \in \mathbb{B}^*(X)$;
\item for all $X \in \mathcal{C}_a$ and all vector bundles $E$ over $X$, there exists $\eta \in \mathbb{B}^*(\mathbb{P}(E))$ pushing forward to $1_X \in \mathbb{B}^*(X)$;
\item for all $X \in \mathcal{C}_a$ and all vector bundles $E$ over $X$, the pullback morphism $\mathbb{B}^*(X) \to \mathbb{B}^*(\mathbb{P}(E))$ is injective.
\end{enumerate}
\end{lem}
\begin{proof}
The implications $1 \Rightarrow 2 \Rightarrow 3$ follow directly from the definition of $\epsilon_{E,X}$. The implication $3 \Rightarrow 4$ follows easily from the projection formula, and the implication $4 \Rightarrow 1$ follows from the fact that
$$\nu \big(e(E) \big) = \nu\big(e(\mathcal{O}(-1)) \bullet e(Q) \big) \in \mathbb{B}^*(\mathbb{P}(E)),$$
which is clearly nilpotent.
\end{proof}
We can then prove the following universal properties (compare this to Theorem \ref{UnivPropOfPCobThm} and Corollary \ref{UnivPropOfCobCor}).
\begin{thm}\label{UnivPropOfModiefiedCobThm}
The bivariant theory ${\underline\Omega}'^\bullet$ ($\Omega'^\bullet$) is the universal stably oriented additive bivariant theory with functoriality $\mathscr{F}_a$ satisfying the section and the formal group law (and the snc) axioms where the pullback morphism along $\mathbb{P}(E) \to X$ is injective for all $X \in \mathcal{C}_a$ and all vector bundles $E$ on $X$. \qed
\end{thm}
In particular both ${\underline\Omega}'^\bullet$ and $\Omega'^\bullet$ admit a unique orientation preserving Grothendieck transformation to the bivariant algebraic K-theory of derived schemes restricted to $\mathscr{F}_a$. We end with the following natural question.
\begin{quest}
Does the bivariant theory ${\underline\Omega}^\bullet$ coincide with ${\underline\Omega}'^\bullet$? In other words, does ${\underline\Omega}^\bullet$ satisfy any of the equivalent conditions of Lemma \ref{EquivalentModificationsLem}?
\end{quest}
\iffalse
\subsubsection{Chern classes of vector bundles and projective bundle formula on divisorial schemes}
The purpose of this section is to prove the following theorem modifying the arguments of \cite{annala-chern} Section 4.
\begin{thm}\label{GeneralChernClassThm}
Theorem \ref{ChernClassThm} is true for all $X \in \mathcal{C}_a$ after replacing ${\underline\Omega}^\bullet$ with ${\underline\Omega}'^\bullet$.
\end{thm}
Since Euler classes of vector bundles are nilpotent in ${\underline\Omega}'^\bullet$, all the proofs of \cite{annala-chern} Sections 4.1 and 4.2 go through immediately; in particular, we can construct Chern classes $c_i(E) \in {\underline\Omega}'^i(X)$ which are natural in pullbacks and satisfy normalization. In order to prove that these Chern classes satisfy Whitney sum formula, we need to find a new proof for the following result, which is Lemma 4.12 from \cite{annala-chern}.
\begin{lem}
Let $X \in \mathcal{C}_a$ and let $E$ be a vector bundle of rank $r$ on $X$. Then the morphism
$$\bigoplus_{i=0}^{r-1} {\underline\Omega}'^*(X) \to {\underline\Omega}'^*(\mathbb{P}(E))$$
defined by
$$(\alpha_0, ..., \alpha_{r-1}) \mapsto \sum_{i=0}^{r-1} \pi^*(\alpha_i) \bullet e(\mathcal{O}(1))^i$$
is an injection, where $\pi$ is the natural map $\mathbb{P}(E) \to X$.
\end{lem}
The above lemma is a consequence of a stronger result, which will be useful also later when proving the projective bundle formula.
\subsubsection{Conner--Floyd theorem for divisorial schemes and applications}
\begin{thm}\label{GeneralCobMainThm}
Define $\Omega'^\bullet$ as the bivariant theory that is the quotient of ${\underline\Omega}'^\bullet$ by the relations of Definition \ref{BivCobDef}. Then, replacing ${\underline\Omega}^\bullet$ by ${\underline\Omega}'^\bullet$ and $\Omega^\bullet$ by $\Omega'^\bullet$, Theorems \ref{ChernClassThm}, \ref{CohCFThm}, \ref{PBFThm} and \ref{GeneralRR} generalize to arbitrary objects/morphisms in $\mathcal{C}_a$.
\end{thm}
\begin{proof}
The old proof to Theorem \ref{ChernClassThm} goes through in this setting without changes. In Theorem \ref{CohCFThm} we need to modify the proof showing that the Chern character morphisms
$$ch: K^0(X) \to \mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}'^*(X)$$
defined by
$$ch([E]) := \mathrm{rank}(E) - e(E^\vee)$$
commute with Gysin pushforwards along arbitrary quasi-smooth projective morphisms (\cite{annala-chern} Proposition 5.7). Using Theorem \ref{ExistenceOfFactorizationsThm}, we have to deal with two cases: the case of a derived regular embedding, which can be dealt with as in \cite{annala-cob} Lemma 4.4, and the case of the natural map
$$\pi: \mathbb{P}(\mathscr{L}_1 \oplus \cdots \oplus \mathscr{L}_r) \to X,$$
which needs a new argument. Combining the $K$-theoretic projective bundle formula with Lemma \ref{ChCommutesWithProjectionLem} takes care of this instance, and the rest of the proof goes through as in \cite{annala-chern} Section 5.1. Given Conner--Floyd theorem, generalizing Theorem \ref{GeneralRR} is formal, and the proof of \cite{annala-chern} Section 5.2 goes through without change.
We are left to generalize the projective bundle formula (Theorem \ref{PBFThm}). As in \cite{annala-yokura} Definition 6.6, we may define the bivariant theory ${\underline\Omega}'^{\bullet, 1}$ (and $\Omega'^{\bullet, 1}$ as well as their unprimed counterparts) otherwise similarly to ${\underline\Omega}'^{\bullet}$, except that the generating cycles are of form
$$[V \to X; \mathscr{L}],$$
where $\mathscr{L}$ is a line bundle on $X$, and bivariant product is defined using the tensor product of line bundles, so that for example
$$[V_1 \to X; \mathscr{L}_1] \bullet [V_1 \to X; \mathscr{L}_2] = [V_1 \times_X V_2 \to X; \mathscr{L}_1 \otimes \mathscr{L}_2] \in {\underline\Omega}'^{*,1}(X).$$
If we can show that the morphism $\iota: {\underline\Omega}'^\bullet(\mathbb{P}(E) \to Y) \to {\underline\Omega}'^{\bullet,1}(X \to Y)$ defined by linearly extending the formula
$$[V \stackrel{f}{\to} \mathbb{P}(E)] \mapsto [V \to X; f^* \mathcal{O}(1)]$$
is injective, then rest of the proof of the projective bundle formula from \cite{annala-chern} Section 5.3 goes through, and we are done. We will in fact show more: if $\mathbb{B}^\bullet$ is a bivariant quotient of ${\underline\Omega}'^\bullet$, then the induced morphism $\iota: \mathbb{B}^\bullet(\mathbb{P}(E) \to Y) \to \mathbb{B}^{\bullet,1}(X \to Y)$ is an injection, proving the projective bundle formula for $\mathbb{B}^\bullet$.
\end{proof}
We are left to justify the following claims. We note that everything below is stated for the bivariant theory ${\underline\Omega}^\bullet$ instead of ${\underline\Omega}'^\bullet$ in order to give a more general statement.
\begin{lem}\label{ChCommutesWithProjectionLem}
Let $\mathscr{L}_1,...,\mathscr{L}_r$ be line bundles on $X \in \mathcal{C}_a$, let $\pi$ be the projection
$$\mathbb{P}(\mathscr{L}_1 \oplus \cdots \oplus \mathscr{L}_r) \to X,$$
and let $\alpha \in \mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X)$. Then for all $i < r$, the equality
$$\pi_! \big( e(\mathcal{O}(1))^i \bullet \pi^*(\alpha) \big) = \alpha$$
holds in ${\underline\Omega}^*(X)$.
\end{lem}
\begin{proof}
By the projection formula, it is enough to show that for all $m < r$, the equality
$$\pi_! \big( e(\mathcal{O}(1))^m \big) = 1_X$$
holds in ${\underline\Omega}^*(X)$. Since
$$e(\mathcal{O}(1)) = e(\mathscr{L}_i(1)) + e(\mathscr{L}_i^\vee) - e(\mathscr{L}_i(1)) \bullet e(\mathscr{L}_i^\vee)$$
for all $i$, and since $\mathbb{P}\big(\bigoplus_{j \not= i} \mathscr{L}_j \big) \hookrightarrow \mathbb{P}\big(\bigoplus_{j} \mathscr{L}_j \big)$ is the derived vanishing locus of a global section of $\mathscr{L}_i(1)$, we have that
\begin{align*}
\pi_! \big( e(\mathcal{O}(1))^m \big) &= \pi_! \Bigg( \prod_{i=1}^m \Big( \big( 1 - e(\mathscr{L}_i^\vee) \big) \bullet e(\mathscr{L}_i(1)) + e(\mathscr{L}_i^\vee) \Big) \Bigg) \\
&= \sum_{I \subset \{1,...,m\}} \prod_{i \in I} \big(1 - e(\mathscr{L}_i^\vee)\big) \bullet \prod_{j \in I^c} e(\mathscr{L}_j^\vee) & (\text{Lemma} \ref{ProjBundlesPushToUnitLem}) \\
&= 1_X,
\end{align*}
proving the claim.
\end{proof}
\begin{lem}\label{ProjBundlesPushToUnitLem}
Let $E$ be a direct sum of $r$ line bundles on $X \in \mathcal{C}_a$ and let $\pi: \mathbb{P}(E) \to X$. Then
$$\pi_!(1_{\mathbb{P}(E)}) = 1_X \in \mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X).$$
\end{lem}
\begin{proof}
Let $E = \bigoplus_{i=1}^r \mathscr{L}_i$. We will proceed by induction on the number of nontrivial $\mathscr{L}_i$. In the base case $E = \mathcal{O}^{\oplus r}$ it is enough to show that $[\mathbb{P}^{r-1}_X \to X] = 1_X$, which can be proven using the argument of \cite{Levine:2007} Lemma 4.2.3.
Suppose then that $\mathscr{L}_1$ is a nontrivial line bundle, and let $E' = \bigoplus_{i=2}^r \mathscr{L}_i$. Then, by Lemma \ref{LMClassOfPBLem} below, we have that
\begin{align*}
[\mathbb{P}(\mathscr{L}_1 \oplus E') \to X] &= [\mathbb{P}(\mathcal{O} \oplus E') \to X] + e(\mathscr{L}_1) \bullet [\mathbb{P}(\mathscr{L}_1 \oplus \mathcal{O} \oplus E') \to X] \\
&- e(\mathscr{L}_1) \bullet [\mathbb{P}(\mathcal{O} \oplus E') \to X] \\
&= 1_X - e(\mathscr{L}_1) + e(\mathscr{L}_1) \bullet [\mathbb{P}(\mathscr{L}_1 \oplus \mathcal{O} \oplus E') \to X],
\end{align*}
where the last equality uses the inductive assumption. Then, applying Lemma \ref{LMClassOfPBLem} several times more, one shows that for all $n \geq 0$
$$[\mathbb{P}(\mathscr{L}_1 \oplus E') \to X] = 1_X - e(\mathscr{L}_1)^n + e(\mathscr{L}_1)^n \bullet [\mathbb{P}(\mathscr{L}_1 \oplus \mathcal{O}^{\oplus n} \oplus E') \to X],$$
and because $e(\mathscr{L}_1) \in {\underline\Omega}^*(X)$ is nilpotent by Lemma \ref{NilpChernLem}, the claim follows.
\end{proof}
We also needed the following result, which is originally Lemma 4.2.2 of \cite{Levine:2007}. The original argument goes through in our context without essential changes.
\begin{lem}\label{LMClassOfPBLem}
Suppose $\mathscr{L}$ is a line bundle and $E$ is a vector bundle on $X$. Then, for any line bundle $\mathscr{M}$ on $X$, the equality
\begin{align*}
\big[\mathbb{P}\big((\mathscr{M} \otimes \mathscr{L}) \oplus E\big) \to X \big] &= [\mathbb{P}(\mathscr{L} \oplus E) \to X] + e(\mathscr{M}) \bullet \big[\mathbb{P}\big((\mathscr{M} \otimes \mathscr{L}) \oplus \mathscr{L} \oplus E\big) \to X \big] \\
&- e(\mathscr{L}) \bullet [\mathbb{P}(\mathscr{L} \oplus E) \to X]
\end{align*}
holds in $\mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X)$. \qed
\end{lem}
\subsubsection{Projective bundle formula for divisorial schemes}
\fi
\section{Projective bundle formula and applications}\label{GeneralPBFSect}
In this section we prove the projective bundle formula (Theorem \ref{GeneralPBFThm}) for ${\underline\Omega}'^\bullet$ (in fact for all bivariant theories that are quotients of ${\underline\Omega}'^\bullet$) and use this to define Chern classes of vector bundles in ${\underline\Omega}'^*$ satisfying good properties by Theorem \ref{GeneralChernClassThm}. As a standard corollary, we obtain a generalized version of the cohomological Conner--Floyd theorem (Theorem \ref{GeneralCFThm}) and the cohomological Grothendieck--Riemann--Roch theorem (Theorem \ref{GeneralGRRThm}).
The proofs use similar ideas as the proofs of the corresponding claims in \cite{annala-chern}. The two main differences are as follows
\begin{enumerate}
\item since we can no longer assume the existence of an ample line bundle, many of the old proofs had to be modified;
\item we have restructured the arguments from \cite{annala-chern} so that we first prove the projective bundle formula, and only after that construct Chern classes and prove their properties.
\end{enumerate}
The structure of this section is as follows: in Section \ref{FundEmbSubSect} we construct the fundamental embedding, which realizes the bivariant group of $\mathbb{P}(E) \to Y$ as a subgroup of the bivariant group of $\mathbb{P}^\infty \times X \to Y$. In Section \ref{GenPBFSubSect} we prove the projective bundle formula by first proving it for trivial projective bundles, and then use the fundamental embedding to prove the general case. Finally, in Section \ref{GenChernClassSubSect} we use the projective buldne formula to define and study Chern classes of vector bundles, and prove the standard corollaries (Conner--Floyd, Grothendieck--Riemann--Roch).
\subsection{The fundamental embedding}\label{FundEmbSubSect}
The purpose of this section is to construct an injective homomorphism
$$\iota_E: \mathbb{B}^\bullet (\mathbb{P}(E) \to Y) \to \mathbb{B}^\bullet (\mathbb{P}^\infty \times X \to Y) := \mathrm{colim}_{n \geq 0} \mathbb{B}^\bullet (\mathbb{P}^n \times X \to Y),$$
called the \emph{fundamental embedding}, where $\mathbb{B}^\bullet$ is a quotient of ${\underline\Omega}^\bullet$, $X \to Y$ is a morphism in $\mathcal{C}_a$, $E$ is a vector bundle on $X$, and the colimit on the right hand side is given by bivariant pushforwards along the evident sequence of inclusions
$$\mathbb{B}^\bullet(X \to Y) \hookrightarrow \mathbb{B}^\bullet(\mathbb{P}^1 \times X \to Y) \hookrightarrow \mathbb{B}^\bullet(\mathbb{P}^2 \times X \to Y) \to \cdots$$
Note that the structure morphisms are injections by \cite{annala-chern} Proposition 2.14.
To do so, we consider the diagram
$$
\begin{tikzcd}
& \mathbb{B}^\bullet\big( \mathbb{P}^{\infty} \times X \to Y\big) \arrow[hook]{d}{j_E} \\
\mathbb{B}^\bullet\big(\mathbb{P}(E) \to Y\big) \arrow[hook]{r}{i_E} & \mathrm{colim}_{n \geq 0} \mathbb{B}^\bullet\big( \mathbb{P}(\mathcal{O}^{\oplus n} \oplus E) \to Y\big)
\end{tikzcd}
$$
where all the morphisms (the structure morphisms included) are given by pushing forward along the obvious inclusions, and injectivity of the transformations follows from \cite{annala-chern} Proposition 2.14. The main step of the construction is the following result.
\begin{lem}\label{JSurjLem}
Let everything be as above. Then $j_E$ is a bijection.
\end{lem}
Assuming this result for now, we can make the following definition.
\begin{defn}
Let everything be as above. Then we define the \emph{fundamental embedding} as $\iota_E := j_E^{-1} \circ i_E$.
\end{defn}
This construction has many good properties, of which we are going to record only a fraction below. Recall that for any vector bundle $E$ on $X$, the bivariant group $\mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$ has a natural structure of a $\mathbb{B}^*(X)$-module by
$$\alpha . \beta := \pi_E^*(\alpha) \bullet \beta,$$
where $\pi_E$ is the natural morphism $\mathbb{P}(E) \to X$.
\begin{prop}\label{FundamentalEmbProp}
Let everything be as above. Then
\begin{enumerate}
\item $\iota_E$ is a map of $\mathbb{B}^*(X)$-modules;
\item there is a well defined operator $e(\mathcal{O}(1)) \bullet -$ on $\mathbb{B}^\bullet(\mathbb{P}^\infty \times X \to Y)$, and moreover
$$e(\mathcal{O}(1)) \bullet \iota_E(-) = \iota_E\big(e(\mathcal{O}(1)) \bullet -\big);$$
\item $\iota_E$ is compatible with bivariant multiplication from right in the sense that if $\alpha \in \mathbb{B}^\bullet(Y \to Z)$, then the square
$$
\begin{tikzcd}
\mathbb{B}^\bullet(\mathbb{P}(E) \to Y) \arrow[]{d}{- \bullet \alpha} \arrow[]{r}{\iota_E} & \mathbb{B}^\bullet(\mathbb{P}^\infty \times X \to Y) \arrow[]{d}{- \bullet \alpha} \\
\mathbb{B}^\bullet(\mathbb{P}(E) \to Z) \arrow[hook]{r}{\iota_E} & \mathbb{B}^\bullet(\mathbb{P}^\infty \times X \to Z)
\end{tikzcd}
$$
commutes;
\item $\iota_E$ is compatible with bivariant pullbacks;
\item if $\psi: E \hookrightarrow F$ is an inclusion of vector bundles, then
$$\iota_E = \iota_F \circ \mathbb{P}(\psi).$$
\end{enumerate}
\end{prop}
\begin{proof}
The first and the second claim follow easily from bivariant projection formula $A_{123}$. The third and the fourth claim follow from $A_{12}$ and $A_{23}$ respectively. To prove the fifth claim, we consider the diagram
$$
\begin{tikzcd}
& \mathbb{B}^\bullet\big( \mathbb{P}^{\infty} \times X \to Y\big) \arrow[hook]{d}{j_E} \\
\mathbb{B}^\bullet\big(\mathbb{P}(E) \to Y\big) \arrow[hook]{r}{i_E} \arrow[hook]{d}{\mathbb{P}(\psi)_*} & \mathrm{colim}_{n \geq 0} \mathbb{B}^\bullet\big( \mathbb{P}(\mathcal{O}^{\oplus n} \oplus E) \to Y\big) \arrow[hook]{d}{\mathbb{P}(\psi')_*}\\
\mathbb{B}^\bullet\big(\mathbb{P}(F) \to Y\big) \arrow[hook]{r}{i_F} & \mathrm{colim}_{n \geq 0} \mathbb{B}^\bullet\big( \mathbb{P}(\mathcal{O}^{\oplus n} \oplus F) \to Y\big),
\end{tikzcd}
$$
where the square commutes by construction. Since $\mathbb{P}(\psi')_* \circ j_E$ is an isomorphism by Lemma \ref{JSurjLem}, it follows that also $\mathbb{P}(\psi')_*$ is an isomorphism, and the claim follows from the fact that $j_F = \mathbb{P}(\psi')_* \circ j_E$.
\end{proof}
Our next goal is to prove Lemma \ref{JSurjLem}. However, we need some preliminary results and definitions before giving the proof. The following observation shows how the hyperplane bundle $\mathcal{O}(1)$ on $\mathbb{P}(E)$ controls the bivariant cobordism classes.
\begin{lem}\label{PBClassControlledByLLem}
Let $X \to Y$ be a morphism in $\mathcal{C}_a$ and let $E$ a vector bundle on $X$. Let $V \to X$ be a projective morphism with $\mathbb{L}_{V/Y}$ of Tor-dimension 1, and let $s_1,s_2: E^\vee \to \mathscr{L}$ be two surjections of vector bundles on $V$ ($\mathscr{L}$ is a line bundle on $V$) inducing projective morphisms $f_1, f_2: V \to \mathbb{P}(E)$ over $X$. Then
$$[V \stackrel{f_1}{\to} \mathbb{P}(E)] = [V \stackrel{f_2}{\to} \mathbb{P}(E)]$$
in ${\underline\Omega}^\bullet(\mathbb{P}(E) \to Y)$.
\end{lem}
\begin{proof}
Let us define a morphism
$$s := x_0 s_1 + x_1 s_2: E^\vee \to \mathscr{L}(1)$$
of vector bundles on $\mathbb{P}^1 \times V$. As the derived vanishing locus $Z$ of $s^\vee$ is disjoint from the fibres over $0$ and $\infty$, we can use Proposition \ref{BlowUpIsResolutionSchemeProp} to find a surjection
$$E^\vee \to \mathscr{L}(1 - \mathcal{E})$$
of vector bundles on $\mathrm{Bl}_Z(\mathbb{P}^1 \times V)$ inducing a projective morphism $\mathrm{Bl}_Z(\mathbb{P}^1 \times V) \to \mathbb{P}^1 \times \mathbb{P}(E)$ which realizes the desired relation.
\end{proof}
The following notation will be useful in the proof.
\begin{defn}
Given a derived scheme $X$ and a line bundle $\mathscr{L}$ on $X$, we recursively define the following pairs:
$$\big(P_0(X,\mathscr{L}), \mathscr{M}_0(X,\mathscr{L}) \big) := \big(X, \mathscr{L} \big)$$
and
$$\big(P_{i+1}(X,\mathscr{L}), \mathscr{M}_{i+1}(X,\mathscr{L}) \big) := \big(\mathbb{P}_{P_i(X,\mathscr{L})} (\mathscr{M}_i(X,\mathscr{L}) \oplus \mathcal{O}), \mathscr{M}_i(X,\mathscr{L})(H_{i+1}) \big),$$
where $\mathcal{O}(H_{i+1})$ is the hyperplane bundle on $\mathbb{P}_{P_i(X,\mathscr{L})} \big(\mathscr{M}_i(X,\mathscr{L}) \oplus \mathcal{O}\big)$.
\end{defn}
\begin{proof}[Proof of Lemma \ref{JSurjLem}]
To simplify notation, we set
$$\mathbb{B}^\bullet( \mathbb{P}(\mathcal{O}^{\oplus \infty} \oplus E) \to Y) := \mathrm{colim}_{n \geq 0} \mathbb{B}^\bullet( \mathbb{P}(\mathcal{O}^{\oplus n} \oplus E) \to Y).$$
Moreover, by Lemma \ref{PBClassControlledByLLem}, given a projective morphism $V \to X$ with $\mathbb{L}_{V/Y}$ of Tor-dimension 1 and a line bundle $\mathscr{L}$ on $V$ that admits a surjection from $E^\vee \oplus \mathcal{O}^{\oplus n}$ for some $n$, there exists a unique class
$$[V \to X; \mathscr{L}] := [X \stackrel f \to \mathbb{P}(\mathcal{O}^{\oplus \infty} \oplus E)] \in \mathbb{B}^\bullet( \mathbb{P}(\mathcal{O}^{\oplus \infty} \oplus E) \to Y),$$
where $f^* (\mathcal{O}(1)) \simeq \mathscr{L}$. As $\mathbb{B}^\bullet$ was assumed to be a quotient of ${\underline\Omega}^\bullet$, these classes generate, so it is enough to show that all $[V \to X; \mathscr{L}]$ lie in the image of $j_E$. Note that this is clear if $\mathscr{L}$ is globally generated.
Our argument is a variation of the arguments of Section 6.2 of \cite{annala-yokura}. Let $D$ be the derived vanishing locus of a global section of $\mathscr{L}$, $W$ the derived blow up of $\mathbb{P}^1 \times V$ at $\infty \times D$, and consider the line bundle $\mathscr{L}' := \mathscr{L}(H - \mathcal{E})$ on $W$, where $H$ is the pullback of the hyperplane class on $\mathbb{P}^1$. Since the center of the blow up giving $W$ is the derived vanishing locus of a global section of $\mathscr{L} \oplus \mathcal{O}(H)$, there exists by Proposition \ref{BlowUpIsResolutionSchemeProp} a surjection
$$\mathcal{O}(H) \oplus \mathscr{L} \twoheadrightarrow \mathscr{M}_0$$
of vector bundles on $W$. Hence, there exists a surjection of vector bundles
$$\mathcal{O}^{\oplus n + 2} \oplus E^\vee \twoheadrightarrow \mathscr{M}_0.$$
Moreover, since there exists a natural surjection
$$\mathcal{O} \oplus \mathscr{M}_i(W, \mathscr{L}') \twoheadrightarrow \mathscr{M}_{i+1}(W, \mathscr{L}')$$
on $P_{i+1}(W, \mathscr{L}')$, we can recursively find surjections
$$\mathcal{O}^{\oplus n + 2 + i + 1} \oplus E^\vee \twoheadrightarrow \mathscr{M}_{i+1}(W, \mathscr{L}'),$$
which give rise to projective morphisms $P_i(W,\mathscr{L}') \to \mathbb{P}^1 \times \mathbb{P}(\mathcal{O}^{\oplus n + 2 + i} \oplus E)$ for all $i \geq 0$.
Since the line bundle $\mathscr{L}'$ on $W$ restricts to
\begin{enumerate}
\item $\mathscr{L}$ on $0 \times V$;
\item $\mathcal{O}$ on the strict transform of $\infty \times V$;
\item $\mathscr{M}_1(V,\mathscr{L})\vert_\mathcal{E}$ on the exceptional divisor $\mathcal{E} \simeq P_1(V, \mathscr{L}) \vert_D$,
\end{enumerate}
we see that the morphisms $P_i(W,\mathscr{L}') \to \mathbb{P}^1 \times \mathbb{P}(\mathcal{O}^{\oplus n + 2 + i} \oplus E)$ give rise to the relations
\begin{align*}
[P_i(V, \mathscr{L}) \to X; \mathscr{M}_i(V, \mathscr{L})] &= [P_i(V, \mathcal{O}) \to X; \mathscr{M}_i(V, \mathcal{O})] \\
&+ e(\mathscr{L}) \bullet [P_{i+1}(V, \mathscr{L}) \to X; \mathscr{M}_{i+1}(V, \mathscr{L})] \\
&- e(\mathscr{L}) \bullet [\mathbb{P}_{P_i(V, \mathcal{O})}(\mathscr{L} \oplus \mathcal{O}) \to X; \mathscr{M}_i(V, \mathcal{O})]
\end{align*}
in $\mathbb{B}^\bullet\big( \mathbb{P}(\mathcal{O}^{\oplus \infty} \oplus E) \to Y\big)$, and since $e(\mathscr{L})$ is nilpotent, we compute that
\begin{align*}
[V \to X; \mathscr{L}] &= [V \to X; \mathcal{O}] + \sum_{i=1}^\infty e(\mathscr{L})^i \bullet \Big([P_i(V, \mathcal{O}) \to X; \mathscr{M}_i(V, \mathcal{O})] \\
&-[\mathbb{P}_{P_{i-1}(V, \mathcal{O})}(\mathscr{L} \oplus \mathcal{O}) \to X; \mathscr{M}_{i-1}(V, \mathcal{O})] \Big) \in \mathbb{B}^\bullet\big( \mathbb{P}(\mathcal{O}^{\oplus \infty} \oplus E) \to Y\big).
\end{align*}
As $\mathscr{M}_i(V, \mathcal{O})$ are globally generated, the right hand side of the above equation is evidently in the image of $j_E$, and therefore so is $[V \to X; \mathscr{L}]$. This proves the claim.
\end{proof}
\subsection{Proof of the projective bundle formula}\label{GenPBFSubSect}
The goal of this section is to prove the following result.
\begin{thm}\label{GeneralPBFThm}
Let $\mathbb{B}^\bullet$ be a quotient of ${\underline\Omega}'^\bullet$, let $X \to Y$ be a morphism in $\mathcal{C}_a$ and let $E$ be a vector bundle of rank $r$ on $X$. Then the morphism
$$\mathcal{P} roj: \bigoplus_{i=0}^{r-1} \mathbb{B}^\bullet(X \to Y) \to \mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$$
given by
$$(\alpha_0, \alpha_1, ..., \alpha_r) \mapsto \sum_{i=0}^{r-1} e(\mathcal{O}(1))^i \bullet 1_{\mathbb{P}(E)/X} \bullet \alpha_i$$
is an isomorphism.
\end{thm}
Throughout the section, $\mathbb{B}^\bullet$ is a quotient of ${\underline\Omega}'^\bullet$, $X \to Y$ is a morphism in $\mathcal{C}_a$, and $E$ is a rank $r$ vector bundle on $X$, unless otherwise mentioned. We start by proving the result for trivial vector bundles, which follows from the next result.
\begin{lem}\label{GeneralTrivPBFLem}
Let $X \to Y$ be a morphism in $\mathcal{C}_a$, and let $\mathbb{B}^\bullet$ be a quotient of ${\underline\Omega}^\bullet$. Then the morphism
$$\bigoplus_{i=0}^{\infty} \mathbb{B}^\bullet(X \to Y) \to \mathbb{B}^\bullet(\mathbb{P}^\infty \times X \to Y),$$
where the $i^{th}$ morphism is given by
$$[\mathbb{P}^i \times X \hookrightarrow \mathbb{P}^\infty \times X] \bullet -: \mathbb{B}^\bullet(X \to Y) \to \mathbb{B}^\bullet(\mathbb{P}^\infty \times X \to Y),$$
is an isomorphism.
\end{lem}
\begin{proof}
The proof from Sections 6.2 and 6.3 of \cite{annala-yokura} goes through without essential modifications. We outline the argument here as the proof of Lemma \ref{JSurjLem} got us almost there. Indeed, by that proof, $\mathbb{B}^\bullet(\mathbb{P}^{\infty} \times X \to Y)$ is generated as an Abelian group by elements of form
$$[\mathbb{P}_i(X, \mathcal{O}) \stackrel{f_i}{\to} \mathbb{P}^\infty \times X] \bullet \alpha,$$
where $f_i^*(\mathcal{O}(1)) \simeq \mathscr{M}_i(X,\mathcal{O})$ and $\alpha \in \mathbb{B}^\bullet(X \to Y)$. It then follows from a straightforward computation (see \cite{annala-yokura} Lemmas 6.20 and 6.21) that every element of $\mathbb{B}^\bullet(\mathbb{P}^{\infty} \times X \to Y)$ is an integral combination of
$$[\mathbb{P}^i \times X \hookrightarrow \mathbb{P}^\infty \times X] \bullet \alpha,$$
and using the operator $e(\mathcal{O}(1)) \bullet -$ and the ``bivariant pushforward''
$$\mathbb{B}^\bullet(\mathbb{P}^\infty \times X \to Y) \to \mathbb{B}^\bullet(X \to Y)$$
one checks that such an expression is unique, proving the claim.
\end{proof}
Combining this with the fundamental embedding constructed in the previous section, we see that
$$\iota_E(\alpha) = \sum_{i = 0}^\infty [\mathbb{P}^i \times X \hookrightarrow \mathbb{P}^\infty \times X] \bullet u_i(\alpha) \in \mathbb{B}^\bullet(\mathbb{P}^\infty \times X \to Y)$$
for some uniquely determined $u_i(\alpha) \in \mathbb{B}^\bullet(X \to Y)$.
\begin{defn}\label{CoeffDef}
Given $\alpha \in \mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$, we will call $u_i(\alpha)$ the \emph{$i^{th}$ coefficient} of $\alpha$.
\end{defn}
Our first goal is to show that the first $r$ coefficients completely determine an element of $\mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$. To do so, we start with the following observation.
\begin{lem}\label{ChernClassFormulaLem}
Let $X \in \mathcal{C}_a$, and let $E$ be a vector bundle of rank $r$ on $X$. Then there exists $d_i(E) \in \mathbb{B}^*(X)$ so that
$$e(\mathcal{O}(1))^r = \sum_{i=1}^{r} d_i(E) \bullet e(\mathcal{O}(1))^{r-i} \in \mathbb{B}^*(\mathbb{P}(E)).$$
\end{lem}
\begin{proof}
Let $\pi$ be the natural projection $\mathrm{Fl}(E) \to \mathbb{P}(E)$ and let $\eta \in \mathbb{B}^*(\mathrm{Fl}(E))$ be such that $\pi_!(\eta) = 1_{\mathbb{P}(E)}$.
Let
$$0 = E_0 \subset E_1 \subset \cdots \subset E_{r-1} \subset E_r = E$$
be a filtration of $E$ on $\mathrm{Fl}(E)$ with $\mathscr{L}_i := E_{i}/E_{i-1}$ line bundles. Then, since
\begin{align*}
e(\mathscr{L}_1(1)) \bullet \cdots \bullet e(\mathscr{L}_r(1)) &= e(E(1)) \\
&= 0
\end{align*}
it follows that also
\begin{align*}
e(\mathscr{L}_1^\vee(-1)) \bullet \cdots \bullet e(\mathscr{L}^\vee_r(-1)) &= \mathrm{inv}_F \big(e(\mathscr{L}_1(1))\big) \bullet \cdots \bullet \mathrm{inv}_F \big( e(\mathscr{L}_r(1))\big) \\
&= 0
\end{align*}
and therefore
$$\prod_{i=1}^r \big( F(e(\mathscr{L}_i^\vee(-1)), e(\mathcal{O}(1))) - e(\mathcal{O}(1)) \big) = \prod_{i=1}^r \big( e(\mathscr{L}_i^\vee) - e(\mathcal{O}(1)) \big)$$
vanishes. In conclusion,
$$\sum_{i=0}^r (-1)^{r-i} s_i\big(e(\mathscr{L}_1^\vee),...,e(\mathscr{L}_r^\vee)\big) \bullet e(\mathcal{O}(1))^{r-i} = 0 \in {\underline\Omega}^*(\mathrm{Fl}(E)),$$
where $s_i$ are the elementary symmetric polynomials, and the claim follows with
$$d_i(E) := (-1)^{r-i+1} \pi_!\Big( s_i\big(e(\mathscr{L}_1^\vee),...,e(\mathscr{L}_r^\vee)\big) \bullet \eta \Big)$$
from the projection formula.
\end{proof}
It is then easy to prove that the first $r$ coefficients of $\alpha \in \mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$ completely determine $\alpha$.
\begin{lem}\label{RCoeffLem}
Suppose $\alpha \in \mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$ is such that $u_i(\alpha) = 0$ for all $i < r$. Then $\alpha = 0$.
\end{lem}
\begin{proof}
Clearly
$$e(\mathcal{O}(1))^r \bullet \iota_E(\alpha) = \sum_{i=0}^\infty [\mathbb{P}^i \times X \hookrightarrow \mathbb{P}^\infty \times X] \bullet u_{r+i}(\alpha).$$
On the other hand, by Lemma \ref{ChernClassFormulaLem}, we can also compute that
\begin{align*}
e(\mathcal{O}(1))^r \bullet \iota_E(\alpha) &= \sum_{i=1}^r d_i(E) \bullet e(\mathcal{O}(1))^{r-i} \bullet \iota_E(\alpha)
\end{align*}
so that
$$u_{r+i}(\alpha) = \sum_{j=0}^{r-1} d_{r-j}(E) \bullet u_{j+i}(\alpha),$$
and therefore $u_i(\alpha) = 0$ for all $i$. Since the fundamental embedding is an injection, the claim follows.
\end{proof}
In order to finish the proof of the projective bundle formula, we need to show that given $u_0,...,u_r \in \mathbb{B}^\bullet(X \to Y)$, there exists
$$\alpha := \sum_{i=0}^{r-1} e(\mathcal{O}(1))^i \bullet 1_{\mathbb{P}(E)/X} \bullet \alpha_i \in \mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$$
so that $u_i(\alpha) = u_i$ for all $i < r$. To achieve this, we start by investigating the image of the fundamental class under $\iota_E$.
\begin{lem}\label{CoefficientLem1}
Suppose $X \in \mathcal{C}_a$ and let $E$ be a vector bundle of rank $r$ on $X$. Then $u_{r-1}(1_{\mathbb{P}(E)/X})$ is a unit and $u_{i}(1_{\mathbb{P}(E)/X})$ is nilpotent for all other $i$.
\end{lem}
\begin{proof}
We start with the case of a split vector bundle, i.e.,
$$E \simeq \bigoplus_{i=1}^r \mathscr{L}_i.$$
By construction of $\iota_E$, it is enough to understand the image of $1_{\mathbb{P}(E)/X}$ inside
$$\mathbb{B}^*(\mathbb{P}(\mathcal{O}^\infty \oplus E) \to X) := \mathrm{colim}_{n \geq 0} \mathbb{B}^*(\mathbb{P}(\mathcal{O}^n \oplus E) \to X).$$
If $E' = \bigoplus_{i \in I} \mathscr{L}_i$ for some $I \subset \{1,...,r\}$ and $n \geq 0$, we will denote by
$$[\mathbb{P}(\mathcal{O}^{\oplus n} \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)]$$
the class of the obvious linear embedding.
Let $E' := \bigoplus_{i=2}^r \mathscr{L}_i$ and consider $\mathbb{P}(\mathcal{O} \oplus \mathscr{L}_1 \oplus E')$. As $\mathscr{L}_1(1)$ has a section with derived vanishing locus $\mathbb{P}(\mathcal{O} \oplus E')$ and $\mathcal{O}(1)$ has a section with derived vanishing locus $\mathbb{P}(\mathscr{L}_1 \oplus E')$ we conclude that
\begin{align*}
&[\mathbb{P}(\mathcal{O} \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)] \\
=& [\mathbb{P}(\mathscr{L}_1 \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)] \\
+& e(\mathscr{L}_1) \bullet [\mathbb{P}(\mathcal{O} \oplus \mathscr{L}_1 \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^{\oplus\infty} \oplus E)] \\
+& \sum_{i,j \geq 1} a_{ij} e(\mathcal{O}(1))^{i-1} \bullet e(\mathscr{L}_1)^j \bullet [\mathbb{P}(\mathscr{L}_1 \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^{\oplus \infty} \oplus E)]
\end{align*}
or in other words
\begin{align*}
[\mathbb{P}(\mathscr{L}_1 \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)] &= H\big(e(\mathcal{O}(1)), e(\mathscr{L}_1)\big) \bullet \Big( [\mathbb{P}(\mathcal{O} \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)] \\
&- e(\mathscr{L}_1) \bullet [\mathbb{P}(\mathcal{O} \oplus \mathscr{L}_1 \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^{\oplus\infty} \oplus E)]\Big),
\end{align*}
where
$$H(t,x) := {1 \over 1 + \sum_{i,j \geq 1} a_{ij} t^{i-1} \bullet x^j} \in \mathbb{L}[[t,x]].$$
Hence
\begin{align*}
&[\mathbb{P}(\mathscr{L}_1 \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)] \\
=& (-1)^{i_1-1} \sum_{i_1 \geq 1} H\big(e(\mathcal{O}(1)), e(\mathscr{L}_1)\big)^{i_1} \bullet e(\mathscr{L}_1)^{i_1 - 1} \bullet [\mathbb{P}(\mathcal{O}^{\oplus i_1} \oplus E') \hookrightarrow \mathbb{P}(\mathcal{O}^{\oplus\infty} \oplus E)],
\end{align*}
and dealing with the other $\mathscr{L}_i$ similarly, we compute that $[\mathbb{P}(E) \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)]$ equals to
\begin{align*}
\sum_{i_1 \cdots i_r \geq 1} \prod_{k=1}^r (-1)^{i_k-1} \Big( H\big(e(\mathcal{O}(1)), e(\mathscr{L}_k)\big)^{i_k} \bullet e(\mathscr{L}_k)^{i_k - 1} \Big) \bullet [\mathbb{P}(\mathcal{O}^{\oplus i_1 + \cdots + i_r}) \hookrightarrow \mathbb{P}(\mathcal{O}^{\oplus\infty} \oplus E)].
\end{align*}
Therefore, by inspection
$$j_E^{-1}\big([\mathbb{P}(E) \hookrightarrow \mathbb{P}(\mathcal{O}^\infty \oplus E)]\big) = \sum_{i=0}^\infty [\mathbb{P}^i \times X \hookrightarrow \mathbb{P}^\infty \times X] \bullet H_i\big(e(\mathscr{L}_1), ..., e(\mathscr{L}_r)\big),$$
where
$$H_i(x_1,...,x_r) \in \mathbb{L}[[x_1,...,x_r]]$$
are universal symmetric power series, $H_{r-1}$ has constant term $1$ and the other $H_i$ do not have a constant term at all. Hence we are done in the split case.
We are left with the general case. Replacing $X$ by the full flag bundle $\mathrm{Fl}(E)$ of $E$ we can use Lemma \ref{BivInjPullbackLem} below to reduce to the case where $E$ admits a filtration
$$0 = E_0 \subset E_1 \subset \cdots \subset E_{r-1} \subset E_{r} = E$$
with $\mathscr{L}_i := E_{i}/E_{i-1}$ line bundles. Denoting by $E'$ the sum $\bigoplus_{i=1}^r \mathscr{L}_i$ and by $\psi_1$ and $\psi_2$ the natural inclusions $\mathbb{P}(E) \hookrightarrow \mathbb{P}(E \oplus E')$ and $\mathbb{P}(E') \hookrightarrow \mathbb{P}(E \oplus E')$ respectively, we may compute that
\begin{align*}
\iota_E(1_{\mathbb{P}(E)/X}) &= \iota_{E \oplus E'} \big(\psi_{1*}(1_{\mathbb{P}(E)/X})\big) \\
&= \iota_{E \oplus E'} \big(e(E'(1))\big) \\
&= \iota_{E \oplus E'} \big(e(\mathscr{L}_1(1)) \bullet \cdots \bullet e(\mathscr{L}_r(1))\big) \\
&= \iota_{E \oplus E'} \big(e(E(1))\big) \\
&= \iota_{E \oplus E'} \big(\psi_{2*}(1_{\mathbb{P}(E')/X})\big) \\
&= \iota_{E'}(1_{\mathbb{P}(E')/X})
\end{align*}
so the general case follows from the split case.
\end{proof}
We still need to prove the following simple lemma used in the above proof.
\begin{lem}\label{BivInjPullbackLem}
Let $g: Y' \to Y$ be a quasi-smooth projective morphism and suppose there exists $\eta \in \mathbb{B}^*(Y')$ such that $g_!(\eta) = 1_Y$. Then, for any morphism $X \to Y$ in $\mathcal{C}_a$, the bivariant pullback
$$g^*: \mathbb{B}^\bullet(X \to Y) \to \mathbb{B}^\bullet(X' \to Y')$$
is injective.
\end{lem}
\begin{proof}
Let $g': X' \to X$ be the pullback of $g$, and let $\alpha \in \mathbb{B}^\bullet(X \to Y)$. Then we can compute that
\begin{align*}
g'_*\big(g^*(\alpha) \bullet \eta \bullet 1_{Y' / Y} \big) &= \alpha \bullet g_*(\eta \bullet 1_{Y' / Y} ) & (\text{bivariant axiom $A_{123}$}) \\
&= \alpha \bullet g_!(\eta) \\
&= \alpha
\end{align*}
proving the claim.
\end{proof}
Next step is to make the following observation.
\begin{lem}\label{CoefficientLem2}
Suppose $X \in \mathcal{C}_a$ and let $E$ be a vector bundle of rank $r$ on $X$. Then for each $i \leq r-1$ and $j \leq r-1$ there exist unique $\alpha_{j,i}(E) \in \mathbb{B}^*(X)$ so that
\begin{align*}
u_k \Bigg(\sum_{j=0}^{r-1} e(\mathcal{O}(1))^j \bullet 1_{\mathbb{P}(E)/X} \bullet \alpha_{j,i}(E) \Bigg) &=
\begin{cases}
1_X & \text{if k=i;} \\
0 & \text{if $k<r$ and $k \not = i$.}
\end{cases}
\end{align*}
Moreover, $[\alpha_{j,i}(E)]$ is an invertible matrix with coefficients in $\mathbb{B}^*(X)$.
\end{lem}
\begin{proof}
Consider the matrix $A(E)$, whose entries are
\begin{align*}
A(E)_{j,i} :=& u_j\big(e(\mathcal{O}(1))^i \bullet 1_{\mathbb{P}(E)/X} \big) \\
=& u_{i+j}\big(1_{\mathbb{P}(E)/X} \big).
\end{align*}
Then, by construction, given
$$\overline{\alpha} =
\begin{bmatrix}
\alpha_0 \\
\vdots \\
\alpha_{r-1}
\end{bmatrix}
\in \mathbb{B}^*(X)^{\oplus r}
$$
we see that $A(E) \overline{\alpha}$ is the vector of first $r$ coefficients of $\sum_{i=0}^{r-1} e(\mathcal{O}(1))^i \bullet 1_{\mathbb{P}(E)/X} \bullet \alpha_i$. By Lemma \ref{CoefficientLem1} $A(E)$ has units on the anti-diagonal and nilpotent elements everywhere else, so $A(E)$ is invertible by basic linear algebra. Moreover, $[\alpha_{j,i}(E)] := A(E)^{-1}$ is clearly the unique matrix with the desired property.
\end{proof}
Finally, we can prove the result we are after.
\begin{lem}\label{CoefficientLem3}
Let $X \to Y$ be a morphism in $\mathcal{C}_a$, let $u_0,...,u_{r-1} \in \mathbb{B}^\bullet(X \to Y)$, and let
$$\alpha := \sum_{i=0}^{r-1} \sum_{j=0}^{r-1} e(\mathcal{O}(1))^j \bullet 1_{\mathbb{P}(E)/X} \bullet \alpha_{j,i}(E) \bullet u_i.$$
Then $u_i(\alpha) = u_i$ for all $i < r$.
\end{lem}
\begin{proof}
This is immediate from \ref{CoefficientLem2}.
\end{proof}
We are then finally ready to prove the projective bundle formula.
\begin{proof}[Proof of Theorem \ref{GeneralPBFThm}]
By Lemma \ref{CoefficientLem3}, for each $u_0,...,u_{r-1}$, the element
$$\alpha := \sum_{j=0}^{r-1} e(\mathcal{O}(1))^j \bullet 1_{\mathbb{P}(E)/X} \bullet \Bigg( \sum_{i=0}^{r-1} \alpha_{j,i}(E) \bullet u_i \Bigg)$$
has $u_i(\alpha) = u_i$ for $i < r$, and since $[\alpha(E)_{j,i}]$ is an invertible matrix by Lemma \ref{CoefficientLem2}, it follows that the map
$$\mathcal{P} roj: (\alpha_0,...,\alpha_{r-1}) \mapsto \sum_{j=0}^{r-1} e(\mathcal{O}(1))^j \bullet 1_{\mathbb{P}(E)/X} \bullet \alpha_j$$
is injective. On the other hand, by Lemma \ref{RCoeffLem}, there exists at most one element of $\mathbb{B}^\bullet(\mathbb{P}(E) \to Y)$ with first $r$ coefficients $u_0,...,u_{r-1}$, so the surjectivity of $\mathcal{P} roj$ follows from Lemma \ref{CoefficientLem3} as well.
\end{proof}
\subsection{Chern classes and applications}\label{GenChernClassSubSect}
The purpose of this section is to define Chern classes of vector bundles, and to apply them to generalizing Conner--Floyd theorem and Grothendieck--Riemann--Roch theorem to all divisorial Noetherian derived schemes. We start with the fundamental definition.
\begin{defn}\label{GeneralChernClassDefn}
Let $X \in \mathcal{C}_a$ and let $E$ be a vector bundle of rank $r$ on $X$. Then by the projective bundle formula (Theorem \ref{GeneralPBFThm}) the equation
$$\sum_{i=0}^r (-1)^i e(\mathcal{O}(1))^{i} \bullet c_{r-i}(E) = 0 \in {\underline\Omega}'^r(\mathbb{P}(E^\vee))$$
holds for unique $c_i(E) \in {\underline\Omega}'^i(E)$ with $c_0(E)=1_X$. The element $c_i(E)$ is called the \emph{$i^{th}$ Chern class} of $E$.
\end{defn}
These Chern classes satisfy the expected properties, as the following result shows.
\begin{thm}\label{GeneralChernClassThm}
Define the \emph{total Chern class} of a rank $r$ vector bundle $E$ as
$$c(E) = 1 + c_1(E) + \cdots + c_r(E).$$
Then Chern classes satisfy the following properties:
\begin{enumerate}
\item \emph{naturality:} $f^* c_i(E) = c_i(f^*E)$ for $f: Y \to X$ in $\mathcal{C}_a$;
\item \emph{Whitney sum formula:} given a short exact sequence
$$0 \to E' \to E \to E'' \to 0$$
of vector bundles on $X$, we have
$$c(E) = c(E') \bullet c(E'') \in {\underline\Omega}'^*(X);$$
\item \emph{normalization:} if $E$ is a rank $r$ vector bundle on $X \in \mathcal{C}_a$, then $c_r(E) = e(E)$.
\end{enumerate}
\end{thm}
\begin{proof}
Naturality is trivial. To prove the Whitney sum formula, we can pull back to flag bundles to reduce to the case where both $E'$ and $E''$ admit filtrations
$$0 = E'_0 \subset E'_1 \subset \cdots \subset E'_{r'} = E'$$
and
$$0 = E''_0 \subset E''_1 \subset \cdots \subset E''_{r''} = E''$$
with graded pieces line bundles $\mathscr{L}'_i := E'_i/E'_{i-1}$ and $\mathscr{L}''_i := E''_i / E''_{i-1}$ respectively. One then argues as in Lemma \ref{ChernClassFormulaLem} to show that
$$c_i(E') = s_i\big(e(\mathscr{L}'_1), ..., e(\mathscr{L}'_{r'}) \big),$$
$$c_i(E'') = s_i\big(e(\mathscr{L}''_1), ..., e(\mathscr{L}''_{r''}) \big)$$
and
$$c_i(E) = s_i\big(e(\mathscr{L}'_1), ..., e(\mathscr{L}'_{r'}), e(\mathscr{L}''_1), ..., e(\mathscr{L}''_{r''}) \big)$$
proving the Whitney sum formula. Normalization is trivial for line bundles, and follows for general vector bundles from the Whitney sum formula.
\end{proof}
Our next goal is to prove the Conner--Floyd theorem.
\begin{thm}\label{GeneralCFThm}
Let $\mathbb{Z}_m$ be the integers considered as an $\mathbb{L}$-algebra via the multiplicative formal group law $x + y - xy$, and suppose $X \in \mathcal{C}_a$. Then the morphisms
$$\mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X) \to K^0(X)$$
given by
$$[V \stackrel f \to X] \mapsto [f_* \mathcal{O}_V]$$
are isomorphisms of rings that commute with pullbacks and Gysin pushforwards. The theorem remains true if we replace ${\underline\Omega}'^*(X)$ with $\Omega'^*(X)$.
\end{thm}
\begin{proof}
Now that we are armed with a good theory of Chern classes, this result can be proved in almost the same way as in \cite{annala-chern} Section 5.1 and \cite{annala-cob} Section 4.1 with one difference: we have to show that the \emph{Chern character morphisms}
$$ch: K^0(X) \to \mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}^*(X)$$
defined by
$$[E] \mapsto \mathrm{rank}(E) - c_1(E^\vee)$$
commute with pushforwards along natural projections
$$\pi: \mathbb{P}\big(\mathscr{L}_1 \oplus \cdots \oplus \mathscr{L}_r \big) \to X$$
(this is part of \cite{annala-cob} Lemma 4.2, but the proof of the other part generalizes immediately). Applying the $K$-theoretic projective bundle formula, and the projection formula, this reduces to showing that for all $i < r$
\begin{align*}
1_X &= ch \big(1_X \big) \\
&= ch \Big( \pi_!\big(c_1(\mathcal{O}(1))^i \big) \Big) \\
&= \pi_! \Big(ch \big(c_1(\mathcal{O}(1))^i \big) \Big) \\
&= \pi_! \big(c_1(\mathcal{O}(1))^i \big),
\end{align*}
where the second equality follows from a simple $K$-theory computation, and the last equality follows from the fact that $ch$ preserves Chern classes and multiplications (see \cite{annala-cob} Lemma 4.1 and the discussion preceding it). But this is easy since
\begin{align*}
c_1(\mathcal{O}(1))^i &= \prod_{k=1}^i \big(c_1(\mathscr{L}_k(1)) + c_1(\mathscr{L}_k^\vee) - c_1(\mathscr{L}_k(1)) \bullet c_1(\mathscr{L}_k^\vee)\big) \\
&= \prod_{k=1}^i \big((1 - c_1(\mathscr{L}_k^\vee)) \bullet c_1(\mathscr{L}_k(1)) + c_1(\mathscr{L}_k^\vee)\big)
\end{align*}
which pushes forward to $1_X$ by Lemma \ref{ProjBundlesPushToUnitLem} below. The same proof goes through for $\mathbb{Z}_m \otimes_\mathbb{L} \Omega^*(X)$ because all the relevant maps stay well defined.
\end{proof}
\begin{lem}\label{ProjBundlesPushToUnitLem}
Let $E$ be a vector bundle of rank $r$ on $X \in \mathcal{C}_a$ and let $\pi: \mathbb{P}(E) \to X$. Then
$$\pi_!(1_{\mathbb{P}(E)}) = 1_X \in \mathbb{Z}_m \otimes_\mathbb{L} {\underline\Omega}'^*(X).$$
\end{lem}
\begin{proof}
Pulling back to the flag bundle, we may assume without loss of generality that $E$ has a filtration
$$0 = E_0 \subset E_1 \subset \cdots \subset E_{r} = E$$
with $\mathscr{L}_i := E_i / E_{i-1}$ line bundles. Letting $E' = \bigoplus_{i=1}^r \mathscr{L}_i$, and noticing that
\begin{align*}
[\mathbb{P}(E) \hookrightarrow \mathbb{P}(E \oplus E')] &= e(E'(1)) \bullet 1_{\mathbb{P}(E \oplus E')} \\
&= e(\mathscr{L}_1(1)) \bullet \cdots \bullet e(\mathscr{L}_r(1)) \bullet 1_{\mathbb{P}(E \oplus E')} \\
&= e(E(1)) \bullet 1_{\mathbb{P}(E \oplus E')} \\
&= [\mathbb{P}(E') \hookrightarrow \mathbb{P}(E \oplus E')]
\end{align*}
we are reduced to the case where $E$ is a direct sum of line bundles $\mathscr{L}_1,...,\mathscr{L}_r$. But it was computed in Lemma \ref{CoefficientLem1} (use multiplicative formal group law in the computations) that
$$[\mathbb{P}(\mathscr{L}_1 \oplus \cdots \oplus \mathscr{L}_r) \to X] = \sum_{i_1,...,i_r \geq 1} \prod_{k=1}^r {\big(-c_1(\mathscr{L}_k)\big)^{i_k - 1} \over \big(1 - c_1(\mathscr{L}_k)\big)^{i_k}} \bullet [\mathbb{P}^{i_1 + \cdots + i_k - 1} \times X \to X]$$
and since
$$\sum_{i=0}^\infty {(-x)^i \over (1-x)^i} = 1-x,$$
we are reduced to showing that $[\mathbb{P}^n \times X \to X] = 1_X$ for all $n \geq 0$. But this can be proven as in the proof of \cite{Levine:2007} Lemma 4.2.3.
\end{proof}
We end by the following formal corollary of Theorem \ref{GeneralCFThm}. We will denote by $\mathbb{Z}_a$ ($\mathbb{Q}_a$) the integers (rationals) considered as an $\mathbb{L}$-algebra via the additive formal group law $x + y$.
\begin{thm}\label{GeneralGRRThm}
For each $X \in \mathcal{C}_a$, there exists a Chern character morphism
$$ch_a: K^0(X) \to \mathbb{Q}_a \otimes {\underline\Omega}'^*(X)$$
which is a homomorphism of rings and commutes with pullbacks. Moreover, given $f: X \to Y$ projective and quasi-smooth, we have that
$$
f_! \bigl( ch_a(\alpha) \bullet \mathrm{Td}(\mathbb{L}_{X/Y}) \bigr) = ch_a(f_!(\alpha))
$$
for all $\alpha \in K^0(X)$, where $\mathrm{Td}(\mathbb{L}_{X/Y})$ denotes the Todd class of $\mathbb{L}_{X/Y}$. Moreover, the induced map
$$ch_a: \mathbb{Q} \otimes K^0(X) \to \mathbb{Q} \otimes {\underline\Omega}_a^*(X)$$
is an isomorphism. The result remains true if we replace ${\underline\Omega}'^*(X)$ with $\Omega'^*(X)$.
\end{thm}
\begin{proof}
As in Section 5.2 of \cite{annala-chern}, this is a formal consequence of the Conner--Floyd theorem (Theorem \ref{GeneralCFThm}).
\end{proof}
\bibliographystyle{alphamod}
| d241185b5adc3294f56a6c6ce5dd62450d1e883c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:intro}
The present work proposes a novel smoothed particle hydrodynamics (SPH) formulation for general thermo-capillary phase change problems involving solid, liquid and gaseous phases. A special focus lies on the mesoscale melt pool modeling in metal powder bed fusion additive manufacturing (PBFAM) processes, e.g. selective laser melting (SLM) or electron beam melting (EBM), requiring some additional model constituents that are specific for this application. Since the governing physics are similar, also the melt pool dynamics in laser beam welding (LBW) or electron beam welding (EBW) processes~\cite{Chang2015,Geiger2009,Ki2002a,Ki2002b,Rai2008,Semak1999} lie in the scope of application of the proposed model.
Basically, two main modeling approaches for surface tension effects can be distinguished in the context of SPH: formulations considering the microscale origin of surface tension in form of discrete, phase-dependent inter-particle potentials~\cite{Nugent2000, Tartakovsky2005, Tartakovsky2016} as well as macroscale surface tension models relying on the continuum surface force (CSF) method proposed by Brackbill and Kothe~\cite{Brackbill1996} and widely used also in combination with other spatial discretization schemes such as finite differences, finite volumes or finite elements. The CSF approaches can be further subdivided into formulations that directly discretize the surface tension stress tensor and subsequently determine its divergence as contribution to the discrete momentum equation~\cite{Lafaurie1994, Hu2006} and formulations that rely on the divergence of the continuous surface tension stress tensor resulting in the well-known curvature-proportional surface tension forces in interface normal direction and tangential interface forces proportional to surface tension gradients. The present work will focus on the second category for which the first SPH discretization has been proposed by Morris~\cite{Morris2000}. Subsequently, this formulation has been extended by density-weighted color field gradients~\cite{Adami2010} as well as different interface reconstruction and smoothing techniques~\cite{Andersson2010, Zhang2010, Zhang2015a, Zhang2015b}. There are only very few approaches to incoorporate wetting effects into this type of SPH formulation as e.g. proposed by Breinlinger et al.~\cite{Breinlinger2013} or by Das and Das~\cite{Das2010}. One of the first SPH formulations for thermo-capillary flow, i.e. surface tension effects coupled with a thermal field, has been proposed by Tong and Browne~\cite{Tong2014} and extented by Hopp-Hirschler et al.~\cite{Hopp-Hirschler2018}. {Recently, also several SPH formulations for thermo-capillary phase change problems in the context of PBFAM melt pool modeling have been proposed~\cite{Russell2018,Wessels2018,trautmann2018numerical,Weirather2019,shah2020simulations,furstenau2020generating,dao2021simulations}}. To the best of the authors' knowledge non of the aforementioned thermo-capillary SPH formulations has incorporated wetting effects so far, which are expected, however, to play an important role on the length scales relevant for metal PBFAM.
In metal PBFAM, a focused laser beam, typically within an inert gas atmosphere, melts pre-defined contours into thin layers of pre-applied metal powder to create the cross-section of a final solid part in a repeated layer-wise buildup procedure. Under typical processing conditions the peak temperatures on the melt pool surface exceed the boiling temperature of the liquid metal. The density jump and accompanied recoil pressure in the phase transition from liquid metal to metal vapor results in a considerable distortion and highly dynamic topology changes of the liquid-gas interface at the melt pool surface giving rise to defects such as spatter, i.e. ejection of melt drops, or pores, i.e. gas bubble inclusions~\cite{Meier2017}. Pioneering modeling approaches in this field are e.g. given by the thermo-hydrodynamics finite element model proposed by Khairallah et al.~\cite{Khairallah2014, Khairallah2016, Khairallah2020}, who considered temperature-dependent surface tension and evaporation-induced recoil pressure forces, based on a phenomenological recoil pressure model~\cite{Anisimov1995}, as primary driving forces of the process. Comparable models based on finite difference, finite volume, finite element, Lattice Boltzmann or meshfree discretizations are e.g. given by~\cite{Geiger2009,Russell2018, Weirather2019, Lee2015, Leitz2018, Qiu2015, Panwisawas2017, Wessels2018, Otto2012, Yan2018, egorov2020, Gurtler2013,Yu2016,Yuan2015}. A more refined model has been proposed by~\cite{Tan2013, Tan2014, Kouraytem2019}, where the gas / vapor phase is explicitly resolved. Typically, the aforementioned models do not account for wetting effects at the triple line solid-liquid-gas. On the contrary, the works~\cite{Korner2011, Korner2013, Markl2015} specifically focus on the interplay between wetting effects and different power particle configurations, without considering however evaporation-induced recoil pressure.
The present work proposes a weakly compressible SPH formulation for thermo-capillary phase change problems involving solid, liquid and gaseous phases. Specifically, evaporation-induced recoil pressure, temperature-dependent surface tension and wetting forces are considered as liquid-gas interface fluxes in the Navier Stokes equation. In the thermal problem, a Gaussian laser beam heat source as well as evaporation-induced heat losses are considered as liquid-gas interface fluxes, while convection boundary conditions are obsolete due to the explicit modeling of the atmospheric gas phase. All mechanical and thermal interface fluxes are modeled in a diffuse sense in analogy to the CSF approach. The following original contributions of the present work can be identified: The first SPH formulation for thermo-capillary problems is proposed that also considers wetting effects. A novel interface stabilization scheme based on viscous interface forces is proposed, which is shown to allow for a stable and smooth liquid-gas interface by effectively damping spurious interface flows well-known for the CSF approach. Moreover, different SPH discretizations for the tangential projection of the temperature gradient, as required for the discrete Marangoni forces, are reviewed. Based on a thorough analysis it is shown that standard \textit{two-sided} gradient approximations are sufficient for this purpose as long as zero-order consistency is satisfied, e.g. by anti-symmetric gradient construction. In the context of metal AM melt pool modeling, the present approach is - to the best of the authors' knowledge - the first model that i) considers the full range of relevant interface forces consisting of evaporation-induced recoil pressure, temperature-dependent surface tension and wetting forces, and ii) resolves the atmospheric gas phase and, thus, can consistently account for defects such as gas inclusions.
The remainder of this work is organized as follows: Section~\ref{sec:goveq} presents the governing equations, i.e. continuity equation, momentum equation, energy equation and equation of state, in space-continuous form. Discretization in space, based on SPH, and in time, based on an explicit velocity-Verlet scheme, is presented in Sections~\ref{sec:nummeth_sph} and~\ref{sec:nummeth_sph_timint}. In Section~\ref{sec:numex_tantempgrad} different SPH approximations for the tangential temperature gradient are thoroughly analyzed and compared. {Finally, in Section~\ref{sec:numex}, the accuracy of the individual model and method components is verified by means of selected benchmark examples with analytical/numerical reference solutions. Eventually, the suitability of the proposed melt pool model for typical metal AM application scenarios is verified by means of point and line melting examples with and without resolved powder particles. Here, a special focus lies on the robustness of the computational model, i.e. the ability to represent challenging and practically relevant scenarios of dynamically changing interface topologies (e.g. generation of melt spatter or gas inclusions) without inducing spurious interface flows or instabilities of the discretization scheme.}
\section{Governing Equations} \label{sec:goveq}
Throughout this work two-phase flow problems of a liquid phase $\Omega^{l}$ and a gas phase $\Omega^{g}$ are considered that interact with a solid phase $\Omega^{s}$ and allow for reversible phase transition between liquid and solid phase. The overall problem domain splits according to $\Omega = \Omega^{l} \cup \Omega^{g} \cup \Omega^{s}$ and the two-phase fluid domain is given by $\Omega^{f} = \Omega^{l} \cup \Omega^{g}$. In the context of metal AM melt pool modeling the solid, liquid and gas phase correspond to the solid metal, the molten metal and the atmospheric gas in the build chamber
of an AM device.
\subsection{Fluid phases} \label{subsec:goveq_fluid}
The liquid and gas phase are governed by the \textit{weakly compressible}, instationary and anisothermal Navier-Stokes equations in the domain~$\Omega^{f}=\Omega^{l} \cup \Omega^{g}$. The problem shall be described by the continuity equation
\begin{equation} \label{eq:fluid_conti}
\dv{\rho}{t} = -\rho \div \vectorbold{u} \qin \Omega^{f},
\end{equation}
the Navier-Stokes momentum equation
\begin{equation} \label{eq:fluid_momentum}
\dv{\vectorbold{u}}{t} = \frac{1}{\rho} \left(-\grad{p} + \vectorbold{f}_{\nu} + \tilde{\vectorbold{f}}^{{lg}}_{s} + \tilde{\vectorbold{f}}^{{slg}}_{w} + \tilde{\vectorbold{f}}^{{lg}}_{v} \right) + \vectorbold{g} \qin \Omega^{f},
\end{equation}
as well as the energy equation:
\begin{equation} \label{eq:fluid_energy}
c _p\dv{T}{t} = \frac{1}{\rho} \left(-\div \vectorbold{q} + \tilde{s}^{lg}_v + \tilde{s}^{lg}_l \right) \qin \Omega.
\end{equation}
Following a weakly compressible approach, density~$\rho$ and pressure~$p$ are linked via the equation of state
\begin{equation} \label{eq:fluid_eos}
p\qty(\rho) = c^{2} \qty(\rho - \rho_{0}) = p_{0} \qty(\frac{\rho}{\rho_{0}} - 1) \qin \Omega^{f},
\end{equation}
which closes the system of equations for the six unknowns velocity $\vectorbold{u}$ (three components), density~$\rho$, pressure~$p$ and temperature~$T$. The individual contributions to these equations
will be discussed in the following.
\subsubsection{Momentum equation} \label{subsec:goveq_fluid_momentum}
In equation~\eqref{eq:fluid_momentum}, contributions from viscous forces~$\vectorbold{f}_{\nu}$, surface tension forces~$\tilde{\vectorbold{f}}^{{lg}}_{s}$, wetting forces~$\tilde{\vectorbold{f}}^{{slg}}_{w}$ and evaporation-induced recoil pressure forces $\tilde{\vectorbold{f}}^{{lg}}_{v}$, each per unit volume, as well as body forces~$\vectorbold{g}$ per unit mass, can be identified. For incompressible Newtonian fluids the viscous forces read
\begin{equation}
\vectorbold{f}_{\nu} = \eta \laplacian{\vectorbold{u}},
\end{equation}
with dynamic viscosity~$\eta$. Following the continuum surface force (CSF) approach by Brackbill and Kothe~\cite{Brackbill1996} we consider surface tension and wetting effects in the momentum equation~\eqref{eq:fluid_momentum} as volumetric forces distributed across an interfacial volume of finite width instead of additional boundary conditions at the liquid-gas interface area and the triple line
solid-liquid-gas. In the following, these interface forces are marked by a tilde symbol and by a superscript indicating the relevant interface, e.g. $\tilde{\vectorbold{f}}^{{lg}}$ for forces on the 2D liquid-gas interface or $\tilde{\vectorbold{f}}^{{slg}}$ for forces on the 1D solid-liquid-gas interface (triple line). Specifically, the distributed surface tension forces consist of the following two contributions in interface normal and tangential direction
\begin{equation} \label{eq:fluid_surfacetension}
\tilde{\vectorbold{f}}^{{lg}}_{s} = -\alpha \kappa \vectorbold{n}^{lg} \delta^{lg} + \left(\vectorbold{I} - \vectorbold{n}^{lg} \otimes \vectorbold{n}^{lg} \right) \grad{\alpha} \delta^{lg},
\end{equation}
with the surface tension coefficient $\alpha$, the interface curvature $\kappa:=\div \vectorbold{n}^{lg}$, the liquid-gas interface normal $\vectorbold{n}^{lg}:= \grad{c^{lg}}/||\grad{c^{lg}}||$, the phase-specific color field $c^{lg}$ between the liquid and gas phase, to be defined in Section~\ref{subsec:nummeth_sph_colorfield}, as well as the surface delta function $\delta^{lg}:=||\grad{c^{lg}}||$ between liquid and gas phase. The surface delta function is employed to distribute interface surface forces across interface domains of finite thickness. It is non-zero only on these interface domains and its integral over the interface thickness direction is normalized to one (see also Section~\ref{subsec:nummeth_sph_colorfield}). Throughout this work a purely temperature-dependent surface tension coefficient, i.e. $\grad{\alpha}= \alpha'(T) \grad{T}$ with $\alpha'(T)=d \alpha(T) / d T$, is considered. Specifically, a linear temperature-dependence of the surface tension is considered in the examples in Section~\ref{sec:numex} according to
\begin{equation} \label{eq:fluid_surfacetension_tempdepend}
\alpha(T) = \alpha_0 - \alpha'_0 (T-T_{\alpha_0}),
\end{equation}
where $\alpha_0$ is the surface tension at reference temperature $T_{\alpha_0}$. Moreover, the wetting forces~{\cite{Breinlinger2013}} are given by
\begin{equation} \label{eq:fluid_wetting}
\tilde{\vectorbold{f}}^{{slg}}_{w} = \alpha \left( \cos \theta - \cos \theta_0 \right) \vectorbold{t}^{sf} \delta^{lg} \delta^{sf} ,
\end{equation}
with the equilibrium wetting angle $\theta_0$ and the current wetting angle $\theta$ defined via $\cos \theta := \vectorbold{n}^{lg} \cdot \vectorbold{n}^{sf}$. Here, the solid-fluid interface normal vector between the domains $\Omega^{s}$ and $\Omega^{f}$ is defined as $\vectorbold{n}^{sf}:= \grad{c^{sf}}/||\grad{c^{sf}}||$ on the basis of a phase-specific color field $c^{sf}$ between the solid and fluid phase to be defined in Section~\ref{subsec:nummeth_sph_colorfield}. Similar to the liquid-gas interface, also the surface delta function of the solid-fluid interface follows the relation $\delta^{sf}:=||\grad{c^{sf}}||$. Moreover, the tangent vector $\vectorbold{t}^{sf}$ is defined as the projection of the liquid-gas interface normal vector $\vectorbold{n}^{lg}$ onto the solid-fluid interface surface, {defined by its normal vector $\vectorbold{n}^{sf}$~\cite{Breinlinger2013}:}
\begin{equation} \label{eq:fluid_wetting2}
\vectorbold{t}^{sf} = \frac{\vectorbold{n}^{lg} - (\vectorbold{n}^{lg} \cdot \vectorbold{n}^{sf}) \vectorbold{n}^{sf}}{|| \vectorbold{n}^{lg} - (\vectorbold{n}^{lg} \cdot \vectorbold{n}^{sf}) \vectorbold{n}^{sf} ||}.
\end{equation}
Besides these standard capillary force contributions, the high peak temperatures at the melt pool surface at typical metal AM process conditions give rise to considerable evaporation effects. As common in the modeling of these processes, a phenomenological model for the evaporation-induced recoil pressure forces acting on the melt pool surface according to the work by Anisimov~\cite{Anisimov1995} is employed:
\begin{equation} \label{eq:fluid_recoil}
\tilde{\vectorbold{f}}^{{lg}}_{v} = -p_v(T) \vectorbold{n}^{lg} \delta^{lg} \quad \text{with} \quad p_v(T) = C_P \exp \left[ - C_T \left( \frac{1}{T} - \frac{1}{T_v} \right)\right],
\end{equation}
where the constants $C_P = 0.54 p_a$ and $C_T=\bar{h}_v/R$ contain the atmospheric pressure $p_a$, the molar latent heat of evaporation $\bar{h}_v$ and the molar gas constant $R$. Moreover, $T_v$ is the boiling temperature.
\begin{rmk}
The working principle of the employed phenomenological evaporation model relies on local peak temperatures in the melt pool that exceed the boiling temperature of the liquid metal. In scenarios with very high laser powers the resulting peak temperatures might lead to negative surface tension coefficients if~\eqref{eq:fluid_surfacetension_tempdepend} is applied without further correction. Thereto, in this work a regularized version of~\eqref{eq:fluid_surfacetension_tempdepend} is applied which keeps the surface tension coefficient $\alpha(T)$ constant if it has already decreased to $10 \%$ of its reference value $\alpha_0$ and the temperature is further increasing.
\end{rmk}
\subsubsection{Energy equation} \label{subsec:goveq_fluid_energy}
The energy equation~\eqref{eq:fluid_energy} contains the mass-specific heat capacity $c_p$, the heat flux $\vectorbold{q}:= -k \grad{T}$ according to Fourier's law with thermal conductivity $k$ as well as heat fluxes stemming from the laser beam heat source $\tilde{s}^{lg}_l$ and from evaporation-induced heat losses $\tilde{s}^{lg}_v$, each per unit volume. The former is given by
\begin{equation} \label{eq:fluid_heatsource}
\tilde{s}^{lg}_l = \chi_l <\! -\vectorbold{n}^{lg} \! \cdot \! \vectorbold{e}_{l} \! > \, s^{lg}_l(\vectorbold{x}) \, \delta^{lg} \quad \text{with} \quad s^{lg}_l(\vectorbold{x})=s^{lg}_{l0} \, \exp \! \left[- 2 \left( \frac{||\vectorbold{x}-\vectorbold{x}_0||}{r_w} \right)^2 \right],
\end{equation}
where the Macauley bracket $<...>$ returns the value of its argument if the argument is positive and zero otherwise. The irradiance $s^{lg}_l(\vectorbold{x})$ describes the incident laser power per unit area at position $\vectorbold{x}$ as a function of the laser beam center position $\vectorbold{x}_0$ and has the form of a Gauss distribution, from which $s^{lg}_{l0}$ is the peak value and $r_{w}=2 \sigma$ represents two times the standard deviation $\sigma$. The corresponding diameter $d_w=2 r_{w}$ is a frequently used measure for the effective laser beam diameter. In addition, $\vectorbold{e}_{l}$ is the unit vector representing the laser beam direction and $\chi_l$ the laser energy absorptivity. Eventually, following the same phenomenological model as for the recoil pressure~\eqref{eq:fluid_recoil}, the evaporation-induced heat loss reads
\begin{equation} \label{eq:fluid_evaporation}
\tilde{s}^{lg}_v = s^{lg}_v \, \delta^{lg} \quad \text{with} \quad s^{lg}_v = - \dot{m}^{lg}_v [h_v+ h(T)], \quad \dot{m}^{lg}_v = 0.82 \, c_s \, p_v(T) \, \sqrt{\frac{C_M}{T}}, \quad h(T)=\int \limits_{T_{h,0}}^T c_p \, \, d\bar{T},
\end{equation}
where the enthalpy rate per unit area $s^{lg}_v$ results from the vapor mass flow per unit area $\dot{{m}}^{lg}_v$ at the melt pool surface and the sum of the specific enthalpy $h(T)$ and the latent heat of evaporation $h_v$, both per unit mass. Moreover, $T_{h,0}$ is a reference temperature of the specific enthalpy and the constant $C_M=M/(2\pi R)$ contains the molar mass $M$ and the molar gas constant $R$. Finally, $p_v(T)$ is the recoil pressure defined in~\eqref{eq:fluid_recoil} and $c_s$ the so-called sticking constant which takes on a value close to one, i.e. $c_s \approx 1$ for metals~\cite{Khairallah2016, Weirather2019}.
\begin{rmk}
According to equation~\eqref{eq:fluid_heatsource} the laser beam heat source is applied at the liquid-gas interface, which is characterized by $\delta^{lg} \neq 0$. In the proposed model, the laser heat source also acts on solid surfaces, i.e. at solid-gas interfaces characterized by $\delta^{sg} \neq 0$. Here, the surface delta function $\delta^{sg}$ of the solid-gas interface is defined in analogy to the surface delta function $\delta^{lg}$ of the liquid-gas interface.
\end{rmk}
\subsubsection{Equation of state} \label{subsec:goveq_fluid_eos}
In the equation of state~\eqref{eq:fluid_eos} the reference density~$\rho_{0}$, which equals the initial density throughout this work, the reference pressure~$p_{0} = \rho_{0} c^{2}$ and the artificial speed of sound~$c$ can be indentified. Note that the commonly applied weakly compressible approach only represents deviations from the reference pressure, i.e., $p\qty(\rho_{0}) = 0$, and not the total pressure. In order to limit density fluctuations to an acceptable level, while still avoiding too severe time step restrictions, Morris et al.~\cite{Morris1997} discussed strategies on how to determine an appropriate value of the artificial speed of sound.
\subsection{Solid phase} \label{subsec:goveq_solid}
Since the focus of this work lies on melt pool thermo-hydrodynamics, the assumption of a rigid and immobile solid phase (substrate and powder grains in the context of PBFAM processes), which is typical for
mesoscale PBFAM models, is made. Thus, only the energy equation~\eqref{eq:fluid_energy} is solved for the solid phase.
\subsection{Phase transition} \label{subsec:goveq_phase_transition}
As presented in Section~\ref{sec:nummeth_sph}, the spatial discretization will be based on smoothed particle hydrodynamics (SPH). Due the Lagrangian nature of this scheme, each (material) particle directly carries its phase information. Based on this information, the corresponding field equations with phase-specific parameter values are evaluated for each particle. Material particles undergo the phase transition \textit{solid} $\leftrightarrow$ \textit{liquid} when passing the melt temperature $T_m$. Since the vapor phase is not modeled explicitly, the phase transition \textit{liquid} $\leftrightarrow$ \textit{vapor} is only considered implicitly in terms of evaporation-induced recoil pressure forces~\eqref{eq:fluid_recoil} and heat losses~\eqref{eq:fluid_evaporation}. While the latent heat of evaporation is already contained~\eqref{eq:fluid_evaporation}, the latent heat of melting could be considered in a straightforward manner as well by employing e.g. an apparent capacity scheme relying on an increased heat capacity $c_p$ within a finite temperature interval~\cite{proell2020phase}. However, for simplicity temperature-independent parameter values are considered in the present work (except for the surface tension coefficient).
\subsection{Initial and boundary conditions}
In general, the system of partial differential equations~\eqref{eq:fluid_conti}-\eqref{eq:fluid_energy} is subject to the following initial conditions
\begin{equation}
\rho = \rho_0, \quad \vectorbold{u} = \vectorbold{u}_{0}, \quad T=T_0 \qin \Omega \qq{at} t = 0.
\end{equation}
Throughout this work, only systems that are initially in static equilibrium, i.e. $\vectorbold{u}_{0}=\vectorbold{0}$, are considered.
In addition, Dirichlet and Neumann boundary conditions are required on the domain boundary $\Gamma = \partial\Omega$:
\begin{equation}
\vectorbold{u} = \hat{\vectorbold{u}} \qq{on} \Gamma^{\vectorbold{u}}_{D}, \quad \quad
\vectorbold{t} = \hat{\vectorbold{t}} \qq{on} \Gamma^{\vectorbold{u}}_{N}, \quad \quad
T = \hat{T} \qq{on} \Gamma^T_{D}, \quad \quad
\vectorbold{q} = \hat{\vectorbold{q}} \qq{on} \Gamma^T_{N} \, ,
\end{equation}
with boundary velocity~$\hat{\vectorbold{u}}$, boundary traction~$\hat{\vectorbold{t}}$, boundary temperature $T$ and boundary heat flux $\hat{\vectorbold{q}}$ on the Dirichlet and Neumann boundaries $\Gamma = \Gamma^{\vectorbold{u}}_{D} \cup \Gamma^{\vectorbold{u}}_{N}$ and $\Gamma^{\vectorbold{u}}_{D} \cap \Gamma^{\vectorbold{u}}_{N} = \emptyset$ as well as $\Gamma = \Gamma^T_{D} \cup \Gamma^T_{N}$ and $\Gamma^T_{D} \cap \Gamma^T_{N} = \emptyset$.
\section{Spatial discretization via smoothed particle hydrodynamics} \label{sec:nummeth_sph}
\subsection{Approximation of field quantities via smoothing kernel} \label{subsec:nummeth_sph_kernel}
The fundamental concept of SPH is based on the approximation of a field quantity~$f$ via a smoothing operation and on the discretization of the domain~$\Omega$ with discretization points following the fluid motion and therefore being denoted as particles. Introducing a smoothing kernel~$W\qty(r,h)$ (see e.g.~\cite{Liu2010,Monaghan2005}) leads to the following sequence of approximations of an arbitrary field quantity~$f$:
\begin{equation}
\label{eq:sph_approximation}
f\qty(\vectorbold{r}) \approx \int_{\Omega} f\qty(\vectorbold{r}') W\qty(\qty| \vectorbold{r} - \vectorbold{r}' |, h) \dd{\vectorbold{r}'} \approx \sum_{j} V_{j} f\qty(\vectorbold{r}_{j}) W\qty(\qty| \vectorbold{r} - \vectorbold{r}_{j} |, h),
\end{equation}
commiting a \textit{smoothing error} in the first and an \textit{integration error} in the second approximation step~\cite{Quinlan2006}. The smoothing kernel~$W\qty(r,h)$ is a monotonically decreasing, smooth function that dependents on the distance~$r$ from the kernel center and a smoothing length~$h$. The smoothing length~$h$ together with a scaling factor~$\kappa$ define the support radius of the smoothing kernel~$r_{c} = \kappa h$. The Dirac delta function property $\lim_{h \rightarrow 0}{ W\qty(r, h) } = \delta\qty(r)$ ensures an exact representation of a field quantity~$f$ in the limit $h \rightarrow 0$. Compact support, i.e., $W\qty(r, h) = 0$ for $r > r_{c}$, as well as positivity, i.e., $W\qty(r, h) \geq 0$ for $r \leq r_{c}$, are typical properties of standard smoothing kernels~$W\qty(r,h)$. In addition, the normalization property requires that $\int_{\Omega} W\qty(\qty| \vectorbold{r} - \vectorbold{r}' |, h) \dd{\vectorbold{r}'} = 1$. In the second approximation step, the domain integral is replaced by a summation over discrete volumes~$V_{j}$ located at the positions of the material discretization points (particles)~$j$. A straightforward approximation for the gradient of quantity~$f$ follows directly by differentiation of \eqref{eq:sph_approximation}:
\begin{equation}
\label{sph_general_gradient}
\grad{f}\qty(\vectorbold{r}) \approx \int_{\Omega} f\qty(\vectorbold{r}') \grad{W \qty(\qty| \vectorbold{r} - \vectorbold{r}' |, h)} \dd{\vectorbold{r}'} \approx \sum_{j} V_{j} f\qty(\vectorbold{r}_{j}) \grad{W \qty(\qty| \vectorbold{r} - \vectorbold{r}_{j} |, h)},
\end{equation}
Note that this (simple) gradient approximation shows some particular disadvantages, hence, more advanced approximations for gradients are given in the literature \cite{Monaghan2005} and will also be applied in the following. Applying these gradient approximations reduces the partial differential equations~\eqref{eq:fluid_conti} and~\eqref{eq:fluid_energy} to ordinary differential equations that are solved, i.e., evaluated and integrated in time, for all particles~$i$ in the domain~$\Omega$ (cf. Sections \ref{subsec:nummeth_sph_momentum}-\ref{sec:nummeth_sph_timint}). As a result, all fluid quantities are evaluated at and associated with particle positions, meaning each particle carries its corresponding fluid quantities.
\begin{rmk}
In the following, a quantity~$f$ evaluated for particle~$i$ at position~$\vectorbold{r}_{i}$ is written as~$f_{i} = f\qty(\vectorbold{r}_{i})$. In addition, the short notation $W_{ij} = W\qty(r_{ij}, h)$ denotes the smoothing kernel~$W$ evaluated for particle~$i$ at position~$\vectorbold{r}_{i}$ with neighboring particle~$j$ at position~$\vectorbold{r}_{j}$, where $r_{ij} = \qty|\vectorbold{r}_{ij}| = \qty|\vectorbold{r}_{i} - \vectorbold{r}_{j}|$ is the absolute distance between particles~$i$ and~$j$. Similarly, the derivative of the smoothing kernel~$W$ with respect to the absolute distance~$r_{ij}$ is denoted by $\pdv*{W}{r_{ij}} = \pdv*{W\qty(r_{ij}, h)}{r_{ij}}$.
\end{rmk}
\begin{rmk}
In this work, a quintic spline smoothing kernel~$W\qty(r, h)$ is considered as defined in \cite{Morris1997} with smoothing length~$h$ and compact support with support radius~$r_{c} = \kappa h$ and scaling factor~$\kappa = 3$.
\end{rmk}
\subsection{Initial particle spacing} \label{subsec:nummeth_sph_spacing}
Within this contribution, the domain $\Omega$ is initially filled with particles located on a regular grid with particle spacing~$\Delta{}x$, thus in $d$ dimensional space each particle initially occupies an effective volume of~$\qty(\Delta{}x)^{d}$. The mass of a particle~$i$ is then set using the reference density according to $m_{i} = \rho_{0} \qty(\Delta{}x)^{d}$ and remains constant throughout the simulation. In general, the initial particle spacing~$\Delta{}x$ can be freely chosen, however, within this work the initial particle spacing~$\Delta{}x$ is set equal to the smoothing length~$h = \flatfrac{r_{c}}{\kappa}$.
\subsection{Density summation} \label{subsec:nummeth_sph_summation}
The density of a particle~$i$ is determined via summation of the respective smoothing kernel contributions of all neighboring particles~$j$ within the support radius~$r_{c}$
\begin{equation} \label{eq:sph_densum}
\rho_{i} = m_{i} \sum_{j} W_{ij} \, .
\end{equation}
This approach is typically denoted as density summation and results in an exact conservation of mass in the fluid domain, which can be shown in a straight-forward manner considering the commonly applied normalization of the smoothing kernel to unity. It shall be noted that the density field may alternatively be obtained via SPH discretization and time integration of the continuity equation~\eqref{eq:fluid_conti}, cf. Liu and Liu~\cite{Liu2010}.
\subsection{Momentum equation} \label{subsec:nummeth_sph_momentum}
Following the standard SPH discretization procedure the discrete version of~\eqref{eq:fluid_momentum} can be formulated as
\begin{equation} \label{eq:sph_momentum}
\vectorbold{a}_{i} = \frac{1}{m_{i}} \left[ \vectorbold{F}_{p,i} + \vectorbold{F}_{\nu,i} + \vectorbold{F}_{s,i} + \vectorbold{F}_{w,i} + \vectorbold{F}_{v,i} + \vectorbold{F}_{d,i} \right] + \vectorbold{b}_{i}
\, ,
\end{equation}
where $\vectorbold{a}_{i} = \dv*{\vectorbold{u}_{i}}{t}$ represents the total acceleration of particle~$i$ whereas the pressure force $\vectorbold{F}_{p,i}$, viscous force $\vectorbold{F}_{\nu,i}$, surface tension force $\vectorbold{F}_{s,i}$, wetting force $\vectorbold{F}_{w,i}$ as well as vapor-induced recoil pressure force $\vectorbold{F}_{v,i}$ acting on particle~$i$ result from summation of all particle-to-particle interaction contributions with neighboring particles~$j$. Optionally, additional viscous dissipation forces $\vectorbold{F}_{d,i}$ are applied at the interfaces solid-liquid and liquid-gas, which will motivated and further detailled in Section~\ref{subsec:nummeth_sph_dissipation}. The pressure and viscous forces in the momentum equation~\eqref{eq:sph_momentum} are discretized following a formulation proposed by Adami et al.~\cite{Adami2012, Adami2013}:
\begin{equation} \label{eq:sph_momentum_pressure_and_viscous}
\vectorbold{F}_{p,i} + \vectorbold{F}_{\nu,i} = \sum_{j} \qty(V_{i}^{2}+V_{j}^{2}) \qty[ - \bar{p}_{ij} \pdv{W}{r_{ij}} \vectorbold{e}_{ij} + \bar{\eta}_{ij} \frac{\vectorbold{u}_{ij}}{r_{ij}} \pdv{W}{r_{ij}} ]\, ,
\end{equation}
with volume~$V_{i} = m_{i}/\rho_{i}$ of particle~$i$, unit vector $\vectorbold{e}_{ij} = \flatfrac{\vectorbold{r}_{i} - \vectorbold{r}_{j}}{\qty|\vectorbold{r}_{i} - \vectorbold{r}_{j}|} = \flatfrac{\vectorbold{r}_{ij}}{r_{ij}}$, relative velocity $\vectorbold{u}_{ij} = \vectorbold{u}_{i}-\vectorbold{u}_{j}$ as well as inter-particle averaged pressure and dynamic viscosity:
\begin{equation} \label{eq:sph_mom_wght_press_and_visc}
\bar{p}_{ij} = \frac{\rho_{j}p_{i}+\rho_{i}p_{j}}{\rho_{i} + \rho_{j}} \, , \quad \quad \bar{\eta}_{ij} = \frac{2\eta_{i}\eta_{j}}{\eta_{i}+\eta_{j}} \, .
\end{equation}
Also the transport velocity formulation proposed in~\cite{Adami2013}, which utilizes a constant background pressure $p_b$ to suppress the problem of tensile instability, is employed in the present work. For the sake of briefity, the definition of the modified advection velocity and the additional terms in the momentum equation stemming from the aforementioned transport velocity contribution are not further delineated and the reader is kindly referred to the original publication~\cite{Adami2013}. The accuracy and stability of this formulation has readily been demonstrated on the basis of well-known benchmark tests in the context of computational fluid dynamics~\cite{Adami2013} and fluid-structure interaction~\cite{Fuchs2020}. The remaining force contributions will be discussed in the following.
\subsubsection{Discretization of phase interfaces} \label{subsec:nummeth_sph_colorfield}
The interface forces to be defined in the following rely on a representation of the different phase domains and interfaces via a color field function. Here, we define the color field according to
\begin{equation} \label{eq:sph_colorfield}
c_i^j = \left\{\begin{array}{ll}
1 , & \text{if particles i and j belong to different phases}\\
0, & \text{else}
\end{array}\right.
\end{equation}
as well as the density-weighted color field gradient according to
\begin{equation} \label{eq:sph_colorfieldgradient}
\grad{c}_i = \frac{1}{V_{i}} \sum_{j} \qty(V_{i}^{2}+V_{j}^{2}) \bar{c}_{ij} \pdv{W}{r_{ij}} \vectorbold{e}_{ij}
\quad \text{with} \quad \bar{c}_{ij} = \frac{\rho_j}{\rho_i+\rho_j} c_i^i + \frac{\rho_i}{\rho_i+\rho_j} c_j^i
\end{equation}
following the approach proposed by Adami et al.~\cite{Adami2010}. Note that $ c_i^i \equiv 0$ according to~\eqref{eq:sph_colorfield}. Based on the definition of the color field gradient
the interface normal and the surface delta function of particle $i$ read
\begin{equation} \label{eq:sph_normalanddelta}
\vectorbold{n}_{i} = \frac{\grad{c}_i}{|| \grad{c}_i ||} \quad \text{and} \quad \delta_i = || \grad{c}_i ||.
\end{equation}
Note that this procedure leads to an \textit{outward-pointing} interface normal vector with respect to the phase of particle $i$. Moreover, it has to be emphasized that these metrics are exclusively used to define the interface between two phases. Consequently, in the calculations according to~\eqref{eq:sph_colorfield}-\eqref{eq:sph_normalanddelta} only two different phases are distinguished (and \textit{not} three independent phases which might occur at the triple line solid-liquid-gas). Specifically, the liquid-gas interface (superscript $"lg"$) is defined by only considering particles of the liquid and the gas phase (i.e. no contribution of solid particles). The solid-gas interface (superscript $"sg"$) is defined by only considering particles of the solid and the gas phase. The solid-fluid interface (superscript $"sf"$) is defined by considering all particles but only distinguishing between either particles of the solid phase or the (combined) fluid phase (sum of particles from the liquid and gas phase).
\begin{rmk}
In case of high density ratios between two phases the definition of $\delta_i$ according to~\eqref{eq:sph_colorfieldgradient}-\eqref{eq:sph_normalanddelta} ensures that the majority of a flux contribution (i.e. of mechanical interface forces or thermal heat fluxes) distributed over the interface via $\delta_i$ acts on the side of the interface associated with the \textit{heavier} phase. Thus, in the melt pool examples in Section~\ref{sec:numex_am}, where typical density ratios $\rho_l / \rho_g > 100$ between melt and gas phase are considered, more than $99\%$ of a flux term is carried by the melt phase. Therefore, for simplicity only the contributions to the melt phase are considered for the mechanical and thermal flux terms in~\eqref{eq:fluid_surfacetension},~\eqref{eq:fluid_wetting},~\eqref{eq:fluid_recoil},~\eqref{eq:fluid_heatsource} and~\eqref{eq:fluid_evaporation}. This approximation seems reasonable when considering typical parameter uncertainties and accuracy requirements in AM melt pool modeling. However, it is emphasized that this procedure is by no means an inherent limitation of the proposed methodology and can easily be changed.
\end{rmk}
\subsubsection{Surface tension forces} \label{subsec:nummeth_sph_surface tension}
In the following, the surface tension forces are split into two contributions $\vectorbold{F}_{s,i}=\vectorbold{F}_{s \kappa,i}+\vectorbold{F}_{s m,i}$, with $\vectorbold{F}_{s \kappa,i}$ representing the curvature-proportional surface tension normal forces and $\vectorbold{F}_{s m,i}$ representing the tangential Marangoni forces due to surface tension gradients. The first contribution is given by
\begin{equation} \label{eq:sph_surfacetension1}
\vectorbold{F}_{s \kappa,i} = -V_i \alpha_i \kappa_i \vectorbold{n}_i^{lg} \delta_i^{lg}.
\end{equation}
As proposed by Morris~\cite{Morris2000}, the curvature $\kappa_i$ is discretized according to
\begin{equation} \label{eq:sph_curvature}
\kappa_i = (\div{\vectorbold{n}^{lg}})_i = -\frac{\sum_{j} N_i N_j V_{j} \vectorbold{n}^{lg}_{ij} \pdv{W}{r_{ij}} \vectorbold{e}_{ij}}{\sum_{j} N_i N_j V_{j} W_{ij}}
\quad \text{with} \quad
N_k = \left\{\begin{array}{ll}
1 , & \text{if} || \grad{c}^{lg}_k || > \epsilon\\
0, & \text{else}
\end{array}\right.
\end{equation}
with $\vectorbold{n}^{lg}_{ij}=\vectorbold{n}^{lg}_{i}-\vectorbold{n}^{lg}_{j}$. Here, $\epsilon \ll 1$ is a user-defined tolerance applied to avoid contributions from particles far away from the interface with erroneous normal vectors~\cite{Morris2000}. The discrete Marangoni forces read
\begin{equation} \label{eq:sph_surfacetension2}
\vectorbold{F}_{s m,i} = V_i \underbrace{\left(\vectorbold{I} - \vectorbold{n}^{lg}_i \otimes \vectorbold{n}^{lg}_i \right) (\grad{T})_i}_{\nabla_T {T}_i} \, \alpha_i' \delta^{lg}_i
\end{equation}
where the operator $\nabla_T$ represents the projection of the nabla operator into the interface tangential plane. Liquid-gas interfaces typically go along with considerably jumps in the mechanical and thermal constitutive parameters. It will be demonstrated in Section~\ref{sec:numex_tantempgrad} that a proper
discretization of the temperature gradient $(\grad{T})_i$ is of upmost importance to represent the temperature field, and its inherent kink across the interface, with sufficient accuracy. Thereto, three
different SPH gradient approximations typically applied in the literature (see also~\cite{Tong2014, Russell2018} in the context of tangential (Marangoni) surface tension forces) shall be considered in the following. The first variant is given by the standard SPH gradient approximation according to~\eqref{sph_general_gradient}:
\begin{equation} \label{eq:sph_tempgrad1}
(\grad{ T})_i \approx \sum_{j} V_{j} T_{j} \pdv{W}{r_{ij}} \vectorbold{e}_{ij} \,.
\end{equation}
The second variant is given by a symmetric gradient approximation as typically applied to gradients (e.g. of the pressure field) in the momentum equation to guarantee for conservation of momentum~\cite{Adami2012, Adami2013}:
\begin{equation} \label{eq:sph_tempgrad2}
(\grad{ T})_i \approx \frac{1}{V_{i}} \sum_{j} (V_{i}^2+V_{j}^2) \frac{T_{i}+T_{j}}{2} \pdv{W}{r_{ij}} \vectorbold{e}_{ij} \,.
\end{equation}
The third variant is given by an anti-symmetric gradient approximation as typically applied to velocity gradients in the continuity equation to guarantee for zero-order consistency~\cite{Monaghan2005}:
\begin{equation} \label{eq:sph_tempgrad3}
(\grad{ T})_i \approx \sum_{j} V_{j} (T_{j}- T_{i}) \pdv{W}{r_{ij}} \vectorbold{e}_{ij} \,.
\end{equation}
In Section~\ref{sec:numex_tantempgrad} it will be demonstrated that only the third variant~\eqref{eq:sph_tempgrad3} leads to reasonable approximations and small discretization errors for the tangential projection of the temperature gradient. Consequently, this variant is applied in all the remaining examples in Section~\ref{sec:numex}. Moreover, it will be derived why this formulation results in a good approximation quality for the tangential projection of the temperature gradient - as required in~\eqref{eq:sph_surfacetension2} - even though it is not suitable to represent the total temperature gradient.
\subsubsection{Wetting forces} \label{subsec:nummeth_sph_wetting}
As indicated by Breinlinger et al.~\cite{Breinlinger2013}, a direct SPH discretization of~\eqref{eq:fluid_wetting} is typically not preferable due to the erroneous representation of the liquid-gas interface normal
vector $\vectorbold{n}^{lg}$ close to the triple line solid-liquid-gas, which can be traced back to a lack of liquid and gas particle support in this region. Therefore, we follow an alternative strategy
proposed in~\cite{Breinlinger2013} prescribing this normal vector in the triple line region on the basis of the equilibrium wetting angle $\theta_0$ according to:
\begin{equation} \label{sph:fluid_wetting1}
\hat{\vectorbold{n}}^{lg}_i = \vectorbold{t}^{sf}_i \sin \theta_0 - \vectorbold{n}^{sf}_i \cos \theta_0,
\end{equation}
where $\vectorbold{n}^{sf}_i$ is the normal vector of the solid-fluid interface according to~\eqref{eq:sph_colorfield}-\eqref{eq:sph_normalanddelta} and $\vectorbold{t}^{sf}_i$ is given by~\eqref{eq:fluid_wetting2} evaluated for particle $i$. In order to prescribe the value $\hat{\vectorbold{n}}^{lg}_i$ for the liquid-gas interface normal vector $\vectorbold{n}^{lg}_i$ in the triple point region and to have a smooth transition from the interface
region (with prescribed normal) to the interior domain (with solution-dependent normal), the following correction scheme is employed:
\begin{equation} \label{sph:fluid_wetting3}
\vectorbold{n}^{lg}_i = \frac{f_{\vectorbold{n},i} (\grad{c^{lg}}_i / || \grad{c^{lg}}_i ||) + (1-f_{\vectorbold{n},i}) \hat{\vectorbold{n}}^{lg}_i }{|| f_{\vectorbold{n},i} (\grad{c^{lg}}_i / || \grad{c^{lg}}_i ||) + (1-f_{\vectorbold{n},i}) \hat{\vectorbold{n}}^{lg}_i ||} \quad
\text{with} \quad
f_{\vectorbold{n},i} = \left\{\begin{array}{ll}
0 , & \text{if} \,\, d_{w,i} < 0\\
\frac{d_{w,i}}{d_{max}} , & \text{if} \,\, 0 \leq d_{w,i} \leq d_{max}\\
1 , & \text{if} \,\, d_{w,i} > d_{max}.
\end{array}\right.
\end{equation}
In the present work, $d_{max}=h$ has been chosen as the kernel smoothing length (which differs from~\cite{Breinlinger2013}, where the kernel support radius has been chosen) and the distance function $d_{w,i}$ is defined as the distance of fluid particle $i$ from the closest wall particle $j$ minus the initial particle spacing $h$ according to $d_{w,i} = \min ( (\vectorbold{r}_i-\vectorbold{r}_j) \cdot \vectorbold{n}^{sf}_i) - h$. From~\eqref{sph:fluid_wetting3} it becomes clear that for particles closer to the wall than $h$ the
interface normal is prescribed as $\hat{\vectorbold{n}}^{lg}_i$, while for particles with wall distance larger than $2h$ the conventional (solution-dependent) calculation of the interface normal via the color field
gradient is applied.
\subsubsection{Recoil pressure forces} \label{subsec:nummeth_sph_recoil}
The discrete version of the evaporation-induced recoil pressure forces occurring in~\eqref{eq:sph_momentum} is given by
\begin{equation} \label{eq:sph_recoil}
\vectorbold{F}_{v,i} = -V_i p_{v,i} \vectorbold{n}_i^{lg} \delta_i^{lg},
\end{equation}
where $p_{v,i}$ is the recoil pressure according to~\eqref{eq:fluid_recoil} evaluated for particle $i$.
\subsubsection{Viscous interface forces} \label{subsec:nummeth_sph_dissipation}
Monaghan and Gingold~\cite{Monaghan1983} proposed a stabilization term, denoted as artificial viscosity, as additional contribution to the momentum equation to reduce
spurious flow oscillations in the numerical solution. In this work we propose to employ this stabilization term selectively only at the solid-liquid and liquid-gas interface
to avoid oscillations originating from the phase transition solid-liquid and from the high liquid-gas interface forces typical for metal AM melt pool hydrodynamics. Moreover, it will
be shown that the resulting interface viscosity contributions can also be motivated from a physical point of view. In its general form the discrete version of these dissipative
interface forces is given by
\begin{equation} \label{eq:sph_dissipation}
\vectorbold{F}_{d,i} = - m_i {\zeta_i} \sum_{j} m_j \bar{h}_{ij} \bar{c}_{ij} \frac{{\vectorbold{u}_{ij}} \cdot \vectorbold{r}_{ij}}{\bar{\rho}_{ij} (r_{ij}^2 + \epsilon h_{ij}^2 )} \pdv{W}{r_{ij}},
\end{equation}
with the inter-particle averaged particle spacing $\bar{h}_{ij}=(h_i+h_j)/2$, speed of sound $\bar{c}_{ij}=(c_i+c_j)/2$ and density $\bar{\rho}_{ij}=(\rho_{i}+\rho_{j})/2$ as well as relative velocity {$\vectorbold{u}_{ij}=\vectorbold{u}_{i}-\vectorbold{u}_{j}$} and distance ${r}_{ij}=||\vectorbold{r}_{i}-\vectorbold{r}_{j}||$. The constant $\epsilon \ll 1$ is introduced to ensure a non-zero denominator. The viscosity factor is split into two contributions ${\zeta_i}={\zeta}^{lg}_i+{\zeta}^{sl}_i$. The first one, acting on (the liquid side of) the liquid-gas interface is given by:
\begin{equation} \label{eq:sph_dissipation_lg}
{\zeta}^{lg}_i = {\zeta}^{lg}_0 \delta^{lg}_i.
\end{equation}
Spurious interface flows are a well-known problem of continuum surface force (CSF) formulations~\cite{Breinlinger2013,Brackbill1996}. Due to the high magnitude of surface tension and recoil pressure
forces this undesirable effect is in particular critical for metal AM melt pool problems. As demonstrated in Section~\ref{sec:numex_droplet_oscillation}, the introduction of an additional viscous term acting exactly (and only) at the origin of these spurious interface flows, i.e. selectively at the liquid-gas interface, enables to effectively reduce this numerical artifact {without introducing additional dissipation of physically relevant flow characteristics in the interior fluid domain}. Since these spurious interface flows are known to decrease with increasing discretization resolution~\cite{Brackbill1996}, we recommend to scale the viscosity factor ${\zeta}^{lg}_0$ with the smoothing length $h_i$. With this strategy the maximal viscous forces acting on interface particles is discretization-independent (as the maximal magnitude of $\delta^{lg}_i$ scales with $1/h_i$), while the overall influence of the viscous interface forces on the global system behavior decreases (as the interface thickness decreases with $h_i$). Similar to slope limiting techniques~\cite{Berger2005}, the numerical scheme could be further refined by applying this interface stabilization term only at locations with extremely high (or fast changing) velocity gradients or in case of metal AM melt pool simulations only at locations with very high temperatures (and thus very high recoil pressure forces). The viscosity factor on the solid-liquid interface is defined as:
\begin{equation} \label{eq:sph_dissipation_sl}
{\zeta}^{sl}_i = {\zeta}^{sl}_0 f_{{\zeta}^{sl}}(T_i) \quad \text{with} \quad
f_{{\zeta}^{sl}}(T_i) = \left\{\begin{array}{ll}
0 , & \text{if} \,\, T_i > T_{max}\\
\frac{T_i-T_{max}}{T_m-T_{max}} , & \text{if} \,\, T_{max} \geq T_i \geq T_{m}\\
1 , & \text{if} \,\, T_i < T_{m},
\end{array}\right.
\end{equation}
where $T_{m}$ is the melt temperature and $T_{max}$ a temperature chosen slightly above the melt temperature. Thus, the viscous force~\eqref{eq:sph_dissipation_sl} only
acts on particles with temperatures close to the melting point and thus effectively damps potential instabilities arising from the jump of material parameters and state variables of a particle when undergoing the phase change solid-liquid. Due to the no-slip condition already applied to the fluid velocity field at this interface, the influence of this additional viscous force on the global system behavior is small as long as $T_{max}$ is chosen sufficiently close to $T_{m}$. As demonstrated in~\cite{Monaghan2006}, the action of the viscous force~\eqref{eq:sph_dissipation} can be associated with an equivalent physical viscous force with effective kinematic viscosity $\nu_i = 0.5 {{\zeta}_i} \bar{h}_{ij} \bar{c}_{ij} / (d+2)$, where $d=2,3$ is the spatial dimension. Thus, besides their stabilizing effect, the contributions~\eqref{eq:sph_dissipation_lg} and~\eqref{eq:sph_dissipation_sl} can also be interpreted from a physical point of view. For example,~\eqref{eq:sph_dissipation_lg} can be thought of as part of a non-conservative surface tension formulation with interface viscosity~\cite{Gounley2016}. In particular,~\eqref{eq:sph_dissipation_sl} can be interpreted as a physical model for the gradual phase transition of alloys between solidus and liquidus temperature {such that the viscosity decreases with increasing temperature during melting.}
\subsection{Energy equation} \label{subsec:nummeth_sph_energy}
The discrete version of the energy equation~\eqref{eq:fluid_energy} has the following general form:
\begin{equation} \label{eq:sph_energy}
\dv{T_i}{t} = \frac{1}{c_{p,i} \rho_i} [- (\div \vectorbold{q})_i + \tilde{s}^{lg}_{v,i} + \tilde{s}^{lg}_{l,i}]
\end{equation}
For the discretization of the conductive term, we follow a formulation proposed by Cleary and Monaghan~\cite{Cleary1999}, which is especially suited for problems
involving a jump of the thermal conductivity $k$ across an interface:
\begin{equation} \label{eq:sph_divq}
(\div \vectorbold{q})_i = \sum_{j} \frac{m_{j} 4k_ik_j (T_j-T_i)}{\rho_j (k_i+k_j) r_{ij}} \pdv{W}{r_{ij}}
\end{equation}
The discrete versions of the laser beam source term $\tilde{s}^{lg}_{l,i}$ and the evaporation-induced heat loss term $\tilde{s}^{lg}_{v,i}$ result directly from
evaluating~\eqref{eq:fluid_heatsource} and~\eqref{eq:fluid_evaporation} for the discrete particle $i$.
\subsection{Equation of state} \label{subsec:nummeth_sph_eos}
The discrete version of the equation of state results from evaluating~\eqref{eq:fluid_eos} for the discrete particle $i$.
\subsection{Boundary conditions} \label{subsec:nummeth_sph_bdrycond}
\paragraph{Rigid wall boundary conditions}
Following the approach of Adami et al.~\cite{Adami2012} rigid wall boundary conditions are modeled using fixed boundary particles with quantities extrapolated from the fluid field based on a local force balance. The same approach is used to model the mechanical interaction between fluid particles and the solid phase. For more details the interested reader is referred to the aforementioned literature.
\paragraph{Periodic boundary conditions}
Imposing a periodic boundary condition in a specific spatial direction allows for particle interaction evaluation across the lower and upper domain border. Moreover, particles leaving the domain on one side are re-injecting on the opposite side.
\section{Time integration scheme} \label{sec:nummeth_sph_timint}
The momentum equation~\eqref{eq:sph_momentum} is integrated in time applying an explicit velocity-Verlet time integration scheme in kick-drift-kick form, also denoted as leapfrog scheme, as proposed by Monaghan~\cite{Monaghan2005}. In the absence of dissipative effects the velocity-Verlet scheme is of second order accuracy and reversible in time~\cite{Monaghan2005}. In a first kick-step the particle accelerations $\vectorbold{a}_{i}^{n} = \qty(\dv*{\vectorbold{u}_{i}}{t})^{n}$ determined in the previous time step~$n$ are used to compute intermediate particle velocities at~$n+1/2$
\begin{equation}
\vectorbold{u}_{i}^{n+1/2} = \vectorbold{u}_{i}^{n} + \frac{\Delta{}t}{2} \, \vectorbold{a}_{i}^{n} \, ,
\end{equation}
where~$\Delta{}t$ is the time step size, before the particle positions at~$n+1$ are updated in a drift-step
\begin{equation}
\vectorbold{r}_{i}^{n+1} = \vectorbold{r}_{i}^{n} + \Delta{}t\vectorbold{u}_{i}^{n+1/2} \, .
\end{equation}
With the particle positions~$\vectorbold{r}_{i}^{n+1}$ the densities~$\rho_{i}^{n+1}$ are determined on the basis of~\eqref{eq:sph_densum}. Based on the temperatures~${T}_{i}^{n}$
as well as the updated particle positions~$\vectorbold{r}_{i}^{n+1}$ and densities~$\rho_{i}^{n+1}$ the temperature rate $\left(d T_i / d t\right)^{n+1}$ is calculated
on the basis of~\eqref{eq:sph_energy} and the temperature is updated according to:
\begin{equation}
{T}_{i}^{n+1} = {T}_{i}^{n} + \Delta{}t \left(\frac{d T_i}{d t}\right)^{n+1} \, .
\end{equation}
Using the updated particle temperatures~${T}_{i}^{n+1}$, positions~$\vectorbold{r}_{i}^{n+1}$ and densities~$\rho_{i}^{n+1}$ as well as the intermediate velocities $\vectorbold{u}_{i}^{n+1/2}$ the accelerations~$\vectorbold{a}_{i}^{n+1}$ are calculated from~\eqref{eq:sph_momentum}. In a final kick-step the particle velocities at time step~$n+1$ are determined via
\begin{equation}
\vectorbold{u}_{i}^{n+1} = \vectorbold{u}_{i}^{n+1/2} + \frac{\Delta{}t}{2} \, \vectorbold{a}_{i}^{n+1} \, .
\end{equation}
To maintain stability of the time integration scheme the time step size~$\Delta{}t$ is restricted by the Courant-Friedrichs-Lewy (CFL) condition, the viscous condition, the body force condition, the surface tension condition, and the conductivity-condition refer to~\cite{Adami2010, Morris1997, Adami2013, Cleary1998} for more details,
\begin{equation} \label{eq:sph_timestepcond}
\Delta{}t \leq \min\qty{ 0.25\frac{h}{c+\qty|\vectorbold{u}_{max}|}, \quad 0.125\frac{h^{2}}{\nu}, \quad 0.25\sqrt{\frac{h}{\qty|\vectorbold{b}_{max}|}} ,
\quad 0.25\sqrt{\frac{\rho h^3}{2 \pi \alpha}} , \quad 0.125 \frac{\rho c_p h^2}{k} } \, ,
\end{equation}
with kinematic viscosity $\nu=\eta/\rho$, maximum fluid velocity~$\vectorbold{u}_{max}$ and maximum body force~$\vectorbold{b}_{max}$.\\
\section{Comparison of different temperature gradient approximations} \label{sec:numex_tantempgrad}
In this section, the approximation quality of the different temperature gradient discretizations~\eqref{eq:sph_tempgrad1}-\eqref{eq:sph_tempgrad3} will be investigated. For this
purpose the temperature field in the liquid and gas phase of a liquid drop resting on a solid substrate and surrounded by a gas atmosphere will be considered at a representative time step (see Figure~\ref{fig:example1_tempandcolorfield_a}). A detailed description of the problem setup is given in Section~\ref{sec:numex_heateddrop}, where the
full thermo-hydrodynamic interaction within this problem is studied. In Figure~\ref{fig:example1_tempandcolorfield_b}, the total color field $\hat{c}_i:=\sum_{j} V_{j} W_{ij}$, which considers contributions from all possible types (i.e. gas, liquid and solid phase) of neighbor particles $j$, is displayed for the considered droplet example. The fact that the color field is close to one throughout the entire domain suggests that all particles have full support and, thus, \textit{two-sided} gradient approximations such as~\eqref{eq:sph_tempgrad1}-\eqref{eq:sph_tempgrad3}
might be applicable in general.
\begin{figure}[htbp]
\centering
\subfigure [Temperature field ranging from $1700$ (blue) to $2700$ (red).]
{
\includegraphics[width=0.47\textwidth]{figures/example1_tempfield.png}%
\label{fig:example1_tempandcolorfield_a}
}
\hspace{0.01\textwidth}
\subfigure [Total color field ranging from $0.99$ (blue) to $1.01$ (red).]
{
\includegraphics[width=0.47\textwidth]{figures/example1_colorfield.png}%
\label{fig:example1_tempandcolorfield_b}
}
\caption{Laser heating of liquid drop (large particles) on solid substrate (not displayed) surrounded by gas (small particles).}
\label{fig:example1_tempandcolorfield}
\end{figure}
It is typically argued that \textit{one-sided} SPH gradient approximations such as the corrective smoothed particle method (CSPM) by Chen et al.~\cite{Chen2001} or the Corrected SPH (CSPH) scheme by Bonet and Lok~\cite{Bonet1999}, which allow for exact representation of first-order gradients even in boundary or interface regions with incomplete particle support, are required to
accurately capture a kink in the temperature field resulting from the jump in the thermal conductivity $k$ at fluid-gas interfaces~\cite{Tong2014}. Therefore, the first-order consistent schemes \textit{CSPM} as well as \textit{CSPH} will be considered as reference solutions in this section and compared to the temperature gradient discretizations according to~\eqref{eq:sph_tempgrad1} (variant \textit{Standard}),~\eqref{eq:sph_tempgrad2} (variant \textit{Symmetric}) and~\eqref{eq:sph_tempgrad3} (variant \textit{Asymmetric}). In Figure~\ref{fig:example1_tempgrad}, the temperature gradients
associated with the temperature field~\ref{fig:example1_tempandcolorfield_a} are displayed for the variants \textit{CSPM}, \textit{Standard}, \textit{Symmetric} and \textit{Asymmetric} and compared
to the variant \textit{CSPH} (shaded). As expected, there is no visual difference of the two \textit{one-sided} gradient approximations (variants \textit{CSPM} and \textit{CSPH} in Figure~\ref{fig:example1_tempgrad_a}). Indeed, the results of these two formulations are identical up to machine precision. However, the non-smooth temperature field across the interface leads to deviations for the other three variants in this region (Figures~\ref{fig:example1_tempgrad_b}-~\ref{fig:example1_tempgrad_d}). While these visible deviations vanish in the interior of the drop for the variant \textit{Asymmetric}, the variants \textit{Standard} and \textit{Symmetric} show large deviations also in this domain, which seems to contradict first intuition. However, this observation can be explained by the fact that the asymmetric formulation exactly filters out constant temperature contributions (exactly vanishing gradient for constant fields), while large constant temperature contributions lead to considerable discretization errors for the variants \textit{Standard} and \textit{Symmetric} due to the lack of zero-order consistency, i.e. $\int_{\Omega} \nabla W\qty(\qty| \vectorbold{r} - \vectorbold{r}' |, h) \dd{\vectorbold{r}'} = 0$ but $\sum_{j} V_{j} \pdv{W}{r_{ij}} \vectorbold{e}_{ij}\neq 0$. The avoidance of large constant contributions is also the reason why the reference pressure is typically set to zero in SPH formulations where the pressure gradient is approximated by the momentum-conserving variant \textit{Symmetric}~\cite{Morris1997}.
\begin{figure}[htbp]
\centering
\subfigure [Variants \textit{CSPM}~\cite{Chen2001} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgrad_Bonet_vs_CSPM.png}%
\label{fig:example1_tempgrad_a}
}
\hspace{0.01\textwidth}
\subfigure [Variants \textit{Standard}~\eqref{eq:sph_tempgrad1} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgrad_Bonet_vs_varg.png}%
\label{fig:example1_tempgrad_b}
}
\subfigure [Variants \textit{Symmetric}~\eqref{eq:sph_tempgrad2} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgrad_Bonet_vs_varb.png}%
\label{fig:example1_tempgrad_c}
}
\hspace{0.01\textwidth}
\subfigure [Variants \textit{Asymmetric}~\eqref{eq:sph_tempgrad3} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgrad_Bonet_vs_vard.png}%
\label{fig:example1_tempgrad_d}
}
\caption{Laser heating of liquid drop: different approximations for total temperature gradient. Displayed is half of the droplet. The total temperature gradient is visualized by bright (investigated solution) and shaded (reference solution) red arrows.}
\label{fig:example1_tempgrad}
\end{figure}
When considering the tangential projection of the temperature gradient $\nabla_T {T}_i$ as defined in~\eqref{eq:sph_surfacetension2} and displayed in Figure~\ref{fig:example1_fig:example1_tempgradtan}, the observations made above for the variants \textit{CSPH}, \textit{CSPM}, \textit{Standard}, and \textit{Symmetric} can be confirmed. Specifically the latter two variants are not suitable to represent the
tangential temperature gradient. In contrast, the variant \textit{Asymmetric} represents the tangential temperature gradient very well without visual difference to the variants \textit{CSPH} and \textit{CSPM} - except for a small deviation at the triple point, which can be expected due to the non-smooth interface line in this region. Therefore, the variant \textit{Asymmetric} according to~\eqref{eq:sph_tempgrad3} will be employed - and favored over the computationally more involved \textit{CSPH} and \textit{CSPM} schemes - in all the remaining numerical examples in this work. In the remainder of this section, a brief analytical explanation will be given on why the \textit{two-sided} gradient approximation~\eqref{eq:sph_tempgrad3} is capable of correctly representing the tangential gradient projection of a temperature field with a kink at the interface.
\begin{figure}[htbp]
\centering
\subfigure [Variants \textit{CSPM}~\cite{Chen2001} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgradtan_Bonet_vs_CSPM.png}%
\label{fig:example1_fig:example1_tempgradtan_a}
}
\hspace{0.01\textwidth}
\subfigure [Variants \textit{Standard}~\eqref{eq:sph_tempgrad1} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgradtan_Bonet_vs_varg.png}%
\label{fig:example1_fig:example1_tempgradtan_b}
}
\subfigure [Variants \textit{Symmetric}~\eqref{eq:sph_tempgrad2} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgradtan_Bonet_vs_varb.png}%
\label{fig:example1_fig:example1_tempgradtan_c}
}
\hspace{0.01\textwidth}
\subfigure [Variants \textit{Asymmetric}~\eqref{eq:sph_tempgrad3} vs. \textit{CSPH}~\cite{Bonet1999} (shaded).]
{
\includegraphics[width=0.45\textwidth]{figures/example1_tempgradtan_Bonet_vs_vard.png}%
\label{fig:example1_fig:example1_tempgradtan_d}
}
\caption{Laser heating of liquid drop: different approximations for tangential temperature gradient. Displayed is half of the droplet. The tangential gradient is visualized by bright (investigated solution) and shaded (reference solution) red arrows.}
\label{fig:example1_fig:example1_tempgradtan}
\end{figure}
Let's assume a particle $i$ close to the liquid-gas interface with closest / orthogonal projection point $c$ onto this interface. Furthermore, a coordinate system with axes
$\vectorbold{t}_1$ and $\vectorbold{t}_2$ tangential to the interface and axis $\vectorbold{n}$ normal to the interface is defined in point $c$. The position of a material particle expressed in this system
shall be denoted as $\tilde{\vectorbold{x}}$, and a quantity evaluated at position $c$ shall be marked by a subscript $(...)_c$. Moreover, subscripts $(...)_n$ and $(...)_t$ represent projections of vectors in interface normal and tangential direction. The first-order Taylor series expansion of the temperature field in the neighborhood of point $c$ is given as:
\begin{equation} \label{eq:examp1_tempfield}
T(\tilde{\vectorbold{x}}) \approx T_c + \grad{T}_c \cdot \tilde{\vectorbold{x}} = T_c + \boldsymbol{\nabla}_t T_c \cdot \tilde{\vectorbold{x}}_t + \boldsymbol{\nabla}_n T_c \cdot \tilde{\vectorbold{x}}_n
\quad \text{with} \quad
\boldsymbol{\nabla}_n T_c = \left\{\begin{array}{ll}
\boldsymbol{\nabla}_n T_c^+ , & \text{if} \,\, \tilde{\vectorbold{x}}_n \cdot \vectorbold{n} > 0\\
\boldsymbol{\nabla}_n T_c^-, & \text{else}.
\end{array}\right.
\end{equation}
Note that~\eqref{eq:examp1_tempfield} accounts for a jump of the temperature gradient in normal direction as typical for liquid-gas interfaces. According to~\eqref{sph_general_gradient} the projection of the temperature gradient in the tangential directions $\vectorbold{t}_k$ reads:
\begin{equation}
\label{eq:examp1_gradient1}
\vectorbold{t}_k \cdot \grad{T}\qty(\vectorbold{r}_i) \approx \vectorbold{t}_k \cdot \int_{\Omega_i} [T_c + \boldsymbol{\nabla}_t T_c \cdot \tilde{\vectorbold{x}}_t + \boldsymbol{\nabla}_n T_c \cdot \tilde{\vectorbold{x}}_n] \, \grad{W} \dd{\tilde{\vectorbold{x}}} \quad \text{for} \quad k=1,2.
\end{equation}
In the following, only the third of the three summands will be considered. For simplicity, the integral is reformulated in spherical coordinates $(r,\alpha, \beta)$ using $\tilde{\vectorbold{x}}= r \vectorbold{e}$ with $\vectorbold{e}=\sin \beta \cos \alpha \vectorbold{t}_1 + \sin \beta \sin \alpha \vectorbold{t}_2 + \cos \beta \vectorbold{n}$:
\begin{equation}
\label{eq:examp1_gradient2}
\begin{split}
& \vectorbold{t}_k \cdot \int_{\Omega_i} \boldsymbol{\nabla}_n T_c \cdot \tilde{\vectorbold{x}}_n \, \grad{W} \dd{\tilde{\vectorbold{x}}} \\
= & \boldsymbol{\nabla}_n T_c \cdot \int_{r=0}^{r_c} \int_{\beta=0}^{\pi} \int_{\alpha=0}^{2\pi} \underbrace{r \cos \beta \vectorbold{n}}_{\tilde{\vectorbold{x}}_n} \underbrace{\pdv*{W}{r} \vectorbold{e}}_{\grad{W}} \cdot \vectorbold{t}_k \underbrace{r^2 \sin \beta \dd{\alpha} \dd{\beta} \dd{r}}_{\dd{\tilde{\vectorbold{x}}}} \\
= & \boldsymbol{\nabla}_n T_c \! \cdot \! \vectorbold{n} \int_{r=0}^{r_c} r^3 \pdv*{W}{r} \int_{\beta=0}^{\pi} \cos \beta \sin^2 \beta
\underbrace{\int_{\alpha=0}^{2\pi} [\cos \alpha \vectorbold{t}_1 \! \cdot \! \vectorbold{t}_k + \sin \alpha \vectorbold{t}_2 \! \cdot \! \vectorbold{t}_k] \dd{\alpha}}_{=0} \dd{\beta} \dd{r} = 0.
\end{split}
\end{equation}
From the second to the third line in~\eqref{eq:examp1_gradient2} use have been made of the spherical symmetry of the kernel $W$, i.e. $\pdv*{W}{r}$ is independent of $\alpha$ and $\beta$. {It is emphasized that only for spherically symmetric kernels the tangential projection of the SPH discretization of a gradient contains only information of the tangential component of the space-continuous gradient, i.e. it is independent of the evolution of the underlying field in normal direction and the third summand in~\eqref{eq:examp1_gradient1} vanishes.} Thus, it can be concluded that standard (two-sided) gradient discretization approaches can be applied since the gradient jump in normal direction does not enter the SPH formulation for the tangential gradient discretization. While this derivation has been made for the \textit{standard} temperature gradient approximation~\eqref{eq:sph_tempgrad1}, the results are equally valid for the \textit{symmetric}~\eqref{eq:sph_tempgrad2} and \textit{asymmetric}~\eqref{eq:sph_tempgrad3} gradient approximation since the main difference is a constant temperature value $T_i$ added to or subtracted from the standard gradient approximation.
\section{Numerical examples} \label{sec:numex}
\subsection{Liquid droplet in surrounding fluid} \label{sec:numex_migratingbubble}
To verify the proposed formulation for temperature-dependent surface tension, cf. Section~\ref{subsec:nummeth_sph_surface tension}, the migration of a liquid droplet as proposed by Ma and Bothe~\cite{Ma2011} is considered. While a finite volume scheme is employed in~\cite{Ma2011}, the same problem has been studied by Tong and Browne~\cite{Tong2014} using an incompressible SPH formulation. The problem consists of a circular droplet (radius $a = 1.44$) of fluid~1 (density $\rho^1=0.25$, dynamic viscosity $\mu^1 = 12.0$, thermal conductivity $k^1=1.2 \times 10^{3}$, heat capacity $c_p^{1} = 50.0$) that is initially resting at the center of a quadratic domain (side length $4a$) filled with fluid~2 ($\rho^2=2\rho^1$, $\mu^2 = 2 \mu^1$, $k^2=2 k^1$, $c_p^{2} = 2 c^{1}$). The temperature-dependent surface tension is defined via $\alpha_0=1.0 \cdot 10^{4}$, $\alpha'_0=2.0 \cdot 10^{3}$ and $T_{\alpha_0}=290$. All properties are given in the units $mm$, $mg$, $s$, and $K$. On the left and right side of the quadratic domain periodic boundary conditions are applied. At the top and bottom of the quadratic domain no-slip boundary conditions for the fluid field as well as prescribed temperatures $\hat{T}^{1}$ (bottom) and $\hat{T}^{2}$ (top) are applied via three layers of boundary particles according to~\cite{Adami2012}. The 2D domain is discretized by a total of 4096 particles (812 particles for fluid 1, 3284 particle for fluid 2) resulting in an initial particle spacing of $\Delta x = 0.09$. Initially, the particles are at rest, i.e. $\boldsymbol{u}_0^1=\boldsymbol{u}_0^2=\boldsymbol{0}$. Like in the original works, the surface tension acts with its full magnitude during the entire simulation time, i.e. no initial ramp function is used. Moreover, the initial temperature profile $T_0$ is chosen
as linear interpolation between $\hat{T}^{1}$ and $\hat{T}^{2}$ leading to an initial gradient of $||\grad{T_0}||=(\hat{T}_{2}-\hat{T}_{1})/(4a)$.
\subsubsection{Static equilibrium of droplet}
As a first test case the pressure jump across the interface of the bubble in a static equilibrium configuration shall be investigated. Thereto the temperatures at the bottom and top wall are prescribed to $\hat{T}_{1}=\hat{T}_{2}=290$. The weakly compressible approach is realized with bulk moduli $K^{1} = 78.125 \cdot 10^{3}$ and $K_{2} = 156.25 \cdot 10^{3}$ (artificial speed of sound $c^1=c^2 \approx 562.5$). The time step size has been chosen to $\Delta t = 4.0 \cdot 10^{-5}$. The analytical solution for this problem setup is a constant temperature field at $T=290$ with the circular droplet of radius $a$ resting at the center of the quadratic domain. While the Marangoni forces vanish for this test case, due to the constant surface tension contribution of $\alpha=\alpha_0=1.0 \cdot 10^{4}$ a pressure jump of $\Delta p = \frac{\alpha_0}{a} \approx 6944$ is expected at the interface of the circular droplet.
\begin{figure}[htbp]
\centering
\subfigure [Static equilibrium: pressure jump $\Delta p := p - p_{\infty}$ over distance $r$ from droplet center for sections horizontal (black) and diagonal (blue) through the droplet.]
{
\includegraphics[width=0.45\textwidth]{figures/static_droplet1.png}%
\label{fig:label_subfig_1}
}
\hspace{0.01\textwidth}
\subfigure [Thermo-capillary migration: dimensionless centroid velocity $U/U_{r}$ over dimensionless time $t/t_{r}$ for present formulation (solid) and reference solution ($+$ marks,~\cite{Ma2011}).]
{
\includegraphics[width=0.45\textwidth]{figures/static_droplet2.png}%
\label{fig:label_subfig_2}
}
\caption{Different configurations of a droplet within surrounding fluid: static equilibrium and thermo-capillary migration.}
\label{fig:example_hydrostatic_pressure}
\end{figure}
In Figure~\ref{fig:label_subfig_1} the pressure jump $\Delta p := p - p_{\infty}$, with $p_{\infty}$ representing the pressure in fluid~2 at sufficient distance from the droplet interface, is displayed over the distance $r$ from the droplet center for a horizontal (black) as well as a diagonal (45 degrees) section (blue) through the droplet center.
Besides slight oscillations at the inner and outer fringe of the interface region, the numerical results show very good agreement with the analytical prediction of the maximal pressure jump at the droplet center given by $\Delta p \approx 6944$.
\subsubsection{Thermo-capillary migration of droplet}
As a second test case the thermo-capillary migration of the droplet will be considered. Thereto the temperatures at the bottom and top wall are prescribed to $\hat{T}_{1}=290$ and
$\hat{T}_{2}=T_{1} + 4 a \qty|\nabla T| = 291.152$ such that the stationary initial temperature field is characterized by a constant gradient $||\grad{T_0}|| = 0.2$. For the
simulation of this problem the bulk moduli $K^{1} = 5.0 \cdot 10^{6}$ and $K^{2} = 1.0 \cdot 10^{7}$ (artificial speed of sound $c^1=c^2 \approx 4472.1$) as well as a time step size
of $\Delta t = 4.0 \cdot 10^{-5}$ have been chosen.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.7\textwidth]{figures/screenshot_t0_08_3}
\caption{Thermo-capillary migration of a droplet (in vertical direction only part of the domain is shown): velocity profile within droplet indicated by arrows and temperature profile of both fluid phases colored in the temperature range from $T_{1} = 290.0$ (blue) to $T_{2} = 291.152$ (red).}
\label{fig:example_migrating_droplet}
\end{figure}
In this test case, temperature-dependent surface tension in combination with the prescribed temperature gradient results in shear stresses (second term in~\eqref{eq:fluid_surfacetension}) and consequently a shear flow (Marangoni convection) at the droplet interface (see Figure~\ref{fig:example_migrating_droplet}). On the one hand, this shear flow induces an upwards migration of the droplet. On the other hand, the shear flow induced in the surrounding fluid~2 redistributes the initially linear temperature profile. In terms of dimensionless numbers the considered choice of parameters leads to a Reynolds number of $Re = \frac{\rho a U_{r}}{\mu} = 0.72$, a Marangoni number of $Ma = \frac{\rho c_{p} a U_{r}}{\lambda} = 0.72$ and a capillary number of $Ca = \frac{\mu U_{r}}{\sigma_{0}} = 0.0576$, where the characteristic velocity $U_{r} = \sigma_{t} \qty|\nabla T| \frac{a}{\mu} = 24$ has been employed. In Figure~\ref{fig:label_subfig_2} the dimensionless velocity $U/U_{r}$, with $U$ calculated via numerical differentiation of the droplet centroid position with respect to time, is plotted over the dimensionless time $t/t_{r}$ with $t_{r}= \frac{a}{U_{r}} = 0.06$. The resulting velocity evolution shows good agreement with the reference solution~\cite{Ma2011}. The slight fluctuations in the steady state regime can be traced back to the employed weakly compressible formulation, while the reference solution in~\cite{Ma2011} has been calculated on the basis of an incompressible formulation.
\subsection{Oscillation of liquid dropled surrounded by gas atmosphere} \label{sec:numex_droplet_oscillation}
This example aims to investigate the influences of the viscous interface force introduced in Section~\ref{subsec:nummeth_sph_dissipation} as well as of approximated material properties of the gas phase on the solution accuracy of surface tension-dominated problems with parameter values as typical for metal AM. For this purpose the well-known example of surface tension-driven oscillations of a liquid droplet in gas atmosphere is considered~{\cite{Nugent2000, Morris2000, Hu2006, Adami2010}}. The droplet is initially resting at the center of a quadratic gas domain with side length $400 \mu m$ and has an elliptic initial shape (larger semiaxis $a=3/2R$, smaller semiaxis $b=2/3R$, where $R=100 \mu m$ is the droplet radius in the static equilibrium configuration; see Figure~\ref{fig:example2b_tempandcolorfield_a}). The reference pressure of the weakly compressible model is set to $p_0=1.0 \cdot 10^7 N m^{-2}$ and the background pressure of the transport velocity formulation to $p_b=5p_0$ for both phases. Moreover, a time step size of $\Delta t = 10^{-6} ms$ and an initial particle spacing of $\Delta x = 5/3 \mu m$ is applied. Again, the temperature field is kept constant such that Marangoni forces vanish. The material properties of the liquid phase inside the droplet are identical to the ones used for the subsequent melt pool simulations and are given in Table~\ref{tab:material_params}. Moreover, different material properties of the gas phase will be investigated in the following. As reference parameter set the gas phase is modeled such that the density and dynamic viscosity are by a factor of $1000$ smaller as compared to the values of the liquid phase. Driven by the surface tension forces the droplet executes oscillations with period T. Figure~\ref{fig:example2b_tempandcolorfield} illustrates the droplet shapes resulting from the reference parameter set at times $t=0$, $t=T/2$ and $t=T$.
\begin{figure}[htbp]
\centering
\subfigure [Configuration at t=0.]
{
\includegraphics[width=0.3\textwidth]{figures/example2b_test_bubble_elliptic_2b_step4.png}%
\label{fig:example2b_tempandcolorfield_a}
}
\hspace{0.01\textwidth}
\subfigure [Configuration at t=T/2.]
{
\includegraphics[width=0.3\textwidth]{figures/example2b_test_bubble_elliptic_2b_step788.png}%
\label{fig:example2b_tempandcolorfield_b}
}
\hspace{0.01\textwidth}
\subfigure [Configuration at t=T.]
{
\includegraphics[width=0.3\textwidth]{figures/example2b_test_bubble_elliptic_2b_step1433.png}%
\label{fig:example2b_tempandcolorfield_c}
}
\caption{Oscillation of liquid dropled surrounded by gas atmosphere for reference parameter set at different times.}
\label{fig:example2b_tempandcolorfield}
\end{figure}
For the case of small amplitude oscillations there is an analytical solution for the resulting oscillation period~{\cite{rayleigh1879capillary}} given by $T_a=2\pi \sqrt{R^3 \rho_0^l / (6 \alpha)}$, where $\rho_0^l$ the initial density of the liquid phase. For the given set of parameters the analytical solution for the oscillation period takes on a value of $T_a=0165ms$. For the numerical solution based on the reference parameter set as illustrated in Figure~\ref{fig:example2b_tempandcolorfield} an oscillation period of
$T_{ref}=0.179ms$ can be determined. The deviation can be explained by the fact that in the present problem setup, which is considered more representative for melt pool hydrodynamics, \textit{large} amplitude oscillations occur and the damping of the oscillation amplitude resulting from the viscosity of the liquid metal is not negligible in this scenario. Specifically, after one oscillation (i.e. at $t=T$) the initial droplet length $l(t=0)=300 \mu m$ has already decreased to $l(t=T)=280 \mu m$. To increase the computational efficiency (in terms of larger critical time step sizes) and robustness of the numerical formulation the density and dynamic viscosity of the gas phase shall be increased. This procedure is considered reasonable since the focus of this work lies on the thermo-hydrodynamics in the melt phase while an exact representation of detailed flow patterns in the gas phase is not of primary interest. When decreasing the density ratio between liquid and gas phase to $100$ and the dynamic viscosity ratio to $10$ the droplet length after one oscillation is $l(t=T)=274 \mu m$, representing a relative error of $\approx 2 \%$ with respect to the reference parameter set (density and dynamic viscosity ratio of $1000$). Since an error of $\approx 2 \%$ seems reasonable for the intended
melt pool simulations, this parameter set ($\rho_0^l/\rho_0^g=100$, $\eta^l/\eta^g=10$) is taken as default value for the subsequent melt pool simulations. Finally, the influence of the viscous interface forces (Section~\ref{subsec:nummeth_sph_dissipation}), which should stabilize the numerical scheme without significantly
changing the physical system behavior, will be investigated. For the discretization resolutions and recoil pressure magnitudes considered in the subsequent numerical examples values of the viscosity constant in the range of ${\zeta}_0^{lg} \sim 10^{-4}$ turned out to effectively stabilize the scheme. Considering the default parameter set ($\rho_0^l/\rho_0^g=100$, $\eta^l/\eta^g=10$) and adding additional viscous interface forces with ${\zeta}_0^{lg}=2.5 \cdot 10^{-4}$ and ${\zeta}_0^{lg}=1.0 \cdot 10^{-4}$ results in relative errors of $\approx 6 \%$ and $\approx 2 \%$ in the droplet length $l(T)$ with respect to the solution resulting from the default parameter set ($\rho_0^l/\rho_0^g=100$, $\eta^l/\eta^g=10$) with ${\zeta}_0^{lg}=0$. Again, this accuracy is considered reasonable for the intended melt pool simulations.
\subsection{Representative test cases for metal additive manufacturing applications} \label{sec:numex_am}
The remaining test cases are designed to verify the proposed formulation with respect to specific effects, conditions and material properties as typical for metal additive manufacturing processes.
As representative material properties values close to the parameters for stainless steel at the melting point ($T_m=1700K$) according to~\cite{Khairallah2014} and~\cite{Khairallah2016} are considered. Table~\ref{tab:material_params} represents the corresponding material parameters considered for the liquid metal phase, i.e. for the melt phase. While the solid phase is modeled as rigid and immobile in this work, the thermal problem in the solid phase is solved in the same manner as for the liquid metal phase employing the same thermal material parameters.
\begin{table}[htbp]
\centering
\begin{tabular}{|p{1.5cm}|p{7.0cm}|p{3.0cm}|p{3.0cm}|} \hline
Symbol & Property & Value & Units \\ \hline
$\rho_0$ & Initial / reference density& $7430$ & $kg \, m^{\!-3}$ \\ \hline
$\eta$ & Dynamic viscosity& $6 \cdot 10^{-3}$ & $kg \, m^{\!-1} \, s^{\!-1}$ \\ \hline
$\alpha_0$ & Surface tension at reference temperature & $1.8$ & $N \, m^{\!-1}$ \\ \hline
$T_{\alpha_0}$ & Reference temperature for surface tension& $1700$ & $K$ \\ \hline
$\alpha'_0$ & Surface tension gradient coefficient& $1.0 \cdot 10^{-3}$ & $N \, m^{\!-1} \, K^{\!-1}$ \\ \hline
$\theta_0$ & Wetting angle & $60$ & degrees \\ \hline
$T_0$ & Initial temperature & $500$ & $K$ \\ \hline
$\hat{T}$ & Temperature at Dirichlet boundaries & $500$ & $K$ \\ \hline
$T_m$ & Melting temperature & $1700$ & $K$ \\ \hline
$T_v$ & Evaporation / boiling temperature & $3000$ & $K$ \\ \hline
$c_p$ & Heat capacity & $965$ & $J \, kg^{-1} \, K^{-1}$ \\ \hline
$k$ & Thermal conductivity & $35.95$ & $ W\, m^{-1} \, K^{-1}$ \\ \hline
$\chi_l$ & Laser absorptivity & $0.5$ & $ - $ \\ \hline
$C_P$ & Pressure constant of recoil pressure model & $20$ & $ N \, m^{-2}$ \\ \hline
$C_T$ & Temperature constant of recoil pressure model & $1 \cdot 10^5$ & $K$ \\ \hline
$h_v$ & Latent heat of evaporation & $6 \cdot 10^6$ & $J \, kg^{-1}$ \\ \hline
$T_{h,0}$ & Reference temperature for specific enthalpy & $663.731$ & $K$ \\ \hline
$C_M$ & Constant for vapor mass flow & $1.0 \cdot 10^{-3}$ & $K \, s^2 \, m^{-2}$ \\ \hline
\end{tabular}
\caption{Representative material parameters for stainless steel taken for the liquid metal phase.}
\label{tab:material_params}
\end{table}
Note that the enthalpy reference temperature $T_{h,0}$ has been chosen such that the specific enthalpy at the melting point yields $h(T=T_m)=1 \cdot 10^6 J \, kg^{-1}$ according to~\eqref{eq:fluid_evaporation}. In order to increase the effect of recoil pressure and to make the numerical examples more challenging with respect to the robustness of the proposed formulation, the constants $C_P$ and $C_T$ of the recoil pressure model as well as the laser powers employed in Sections~\ref{sec:numex_2Dmelt_rec} and~\ref{sec:numex_3Dmelt} have been adapted compared to typical values for SLM processes. {Thus, the following examples are intended to demonstrate the robustness of the proposed SPH formulation and its general suitability for challenging application scenarios in metal AM, but not to represent the system behavior of a specific AM experiment with given process and material parameters. In our continued research work additional modeling aspects such as temperature-dependent material parameters or laser beam absorption via ray tracing, capturing e.g. multiple reflection and shading effects, will be incorporated. While these aspects are crucial to resemble detailed characteristics of experimental measurements, they are expected to not significantly influence the general accuracy and robustness of the numerical formulation, which is in the focus of the present study.} For reasons of numerical efficiency, the gas phase is approximated by an reduced density and dynamic viscosity ratio of $\rho_0^l/\rho_0^g=100$ and $\eta^l/\eta^g=10$ as analyzed in the last section. As thermal properties $c_p=10 J \, kg^{-1} \, K^{-1}$, $k=0.026 W\, m^{-1} \, K^{-1}$ and $\chi_l=0$ are chosen for the gas phase. Unless stated otherwise, the following standard discretization and regularization strategy is applied: The reference pressure of the weakly compressible model is set to $p_0=1.0 \cdot 10^7 N m^{-2}$ and the background pressure of the transport velocity formulation to $p_b=5p_0$ for both phases. Moreover, a standard time step size of $\Delta t = 10^{-6} ms$ and an initial particle spacing of $\Delta x = 5/3 \mu m$ is typically applied. Finally, the considered computational domains are surrounded by three layers of boundary particles with prescribed temperature $\hat{T}$ as well as free-slip boundary conditions for the fluid field.
\subsubsection{Laser heating of melt drop on solid substrate} \label{sec:numex_heateddrop}
\begin{figure}[htbp]
\centering
\subfigure [Static equilibrium configuration: surface tension gradient coefficient $\alpha'_0/ \rho_0=0.0$. Analytical solution in grey.]
{
\includegraphics[width=0.95\textwidth]{figures/example3_static_t1_80_new.png}%
\label{fig:example3_1}
}
\subfigure [Low Marangoni convection strength: surface tension gradient coefficient $\alpha'_0/ \rho_0=2.0 \cdot 10^{-8}$.]
{
\includegraphics[width=0.95\textwidth]{figures/example3_variant_a_t4_62.png}%
\label{fig:example3_2}
}
\subfigure [Medium Marangoni convection strength: surface tension gradient coefficient $\alpha'_0/ \rho_0=5.0 \cdot 10^{-8}$.]
{
\includegraphics[width=0.95\textwidth]{figures/example3_variant_b_t6_00.png}%
\label{fig:example3_3}
}
\subfigure [High Marangoni convection strength: surface tension gradient coefficient $\alpha'_0/ \rho_0=10.0 \cdot 10^{-8}$.]
{
\includegraphics[width=0.95\textwidth]{figures/example3_variant_c_t6_00.png}%
\label{fig:example3_4}
}
\caption{Laser heating of melt drop: Melt drop shape, temperature and velocity field for static contact angle $\theta_0=60^o$ and different Marangoni convection strengths ($\alpha'_0/ \rho_0$ in [$(N \, m^{-1} \, K^{-1})/(kg \, m^{-3})$]). Temperature range (left) from $1700K$ (blue) to $2800K$ (red). Velocity range (right) from $0.0 m/s$ (blue) to $4.2 m/s$ (red). Gas and boundary phase not displayed.}
\label{fig:example3}
\end{figure}
In this example, the effects already considered in Section~\ref{sec:numex_migratingbubble} shall be extended by wetting forces and a laser beam heat source. Thereto, a rectangular domain with dimensions $x \in [-100;100], \, y \in [-40;40]$ (all length dimensions given in $\mu m$) is considered. The sub-domain $x \in [-50;50], \, y \in [-40;-10]$ is initially covered with liquid melt, representing a droplet of initially rectangular shape, the remaining domain with gas. The overall domain is surrounded by three layers of boundary particles. In this example, the initial as well as the wall temperature (Dirichlet boundary conditions) have been chosen to $T_0=\hat{T}=1700K$. Note that this choice for the temperature initial and boundary conditions ensures that the melt drop remains liquid at all times and no phase change problem has to be considered in this example. The surface tension coefficient $\alpha_0$ is linearly increased from zero to $1.8 N \, m^{\!-1}$ in the time interval $t \in [0;0.05]$ (all time units in $ms$) and then kept constant throughout the rest of the simulation. Moreover, a static wetting angle of $\theta_0=60^o$ is considered. In the time interval $t \in [0;0.3]$, the laser heat source is set to zero such that the system can reach a static equilibrium configuration under the action of surface tension and wetting forces (see Figure~\ref{fig:example3_1}). In order to enable a fast transition into the equilibrium state, the dynamic viscosity of the melt phase is increased by a factor of ten compared to its standard value (only) in this first time interval $t \in [0;0.3]$. In the remaining simulation time $t \in [0.3;0.6]$, the surface tension, the dynamic viscosity as well as all the other material parameters are kept constant at their standard values according to Table~\ref{tab:material_params}. Additionally, the laser beam heat source is (instantaneously) switched on at $t=0.3$. The laser beam points into negative $y-$direction and its center is located at $x=0$. The irradiance is set to $s^{lg}_{l0} \approx 5.94 \cdot 10^{-3} W \, \mu m^{-2}$ ($s^{lg}_{l0}/\rho_0 = 8 \cdot 10^5 W \, m \, kg^{-1}$) and the characteristic radius to $r_w=40 \mu m$. The reference pressure of the weakly compressible model is set to $p_0= 1.0 \cdot 10^7 N m^{-2}$ for the melt and to $p_0=1.0 \cdot 10^6 N m^{-2}$ for the gas phase. The background pressure of the transport velocity formulation is set to $p_b=p_0$ for both phases. Moreover, a time step size of $\Delta t = 5.0 \cdot 10^{-7} ms$ and an initial particle spacing of $\Delta x = 5/12 \mu m$ is applied.\\
The analytical solution for the shape of the melt domain at the equilibrium state $t=0.3 ms$ is given by a circular segment as illustrated in Figure~\ref{fig:example3_1} in grey color. The numerical solution to this problem is characterized by a highly regular particle distribution and interface normal vector field, which represent the analytical solution with high accuracy (see also Figure~\ref{fig:example3_1}). Figures~\ref{fig:example3_2}-\ref{fig:example3_4} represent the steady state solutions resulting from the laser heating at $t \approx 0.6 ms$ for different choices of the surface tension gradient coefficient $\alpha'_0$ according to~\eqref{eq:fluid_surfacetension_tempdepend}. For simplicity, the ratio $\alpha'_0 / \rho_0$ has been prescribed in the simulations, where the four different variants $\alpha'_0 / \rho_0 = \{0.0, 2.0 \cdot 10^{-8}, 5.0 \cdot 10^{-8}, 10.0 \cdot 10^{-8}\}$ in $[(N \, m^{-1} \, K^{-1})/(kg \, m^{-3})]$ (corresponding to $\alpha'_0 = \{0.0, 1.486 \cdot 10^{-4}, 3.715 \cdot 10^{-4}, 7.43 \cdot 10^{-4}\}$ in $[N \, m^{-1} \, K^{-1}]$) have been analyzed. It can be observed that an increasing surface tension gradient coefficient (from Figure~\ref{fig:example3_2} to \ref{fig:example3_4}) leads to increased Marangoni convection characterized by higher velocity magnitudes and flatter droplet shapes. The combination of very flat droplet shapes and the static wetting angle constraint at the triple points even induces local regions with concave interface curvature for the highest surface tension gradient coefficient $\alpha'_0$ under consideration (Figure~\ref{fig:example3_4}). Besides the flatter droplet shape a higher Marangoni convection also results in a more homogeneous temperature distributions with lower peak temperature values. Both effects, i.e. flat droplets with homogeneous temperature distribution, can be considered as beneficial for typical additive manufacturing processes since problems related to non-smooth surface patterns (e.g. the balling effect due to surface tension-driven Plateau-Rayleigh instabilities~\cite{gusarov2010}) and overheating (e.g. excessive evaporation and keyhole instabilities~\cite{King2014}) are expected to be less pronounced.
\subsubsection{2D laser melting with wetting and Marangoni forces} \label{sec:numex_2Dmelt_wetmar}
\begin{figure}[htbp]
\centering
\subfigure [$\alpha'_0/ \rho_0\!=\!0.0$, $t\!=\!0.2ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni00_t2_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!0.0$, $t\!=\!0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni00_t5_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!0.0$, $t\!=\!0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni00_t6_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!0.0$, $t\!=\!0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni00_t7_00}
}
\\
\subfigure [$\alpha'_0/ \rho_0\!=\!2.0$, $t\!=\!0.2ms$.]
{
\includegraphics[width=0.22\textwidth]{figures/example4_wetangle75_marangoni02_t2_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!2.0$, $t\!=\!0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni02_t5_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!2.0$, $t\!=\!0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni02_t6_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!2.0$, $t\!=\!0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni02_t7_00}
}
\\
\subfigure [$\alpha'_0/ \rho_0\!=\!5.0$, $t\!=\!0.2ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t2_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!5.0$, $t\!=\!0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t5_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!5.0$, $t\!=\!0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t6_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!5.0$, $t\!=\!0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t7_00}
}
\\
\subfigure [$\alpha'_0/ \rho_0\!=\!10.0$, $t\!=\!0.2ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni10_t2_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!10.0$, $t\!=\!0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni10_t5_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!10.0$, $t\!=\!0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni10_t6_00}
}
\subfigure [$\alpha'_0/ \rho_0\!=\!10.0$, $t\!=\!0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni10_t7_00}
}
\caption{2D laser melting with wetting and Marangoni forces: Different Marangoni convection strengths ($\alpha'_0/ \rho_0$ in $[10^{-8} (N \, m^{-1} \, K^{-1})/(kg \, m^{-3})]$) and time steps. Static wetting angle $\theta_0 = 75^o$. Temperature range from $1700K$ (blue) to $3500K$ (red). Solid phase in black. Gas and boundary particles not displayed.}
\label{fig:example4_0}
\end{figure}
\begin{figure}[b!!!]
\centering
\subfigure [$\theta_0 = 30^o$, $t=0.2ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle30_marangoni05_t2_00}
}
\subfigure [$\theta_0 = 30^o$, $t=0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle30_marangoni05_t5_00}
}
\subfigure [$\theta_0 = 30^o$, $t=0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle30_marangoni05_t6_00}
}
\subfigure [$\theta_0 = 30^o$, $t=0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle30_marangoni05_t7_00}
}
\\
\subfigure [$\theta_0 = 60^o$, $t=0.2ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle60_marangoni05_t2_00}
}
\subfigure [$\theta_0 = 60^o$, $t=0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle60_marangoni05_t5_00}
}
\subfigure [$\theta_0 = 60^o$, $t=0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle60_marangoni05_t6_00}
}
\subfigure [$\theta_0 = 60^o$, $t=0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle60_marangoni05_t7_00}
}
\\
\subfigure [$\theta_0 = 75^o$, $t=0.2ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t2_00}
}
\subfigure [$\theta_0 = 75^o$, $t=0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t5_00}
}
\subfigure [$\theta_0 = 75^o$, $t=0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t6_00}
}
\subfigure [$\theta_0 = 75^o$, $t=0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle75_marangoni05_t7_00}
}
\\
\subfigure [$\theta_0 = 90^o$, $t=0.2ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle90_marangoni05_t2_00}
}
\subfigure [$\theta_0 = 90^o$, $t=0.5ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle90_marangoni05_t5_00}
}
\subfigure [$\theta_0 = 90^o$, $t=0.6ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle90_marangoni05_t6_00}
}
\subfigure [$\theta_0 = 90^o$, $t=0.7ms$.]
{
\includegraphics[width=0.23\textwidth]{figures/example4_wetangle90_marangoni05_t7_00}
}
\caption{2D laser melting with wetting and Marangoni forces: Different static wetting angles and time steps. Surface tension gradient coefficient $\alpha'_0/ \rho_0 = 5.0 \cdot 10^{-8} (N \, m^{-1} \, K^{-1})/(kg \, m^{-3})$. Temperature range (left) from $1700K$ (blue) to $3500K$ (red). Solid phase in black. Gas and boundary particles not displayed.}
\label{fig:example4_1}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure [Low Marangoni convection strength: surface tension gradient coefficient $\alpha'_0/ \rho_0=2.0$.]
{
\includegraphics[width=1.0\textwidth]{figures/example4_detail_wetangle75_marangoni02_t5_03.png}
\label{fig:example4_2a}
}
\\
\subfigure [Medium Marangoni convection strength: surface tension gradient coefficient $\alpha'_0/ \rho_0=5.0$.]
{
\includegraphics[width=1.0\textwidth]{figures/example4_detail_wetangle75_marangoni05_t5_01.png}
\label{fig:example4_2b}
}
\\
\subfigure [High Marangoni convection strength: surface tension gradient coefficient $\alpha'_0/ \rho_0=10.0$.]
{
\includegraphics[width=1.0\textwidth]{figures/example4_detail_wetangle75_marangoni10_t5_00.png}
\label{fig:example4_2c}
}
\caption{2D laser melting with wetting and Marangoni forces: Time step $t=0.5ms$. Static wetting angle $\theta_0 = 75^o$. Different Marangoni convection strengths ($\alpha'_0/ \rho_0$ in $[10^{-8} (N \, m^{-1} \, K^{-1})/(kg \, m^{-3})]$). Temperature range from $1700K$ (blue) to $3500K$ (right). Velocity range (right) from $0.0m/s$ to $0.8m/s$. Solid phase in black. Gas and boundary particles not displayed.}
\label{fig:example4_2}
\end{figure}
In a next step, the system considered in the last example is extended by the phase change problem. Thereto, the problem setup will be changed such that a solid domain exists initially instead of a melt phase domain. Specifically, a rectangular domain with dimensions $x \in [-100;100], \, y \in [-60;60]$ (all length dimensions given in $\mu m$) is considered. The lower half of the domain, i.e. $x \in [-100;100], \, y \in [-60; 0]$ is initially covered with a solid phase, the remaining domain with gas. The overall domain is surrounded by three layers of boundary particles. The initial as well as the wall temperature (Dirichlet boundary conditions) have been chosen to $T_0=\hat{T}=500K$. In the time interval $t \in [0;0.5]$ (all time units given in $ms$), a laser heat source ($s^{lg}_{l0} \approx 7.43 \cdot 10^{9} W \, m^{-2}$, i.e. $s^{lg}_{l0}/\rho_0 = 1.0 \cdot 10^6 W \, m \, kg^{-1}$; $r_w=60 \mu m$; located at $x_0=0$; pointing in negative $y-direction$) is acting onto the surface of the melt and solid material. In the time interval $t \in [0.5;1.0]$ the laser beam is switched off to study the subsequent solidification problem. Since melting is relevant in this test case, an additional viscous force contribution at the solid-liquid interface according to~\eqref{eq:sph_dissipation_sl} with the dimensionless viscosity parameter ${\zeta}^{sl}_0=10$ as well as $T_{max}=3500K$ applied. Note that for this specific example an untypically high value of $T_{max}$ has been chosen to increase the overall viscosity within a large portion of the melt pool, which in turn yields a laminar flow and a smooth velocity profile. This parameter setting has proven beneficial to analyze and isolate the influence of the static wetting angle and of the surface tension gradient coefficient on the resulting flow patterns. All remaining material parameters are chosen according to Table~\ref{tab:material_params}. The reference pressure of the weakly compressible model is set to $p_0=1.0 \cdot 10^7 N m^{-2}$ for the melt and to $p_0=1.0 \cdot 10^6 N m^{-2}$ for the gas phase. The background pressure of the transport velocity formulation is set to $p_b=p_0$ for both phases. Moreover, a time step size of $\Delta t = 2.5 \cdot 10^{-6} ms$ and an initial particle spacing of $\Delta x = 5/6 \mu m$ is applied. In Figure~\ref{fig:example4_0}, the melting and solidification process is displayed at different time instances and for different values of the surface tension gradient coefficient $\alpha'_0$ dictating the strength of the Marangoni convection. As already observed in the last example, the Marangoni convection fosters material transport from the melt pool center (high temperatures) to the edge of the melt pool (lower temperatures) leading to a depression at the melt pool center. Moreover, Marangoni convection (from first to fourth row of Figure~\ref{fig:example4_0}) is an effective means of heat transfer and reduces temperature gradients with increasing surface tension gradient coefficient $\alpha'_0$. During the solidification process (third and fourth column of Figure~\ref{fig:example4_0}) the triple point solid-liquid-gas moves with the solidification front towards the center of the melt pool. The wetting forces, tending to enforce the static wetting angle at the current triple point position, induce an upward warping of the liquid-gas interface while the triple point moves towards the melt pool center. As expected, the resulting peak at the center of the solidified melt pool is less pronounced for the higher values of $\alpha'_0$, since the initial depression of the liquid melt pool due to the Marangoni convection (partly) compensates this warping effect.\\
Similarly, in Figure~\ref{fig:example4_1}, the melting and solidification process is displayed at different time instances and for different values of the static wetting angle $\theta_0$. When comparing the different rows in the first and second column of Figure~\ref{fig:example4_1}, only minor differences can be observed in the resulting melt pool shapes. This observation can be explained by the fact that in all considered cases the triple point lies on (or is very close to) a sharp edge of the solid domain for which no unique normal vector $\vectorbold{n}^{sf}$ can be defined. On the other hand, during the solidification of the melt pool the same warping effect of the liquid-gas interface as in Figure~\ref{fig:example4_0} can be observed. As expected the dimension of the resulting peak at the center of the solidified melt pool increases with increasing values of the static wetting angle $\theta_0$.\\
Eventually, in Figure~\ref{fig:example4_2} also the resulting velocity profiles at $t=0.5ms$ are illustrated for the three different values of the surface tension gradient coefficient $\alpha'_0$ as considered above. The resulting velocity profiles are similar to the ones already observed in Section~\ref{sec:numex_heateddrop} with increasing velocity magnitudes for increasing values of the surface tension gradient coefficient $\alpha'_0$. {While the accuracy of the wetting as well as normal and tangential surface tension force contributions has already been verified individually in Sections~\ref{sec:numex_migratingbubble} and~\ref{sec:numex_droplet_oscillation} (and Figure~\ref{fig:example3_1} of the last section), the remaining examples in this and the last section mainly demonstrate the robustness of the numerical scheme when all these interface forces occur simultaneously. Specifically, the example of this section considers melting and solidification, i.e. a dynamic change of the size/shape of the liquid domain, as additional complexity. While the observed influence of wetting and Marangoni effects on the melt pool shape and dynamics is plausible (see discussion above), a comparison with experimentally observed behavior will be considered in our ongoing research work as additional means of validation.}
\subsubsection{2D laser melting with recoil pressure} \label{sec:numex_2Dmelt_rec}
\begin{figure}[b!!!]
\centering
\subfigure [Exemplary melt pool shape from laser melting experiment; taken from~\cite{cunningham2019keyhole}.]
{
\includegraphics[width=0.31\textwidth]{figures/example5_expsimcomp_exp.png}
\label{fig:example5_0a0}
}
\subfigure [Simulation with SPH model: particle distribution \textit{with} interface viscosity.]
{
\includegraphics[width=0.31\textwidth]{figures/example5_expsimcomp_sim1.png}
\label{fig:example5_0a1}
}
\subfigure [Simulation with SPH model: particle distribution \textit{without} interface viscosity.]
{
\includegraphics[width=0.31\textwidth]{figures/example5_expsimcomp_sim2.png}
\label{fig:example5_0a2}
}
\caption{{Melt pool shape from experiment (left). Simulation results for example "2D laser melting with recoil pressure" (\textit{variant 1}) at $t=0.072ms$: Visualization of spurious interface currents and stabilization via viscous interface forces (middle and right). Temperature range from $1700K$ (blue) to $3500K$ (red). Display of solid phase in black and gas phase in light blue.}}
\label{fig:example5_0a}
\end{figure}
The effects studied in Section~\ref{sec:numex_2Dmelt_wetmar} will now be extended by evaporation-induced recoil pressure forces according to~\eqref{eq:fluid_recoil}. In order to study deep keyhole dynamics in the range of high recoil pressure magnitudes, a larger domain with dimensions $x \in [-200;200], \, y \in [-200;200]$ (length dimensions given in $\mu m$) is considered. The lower region with $x \in [-200;200], \, y \in [-200; 40]$ is initially covered with a solid phase, the remaining domain with gas. The overall domain is surrounded by three layers of boundary particles. The initial as well as the wall temperature (Dirichlet boundary conditions) have been chosen to $T_0=\hat{T}=500K$. A laser heat source ($s^{lg}_{l0} \approx 16.0 W \, \mu m^{-2}$; $r_w=30 \mu m$; located at $x_0=0$; pointing in negative $y-$direction) is acting onto the surface of the melt and solid material. Again, an additional viscous force contribution at the solid-liquid interface according to~\eqref{eq:sph_dissipation_sl} with the dimensionless viscosity parameter ${\zeta}^{sl}_0=5$ as well as $T_{max}=2000K$ is applied. Moreover, to reduce numerical instabilities at the liquid-gas interface due to the high recoil pressure forces acting in this example, at this interface an additional viscous force according to~\eqref{eq:sph_dissipation_lg} with ${\zeta}^{lg}_0=2.5 \cdot 10^{-4}$ is applied. To isolate the effect of surface tension-recoil pressure interaction, no wetting and Marangoni forces are considered. All remaining material parameters are chosen according to Table~\ref{tab:material_params}. The reference pressure of the weakly compressible model is set to $p_0=1.0 \cdot 10^7 N m^{-2}$ for the melt and gas phase.
\begin{figure}[b!!!]
\centering
\begin{minipage}{0.95\textwidth}
\centering
\subfigure [Particle distribution \textit{with} interface viscosity.]
{
\includegraphics[width=0.45\textwidth]{figures/example5_0/deep5_img0050.png}
\label{fig:example5_0b1}
}
\subfigure [Particle distribution \textit{without} interface viscosity.]
{
\includegraphics[width=0.45\textwidth]{figures/example5_0/deep2b_img0050.png}
\label{fig:example5_0b2}
}
\subfigure [Velocity field \textit{with} interface viscosity.]
{
\includegraphics[width=0.45\textwidth]{figures/example5_0/deep5_velarrow_img0050.png}
\label{fig:example5_0b3}
}
\subfigure [Velocity field \textit{without} interface viscosity.]
{
\includegraphics[width=0.45\textwidth]{figures/example5_0/deep2b_velarrow_img0050.png}
\label{fig:example5_0b4}
}
\caption{2D laser melting with recoil pressure according to \textit{variant 2} at time $t=0.0255ms$: Visualization of spurious interface currents and stabilization via viscous interface forces. Temperature range from $1700K$ (blue) to $3500K$ (red). Velocity range from $0ms^{-1}$ (blue) to $10ms^{-1}$ (red). Display of solid phase in black and gas phase in light blue. Note that only a subdomain with $x \in [-100;100], \, y \in [-100; 100]$ is displayed.}
\label{fig:example5_0b}
\end{minipage}
\end{figure}
The background pressure of the transport velocity formulation is set to $p_b=5p_0$ for both phases. Moreover, a time step size of $\Delta t = 1.0 \cdot 10^{-6} ms$ and an initial particle spacing of $\Delta x = 5/3 \mu m$ is applied. The problem setup described by these parameters is denoted as \textit{variant 1} in the following. In order to study the influence of the evaporative mass loss~\eqref{eq:fluid_evaporation} a second variant (\textit{variant 2}) of this problem is considered where this term is neglected (while the recoil pressure is still considered). In order to end up with comparable thermal characteristics the laser irradiance has been decreased to a value of $s^{lg}_{l0} \approx 0.08 W \, \mu m^{-2}$ for \textit{variant 2}, which turned out to result in comparable peak temperatures in the melt pool center. Due to the higher interface dynamics in \textit{variant 2} the viscosity factor in~\eqref{eq:sph_dissipation_lg} is increased to ${\zeta}^{lg}_0=1.0 \cdot 10^{-3}$. Moreover, for \textit{variant 2} a smaller domain with dimensions $x \in [-200;200], \, y \in [-150; 150]$ is considered.\\
To study the influence of the viscous interface force, simulations with and without this additional stabilization term shall be compared. In Figure~\ref{fig:example5_0a} the resulting melt pool shape at $t=0.072ms$ is displayed for these two cases and problem \textit{variant 1}. Without interface stabilization term (Figure~\ref{fig:example5_0a2}) the spurious interface flows lead to a non-smooth, strongly distorted interface topology fostering nonphysical interface dynamics and eventually instability of the numerical scheme. In contrast, the additional interface stabilization term (Figure~\ref{fig:example5_0a1}) results in a smooth and physically reasonable interface topology by damping out the spurious interface flows and thus effectively stabilizing the numerical scheme. The high recoil pressure forces form a deep depression at the center of the melt pool, typically denoted as keyhole. Note that the configuration displayed in Figure~\ref{fig:example5_0a1} is no stationary configuration of the melt pool but rather characterized by high-frequency fluctuations of the liquid-gas interface especially at the bottom of the keyhole. {Qualitatively, Figure~\ref{fig:example5_0a1} shows good agreement with representative experimental results (see Figure~\ref{fig:example5_0a0}, taken from~\cite{cunningham2019keyhole}; note that the specific process parameters from experiment and simulation don't match.)}\\
In Figure~\ref{fig:example5_0b} the melt pool shape resulting from simulations with and without additional stabilization term is displayed for problem \textit{variant 2} at $t=0.0255ms$. The neglect of evaporative heat losses in problem \textit{variant 2} leads to stronger recoil pressure dynamics and, consequently, the spurious interface flows are even more pronounced. The result is a strongly distorted interface topology and velocity field as illustrated in Figures~\ref{fig:example5_0b2} and~\ref{fig:example5_0b4}. Again, the additional interface stabilization term results in a smooth and physically reasonable interface topology and velocity field as shown in Figures~\ref{fig:example5_0b1} and~\ref{fig:example5_0b3}.
\begin{figure}[htbp]
\centering
\subfigure [time: $t=0.079ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largeandhighdomain_coarse_highbp_highrecoil_evaploss_5_img0078.png}
\label{fig:example5_2_1}
}
\subfigure [time: $t=0.083ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largeandhighdomain_coarse_highbp_highrecoil_evaploss_5_img0082.png}
\label{fig:example5_2_2}
}
\subfigure [time: $t=0.085ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largeandhighdomain_coarse_highbp_highrecoil_evaploss_5_img0084.png}
\label{fig:example5_2_3}
}
\subfigure [time: $t=0.089ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largeandhighdomain_coarse_highbp_highrecoil_evaploss_5_img0088.png}
\label{fig:example5_2_4}
}
\caption{2D laser melting with recoil pressure according to \textit{variant 1}: creation mechanism of a gas inclusion at different time steps. Temperature range from $1700K$ (blue) to $3500K$ (red). Display of solid phase in black and gas phase in light blue.}
\label{fig:example5_2}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure [time: $t=0.056ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largedomain_coarse_highbp_2_img0055.png}
\label{fig:example5_1_1}
}
\subfigure [time: $t=0.082ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largedomain_coarse_highbp_2_img0081.png}
\label{fig:example5_1_2}
}
\subfigure [time: $t=0.109ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largedomain_coarse_highbp_2_img0108.png}
\label{fig:example5_1_3}
}
\subfigure [time: $t=0.135ms$]
{
\includegraphics[width=0.23\textwidth]{figures/recoil_2d_largedomain_coarse_highbp_2_img0134.png}
\label{fig:example5_1_4}
}
\caption{2D laser melting with recoil pressure according to \textit{variant 2}: process of melt drop ejection at different time steps. Temperature range from $1700K$ (blue) to $3500K$ (red). Display of solid phase in black and gas phase in light blue.}
\label{fig:example5_1}
\end{figure}
Eventually, the melt pool thermo-hydrodynamics resulting from the two considered problem variants shall be studied for longer physical time spans. Figure~\ref{fig:example5_2} displays the system behavior of \textit{variant 1}. Accordingly, the interaction of surface tension and recoil pressure forces leads to oscillations of the liquid-gas interface with maximal amplitudes slightly above the bottom of the keyhole. Once, the amplitudes of these oscillations are large enough such that the opposite keyhole walls gets into touch a liquid bridge forms, which effectively encloses the gas material at the keyhole base - a gas inclusion is formed. Note that this example mainly aims at demonstrating the robustness of the proposed formulation in representing the highly dynamic surface tension-recoil pressure interaction and eventually the formation of a gas bubble. Of course, the employed phenomenological recoil pressure model does not explicitly resolve the high-velocity vapor jet that would arise from the keyhole base in the real physical system. The latter might considerably influence the described creation mechanism of a liquid bridge eventually resulting in a gas inclusion. Moreover, the present formulation does not apply any form of ray tracing for the laser heat source. Specifically, once the gas bubble has formed, the present model applies a heat source term not only on the upper interface of the arising liquid bridge but also at the bottom interface of the gas inclusion (which is visible in Figure~\ref{fig:example5_2_4} through the high peak temperatures at these locations). While an elimination of this deficiency (e.g. via ray tracing) can be considered as rather straight-forward process, the present formulation is still deemed suitable to demonstrate the robustness and general working principle of the proposed model and to study the problem-specific physical phenomena except for the subsequent motion of the closed gas bubble. Figure~\ref{fig:example5_1} displays the melt pool dynamics resulting from problem \textit{variant 2}. The high recoil pressure magnitudes in this variant result in periodic flow patterns consisting of recoil pressure-driven high-velocity waves from the center to the edges of the melt pool and surface tension-driven back flow from the edges to the center. The continued energy input by the laser beam heat sources results in increasing amplitudes of these flow oscillations until eventually melt droplets are ejected at the melt pool edges (Figures~\ref{fig:example5_2_3} and~\ref{fig:example5_2_4}). This example demonstrates that also the highly dynamic evolution of the interface topology during melt drop ejection can be captured by the proposed formulation in a robust manner.
\subsubsection{3D laser melting problem with explicitly resolved powder particle geometry} \label{sec:numex_3Dmelt}
\begin{figure}[htbp]
\centering
\subfigure [time: $0.0ms$]
{
\includegraphics[width=0.44\textwidth]{figures/example6_4c-restart-temperature0000.jpg}
\label{fig:example6_1a}
}
\subfigure [time: $0.080ms$]
{
\includegraphics[width=0.44\textwidth]{figures/example6_4c-restart-temperature0015.jpeg}
\label{fig:example6_1b}
}
\subfigure [time: $0.180ms$]
{
\includegraphics[width=0.44\textwidth]{figures/example6_4c-restart-temperature0035.jpeg}
\label{fig:example6_1c}
}
\subfigure [time: $0.455ms$]
{
\includegraphics[width=0.44\textwidth]{figures/example6_4c-restart-temperature0090.jpeg}
\label{fig:example6_1d}
}
\subfigure [time: $0.630ms$]
{
\includegraphics[width=0.44\textwidth]{figures/example6_4c-restart-temperature0125.jpeg}
\label{fig:example6_1e}
}
\subfigure [time: $1.000ms$]
{
\includegraphics[width=0.44\textwidth]{figures/example6_4c-restart-temperature0199.jpeg}
\label{fig:example6_1f}
}
\caption{3D point melting problem with explicitly resolved powder particle geometry: resulting melt pool shape and final topology of solidified surface (post-processed) as well as temperature field in the range from $300K$ (blue) to $3700K$ (red).}
\label{fig:example6_1}
\end{figure}
In this last example, the problem considered in the last section shall be extended to 3D as well as more complex and realistic geometries by resolving individual particles of the metal powder in PBFAM as illustrated in Figure~\ref{fig:example6_1a}. The problem domain is confined by $x \in [-100;100], \, y \in [-100;100 ], \, z \in [-50;50]$ (length dimensions given in $\mu m$). The lower half with $x \in [-100;100], \, y \in [-100;100], \, z \in [-50;0],$ is initially covered with a solid phase, the remaining domain with gas. In addition, (spatially fixed) powder particles with diameters between $160 \mu m$ and $320 \mu m$ are randomly placed on top of the solid substrate. The overall domain is surrounded by three layers of boundary particles. The initial as well as the wall temperature (Dirichlet boundary conditions) have been chosen to $T_0=\hat{T}=500K$. A laser heat source ($s^{lg}_{l0} \approx 1.5 W \, \mu m^{-2}$; $r_w=60 \mu m$; located at $x_0=0$; pointing in negative $z-$direction) is acting onto the surface of the melt and solid material. The material parameters are chosen according to Table~\ref{tab:material_params}. The reference pressure of the weakly compressible model is set to $p_0=1.0 \cdot 10^7 N m^{-2}$ for the melt and gas phase. The background pressure of the transport velocity formulation is set to $p_b=5p_0$ for both phases. Moreover, a time step size of $\Delta t = 1.0 \cdot 10^{-6} ms$ and an initial particle spacing of $\Delta x = 5/3 \mu m$ is applied. The laser beam is switched on only for the time interval $t \in [0;0.5]$ and the cooling / solidification process is considered in the remaining time interval $t \in [0.5;1]$. Thus, in total the discrete problem consists of one million time steps and approximately one million SPH particles. Figure~\ref{fig:example6_1} illustrates the melting and solidification process at different time steps. Surface tension forces dominate volumetric forces such as gravity at the considered length scales and smoothen out the original particle contours almost immediately after melting (see Figures~\ref{fig:example6_1b} and~\ref{fig:example6_1c}). The peak temperatures at the melt pool center exceed the boiling temperature of the liquid metal with ongoing exposure time. The resulting recoil pressure forces foster an increasingly deep depression at the melt pool center as illustrated in Figures~\ref{fig:example6_1c} and~\ref{fig:example6_1d}. For the chosen set of parameters the time scales governing the surface tension forces dominate the time scales of (conductive) heat transfer. Thus, once the laser is switched off, the surface tension forces close the melt pool depression before the liquid metal solidifies again resulting in a smooth and rather plain surface topology at the former melt pool center (see Figure~\ref{fig:example6_1f}).
\begin{figure}[htbp]
\centering
\subfigure [time: approx. $0.12ms$ \hspace{0.22\textwidth}]
{
\includegraphics[width=0.48\textwidth]{figures/linescan_config1.png}
\label{fig:example6_2a}
}
\subfigure [time: approx. $0.24ms$ \hspace{0.22\textwidth}]
{
\includegraphics[width=0.48\textwidth]{figures/linescan_config2.png}
\label{fig:example6_2b}
}
\subfigure [time: approx. $0.36ms$ \hspace{0.22\textwidth}]
{
\includegraphics[width=0.48\textwidth]{figures/linescan_config3.png}
\label{fig:example6_2c}
}
\subfigure [time: approx. $0.48ms$ \hspace{0.22\textwidth}]
{
\includegraphics[width=0.48\textwidth]{figures/linescan_config4.png}
\label{fig:example6_2d}
}
\caption{3D line melting problem with explicitly resolved powder particle geometry: resulting melt pool shape and final topology of solidified surface (post-processed) as well as temperature field in the range from $300K$ (blue) to $3700K$ (red).}
\label{fig:example6_2}
\end{figure}
Eventually, the example is extended to line scanning by increasing the total length of the domain by a factor of four in $x-$direction, increasing the laser energy density by a factor of two and moving the laser with a constant velocity of $v_x=1m/s$ in $x-$direction. The results are displayed in Figure~\ref{fig:example6_2}. All in all, both variants of this example, i.e. point and line scanning, demonstrate the robustness and general suitability of the proposed SPH formulation for 3D thermo-capillary phase change problems with geometrically complex features as occurring in PBFAM.
\section{Conclusion and Outlook} \label{sec:concl}
In the present work, a weakly compressible SPH formulation for thermo-capillary phase change problems involving solid, liquid and gaseous phases has been proposed with special focus on laser melting processes such as metal additive manufacturing. Evaporation-induced recoil pressure, temperature-dependent surface tension and wetting forces have been considered as mechanical interface fluxes, while a Gaussian laser beam heat source and evaporation-induced heat losses have been considered as thermal interface fluxes.\\
Specifically, a novel interface stabilization scheme based on viscous interface forces has been proposed, which was shown to effectively damp spurious interface flows well-known for continuum surface force approaches. It was demonstrated that only by this means recoil pressure-dominated problems, which typically arise in metal AM processes in the range of high laser powers, can be represented in a stable and physically accurate manner. Moreover, different SPH discretizations for the tangential projection of the temperature gradient, as required for the discrete Marangoni forces, have been reviewed. It was demonstrated analytically and numerically that standard \textit{two-sided} gradient approximations are sufficient for this purpose as long as zero-order consistency is satisfied, e.g. by anti-symmetric gradient construction.\\
In the context of metal AM melt pool modeling, the proposed formulation has been employed to study the influence of wetting forces and Marangoni convection of different strength on the resulting melt pool dynamics and shape as well as on the topology of the final solidified metal surface. Moreover, the robustness of the proposed scheme has been demonstrated by representing highly dynamic recoil-pressure-induced fluctuations and significant topology changes of the liquid-gas interface, e.g. during the ejection of liquid droplets from the melt pool or the inclusion of gas bubbles into the melt pool. Finally, the robustness, efficiency and general suitability of the proposed SPH formulation for 3D thermo-capillary phase change problems with geometrically complex features has been indicated by means of a 3D laser melting problem with explicit resolution of individual metal powder particles.
\section*{Acknowledgements} \label{sec:Acknowledgements}
This work was supported by a postdoc fellowship of the German Academic Exchange Service (DAAD) and by funding of the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) within project 437616465 and project 414180263. The authors would like to thank Lennart Schulze for his contributions in the code implementation and comparison of different temperature gradient discretization strategies and Yushen Sun for his support in the visualization of 3D melt pool simulations.
\bibliographystyle{elsarticle-num}
| c8eff699e9f14c5ecc6aee82b417d24327e6f5ee | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{S:Introduction}
Let $G=(V,E)$ be a simple connected graph. The {\it Wiener index} of a graph $G$ is defined as the sum of distances between all pairs of vertices of $G$:
$$W(G)=\frac{1}{2}\sum_{v,u\in V, v\neq u}{d_G(u,v)},$$
where $d_G(x,y)$ is a usual graph distance, i.e. the length of a shortest path from $x$ to $y$ in $G$.
This index was introduced in $1947$ by Wiener \cite{W1947} for applications in chemistry, where he used similar expression to determine the boiling point of the paraffin. In $1971$, Hosoya \cite{H1971} first defined this index in terms of graph theory. After that, the index for different families of graphs was investigated because of pure mathematical interest. Very recently, Egorov and Vesnin \cite{EgorVesn} investigated correlation of hyperbolic volumes of fullerenes (fullerene is a molecule consisting entirely of carbon atoms) with Wiener index and other related indexes. We also refer an interested reader to survey papers by Dobrynin, Entringer and Gutman \cite{DEG2001} and Knor, \v{S}krekovski and Tepeh \cite{KST2016}.
Let us define the {\it transmission} of a vertex $v\in V$ in $G$ as $$t_G(v)=\sum_{u\in V\setminus{\{v\}}}{d_G(v,u)}.$$
Clearly, the Wiener index of a graph $G$ may be expressed in terms of vertices transmissions:
\begin{equation*}
W(G)=\frac{1}{2}\sum_{v\in V}{t_G(v)}.
\end{equation*}
Let us denote by $G-v$ the graph obtained by removing the vertex $v$ from $G$. A vertex $v\in V$ is called {\it good} if $G-v$ is connected and $W(G) = W(G - v)$. In 1991, \v{S}olt\'{e}s \cite{S1991} posed the following problem.\\
\begin{problem}\label{Prob:1}
Find all graphs $G$ such that the equality $W(G) = W(G - v)$ holds for all their vertices $v$.
\end{problem}
Today the only known graph with this property is $C_{11}$. In other words, \v{S}olt\'{e}s asked about graphs such that all their vertices are good. After thirty years of study of this problem, which led to lots of elegant ideas, the problem is far from being solved. However, there are some partial results. In $2018$, Knor, Majstorovi\'{c} and \v{S}krekovski in \cite{KMS2018dam} constructed an infinite series of graphs with one good vertex. The same authors also showed \cite{KMS2018amc} that for $k\geq 3$ there are infinitely many graphs with one good vertex of degree $k$. In \cite{BJM2019}, for any $k\in \mathbb{N}$, Bok, Jedli\v{c}kov\'{a} and Maxov\'{a} found an infinite series of graphs with exactly $k$ good vertices.
For a graph $G=(V,E)$ and vertices $v,u,w\in V$, let us define functions $\Delta_v(G)=W(G)-W(G-v)$ and $\delta_G^v(u,w)=d_G(u,w)-d_{G-v}(u,w)$. From the definitions we directly have the following
\begin{proposition}\label{Prop:1}
Given a graph $G=(V,E)$, $v\in V$. If $G-v$ is connected then the following equality holds:
$$\Delta_v(G)=t_G(v)+\frac{1}{2}\sum_{u,w \in V\setminus \{v\}}{\delta^v_G(u,w)}.$$
\end{proposition}
Instead of finding graphs with a fixed number of good vertices, one may try to find graphs with a fixed proportion of such vertices. So we denote a {\it proportion of good vertices} in $G$ as $\frac{|\{v\in V|\Delta_v(G)=0\}|}{|V|}.$ Now we are ready to formulate the following relaxation of Problem \ref{Prob:1}.
\begin{problem}\label{Prob:2}
For a fixed $\alpha\in (0,1]$ construct an infinite series $S$ of graphs such that for all $G=(V,E)$ from $S$ the following inequality takes place: $$ \frac{|\{v\in V|\Delta_v(G)=0\}|}{|V|} \geq \alpha.$$
\end{problem}
Clearly, a solution to this problem for $\alpha=1$ would give an infinite series of solutions to the Problem \ref{Prob:1}.
Let us note that the construction from \cite{BJM2019} after a slight modification yields an infinite series of graphs with the proportion of good vertices tending to $\frac{1}{3}$ (we will discuss it in more details in Section \ref{S:Lily}).
In this work, we provide two constructions based on different ideas in order to improve this constant.
In Section 2, we find an infinite series of graphs with the proportion of good vertices tending to $\frac{2}{5}$ as the number of vertices tends to infinity. In Section 3, we find one more series with the proportion tending to $\frac{1}{2}$. In Section 4, we discuss results and open problems.
\section{Bunch of $11$-cycles}\label{S:Bunch}
As it was mentioned in Section \ref{S:Introduction}, $C_{11}$ is the only one known graph where all vertices are good. Now we provide an intriguing construction of graphs with many $C_{11}$ as induced subgraphs.
\begin{construction}\label{C:bunch} Given $k\in \mathbb{N}$, $k>1$. Let us define a graph $B(k)$.
\begin{itemize}
\item Step $1$. Take a path of length $5$ and denote its consecutive vertices by $w_0$, $w_1$, $w_2$, $w_3$, $w_4$ and $w_5$.
\item Step $2$. Take $k$ distinct paths $(v_{i1},v_{i2},v_{i3},v_{i4},v_{i5})$, $i\in \{1,2, \dots, k\}$.
\item Step $3$. For $i\in \{1,2, \dots, k\}$, add edges $(w_0, v_{i1})$ and $(w_5, v_{i5})$ to our graph.
\end{itemize}
\end{construction}
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={scale=0.9}, rotate=270, yscale=-1, xscale=-1 ]
\node[draw, circle] (w0) at (0,1) {$w_0$};
\node[draw, circle] (w5) at (0,-7) {$w_5$};
\node[draw, circle] (w1) at (-4,-1) {$w_1$};
\node[draw, circle] (w2) at (-4,-2.33) {$w_2$};
\node[draw, circle] (w3) at (-4,-3.66) {$w_3$};
\node[draw, circle] (w4) at (-4,-5) {$w_4$};
\draw [thick] (w0) to (w1);
\draw [thick] (w1) to (w2);
\draw [thick] (w2) to (w3);
\draw [thick] (w3) to (w4);
\draw [thick] (w4) to (w5);
\node[draw, circle] (v11) at (-2,-1) {$v_{11}$};
\node[draw, circle] (v12) at (-2,-2) {$v_{12}$};
\node[draw, circle] (v13) at (-2,-3) {$v_{13}$};
\node[draw, circle] (v14) at (-2,-4) {$v_{14}$};
\node[draw, circle] (v15) at (-2,-5) {$v_{15}$};
\draw [thick] (w0) to (v11);
\draw [thick] (v11) to (v12);
\draw [thick] (v12) to (v13);
\draw [thick] (v13) to (v14);
\draw [thick] (v14) to (v15);
\draw [thick] (v15) to (w5);
\node[draw, circle] (v21) at (-1,-1) {$v_{21}$};
\node[draw, circle] (v22) at (-1,-2) {$v_{22}$};
\node[draw, circle] (v23) at (-1,-3) {$v_{23}$};
\node[draw, circle] (v24) at (-1,-4) {$v_{24}$};
\node[draw, circle] (v25) at (-1,-5) {$v_{25}$};
\draw [thick] (w0) to (v21);
\draw [thick] (v21) to (v22);
\draw [thick] (v22) to (v23);
\draw [thick] (v23) to (v24);
\draw [thick] (v24) to (v25);
\draw [thick] (v25) to (w5);
\draw [fill=black!100, color=black!100] (-0.2,-3) circle (0.01cm);
\draw [fill=black!100, color=black!100] (0,-3) circle (0.01cm);
\draw [fill=black!100, color=black!100] (0.2,-3) circle (0.01cm);
\node[draw, circle] (vk1) at (1,-1) {$v_{k1}$};
\node[draw, circle] (vk2) at (1,-2) {$v_{k2}$};
\node[draw, circle] (vk3) at (1,-3) {$v_{k3}$};
\node[draw, circle] (vk4) at (1,-4) {$v_{k4}$};
\node[draw, circle] (vk5) at (1,-5) {$v_{k5}$};
\draw [thick] (w0) to (vk1);
\draw [thick] (vk1) to (vk2);
\draw [thick] (vk2) to (vk3);
\draw [thick] (vk3) to (vk4);
\draw [thick] (vk4) to (vk5);
\draw [thick] (vk5) to (w5);
\node[below,font=\large\bfseries] at (current bounding box.south) {Figure 1. Graph $B(k)$ on $5k+6$ vertices};
\end{tikzpicture}
\end{center}
\begin{theorem}\label{T:bunch}
For $k\in \mathbb{N}$, $k\geq 2$, the proportion of good vertices in $B(k)$ equals $\frac{2k}{5k+6}$.
\end{theorem}
\begin{proof}
Consider a vertex $v=v_{11}$ of the graph $G=B(k)=(V,E)$. Our goal is to show that $\Delta_{v}(B(k))=0$. According to Proposition \ref{Prop:1}, one need to find $t_G(v)$ and $\delta^{v}_G(u,w)$ for all vertices $u,w\in V\setminus \{v\}$.
By direct calculations we have that $$t_G(v)=1+2+3+4+5+5+4+3+2+1+(k-1)(2+3+4+5+6)=20k+10.$$ It is easy to see from the construction of the graph that $\delta^{v}_G(u,w)=0$ whenever $u,w\notin \{v_{12},v_{13},v_{14},v_{15}\}$. All nonzero values of $\delta^{v}_G(u,w)$ are presented in the Table \ref{Tab:bunch} (it also contains some zeros):
\begin{table}[h]
\centering
\begin{tabular}{ |c|c|c|c|c|c|c|c|}
\hline
Vertices & $w_0$ & $w_1$ & $w_2$ & $w_3$ & $v_{i1}$ & $v_{i2}$ & $v_{i3}$ \\
\hline
$v_{12}$ & $-7$ & $-5$ & $-3$ & $-1$ & $-6$ & $-4$ & $-2$ \\
$v_{13}$ & $-5$ & $-3$ & $-1$ & $0$ & $-4$ & $-2$ & $0$ \\
$v_{14}$ & $-3$ & $-1$ & $0$ & $0$ & $-2$ & $0$ & $0$ \\
$v_{15}$ & $-1$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$\\
\hline
\end{tabular}
\caption{Values of $\delta^{v}_G(u,w)$, $i\in \{2,3,\dots, k\}$}
{\label{Tab:bunch}}
\end{table}
Consequently, $$\frac{1}{2}\sum_{u,w \in V\setminus \{v\}}{\delta^{v}_G(u,w)}=-(7+5+3+1+5+3+1+3+1+1)-(k-1)(6+4+2+4+2+2)=-20k-10.$$
By Proposition \ref{Prop:1} we have that $\Delta_{v}(G)=0$.
Based on the symmetry of $B(k)$ we conclude that for all $z\in \{v_{11}, v_{21}, \dots, v_{k1}, v_{15}, v_{25}, \dots, v_{k5}\}$ we have $\Delta_{z}(G)=0$. By similar calculations it is easy to check, that for other vertices $z\in V$ we have $\Delta_{z}(G)\neq 0$. Therefore, $B(k)$ has exactly $2k$ good vertices among its $5k+6$ ones.
\end{proof}
\begin{corollary}\label{Cor:bunch}
The series of graphs $B(k)$, $k\in \mathbb{N}$, has a proportion of good vertices tending to $\frac{2}{5}$ as $k$ tends to infinity.
\end{corollary}
From Construction \ref{C:bunch} one may see that the graph $B(k)$ is $2$-connected, so the value $\Delta_v(B(k))$ is defined for any vertex $v$. In the next Section, we let a vertex-connectivity be $1$. It allows us to present one more construction with a proportion of good vertices tending to $\frac{1}{2}$.
\section{Lily-shaped construction}\label{S:Lily}
In this Section, we find one more series of graphs with a proportion of good vertices tending to $\frac{1}{2}$.
\begin{construction}\label{C:lily} Given $k,m\in \mathbb{N}$, $m\geq 7$ and $k\geq \frac{m-3}{m-6}$. Let us define a graph $L(k,m)$.
\begin{itemize}
\item Step $1$.
Define a graph containing $k$ distinct paths of length $3$ of the form $(u_0, v^i_{1j}, v^i_{2j}, v_i)$, $j\in \{1,2,\dots, k\}$, $i=1$. Note, that all these paths have two common vertices $u_0$ and $v_i$. In the rest of the paper, this graph will be called a {\it block}.
\item Step $2$. Add $m-1$ distinct blocks of the same form to our graph. The only difference is a changing index $i\in \{2,3,\dots, m\}$. Note, that all these blocks contain exactly one common vertex $u_0$.
\item Step $3$. For all $i_1,i_2\in \{1,2,\dots m\}$ and $j_1,j_2\in \{1,2,\dots k\}$ such that $i_1 \neq i_2$, add the edge $(v^{i_1}_{1j_1},v^{i_2}_{1j_2})$ to our graph. In other words, the graph induced by the set of vertices $\{v^i_{1j}| i\in \{1,2,\dots m\}, j\in \{1,2,\dots k\}\}$ is a the complete $m$-partite graph with parts $\{v^i_{11}, v^i_{12}, \dots , v^i_{1k} \}$, $i\in \{1,2,\dots, m\}$.
\item Step $4$. Add a path $(u_0, u_1, u_2, \dots, u_d)$ to our graph, where $$d=\Big\lfloor{\frac{\sqrt{8km-48k-8m+33}+1}{2}}\Big\rfloor.$$
\item Step $5$. Add one neighbour $u'_{t}$ to the vertex $u_t$, where $$t=km-6k-m+3-\frac{(d+1)(d-2)}{2}.$$
\end{itemize}
\end{construction}
\begin{center}
\begin{tikzpicture}[thick, scale=1, every node/.style={scale=0.9}]
\def \l {1}
\def 3.3 {3.3}
\def 30 {30}
\node[draw, rectangle] (w) at (0, 0) {$u_0$};
\node[draw, rectangle] (b1v1) at ({3*3.3*cos(30)}, {3*3.3*sin(30))}) {$v_1$};
\def (\d^2+(3*\l/2)^2)^(1/2) {(3.3^2+(3*\l/2)^2)^(1/2)}
\def atan((3/2*\l)/(\d)) {atan((3/2*\l)/(3.3))}
\node[draw, rectangle] (b1v11) at ({(\d^2+(3*\l/2)^2)^(1/2)*cos(30+atan((3/2*\l)/(\d)))} , {(\d^2+(3*\l/2)^2)^(1/2)*sin(30+atan((3/2*\l)/(\d)))}) {$v^1_{11}$};
\def ((2*\d)^2+(3*\l/2)^2)^(1/2) {((2*3.3)^2+(3*\l/2)^2)^(1/2)}
\def atan((3/2*\l)/(2*\d)) {atan((3/2*\l)/(2*3.3))}
\node[draw, rectangle] (b1v21) at ({((2*\d)^2+(3*\l/2)^2)^(1/2)*cos(30+atan((3/2*\l)/(2*\d)))} , {((2*\d)^2+(3*\l/2)^2)^(1/2)*sin(30+atan((3/2*\l)/(2*\d)))}) {$v^1_{21}$};
\def (\d^2+(\l/2)^2)^(1/2) {(3.3^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(\d)) {atan((\l/2)/(3.3))}
\node[draw, rectangle] (b1v12) at ({(\d^2+(\l/2)^2)^(1/2)*cos(30+atan((\l/2)/(\d)))} , {(\d^2+(\l/2)^2)^(1/2)*sin(30+atan((\l/2)/(\d)))}) {$v^1_{12}$};
\def ((2*\d)^2+(\l/2)^2)^(1/2) {((2*3.3)^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(2*\d)) {atan((\l/2)/(2*3.3))}
\node[draw, rectangle] (b1v22) at ({((2*\d)^2+(\l/2)^2)^(1/2)*cos(30+atan((\l/2)/(2*\d)))} , {((2*\d)^2+(\l/2)^2)^(1/2)*sin(30+atan((\l/2)/(2*\d)))}) {$v^1_{22}$};
\def ((3*\d/2)^2+(\l/2)^2)^(1/2) {((3*3.3/2)^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(3*\d/2)) {atan((\l/2)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(\l/2)^2)^(1/2)*cos(30-atan((\l/2)/(3*\d/2)))} , {((3*\d/2)^2+(\l/2)^2)^(1/2)*sin(30-atan((\l/2)/(3*\d/2)))}) circle (0.01cm);
\def ((3*\d/2)^2+(\l/4)^2)^(1/2) {((3*3.3/2)^2+(\l/4)^2)^(1/2)}
\def atan((\l/4)/(3*\d/2)) {atan((\l/4)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(\l/4)^2)^(1/2)*cos(30-atan((\l/4)/(3*\d/2)))} , {((3*\d/2)^2+(\l/4)^2)^(1/2)*sin(30-atan((\l/4)/(3*\d/2)))}) circle (0.01cm);
\def ((3*\d/2)^2+(3*\l/4)^2)^(1/2) {((3*3.3/2)^2+(3*\l/4)^2)^(1/2)}
\def atan((3*\l/4)/(3*\d/2)) {atan((3*\l/4)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(3*\l/4)^2)^(1/2)*cos(30-atan((3*\l/4)/(3*\d/2)))} , {((3*\d/2)^2+(3*\l/4)^2)^(1/2)*sin(30-atan((3*\l/4)/(3*\d/2)))}) circle (0.01cm);
\def (\d^2+(3*\l/2)^2)^(1/2) {(3.3^2+(3*\l/2)^2)^(1/2)}
\def atan((3*\l/2)/(\d)) {atan((3*\l/2)/(3.3))}
\node[draw, rectangle] (b1v1k) at ({(\d^2+(3*\l/2)^2)^(1/2)*cos(30-atan((3*\l/2)/(\d)))} , {(\d^2+(3*\l/2)^2)^(1/2)*sin(30-atan((3*\l/2)/(\d)))}) {$v^1_{1k}$};
\def ((2*\d)^2+(3*\l/2)^2)^(1/2) {((2*3.3)^2+(3*\l/2)^2)^(1/2)}
\def atan((3*\l/2)/(2*\d)) {atan((3*\l/2)/(2*3.3))}
\node[draw, rectangle] (b1v2k) at ({((2*\d)^2+(3*\l/2)^2)^(1/2)*cos(30-atan((3*\l/2)/(2*\d)))} , {((2*\d)^2+(3*\l/2)^2)^(1/2)*sin(30-atan((3*\l/2)/(2*\d)))}) {$v^1_{2k}$};
\draw [thick] (w) to (b1v11);
\draw [thick] (w) to (b1v12);
\draw [thick] (w) to (b1v1k);
\draw [thick] (b1v11) to (b1v21);
\draw [thick] (b1v12) to (b1v22);
\draw [thick] (b1v1k) to (b1v2k);
\draw [thick] (b1v21) to (b1v1);
\draw [thick] (b1v22) to (b1v1);
\draw [thick] (b1v2k) to (b1v1);
\def 150 {150}
\node[draw, rectangle] (b2v1) at ({3*3.3*cos(150)}, {3*3.3*sin(150))}) {$v_2$};
\def (\d^2+(3*\l/2)^2)^(1/2) {(3.3^2+(3*\l/2)^2)^(1/2)}
\def atan((3/2*\l)/(\d)) {atan((3/2*\l)/(3.3))}
\node[draw, rectangle] (b2v11) at ({(\d^2+(3*\l/2)^2)^(1/2)*cos(150+atan((3/2*\l)/(\d)))} , {(\d^2+(3*\l/2)^2)^(1/2)*sin(150+atan((3/2*\l)/(\d)))}) {$v^2_{11}$};
\def ((2*\d)^2+(3*\l/2)^2)^(1/2) {((2*3.3)^2+(3*\l/2)^2)^(1/2)}
\def atan((3/2*\l)/(2*\d)) {atan((3/2*\l)/(2*3.3))}
\node[draw, rectangle] (b2v21) at ({((2*\d)^2+(3*\l/2)^2)^(1/2)*cos(150+atan((3/2*\l)/(2*\d)))} , {((2*\d)^2+(3*\l/2)^2)^(1/2)*sin(150+atan((3/2*\l)/(2*\d)))}) {$v^2_{21}$};
\def (\d^2+(\l/2)^2)^(1/2) {(3.3^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(\d)) {atan((\l/2)/(3.3))}
\node[draw, rectangle] (b2v12) at ({(\d^2+(\l/2)^2)^(1/2)*cos(150+atan((\l/2)/(\d)))} , {(\d^2+(\l/2)^2)^(1/2)*sin(150+atan((\l/2)/(\d)))}) {$v^2_{12}$};
\def ((2*\d)^2+(\l/2)^2)^(1/2) {((2*3.3)^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(2*\d)) {atan((\l/2)/(2*3.3))}
\node[draw, rectangle] (b2v22) at ({((2*\d)^2+(\l/2)^2)^(1/2)*cos(150+atan((\l/2)/(2*\d)))} , {((2*\d)^2+(\l/2)^2)^(1/2)*sin(150+atan((\l/2)/(2*\d)))}) {$v^2_{22}$};
\def ((3*\d/2)^2+(\l/2)^2)^(1/2) {((3*3.3/2)^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(3*\d/2)) {atan((\l/2)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(\l/2)^2)^(1/2)*cos(150-atan((\l/2)/(3*\d/2)))} , {((3*\d/2)^2+(\l/2)^2)^(1/2)*sin(150-atan((\l/2)/(3*\d/2)))}) circle (0.01cm);
\def ((3*\d/2)^2+(\l/4)^2)^(1/2) {((3*3.3/2)^2+(\l/4)^2)^(1/2)}
\def atan((\l/4)/(3*\d/2)) {atan((\l/4)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(\l/4)^2)^(1/2)*cos(150-atan((\l/4)/(3*\d/2)))} , {((3*\d/2)^2+(\l/4)^2)^(1/2)*sin(150-atan((\l/4)/(3*\d/2)))}) circle (0.01cm);
\def ((3*\d/2)^2+(3*\l/4)^2)^(1/2) {((3*3.3/2)^2+(3*\l/4)^2)^(1/2)}
\def atan((3*\l/4)/(3*\d/2)) {atan((3*\l/4)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(3*\l/4)^2)^(1/2)*cos(150-atan((3*\l/4)/(3*\d/2)))} , {((3*\d/2)^2+(3*\l/4)^2)^(1/2)*sin(150-atan((3*\l/4)/(3*\d/2)))}) circle (0.01cm);
\def (\d^2+(3*\l/2)^2)^(1/2) {(3.3^2+(3*\l/2)^2)^(1/2)}
\def atan((3*\l/2)/(\d)) {atan((3*\l/2)/(3.3))}
\node[draw, rectangle] (b2v1k) at ({(\d^2+(3*\l/2)^2)^(1/2)*cos(150-atan((3*\l/2)/(\d)))} , {(\d^2+(3*\l/2)^2)^(1/2)*sin(150-atan((3*\l/2)/(\d)))}) {$v^2_{1k}$};
\def ((2*\d)^2+(3*\l/2)^2)^(1/2) {((2*3.3)^2+(3*\l/2)^2)^(1/2)}
\def atan((3*\l/2)/(2*\d)) {atan((3*\l/2)/(2*3.3))}
\node[draw, rectangle] (b2v2k) at ({((2*\d)^2+(3*\l/2)^2)^(1/2)*cos(150-atan((3*\l/2)/(2*\d)))} , {((2*\d)^2+(3*\l/2)^2)^(1/2)*sin(150-atan((3*\l/2)/(2*\d)))}) {$v^2_{2k}$};
\draw [thick] (w) to (b2v11);
\draw [thick] (w) to (b2v12);
\draw [thick] (w) to (b2v1k);
\draw [thick] (b2v11) to (b2v21);
\draw [thick] (b2v12) to (b2v22);
\draw [thick] (b2v1k) to (b2v2k);
\draw [thick] (b2v21) to (b2v1);
\draw [thick] (b2v22) to (b2v1);
\draw [thick] (b2v2k) to (b2v1);
\def 270 {270}
\node[draw, rectangle] (bmv1) at ({3*3.3*cos(270)}, {3*3.3*sin(270))}) {$v_m$};
\def (\d^2+(3*\l/2)^2)^(1/2) {(3.3^2+(3*\l/2)^2)^(1/2)}
\def atan((3/2*\l)/(\d)) {atan((3/2*\l)/(3.3))}
\node[draw, rectangle] (bmv11) at ({(\d^2+(3*\l/2)^2)^(1/2)*cos(270+atan((3/2*\l)/(\d)))} , {(\d^2+(3*\l/2)^2)^(1/2)*sin(270+atan((3/2*\l)/(\d)))}) {$v^m_{11}$};
\def ((2*\d)^2+(3*\l/2)^2)^(1/2) {((2*3.3)^2+(3*\l/2)^2)^(1/2)}
\def atan((3/2*\l)/(2*\d)) {atan((3/2*\l)/(2*3.3))}
\node[draw, rectangle] (bmv21) at ({((2*\d)^2+(3*\l/2)^2)^(1/2)*cos(270+atan((3/2*\l)/(2*\d)))} , {((2*\d)^2+(3*\l/2)^2)^(1/2)*sin(270+atan((3/2*\l)/(2*\d)))}) {$v^m_{21}$};
\def (\d^2+(\l/2)^2)^(1/2) {(3.3^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(\d)) {atan((\l/2)/(3.3))}
\node[draw, rectangle] (bmv12) at ({(\d^2+(\l/2)^2)^(1/2)*cos(270+atan((\l/2)/(\d)))} , {(\d^2+(\l/2)^2)^(1/2)*sin(270+atan((\l/2)/(\d)))}) {$v^m_{12}$};
\def ((2*\d)^2+(\l/2)^2)^(1/2) {((2*3.3)^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(2*\d)) {atan((\l/2)/(2*3.3))}
\node[draw, rectangle] (bmv22) at ({((2*\d)^2+(\l/2)^2)^(1/2)*cos(270+atan((\l/2)/(2*\d)))} , {((2*\d)^2+(\l/2)^2)^(1/2)*sin(270+atan((\l/2)/(2*\d)))}) {$v^m_{22}$};
\def ((3*\d/2)^2+(\l/2)^2)^(1/2) {((3*3.3/2)^2+(\l/2)^2)^(1/2)}
\def atan((\l/2)/(3*\d/2)) {atan((\l/2)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(\l/2)^2)^(1/2)*cos(270-atan((\l/2)/(3*\d/2)))} , {((3*\d/2)^2+(\l/2)^2)^(1/2)*sin(270-atan((\l/2)/(3*\d/2)))}) circle (0.01cm);
\def ((3*\d/2)^2+(\l/4)^2)^(1/2) {((3*3.3/2)^2+(\l/4)^2)^(1/2)}
\def atan((\l/4)/(3*\d/2)) {atan((\l/4)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(\l/4)^2)^(1/2)*cos(270-atan((\l/4)/(3*\d/2)))} , {((3*\d/2)^2+(\l/4)^2)^(1/2)*sin(270-atan((\l/4)/(3*\d/2)))}) circle (0.01cm);
\def ((3*\d/2)^2+(3*\l/4)^2)^(1/2) {((3*3.3/2)^2+(3*\l/4)^2)^(1/2)}
\def atan((3*\l/4)/(3*\d/2)) {atan((3*\l/4)/(3*3.3/2))}
\draw [fill=black!100, color=black!100] ({((3*\d/2)^2+(3*\l/4)^2)^(1/2)*cos(270-atan((3*\l/4)/(3*\d/2)))} , {((3*\d/2)^2+(3*\l/4)^2)^(1/2)*sin(270-atan((3*\l/4)/(3*\d/2)))}) circle (0.01cm);
\def (\d^2+(3*\l/2)^2)^(1/2) {(3.3^2+(3*\l/2)^2)^(1/2)}
\def atan((3*\l/2)/(\d)) {atan((3*\l/2)/(3.3))}
\node[draw, rectangle] (bmv1k) at ({(\d^2+(3*\l/2)^2)^(1/2)*cos(270-atan((3*\l/2)/(\d)))} , {(\d^2+(3*\l/2)^2)^(1/2)*sin(270-atan((3*\l/2)/(\d)))}) {$v^m_{1k}$};
\def ((2*\d)^2+(3*\l/2)^2)^(1/2) {((2*3.3)^2+(3*\l/2)^2)^(1/2)}
\def atan((3*\l/2)/(2*\d)) {atan((3*\l/2)/(2*3.3))}
\node[draw, rectangle] (bmv2k) at ({((2*\d)^2+(3*\l/2)^2)^(1/2)*cos(270-atan((3*\l/2)/(2*\d)))} , {((2*\d)^2+(3*\l/2)^2)^(1/2)*sin(270-atan((3*\l/2)/(2*\d)))}) {$v^m_{2k}$};
\draw [thick] (w) to (bmv11);
\draw [thick] (w) to (bmv12);
\draw [thick] (w) to (bmv1k);
\draw [thick] (bmv11) to (bmv21);
\draw [thick] (bmv12) to (bmv22);
\draw [thick] (bmv1k) to (bmv2k);
\draw [thick] (bmv21) to (bmv1);
\draw [thick] (bmv22) to (bmv1);
\draw [thick] (bmv2k) to (bmv1);
\def ((3*\d/2) {((3*3.3/2)}
\def (\alfabt+\alfabs)/2 {(270+150)/2}
\draw [fill=black!100, color=black!100] ({((3*\d/2)*cos((\alfabt+\alfabs)/2)} , {((3*\d/2)*sin((\alfabt+\alfabs)/2)}) circle (0.01cm);
\def ((\alfabt+\alfabs)/2*3+(\alfabs+atan((3*\l/2)/\d)))/4 {((270+150)/2*3+(150+atan((3*\l/2)/3.3)))/4}
\def ((\alfabt+\alfabs)/2*3+(\alfabt-atan((3*\l/2)/\d)))/4 {((270+150)/2*3+(270-atan((3*\l/2)/3.3)))/4}
\draw [fill=black!100, color=black!100] ({((3*\d/2)*cos(((\alfabt+\alfabs)/2*3+(\alfabs+atan((3*\l/2)/\d)))/4)} , {((3*\d/2)*sin(((\alfabt+\alfabs)/2*3+(\alfabs+atan((3*\l/2)/\d)))/4)}) circle (0.01cm);
\draw [fill=black!100, color=black!100] ({((3*\d/2)*cos(((\alfabt+\alfabs)/2*3+(\alfabt-atan((3*\l/2)/\d)))/4)} , {((3*\d/2)*sin(((\alfabt+\alfabs)/2*3+(\alfabt-atan((3*\l/2)/\d)))/4)}) circle (0.01cm);
\def 0 {0}
\def -\d/3.5 {-3.3/3.5}
\def 1.5*\d-\xstep {1.5*3.3-0}
\def -1*\d-\ystep {-1*3.3--\d/3.5}
\node[draw, rectangle] (u1) at ({1.5*\d-\xstep+0}, {-1*\d-\ystep+-\d/3.5}) {$u_1$};
\node[draw, rectangle] (u2) at ({1.5*\d-\xstep+2*0}, {-1*\d-\ystep+2*-\d/3.5}) {$u_2$};
\draw [fill=black!100, color=black!100] ({1.5*\d-\xstep+3*0},{-1*\d-\ystep+3*-\d/3.5}) circle (0.01cm);
\draw [fill=black!100, color=black!100] ({1.5*\d-\xstep+2.8*0},{-1*\d-\ystep+2.8*-\d/3.5}) circle (0.01cm);
\draw [fill=black!100, color=black!100] ({1.5*\d-\xstep+3.2*0},{-1*\d-\ystep+3.2*-\d/3.5}) circle (0.01cm);
\node[draw, rectangle] (utl) at ({1.5*\d-\xstep+4*0}, {-1*\d-\ystep+4*-\d/3.5}) {$u_{t-1}$};
\node[draw, rectangle] (ut) at ({1.5*\d-\xstep+5*0}, {-1*\d-\ystep+5*-\d/3.5}) {$u_{t}$};
\node[draw, rectangle, dashed] (u't) at ({1.5*\d-\xstep+5*0-1}, {-1*\d-\ystep+5*-\d/3.5}) {$u'_{t}$};
\node[draw, rectangle] (utr) at ({1.5*\d-\xstep+6*0}, {-1*\d-\ystep+6*-\d/3.5}) {$u_{t+1}$};
\draw [fill=black!100, color=black!100] ({1.5*\d-\xstep+7*0},{-1*\d-\ystep+7*-\d/3.5}) circle (0.01cm);
\draw [fill=black!100, color=black!100] ({1.5*\d-\xstep+6.8*0},{-1*\d-\ystep+6.8*-\d/3.5}) circle (0.01cm);
\draw [fill=black!100, color=black!100] ({1.5*\d-\xstep+7.2*0},{-1*\d-\ystep+7.2*-\d/3.5}) circle (0.01cm);
\node[draw, rectangle] (ud) at ({1.5*\d-\xstep+8*0}, {-1*\d-\ystep+8*-\d/3.5}) {$u_{d}$};
\draw [thick] (w) to (u1);
\draw [thick] (u1) to (u2);
\draw [thick] (utl) to (ut);
\draw [thick] (ut) to (utr);
\draw [thick] (ut) to (u't);
\draw [thick] (u2) to ({1.5*\d-\xstep+2.6*0},{-1*\d-\ystep+2.6*-\d/3.5});
\draw [thick] (utl) to ({1.5*\d-\xstep+3.4*0},{-1*\d-\ystep+3.4*-\d/3.5});
\draw [thick] (utr) to ({1.5*\d-\xstep+6.6*0},{-1*\d-\ystep+6.6*-\d/3.5});
\draw [thick] (ud) to ({1.5*\d-\xstep+7.4*0},{-1*\d-\ystep+7.4*-\d/3.5});
\draw [thick, loosely dashed] (b1v11) to (b2v11);
\draw [thick, loosely dashed] (b1v11) to (b2v12);
\draw [thick, loosely dashed] (b1v11) to (b2v1k);
\draw [thick, loosely dashed] (b1v12) to (b2v11);
\draw [thick, loosely dashed] (b1v12) to (b2v12);
\draw [thick, loosely dashed] (b1v12) to (b2v1k);
\draw [thick, loosely dashed] (b1v1k) to (b2v11);
\draw [thick, loosely dashed] (b1v1k) to (b2v12);
\draw [thick, loosely dashed] (b1v1k) to (b2v1k);
\draw [thick, loosely dashed] (b1v11) to (b2v11);
\draw [thick, loosely dashed] (b1v11) to (b2v12);
\draw [thick, loosely dashed] (b1v11) to (b2v1k);
\draw [thick, loosely dashed] (b1v12) to (b2v11);
\draw [thick, loosely dashed] (b1v12) to (b2v12);
\draw [thick, loosely dashed] (b1v12) to (b2v1k);
\draw [thick, loosely dashed] (b1v1k) to (b2v11);
\draw [thick, loosely dashed] (b1v1k) to (b2v12);
\draw [thick, loosely dashed] (b1v1k) to (b2v1k);
\draw [thick, loosely dashed] (bmv11) to (b1v11);
\draw [thick, loosely dashed] (bmv11) to (b1v12);
\draw [thick, loosely dashed] (bmv11) to (b1v1k);
\draw [thick, loosely dashed] (bmv11) to (b2v11);
\draw [thick, loosely dashed] (bmv11) to (b2v12);
\draw [thick, loosely dashed] (bmv11) to (b2v1k);
\draw [thick, loosely dashed] (bmv12) to (b1v11);
\draw [thick, loosely dashed] (bmv12) to (b1v12);
\draw [thick, loosely dashed] (bmv12) to (b1v1k);
\draw [thick, loosely dashed] (bmv12) to (b2v11);
\draw [thick, loosely dashed] (bmv12) to (b2v12);
\draw [thick, loosely dashed] (bmv12) to (b2v1k);
\draw [thick, loosely dashed] (bmv1k) to (b1v11);
\draw [thick, loosely dashed] (bmv1k) to (b1v12);
\draw [thick, loosely dashed] (bmv1k) to (b1v1k);
\draw [thick, loosely dashed] (bmv1k) to (b2v11);
\draw [thick, loosely dashed] (bmv1k) to (b2v12);
\draw [thick, loosely dashed] (bmv1k) to (b2v1k);
\node[below,font=\large\bfseries] at (current bounding box.south) {Figure 2. Graph $L(k,m)$};
\end{tikzpicture}
\end{center}
\begin{theorem}\label{T:lily}
For $k,m\in \mathbb{N}$, $m\geq 7$ and $k\geq \frac{m-3}{m-6}$, the proportion of good vertices of $L(k,m)$ equals $$\frac{km}{2km+m+2+\Big\lfloor{\frac{\sqrt{8km-48k-8m+33}+1}{2}}\Big\rfloor}.$$
\end{theorem}
\begin{proof}
Let $G=(V,E)$ the a graph obtained after steps $1-3$ of Construction \ref{C:lily}. Consider a vertex $v=v^1_{11}$ of the graph $G=(V,E)$. Take an arbitrary graph $G'=(V',E')$ such that $V\cap V'=\{u_0\}$. Then, let us define a graph $\tilde{G}=(\tilde{V},\tilde{E})=(V\cup V', E\cup E')$. We do not specify sets $V'$ and $E'$ by now in order to see what conditions on these sets one should impose.
Our final goal is to find a graph $\tilde{G}$ such that $\Delta_{v}(\tilde{G})=0$. According to Proposition \ref{Prop:1}, we know that $\Delta_v(\tilde{G})=t_{\tilde{G}}(v)+\frac{1}{2}\sum_{u,w \in \tilde{V}\setminus \{v\}}{\delta^v_{\tilde{G}}(u,w)}$. Therefore we proceed by determining $t_{\tilde{G}}(v)$ and $\delta^{v}_{\tilde{G}}(u,w)$ for all vertices $u,w\in \tilde{V}\setminus \{v\}$.
By direct calculations we have that $$t_{\tilde{G}}(v)=(1+2)+2(k-1)+3(k-1)+(m-1)(k+2k+3)+t_{G'}(u_0)+|V'|.$$ Therefore, $$t_{\tilde{G}}(v)=3km+2k+3m-5+t_{G'}(u_0)+|V'|.$$ It is easy to see from the construction of the graph, that $\delta^{v}_{\tilde{G}}(u,w)=0$ whenever $u\neq v^1_{21}$ and $w\neq v^1_{21}$. All nonzero values of $\delta^{v}_{\tilde{G}}(u,w)$ are presented in the Table \ref{Tab:lily}.
\begin{table}[h]
\centering
\begin{tabular}{ |c|c|c|c|c|}
\hline
Vertices & $z$ & $v^i_{1j}$ & $v^i_{2j}$ & $v^i$ \\
\hline
$v^1_{21}$ & $-2$ & $-2$ & $-2$ & $-2$ \\
\hline
\end{tabular}
\caption{Values of $\delta^{v}_{\tilde{G}}(u,w)$, $z\in V'$, $i\in \{2,3,\dots, m\}$, $j\in \{1,2,\dots, k\}$.}
{\label{Tab:lily}}
\end{table}
Consequently, $$\frac{1}{2}\sum_{u,w \in V\setminus \{v\}}{\delta^v_{\tilde{G}}(u,w)}=-2\Big(|V'|+k(m-1)+k(m-1)+(m-1)\Big)=-2|V'|-4km+4k-2m+2.$$
By Proposition \ref{Prop:1} we have that $\Delta_{v}(\tilde{G})=t_{G'}(u_0)-|V'|-km+6k+m-3$. Evidently, $\Delta_{v}(\tilde{G})$ depends on $t_{G'}(u_0)-|V'|$, but not on the structure of a graph $G'$. Therefore, we are free to take arbitrary graph $G'$ such that the following equation holds
\begin{equation}\label{eq:lily}
t_{G'}(u_0)-|V'|=km-6k-m+3.
\end{equation}
According to steps $4$ and $5$ of Construction \ref{C:lily}, we take a path $(u_0, u_1, u_2, \dots, u_d)$, with one additional vertex $u'_t$ and edge $(u_t,u'_t)$ and finally construct the graph $L(k,m)$, where $$d=\Big\lfloor{\frac{\sqrt{8km-48k-8m+33}+1}{2}}\Big\rfloor ,\, t=km-6k-m+3-\frac{(d+1)(d-2)}{2}.$$
In other words, $d$ is a floor of solution to the equation (\ref{eq:lily}) for the path $(u_0, u_1, u_2, \dots, u_d)$: $$(1+2+\dots+d)-(d+1)\leq km-6k-m+3 < (1+2+\dots+d+(d+1))-(d+2).$$ By the assumption of the theorem we know that $m\geq 7$ and $k\geq \frac{m-3}{m-6}$. Hence, $t$ is a non-negative integer and $t<d+1$, so $t$ is defined correctly. Note that if $t=0$ then one does not need to add a vertex $u'_t$ (equation (\ref{eq:lily}) holds). However, in this case we add $u'_0$ as a neighbour of $u_0$ to obtain the resulting graph on the same number of vertices.
Finally, we conclude that $\Delta_{v_{11}}(L(k,m))=0$.
Based on symmetry of $L(k,m)$, we conclude that for all $z\in \{v^i_{1j}|i\in \{1,2\dots m\}, j\in \{1,2,\dots k \}\}$ the equation $\Delta_{z}(G)=0$ holds.
One may notice that $v_{11}$ and other neighbours of $u_0$ (except $u_1$) are good because its deletion makes a lot of shortest paths longer. It is easy to see from the construction, that other vertices of the graph do not have this property, so either a corresponding transmission is always greater than possible changes of lengths of the shortest paths or the graph with deleted vertex is disconnected. Thus, other vertices can not be good and the theorem is proved.
\end{proof}
From the theorem we directly deduce the following
\begin{corollary}\label{Cor:lily}
For a fixed $m\in \mathbb{N}$, $m\geq 7$, the series of graphs $L(k,m)$, $k\in \mathbb{N}$ has a proportion of good vertices tending to $\frac{1}{2}$ as $k$ tends to infinity.
\end{corollary}
As one could notice from the proof of the last theorem, values of $t$ and $d$ in steps $4$ and $5$ respectively are chosen in order to have $W(L(k,m))-W(L(k,m)-v)=0$. In a similar way for any fixed $z\in \mathbb{Z}$, one can choose $t'$ and $d'$ (denote the resulting graph $W(L'(k,m,z))$) in order to obtain the following equation
$$W(L'(k,m,z))-W(L'(k,m,z)-v)=z.$$
Since the value $-km+6k+m-3$ for fixed $m\geq 7$ is a decreasing function on $k$ this choice is always possible for $k$ big enough. These arguments allow us to prove the following
\begin{corollary}\label{Cor:lilyz}
For fixed $m\in \mathbb{N}$, $m\geq 7$, $z\in \mathbb{Z}$, the series of graphs $L'(k,m,z)$, $k\in \mathbb{N}$ has a proportion of vertices $v$, such that $W(L'(k,m,z))-W(L'(k,m,z)-v)=z$, tending to $\frac{1}{2}$ as $k$ tends to infinity.
\end{corollary}
Construction \ref{C:lily} is conceptually close to the one from \cite{BJM2019}. Actually, steps $1$, $2$ and $4$ are similar. The authors also used a number of {\it blocks} (cycles of length at least $7$) having one common vertex and a tree of special structure outgoing from this vertex. Authors were solving a different problem, so they were not interested in analysing and minimizing the size of this tree depending on the number of blocks.
In fact, if one uses the cycle $C_7$ as a block and a tree of the same structure as it is used in Construction \ref{C:lily} then in every block of the graph there are two vertices that would be good and the whole construction would give a proportion tending to $\frac{1}{3}$ as the number of blocks tends to infinity.
Thus, the better choice of a structure of a block and step $3$ in Construction \ref{C:lily} play a crucial role in Theorem \ref{T:lily}. After numerous attempts to improve presented constructions in order to increase $\alpha$, we tend to think that $\frac{1}{2}$ is the best possible value for constructions of this type.
\section{Conclusion}\label{S:Conclusion}
In this paper, we consider a problem of constructing graphs $G$ with the given proportion of good vertices, i.e. vertices $v$ such that the following equality takes place $$W(G)=W(G-v).$$
As the main result we built a series of graphs with a proportion tending to $\frac{1}{2}$ with the number of vertices tending to infinity.
Beside that, we present one intriguing example of a graph with the proportion of good vertices $\frac{2}{3}$.
\begin{example}
Consider a graph $G$ obtained by adding to the cycle $(v_1,v_2, \dots , v_{12})$ of length $12$ four additional edges: $(v_1,v_3)$, $(v_4,v_6)$, $(v_7,v_9)$, $(v_{10},v_{12})$. This graph on $12$ vertices has $8$ good vertices, namely those of degree $3$.
\end{example}
\begin{center}
\begin{tikzpicture}[scale=0.75]
\def 12 {12}
\def 3cm {3cm}
\foreach [count=\j] \s in {1,...,12}
{
\node[draw, circle] (\j) at ({360/12 * (\s +1)}:3cm) {};
}
\foreach [remember=\i as \j (initially 1)] \i in {2,...,12} {
\draw [thick] (\j) -- (\i);
}
\draw [thick] (1) -- (12);
\draw [thick] (1) .. controls +(down:1cm) and +(down:1cm) .. (3);
\draw [thick] (7) .. controls +(up:1cm) and +(up:1cm) .. (9);
\draw [thick] (4) .. controls +(right:1cm) and +(right:1cm) .. (6);
\draw [thick] (10) .. controls +(left:1cm) and +(left:1cm) .. (12);
\node[below,font=\large\bfseries] at (current bounding box.south) {Figure 3. Graph with a proportion $\frac{2}{3}$.};
\end{tikzpicture}
\end{center}
Despite the fact that we have not generalized this example for a bigger number of vertices, we expect that there exist an infinite series of graphs with a proportion $\alpha>\frac{1}{2}$, or perhaps even $\alpha$ tending to $1$.
Corollary \ref{Cor:lilyz} shows that instead of solving Problems \ref{Prob:1} and \ref{Prob:2} it is reasonable to consider the following more general problem statements.
\begin{problem}\label{Prob:3}
For a fixed $z\in \mathbb{Z}$, find all such graphs $G$ that the equality $W(G) - W(G - v) = z$ holds for all their vertices $v$.
\end{problem}
\begin{problem}\label{Prob:4}
For a fixed $z\in \mathbb{Z}$ and $\alpha\in (0,1]$, construct an infinite series $S$ of graphs such that for all $G=(V,E)$ from $S$ the following inequality takes place: $$ \frac{|\{v\in V|\Delta_v(G)=z\}|}{|V|} \geq \alpha.$$
\end{problem}
If there are arguments that allow to solve Problems \ref{Prob:1} and \ref{Prob:2} then they would probably work for these ones too.
\section{Acknowledgements}
The main part on this project was done during the research workshop "Open problems in Combinatorics and Geometry II", held in Adygea in September and October 2020.
The results from Section \ref{S:Bunch} are supported by the Ministry of Science and Higher Education of the Russian Federation in the framework of MegaGrant no 075-15-2019-1926. The work from Section \ref{S:Lily} was carried out within the framework of the state contract of the Sobolev Institute of Mathematics (project no. 0314-2019-0016).
Authors are grateful to Andrey Dobrynin for interesting discussions on the theme of this paper.
| cbf182e2e39b4f1e98be62b5028687d22c4c726e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{section:intro}
Let $X$ be a smooth projective genus $g$ curve over a field $k$ of characteristic $p$. The Torelli map associates $X$ with its Jacobian $J_X$, a principally polarized abelian variety of dimension $g$. The map embeds the moduli space $\mathcal{M}_g$ of curves of genus $g$ into $\mathcal{A}_g$, the moduli space of principally polarized abelian varieties of dimension $g$ over $k$. In consequence, it allows us to study the stratification of $\mathcal{M}_g$ by looking at the group scheme structure of $J_X[p]$, the $p$-torsion part of the Jacobian. This is called the Ekedahl--Oort stratification.
For $g=2,3$, the Torelli locus is open and dense in $\mathcal{A}_g$. For $p\ge 3$ and $g\leq 3$ this can be used to show that all Ekedahl--Oort types occur for the Jacobians of smooth curves $X/\overline{\mathbb{F}}_p$ (\cite{Pries_sh}). The same is not known for $g\geq 4$.
Motivated by this and other similar open questions related to the $p$-torsion part of the Jacobian, we study smooth irreducible curves with $g=4$. We focus on the non-hyperelliptic kind. In particular, we look at the $a$-number and the $p$-rank, which are two invariants of $J_X[p]$.
In order to obtain a database of smooth, irreducible, genus $4$ non-hyperelliptic curves, we restrict our analysis to what we define as curves in \textit{standard form}. Recall that if $X$ is a curve with the above properties, then it has a model given by the zero locus of a quadratic and a cubic homogeneous polynomials in $k[x,y,z,w]$. Kudo and Harashita show in \cite{Kudo&Harashita} that under some assumptions, the defining equations can be simplified to reduce the number of cases. The curves given by these simplified equations are \textit{curves in standard form}.
We gather a statistical sample of curves in standard form defined over $\mathbb{F}_p$ for $p\in \{3,5,7,11\}$. For each of them we find the Hasse--Witt matrix $H$ and use it to compute the $a$-number and $p$-rank: the $a$-number is $g-\text{rank}(H)$ and the $p$-rank is $f=\text{rank} \left(HH^{(p)}\cdots H^{\left(p^{g-1}\right)}\right)$. As one should expect, the majority of curves appear in the sample are ordinary, and the percentages decrease as the $a$-number increases (or similarly, as the $p$-rank decreases).
We also explore in this paper the concept of Cartier point. We say that $P\in X(\overline{k})$ is a Cartier point if the hyperplane of regular differentials of $X$ vanishing at $P$ is stable under the Cartier operator. Baker introduces the definition in \cite{Baker} and remarks that they are related to the $p$-torsion points of the Jacobian.
If $X$ has $a$-number $0\leq a<g$ then there is an upper bound on the number of Cartier points of $X$ given by Baker \cite{Baker}. When $a\ne 0$, we classify Cartier points in Type 1 and Type 2 (see Definition~\ref{def:T1andT2}). The maximum number of Type 1 points depends on the $a$-number and the maximum number of Type 2 points depends on the $p$-rank. We are interested in determining the conditions under which these bounds are attained when $X$ is non-ordinary. Therefore, we develop algorithms to find all of the Cartier points on curves in standard form and apply them to our database.
The Cartier points are particularly interesting when $a=g-1$, because we can assign multiplicity to each of them. This is why we later focus on curves with $a=3$. We find all of the curves in standard form with $a=3$, defined over $\mathbb{F}_p$ for $p=3,5$ and a subset of them over $\mathbb{F}_7$. We explore the possible degrees and multiplicity distributions of these points.
Here are some of the most relevant conclusions from our work, concerning non-hyperelliptic genus $4$ curves in standard form:
\begin{enumerate}
\item In our smooth sample, the are no curves with $(a,f)=(1,0)$ over $\mathbb{F}_p$ for $p\in \{3,5,7,11\}$. (Corollary~\ref{cor:no_a1_f0_curves}).
\item There are, up to $\mathbb{F}_3$-isomorphism, exactly 27 curves with $a$-number $3$ over $\mathbb{F}_3$ in standard form. All of them have $p$-rank $1$. (Corollaries~\ref{cor:exhaustive_search_p3} and~\ref{cor:no_prank0_p3_curves}).
\item There are, up to $\mathbb{F}_5$-isomorphism, exactly $134$ curves with $a=3$ over $\mathbb{F}_5$. (Corollary~\ref{cor:exhaustive_search_p5}).
\item In our smooth sample, no curve with $a$-number $2$ reaches the bound of $6$ Type $1$ Cartier points. Moreover, the maximum number of Type $1$ points attained on curves with $a$-number $2$ is $3$ for $p\in \{5,7,11\}$ and $2$ for $p=3$. (Corollary~\ref{cor:no_a2_curves_reach_UB_of_T1}).
\item In our smooth sample no curve with $p$-rank $2$ or $3$ reaches the bound of $6$ Type $2$ Cartier points. The maximum number of points that occurs is $3$ and $4$, respectively. (Corollary~\ref{cor:no_f2_curves_reach_UB_of_T2}).
\item When $a=3$, the bound on Type 1 points is sharp for $p\in \{5,7,11\}$ and the total bound for both Types is sharp for $p=7$. (Corollaries~\ref{cor:summary_T1_bounds} and~\ref{cor:bounds_reached_p7}).
\item There are no curves in standard form over $\mathbb{F}_3$ with $a=3$ that attain either of the upper bounds for Cartier points. (Lemma~\ref{lem:T1_points_in_a3_p3_curves}).
\end{enumerate}
\section{Acknowledgements}
This paper is part of my doctoral dissertation.
The work presented here would not have been possible without the support of my advisor Rachel Pries. I would like to thank Dr. Pries for suggesting this topic and sharing her knowledge throughout my time in graduate school.
I was financially supported by University of Costa Rica while I worked on this project.
\section{Preliminaries}\label{sec:preliminaries}
This section includes the background information related to the Cartier operator, Hasse--Witt matrix, $p$-rank and $a$-number of a curve. Unless otherwise stated, $p$ will be an odd prime number and $k$ a perfect field of characteristic $p$.
\subsection{The $p$-rank }\label{subsec:anumber_prank}
Let $X$ be a smooth irreducible genus $g$ curve over $k$.
Denote by $J_X$ the Jacobian variety associated to $X$. Then $J_X$ is an abelian variety of dimension $g$ isomorphic to $\text{Pic}^0_{X/k}$, equipped with a principal polarization.
The $p$-torsion part of the Jacobian, denoted by $J_X[p]$, is a group scheme and here we will study two invariants associated to it. The first one is the $p$-rank, defined as the integer $f$ such that $\# J_X[p](k)=p^f$. The $p$-torsion subgroup of an abelian variety of dimension $g$ has order at most $g$, hence $0\leq f\leq g$.
\subsection{Cartier and Frobenius operators and the $a$-number}\label{subsec:CartierFrob}
Suppose that $x$ is a separating variable of $k(X)/k$, then every $t \in k(X)$ can be written as
\begin{align}\label{eq:f_inK(X)}
t=t_0^p+t_1^px+\ldots +t_{p-1}^px^{p-1},
\end{align}
with $t_i \in k(X)$.
\begin{definition}\label{def:cartier_operator}
The Cartier operator $\mathcal{C}$ is defined on $\Omega_X^1$ for $t$ as above by
\begin{align}
\mathcal{C}(tdx)=t_{p-1}dx.
\end{align}
\end{definition}
The Cartier operator is $1/p$-linear, meaning that $\mathcal{C}(a^p\omega_1+b^p\omega_2)=a\mathcal{C}(\omega_1)+b\mathcal{C}(\omega_2)$, for $a,b$ in $k(X)$ and $\omega_1, \omega_2 \in \Omega_X^1$. It induces a well defined map $\mathcal{C}$ on the $k$-vector space of regular differentials $H^0(X,\Omega_X^1)$.
%
%
\begin{definition}\label{def:CartierMatrix}
If $B=\left\lbrace\omega_1, \ldots, \omega_g\right\rbrace$ is a $k$-basis for $H^0(X,\Omega_X^1)$ and $\mathcal{C}(\omega_j)=\sum_{i=1}^g{c_{ij}\omega_i}$ then the Cartier--Manin matrix of $X$ with respect to this basis is the matrix $(c_{ij}^p)_{ij}$.
\end{definition}
\begin{definition}\label{def:frobenius_in H1}
The absolute Frobenius of $X$ is the morphism $\mathcal{F}:X\to X$ given by the identity on the underlaying topological space and $t\mapsto t^p$ on $\mathcal{O}_X$. Let $\mathcal{F}_X$ be the induced endomorphism in $H^1(X,\mathcal{O}_X)$. We call $\mathcal{F}_X$ the Frobenius endomorphism.
\end{definition}
The Frobenius endomorphism is $p$-linear, that is $\mathcal{F}_X(a\xi)=a^p\mathcal{F}_X(\xi)$ for all $a\in k$ and all $\xi \in H^1(X,\mathcal{O}_X)$.
\begin{definition}\label{def:HW_matrix}
Let $B'=\{\xi_1, \ldots, \xi_g\}$ be a $k$-basis of $H^1(X,\mathcal{O}_X)$ and $\mathcal{F}_X(\xi_j)=\sum_{i=1}^{g}a_{ij}\xi_i$ for some $a_{ij}\in k$. Then the Hasse--Witt matrix of $X$ with respect to $B'$ is the matrix $(a_{ij})_{ij}$.
\end{definition}
The space $H^1(X,\mathcal{O}_X)$ is the dual of $H^0(X,\Omega_X^1)$ and there is a perfect paring $<,>$ on $H^1(X,\mathcal{O}_X)\times H^0(X,\Omega_X^1)$ such that $<\mathcal{F}_X\xi,\omega>=<\xi,\mathcal{C}\omega>^p$.
When $B$ and $B'$ are dual basis,the Cartier--Manin matrix $M$ with respect to $B$ is the transpose of the Hasse--Witt matrix $H$ with respect to $B'$.
By Serre \cite{serre1972}, the $p$-rank $f$ is the stable rank of the Frobenius and since this operator is $p$-linear, it implies that
\begin{equation}\label{eq:p-rank-stablerank}
f=\text{rank} \left(HH^{(p)}\cdots H^{\left(p^{g-1}\right)}\right),
\end{equation}
where $H^{(i)}$ is the matrix obtained by raising every entry of $H$ to the $i$-th power.
The second invariant of $J_X[p]$ that we study here is the $a$-number, defined as the rank of the Cartier operator (see Oort \cite{oort_li}), i.e.,
\begin{equation}\label{eq:a_number_rank}
a=g-\text{rank}(H)=g-\text{rank}(M).
\end{equation}
It is known that $0\leq a+f \leq g$. Generically $f=g$, in which case $X$ is said to be \textit{ordinary}.
The other extreme case is when $X$ is \textit{superspecial} and it occurs when $a=g$ or equivalently, when the Cartier operator is identically $0$ on $H^0(X,\Omega_X^1)$.
In Section~\ref{subsec:HWmatrix_genus4} we follow the work from \cite{Kudo&Harashita} to describe how to explicitly find the Hasse--Witt matrix, when $X$ is the complete intersection of two homogeneous polynomials and use it to compute the $p$-rank and $a$-number.
\subsection{Previous results}\label{subsec:previous_results}
In this section we review some of the main and more recent results with respect to genus $g$ curves of positive characteristic and possible values for the $a$-number and $p$-rank that occur.
\begin{thm}[Ekedahl \cite{Ekedahl:Supersingular},Theorem 1.1]\label{thm:Ekedahl}
Let $X$ be a smooth curve of genus $g$ over an algebraically closed field $k$ of characteristic $p>0$. If $X$ is superspecial then
\begin{enumerate}
\item $g\leq \frac{1}{2}(p^2-p)$ and
\item $g\leq \frac{1}{2}(p-1) $ if $X$ is hyperelliptic and $(p,g)\neq (2,1).$
\end{enumerate}
\end{thm}
Baker gives in \cite{Baker} an alternative proof for Theorem~\ref{thm:Ekedahl}, based on the existence of Cartier points (see Section~\ref{sec:CartierPoints} for definition). Let $m$ be the rank of the Cartier operator. Re (\cite{RE}, Theorem 3.1 and Proposition 3.1) provides a generalization of this result to any value of $m$. The author proves that in fact $ g\leq (m+1)p\frac{(p-1)}{2}+pm$ and if $X$ is also hyperelliptic then $g<\frac{p+1}{2}+mp$.
Zhou (\cite{Zhou}, Theorem 1.1) gives a strengthening of this result for the case when $m=1$ that gives a bound for $g$ of $p+\frac{p(p-1)}{2} $. Moreover, Frei (\cite{Frei18}, Theorem 3.1]) proved that if $g\geq p$ where $p$ is an odd prime, then there are no smooth hyperelliptic curves of genus $g$ defined over a field of characteristic $p$ with $a$-number equal to $g - 1$.
There are examples, however, of curves with $a$-number $g-1$ that are non-hyperelliptic. Also Zhou finds in \cite{Zhou_anumber} a family of Artin--Schreier curves with these properties.
More generally, Pries \cite{PriesPtorsion09} proves the existence of smooth curves with $a$-number $1,2$ and $3$, under certain conditions.
\section{Genus $4$ non-hyperelliptic curves}\label{sec:genus4}
There are currently many open questions concerning the existence of curves with certain $p$-ranks and $a$-numbers, given a fixed genus $g$. For instance, there exist curves of genus $2$ and 3 with any possible $p$-rank and $a$-number over fields of characteristic $p$, with the exception of superspecial curves of genus $2$ when $p=2$ and superspecial curves of genus $3$ when $p=2,3$. For $g\geq 4$, however, it is not known if this happens. For example, consider Question 3.6 in \cite{PriesCRN}: \textit{For all $p$, does there exist a smooth curve of genus $4$ with $p$-rank $0$ and $a$-number at least $2$?}
One can find in the literature partial answers to the last and similar questions. For example, if $p=3$ then by Ekedahl's Theorem (\ref{thm:Ekedahl}), there is no curve of genus $4$, with $a$-number 4. In \cite{Zhou_EOg4}, Zhou (building on work from \cite{Frei18}, \cite{GlassPries} and \cite{PriesPtorsion09}) shows, by studying a family of Artin--Schreier curves that in characteristic 3, there are genus $4$ curves with $a$-number $a$ and $p$-rank $f$ for all $a\leq 2$ and $ f\leq a$ and for $(a,f)=(3,1)$.
In fact, the author computes the Ekedahl-Oort types to show that the corresponding locus of $\mathcal{M}_g$ is non empty of codimension at most 6. In Section~\ref{sec:CP_on_GC_p=3} we provide additional examples of genus $4$ curves with $a$-number $3$ and $p$-rank $1$ over $\mathbb{F}_{3}$. We know by Frei ~\cite{Frei18} that there are no $a$-number $3$ and $p$-rank $0$ genus $4$ hyperelliptic curves, so one can ask whether it is possible to have a non-hyperelliptic genus $4$ with those invariants.
Kudo and Harashita \cite{Kudo&Harashita} also studied genus $4$ curves, they prove two results related to non-hyperelliptic superspecial curves of genus $4$. In particular they show that there are no superspecial genus $4$ curves in characteristic $7$, and that over $\mathbb{F}_{25}$ they are all isomorphic to
$$2yw+z^2=0, x^3+a_1y^3+a_2w^3+a_3zw^2=0, $$
in $\mathbb{P}_{\mathbb{F}_{25}}^3$, where $a_1,a_2\in \mathbb{F}^\times_{25}$ and $a_3\in \mathbb{F}_{25}.$
\subsection{Defining equations of genus $4$ non-hyperelliptic curves}\label{subsec:equations_genus4}
If $X$ is a genus $4$, smooth, irreducible and non-hyperelliptic curve, then the canonical map embeds $X$ into $\mathbb{P}^3_k$ as the intersection of the zero loci of a quadratic and a cubic homogeneous polynomial in four variables (see \cite{Hartshorne77}), that we detail below.
We restrict our computations to what we will define as genus $4$ curves in \textit{standard form}. These are based on equations given by Kudo and Harashita \cite{Kudo&Harashita}.
\subsubsection{Quadratic forms and reduction of cubics}
As explained in \cite{Kudo&Harashita}, any irreducible quadratic form in $k[x,y,z,w]$ is equivalent to one of $ F_1=2xw+2yz$, $F_2=2xw+y^2-\epsilon z^2$ or $F_d=2yw+ z^2$ with some $\epsilon\notin (k^\times)^2$. Therefore we can assume that $X$ has a model given by $V(F,G)$, with $F $ being one of $F_1$, $F_2$ or $F_d$, and $G$ a homogeneous polynomial of degree $3$. The cubic $G$ can be reduced by changes of variables, induced by the action of the orthogonal similitude groups associated to the quadratic forms. This is done in detail in Section 4 of \cite{Kudo&Harashita}. The simplified equations provide the following definition.
\begin{definition}\label{def:standardform}
Let $ F_1=2xw+2yz$, $F_2=2xw+y^2-\epsilon z^2$ or $F_d=2yw+ z^2$ with $\epsilon\notin (k^\times)^2$. We say that a curve $X$ of genus $4$ over $k$ is in \textit{standard form} if it is non-hyperelliptic, irreducible, smooth and $X=V(F,G)$ with \\
\textbf{(Case D)} $F=F_d$ and
\begin{align*}
G& =a_0x^3+(a_1y^2+a_2z^2+a_3w^2+a_4yz+a_5zw)x+a_6y^3+a_7z^3+a_8w^3+a_9yz^2\\
&+ b_1z^2w+b_2zw^2,
\end{align*}\label{eq:FandG2_p=5}
for $a_i \in k$ and $a_0,a_6\in k^\times$, with $b_1, b_2 \in \{0,1\}$ and the leading coefficient of $r=a_1y^2+a_2z^2+a_3w^2+a_4yz+a_5zw$ is $1$ or $r=0$; or\\
\textbf{(Case N1i)} $F=F_1$ and
\begin{align*}
G&=(a_1y +a_2z)x^2 +a_3yzx+y^3 +a_4z^3 +b_1y^2z +a_5yz^2 \\ &+(a_6y^2 +a_7yz +b_2z^2)w+(a_8y +a_9z)w^2 +a_{10}w^3,
\end{align*}
for $a_i \in k $ with $a_1\neq 0$, $a_2\neq 0$ and for $b_1\in \{0\} \cup k^{\times}/(k^{\times})^2$ and $b_2 \in \{0,1\}$; or \\
\textbf{(Case N1ii)} $F=F_1$ and
\begin{align*}
G=& (a_1y +a_2z)x^2 +a_3yzx+b_1y^2z +b_2yz^2 +(a_4y^2 +a_5yz +b_3z^2)w+(a_6y +\\ &a_7z)w^2 +a_8w^3,
\end{align*}
for $a_i \in k $ with $a_1a_2\neq 0$ and for $b_1,b_3 \in \{0,1\}$ and $b_2 \in \{0\} \cup k^{\times}/(k^{\times})^2$; or\\
\textbf{(Case N2)} $F=F_2$ and
\begin{align*}
\begin{split}
G=&(a_1y + a_2z)x^2 + a_3(y^2 - \epsilon z^2)x + b_1y(y^2 -\epsilon z^2) + a_4y(y^2 + 3\epsilon z^2) + a_5z(3y^2 \\&+\epsilon z^2)+(a_6y^2 + a_7yz + b_2z^2)w + (a_8y + a_9z)w^2 + a_{10}w^3,
\end{split}
\end{align*}
for $a_i \in k$, with $(a_1,a_2)\neq (0,0)$ and $b_1,b_2 \in \{0,1\}$ and $\epsilon$ a non-trivial fixed representative of $k^\times/(k^\times)^2$,
\end{definition}
We remark that every genus $4$ non-hyperelliptic can be written as\linebreak $X=V(F,G)$, where $F$ is one of $F_1, F_2, F_d$. Moreover, $G$ can be simplified to standard form if the curve has least 37 points. If the $F=F_d$ then it is enough that $\#k>5$ (\cite{Kudo&Harashita} Lemmas 4.3.1, 4.4.1 and 4.5.1).
%
\subsection{Hasse--Witt Matrix of genus $4$ non-hyperelliptic curves}\label{subsec:HWmatrix_genus4}
Let $X=V(F,G)$ be the complete intersection on $\mathbb{P}^3_k$ defined by homogeneous polynomials $F$ and $G$ in $k[x,y,z,w]$ of degrees $d$ and $c$, respectively. Following \cite{Hartshorne77} and \cite{Baker} we see that $H^1(X,\mathcal{O}_X)\cong H^3(\mathbb{P}^3,\mathcal{O}_\mathbb{P}^3(-c-d))$, where the basis $\mathcal{B}$ of $H^1(X,\mathcal{O}_X)$ that corresponds to the coordinates $x,y,z,w$ is associated to the basis of $ H^3(\mathbb{P}^3,\mathcal{O}_\mathbb{P}^3(-c-d))$ given by
\begin{align}\label{eq:basis_h3}
\{x^ky^lz^mw^n : (k,l,m,n)\in (\mathbb{Z}_{<0})^4 \text{ and } -k-l-m-n=c+d \}.
\end{align}
Using this fact, Kudo and Harashita present an algorithm to compute the Hasse--Witt of such curves, which can be generalized to compute the corresponding matrix for any complete intersection over a perfect field of positive characteristic.
\begin{prop}[Kudo and Harashita \cite{Kudo&Harashita}, Proposition 3.1.4]\label{prop:computation_hw}
Let $X$ of genus $g$ be defined as above and suppose $(FG)^{p-1}=\sum c_{i_1,i_2,i_3,i_4}x^{i_1}y^{i_2}z^{i_3}w^{i_4}$. Then the Hasse--Witt matrix of $X$ is given by
\begin{eqnarray}\label{eq:HWmatrix:general}
\left(c_{-k_jp+k_i,-l_jp+l_i,-m_jp+m_1i-n_jp+n_i}\right)_{i,j}.
\end{eqnarray}
\end{prop}
This formula gives a matrix $H$ that represents the action of $\mathcal{F}_X$ by left matrix multiplication. Let $\mathbf{v}$ be the column vector corresponding to an element of $H^1(X,\mathcal{O}_X)$ expressed in terms of the basis $\mathcal{B}$. Then the image of $\mathbf{v}$ under $\mathcal{F}_X$ is given by $H\cdot \mathbf{v}^{(p)}$, since $\mathcal{F}_X$ is $p$-linear.
In the next section, we use Proposition~\ref{prop:computation_hw} together with the equations from Section~\ref{sec:genus4} to compute examples of non-hyperelliptic smooth curves of genus $4$ with $a$-number $3$.
\section{A database of curves in standard form over $\mathbb{F}_p$ }\label{sec:examples_goodcurves_fp}
In this section we construct a database of genus 4 curves in standard form (Definition~\ref{def:standardform}) over $\mathbb{F}_p$ for $p \in \{3,5,7,11\}$. We restrict our data collection to non-ordinary and non-superspecial curves, that is, curves with $a$-number equal to $1,2$ or $3$. First let us explain the notation used:
\begin{itemize}
\item Let ${\vec{a}} \in \mathbb{F}_p^{20}$. We denote by $G_{{\vec{a}}}$ the cubic whose coefficients correspond to the entries of ${\vec{a}}$, assigned to the monomials of degree 3 in $\mathbb{F}_p[x,y,z,w]$ in graded lexicographic order, which means that we consider $x>y>z>w$ and to order the monomials we first compare the exponents of $x$, then those of $y$ and so on.
\item For each one of the cases D, N1i, N1ii and N2 in Definition~\ref{def:standardform} we define a subset of $\mathbb{F}_p^{20}$ such that for all ${\vec{a}}$ in that subset, the cubic $G_{{\vec{a}}}$ has conditions. We denote each subset by $\Ds{p},\Ni{p},\Nii{p}$ and $\Ntwo{p}$.
\end{itemize}
As a reference, the cardinality of each of the sets above is shown in Table~\ref{tab:card_vectors_sets}.
\begin{table}[h]
\caption{Cardinality of the sets $\Ds{p}$, $\Ni{p}$, $\Nii{p}$, $\Ntwo{p}$.}\label{tab:card_vectors_sets}
\begin{center}
\begin{tabular}{cc}
Set & Cardinality \\ \hline \hline
$\Ds{p}$ & $4p^3(p-1)(p^5+p-2)$ \\
$\Ni{p}$ & $6p^8(p-1)^2$ \\
$\Nii{p}$ & $12p^6(p-1)^2$ \\
$\Ntwo{p}$ & $4p^8(p^2-1)$ \\ \hline
\end{tabular}
\end{center}
\end{table}
To construct a curve in standard form we select an element ${\vec{a}}$ in one of the subsets above together with the corresponding $F \in \{F_1,F_2,F_d\}$ and verify if $V(F,G_{\vec{a}})$ is non-singular, irreducible and has genus $4$. In practice, we implement the algorithm in \texttt{Magma}, where we check each of the conditions as follows:
\begin{itemize}
\item Irreducibility: using the intrinsic \texttt{Magma} function \texttt{IsIrreducible}, which verifies the condition over the base field by a Gr\"{o}bner basis computation.
\item Genus: the command \texttt{Genus} finds the arithmetic genus of the projective normalization of the curve.
\item Nonsingularity: Implementing Algorithm \texttt{DetermineNonSingularity} (Algorithm 3.2.1 in \cite{Kudo&Harashita}), which is based on solving a radical membership problem on the minors of the Jacobian matrix of $V(F,G_{\vec{a}})$. The later is done by a \texttt{RadicalMembership} algorithm (\cite{Kudo&Harashita}, Appendix A)
\end{itemize}
Next, suppose that $X=V(F,G_{\vec{a}})$ is a curve in standard form, then we need to determine its $a$-number and $p$-rank. We do this by computing the Hasse--Witt matrix $H$ of $X$ with respect to the basis of $H^1(X,\calO_X)$ given by (\ref{eq:basis_h3}), as in Proposition~\ref{prop:computation_hw}. The $a$-number is equal to $4-\text{rank}(H)$ and the $p$-rank $f$ is the rank of $HH^{(p)}\cdots H^{\left(p^{g-1}\right)}$. In our case, since $F$ and $G_{\vec{a}}$ are defined over $\mathbb{F}_p$, then $f=\text{rank}(H^{g})$.
\subsection{Collecting the data}\label{sec:data_collection_curves}
We apply the above procedure to random samples of tuples in $\Ds{p}$, $\Ni{p}$, $\Nii{p}$ and $\Ntwo{p}$ for $p\in \{3,5,7,11\}$ in order to gather statistical information. We construct the sample using the intrinsic \texttt{Random} function in \texttt{Magma}. In addition, we classify the curves by $a$-number and $p$-rank.
When we analyze the samples of curves in our data, we will often compare the number of curves with certain $p$-rank and/or $a$-number with the total of smooth curves in standard form obtained. In our search we classify the curves with $a$-number $0,1,2$ and $3$. We ignore the curves with $a=4$ because they represent a very small proportion of the curves and hence they do not affect the percentages in a significant way.
Also, our focus will be studying the occurrence of Cartier points (see Section~\ref{sec:CartierPoints}), and we already know that superspecial curves have infinitely many of them (Baker \cite{Baker}). Here is how we will refer to our different samples:
\begin{itemize}
\item \textbf{Total set}: the set $\Ds{p}$, $\Ni{p}$, $\Nii{p}$ or $\Ntwo{p}$, depending on the case.
\item\textbf{Sampled set}: the curves of the sets above that are included in our random search.
\item \textbf{Smooth sample}: the set of cubics from the sampled set that give smooth, irreducible, genus $4$ curves (excluding superspecial curves).
\item \textbf{Singular sample}: the set of cubics from the sampled set that give curves that fail either one of the conditions for smoothness, irreducibility or genus.
\end{itemize}
If we do not specify the subcase, then \textbf{ Sampled set}, \textbf{Smooth sample} and \textbf{Singular sample} will refer to the total samples throughout the four cases.
\subsection{Summary of results of sampling search}\label{sec:summary_sampling}
In this section we display the overall results from the sampling search. The sizes of our final Sampled sets by case are shown in \tref{tab:total_sample_sizes}.
\begin{table}[h]
\caption{Total sampled sizes over $\mathbb{F}_p$.}\label{tab:total_sample_sizes}
\begin{center}
\begin{tabular}{c|cccc|c}
$p$ & D & N1i & N1ii & N2 & Total sample \\ \hline \hline
$3$ & 52704 & 92123 & 34992 & 6447 & 186266 \\
$5$ & 179728 & 179970 & 179434 & 179970 & 719102 \\
$7$ & 215957 & 225193 & 206890 & 214998 & 863038 \\
$11$ & 89999 & 91100 & 89999 & 90000 & 361098 \\ \hline
\end{tabular}
\end{center}
\end{table}
In Table~\ref{tab:summary_allcurves} we list the number of curves with $p$-ranks $f=0,1,2,3$ and $a$-number $a=1,2,3$, over $\mathbb{F}_p$ for $p\in \{3,5,7,11\}$. We include the totals for the singular samples and the ordinary curves (that is, those with $a$-number $0$). From this classification we have the following statement:
\begin{cor}\label{cor:no_a1_f0_curves}
In our data set of genus $4$ curves in standard form there are no curves with $p$-rank $0$ and $a$-number $1$, when $p\in \{3,5,7,11\}$.
\end{cor}
\begin{proof}
See Table~\ref{tab:summary_allcurves}.
\end{proof}
\begin{table}[h]
\caption{ Sample of curves in standard over $\mathbb{F}_p$.}\label{tab:summary_allcurves}
\begin{center}
\begin{tabular}{cccccc}
$a$ & $f$ & 3 & 5 & 7 & 11 \\ \hline\hline
\multicolumn{2}{c}{Sampled set} & 186266 & 719102 & 863038 & 361098 \\ \hline
\multicolumn{2}{c}{Singular sample} & 92654 & 251584 & 191925 & 81845 \\ \hline
0 & & 56983 & 370476 & 529394 & 253627 \\ \hline
1 & 0 & 0 & 0 & 0 & 0 \\
1 & 1 & 1679 & 3592 & 1652 & 217 \\
1 & 2 & 4134 & 14615 & 10687 & 2146 \\
1 & 3 & 23485 & 74585 & 76142 & 23044 \\
& Total & 29298 & 92792 & 88481 & 25407 \\ \hline
2 & 0 & 1157 & 183 & 44 & 3 \\
2 & 1 & 2095 & 790 & 231 & 21 \\
2 & 2 & 3379 & 3231 & 1624 & 194 \\
& Total & 6631 & 4204 & 1899 & 218 \\ \hline
3 & 0 & 0 & 10 & 0 & 0 \\
3 & 1 & 700 & 36 & 5 & 1 \\
& Total & 700 & 46 & 5 & 1 \\ \hline
\multicolumn{2}{c}{Smooth sample} & 93612 & 467518 & 619779 & 279253 \\ \hline
\end{tabular}
\end{center}
\end{table}
Next we discuss the results from the search for each $p$. We include the sizes of the total set, sampled set, smooth sample and singular sample, normalized by $\log_{p}$. Then we show the break down of curves with $a=1,2,3$ and specify the percentage of the sampled set and smooth sample they represent.
\subsubsection{Case $p=3$ .}\label{sec:goodcurvesp=3}
We checked a total of $186266$ tuples, which corresponds to approximately $71\%$ of all the possible tuples. We saw that $50.26\%$ of them gave smooth curves and $19.66\%$ of the total had $a$-number $a=1,2,3$.
Notice that we were able to sort all of the cubics from the total sets $\Ds{3}$ and $\Nii{3}$.
\subsubsection{Case $p=5$. }\label{sec:goodcurvesp5}
We selected a random sample of $719102$ tuples in $\Ds{5}, \Ni{5}, \Nii{5}$ and $\Ntwo{5}$. This is around $1.49\%$ of the total set. A $61.01\%$ of that sample corresponds to the Smooth sample, and $13.49\%$ of the total are non-ordinary curves.
\subsubsection{Case $p=7$. }\label{sec:goodcurvesp=7}
We analyzed a random sample of $863038$ pairs of tuples in $\Ds{7}$, $\Ni{7}$, $\Nii{7}$ and $\Ntwo{7}$, which corresponds to $0.06\%$ of the total. The smooth sample from this set has $619779$ tuples and $90385$ of them are curves in standard form with $a$-number $1,2$ or $3$. These correspond to $71.81\%$ and $10.47\%$ of the total, respectively. %
\subsubsection{Case $p=11$}\label{sec:goodcurves_p11}
In this case we did a random search that included $361098$ tuples in $\Ds{11}, \Ni{11}$, $ \Nii{11}$ and $\Ntwo{11}$, this is approximately $0.0002\%$ of the total set. Of this sample, $77.33\%$ are smooth and $7.10\%$ of the total are non-ordinary.
\subsection{The case $a=3$}\label{sec:summary_exhaustive}
In order to focus on the analysis of some aspects of genus $4$ curves with $a$-number $3$, we apply the procedure above to all the tuples in $\Ds{p}, \Ni{p}, \Nii{p}$ and $\Ntwo{p}$, but only store the smooth, irreducible curves with $a=3$. We did this for $p=3,5$. The search is also done for $p=7$ but only for tuples in $\Ds{7}$ and a subset of $\Ni{7}$, because of the long computing times. It is important to remark that after the search, we classify the curves by $\mathbb{F}_p$-isomorphism classes.
In the cases $p=3,5$ we also have a complete list of all curves in standard form that have $a$-number $3$. For $p=7$ we have a subset of them: the ones in case D and those in case N1i where $(b_1,b_2)=(0,0)$, we denote these lists by 7(D) and 7(N1i'). The information on these curves, classified by isomorphism classes, is shown in Table~\ref{tab:isom_classes_p3_a3}. A first conclusion that we can draw from this data is the following:
\begin{cor}\label{cor:no_prank0_p3_curves}
There are no genus $4$ curves in standard form over $\mathbb{F}_3$ with $p$-rank $0$ and $a$-number $3$.
\end{cor}
\begin{table}[h]
\caption{Isomorphism classes of curves in standard form with $a=3$ over $\mathbb{F}_p$. }\label{tab:isom_classes_p3_a3}
\begin{center}
\begin{tabular}{ccccc}
$p$-rank & 3 & 5 & 7(D) &7(N1i') \\ \hline \hline
0 & 0 & 36 & 9&2 \\
1 & 27 & 98 & 56& 27\\
\hline
Total & 27 & 134 & 65& 29\\ \hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{Case $p=3$}
There are a total of $1188$ vectors ${\vec{a}}$ in $\Ds{3}$, $\Ni{3}$, $\Nii{3}$ and $\Ntwo{3}$ that give curves of genus $4$ and $a$-number $3$ over $\mathbb{F}_3$. We give a summary of the number of vectors classified by case and some restrictions that occur in the case D. We use {\texttt{Magma}} to classify these curves in $\mathbb{F}_3$-isomorphism classes, which we detail in Table~\ref{table:isom_classes_p3} and Lemma~\ref{lem:standard_form_a3_p3_curves}.
\begin{table}[h]
\caption{Isomorphism classes of genus $4$ and $a=3$ curves over $\mathbb{F}_3$.}
\label{table:isom_classes_p3}
\begin{center}
\begin{tabular}{cccc}
Case &$\#$ Isomorphism classes & $\#$ $p$-rank $1$& $\#$ $p$-rank $0$\\
\hline
\hline
D&6 & 6&0 \\
N1(i) & 7& 7&0 \\
N1(ii) & 3& 3&0 \\
N2 & 11& 11& 0\\
\hline
Total & 27& 27&0 \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{cor}\label{cor:exhaustive_search_p3}
There are, up to $\mathbb{F}_3$-isomorphism, exactly $27$ curves of genus $4$ with $a=3$ in standard form.
\end{cor}
\begin{proof}
See \tref{table:isom_classes_p3}.
\end{proof}
\begin{lem}\label{lem:standard_form_a3_p3_curves}
Let $X$ be a curve in standard form with $a$-number $3$ defined over $\mathbb{F}_3$.
Then $X$ is isomorphic to one of the following:
\begin{itemize}
\item $V(F_d,G)$ with $G=x^3+y^3+xyz+c_1yz^2+xw^2+c_2w^3,$
where $c_1 \in \mathbb{F}_3^\times$ and $c_2 \in \mathbb{F}_3$.
\item $V(F_1,G)$ with $G=x^2y+c_1y^3+x^2z+c_2y^2w+c_3w^3+c_4z^3 +z^2w,$
where $c_1 \in \{0,1\}$, $c_2,c_3 \in \mathbb{F}_3^\times$ and $c_4 \in \mathbb{F}_3$.
\item $V(F_2,G)$ with $G=x^2y+c_1y^3+c_2x^2z+c_3z^3+c_4yzw+c_5w^3+c_6(z^2w+y^2w),$
where $c_1,c_3 \in \mathbb{F}_3$, $c_2,c_4,c_5 \in \mathbb{F}_3^\times$ and $c_6 \in \{0,1\}$.
\end{itemize}
\end{lem}
\subsubsection{Case $p=5$}
From the exhaustive search we conclude that there are $134$ $\mathbb{F}_5$-isomorphism classes of standard form curves over $\mathbb{F}_5$ with $a$-number $3$. Table~\ref{table:isom_classes_p5} contains the summary of the isomorphism classes and the number of curves with $p$-rank $1$ and $p$-rank $0$.
\begin{table}[h]
\caption{Isomorphism classes of genus $4$ and $a=3$ curves over $\mathbb{F}_5$.}
\label{table:isom_classes_p5}
\begin{center}
\begin{tabular}{cccc}
Case &$\#$ Isomorphism classes & $\#$ $p$-rank $1$& $\#$ $p$-rank $0$\\
\hline
\hline
D&59 &46 &13 \\
N1(i) & 60& 48&12 \\
N1(ii) & 6& 4&2 \\
N2 & 9& 0& 9\\
\hline
Total & 134& 98& 36\\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{cor}\label{cor:exhaustive_search_p5}
There are, up to $\mathbb{F}_5$-isomorphism, exactly $134$ curves of genus $4$ and $a=3$ in standard form.
\end{cor}
\begin{proof}
See \tref{table:isom_classes_p5}.
\end{proof}
\subsubsection{Case $p=7$}
The exhaustive search for all curves in standard form with $a=3$ proved to be too time consuming for $p=7$. So it was only possible to find the curves in the case D and a subset of N1 curves. We conclude from this search that there are at least $94$ $\mathbb{F}_7$-isomorphism classes of curves of $a$-number $3$, where $65$ correspond to the case D and $29$ to the N1. Now, we know by Lemma 4.5.1 in \cite{Kudo&Harashita} that any smooth, irreducible, genus $4$ curve $X=V(F_d,G)$ can be written in standard form. This is because condition (A3) is satisfied over $\mathbb{F}_7$ and $G$ can be reduced by a change of variables to the form of Definition~\ref{def:standardform}. This implies that the list of curves that we found in the case D actually includes all of the curves where the quadratic polynomial is degenerate.
\begin{cor}\label{cor:degenerate_curves_over_F7}
There are, up to $\mathbb{F}_7$-isomorphism, exactly $65$ genus $4$ and $a$-number $3$, smooth, irreducible, non-hyperelliptic curves over $\mathbb{F}_7$, given as $V(F,G)$, where $G$ is a cubic homogeneous polynomial and $F$ is a degenerate quadratic form. In addition, there are, up to $\mathbb{F}_7$-isomorphism, at least $29$ genus $4$ curve in standard form over $\mathbb{F}_7$ with $a$-number $3$ where $F$ is non-degenerate.
\end{cor}
\begin{proof}
See Table~\ref{table:isom_classes_p7}.
\end{proof}
In the case D there are $1440$ curves with $a$-number $3$, which are divided into $65$ classes, $56$ of them have $p$-rank $1$ and $9$ have $p$-rank $0$. The subset of curves of the case N1i that we computed corresponds to those where the the cubic is of the form where $b_1=b_2=0$ (see Definition~\ref{def:standardform}). There are $432$ of these $G$ such that $V(F_1,G)$ is a smooth genus $4$ smooth curve with $a$-number $3$, only $16$ have $p$-rank $0$. These curves are distributed in $29$ ${\mathbb{F}}_7$-isomorphism class with $2$ of them having $p$-rank $0$ and $27$ of $p$-rank $1$.
\begin{table}[h]
\caption{Isomorphism classes of genus $4$ and $a=3$ curves over $\mathbb{F}_7$.}
\label{table:isom_classes_p7}
\begin{center}
\begin{tabular}{cccc}
Case &$\#$ Isomorphism classes & $\#$ $p$-rank $1$& $\#$ $p$-rank $0$\\
\hline
\hline
D&65 &56 &9 \\
N1i' & 29& 27&2 \\
\hline
Total &94 &83 &11 \\ \hline
\end{tabular}
\end{center}
\end{table}
\section{Cartier points }\label{sec:CartierPoints}
\subsection{Definition and properties}\label{sect:CartierPoints_definitions and properties}
Theorem~\ref{thm:Ekedahl} states that the genus of a superspecial curve in characteristic $p$ is bounded by $p(p-1)/2$. Ekedahl \cite{Ekedahl:Supersingular} bases the proof of this result on the fact that a curve is superspecial if and only if its Jacobian is isomorphic to the product of supersingular elliptic curves (Oort, \cite{Oort75}). Baker, on the other hand, presents in \cite{Baker} an alternative proof which makes use of an equivalent definition: a curve is superspecial if and only if the Cartier operator annihilates $H^0(X,\Omega_X^1)$. The other component of his proof is the existence of linear systems of dimension $1$, associated to a certain type of points on $X$, defined as Cartier points.
\begin{definition}\label{def:cartier_points}
A closed point $P$ of $X$ is said to be a Cartier point if the hyperplane of regular differentials vanishing at $P$ is stable under the Cartier operator.
\end{definition}
The following result yields Theorem~\ref{thm:Ekedahl} as a corollary.
\begin{thm}[Baker \cite{Baker}, Theorem 2.8]\label{thm:baker}
Let $X$ be a curve of genus $g$ over an algebraically closed field of characteristic $p$.
\begin{enumerate}
\item If $X$ has at least $p$ distinct Cartier points, no two of which differ by a $p$-torsion point on $J_X$, then $g \leq p(p-1)/2$.
\item If $X$ is hyperelliptic of genus $g$, $p$ is odd, and some hyperelliptic branch point of $X$ is a Cartier point, then $g \leq (p-1)/2$.
\end{enumerate}
\end{thm}
\begin{notation}\label{remark:on_notation}
Let $k$ be a field of characteristic $p$. Let $X$ be a non-hyperelliptic curve of genus $g$ over $k$ embedded in $\mathbb{P}(H^0(X,\Omega_X^1))=\mathbb{P}^{g-1}$ by a basis \linebreak$\mathcal{B}'=\{\omega_1,\ldots, \omega_g\}$ of $H^0(X,\Omega_X^1)$. Suppose $x_1,\ldots, x_g$ are the coordinates of $\mathbb{P}^{g-1}$ given by this basis and that $\mathcal{B}$ is the basis of $H^1(X,\mathcal{O}_X)$ dual to $\mathcal{B}'$. Given a point $P=[a_1:\cdots :a_g] $ of $X$ we denote by $\mathbf{v}_P$ the vector $(a_1,\ldots, a_g)^T$ in $H^1(X,\mathcal{O}_X)$ expressed in terms of $\mathcal{B}$.
\end{notation}
From here on, we consider the Hasse--Witt matrix $H$ of $X$ to be in terms of the basis $\mathcal{B}$, unless otherwise stated.
\begin{prop}\label{prop:find-c-pts-with-HW}
Let $X$ be a non-hyperelliptic curve of genus $g$ over $k$ embedded in $\mathbb{P}^{g-1}$ as in Notation~\ref{remark:on_notation}. A point $P=[a_1:\cdots :a_g] $ of $X(\overline{k})$ is a Cartier point if and only if there exists $c\in \overline{k}$ such that
\begin{equation}\label{eq:Hv^p=cv_is CP}
H\mathbf{v}_P^{(p)}=c\mathbf{v}_P,
\end{equation}
where $\mathbf{v}_P^{(i)}$ indicates that each entry of the vector is raised to the $i$-th power.
\end{prop}
\begin{proof}
Since $X$ is embedded in $\mathbb{P}^{g-1}$ by $\{\omega_1,\ldots, \omega_g\}$ then $\omega_i(P)=a_i$, for \linebreak $i=1,\ldots, g$. A regular $1$-form $\omega=b_1\omega_1+\cdots + b_g\omega_g$ vanishes at $P$ if and only if
\begin{equation}
b_1a_1+\cdots + b_ga_g=0.
\end{equation}
Then the hyperplane of $1$-forms vanishing at $P$ is
%
\begin{equation}
L_P:= \left\lbrace (b_1,\ldots ,b_g)^T: b_1a_1+\cdots + b_ga_g=0\right\rbrace.
\end{equation}
%
Let $L_P^0$ be the annihilator of $L_P$. We know that this is a $1$-dimensional subspace of ${H^0(X,\Omega_X^1)}^{*}\cong H^1(X,\mathcal{O}_X)$, so it is generated by the vector $\mathbf{v}=(a_1,\ldots, a_g)^T$.
Now, $L_P$ is stable under the action of $\mathcal{C}$ if $\mathcal{C}(\omega)\in L_P$, for every $\omega \in L_P$. By duality, this is equivalent to $\mathcal{F}_X(\mathbf{v}_P)=c\mathbf{v}_P$, that is, $H\mathbf{v}_P^{(p)}=c\mathbf{v}_P.$
\end{proof}
\begin{definition}\label{def:T1andT2}
We say that a Cartier point $P\in X$ as above is of\textit{ Type $1$} if $c=0$ and of \textit{Type $2$} otherwise.
\end{definition}
Notice that if $k$ is not algebraically closed, then the Cartier points of $X/k$ might not be defined over $k$, but over some extension. In general we consider Cartier points as points in $X(\overline{k})$. The next lemma gives us a way to find Type $1$ points.
\begin{lem}\label{lem:Type_1Points_H1/pv=0}
A point $P$ is a Type $1$ Cartier point of $X$ if and only if \linebreak$H^{(1/p)}\mathbf{v}_P=0$.
\end{lem}
\begin{proof}
By definition $P$ is a Type $1$ point of $X$ if and only if $H\mathbf{v}^{(p)}=0$. By applying the inverse of the $p$-th Frobenius morphism we see that this is equivalent to $H^{(1/p)}\mathbf{v}_P=0$.
\end{proof}
\begin{cor}\label{cor:find_type1_points_Hv=0_Fp}
Supose $X$ is defined over $\mathbb{F}_p$. Then $P\in X$ is a Type $1$ point if and only if $H\mathbf{v}_P=0$.
\end{cor}
Suppose that $X$ is defined over $\mathbb{F}_{q}$, with $p^r=q$ for some positive integer $r$. For a point $Q\in X$ we denote by $\sigma(Q)$ the action of the $r$-th power of the Frobenius morphism on $Q$
\begin{lem}\label{lem:frob_cartierpoint}
Let $Q$ be a Cartier point of $X/\mathbb{F}_q$ and let $P=\sigma(Q)$. Then $Q$ is a Cartier point of $X$ if and only if $P$ is a Cartier point of $X$. Furthermore, if $H\mathbf{v}_Q^{(p)}=c\mathbf{v}_Q$, then $H\mathbf{v}_P^{(p)}=c^{q}\mathbf{v}_P$.
\end{lem}
\begin{proof}
First we note that $P$ is also a point on $X$, because $X=V(F,G)$ and $F$ and $G$ are defined over $\mathbb{F}_q$. Let $H$ be the Hasse--Witt matrix of $X$. After scaling, we can assume that $\mathbf{v}_Q^{(q)}=\mathbf{v}_P$. Then
$$\begin{array}{ccl}
H\mathbf{v}_Q^{(p)}=c\mathbf{v}_Q& \Leftrightarrow&(H\mathbf{v}_Q^{(p)})^{(q)}=(c\mathbf{v}_Q )^{(q)}\\
&\Leftrightarrow& H(\mathbf{v}_Q^{(q)})^{(p)}=c^{q}(\mathbf{v}_Q)^{(q)}\\
&\Leftrightarrow & H\mathbf{v}_P^{(p)}=c^{q}\mathbf{v}_P,
\end{array}$$
where the second equivalence is true because $X$ is defined over $\mathbb{F}_q$ then so is $H$.
\end{proof}
We apply Lemma~\ref{lem:frob_cartierpoint} to reduce the search of Cartier points to a computation of eigenvectors.
\begin{lem}\label{lem:T_points_are_eigenvectors}
Suppose $X$ is defined over $\mathbb{F}_p$ with Hasse--Witt matrix $H$. Let $Q$ be a Cartier point of $X$ defined over $\mathbb{F}_{p^e}$ for some positive integer $e$. There exists $\lambda \in \mathbb{F}_{p}$ such that $H^e\mathbf{v}_Q=\lambda\mathbf{v}_Q$.
\end{lem}
\begin{proof}
Since $Q$ is a Cartier point in of $X$ then $H\mathbf{v}_Q^{(p)}=c\mathbf{v}_Q$ for some $c\in \overline{\F}_p$. Now, since $Q\in X(\mathbb{F}_{p^e})$, then $c\in\mathbb{F}_{p^e}$. Let $P_i:=\sigma^i(Q)$, then $P_1, P_2, \ldots, P_e=Q$ are distinct Cartier points. Also, after scaling we can assume that $\mathbf{v}_{P_i}^{(p)}=\mathbf{v}_{P_{i+1}} $. After applying $i$ times the result from Lemma~\ref{lem:frob_cartierpoint} with $q=p$, we get that $H\mathbf{v}_{P_i}^{(p)}=c^{p^i}\mathbf{v}_{P_i} $. To ease notation, we write $\mathbf{v}_i$ instead of $\mathbf{v}_{P_i}$. Then
%
\begin{eqnarray*}
H^e\mathbf{v}_e=&H^e\mathbf{v}_{e-1}^{(p)}=H^{e-1}\left(H\mathbf{v}_{e-1}^{(p)}\right)= H^{e-1}\left(c^{p^{e-1}}\mathbf{v}_{e-1}\right)= c^{p^{e-1}}H^{e-1}\mathbf{v}_{e-2}^{(p)}.
\end{eqnarray*}
By an inductive process, we get that
$H^e\mathbf{v}_Q= c^{p^{e-1}+p^{e-2}\cdots +p+1}\mathbf{v}_Q=c^{\frac{p^e-1}{p-1}}\mathbf{v}_Q$. Let $\lambda:=c^{\frac{p^e-1}{p-1}}$. Since $c\in \mathbb{F}_{p^e}$ then $\lambda$ is a $(p-1)$-root of unity, if $c\neq 0$ and $\lambda=0$ if $c=0$. Hence $\lambda \in \mathbb{F}_p$.
\end{proof}
Baker provides in \cite{Baker} an upper bound for the number of Cartier points on a smooth irreducible curve that is not ordinary nor superspecial.
\begin{prop} [Baker, \cite{Baker} Prop. 3.3]\label{prop:bound_cartier_points}
Let $X$ be a smooth, irreducible curve of genus $g$ with $p$-rank $f$, which is not ordinary nor superspecial.
\begin{enumerate}
\item The number of Type $2$ points on $X$ is bounded by
\begin{align*}\label{eq:bound-on-CP}
b:=b_{g,p,f,\delta_X}= \min \left(2g -2, \delta_X\frac{p^f-1}{p-1} \right),
\end{align*}
where $\delta_X$ is $2$ if $X$ is hyperelliptic and $1$ otherwise.
\item The number of Type $1$ points on $X$ is bounded by $2g-2$. Furthermore, if the $a$-number of $X$ is $g-1$ then there is at least one Type $1$ Cartier point on $X$.
\end{enumerate}
\end{prop}
Let us explore the geometric meaning of this bound, following the proof of Proposition~\ref{prop:bound_cartier_points} in \cite{Baker}. We will only be concerned with non-hyperelliptic curves, so here $\delta_X=1$.
We know by Proposition~\ref{prop:find-c-pts-with-HW} that $P\in X(\overline{k})$ is a Cartier point if and only if there exists $c \in \overline{k}$ such that $H\mathbf{v}_P^{(p)}=c\mathbf{v}_P$. If $c=0$ then this equation is equivalent to $H\mathbf{v}_P^{(1/p)}=0$. Hence the Type $1$ points are those in the intersection of $X(\overline{k})$ and the subspace spanned by the kernel of $H^{(1/p)}$ in $\mathbb{P}^{g-1}$.
This subspace is linear and has codimension at least $1$, so it is contained in a hyperplane. From where we conclude that the number of points in the intersection is at most the degree of the curve $2g-2$.
On the other hand, if $c\neq 0$ then we can rewrite Equation \eref{eq:Hv^p=cv_is CP} as $H\mathbf{w}^{(p)}=\mathbf{w}$ by setting $\mathbf{w}=\lambda\mathbf{v}_P$
for $\lambda$ equal to some $(p-1)$-th root of $c^{-1}$. Now the element of $H^1(X,\mathcal{O}_X)$ given by $\mathbf{w}$ is fixed by the Frobenius operator. By definition the $p$-rank $f$ is the dimension of the subspace of $H^1(X,\calO_X)$ where $\mathcal{F}$ is bijective. So there are $p^f-1$ non trivial solutions, that yield at most $\frac{p^f-1}{p-1}$ Type $2$ points. After maybe doing a base extension on $X$ one can choose a basis of $H^0(X,\Omega_X^1)$ given by $\{\xi_1, \ldots, \xi_g\}$ such that $\mathcal{C}(\xi_i)=\xi_i$ for $1\leq i\leq f$. Assume that the coordinates of a Type $2$ point are given by this basis. Then,some coordinate, say $x_g$ of every such point must be zero, since $f<g$. Then the point lies on the hyperplane $x_g=0$. Again, there can only be $2g-2$ such points, so the number of Type $2$ points is $\min \{ 2g-2,\frac{p^f-1}{p-1} \}$.
%
As a direct consequence of the existence of the upper bound on the number of Cartier points, we also get a bound on the degree of the field of definition of the point. In particular, we have Corollary~\ref{cor:degree_of_definition_C_points} for the case $k=\mathbb{F}_q$.
\begin{cor}\label{cor:degree_of_definition_C_points}
Let $X$ be a genus $g$ non-hyperelliptic curve that is not ordinary nor superspecial, defined over $\mathbb{F}_q$.
\begin{itemize}
\item If $P$ is a Type $1$ Cartier point of $X$, then $P\in X(\mathbb{F}_{q^e})$ with $1\le e \le 2g-2$.
\item If $P$ is a Type $2$ Cartier point of $X$, then $P\in X(\mathbb{F}_{q^e})$ with $1\le e \le b$, where $b$ is as in Proposition~\ref{prop:bound_cartier_points}.
\end{itemize}
\end{cor}
\begin{proof}
Let $P$ be a Cartier point of $X$. Let $e$ be the minimum positive integer such that $P\in X(\mathbb{F}_{q^e})$. By Lemma~\ref{lem:frob_cartierpoint}, the $e$ distinct points \linebreak $\{P, \sigma(P), \sigma^2(P), \ldots, , \sigma^{e-1}(P)\}$ are all Cartier points of the same type. By Proposition~\ref{prop:bound_cartier_points}, there are at most $2g-2$ Type $1$ points and $b$ Type $2$ points. Hence, if $P$ is a Type $1$ point (resp. Type $2$), then $e\leq 2g-2$ (resp. $e\leq b$).
%
\end{proof}
\subsubsection{Type $1$ points in the case $a=g-1$} \label{sec:T1_a=g-1}
The behavior of the Cartier points when the $a$-number is $g-1$ has an additional feature which is the multiplicity. In this case, the subspace $S$ generated by the kernel of $H^{(1/p)}$ is a hyperplane, then assuming $X\nsubseteq S$, the intersection $X \cap S$ is proper. Then we can define the intersection multiplicity of $P$ in $ X \cap S$, that is, of the Type $1$ Cartier points.
If $X\cap S=\{P_1,\ldots, P_n\}$ and $m_i$ denotes the multiplicity of the point $P_i$, then $\sum_{i=1}^{n}m_i=2g-2 $. Hence the possible multiplicity distributions of the Type 1 points correspond to the partitions of $2g-2$. Both the multiplicity and the degree of $P$ are preserved under $\sigma$, so if $d_i$ is the degree of the point $P_i$, then $P_i, \sigma(P_i),\sigma^2(P_i),\ldots, \sigma^{d_i-1}(P_i)$ are $d_i$ distinct points of the same degree and multiplicity.
\subsection{Cartier points on genus $4$ curves}\label{sec:CartierPoints_g4_curves}
Suppose $q=p^r$ for some positive integer $r$ and let $X=V(F,G)$ be a non-ordinary and non-superspecial smooth, irreducible genus $4$ non-hyperelliptic curve over $\mathbb{F}_q$. We want to determine the sharpness of the bound given by Baker in Proposition~\ref{prop:bound_cartier_points} for the number of Cartier points on $X$. We will first make some remarks about the possible bounds depending on the $a$-number and the $p$-rank.
\begin{cor}[Type $2$ Cartier points and the $p$-rank]\label{cor:type2CP}
Let $X$ be a non-hyperelliptic curve of genus $4$ defined over $\mathbb{F}_q$ with $p$-rank $f$.
\begin{enumerate}
\item [(i)] There are at most six Type $2$ points on $X$ and they are defined over $\mathbb{F}_{q^e}$ for some $1\le e \le 6$. Moreover:
\item[(ii)] if $f=0$, then there are no Type $2$ points;
\item[(iii)] if $f=1$, then there is at most one Type $2$ point and it must be defined over $\mathbb{F}_q$;
\item[(iv)] if $f=2$, then there are at most six Type $2$ points, at most three if $p=2$ and at most four if $p=3$.
\end{enumerate}
\end{cor}
\begin{proof}
Parts \textit{(i)}, \textit{(ii)} and \textit{(iv)} are direct consequences of the discussion following Proposition~\ref{prop:bound_cartier_points}. Indeed, the minimum of $2g-2=6$ and $\frac{p^f-1}{p-1}$ is 6, unless $f\leq 1$ or $p=2,3$ and $f= 2$.
For \textit{(iii)} let $Q$ be a Type $2$ Cartier point of $X$ and suppose that $\sigma(Q)\neq Q$. By Lemma~\ref{lem:frob_cartierpoint} we have that $\sigma(Q)$ is another Cartier point of Type 2, which is a contradiction.
\end{proof}
\begin{cor}[Type $1$ points and the $a$-number]\label{cor:type1_genus4}
If $X$ is non-hyperelliptic of genus $4$ with $a$-number $1,2$ or $3$ defined over $\mathbb{F}_q$, then
\begin{enumerate}
\item[(i)] $X$ has at most six Cartier points of Type $1$ and they are defined over $\mathbb{F}_{q^e}$ for some $e\leq 6$. Moreover,
\item[(ii)] if $a=1$, then there is at most one Type $1$ Cartier point
\item[(iii)] if $a=3$, there are exactly six Type $1$ Cartier points on $X$, counting with multiplicity.
\end{enumerate}
\end{cor}
\begin{proof}
Part \textit{(i)} follows from Proposition~\ref{prop:bound_cartier_points} and
Corollary~\ref{cor:degree_of_definition_C_points}, \linebreak using $2g-2=6$.
Now, if $a=1$, the kernel of $H^{(1/p)}$ has dimension $1$, because $a$ equals $\dim(\ker(H))$. Then $\ker(H)$ spans a point in $\mathbb{P}^3$, and this is the only possible Type $1$ point.
When $a=3$, on the other hand, the subspace spanned by the same kernel is a hyperplane, so its intersection with $X$ consists on exactly as many points as the degree of the curve.
\end{proof}
\subsubsection{Computing Cartier points}
We explain here an algorithm to find the Cartier points on smooth, irreducible genus $4$ non-hyperelliptic curves over $\mathbb{F}_p$, given their quadratic and cubic defining homogeneous polynomials. Before we detail the procedure, we will revisit some facts about Cartier points. Let $X=V(F,G)$ be a curve as before.
\begin{enumerate}
\item Let $H$ be the Hasse--Witt matrix of $X$ as in Proposition~\ref{prop:computation_hw}. Recall that $H$ represents the action of Frobenius on $H^1(X,\mathcal{O}_X)$ with respect to the basis that corresponds to the coordinates $x,y,z,w$. This implies that we can use $H$ and Proposition~\ref{prop:find-c-pts-with-HW} to find the Cartier points of $X$, by solving $H\mathbf{v}^{(p)}=c\mathbf{v}$.
\item If $P \in X(\mathbb{F}_{p^e})$ is a Cartier point, then there exists some $\lambda \in \mathbb{F}_{p}$ such that $H^e\mathbf{v}_P=\lambda\mathbf{v}_P$ (Lemma~\ref{lem:T_points_are_eigenvectors}).
\item If $P$ is a Type $1$ point, then $H\mathbf{v}_P=0$ (Corollary~\ref{cor:find_type1_points_Hv=0_Fp}).
\item The eigenvalues of $H^e$ are exactly the $e$-powers of the eigenvalues of $H$. If $\lambda $ is as in (2), then there exists an eigenvalue $\mu$ of $H$ such that $\lambda=\mu^e\in \mathbb{F}_{p}$
\item Let $h(x)$ be the characteristic polynomial of $H$. The splitting field of $h(x)$ is either $\mathbb{F}_p$, $\mathbb{F}_{p^2}$ or $\mathbb{F}_{p^3}$. Indeed, $h(x)$ has degree $4$, but since $H$ has rank at most $3$, then $x$ is a factor of $h(x)$.
\end{enumerate}
We use these facts to compute the Type $1$ and Type $2$ Cartier points of $X$. Algorithm~\ref{alg:Type2_points} is restricted to the case when $q=p$, to simplify the computations.
\begin{algo}\label{alg:Type1_points}[Type $1$ Cartier points]
\noindent
Input: $F$ and $G$ in $\mathbb{F}_q[x,y,z,w]$.\\
Output: List of Type $1$ Cartier points of $X=V(F,G)$.
\begin{enumerate}
\item Compute the Hasse--Witt matrix $H$ of $X$, as in Proposition~\ref{prop:computation_hw}.
\item Let $M=H^{(1/p)}$. Construct the linear forms $L_i=(M)_{i,1}x+(M)_{i,2}y+(M)_{i,3}z+(M)_{i,4}w$, for $1\leq i\leq 4$.
\item Let $I$ be the ideal generated by $\{L_1,L_2,L_3,L_4,F,G\}$ and let $T=V(I)$.
\item For each $1\le e \le 6$, find the points in $T_e=T(\mathbb{F}_{q^e})$.
\item The set of Type $1$ points is $\bigcup_{e}(T_e)$.
\end{enumerate}
\end{algo}
\begin{algo}\label{alg:Type2_points}[Type $2$ Cartier points]
\noindent
Input: $F$ and $G$ in $\mathbb{F}_p[x,y,z,w]$.\\
Output: List of Type $2$ Cartier points of $X=V(F,G)$.
\begin{enumerate}
\item Compute the Hasse--Witt matrix $H$ of $X$, as in Proposition~\ref{prop:computation_hw}.
\item Compute $h(x)$, the characteristic polynomial of $H$ and find the roots of $h(x)$ in its splitting field. For each $\mu$ non-zero root of $h(x)$ and each $1\le e \le 6$ such that $ \mu^e \in \mathbb{F}_{p}$:
\begin{enumerate}
\item Let $M=H^e-\mu^eI$, where $I$ is the identity matrix.
\item For $1\leq i\leq 4$, construct the linear forms $L_i=(M)_{i,1}x+(M)_{i,2}y+(M)_{i,3}z+(M)_{i,4}w$.
\item Let $I$ be the ideal generated by $L_1,L_2,L_3,L_4,F,G$ and let $T=V(I)$.
\item For each point in $T(\mathbb{F}_{p^e})$, compute $H\mathbf{v}_P^{(p)}$. If this gives a scalar multiple of $\mathbf{v}$, then $P$ is a Type $2$ Cartier point.
\end{enumerate}
\end{enumerate}
\end{algo}
\subsection{Cartier points on standard curves over $\mathbb{F}_p$}\label{sec:Cartier_points_on_stn_FP}
In this section we present the results from our search of curves in standard form, related to their $a$-number, $p$-rank and Cartier points, both of Type $1$ (T1) and Type $2$ (T2).
\textbf{The upper bound of Type $1$ points: }Our data reflects that, as expected, it is hard to find curves that attain the bounds on the number of Type $1$ Cartier points. For instance when the $a$-number is 1, a curve can have at most one Type $1$ point, but most curves have zero (see Table~\ref{tab:summary_T1}). When $a=2$ the bound is six, but as stated in Corollary~\ref{cor:no_a2_curves_reach_UB_of_T1}, all of the curves in our sample have three or fewer Type $1$ Cartier points.
The bound on Type $1$ points is also six when $a=3$, but here we see a different behavior. This is mainly because in this case, the Type $1$ points come from intersecting the curve with a hyperplane, which guarantees exactly six points, counting with multiplicity. Even though our statistical data does not give a large sample of $a$-number $3$ curves over $\mathbb{F}_{7}$ and $\mathbb{F}_{11}$, we can see in Tables~\ref{tab:method1_curves} and~\ref{tab:summary_T1} that most curves attain the bound, except when $p=3$.
\textbf{The upper bound on Type $2$ points:}
Similarly, it seems unlikely for a curve with $p$-rank $f>0$ to reach the upper bound of Type $2$ points. This is true even when $f=1$ and thus the bound is one.
We can see in Table~\ref{tab:summary_T2} that the majority of curves with $p$-rank $1$ have no such points. The same happens when $f=2$. In this case the bound is six, but we did not find any curve with more than two Type $2$ points, except when $p=5$, where there exist two curves with six Type $2$ points.
It is important to recall that the bound for Type $2$ points in characteristic $3$ is $4$, not $6$, and there are in fact $16$ $p$-rank $3$ curves that reach this bound.
Here we present the overall results from obtained from our data. The details corresponding to each of $p \in \{3,5,7,11\}$ are given in the following subsections.
\begin{cor}\label{cor:no_a2_curves_reach_UB_of_T1}
In our sample of smooth curves, no curve with $a$-number $2$ reaches the bound of six Type $1$ Cartier points. Moreover, the maximum number of Type $1$ Cartier points attained for curves with $a$-number $2$ is three for $p \in\{5,7,11\}$ and two for $p=3$.
\end{cor}
\begin{proof}
See Table~\ref{tab:summary_T1}.
\end{proof}
\begin{cor}\label{cor:summary_T1_bounds}
The bound on the number of Type 1 points is sharp for non-hyperelliptic smooth genus $4$ curves over $\mathbb{F}_p$ when
\begin{itemize}
\item $p\in \{3,5,7,11\}$ and $a=1$.
\item $p\in \{5,7,11\}$ and $a=3$.
\end{itemize}
\end{cor}
\begin{proof}
See Table~\ref{tab:summary_T1}.
\end{proof}
\begin{cor}\label{cor:no_f2_curves_reach_UB_of_T2}
In our sample of smooth curves, no curve with $p$-rank $2$ or $3$ reaches the bound of six Type $2$ Cartier points when $p\in \{5,7,11\}$.
\end{cor}
\begin{proof}
See Table~\ref{tab:summary_T2}.
\end{proof}
\begin{table}[h]
\caption{Summary of Type $1$ points on samples of standard form curves.}\label{tab:summary_T1}
\begin{center}
\begin{tabular}{cccccc}
$a$-number & \#T1 & 3 & 5 & 7 & 11 \\ \hline\hline
1 & 0 & 23971 & 88896 & 86563 & 25205 \\
1 & 1 & 5327 & 3896 & 1918 & 202 \\
\multicolumn{2}{c}{Total} & 29298 & 92792 & 88481 & 25407 \\
\multicolumn{2}{c}{\% that attains UB} & 18.18 \%& 4.20 \%& 2.17 \%& 0.80 \%\\ \hline
2 & 0 & 1690 & 3158 & 1595 & 191 \\
2 & 1 & 3268 & 832 & 262 & 25 \\
2 & 2 & 1673 & 149 & 37 & 1 \\
2 & 3 & 0 & 65 & 5 & 1 \\
\multicolumn{2}{c}{Total} & 6631 & 4204 & 1899 & 218 \\
\multicolumn{2}{c}{\% that attains UB} & 0 \%& 0 \%& 0 \%& 0 \%\\ \hline
3 & 0 & 0 & 0 & 0 & 0 \\
3 & 1 & 660 & 4 & 0 & 0 \\
3 & 2 & 40 & 4 & 0 & 0 \\
3 & 3 & 0 & 14 & 0 & 0 \\
3 & 4 & 0 & 0 & 1 & 0 \\
3 & 5 & 0 & 0 & 0 & 0 \\
3 & 6 & 0 & 24 & 4 & 1 \\
\multicolumn{2}{c}{Total} & 700 & 46 & 5 & 1 \\
\multicolumn{2}{c}{\% that attains UB} & 0 \%& 52.17 \%& 80.00 \%& 100 \%\\ \hline
& Sample size & 186266 & 719102 & 863038 & 361098 \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[h]
\caption{Summary of Type $2$ points on samples of standard form curves. }\label{tab:summary_T2}
\begin{center}
\begin{tabular}{cccccc}
$f$ & \#T2 & 3 & 5 & 7 & 11 \\ \hline\hline
0 & 0 & 1157 & 193 & 44 & 3 \\ \hline
1 & 0 & 4276 & 4272 & 1853 & 239 \\
1 & 1 & 198 & 146 & 35 & 0 \\
\multicolumn{2}{c}{Total} & 4474 & 4418 & 1888 & 239 \\
\hline
2 & 0 & 7137 & 17072 & 12071 & 2320 \\
2 & 1 & 353 & 737 & 235 & 20 \\
2 & 2 & 23 & 35 & 5 & 0 \\
2 & 3 & 0 & 2 & 0 & 0 \\
\multicolumn{2}{c}{Total} & 7513 & 17846 & 12311 & 2340 \\
\hline
3 & 0 & 21951 & 71423 & 74504 & 22837 \\
3 & 1 & 1394 & 3032 & 1597 & 204 \\
3 & 2 & 117 & 113 & 41 & 3 \\
3 & 3 & 14 & 15 & 0 & 0 \\
3 & 4 & 9 & 2 & 0 & 0 \\
\multicolumn{2}{c}{Total} & 23485 & 74585 & 76142 & 23044 \\
\multicolumn{2}{c}{Sample size} & 186266 & 719102 & 863038 & 361098 \\ \hline
\end{tabular}
\end{center}
\end{table}
%
We know by Corollary~\ref{cor:type1_genus4} that the upper bound on the number of Type $1$ Cartier points is given by $2g-2=6$ if $a=2,3$ and $1$ if $a=1$. We want to determine for which $a,f$ and $p$ these bounds are attained. For example, this does not happen for any of the curves with $a=2$ that we found. In Table~\ref{tab:summary_T1}, we break down the number of curves over $\mathbb{F}_p$ by the number of Type $1$ points they have.
On the other hand, the upper bound on the number of Type $2$ points depends on the $p$-rank: it is $\min \{2g-2, \frac{p^f-1}{p-1}\}$. As stated in Corollary~\ref{cor:no_a2_curves_reach_UB_of_T1}, and observed in Table~\ref{tab:summary_T2}, it is unlikely that a curve of $p$-rank $2$ or $3$ reaches the bound of six Type $2$ points.
%
\subsubsection{Case $p=3$}\label{sec:CP_on_GC_p=3}
From the sample of $186266$ curves over $\mathbb{F}_3$ a total of $35472$ have $a$-number $1,2$ or $3$. There are only two instances in which we identified curves that realize the (non-zero) upper bound on Type $1$ and Type $2$ points. These are when $a=1$, for Type $1$ points, and when $f=3$ for Type $2$ points, and the bounds are $1$ and $4$, respectively. Even so, we can observe in Tables~\ref{tab:summary_T1} and ~\ref{tab:summary_T2} that most curves tend to have fewer Cartier points.
\subsubsection{Case $p=5$}
Recall that we sampled a total of $719102$ tuples over $\mathbb{F}_5$ and obtained $97042$ curves of $a$-number $1\leq a \leq 3$.
Only $3896$ out of the $88896$ curves with $a=1$ have a Type $1$ Cartier point. This is expected because for each curve, there is only one point $P$ such that $H\mathbf{v}_P=0$, so it is unlikely for this point to also be on the curve.
\subsubsection{Case $p=7$}\label{subsub:CPsF7}
We sampled a total of $863200$ random tuples, obtaining $619872$ curves in standard form and $90403$ of them with $a$-numbers $1,2$ or $3$. We note that the upper bounds for the number of Type $1$ points are realized when $a=1$ and $a=3$, but not for $a=2$, where the maximum number of points attained is three. With respect to the Type $2$ points, the (non-zero) upper bounds are only attained when the $p$-rank is $1$. In particular, we get the following result:
\begin{cor}\label{cor:bounds_reached_p7}
Baker's bound on the total number of Cartier points for genus $4$ curves with $a=3$ and $p$-rank $1$ is attained over $\mathbb{F}_7$.
\end{cor}
In our data there is only one curve with $a$-number $3$ where both the bounds of Type $1$ and Type $2$ points are attained, we show it in Example~\ref{ex:CPbound_reached_p7_2}. We also include Example~\ref{ex:p7_6T1_points_2}, in which we see a curve with only $4$ Type $1$ points. The Hasse--Witt matrix in each example is computed with the basis given in Proposition~\ref{prop:computation_hw}.
\subsubsection{Case $p=11$}\label{subsub:CPsF11}
The size of the random sample in this case is of $361098$ tuples. Note that, once again, the upper bound for the Type $1$ points is attained for some cases where $a=1,3$. On the other hand, none of the curves realize the bound of Type $2$ points (except, of course when $f=0$). One important observation is that, from all sampled tuples, only one of them resulted in a curve with $a$-number $3$. This curve also achieves the maximum of $6$ Type $1$ points.
\section{Cartier points on curves over $\mathbb{F}_p$ with $a$-number $3$}
In Table~\ref{tab:method1_curves} we display the information of curves in standard form over $\mathbb{F}_p$ for $p \in \{3,5,7\}$, obtained from the exhaustive search. For $p=3,5$ the search was done over all the possible tuples. For $p=7$ we found all the curves in standard form in the degenerate case and all of those in the case N1i when the coefficients of $y^2z$ and $z^2w$ are 0. We denote these lists by 7(D) and 7(N1i'), respectively.
In the second column, we list the total number of isomorphism classes of curves with $a$-number $3$ found for each $p$. The third column indicates the total number of curves that attain the maximum of six Type $1$ Cartier points. The last two columns correspond to the number of curves that reach the maximum of Type $2$ points, over the total with the respective $p$-rank. Notice that all curves with $f=0$ trivially reach this bound, since the bound is $0$.
\begin{table}[h]
\caption{Isomorphism classes in standard form with $a$-number $3$ over $\mathbb{F}_p$. }\label{tab:method1_curves}
\begin{center}
\begin{tabular}{c|c|c|cc}
\multirow{3}{*}{$p$ }&Classes & { Attain max. of T1} &\multicolumn{2}{c}{Attain max. of T2} \\
&$a=3$ &$a=3$&$f=0$&$f=1$\\
& & $(6)$ & $(0)$&$(1)$ \\
\hline
\hline
{3} &27&0&0/0&0/27\\
{5} &134&80&36/36&5/98\\
{7(D)} &65&48&9/9&0/56\\
{7(N1i')} &29&23&2/2&1/27\\
\hline
\end{tabular}
\end{center}
\end{table}
In the case $p=3$ all curves have either one or two Type $1$ points maximum; for $p=5$, the curves show one, two three or six points, and for $p=7$ we saw evidence of curves with any number of Type $1$ points ranging from one to six. In addition, over $\mathbb{F}_5$, every possible degree distribution for curves with six (distinct) Type $1$ points occurs. Also, all except one of them occur among our sample over $\mathbb{F}_7$.
%
%
\subsection{Case $p=3$}
We will now discuss the curves in standard form over $\mathbb{F}_3$ of $a$-number $3$. Since we have a complete list of all of the isomorphism classes of these curves, we know that all of them have $p$-rank $1$.
In Lemma~\ref{lem:standard_form_a3_p3_curves} we specify representatives for the isomorphism classes of curves in standard form over $\mathbb{F}_3$ with $a$-number $3$. It turns out the all the curves from the same kind have the equal multiplicity and degree distribution of Type $1$ points, as a consequence we get the next lemma. We show explicitly the points in Tables ~\ref{table:repres_p3_D},~\ref{table:repres_p3_N1} and~\ref{table:repres_p3_N2}.
\begin{lem}\label{lem:T1_points_in_a3_p3_curves}
A curve $X=V(F,G) $ in standard form with $a$-number $3$ over $\mathbb{F}_3$ has no Type $2$ Cartier points and
\begin{itemize}
\item if $F=F_d=2yw+z^2$, then $X$ has exactly one Type $1$ point of multiplicity 6;
\item if $F=F_1=2xw+2yz$, then $X$ has exactly two Type $1$ points of multiplicity 3, each defined over $\mathbb{F}_3$;
\item if $F=F_2=2xw+y^2+z^2$, then $X$ has exactly two Type $1$ points of multiplicity 3, each defined over $\mathbb{F}_9.$
\end{itemize}
\end{lem}
\begin{proof}
See Tables~\ref{table:repres_p3_D},~\ref{table:repres_p3_N1} and~\ref{table:repres_p3_N2}.
\end{proof}
The proof of Lemma~\ref{lem:T1_points_in_a3_p3_curves} follows from our classification of curves in standard form into isomorphism classes and direct computation of Cartier points. Here we give an overview of the heuristics that go into this procedure, in the degenerate case. The other two cases can be worked out in a similar way.
\begin{example}[Cartier points in the case D]\label{ex:T1_T2_on_D_p3_a3}
Let $X=V(F,G)$ be a curve from Table~\ref{table:repres_p3_D}.
We know that $F=2yw+z^2$ and that $G$ is of the form
$$ G=x^3+y^3+xyz+c_1yz^2+xw^2+c_2w^3,$$
where $c_1 \in \mathbb{F}_3^\times$ and $c_2 \in \mathbb{F}_3$. We will show that $X$ has no Type $2$ Cartier points and that $[2-c_1-c_2:1:2:1]$ is the unique Type $1$ Cartier point.
By Proposition~\ref{prop:computation_hw}, the Hasse-Witt matrix of $X$ is
\begin{equation}
H= \begin{bmatrix}
0& 0& 0 &0\\
0&1 & 2 & 1\\
0& 0 & 0 & 0\\
0&0 & 0&0 \\
\end{bmatrix}.
\end{equation}
We can see right away that $X$ has no Type $2$ Cartier points, since, up to scalar multiplication, the only vector such that $H\mathbf{v}=c\mathbf{v}$ for some $c\in \mathbb{F}_3$ is \linebreak $\mathbf{v}=(0,1,0,0)^T$, but these are not the coordinates of a point on $X$.
Now, the hyperplane $S$ generated by the kernel of $H$ is the zero locus of $L= y+2z+w$. Then by substituting $y=z-w$ in $F=0$ we get $z^2+2zw+w^2=0$. If $w=0$, then $z=y=0$, but $[1:0:0:0] \not\in X$, so assume $w=1$. Then $z=2$ and $y=1$ and the Type $1$ points of $X$ are of the form $P=[\alpha:1:2:1]$. To find $\alpha $ we evaluate $G$ at $P$ and get $\alpha =2-c_1-c_2$. So there is a unique Type $1$ point, which must have multiplicity $6$.
\end{example}
In Tables~\ref{table:repres_p3_D},~\ref{table:repres_p3_N1} and~\ref{table:repres_p3_N2} we show the cubic polynomial for a representative of each isomorphism class in the cases D, N1 and N2, respectively. The curves are of the form $X=V(F,G)$. We also specify the Type $1$ Cartier points.
\begin{table}[h]
\caption{Isomorphism classes of D curves with $a=3$ over $\mathbb{F}_3$.}
\label{table:repres_p3_D}
\begin{center}
\begin{tabular}{cc}
Representative $G$ & Type 1 point\\
\hline
\hline
$x^3 + y^3 + xyz + yz^2 + xw^2$ & $[ 1 : 1 : 2 : 1] $\\
$x^3 + y^3 + xyz - yz^2 + xw^2$ & $ [0 : 1 : 2 : 1]$ \\
$x^3 + y^3 + xyz + yz^2 + xw^2 + w^3$ & $[0 : 1 : 2 : 1]$\\
$x^3 + y^3 + xyz - yz^2 + xw^2 + w^3$ & $[2 : 1 : 2 : 1]$\\
$x^3 + y^3 + xyz + yz^2 + xw^2 - w^3$ & $[2 : 1 : 2 : 1]$\\
$x^3 + y^3 + xyz - yz^2 + xw^2 - w^3$ & $[1 : 1 : 2 : 1 ]$\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[h]
\caption{Isomorphism classes of N1 curves with $a=3$ over $\mathbb{F}_3$.}
\label{table:repres_p3_N1}
\begin{center}
\begin{tabular}{cl}
Representative $G$ & Type 1 points \\
\hline
\hline
$x^2y + y^3 + x^2z + y^2w + z^2w - w^3$ &$[ 1 : 2 : 1 : 1],[ 0 : 0 : 2 : 1 ]$\\
$x^2y + y^3 + x^2z - y^2w + z^2w - w^3$ &$[ 0 : 2 : 0 : 1], [0 : 0 : 1 : 1 ]$\\
$x^2y + y^3 + x^2z + z^3 + y^2w + z^2w + w^3$ & $[ 1 : 2 : 1 : 1],[ 1 : 1 : 2 : 1 ]$\\
$x^2y + y^3 + x^2z + z^3 - y^2w + z^2w + w^3$ & $[0 : 0 : 1 : 1], [2 : 2 : 2 : 1] $\\
$x^2y + y^3 + x^2z - z^3 + y^2w + z^2w + w^3$ & $[0 : 0 : 2 : 1],[ 1 : 0 : 1 : 0] $\\
$x^2y + y^3 + x^2z - z^3 + y^2w + z^2w - w^3$ & $[1 : 1 : 2 : 1],[ 1 : 0 : 1 : 0] $\\
$x^2y + y^3 + x^2z - z^3 - y^2w + z^2w + w^3$ & $[2 : 1 : 1 : 1], [1 : 0 : 1 : 0] $\\
$x^2y + x^2z + y^2w + z^2w + w^3$&$[1 : 2 : 1 : 1],[ 1 : 1 : 2 : 1] $\\
$x^2y + x^2z + y^2w + z^2w - w^3$&$[ 0 : 2 : 0 : 1], [0 : 0 : 2 : 1 ]$\\
$x^2y + x^2z - y^2w + z^2w + w^3$& $ [0 : 2 : 0 : 1], [2 : 1 : 1 : 1] $\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[h]
\caption{Isomorphism classes of N2 curves with $a=3$ over $\mathbb{F}_3$, with $\beta^2 + 2\beta+ 2=0$.}
\label{table:repres_p3_N2}
\begin{center}
\begin{tabular}{cl}
Representative $G$ & Type 1 points \\
\hline
\hline
$ x^2y + y^3 + x^2z - z^3 - yzw - w^3$&$[\beta^5 : 1 : \beta^3 : 1], [\beta^7 : 1 : \beta : 1]$ \\
$x^2y + x^2z - yzw - w^3$&$ [0 : \beta^5 : \beta^7 : 1], [0 : \beta^7 : \beta^5 : 1] $\\
$x^2y + y^3 + x^2z - z^3 - yzw + w^3$&$ [\beta^2 : 0 : \beta : 1], [\beta^6 : 0 : \beta^3 : 1] $\\
$x^2y + x^2z - z^3 - yzw + w^3$&$ [\beta^2 : 0 : \beta : 1], [\beta^6 : 0 : \beta^3 : 1] $\\
$x^2y +y^3 + x^2z + z^3 - yzw + w^3$&$ [1 : \beta^2 : \beta^6 : 1], [1 : \beta^6 : \beta^2 : 1] $\\
$x^2y - y^3 + x^2z - z^3 - yzw - w^3$&$ [\beta^7 : \beta^2 : 1 : 0], [\beta^5 : \beta^6 : 1 : 0] $\\
$x^2y + x^2z + z^3 - y^2w - yzw + z^2w + w^3$&$ [2 : 0 : \beta^2 : 1], [2 : 0 : \beta^6 : 1] $\\
$x^2y - y^3 + x^2z - y^2w - yzw + z^2w - w^3$&$ [\beta^7 : \beta^5 : 1 : 1], [\beta^5 : \beta^7 : 1 : 1] $\\
$x^2y - x^2z - y^2w + yzw + z^2w + w^3$&$ [\beta^2 : \beta^3 : \beta^7 : 1], [\beta^6 : \beta : \beta^5 : 1] $\\
$x^2y + x^2z - z^3 - y^2w - yzw + z^2w + w^3$&$ [\beta^2 : \beta^3 : \beta^7 : 1], [\beta^6 : \beta : \beta^5 : 1] $\\
$x^2y - x^2z - z^3 - y^2w + yzw + z^2w - w^3$&$ [0 : 2 : \beta^2 : 1], [0 : 2 : \beta^6 : 1] $\\
\hline
\end{tabular}
\end{center}
\end{table}
%
%
%
\subsection{Case $p=5$}\label{sec:exhaustive_p5}
We found the set of Cartier points on all curves in standard form over $\mathbb{F}_5$ with $a$-number $3$. There are some behaviors that are different from the case $p=3$. For instance, there are isomorphism classes of curves that attain the bound on the number of Cartier points, both with $f=1$ and $f=0$. It also happens that some of them have Type $2$ Cartier points (see Table~\ref{table:p=5GoodCurves_number_of_CPs} and Proposition~\ref{prop:bound_onCPs_p5}).
Another feature is that every degree distribution occurs when there are six distinct Type 1 points.
%
\begin{table}[h]
\caption{Number of Type $1$ and $2$ Cartier points over $\mathbb{F}_5$.}
\label{table:p=5GoodCurves_number_of_CPs}
\begin{center}
\begin{tabular}{c|cccccc|c}
Case & \multicolumn{6}{c|}{Curves with $n$ Type $1$ points} & Curves with Type $2$ \\
& $1$ &2 &3 &4 & 5 & 6 &1 \\
\hline \hline
D & 14 & 0 &26 & 0 &0 & 19& 0 \\
N1i & 3 & 7 & 0 & 0 &0 &50 & 4 \\
N1ii & 0 & 1& 0 & 0 &0 &5 &1 \\
N2 &2 & $1$ &0 & 0 & 0&6 &0 \\
\hline
Total& 19&9&26&0&0&80&5\\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{example}\label{ex:T2CPs_p5}
There are only $5$ isomorphism classes of curves in standard form with $a$-number $3$ over $\mathbb{F}_5$ that have a Type $2$ Cartier point. Here we show the cubic polynomial of each curve, together with the Type $2$ point. All of these curves have only one Type $1$ Cartier point, which implies there are no $p$-rank $1$ curves over $\mathbb{F}_5$ that reach the bound of seven total Cartier points.
\begin{itemize}
\item $x^2y + y^3 + x^2z + 2xyz + 2yz^2 + z^3 + 2y^2w + 2yzw + yw^2 + w^3$, $[1:0:3:0]$.
\item $2x^2y + y^3 + x^2z + xyz - yz^2 - z^3 + 2yzw - 2yw^2 + zw^2 - w^3 $, $[0:0:1:2]$.
\item $2x^2y + y^3 + 2x^2z + xyz + y^2z - 2yz^2 - 2z^3 + 2y^2w - yzw + yw^2 - 2zw^2 - 2w^3 $, $[1:3:2:4]$.
\item $ x^2y + y^3 - 2x^2z + y^2z - yz^2 - y^2w - yzw + z^2w - yw^2 - 2zw^2 - 2w^3$, $[1:2:0:0]$.
\item $x^2y + 2x^2z - 2xyz + y^2z + yz^2 - 2y^2w + 2yzw + z^2w + yw^2 - 2zw^2 - 2w^3$, $[1:4:1:1]$.
\end{itemize}
\end{example}
\begin{prop}\label{prop:bound_onCPs_p5}
There are no curves in standard form over $\mathbb{F}_5$ with $p$-rank $1$ and $a$-number $3$ that reach Baker's bound of seven Cartier points.
There are, up to $\mathbb{F}_5$-isomorphism, $17$ curves in standard form with $p$-rank $0$ and $a$-number $1$ that reach the bound of six Type $1$ Cartier points.
\end{prop}
\begin{proof}
If a genus $4$ curve with $a=3$ has $p$-rank $0$ then by Proposition~\ref{prop:bound_cartier_points}, it has at most six Cartier points. If the $p$-rank is $1$, then the curve has at most seven Cartier points: six of Type $1$ and one of Type $2$. We use Algorithms~\ref{alg:Type1_points} and~\ref{alg:Type2_points} to compute the Cartier points on all N1, N2 and D curves with $a=3$ over $\mathbb{F}_5$ and find that only seventeen curves with $p$-rank $0$ have a total of six Cartier points. Example~\ref{ex:T2CPs_p5} shows that the only curves with Type $2$ Cartier points have fewer than six Type $1$ points, hence the total of Cartier points is less than the upper bound for all the $p$-rank $1$ curves.
\end{proof}
\subsection{Case $p=7$}
As mentioned in Section~\ref{sec:goodcurvesp=7}, for $p=7$ we also computed all the $a$-number $3$ curves of type D and of type N1i with $(b_1,b_2)=(0,0)$. Table~\ref{table:p=7GoodCurves_number_of_CPs} shows the number of Cartier points on these curves.
\begin{table}[h]
\caption{Number of Cartier points on known D and N1 genus $4$ curves over $\mathbb{F}_7$.}
\label{table:p=7GoodCurves_number_of_CPs}
\begin{center}
\begin{tabular}{c|cccccc|c}
Case & \multicolumn{6}{c|}{\#Curves with $n$ Type $1$ points} & \#Curves with Type $2$ \\
& $1$ &2 &3 &4 & 5 & 6 &1 \\
\hline \hline
D & 0& 0 &3 & 9 &5 & 48& 0 \\
N1i'& 0 & $1$ &0 & 3 &2 &23 & $1$ \\
\hline
Total&0&1&3&12&7&71&1\\
\hline
\end{tabular}
\end{center}
\end{table}
%
\section{Examples}\label{sec:examples}
In this section we show five examples where we explicitly compute all the Cartier points of curves (or families of curves) with genus $4$.
\begin{example}\label{ex:CPbound_reached_p7_2}
Let $X=V(F,G)$ be a genus $4$ curve over $\mathbb{F}_7$ where
\begin{align*}
F&=2yz+2xw,\\
G&=2x^2y + y^3 + x^2z + y^2z + 3z^3 + 2yzw + z^2w + 4yw^2 + 6zw^2 + 4w^3.
\end{align*}
Notice that $X$ belongs to the N1 case from Definition~\ref{def:standardform}. We will see how this curves attains Baker's bound on Cartier points. First, the Hasse--Witt matrix of $X$ is
\begin{align*}
\begin{bmatrix}
2&3&6&1\\
0&0&0&0\\
0&0&0&0\\
0&0&0&0\\
\end{bmatrix}.
\end{align*}
Thus the $a$-number is $3$ and the $p$-rank $1$, in which case the total bound on the number of Cartier point is $7$. By solving $H\mathbf{v}=c\mathbf{v}$ we see that the Type $2$ point is $[1:0:0:0]$.
Also, the Type $1$ points are those in the intersection of $X$ and the hyperplane $2x+3y+6z+w=0$. There are two such points defined over $\mathbb{F}_7$, they are $[0 : 0 : 1 : 1]$ and $[2 : 4 : 3 : 1]$. The other four points are two pairs of $\sigma$-conjugate defined over $\mathbb{F}_{49}$ which are
$[6 : \alpha^{26} : \alpha^{22} : 1], [6 : \alpha^{38} : \alpha^{10} : 1]$ and
$[4 : \alpha^{33 }: \alpha^{23} : 1], [4 : \alpha^{39 }: \alpha^{17 }: 1]$, for $\alpha$ such that $\alpha^2 + 6\alpha + 3=0$.
\end{example}
\begin{example}\label{ex:p7_6T1_points_2}
Let $X=V(F,G)$ be a genus $4$ curve over $\mathbb{F}_7$ where
\begin{align*}
F&= y^2 + 4z^2 + 2xw.\\
G&= 3x^2y + 4xy^2 + 3y^3 + 5y^2z + 2xz^2 + yz^2 + 5z^3 + y^2w + 6yzw + 5zw^2 + 2w^3.
\end{align*}
The Hasse--Witt matrix of $X$ is
\begin{align*}
H=\begin{bmatrix}
1 &4 &1 &6\\
5& 6& 5& 2\\
0& 0& 0& 0\\
1& 4& 1& 6\\
\end{bmatrix}.
\end{align*}
In this case, $X$ has $p$-rank $1$, but the only eigenvector of $H$, up to scaling is $(1, 5, 0, 1)^T$, and this does not give a point on $X$. On the other hand, there are four Type $1$ points: $[0 : \alpha^{14} : \alpha^{34 }: 1]$,$ [\alpha^7 : \alpha^{36 }: \alpha^{34} : 1]$, $[0 : \alpha^2 : \alpha^{46} : 1]$, $[\alpha : \alpha^{12} : \alpha^{46} : 1]$, where $\alpha^2 + 6\alpha + 3=0$. The first two have multiplicity one and the others have multiplicity 2.
\end{example}
\begin{example}\label{ex:CP_on_p11_curve}
Let $X=V(F,G)$ be a genus $4$ curve over $\mathbb{F}_{11}$ where
\begin{align*}
F=&z^2 + 2yw,\\
G=&9x^3 + xy^2 + 4y^3 + 9xyz + 2xz^2 + 8z^3 + 7xzw + 8xw^2 + zw^2 + 3w^3.
\end{align*}
The Hasse--Witt matrix of $X$ is
\begin{align*}
H=\begin{bmatrix}
1 & 5& 3& 10\\
7 & 2& 10& 4\\
2 &10 & 6 & 9\\
4 & 9 & 1& 7
\end{bmatrix}.
\end{align*}
This curve has $a$-number $3$ and $p$-rank $1$. There are no Type $2$ points. Indeed, the only solution, up to scalar multiplication of $H\mathbf{v}=c\mathbf{v}$ is $\mathbf{v}=(1, 7, 2, 4)^T$, and this does not give a point over $X$.
The Type $1$ points are those on the intersection of $X$ and the hyperplane with equation $x+5y+3z+10w=0$, that is $ [6: 9: 9: 1]$,
$[ 7 +2\alpha +7\alpha^2 +9\alpha^3 +10\alpha^4: 6 +4\alpha +5\alpha^2 +4\alpha^3 +1\alpha^4:10 +4\alpha^2 +5\alpha^3 +6\alpha^4:1]$ with $\alpha^5+10\alpha^2+9=0$, and its four conjugate points.
\end{example}
\begin{example} \label{ex:GCd5-301_a4=0}
Let $X$ be the genus $4$ curve over $\mathbb{F}_{5}$ embedded in $\mathbb{P}^3$ as the zero locus of
\begin{equation} \label{eq:example_GCd5-301_a4=0}
F=2yw+z^2, \text{ }
G=a x^3+xw^2 +b y^3+c w^3+zw^2,
\end{equation}
with $a, b \in\mathbb{F}_{5}^\times$ and $c \in \mathbb{F}_{5}$.
By Proposition~\ref{prop:computation_hw} the Hasse-Witt matrix of $X$ is
%
\begin{equation}\label{eq:HWmatrix_ex_GCd5-301_a4=0}
H= \begin{bmatrix}
0& 0& 0 &0\\
0&3ab^2& 4ab & 4a\\
0& 0 & 0 & 0\\
0& 0& 0& 0\\
\end{bmatrix}.
\end{equation}
%
If $H\mathbf{v}=c\mathbf{v}$ for some $c\ne 0$, then $\mathbf{v}$ must be in the subspace spanned by $(0,1,0,0)^T$. But $[0:1:0:0]$ is not a point on $X$. Therefore, $X$ has no Type $2$ point.
Now, suppose $P=[x_0,y_0,z_0,w_0]$ is a Type $1$ Cartier point of $X$, and let $\mathbf{v}=\mathbf{v}_P$. Since $P$ is also on the hyperplane generated by the kernel of $H$, then $3b^2y_0+4bz_0+4w_0=0$, from where $w_0=3b^2y_0+4bz_0$. By evaluating $F$ at $P$ we see that
\begin{equation*}
b^2y^2+3byz+z^2=0,
\end{equation*}
hence $z_0=2by_0$. Note that $y_0\ne 0$, because otherwise $P=[1:0:0:0]$, and this is not a point on $X$. So we can assume that $y_0=1$, and then $P=[x_0:1:2b:b^2]$. Evaluating $P$ at $G$ we get that
\begin{equation*}
ax_0^3+x_0+3b+cb=0.
\end{equation*}
Therefore, $X$ has $1, 2$ or $3$ Type $1$ Cartier points (counting without multiplicity), one for each of the roots of $ax^3+x+3b+cb$, that can be defined over $\mathbb{F}_5$, $\mathbb{F}_{5^2}$ or $\mathbb{F}_{5^3}$.
\end{example}
\begin{example} \label{ex:GCd5-301_a4}
Let $X$ be the genus $4$ curve over ${\mathbb{F}}_{5}$ embedded in $\mathbb{P}^3$ as the zero locus of
\begin{equation} \label{eq:example_GCd5-301_a4}
F=2yw+z^2, \text{ }
G=3a^2 x^3+xw^2 +bxyz+b^2a y^3+a w^3+zw^2,
\end{equation}
with $a, b \in\mathbb{F}_{5}^\times$. The curve $X$ is smooth and irreducible if $a\ne b$. By Proposition~\ref{prop:computation_hw} the Hasse-Witt matrix of $X$ is
\begin{equation}\label{eq:HWmatrix_ex_GCd5-301_a4}
H= \begin{bmatrix}
0& 0& 4ab^3 &4b\\
0& 0 & 2a^3b^2 &2a^2 \\
0& 0 & 0 & 0\\
0& 0& 4b^3& 4a^3b\\
\end{bmatrix}.
\end{equation}
So $X$ has $a$-number and $p$-rank $1$. There are no points on $X$ such that the corresponding vector $\mathbf{v}$ is a solution to $H\mathbf{v}^{(5)}=c\mathbf{v}$ for $c$ in ${\mathbb{F}}_5^{\times}$.
Now, suppose that $P=[x_0,y_0,z_0,w_0]$ is a Type $1$ point. Then $P$ is in the hyperplane given by $ab^2z+w=0$, hence $w_0=4ab^2z$. If $z_0=0$, then $w_0=0$ and since $G(P)=0$, we have $3a^2x_0^3+ab^2y_0^3=0$. Clearly $x_0$ and $y_0$ cannot be zero, so let $y_0=1$ and then $P=[x_0:1:0:0]$ for $x_0$ a root of $3ax^3+b^2=0$. This gives a Type 1 point over $\mathbb{F}_5$ and two over $\mathbb{F}_{25}$.
On the other hand, if we assume $z_0=1$ then $w_0=4ab^2$ and by substituting $P$ in $F=0$ we obtain $y_0=3a^3b^2$. Therefore $P=[x_0:3a^3b^2:1:4ab^2]$, with $x_0$ a root of $3a^2x^3+x(a^2+3a^3b^3)+3a^2+4b^2$. For all possible $(a,b)$, this polynomial has three distinct roots, either over $\mathbb{F}_5$, $\mathbb{F}_{25}$ or $\mathbb{F}_{125}$.
\end{example}
\bibliographystyle{apa}
| b7eab9b50048bd2e0003409adf8b05d30cd7b0ec | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Lasers are nowadays well-established as the workhorse of modern telecommunication, metrology, as well as developing quantum technologies. However, as has been already realized by Schawlow and Townes \cite{PhysRev.112.1940}, the phase noise of a laser is finite and fundamentally limited. Practically, many more effects, such as thermal \cite{NUMATA2012798} or acoustical \cite{CONFORTI2020126286} noise of laser cavities, contribute to the total phase noise.
Many techniques are employed to reduce phase noise, and thus the linewidth of lasers. To great success, external ultra-stable cavities have been employed as references \cite{Kessler2012,PhysRevA.77.053809}, achieving linewidths well below \SI{1}{\Hz}, but broadband phase noise remains a problem in many applications. In contrast to intensity noise, which usually exhibits localized peaks due to relaxation-oscillation \cite{8123620}, phase noise in lasers tends to exhibit both a broadband noise floor and technical noise peaks.
The broadband noise can be either suppressed passively via filtration through a narrow cavity, or using an active feedback signal. As an alternative method to reference cavities, an unbalanced Mach-Zehnder interferometer with an optical fiber delay line in one of its arms can be used. Such setups are routinely employed to characterize laser linewidths \cite{LUDVIGSEN1998180}. One can then use the signal of such a delay-line setup to feedback on the laser light's phase, thereby reducing its noise \cite{Li:17,Chen:89,Cranch:02,Sheard:06,Kefelian:09,Dong:15, _m_d_2015,Shehzad:19,doi:10.1063/1.3606439,Bandutunga:20}. All previous approaches operated in a relatively low-frequency range, starting from very low infrasonic frequencies up to maximally around \SI{100}{\kHz}. This has so far satisfied most needs of spectroscopy experiments, and implementing feedback at higher frequencies becomes challenging.
For a delay-line setup, fiber noise becomes an important matter to consider. At low frequencies, acoustic isolation of fibers has been successfully implemented \cite{Bandutunga:20}, but at higher frequencies the thermomechanical and thermoconductive noises are unavoidable \cite{PhysRevA.86.023817}. Fiber noise also poses limitations on transfer of optical frequency standards, and thus a reverse approach is often employed in which the fiber is stabilized to a narrowband stable laser \cite{Ma:94,PhysRevLett.111.110801}. Several works have also studied fundamental limits to fiber noise and related fiber strain sensing \cite{doi:10.1063/1.4939918,Gagliardi1081,Cranch286}.
Here we employ a 50-meter-long fiber delay line combined with a balanced detection scheme to measure laser phase noise and subsequently use active feedback to reduce it at high frequencies. In particular we are interested in reducing phase noise at frequencies in the vicinity of \SI{1.5}{\MHz} that correspond to the resonance frequency of a membrane mechanical oscillator \cite{Tsaturyan2017,ourOptica}. Using a feedback loop, with the gain concentrated around the frequency of interest, we achieve phase noise of \SI{-164}{\deci\bel(\radian^2/\hertz)} at 1.5 MHz frequency offset of a Titanium-Sapphire laser (M-Squared SolsTiS), providing noise reduction in a previously untackled frequency regime with a very high absolute bandwidth of around 300 kHz, compared with hitherto approaches that target up to 100 kHz bandwidth at baseband. The results are enabled by low-noise detection at high light powers, combined with an optimized fiber length and properly engineered feedback.
Noise at such high frequency offsets is an essential limitation in quantum optomechanics, where the sideband at the resonance frequency of the oscillator leads to significant classical drive forces \cite{PhysRevLett.123.153601,Kippenberg_2013,Safavi_Naeini_2013}. We envisage that using this light for feedback cooling of membrane resonators \cite{Tsaturyan2017,Rossi2018,ourOptica} in a sideband-resolved cavity quantum optomechanics regime will allow to break an important barrier of 0.1 residual occupation of phonons of a membrane resonator mode at liquid helium temperatures ($\sim \SI{4}{\kelvin}$).
The paper is organized as follows. We first introduce the experimental setup, which involves two similar unbalanced interferometers, for feedback and characterization. Next, we introduce the model for detection of the laser's phase noise, including a treatment of spurious fiber noise and photon shot noise, as well as a model for feedback. Finally, we show results for noise suppression in several MHz-level bands and compare them with our model prediction.
\section{Setup}
\label{sec:setup}
\begin{figure}[h!]
\centering\includegraphics[width=1\linewidth]{img/phase_noise_loop_schematic.pdf}
\caption{Schematic of the delay line feedback loop (setup 1) and an additional loop for out-of-loop measurement (setup 2). At the input, the interferometers feature a fiber polarization-maintaing 50:50 beamsplitter, while at the output the beams are combined via free-space polarization optics and sent onto a balanced photodetector. Rotating the half-wave plate ($\lambda/2$) allows fine adjustement of balanace on the detector. Both interferometers are independently stabilized with piezo mirrors at low frequencies, such that the detection remains balanced. Feedback to the laser light is applied from setup 1 via an electronic controller through a free-space EOM. An additional fiber EOM is used for absolute calibration of setup 2.}
\label{fig:exp_schematic}
\end{figure}
As presented in Fig. \ref{fig:exp_schematic}, our test experiment consists of two unbalanced Mach-Zehnder interferometers with the short arm being $~\SI{1}{\meter}$-long and the long arm being \SI{50}{\meter}-long. All fiber in our setup is polarization maintaining (PM), single-mode fiber (PM-780HP), which allows us to maintain good interferometric visibility and polarization stability. We inject equal amounts of light into both arms using a 50:50 fiber beamsplitter. In each interferometer we place an additional piezo-actuated mirror in one of the arms for stabilization of low-frequency phase drifts in order to lock the interferometer at the optimal (balanced) point for sensing phase noise. The beams are combined on a polarizing beamsplitter (PBS), and then sent through an additional half-wave plate and another PBS onto a balanced detector.
Our custom detectors are based on two Hamamatsu S5971 silicon photodiodes connected in series in a differential configuration \cite{TheJurgen}. The photodiodes themselves feature a high quantum efficiency of $\eta\approx0.9$, which is combined with separate amplification paths for DC and AC ($> \SI{150}{\kHz}$) components of the signal. The overall transimpedance gains are \SI{130}{\kV/\A} and \SI{30}{\kV/\A} for AC/DC respectively. The low-frequency DC output is used to actively stabilize the interferometer with the piezo-actuated mirror. This ensures that the detector always operates in the balanced regime, rejecting the amplitude quadrature of the incoming light to better than \SI{20}{\dB} and preventing the amplifiers from saturating. Due to large optical powers used, we maximize the size of the beam at the photodiodes to prevent temporary "bleaching" of the photodiodes by excessive intensity.
In setup 1, the high-frequency AC output is sent to a STEMlab Red Pitaya 125-14 board equipped with PyRPL software \cite{8087380}, which implements the feedback filter in a field-programmable gate array (FPGA), and drives a free-space EOM (electro-optic modulator) through an additional \SI{20}{\dB} attenuator to minimize electronic and quantization noises. The signal is also sent to an FFT spectrum analyzer in order to measure in-loop noise.
The feedback is achieved using an I/Q modulator/demodulator module in the RedPitaya/PyRPL FPGA architecture. We first demodulate the error signal with a carrier wave at a desired frequency $f_\mathrm{c}$. Subsequently, we apply a lowpass filter with a bandwidth $\gamma$ independently to both quadratures and multiply both signals by gain factor $g$. Finally, the signal is modulated with the same carrier wave $f_\mathrm{c}$ shifted in phase by $\vartheta$. Using the same carrier wave at demodulation and modulation leads to cancellation of unwanted carrier phase noise. Stability is guaranteed since the entire scheme is implemented digitally in the FPGA. The phase difference of the carrier wave between demodulator and modulator determines the phase $\vartheta$ of our feedback, which is empirically optimized. Overall, the filter is shaped like a Lorentzian centered around $f_\mathrm{c}$ with a bandwidth $\gamma$. Importantly, it allows us to continuously tune the phase of the feedback at $f_c$, thus compensating the phase lag introduced due to a delay.
In setup 2, we send the AC signal to a spectrum analyzer for out-of-loop noise measurement. It uses an additional fiber EOM which can be used for absolute phase-noise calibration as it features a flat, calibrated response to phase modulation \cite{GorodetksyDOVOCRUFNC}.
Our laser operates at $\lambda=\SI{852}{\nano\meter}$ and outputs approximately \SI{200}{\mW} of light power. We send \SI{120}{\mW} through the free-space EOM for the entire experiment \cite{ourOptica}, part of which is sent to both fiber delay line setups. In setup 1, we obtain $\bar{P}=\SI{11}{\mW}$ of total power impinging on the balanced detector, which allows for fiber-noise limited measurement of laser phase noise, increasing the signal-to-shot-noise ratio. This allows for the fiber noise to be larger than shot noise until up to 2 MHz. Furthermore, we minimize the beam path between setup 1 and the laser, which is less than 3 meters and includes only two mirrors. For setup 2, we use $\bar{P}=\SI{5}{\mW}$ of total power at the detector. Setup 2 is connected using 5 meters of additional fiber. Any noise added in propagation will be treated as an additional detection noise.
\section{Model}
\label{sec:model}
\subsection{Detection}
We consider input laser light with a fluctuating phase given by $\varphi(t)$. The phase of the field at the output of the long delay is given by $\varphi(t-\tau) + \varphi_f(t)$, where $\varphi(t-\tau)$ is the laser phase noise signal delayed due to propagation in the long arm by a constant time $\tau$ and $\varphi_f(t)$ is the noise introduced by the fiber. At the output of the short arm we simply have $\varphi(t)$. At the output of the interferometer we will measure the phase difference given by $\varphi(t)-\varphi(t-\tau)-\varphi_f(t)$.
The (frequency-domain) transfer function for the laser phase noise is therefore given by:
\begin{equation}
G(f)=1-e^{-2\pi i \tau f},
\end{equation}
with delay $\tau=nL/c$, introduced by the fiber of length $L$ with refractive index $n$ and the speed of light in vacuum $c$.
The fiber noise $\varphi_f(t)$ is added independently as a detector noise term, together with shot noise. Active stabilization of the interferometer arms' relative phase preserves power balance, such that we can assume a total power $\bar{P}$ impinging on the balanced differential photodetector to be equally split between the individual diodes. In this configuration, the total power spectral density (PSD) of the registered optical signal is:
\begin{equation}
S=\eta \bar{P}^2 |G(f)|^2 S_{\varphi\varphi} + \eta \bar{P}^2 S_{\varphi_f\varphi_f} + 2 h \nu \bar{P},
\end{equation}
where $\eta$ is the quantum efficiency of the detector.
The first term represents the laser phase noise and the second term represents fiber noise.
Finally, photon shot noise PSD is given by $2h\nu\bar{P}$ with $\nu=c/\lambda$ and $h$ being the Planck's constant.
Taking the transduction to/from phase noise into account, we may express the noise in terms of equivalent laser phase noise:
\begin{equation}
S_{\varphi\varphi}^\mathrm{meas} = \eta^{-1} \bar{P}^{-2} |G(f)|^{-2} S = S_{\varphi\varphi} + |G(f)|^{-2} S_{\varphi_f\varphi_f} + |G(f)|^{-2} \frac{2 h \nu}{\eta\bar{P}}
\label{eq:phimeas}
\end{equation}
From this, we quite clearly observe that using a high optical power $\bar{P}$ diminishes the shot-noise contribution. The signal to noise ratio will indeed scale as $\sim1/(1+2h\nu/\eta\bar{P})$, therefore it is clearly beneficial to use as high power as realistically available.
For the sake of this model, let us assume that the two delay line setups (in-loop and out-of-loop) have identical powers and efficiencies. Conveniently, we can observe the laser phase noise directly in the cross spectral density (CSD) between the two setups:
\begin{equation}
S^{12} = \eta \bar{P}^2 |G(f)|^2 S_{\varphi\varphi}.
\label{eq:csd}
\end{equation}
This is possible as the laser phase noise is the only correlated noise shared between the two setups. In practice, we use this method to identify the laser phase noise contribution to total measured noise.
\subsection{Fiber noise}
In order to estimate the noise of the fiber, we use the theory of Duan from Ref. \cite{PhysRevA.86.023817}. The theoretical model leads to a conclusion that at our frequencies of interest the noise is heavily dominated by the thermoconductive noise, i.e. fluctuations of temperature transduced to phase fluctuations via thermal expansion and temperature dependence of the refractive index. The expected fiber noise is:
\begin{equation}
S_{\varphi_\mathrm{f}\varphi_\mathrm{f}} = \frac{2 k_B T^2 L }{\lambda^2 \kappa} \left(\frac{\mathrm{d}n}{\mathrm{d}T} + n \alpha\right) \frac{}{} \mathrm{Re}\left[\exp(2\pi i f r_0^2/2D) E_1(2\pi i f r_0^2/2D)\right],
\label{eq:fn}
\end{equation}
where $k_B$ is the Boltzmann constant, $T$ is temperature, $\kappa$ is thermal conductivity, $\frac{\mathrm{d}n}{\mathrm{d}T}$ is thermo-optic coefficient, $\alpha$ is coefficient of linear expansion, $D$ is thermal diffusivity, $r_0$ is mode power profile radius and $E_1$ represents the exponential integral.
\subsection{Feedback}
We now introduce feedback, such that in the Fourier domain:
\begin{equation}
\check{\varphi} \rightarrow \check{\varphi} - H(f) G(f) \check{\varphi} - H(f) \check{\varphi}_f - H(f) \eta^{-1/2} \bar{P}^{-1} \check{\zeta},
\end{equation}
where $\check{\zeta}$ represents shot noise with PSD of $2h\nu\bar{P}$, while $\check{\varphi}$ and $\check{\varphi}_\mathrm{f}$ represent stochastic laser and fiber noises. The feedback transfer function $H(f)$ shall include both our desired feedback, as well as an undesired but unavoidable delay given by a prefactor $e^{-2\pi i \tau_\mathrm{D} f}$, mostly coming from the electronic processing delay ($\tau_\mathrm{D}\approx\SI{250}{\nano\second}$).
The feedback acts on light "shared" by both interferometers. In this case, setup 1 is the in-loop setup used for feedback, while setup 2 serves for independent out-of-loop characterization.
As a result of the feedback, the actual laser phase noise becomes:
\begin{equation}
S_{\varphi\varphi}^{\mathrm{fb}} = \frac{S_{\varphi\varphi} + |H(f)|^2 \left(S_{\varphi_f\varphi_f}^1 + \frac{2h\nu}{\eta \bar{P}}\right)}{|1+G(f)H(f)|^2}
\end{equation}
We see that the original phase noise is suppressed, but new noise is added due to detection and fiber noise in setup 1 ($S_{\varphi_f\varphi_f}^1$). This directly allows us to identify the out-of-loop detector (setup 2) noise $S_{\varphi\varphi}^{\mathrm{fb,meas}}$ by substituting $S_{\varphi\varphi}^{\mathrm{fb}}$ into Eq. \ref{eq:phimeas} with uncorrelated fiber and shot noise.
For the in-loop detector (setup 1), however, the same detection noise becomes present in the feedback, which leads to interference. We obtain the following measured noise:
\begin{multline}
S_{\varphi\varphi}^\mathrm{1,fb,meas} = \frac{S_{\varphi\varphi}}{|1+G(f)H(f)|^2} + \\ + \left[\frac{|H(f)|^2}{|1+G(f)H(f)|^2} + \frac{1}{|G(f)|^2} \left(1 - 2 \mathrm{Re} \frac{G(f)H(f)}{1+G(f)H(f)}\right)\right]\left( S_{\varphi_f\varphi_f}^1 + \frac{2h\nu}{\eta \bar{P}} \right),
\end{multline}
where the first term is the expected reduced laser noise, while the second term represents the detection noise itself as well as its self-interference. In particular, the measured noise in this case may fall well below the original detection noise.
In practice, for feedback we employ the I/Q modulation/demodulation technique as described in Sec. \ref{sec:setup}. We can then use the following controller gain model:
\begin{equation}
H(f) = g e^{-2\pi i \tau_\mathrm{D} f + i \vartheta} \frac{\gamma f}{f_\mathrm{c}^2-f^2+i \gamma f},
\end{equation}
where we control the following parameters: gain $g$, phase $\vartheta$, central frequency $f_\mathrm{c}$ and bandwidth $\gamma$.
\section{Experimental results}
\begin{figure}[h!]
\centering\includegraphics[width=1\linewidth]{figures/mnf.pdf}
\caption{Components of noise registered by the balanced detector in setup 1. (a) Optical power spectral density normalized to photon shot noise and (b) the same signal in terms of equivalent laser phase noise as calculated from transduction of phase noise into detected noise. The three components are photon shot noise, detection noise (primarily composed of fiber noise) and the signal which corresponds to laser phase noise. The peak sensitivity due to delay loop transduction is obtained in the vicinity of \SI{2}{MHz}.}
\label{fig:master_noise}
\end{figure}
\begin{figure}[h!]
\centering\includegraphics[width=1\linewidth]{figures/fibernoise.pdf}
\caption{Incoherent part of noise in setup 1 with shot noise subtracted (blue part in \ref{fig:master_noise}) compared with theoretical prediction for noise of the delay fiber. The result is normalized by the equivalent fiber noise per unit fiber length, where we assume that all measured incoherent noise is fiber noise}
\label{fig:fnoise}
\end{figure}
\begin{figure}[h!]
\centering\includegraphics[width=1\linewidth]{figures/reduction_15.pdf}
\caption{Laser phase noise for loop parameters $f_\mathrm{c}=\SI{1.5}{\mega\hertz}$ and bandwidth $\gamma=\SI{78}{\kilo\hertz}$ and $g=1.2,\ 6$ and $12$ for columns from left to right, respectively. The first row presents the laser phase noise as measured by out-of-loop setup 2, with detection noise subtracted. In the second row we show the in-loop noise of setup 1, converted to equivalent laser phase noise units. Bottom row shows the reduction of laser phase noise, as inferred from the out-of-loop measurement.}
\label{fig:red15}
\end{figure}
\begin{figure}[h!]
\centering\includegraphics[width=1\linewidth]{figures/reduction_08.pdf}
\caption{Laser phase noise as measured by out-of-loop setup 2 and respective noise reduction due to feedback with loop parameters: $f_\mathrm{c}=\SI{0.8}{\mega\hertz}$, $\gamma=\SI{78}{\kilo\hertz}$ and $g=10$. Red curve is the original noise, blue represents the noise after feedback, and yellow curve is the theoretical prediction.}
\label{fig:red08}
\end{figure}
\begin{figure}[h!]
\centering\includegraphics[width=1\linewidth]{figures/reduction_11.pdf}
\caption{Laser phase noise as measured by out-of-loop setup 2 and respective noise reduction due to feedback with loop parameters: $f_\mathrm{c}=\SI{1.1}{\mega\hertz}$, $\gamma=\SI{78}{\kilo\hertz}$ and $g=6$. Red curve is the original noise, blue represents the noise after feedback, and yellow curve is the theoretical prediction.}
\label{fig:red11}
\end{figure}
In the experiment we use an additional fiber EOM with $V_\pi=\SI{4.65}{\volt}$ to recover an experimental profile of phase sensitivity curve of setup 2. Simultaneously, we may apply a calibration signal to the free-space EOM, which will be detected by both setups, and obtain a relative calibration of setup 1 with respect to calibrated setup 2. For the calibration we apply a low-frequency square-wave pattern and compare amplitudes of its Fourier components as registered by both setups.
Using such calibration, we can determine the cross-spectral density (CSD) which only contains the common-mode laser phase noise (Eq. \ref{eq:csd}). After expressing all PSDs and the CSD in equivalent laser phase noise units (using calibrated phase-modulation responses), we identify the incoherent (uncorrelated) parts of noise in both setups. Finally, we also measure the photon shot noise by subsequently blocking arms of the interferometer and adding resulting registered noise.
Figure \ref{fig:master_noise} presents the raw measured detector noise (expressed in shot-noise units) and equivalent laser phase noise for setup 1. We decompose the noise into the \emph{coherent} (correlated) laser phase noise and the \emph{incoherent} detection noise part, which is itself composed of shot noise, fiber noise, and other detection noise. While we expect that the incoherent noise is heavily dominated by fiber noise, our analysis does not require this assumption. With $\tau=\SI{250}{\nano\second}$ ($n=1.5$, $L=\SI{50}{\meter}$) we observe maximum sensitivity to phase noise at a \SI{2}{\MHz} offset, and minima of sensitivity at \SI{0}{\MHz} and \SI{4}{MHz}. In terms of equivalent laser phase noise, this means that in the range above \SI{3.5}{\MHz}, where laser phase noise is becoming smaller, we are heavily dominated by detection noise. Below this frequency, however, our detection setup exhibits a signal-to-noise ratio of at least 1. By using an optical power of $\bar{P}=\SI{11}{\milli\watt}$, we make sure that at our main frequency of interest at \SI{1.5}{\MHz}, detection noise is dominated by fiber noise and not shot noise. In particular, fiber noise constitutes around 70\% of detection noise under these conditions at \SI{1.5}{\MHz}. Large laser phase noise peaks around \SI{2.3}{\MHz} make the subtraction procedure sensitive to uncertainties, and thus in order to avoid imprecision we interpolate the detection noise by a second order polynomial in the \SIrange{2}{2.5}{\MHz} region. We have verified that this additional noise is internal to the laser, and most likely originates in the green pump.
With the extracted incoherent noise of setup 1, we may also compare it with a prediction for fiber noise from Eq. \ref{eq:fn}. The following parameters are assumed \cite{PhysRevA.86.023817,6171816}: $T=\SI{298}{\kelvin}$, $\kappa=\SI{1.37}{\watt\per\meter\per\kelvin}$, $\frac{\mathrm{d}n}{\mathrm{d}T}=\SI{1e-5}{\per\kelvin}$, $\alpha=\SI{5e-7}{\per\kelvin}$, $D=\SI{0.82e-6}{\meter^2\per\second}$ and $r_0=\SI{1.5}{\micro\meter}$. In Fig. \ref{fig:fnoise} observe that while the general behavior is well reproduced, the predicted fiber noise lies slightly below our measured incoherent detection noise (with shot noise subtracted). We attribute this discrepancy either to additional noise, or more likely to inaccuracy in used parameters, which are not directly available for the PM-780HP fiber we use, but are rather extracted from general material properties. In particular, some critical parameters of the model, such as for instance the coefficient of linear expansion vary appreciably across literature references and may also vary depending on the exact fiber material. Furthermore, a strong dependence of the model on fiber core radius as well as temperature call for a more systematic and precise study of fiber noise at MHz frequencies, which to the best of our knowledge has not been performed so far.
Next, we proceed to apply feedback and observe the noise registered in both setups. With other parameters optimized, we use a filter with $f_\mathrm{c}=\SI{1.5}{\mega\hertz}$ and bandwidth $\gamma=\SI{78}{\kilo\hertz}$ while changing the loop gain from $g=1.2$ through $g=6$ to $g=12$. In the top panel of figure \ref{fig:red15} we show the laser phase noise as measured with the out-of-loop setup 2. Here we subtract the uncorrelated incoherent noise. We observe that with our feedback we register significantly less noise with up to \SI{10}{\dB} of reduction for the highest gain (see the bottom row). A phase noise peak at \SI{1.54}{\MHz} is suppressed to an even greater degree. Simultaneously, we observe increased noise away from the central frequency, which becomes more pronounced at higher gains. This is due to sub-optimal phase of the feedback controller at those frequencies. For the purpose of noise reduction at the specific frequencies of our optomechanical experiment, this does not pose a problem, as long as no oscillation or feedback instability is observed.
In the middle row we show the simultaneously registered noise of the in-loop setup 1. This noise is the raw noise of the detector, converted into equivalent laser phase noise units. Here we observe that the noise, particularly for the highest gain, reaches well below the detection noise. This noise squashing behavior is expected as the result of noise self-interference. Our experimental results are accompanied by theory curves that use the model from Sec. \ref{sec:model} along with measured responses and detection noise. We observe particularly good agreement for the most important phase noise measured by the out-of-loop detector. A slightly higher noise than expected is registered in the in-loop detector, which may be a result of additional electronic detection noise.
Finally, we apply the feedback at different central frequencies, as shown in Figs. \ref{fig:red08} and \ref{fig:red11}. We observe up to \SI{12}{\dB} of noise reduction and bandwidths of at least \SI{300}{\kHz} in all cases.
\section{Conclusions and prospects}
We have demonstrated a setup that combines a measurement of laser phase noise via a delay line with an electro-optic feedback to actively reduce laser phase noise in the \si{\MHz} frequency range, which lies outside of previously explored frequency domains \cite{Li:17,Chen:89,Cranch:02,Sheard:06,Kefelian:09,Dong:15, _m_d_2015,Shehzad:19,doi:10.1063/1.3606439,Bandutunga:20}. We have demonstrated phase noise reduction of at least \SI{10}{\dB} and broadband operation, achieving relatively low phase noise densities at the level of approximately \SI{-160}{\deci\bel(\radian^2/\hertz)}, or equivalently \SI{-20}{\deci\bel((\radian\per\second)^2/\hertz)} in terms of frequency noise. We have also demonstrated an effective model for predicting the performance of our method, as well as a convenient way to identify noise components in the system based on cross-spectral density evaluation. Our results also include measurements of the fiber noise at high frequencies, which show reasonable agreement with predicted thermoconductive noise.
Our results can find particular applications in optomechanics, where reduction of noise around a specific offset frequency is desired. This applies to both sideband and feedback (cold-damping) cooling of trapped particle oscillators \cite{Delic892,PhysRevLett.122.123602,PhysRevLett.123.153601}, room-temperature integrated resonators \cite{PhysRevLett.123.223602}, membrane-in-the-middle systems \cite{ourOptica,Rossi2018} and others \cite{Kippenberg_2013,Safavi_Naeini_2013}.
Other applications sensitive to noise at a particular offset frequency include the driving of Raman transitions \cite{Arias:17,LeGouet2008} or hybrid electro-opto-mechanical converters \cite{Bagci2014}.
Our main prospective application is generation of low-phase-noise light for optomechanical sideband cooling. As estimated by Kippenberg \emph{et al.} \cite{Kippenberg_2013}, the limit for the expected final phonon occupation (i.e. with intracavity power and detuning optimized) for an optomechanical resonator in the sideband-resolved regime due to phase noise is given approximately by:
\begin{equation}
\bar{n}\approx\sqrt{\frac{\bar{n}_\mathrm{th} \Gamma_\mathrm{m}}{g_0^2} \Omega_\mathrm{m}^2 \bar{S}_{\varphi\varphi}(\Omega_\mathrm{m})}
\end{equation}
Assuming a set of parameters consistent with recent experiments \cite{ourOptica} ($g_0/2\pi=\SI{8}{\hertz}$, $\bar{n}_\mathrm{th}/2\pi=1.5\times10^5$, $\Gamma_\mathrm{m}/2\pi=\SI{2}{\milli\hertz}$ and $\Omega_\mathrm{m}/2\pi=\SI{1.5}{\mega\hertz}$), we obtain $\bar{n}=0.16$ without our noise cancellation device ($\bar{S}_{\varphi\varphi}=\SI{-154}{\deci\bel(\radian^2/\hertz)}$) which is improved to $\bar{n}=0.05$ with reduced laser phase noise ($\bar{S}_{\varphi\varphi}=\SI{-164}{\deci\bel(\radian^2/\hertz)}$). Even more remarkably, at around $\Omega_\mathrm{m}/2\pi=\SI{1}{\mega\hertz}$ we may decrease residual occupation from $\bar{n}=0.24$ to $\bar{n}=0.05$ as well.
It is also worth mentioning an alternative approach to reduce noise at \si{MHz} frequencies. This approach makes use of a filtering cavity, which has been also demonstrated as a method to measure laser phase noise \cite{Kippenberg_2013}. In such setup, one uses light directly transmitted through a narrowband cavity. However, special treatment would be required to make sure that mirror thermal noise, which is widely considered an important limitation of optomechanical setups \cite{Zhao:12}, remains well below the input laser phase noise. We estimate that such cavity would either have to be sufficiently long (we estimate at least 30 cm) to reduce mirror noise transduction, or cryogenically cooled to suppress thermal motion. Therefore, a fiber delay line seems to be a more practical solution in several matters. Another advantage over a filtering cavity is that it requires only a constant power, and does not produce a loss for main experimental light.
Finally, we envisage that the delay line approach can be further improved by reducing fiber thermal noise. One approach would be to embed the delay line in a cryogenic environment. We estimate that due to simultaneous reduction of thermal noise and the thermorefractive coefficient, the fiber noise could be reduced by more than \SI{20}{\dB} at moderate liquid-nitrogen temperatures. Furthermore, the noise can be greatly reduced with an increased mode field diameter. For a given wavelength, this could be possibly achieved in large-mode-area fibers.
\section*{Funding}
Villum Fonden (QMACVillum Investigator Grant 25880); European Research Council (Advanced Grant QUANTUM-N); John Templeton Foundation;
\section*{Acknowledgments}
We thank J. H. Müller, Y. Tsaturyan and G. Enzian for discussions and input.
M.P. is also supported by the MAB/2018/4 project "Quantum Optical Technologies", carried out within the International Research Agendas program of the Foundation for Polish Science co-financed by the European Union under the European Regional Development Fund. M.P. and I.G. contributed equally to this work. E.S.P. is a Villum Investigator.
\section*{Disclosures}
The authors declare no conflicts of interest.
| 015bb1a0f6b36c0a62f02629d0edb9b8948eb966 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\label{sec:intro}Introduction}
The electron momentum distribution, through its dependence on the ground state wavefunctions, is a powerful quantity for understanding many-body effects in solids \cite{schneider:92}. Compton scattering experiments, which involve measuring the energy distribution of inelastically scattered photons which have impinged on electrons within the sample being studied, are able to measure a projection (integration over two momentum components) of the underlying electron momentum distribution \cite{cooper:85}. Since the photons scatter from the occupied momentum states, Compton scattering is sensitive to the Fermi surfaces of metals.
While other experimental techniques such as photoemission spectroscopy (PES, and its angle-resolved counterpart, ARPES) give excellent insight into the many-body interactions present, it is important to remember that the photoemission process is a complex excitation of the whole system. Indeed, any interaction of the photo-hole with the electron quasiparticle would invalidate a claim to be a measurement of the ground-state. Thus Compton scattering is a uniquely powerful probe of the ground state many-body wavefunction \cite{Kaplan_2003}. In recent years, Compton scattering has been used to reveal the electronic structure and Fermi surfaces \cite{dugdale:14} in electronically complex materials such as substitutionally disordered alloys \cite{robarts_2020,billington_2020} and compounds with high vacancy concentrations \cite{ernsting2017vacancies}. Most relevantly, Compton scattering is able to probe the electron correlations within many complex materials \cite{sakurai:95,schulke:01,ruotsalainen2018isotropic,billington2015magnetic}. Therefore, Compton scattering offers a valuable and complementary perspective on electronic structure and, in particular, a window onto electron correlations in different regimes of composition, temperature and magnetic field from those which other probes can reach.
When X-rays are inelastically scattered by the
electrons in solids, the scattered photon energy
distribution is Doppler broadened because of the
electrons' momentum distribution $\rho(\mathbf{p})$~\cite{willi.77,lo.co.96}.
In practice this is measured through the double differential scattering cross-section
${\mathrm{d}^2\sigma/ \mathrm{d}\Omega \mathrm{d}\omega}$ for a given infinitesimal solid angle $\mathrm{d}\Omega$ and energy
$\mathrm{d}\omega$ of the scattered photon, respectively. The incident energy of the monochromatic X-rays and the scattering angle are fixed during the experiment, and the scattering cross-section is measured as a function of the photon energy.
If the scattering event is described within the impulse approximation~\cite{ch.wi.52,as.wi.52}, the
scattering cross-section is proportional to the Compton profile, ${\mathrm{d}^2\sigma/\mathrm{d}\Omega \mathrm{d}\omega \propto J(p_z)}$, which is the 1D projection of the electron momentum distribution, $\rho(\mathbf{p})$, along the scattering vector $p_z$:
\begin{equation}
\label{eq:J}
J(p_z) = \iint \rho(\mathbf{p}) \mathrm{d}p_x\mathrm{d}p_y\,.
\end{equation}
If the incident photon beam has a component of circular polarization, the scattering
cross-section contains a term which is spin dependent. This term may be isolated from
the charge scattering by either flipping the direction of the sample magnetization or
the photon helicity parallel and antiparallel with respect to the scattering vector,
resulting in a magnetic Compton profile (MCP), $J_{\mathrm{mag}}(p_z)$ \cite{duffy:13}.
In analogy to the Compton profile, the MCP is defined as
the 1D projection of the spin-polarized electron momentum density:
\begin{equation}
\label{eq:J_mag}
J_{\mathrm{mag}}(p_z)= \iint
\left[
\rho^\uparrow(\mathbf{p})-\rho^\downarrow(\mathbf{p})
\right]
\mathrm{d}p_x \mathrm{d}p_y\,.
\end{equation}
Electronic structure calculations are particularly useful for the interpretation of MCPs and these can be calculated from the spin-dependent momentum distributions. Density functional theory (DFT)~\cite{ho.ko.64,kohn.99,jo.gu.89,jone.15} is by far the most widely used method with its immense success in predicting properties of the solid state.
However, treating electron correlation in an effective one-particle framework
results in notable discrepancies with experiment even with the best available functionals. Over the last decade, it has been demonstrated that within the combined DFT and dynamical mean-field
theory (DMFT)~\cite{me.vo.89,ge.ko.96,ko.vo.04}, the so called DFT+DMFT
approach~\cite{held.07,ko.sa.06}, many of the electronic ground state
properties of $d$-metal elements, their alloys and compounds can be
well described~\cite{ko.sa.06,held.07,ka.ir.08}.
Early developments of DFT+DMFT were of the one-shot type of calculation in which a DFT computation is first converged and the subsequent DFT Hamiltonian is supplemented with a local Coulomb interaction for the correlated orbitals~\cite{sa.ko.04,le.ge.06,held.07,pa.mi.14}. Then, in a separate step, the interacting problem is solved self-consistently within DMFT.
Fully charge self-consistent approaches have been implemented~\cite{ch.vi.03,mi.ch.05,ha.ch.10,ai.po.11} as well, in which both DFT and DMFT are converged simultaneously. For some materials the correlation-induced modification in the charge density can be significant, while in others this was shown not to be the case.
However, to be consistent with the DFT concepts, charge densities ($\rho(\mathbf{r})$), should be potential ($V(\mathbf{r})$) representable (${\rho \rightleftarrows V}$), which is only achieved by full self consistency~\cite{jone.15}.
One of the most studied simple metallic system that presents signatures of electronic
correlations is the fcc itinerant ferromagnetic Ni. It is known that the DFT alone cannot reproduce the dispersionless feature at a binding energy of about $6$~eV which is known as the ``$6$~eV satellite''~\cite{gu.ba.77}. The valence band photoemission spectrum of Ni shows a $3d$-band width that is about $30\%$ narrower than the value obtained from the DFT
calculations. Similarly, the exchange splitting in both the local spin-density
approximation (LSDA) and the generalized gradient approximations (GGA)~\cite{le.ch.91}
overestimates the experimental splitting by approximately $50\%$ ~\cite{ea.hi.78,di.ge.78,hi.kn.79,ea.hi.80}.
The combined DFT+DMFT describes the occupied $3d$ bandwidth of Ni,
reproduces the exchange splitting and the $6$~eV satellite structure in the valence band~\cite{li.ka.01,br.mi.06,sa.br.12}. In addition to this, DFT+DMFT has shown the consequences of the local moment in ambient and Earth-core-like conditions \cite{hausoel2017local}. Further information about the electronic structure of Ni can be extracted by using Compton scattering.
The MCPs of Ni have been calculated by using various DFT implementations and their extensions.
Features associated with the Fermi surface (as a consequence of bands crossing the Fermi energy) seen in experiment~\cite{di.du.98} were generally reproduced with good agreement, notwithstanding the distinct discrepancy at low momenta which points towards some inaccuracies in the position of the spin-polarized bands with respect to the Fermi level.
It has been also shown that the negative polarization of the itinerant $s$- and $p$-like band electrons can be observed~\cite{ku.as.90,ti.br.90} and the discrepancy with respect to the theoretical predictions were attributed to the insufficient treatment of correlations present in the standard DFT exchange-correlation functionals at low momentum~\cite{ku.as.90}.
The directional Compton and magnetic Compton profiles have also been computed in combination with DMFT ~\cite{be.mi.12,ch.be.14a,ch.be.14b} which facilitated a discussion of the anisotropy of the electronic correlations of Ni as a function of the on-site Coulomb interaction strength, $U$.
Those theoretical comparisons with the experimental data led to the conclusion that the theoretical MCPs improved when the local correlations are taken into account, which also extends to the total Compton profiles.
In this work, we focus on the calculation of the momentum distribution and related quantities within the framework of many-body theory.
We have used two approaches on the DFT side, namely KKR~\cite{eber.00,SPR-KKR2.1.1,mi.ch.05} which is a spin-polarized relativistic multiple-scattering theory implementation, and ELK~\cite{elk} which is a full potential linear augmented planewave (FP LAPW) implementation.
For the many-body solvers we used the perturbative SPT-FLEX~\cite{ka.li.02,po.ka.05} in combination with KKR~\cite{mi.ch.05} and the numerically exact continuous-time quantum Monte Carlo (CT-QMC)~\cite{gu.mi.11,seth2016} in combination with the ELK~\cite{elk}.
We show that while the experimental magnetic moment can be obtained by varying $U$, the Ni MCP shapes have a weak dependence on $U$. These results further indicate the importance of the effects beyond the local approximation of DMFT.
\section{Compton and Magnetic Compton profiles with DFT+DMFT}
Over the last few decades, substantial progress has been made in the development of computational tools and libraries that combine DMFT with electronic structure methods in the framework of DFT+DMFT~\cite{ko.sa.06,held.07}.
All these implementations can be divided into two sub-groups according to the employed schemes for constructing the local orbitals and the definition of the so-called correlation subspace in which the DMFT equations are solved. One of the sub-groups utilizes the local atomic orbitals constructed from the atomic solutions which cover a wide energy window of the DFT (Kohn-Sham) wavefunctions, and is implemented in almost all existing electron codes with different flavors such as LMTO~\cite{gr.ma.07,gr.ma.12}, EMTO~\cite{ch.vi.03,os.vi.17}, KKR~\cite{mi.ch.05}.
The other popular choice is the set of localized Wannier wavefunctions, which is currently realized as an independent interface in various codes, including VASP~\cite{kr.ha.93}, QUANTUM ESPRESSO~\cite{Giannozzi_2009}, SIESTA~\cite{Soler_2002}, ABINIT~\cite{Gonze2009}, WIEN2k~\cite{bl.sc.20}, and ELK.
The ELK code has been interfaced with the TRIQS/DFTTools application~\cite{james_2021,aichhorn2016} to enable DFT+DMFT calculations with ELK and the TRIQS library \cite{triqs} (we refer to this as the ELK-TRIQS package).
Prior to the description of the computational procedure for the (magnetic) Compton profiles, we briefly present the steps of the fully charge-self consistent DFT+DMFT calculation.
\subsection{DFT(+DMFT) using KKR}
The calculation scheme within the Korringa-Kohn-Rostoker (KKR) method is based on the Green’s function formalism of multiple-scattering theory~\cite{eber.00}. The details of the DFT+DMFT implementation within the fully relativistic multiple-scattering KKR method have been reported previously~\cite{mi.ch.05}. It is important to note that the flexibility of KKR lies in utilizing the Dyson equation to relate the Green's function of a perturbed system to the Green's function of a suitable unperturbed reference system.
In KKR, the central quantity is the multiple scattering path-operator, $\hat{\tau}$, which is used to compute the real-space Green function.
In the most general relativistic form, the scattering path-operator is represented in the site basis with site-index $R$, and the total angular momentum with combined index ${\Lambda(\kappa,\mu)}$, with spin-orbit, $\kappa$, and magnetic, $\mu$, quantum numbers.
Electronic correlations can be included by supplementing the scalar real-valued local single-particle potential provided by the DFT with a self-energy that is energy-dependent, complex-valued, and non-local.
The Dyson equation allows for the computation of the single-particle Green's function ${G({\bf r},{\bf r^\prime},E)}$ with respect to a reference system.
The ${\bf r}$ and ${\bf r^\prime}$ are defined relative to the center of an atomic cell corresponding to a specific site, and the reference system is described by a one-electron Hamiltonian containing the DFT potentials located on the regular lattice sites~\cite{mi.ch.05,mi.eb.17}.
The single-site Dirac equation is solved and the wavefunction is matched to the free-electron solution at the boundary of the atomic cell. Subsequently, the single-site scattering matrix, ${t^{R}_{\Lambda\Lambda^\prime}(E)}$ is obtained, which in turn provides the expression for the scattering path operator, ${\tau^{R R^\prime}_{\Lambda\Lambda^\prime}(E)}$, connecting the sites $R$ and $R^\prime$.
Within this basis, employing the four-component wave functions of the regular (${Z^{R(\times)}_\Lambda({\bf r},E)}$) and irregular (${J^{R(\times)}_\Lambda({\bf r},E)}$) solutions of the Dirac equation~\cite{tamu.92}, the Green's function can be written as
\begin{align}
\begin{split}
G({\bf r},{\bf r^\prime},E)
&=
\sum_{\Lambda \Lambda^\prime} Z^{R}_{\Lambda}({\bf r},E) \tau^{R R^\prime}_{\Lambda\Lambda^\prime}(E) Z^{R^\prime\times}_{\Lambda^\prime}({\bf r^\prime},E) \\
&- \sum_{\Lambda}\!\left[Z^R_{\Lambda}({\bf r},E)J^{R\times}_{\Lambda}({\bf r^\prime},E)\theta(r^\prime-r) \right. \\
&\quad\ \ + \left.J^R_{\Lambda}({\bf r},E) Z^{R\times}_{\Lambda}({\bf r^\prime},E)\theta(r-r^\prime)\right]\!\delta_{RR^\prime},
\end{split}
\end{align}
where $\theta(r)$ is the step function.
Note that in the case of a complex non-Hermitian self-energy, ${\Sigma({\bf r},{\bf r^\prime},E)}$, one has to distinguish between the left-and right-hand side solutions of the single-site Dirac equation~\cite{tamu.92}. The left-hand side solutions are denoted by the ``$\times$''-symbol as an upper index.
The DMFT solver used in the current implementation is the relativistic version of the so-called Spin-Polarized T-Matrix (SPT-) Fluctuation Exchange approximation (FLEX)~\cite{ka.li.02,po.ka.05} which is formulated on the complex (Matsubara) axis.
In contrast to the original formulation of FLEX~\cite{bi.sc.89}, the particle-particle and the particle-hole channels are treated differently~\cite{ka.li.02,po.ka.05}.
The particle-particle processes renormalize the effective interaction, which is added into the particle-hole channel. The particle-hole channel itself describes the interaction of electrons with the spin-fluctuations, which represents one of the most relevant correlation effects in Ni.
Here we employed the SPT-FLEX solver which uses a rotationally invariant formulation for the interaction and is self-consist in charge and self-energy~\cite{mi.ch.05}. Once the self-energy is computed within the many-body solver it is returned directly into the Dirac equation~\cite{mi.ch.05}.
In order to compute the electron momentum density, the momentum operator is diagonalized in the crystal basis set, and its eigenfunctions are used to construct the Green function in the momentum representation, ${G_{\sigma}(\mathbf{p}, E)}$~\cite{be.ma.06}.
The spin-projected momentum density ${\rho^{\sigma}(\mathbf{p})}$ can then be directly computed as~\cite{be.ma.06}:
\begin{equation}
\label{eq:rho_KKR}
\rho^{\sigma}(\mathbf{p}) = {-\frac{1}{\pi} \int\limits_0^{E_F}\Im G_{\sigma}(\mathbf{p},E)dE}\,,
\end{equation}
where ${\sigma=\,\uparrow\!(\downarrow)}$ represents the spin projections.
The energy integration in Eq.~\eqref{eq:rho_KKR} is performed in the complex plane along the contour that encloses the poles of the one-particle Green's function. The corresponding Compton and magnetic Compton profiles are computed using Eqs.~\eqref{eq:J} and~\eqref{eq:J_mag}.
\subsection{DFT(+DMFT) using ELK-TRIQS}
The DFT+DMFT framework within the ELK-TRIQS package (which will be referred to as ELK+DMFT throughout) starts with a self-consistent calculation at the DFT level. The DFT density of states (DOS) is then used to identify an appropriate ``correlated'' energy window that contains the desired orbitals.
For this selected energy window, the so-called Wannier projectors~\cite{le.ge.06,aichhorn2009,james_2021} are constructed, which are used to project the lattice Green's function
onto the localized Wannier-orbitals representation.
The resulting local Green's function serves as an input to the DMFT.
Here we employ the CT-QMC solver to obtain the self-consistent solution of the DMFT equations.
The self-consistently obtained self-energy (with the double-counting removed) is then upfolded back to the Bloch basis so that it is suitable to update the lattice Green's function.
The charge density matrix is obtained from the lattice Green's function by summing over the Matsubara frequencies. This is then used to generate the total DFT+DMFT density matrix which is generally non-diagonal in the Kohn-Sham basis.
However, this total density matrix can be diagonalized into the orthonormal L\"{o}wdin-type basis~\cite{lowdin_quantum_1955} with a new set of diagonal DFT+DMFT occupations $N_{\zeta}^{\mathbf{k},\sigma} $ and wavefunctions ${\phi^\sigma_{\mathbf{k}, \zeta}(\mathbf{r})}$, as described in Ref.~\onlinecite{james_2021},
which are then used to update the electron density by
\begin{equation}
\rho(\mathbf{r})=\text{Tr}_\sigma \sum_{\zeta,\mathbf{k}} N_{\zeta}^{\mathbf{k},\sigma}
\phi^\sigma_{\mathbf{k}, \zeta}(\mathbf{r})
\left(\phi^{\sigma}_{\mathbf{k}, \zeta}(\mathbf{r})\right)^\dagger.
\end{equation}
Here $\sigma$ is the spin index, and $\zeta$ is the eigenstate index.
The DFT+DMFT results presented in this work are obtained by the fully charge self-consistent method with spin-polarized DFT inputs.
Fully charge self-consistency is achieved by updating $\rho(\mathbf{r})$
from the DFT+DMFT occupations and wavefunctions. Subsequently, the Kohn-Sham equations are solved once, and the new Wannier projectors are generated for the next fully charge self-consistent cycle.
In the current implementation of the DMFT framework, the effective Anderson impurity problem corresponding to the correlated many-body system is solved by the CT-QMC method~\cite{gu.mi.11} using the TRIQS/CTHYB application~\cite{seth2016}.
The CT-QMC methods have different formulations, namely the interaction expansion (CT-INT)~\cite{ru.sa.05}, the auxiliary-field (CT-AUX)~\cite{gu.we.08}, and the hybridization expansion (CT-HYB)~\cite{we.co.06}.
We use the CT-HYB formulation for all finite temperatures. CT-HYB operates on an imaginary time and frequency (Matsubara) axis. Therefore, analytical continuation is necessary to produce spectral functions on the real-frequency axis.
The latter suffers from the finite-precision arithmetic, which tends to amplify numerical noise, and produce unphysical artefacts. These issues may be especially severe for multi-orbital problems with complicated spectral lines.
The computation of the electron momentum density, $\rho(\mathbf{p})$, however, does not involve analytical continuation.
It is formally derived from the Fourier transformed real space wavefunctions $\psi^\sigma_{\mathbf{k}, \eta}(\mathbf{r})$, and in practice is determined by using the tetrahedron interpolation method for the discrete $\mathbf{k}$-mesh~\cite{Ernsting_2014}. The calculated electron momentum density has the form
\begin{equation}
\rho^\sigma(\mathbf{p})= \sum_{\mathbf{k}, \eta}n^{\sigma}_{\mathbf{k}, \eta}\left|\int\limits_{V} \exp (-i \mathbf{p} \cdot \mathbf{r}) \psi^\sigma_{\mathbf{k}, \eta}(\mathbf{r}) \mathrm{d} \mathbf{r}\right|^{2}, \label{EMD}
\end{equation}
where $n_{\mathbf{k}, \eta}$ are occupation functions with eigenstate index $\eta$. The electron momentum density within the DFT formalism is computed using the Kohn-Sham wavefunctions and occupation functions. Eq.~\eqref{EMD} can also be used in the DFT+DMFT formalism but now with the corresponding DFT+DMFT wavefunctions and occupation functions. By doing so, the direct impact of the non-physical artifacts of the analytical continuation on the electron momentum density can be circumvented. The occupations, as well as other observables, are implicitly dependent on the wavefunctions.
The changes in these quantities are coupled to those in the wavefunctions and are hard to disassociate.
Once the electron momentum density has been calculated, $J(p_{z})$ and/or $J_{\mathrm{mag}}(p_z)$ can be determined by Eqs.~\eqref{eq:J} and~\eqref{eq:J_mag}.
\section{Computational Details and Results}
\label{sec:comp}
Both KKR and ELK self-consistent computations were performed with the same parameters for the crystal structure ($a=3.52$~\AA) and the same parametrization for the DFT exchange-correlation potential, LSDA~\cite{perdew1992}.
The ELK DFT calculations used a 20$\times$20$\times$20 $k$-mesh, which proved to be sufficient for the k-point convergence of the self-consistent calculation.
The KKR calculations within atomic sphere approximation were performed on a 57$\times$57$\times$57 $k$-mesh, and a semicircular complex contour was used with $40$ energy points enclosing the one-particle poles of the Green's function.
The minor differences in density of states and spectral functions can be attributed to the different approaches within ELK and KKR.
Sightly more significant differences are expected to appear at the DFT+DMFT level. Both approaches use a rotationally invariant form for the interacting Hamiltonian. The multi-orbital interaction has been parameterized by the average screened Coulomb interaction $U$ and the Hund's exchange coupling $J$. The values of $U$ and $J$ are sometimes used as fitting parameters, although recent developments allow the computation of the dynamic electron-electron interaction matrix elements exactly~\cite{ar.im.04}.
It was shown \cite{mi.ar.08} that the static limit of the screened energy-dependent Coulomb interaction led to the $U$ parameter being in the energy range of $2$ and $4$~eV for all $3d$ transition metals.
Previous DMFT calculations showed that these $U$ and $J$ parameters provide the best description of the ground state properties related to the structure and different spectroscopic measurements for many of the $3d$ metals~\cite{ma.mi.09,ka.li.02,li.ka.01,ko.pa.12}.
In a considerable number of studies of bulk fcc Ni the excellent agreement with the experimental results were obtained by setting ${J = 0.9}$~eV~\cite{ch.be.14a,ch.be.14b,be.mi.12,li.ka.01,ce.we.16}, the value which we also use here.
Besides, these $U$ and $J$ parameters are in line with constrained random-phase approximation (cRPA) calculations of $3d$ transition metals~\cite{mi.ar.08,sa.fr.11}.
Note, however, that the multi-orbital interacting Hamiltonian is formulated in different basis sets. In KKR+DMFT, the local atomic basis set is used~\cite{mi.ch.05,mi.eb.17}, and consequently, the many-body problem is formulated within the $d$-block. Correlation effects are felt by other orbitals only through the self-consistency cycle.
In contrast, with the ELK+DMFT, the Wannier projectors are constructed such that the Ni-$d$ states, which are completely within the used correlated energy window of ${[-10, 3]}$~eV, are captured.
Further essential parameters for the CT-QMC computation~\cite{seth2016}) are the number of 4.2$\times$10$^{8}$ sweeps and the inverse temperature $\beta$ of $40$~eV$^{-1}$.
In both methods the spin-polarized around-mean-field double-counting term (AMF)~\cite{pe.ma.03,cz.sa.94} was employed.
The ELK+DMFT spectral function presented in Sec.~\ref{sec:sf} was calculated by analytically continuing the DMFT self-energy using the {\em LineFitAnalyzer} technique of the maximum entropy analytic continuation method implemented within the TRIQS/Maxent application \cite{maxent}.
The different descriptions of the potentials, full potential in ELK and the atomic sphere approximation (ASA) in KKR, also lead to the difference in the calculated chemical potentials.
Within the KKR+DMFT method,
the self-energy is added into the Kohn-Sham-Dirac equation~\cite{mi.ch.05,mi.eb.17}, and the chemical potential is updated to conserve the number of valence electrons similarly as in the DFT loop.
The ELK+DMFT, using the Wannier projectors instead, updates the electron density from which a new set of Kohn-Sham eigenvalues and eigenvectors are generated and the corresponding chemical potential is obtained. The difference in the DFT+DMFT chemical potential with respect to the DFT values is at most a few tenths of an eV. The different solvers produce slightly different values for the real parts of the self-energies at the chemical potential.
An important point here is the double counting and even though the functional form is the same for both KKR+DMFT and ELK+DMFT, the slightly different values in the occupation matrix produce slightly different double counting values.
\subsection{$U$-dependent spin and orbital magnetic moments}
\label{sec:spin_mag}
To identify the optimal value of $U$, or at least to narrow the {\it ab-initio} interval, we first analyzed the behavior of the Ni ferromagnetic spin magnetic moment with respect to the on-site Coulomb interaction, $U$ and fixed Hund's rule coupling, ${J=0.9}$~eV.
\begin{figure}[!t
\centerline{\includegraphics[width=0.95\linewidth]{Fig_1.pdf}}
\caption{Spin $m_s(\mu_B)$ and orbital $m_\ell(\mu_B)$ magnetic moment as function of the intra-site Coulomb potential $U$. Blue plus signs and black crosses represent results of ELK(+DMFT) and KKR(+DMFT) spin magnetic moment calculations, respectively. On the right axis we plot the KKR (red asterisks) orbital magnetic moment ($m_l$). In all calculations, the Hund's rule coupling parameter ${J=0.9}$~eV was used. The data points for ${U=0.0}$~eV represent the DFT calculations.}
\label{fig:mag}
\end{figure}
The ferromagnetic spin ($m_s$) and orbital ($m_\ell$) magnetic moment as a function of the on-site Coulomb interaction $U$ are shown in Fig.~\ref{fig:mag}. Both the ELK+DMFT and KKR+DMFT results show a similarly decreasing spin magnetic moment with increasing $U$, in quite close correspondence to each other.
Contrary to the decreasing spin moment over the entire $U$ range, the orbital moment $m_\ell$ obtained in relativistic KKR+DMFT calculations, increases with the $U$ values, passing the maximum value at ${U\approx2.3}$~eV, and decreases upon further increasing the value of $U$. Even for the largest value of $U$ (${U=3.0}$~eV in the presented calculations), the KKR+DMFT orbital magnetic moment is larger than the corresponding DFT value by about $30\%$. Similar results have also been reported previously in Ref.~\onlinecite{mi.ma.14} and were interpreted as a correlation-induced orbital moment enhancement.
Despite the different descriptions, it is satisfying to see the good agreement between the results obtained with both methods. For ${U=2.0}$~eV, the calculated spin moment matches best with experiment {$\approx 0.56~\mu_{\mathrm{B}}$} for both DFT+DMFT methods and is within the {\it ab-initio} predictions for the $3d$ transition elements. These ${U=2.0}$~eV and ${J=0.9}$~eV values are in agreement with that used in the previous spin-polarized Ni ACAR study~\cite{ce.we.16}.
The experimental spin moment originates from the polarized neutron diffraction measurements by Ref.~\onlinecite{Mook_1966}. The total (spin + orbital) measured magnetic moment, which the analysis relied on, was subsequently revised by Ref.~\onlinecite{Danan_1968} to ${0.616\,\mu_B}$, and with which our KKR+DMFT ${U=2.0}$~eV calculation has excellent agreement.
Our chosen $U$ and $J$ values for ferromagnetic fcc Ni have a higher $J$/$U$ ratio compared to previous (DFT+)DMFT studies in the paramagnetic phase~\cite{ge.me.13}.
For other values of ${J}$ two different values of $U$ are required to match either spin or orbital moments, where as for ${J=0.9}$~eV and ${U=2.0}$~eV we obtain an excellent agreement of both.
Magnetic Compton scattering, however, does not directly provide information concerning the orbital moments, but when combined with a SQUID measurement of the total magnetic moment, the orbital contribution can be inferred \cite{du.03}.
\subsection{Magnetic Compton profiles}
\label{sec:mcp}
In the KKR(+DMFT), the magnetic Compton profiles are calculated from the spin-resolved momentum density ${\rho^{\sigma}(\mathbf{p})}$ which in turn is obtained as a contour integral of the Green's function in the momentum representation, Eq.~\ref{eq:rho_KKR}.
In the ELK+DMFT the electron momentum densities (and the MCPs) are computed through the wavefunctions and occupation functions across the Brillouin zone on the imaginary frequency axis.
The method of obtaining the wavefunctions and occupation functions in ELK-TRIQS are described in Ref.~\onlinecite{james_2021}.
In both methods, the MCPs were calculated within a sphere of radius ${16}$~a.u. (${|\mathbf p| \leqslant 16}$~a.u.), and then renormalized such that their areas were equal to the corresponding spin magnetic moment.
\begin{figure*}[!t]
\centerline{\includegraphics[width=\linewidth]{Fig_2.pdf}}
\caption{The Ni magnetic Compton profiles (MCPs) at $[100]$, $[110]$, and $[111]$ high symmetry directions (indicated on each plot) for several intra-site Coulomb potential $U$ and fixed Hund's rule coupling ${J=0.9}$~eV. The KKR+DMFT MCPs results are shown in the (a), (b) and (c) panels (upper row). The ELK+DMFT MCPs are presented in the (d), (e) and (f) panels (lower row). The areas of each MCP have been normalized to their corresponding spin magnetic moment results given in Fig.~\ref{fig:mag}. The DFT+DMFT results are complemented by the LSDA results from the respective ELK and KKR codes (dashed curves) and the experimental measurements from Dixon {\it et~al.} (dots with error bars) \cite{di.du.98}. For clarity, the error bars are shown for every tenth data point. The computed results have been convoluted with a Gaussian with a full-width-at-half-maximum (FWHM) of $0.43$~a.u. to represent the experimental resolution.
}
\label{fig:elk_kkr}
\end{figure*}
\begin{figure}[t!]
\centerline{\includegraphics[width=0.95\linewidth]{Fig_3.pdf}}
\caption{The comparison of the experimental Ni magnetic Compton profiles (MCPs) from Dixon {\it et~al.}~\cite{di.du.98} (dots with error bars) with the DFT results (solid and dashed curves) and the DFT+DMFT results for the chosen ${U=2.0}$~eV and ${J=0.9}$~eV (dash-dotted and dotted curves). For clarity, the error bars are shown for every tenth data point. The (a), (b) and (c) panels show the MCPs for the $[100]$, $[110]$ and $[111]$ high symmetry directions. The computed results have been convoluted with a Gaussian with a full-width-at-half-maximum (FWHM) of $0.43$~a.u. to represent the experimental resolution.
The areas of each MCP have been normalized to their corresponding spin magnetic moments given in Fig.~\ref{fig:mag}.}
\label{fig:Exp_comp}
\end{figure}
\begin{figure}[t!]
\centerline{\includegraphics[width=0.965\linewidth]{Fig_3_unconv.pdf}}
\caption{The comparison of the unconvoluted high symmetry direction ([100], [110], and [111]) MCPs from the DFT (solid and dashed curves) and the DFT+DMFT with ${U=2.0}$~eV, ${J=0.9}$~eV (dash-dotted and dotted curves) calculations. The profiles in this figure are the unconvoluted counterparts of the MCPs which are in Fig.~\ref{fig:Exp_comp}}
\label{fig:comp_unconv}
\end{figure}
To analyze the effects of correlation on the MCPs, we calculated MCPs with the DFT+DMFT method for a series of on-site interaction values $U$ and Hund's rule coupling ${J=0.9}$~eV by employing both KKR+DMFT and ELK+DMFT. Fig.~\ref{fig:elk_kkr} shows the Ni MCPs along the cubic high symmetry directions, obtained using the KKR(+DMFT) (Figs.~\ref{fig:elk_kkr}~(a)-(c)), and the ELK(+DMFT) (Figs.~\ref{fig:elk_kkr}~(d)-(f)) in the momentum range ${0 \leqslant p_z \leqslant 8}$~a.u..
The theoretical MCPs have been convoluted with a Gaussian with a full-width-at-half-maximum (FWHM) of $0.43$~a.u. to represent the experimental resolution.
Starting with the presented DFT results, the MCPs show good agreement with the experiment for ${p_z > 2}$~a.u. but these MCPs do not match the low-momentum region for any of these high symmetry directions. Our DFT results are in good agreement with those previously presented in Ref.~\onlinecite{di.du.98}.
The MCP peak structures within the first Brillouin zone are due to the exchange splitting, which in turn causes the majority and minority spin bands to cross the Fermi level at different $\mathbf{k}_F$ values (see Fig.~\ref{fig:comp_unconv} and Fig.~\ref{fig:Akw}).
These peaks are periodically repeated in the MCPs as these are the umklapp contributions from higher zones (i.e., ${\mathbf{k}+\mathbf{G}}$ where $\mathbf{G}$ is the reciprocal lattice vector).
One of the advantages of the effective one-particle framework of DFT calculations is the possibility to decompose the total MCP into the contributions originating from individual bands~\cite{di.du.98,ku.as.90,ti.br.90,ka.ku.03}.
The dip in this low-momentum region has been attributed partly to the contribution of the so-called negative polarization of the $s$- and $p$-like bands with respect to the positive contribution of the $d$-bands.
At the same time, Refs.~\onlinecite{di.du.98,ka.ku.03} note that another source of discrepancy may be due to the $d$-like fifth band (band numbering according to Ref.~\onlinecite{di.du.98}), where Ref.~\onlinecite{ka.ku.03} attributes the shape of the contribution of this band to the inconsistencies between the theoretical and the true Fermi surface.
These interpretations, based on the DFT band structures, raise some interesting unsolved questions about the origin of the discrepancy at low momentum.
From the DFT results, the predicted negative polarization contributions are not sufficient to explain the low momentum dip seen in the experimental results.
Dixon {\it et~al.}, Ref.~\onlinecite{di.du.98}, suggested that it was the deficient representation of the $d$-electron correlations in LSDA (and GGA), not just the negative polarization from the $s$- and $p$-electrons, which was the potential cause for the low momentum experiment-theory disagreement ~\cite{di.du.98}. Artificial shifts of the bands around the Fermi level~\cite{ma.du.04} showed improved agreement with the low momentum MCP region. As correlation effects lead to the shift of those bands naturally, improved theoretical description of the Ni MCPs can be obtained by taking them into account.
Recent studies~\cite{be.mi.12,ch.be.14a,ch.be.14b} also demonstrated that including the local correlations through the DMFT framework reduces the discrepancy between theoretical and experimental MCPs of Ni.
Moving onto the DFT+DMFT results,
the large dips near ${p_z= 0}$~a.u. in the high symmetry directions are better reproduced by the DFT+DMFT MCP for ${U>2.3}$~eV. On the other hand, for high-momentum, ${p_z>2}$~a.u., region ${U<2.3}$~eV is a better choice.
Although we are able to produce improved agreement (with respect to the DFT MCPs) with the experiment at low momentum, ${p_z < 1}$~a.u., DFT+DMFT fails to reproduce the experimental MCP for the $[100]$ and $[110]$ directions (see Fig.~\ref{fig:elk_kkr}~(a),(d) and (b),(e)). Along the nearest-neighbor direction $[110]$, no $U$ value was found to suppress the peak at around ${p_z = 0.6}$~a.u..
Although the general low momentum disagreement is the case for both implementations, there are some notable differences between the ELK+DMFT and KKR+DMFT results. Along $[100]$ direction, ${J_{\mathrm{mag}}(p_z)}$ for ${p_z<1}$~a.u calculated with ELK+DMFT for increasing values of $U$ matches the experimental MCP better than those obtained with KKR+DMFT. The latter visibly overestimates the ${J_{\mathrm{mag}}(p_z)}$ (by almost the same amount) for all $U$ values considered.
The opposite happens for the $[111]$ direction. In this case, ${J_{\mathrm{mag}}(p_z)}$ obtained with KKR+DMFT matches the experimental values in ${p_z\lesssim 2}$~a.u. region for ${U\geqslant 2.3}$~eV, while ELK+DMFT results overestimate the experimental values for ${p_z\lesssim 1}$~a.u for all considered values of $U$.
Although in general the low momentum is better described with higher $U$ values (see Fig.~\ref{fig:elk_kkr}), the costs of this is the poorer agreement with the experiment from $1$~a.u. to about $5$~a.u..
This is because the area under the MCP, which is equal to the corresponding spin moment for each $U$ value in Fig.~\ref{fig:mag}, reduces with increasing $U$ and is less than the experimental value for about ${U>2.0}$~eV. Therefore, for the different $U$ values, an improvement in one momentum region of the MCP causes another region to worsen in order to conserve the area.
We did not find a single $U$ value, within the {\it ab-initio} range of $U$ values, which would simultaneously match both, low- and high-momentum regions of experimental profile within its error.
On the other hand, in the previous section, we identified that the DFT+DMFT calculation with ${U=2.0}$~eV and ${J=0.9}$~eV produces the best match between the calculated and experimental magnetic moment.
To see how well the DFT+DMFT MCPs for ${U=2.0}$~eV match the experimental MCPs from Ref.~\onlinecite{di.du.98}, and also to compare the results obtained by two different packages
and two distinct frameworks
in Fig.~\ref{fig:Exp_comp}, we show the corresponding MCPs.
Although the MCPs calculated in the DFT+DMFT framework for ${U=2.0}$~eV deviate from the experimental results in the momentum range ${0< p_z < 1}$~a.u., extending DFT with the DMFT framework significantly improves the description of the experiment in the range ${1< p_z< 2}$~a.u..
For ${U=2.0}$~eV, the structure of the MCPs is well reproduced in all three high symmetry directions in this region where the dominant contributions are made.
DFT+DMFT results also stay in reasonably good agreement with the experiment for higher values of $p_z$, from ${p_z=2}$~a.u. onwards, but they tend to slightly underestimate the tails, although they are within the experimental error. This is also a consequence of the calculations overestimating the low momentum region.
\begin{figure*}[t!]
\centerline{\includegraphics[width=0.95\linewidth]{Fig_5.pdf}}
\caption{The ELK DFT band structure and ELK+DMFT DFT+DMFT (${U=2.0}$~eV and ${J=0.9}$~eV) spectral function for the majority (left) and minority (right) spin. The DFT bands have been broken down in terms of their indices (using the same numbering as Dixon {\it et~al.}~\cite{di.du.98}) for discussions about the band contributions to the MCPs and resemble those of Dixon {\it et~al.}. The insets are a zoomed image around the X symmetry point (indicated by the grey outline) showing the differences between the theoretical treatments.}
\label{fig:Akw}
\end{figure*}
Overall, dynamic correlations improve the agreement with the experimental data beyond the LSDA results.
The results including dynamic correlations also show the correct trend for low momentum region ${p_z \lesssim 2}$~a.u. where better MCPs are obtained in comparison to the LSDA.
LSDA overestimates MCP values almost in the entire region.
As mentioned earlier, since the areas under the MCPs directly equal the spin moment ($m_s$), the areas reduce with increasing $U$ as per Fig.~\ref{fig:mag}.
Nevertheless, since DFT+DMFT also overestimates the experimental MCP values in the ${p_z \lesssim 1}$~a.u. region, the consequence is that the tails (high-momentum region) are underestimated --- after all, the areas beneath the MCP and ${U=2.0}$~eV DFT+DMFT curves are almost equal.
The current results, however, do not allow us to infer the optimal value for the on-site Coulomb interaction necessary to obtain the best agreement with the experimental measurements. Nevertheless, we see that $U$ values in the range $[1.7,2.3]$~eV describe the on-site Coulomb interaction reasonably well (almost exactly within experiment error bars in the ${1\lesssim p_z \lesssim 2}$~a.u. range), in agreement with positron annihilation measurements~\cite{ce.we.16}. Similar conclusions have been drawn in previous papers reporting the correlation effects upon the MCP of Ni~\cite{be.mi.12,ch.be.14a,ch.be.14b}.
A direct comparison between the methods can be seen in Fig.~\ref{fig:comp_unconv}, where we plot the theoretical MCPs which have not been convoluted with the experimental resolution.
The results produced with the two DFT+DMFT implementations are in excellent agreement. Therefore, we are confident that the effect of the resolution on the MCPs does not hide any glaring disagreements between the implementations.
Finally, we conclude that neither implementation produces results (for all $U$ values) that have a better overall agreement with the experimental data than the other (within the experimental error). The level of experiment-theory agreement between the MCPs from both implementations varies in different regions of momentum.
Overall, the results from the two implementations are in good agreement with minor discrepancies due to the aforementioned differences discussed at the beginning of Sec.~\ref{sec:comp}.
\subsection{Spectral function}
\label{sec:sf}
Features in the MCPs can be traced back to the form of the spectral function which, for the non-interacting case, is represented by the band structure.
Fig.~\ref{fig:Akw} shows the DFT band structure together with the DFT+DMFT $\mathbf{k}$-resolved spectral function along the high symmetry direction in the Brillouin zone from the ELK and ELK+DMFT calculations.
In the present DFT calculation, we confirm that the bands $5$ and $6$ of Fig.~\ref{fig:Akw} only give a positive contribution to the MCPs whereas bands $1$-$4$ have negative contributions to the MCPs at low momentum.
In a many-body picture, however, such a band-resolved interpretation is not possible. The spectral function in Fig.~\ref{fig:Akw} shows the quasi-particle dispersion. The self-energy affects the two spin channel spectral functions differently. A significant part of the energy dependence of the self-energy is related to the different occupations of the spin-polarized $d$-states, on which the MCPs are also dependent. Scattering processes involving $s$-electrons may be neglected as the corresponding orbitals are almost completely filled~\cite{lieb.79}.
Within the DMFT approximation, the self-energy matrix is diagonal in the angular momentum representation and is independent of ${\mathbf{k}}$.
It is the orbital dependence of the self-energy that produces a coupling between the terms of the $d^8$-multiplets
~\cite{lieb.79}, where the neglected $k$-dependence of the self-energy amounts to disregarding the hopping processes of the two holes bound to the same Ni-site. The CT-HYB impurity solver captures the self-energy contributions relevant for the strong ferromagnetic state such as repeated scattering of paired holes, hole-hole and hole-electron interactions as these processes enter in the fully rotationally invariant formulation of the Hubbard model and parameterized by the $U$ and $J$ parameters~\cite{ko.pa.12}. As Ni has a relatively large band width, the atomic multiplet structure is extended in the energy range around ${-6}$~eV. Therefore, the expected satellite in our treatment is a broad feature instead.
The prominent correlation effect of the DFT+DMFT $\mathbf{k}$-resolved spectral function is to renormalize the position and width of the $d$-bands and significantly reduce the exchange splitting to about $0.3$~eV at the L-point (which we measured as the difference between the majority and minority band centers). These are direct consequences of the presence of the real part of the DMFT self-energy having a negative slope at $E_F$.
These features are in good agreement with experiments \cite{hi.kn.79,ea.hi.80} and are in line with previous studies \cite{li.ka.01,br.mi.06,hausoel2017local}.
We observe that the crossing of the bands at the Fermi level hardly changes for the majority spin channel (see the left panel in Fig.~\ref{fig:Akw}).
In the minority bands, however, there are subtle changes around the X-point, where two X-hole pockets reside (see the inset in the right panel in Fig.~\ref{fig:Akw}).
These changes are less significant for the MCPs but are relevant in other experiments such as de Haas-van Alphen and ARPES~\cite{Wang_1974,weling_1982}. Previous experiment-theory comparisons~\cite{Wang_1974,weling_1982} have shown that DFT predicts a second shallow minority hole pocket around X. This is referred to as the minority X$_2$ hole pocket (related to minority band $3$ in Fig.~\ref{fig:Akw}) but there is no strong evidence of its presence in the experiments.
The present DFT+DMFT calculation with ${U=2.0}$~eV shows that the size of the minority X$_2$ hole pocket shrinks and also becomes shallower as compared to DFT results, but it does not vanish.
These may indicate that other correlation effects are required to suppress this band below the Fermi level, or that the large effective mass of X$_2$ hole pocket due to the shallowness of the corresponding band around X (see the inset in the right panel in Fig.~\ref{fig:Akw}) might have made its observation more challenging.
Contrary to previous interpretations based on the one-particle description it is not obvious that the negative polarization contributions to the MCP (by the $s$- and $p$-electrons) is the cause for the disagreement between the experiment and the DFT and DFT+DMFT computations.
The low momentum disagreement is likely the consequence of the other missing correlation effects beyond DFT+DMFT, such as screening.
As screening is a genuine many-body effect, it requires methods like quasiparticle GW (QSGW).
Such a calculation for Ni has been performed recently by L. Sponza {\it et~al.}~\cite{sponza_2017}. The QSGW calculations produce an enhanced value for the magnetic moment and exchange splittings.
Nonetheless, in supplementing the computations with DMFT in the combined QSGW+DMFT, the values for the magnetic moment and exchange splitting are in good agreement with the experiment.
We expect that a QSGW+DMFT calculations would likely improve the MCPs, as these incorporate non-local and screening effects.
\section{Conclusion and Outlook}
To conclude, we have presented results of two different DFT+DMFT implementations to calculate the spin-resolved momentum distributions ${\rho^\sigma({\bf p})}$, and the magnetic Compton profile ${J_\mathrm{mag}(p_z)}$.
Both of these implementations show excellent agreement with each other considering the differences in their approaches to applying both DFT and DMFT and the different challenges that these contribute to the calculations.
The DFT+DMFT spin moment calculations have the same $U$ dependence in both setups, the slight difference in magnitude likely being due to the details of the implementations.
Although the spin moment improves to be comparable with the experimental value, the shape of the MCP has a weak $U$ dependence, features in the profile such as umklapp peaks, remain relatively unchanged and only the MCP contributions are redistributed compared to the calculated DFT profiles.
For the ${U=2.0}$~eV calculation, which reproduces the experimental spin (and total) magnetic moments, the corresponding spectral function reveals that the minority X$_2$ pocket shrinks and gets shallower with respect to the DFT calculations, but nevertheless still survives. This small X$_2$ pocket is likely to have a large effective mass and this may explain why it was not observed in the de Haas-van Alphen experiment.
According to our combined DFT+DMFT approaches, some arguments in previous DFT studies built upon the existence of negative polarization description are not sufficient to explain the discrepancy between the theoretical and experimental MCP and low-momentum region.
Instead, theories including a non-local description of interaction and retardation effects (i.e., energy-dependent screening) such as cluster-DMFT, GW (QSGW) and beyond might be more suitable to deliver a better description of the MCP in ferromagnetic metals such as Ni.
To truly resolve the intricacies which may arise between the aforementioned theoretical frameworks, it would be essential to remeasure the Ni MCPs with a higher resolution. This will lead to further valuable understanding of the many-body groundstate properties probed in momentum space.
\begin{acknowledgments}
Financial support by the Deutsche Forschungsgemeinschaft through TRR80 (project E2)
Project number 107745057 is gratefully acknowledged.
A.D.N.J.~acknowledges funding and support from the
Engineering and Physical Sciences Research Council (EPSRC), Grant No.~EP/L015544/1.
M.S.~was partially supported by the Shota Rustaveli Georgian National Science Foundation through the grant N~FR-19-11872.
We would like to thank E.~I.~Harris-Lee for his insights and also J.~Min\`ar and H.~Ebert for a fruitful collaboration.
\end{acknowledgments}
| a8cc6daa3d277ca5fbd0ceda71868d78e827be7e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
To study the knots and links occurring in flows on $S^3$, Birman and Williams \cite{BW1, BW2} introduced the template (they called it the knot holder) as a geometric model. Roughly speaking, a template is a compact branched surface with an associated semi-flow on it, whose periodic orbits are in bijective correspondence with the periodic orbits of the original $3$-dimensional flow.
As in knot theory, the embedded template is studied by the template diagram with planar moves \cite{KSS}. Two template diagrams correspond to isotopic templates if and only if one can be obtained from the other by a finite sequence of plane isotopies and planar moves. However, from the point of view of dynamics, two embedded templates are identified if they are isotopic in ambient space with additional switch move and splitting move \cite{BW2}. These two template moves may change the homeomorphism type of a template, thus bring an obstruction to define topological invariant of embedded templates.
In this paper, we fix the homeomorphism type of an embedded template.
Then the boundary of the embedded template is a trivalent spatial graph, which is in fact a complete invariant of the original template with specifical homeomorphism type (Theorem~\ref{crit}). Furthermore, we use Kauffman's invariant of spatial graphs \cite{K} to make it easy for comparing template boundaries. It leads an isotopic invariant which is a set of knots and links (Theorem~\ref{taut}). Another advantage of this link-set invariant is the veracity of describing the template with twisted and knotted bands. Knots as invariants of dynamics are not unusual such as in \cite{BG}.
The isotopy of flows is also under consideration. Particularly for simple Smale flows, we construct a spatial graph invariant and a generalization of Kauffman's invariant involving the unique attractor (Theorem~\ref{lemg}). They can classify the simple Smale flows on $S^3$ with Lorenz-like templates \cite{Su, Y} and the universal template with its analogues \cite{HS, AS}, and the former is a complete isotopic invariant for general case (Theorem~\ref{flowg}) while the completeness of the later is unknown.
This paper is organized as follows. In Section 2, we give a brief account of template theory, and fix some notions. The boundary spatial graph of a template is discussed in Section 3. Our main result lies in Section 4. We show that the isotopy class of a template is totally determined by that of its boundary spatial graph. Section 5 devotes an application of our results into simple Smale flow on $S^3$.
\section{Templates and ambient isotopy} \label{secBCHD}
In this section, we concentrate on $3$-dimensional manifolds. We state the template theory briefly and discuss ambient isotopy acting on embedded templates.
\begin{definition} \cite{BW2, GHS}
A template $(\fT, \varphi_t)$ is a compact branched $2$-manifold $\fT$ constructed from two types of charts, called joining charts and splitting charts, together with a semi-flow $\varphi_t\col \fT \to \fT$. The gluing maps between charts must respect the semi-flow and act linearly on the edges.
\end{definition}
Here a semi-flow is the same as a flow except that one cannot go backward in time uniquely, and in a template one cannot back up uniquely at a branch line. The semi-flows are usually indicated by arrows on charts.
\begin{example}
Lorenz template $\mathcal{L}(0,0)$, horseshoe template $\mathcal{L}(0,1)$, and template $\mathcal{L}(1,1)$ as bellow. Any two of them are not homeomorphic as branched surfaces, since the numbers of non-orientable bands of them are $0,1$ and $2$, respectively.
\end{example}
\begin{center}
\setlength{\unitlength}{0.26mm}
\begin{picture}(150,90)(-75,-40)
\qbezier(-70,0)(-70,50)(-30,50)
\qbezier(-30,50)(-15,50)(0,40)
\qbezier(0,40)(30,20)(30,0)
\qbezier(70,0)(70,50)(30,50)
\qbezier(30,50)(15,50)(0,40)
\bezier{30}(0,40)(-30,20)(-30,0)
\qbezier(-70,-10)(-70,-40)(-40,-40)
\qbezier(-40,-40)(-10,-40)(-10,-10)
\qbezier(70,-10)(70,-40)(40,-40)
\qbezier(40,-40)(10,-40)(10,-10)
\qbezier(-50,0)(-50,10)(-40,10)
\qbezier(-40,10)(-30,10)(-30, 0)
\qbezier(-50,-10)(-50,-20)(-40,-20)
\qbezier(-40,-20)(-30,-20)(-30, -10)
\qbezier(50,0)(50,10)(40,10)
\qbezier(40,10)(30,10)(30, 0)
\qbezier(50,-10)(50,-20)(40,-20)
\qbezier(40,-20)(30,-20)(30, -10)
\put(-70,0){\line(0,-1){10}}
\put(-50,0){\line(0,-1){10}}
\put(-30,0){\line(0,-1){10}}
\put(-10,0){\line(0,-1){10}}
\put(70,0){\line(0,-1){10}}
\put(50,0){\line(0,-1){10}}
\put(30,0){\line(0,-1){10}}
\put(10,0){\line(0,-1){10}}
\put(-30,0){\line(1,0){60}}
\end{picture}
\ \
\begin{picture}(150,90)(-75,-40)
\qbezier(-70,0)(-70,50)(-30,50)
\qbezier(-30,50)(-15,50)(0,40)
\qbezier(0,40)(30,20)(30,0)
\qbezier(70,0)(70,50)(30,50)
\qbezier(30,50)(15,50)(0,40)
\bezier{30}(0,40)(-30,20)(-30,0)
\qbezier(-70,-10)(-70,-40)(-40,-40)
\qbezier(-40,-40)(-10,-40)(-10,-10)
\qbezier(70,-10)(70,-40)(40,-40)
\qbezier(40,-40)(10,-40)(10,-10)
\qbezier(-50,0)(-50,10)(-40,10)
\qbezier(-40,10)(-30,10)(-30, 0)
\qbezier(-50,-10)(-50,-20)(-40,-20)
\qbezier(-40,-20)(-30,-20)(-30, -10)
\qbezier(50,0)(50,10)(40,10)
\qbezier(40,10)(30,10)(30, 0)
\qbezier(50,-10)(50,-20)(40,-20)
\qbezier(40,-20)(30,-20)(30, -10)
\put(-70,0){\line(0,-1){10}}
\put(-50,0){\line(0,-1){10}}
\put(-30,0){\line(0,-1){10}}
\put(-10,0){\line(0,-1){10}}
\put(30,0){\line(0,-1){10}}
\put(10,0){\line(0,-1){10}}
\qbezier(63.916, -4.2432)(70., -2.73)(70., 0.)
\qbezier(56.084, -5.7568)(50., -7.27)(50., -10.)
\qbezier(60., -5.)(50., -3.5)(50., 0.)
\qbezier(60., -5.)(70., -6.5)(70., -10.)
\put(-30,0){\line(1,0){60}}
\end{picture}
\ \
\begin{picture}(150,90)(-75,-40)
\qbezier(-70,0)(-70,50)(-30,50)
\qbezier(-30,50)(-15,50)(0,40)
\qbezier(0,40)(30,20)(30,0)
\qbezier(70,0)(70,50)(30,50)
\qbezier(30,50)(15,50)(0,40)
\bezier{30}(0,40)(-30,20)(-30,0)
\qbezier(-70,-10)(-70,-40)(-40,-40)
\qbezier(-40,-40)(-10,-40)(-10,-10)
\qbezier(70,-10)(70,-40)(40,-40)
\qbezier(40,-40)(10,-40)(10,-10)
\qbezier(-50,0)(-50,10)(-40,10)
\qbezier(-40,10)(-30,10)(-30, 0)
\qbezier(-50,-10)(-50,-20)(-40,-20)
\qbezier(-40,-20)(-30,-20)(-30, -10)
\qbezier(50,0)(50,10)(40,10)
\qbezier(40,10)(30,10)(30, 0)
\qbezier(50,-10)(50,-20)(40,-20)
\qbezier(40,-20)(30,-20)(30, -10)
\put(-30,0){\line(0,-1){10}}
\put(-10,0){\line(0,-1){10}}
\put(30,0){\line(0,-1){10}}
\put(10,0){\line(0,-1){10}}
\qbezier(-63.916, -4.2432)(-70., -2.73)(-70., 0.)
\qbezier(-56.084, -5.7568)(-50., -7.27)(-50., -10.)
\qbezier(-60., -5.)(-50., -3.5)(-50., 0.)
\qbezier(-60., -5.)(-70., -6.5)(-70., -10.)
\qbezier(63.916, -4.2432)(70., -2.73)(70., 0.)
\qbezier(56.084, -5.7568)(50., -7.27)(50., -10.)
\qbezier(60., -5.)(50., -3.5)(50., 0.)
\qbezier(60., -5.)(70., -6.5)(70., -10.)
\put(-30,0){\line(1,0){60}}
\end{picture}
\end{center}
Let $M$ be a connected compact smooth $3$-manifold with a smooth flow $\phi_t$, which has a hyperbolic chain-recurrent set $\mathcal{R}(\phi_t)$. By the Template Theorem of Birman and Williams \cite{BW2}, there is a template $\fT$ embedded in $M$ with a semi-flow $\varphi_t$ such that the periodic orbits of saddle of $\phi_t$ are in bijective correspondence with the periodic orbits of $\varphi_t$, and for any finitely many orbits the correspondence can be via ambient isotopy.
Here an ambient isotopy is a family of diffeomorphisms $\{h_{\lambda}\}_{\lambda \in I}$ of the ambient manifold $M$, depending smoothly on $\lambda$ and satisfying that $h_0 = 1_M$. We can define the ambient isotopy of embedding templates.
\begin{definition}
Two embedded templates $(\fT,\varphi_t)$ and $(\fT' ,\varphi'_t)$ in a $3$-manifold $M$ are said to be ambient isotopic if there is an ambient isotopy
$\{ h_\la \}_{\la \in I}$ such that
$h_1 (\fT)= \fT'$
and
$h_1\circ \varphi_t= \varphi_t '$.
\end{definition}
If two templates $(\fT,\varphi_t)$ and $(\fT' ,\varphi_t ')$ are ambient isotopic, then they are homeomorphic as branched surfaces and topologically equivalent as dynamical systems, as well as have the same embedding properties.
Next we consider the effect of an ambient isotopy acting on a template of a flow.
\begin{proposition} \label{isot}
Let $\phi_t$ be a flow on $M^3$ with hyperbolic chain-recurrent set. Suppose that $\{ h_\la \}_{\la \in I}$ is an ambient isotopy of $M$. Then the chain-recurrent set of the flow $h_1\circ \phi_t$ is hyperbolic. Moreover, if $\fT$ is a template of $\phi_t$, then $h_1 (\fT)$ is a template of $h_1\circ \phi_t$.
\end{proposition}
\begin{proof}
The differential of a diffeomorphism keeps the $\phi_t$-invariant decomposition of tangent bundle
$$\mathrm{T}M = \mathrm{E}^c\oplus \mathrm{E}^s\oplus \mathrm{E}^u.$$
Thus ambient isotopy preserves the chain-recurrent set decomposition clearly. It follows that $\{ h_\la \circ \phi_t\}_{\la \in I}$ is a family of flows on $M$, all of which have hyperbolic chain-recurrent sets. Particularly, so does $h_1\circ \phi_t$. Then from the constructing procedure of a template, following \cite{BW2, GHS}, $h_1 (\fT)$ is a template of $h_1\circ \phi_t$.
\end{proof}
Besides the ambient isotopy, there are two additional template moves adopted for defining equivalence among embedded templates, the splitting move and the switch move \cite{BW2, KSS}.
\begin{center}
\setlength{\unitlength}{0.35mm}
\begin{picture}(100,50)(-50,-20)
\put(-20,0){\line(0,-1){20}}
\put(-5,0){\line(0,-1){20}}
\put(20,0){\line(0,-1){20}}
\put(5,0){\line(0,-1){20}}
\bezier{15}(-20,0)(-10,10)(0,20)
\qbezier(10,30)(5,25)(0,20)
\put(-20,0){\line(-1,1){30}}
\put(20,0){\line(-1,1){30}}
\put(20,0){\line(1,1){30}}
\put(-5,0){\line(1,0){10}}
\linethickness{.1pt}
\put(-20,0){\line(1,0){15}}
\put(5,0){\line(1,0){15}}
\put(9,0){\makebox(0,0)[lc]{$\quad \quad \quad \quad \iff$}}
\end{picture}
\ \ \ \ \ \
\begin{picture}(100,50)(-50,-20)
\put(-20,0){\line(0,-1){20}}
\put(-5,0){\line(0,-1){20}}
\put(20,0){\line(0,-1){20}}
\put(5,0){\line(0,-1){20}}
\bezier{19}(-20,0)(-10,10)(0,20)
\qbezier(-12.5,7.5)(-10,10)(-7.5,12.5)
\qbezier(10,30)(5,25)(0,20)
\put(-20,0){\line(-1,1){30}}
\put(20,0){\line(-1,1){30}}
\put(20,0){\line(1,1){30}}
\put(-5,0){\line(-1,1){20}}
\put(5,0){\line(-1,1){20}}
\put(-25,20){\line(1,0){10}}
\qbezier(-5,0)(-2.5,2.5)(0,5)
\bezier{8}(0,5)(3.75,8.75)(7.5,12.5)
\qbezier(7.5,12.5)(12.5,17.5)(17.5,22.5)
\bezier{8}(5,0)(8.75,3.75)(12.5,7.5)
\qbezier(12.5,7.5)(20,15)(27.5,22.5)
\put(17.5,22.5){\line(1,0){10}}
\linethickness{.1pt}
\put(-20,0){\line(1,0){15}}
\put(5,0){\line(1,0){15}}
\end{picture}
\setlength{\unitlength}{1.0mm}
\begin{picture}(20,25)(-10,-10)
\put(0,0){\line(0,-1){10}}
\put(0,0){\line(1,1){7}}
\put(0,0){\line(-1,1){7}}
\put(-3,3){\line(1,1){5}}
\put(9,0){\makebox(0,0)[lc]{$\times I \quad \iff$}}
\end{picture}
\ \ \ \ \ \ \ \ \ \ \ \ \ \
\begin{picture}(20,25)(-10,-10)
\put(0,0){\line(0,-1){10}}
\put(0,0){\line(1,1){7}}
\put(0,0){\line(-1,1){7}}
\put(3,3){\line(-1,1){5}}
\put(9,0){\makebox(0,0)[lc]{$\times I$}}
\end{picture}
\end{center}
A dynamical reason for introducing these two moves is from the Conley index theory \cite{C}, that the dynamics of a invariant set of a flow is invariant if the relative homotopy type of the index pair of this invariant set is unchanged \cite{Sa}.
However, obviously the splitting move changes the homeomorphism type of templates, and still less the isotopy type. This is the essential difficulty to define an isotopic invariant for embedded templates, such that equivalent templates have the same invariant.
In the next section, we remedy it by choosing a fixed homeomorphism type of an embedded template, and then merely consider the action of ambient isotopy on the template without involving the two template moves.
\section{Templates and spatial graphs}
This section focuses on the boundary of templates. If the ambient manifold is a $3$-space, \ie $\R^3$ or $S^3$, the boundary of a template can be regarded as a trivalent spatial graph. We have the following observation.
\begin{proposition} \label{isog}
Let $\fT$ be a template in $\R^3$ or $S^3$, and $\{ h_\la \}_{\la \in I}$ be an ambient isotopy. Then $\pa \fT$ is isotopic to $\pa h_1 (\fT)$ as spatial graph.
\end{proposition}
\begin{proof}
The proof is straightforward by
$$\pa h_1(\fT)= h_1 (\pa \fT)\sim \pa \fT,$$
where the equality holds because of the diffeomorphism $h_1$ commuting with $\pa$, and $\sim$ means isotopy of spatial graphs due to the ambient isotopy.
\end{proof}
From the viewpoint of topology, the spatial graph is a bit more complex object than the knot or link. Now we use Kauffman's invariant of trivalent spatial graphs to dispose the boundary of a template, and get a collection of knots and links as our isotopic invariant of embedded templates.
Let us recall the definition of Kauffman's invariant in \cite{K}. For a trivalent spatial graph $G$ in a $3$-space, there are three types of local replacement at a vertex $v$ of $G$ as follows.
\begin{center}
\setlength{\unitlength}{1.0mm}
\begin{picture}(20,28)(-10,-14)
\put(0,0){\line(0,-1){10}}
\put(0,0){\line(1,1){7}}
\put(0,0){\line(-1,1){7}}
\put(4,0){\makebox(0,0)[lc]{$v \quad \longrightarrow$}}
\end{picture}
\ \ \ \
\begin{picture}(20,28)(-10,-14)
\put(0,0){\line(0,-1){10}}
\put(0,0){\line(1,1){7}}
\put(-1,1){\line(-1,1){6}}
\end{picture}
\
\begin{picture}(20,28)(-10,-14)
\put(0,0){\line(0,-1){10}}
\put(1,1){\line(1,1){6}}
\put(0,0){\line(-1,1){7}}
\end{picture}
\
\begin{picture}(20,28)(-10,-14)
\put(0,-1.5){\line(0,-1){8.5}}
\put(0,0){\line(1,1){7}}
\put(0,0){\line(-1,1){7}}
\end{picture}
\end{center}
Let $\tau(G)$ be the collection of knots and links obtained by making a local replacement of all the three types at each vertex of $G$.
Moreover, for a path or cycle $\alpha$ in $G$, we define the generalized Kauffman's invariant of $G$ by
$$\tau_\alpha (G)\defeq \{l\in \tau(G):\alpha \subset l\}.$$
That is, the subset of knots and links in $\tau(G)$ which contain $\alpha$.
For example, for a unknotted $\theta$-curve $G$ with three edges $a,b$ and $c$, $\tau(G)$ is the set of three trivial knots $\{ab,bc,ac \}$. Hence, the generalized Kauffman's invariant $\tau_b(G)$ is the set $\{ab,bc \}$ consisting of two unknots.
We say that two Kauffman's invariants $\tau(G)$ and $\tau(G')$ are equal if they are the same set of link types. Similar to generalized Kauffman's invariants.
\begin{theorem}(Kauffman \cite{K}) \label{tau}
For a trivalent spatial graph $G$ in a $3$-space, $\tau(G)$, taken up to ambient isotopy, is a topological invariant of $G$.
\end{theorem}
Here "topological" means embedding.
\begin{corollary} \label{taua}
$\tau_\alpha (G)$ is an isotopic invariant of $G$ for any path or cycle $\alpha$ in $G$.
\end{corollary}
Combining Proposition~\ref{isog} and Kauffman's theorem~\ref{tau}, we get the following consequence immediately.
\begin{theorem} \label{taut}
Let $\fT$ ba a template in a $3$-space, and $\{ h_\la \}_{\la \in I}$ be an ambient isotopy. Then the Kauffman's invariants of $\pa \fT$ and $\pa h_1 (\fT)$ are equal, i.e.
$\tau (\pa \fT) = \tau (\pa h_1(\fT))$
as sets of link types.
\end{theorem}
\begin{remark}
The effect of the splitting and the switch moves on the Kauffman's invariant of a template boundary is explicit. For an embedded template $\fT$, the set $\tau (\pa \fT)$ of knots and links is invariant under the switch move. On the other hand, from the template diagram \cite{KSS} we can see that doing a splitting move induces adding or decreasing exactly one link to $\tau (\pa \fT)$, which is split into a link in the original set with a unlinked trivial component.
\end{remark}
In Section 5, we will apply this invariant to study the isotopic classification of simple Smale flows on $S^3$.
\section{A complete invariant of templates}
Proposition~\ref{isog} states that ambient isotopic templates have isotopic boundaries. Conversely, we have the following criterion on determining if two templates in $3$-sphere are ambient isotopic. Therefore, it is equivalent to say that the template boundary, as a spatial graph, is a complete invariant of embedded templates with a fixed homeomorphism type.
\begin{theorem} \label{crit}
Let $\fT$ and $\fT'$ be two embedded templates in $S^3$, which are topologically equivalent as dynamical systems. If they have isotopic boundaries
$\pa \fT\sim \pa \fT'$,
then $\fT$ and $\fT'$ are ambient isotopic templates.
\end{theorem}
The construction of ambient isotopies relies on the Isotopy Extension Theorem \cite{Hir}, which states that an isotopy of a compact submanifold of $S^3$ always extends to an ambient isotopy having compact support. To prove Theorem~\ref{crit}, we need the following two lemmas on embedded discs, annuli, and M\"{o}bius strips.
\begin{lemma}\label{lemd}
If two embedded discs $D$ and $D'$ in $S^3$ have the same boundaries $\pa D=\pa D'$, then there is an ambient isotopy sending $D$ to $D'$.
\end{lemma}
\begin{proof}
By transversality, $D$ can be perturbed by a small isotopy to intersect $D'$ transversally. The intersection is the union of a submanifold of dimension $1$ and the common boundary, then it consists of three parts: $\pa D=\pa D'$, several families of finite concentric circles, and a finite number of intervals whose endpoints lie on the boundary.
First, consider the circles in $D\cap D'$. Choose a family of concentric circles in $D$ with innermost one $C$, which is also a innermost one in $D'$. The two discs $D_1$ and $D_1'$, bounded by $C$ in $D$ and $D'$ respectively, form an embedded $2$-sphere in $S^3$. By Alexander's theorem on sphere \cite{Ha}, the $2$-sphere $D_1\cup D_1'$ bounds an embedded $3$-ball $B$ in $S^3$. Then by an isotopy of $D$ supported near $C$ we can push $D$ across $B$, eliminating $C$ and decreasing by $1$ the number of components of $D\cap D'$. Repeat this step of eliminating circle components of $D\cap D'$ as long as only $\pa D=\pa D'$ remains.
Second, consider the intervals in $D\cap D'$. We can find out an "outmost" interval $\alpha$ in $D\cap D'$. Hence, there is a sub-arc $\beta$ in $\pa D$ such that $\alpha \cup \beta$ bounds two discs $D_2$ and $D'_2$ in $D$ and $D'$ respectively. $D_2$ and $D'_2$ intersect exactly at the circle $\alpha \cup \beta$. As for $D_1$ and $D_1'$, an isotopy of $D$, supported near $\alpha \cup \beta$, eliminates $\alpha$ and decreases by $1$ the number of intervals in $D\cap D'$. Thus doing this step repeatedly allows us to eliminate all the intersection intervals.
Now $D$ and $D'$ intersect only at $\pa D=\pa D'$ and form an embedded $2$-sphere in $S^3$, which bounds an embedded $3$-ball giving an isotopy of $D$ and $D'$.
According to the Isotopy Extension Theorem, all the above isotopies extend to ambient isotopies of $S^3$. Gluing these finite number of isotopies gives an ambient isotopy sending $D$ to $D'$.
\end{proof}
\begin{lemma}\label{lema}
Let $A$ and $A'$ be two embedded annuli or two embedded M\"{o}bius strips in $S^3$. If $\pa A=\pa A'$, then there is an ambient isotopy sending $A$ to $A'$.
\end{lemma}
\begin{proof}
Let $A$ and $A'$ be two embedded annuli in $S^3$. First consider the special case that $A=A_0$ is a standard annulus, \ie $A_0$ is unknotted and its two boundary circles are unlinked.
By transversality, after a small isotopy $A_0$ and $A'$ intersect transversally. The intersection is composed of $\pa A_0 =\pa A$, a finite number of circles, and a finite number of intervals with endpoints lying on the boundary. We can repeat the procedure in the proof of the previous lemma to eliminate all the inessential circles in $A_0\cap A'$.
For an interval $\alpha$ in $A_0\cap A'$, if it connects the two components of $\pa A_0$, then $A_0\cap A'$ contains no essential circles. $\alpha$ cuts $A_0$ and $A$ into discs, and by Lemma~\ref{lemd} $A_0$ and $A$ are isotopic. If the two endpoints of $\alpha$ lie on one of the two boundary circles, there is a sub-arc $\beta$ in this component satisfying that $\alpha \cup \beta$ bounds two discs $D$ and $D'$ in $A_0$ and $A'$ respectively, and
$$D\cap D' =\alpha \cup \beta.$$
Then as in Lemma~\ref{lemd}, an isotopy of $A_0$, supported near $\alpha \cup \beta$, eliminates $\alpha$ and decreases by $1$ the number of intervals in $A_0\cap A'$. Thus by repeating this step we can eliminate all the intersection intervals.
For an essential circle $C$ in $A_0\cap A'$, if there is another essential one $C'$ adjacent to $C$ and not in the boundary, $C$ together with $C'$ bounds two annuli $A_1$ and $A_1'$ in $A_0$ and $A'$ respectively. $A_1$ and $A_1'$ intersect transversally only at boundaries thus form an embedded torus in $S^3$. By Alexander's theorem on torus \cite{Ha}, the torus $A_1\cup A_1'$ bounds a solid torus $X$ in $S^3$. Then by an isotopy of $A_0$ supported near $C\cup C'$, we can push $A_0$ across $X$, eliminating $C$ and $C'$, and decreasing by $2$ the number of circle components of $A_0\cap A'$. Repeat this step of eliminating circles in $A_0\cap A'$, as long as only $\pa A_0=\pa A'$ remains, or with a single essential circle.
If no other circle than $\pa A_0=\pa A'$ remains, $A_0$ and $A'$ form an embedded torus in $S^3$, which bounds an embedded solid torus giving an isotopy of $A_0$ and $A'$. If there is exactly one essential circle $C''$ in $A_0\cap A'$ but not in boundary, $C''$ together with a component of $\pa A_0=\pa A'$ bounds two annuli in the original annuli respectively. As above, they bound an embedded solid torus giving an isotopy of $A_0$ and $A'$ to eliminate $C''$, and then no other circles than $\pa A_0=\pa A'$ exists.
According to the Isotopy Extension Theorem, all the above isotopies extend to ambient isotopies of $S^3$. Gluing these finite number of isotopies gives an ambient isotopy sending $A_0$ to $A'$.
For general annulus $A$, there is a homeomorphism $h\col S^3\to S^3$ satisfying $h(A) =A_0$. By the special case, there is an isotopy $h_\la \col S^3\to S^3$ such that
$$h_0= 1_{S^3},\quad h_0 (A_0)= A_0,\quad h_1 (A_0)= h(A').$$
Then $h^{-1}\circ h_\la \circ h$ is an ambient isotopy sending $A$ to $A'$.
The proof for the case of M\"{o}bius strips is similar, and is simpler than the case of annuli by the fact that if there is an interval in $A\cap A'$ with endpoints lying on the boundary, then the intersection contains no essential circles.
\end{proof}
We can now prove Theorem~\ref{crit}.
\begin{proof}
$\fT$ and $\fT'$ are topologically equivalent templates thus are homeomorphic as branched surfaces. We construct an ambient isotopy of $S^3$ sending $\fT$ to $\fT'$ in three steps. Then the semi-flow condition holds clearly, and the two templates are ambient isotopic.
Assume that $\fT$ and $\fT'$ are both connected. Otherwise, for each component we construct a desired isotopy supported near it, and glue them together.
(1) Since $\fT$ and $\fT'$ have isotopic boundaries, after a suitable isotopy of $S^3$ they have common boundary
$\pa \fT =\pa \fT'$.
(2) Let us tackle branch lines. Let $\alpha_1,\cdots, \alpha_n$ be the branch lines of $\fT$ while $\alpha'_1,\cdots, \alpha'_n$ be that of $\fT'$, satisfying that $\alpha_i$ and $\alpha'_i$ have the same endpoints on
$\pa \fT =\pa \fT'$
for $i=1,\cdots, n$. According to the template diagrams of $\fT$ and $\fT'$ \cite{KSS}, for every $i$, the circle consisting of $\alpha_i$ and $\alpha'_i$ cannot be knotted. As in the proof of Lemma~\ref{lemd} and by the Schoenflies Theorem, $\alpha_i$ and $\alpha'_i$ form the boundary circle of an embedded disc in $S^3$, which gives an isotopy of the two arcs relative the endpoints and with support near the circle. Extending these $n$ isotopies and gluing them together give an isotopy of $S^3$ sending each $\alpha_i$ to $\alpha'_i$ and preserving
$\pa \fT= \pa \fT'$.
(3) The branch lines divide the template into pieces which can be viewed as splitting charts. For a piece $S$ of $\fT$, let $S'$ be the corresponding one in $\fT'$, then one of the following three cases occurs.
(i) If the two bands of $S$ are disjoint with its branch line, $S$ and $S'$ are embedded discs in $S^3$. By Lemma~\ref{lemd}, there is an ambient isotopy sending $S$ to $S'$ and preserving
\begin{equation*}
\pa \fT =\pa \fT', \quad \alpha_i=\alpha'_i, i=1,\cdots, n.
\end{equation*}
(ii) If one band is disjoint with the branch line while the other one is not, both $S$ and $S'$ have a closed band, and are either embedded annuli or M\"{o}bius strips. Then by Lemma~\ref{lema}, an ambient isotopy sends $S$ to $S'$ and preserves the coincidence of the boundary and branch lines.
(iii) If all the two bands are closed, by connectivity both $S$ and $S'$ have exactly one branch line thus are Lorenz-like templates.
The exit line $\beta$ of $S =\fT$ sweeps a square domain $B$ by going backward through the semi-flow until touching the branch line $\alpha$. So does for
$S' =\fT'$ to get a square domain $B'$. Then $B$ and $B'$ have a pair of opposite common sides, the exit line $\beta =\beta'$ and an interval $\gamma =\gamma'$ of the branch line, while the other sides, $\delta_1, \delta_2$ in $\fT$ and $\delta'_1, \delta'_2$ in $\fT'$, are part of flow lines with the same endpoints
$$\pa \delta_1=\pa \delta'_1, \quad \pa \delta_2=\pa \delta'_2.$$
As in Step (2),
there is an ambient isotopy sending each $\delta_i$ to $\delta'_i$ and preserving
$$\pa \fT= \pa \fT', \quad \alpha =\alpha'.$$
Now the two discs $B$ and $B'$ have common boundary, Lemma~\ref{lemd} implies they are ambient isotopic. Both the complements $\fT -B$ and $\fT' -B'$ consist of two bands, we can cut them off along branch lines and use Lemma~\ref{lema} twice to isotope $\fT -B$ to $\fT' -B'$.
To sum up, there is always an ambient isotopy $h_t$ sending $S$ to $S'$ and preserving $\pa \fT =\pa \fT'$. Moreover, we can ask it to be supported near $S$ by the compactness.
For another piece of $\fT$ an ambient isotopy, sending it to the corresponding one of $\fT'$, can be constructed as above. It can be required to preserve the coincidence of the boundary and branch lines, as well as
$$h_1(S) =S'.$$
Thus gluing $n$ ambient isotopies for all the $n$ pieces of $\fT$ gives an isotopy of $S^3$ sending $\fT$ to $\fT'$. Now we finish Step (3).
All the ambient isotopies constructed in the three steps yield an ambient isotopy sending $\fT$ to $\fT'$. This completes the proof.
\end{proof}
The proof of the previous two lemmas works as well as for ambient manifolds with trivial $\pi_2$. This implies that the conclusion of Theorem~\ref{crit} holds in those cases. We have the following corollary.
\begin{corollary}
Let $M$ be an irreducible closed $3$-manifold. Let $\fT$ and $\fT'$ be two embedded templates in $M$, which are topologically equivalent as dynamical systems. If they have isotopic boundaries
$\pa \fT\sim \pa \fT'$,
then $\fT$ and $\fT'$ are ambient isotopic as templates.
\end{corollary}
\begin{remark}
Theorem~\ref{crit} can be viewed as an isotopic version of Conley index theory. For a flow $\phi_t$ on $S^3$, a saddle set $\Om$ induces a template $\fT$. Then the thicken template $\bar{\fT}$ is an isolated neighborhood of $\Om$, while the boundary $\pa \fT$ is exactly the core of the exit set $\pa_- \bar {\fT}$. Our theorem shows that the embedding of exit set $\pa_- \bar {\fT}$ totally determines not only the homeomorphism type of the Conley index pair $(\bar{\fT}, \pa_- \bar {\fT})$, but also the isotopy type of their embedding in the underlying $3$-manifold.
\end{remark}
\section{Isotopy of simple Smale flows}
\label{secMldCalc}
In dynamics, topologically equivalent flows have the same orbit structure hence the same dynamical properties, but their corresponding orbits may be embedded in the underlying manifold in different ways, or in other words by physicists, they may have different orbit organizations \cite{Gil}. Embedding of objects in the ambient space is usually under consideration in topology up to isotopy \cite{Hir}.
In this section, we define the isotopy of flows, which is more subtle than the topological equivalence in the viewpoint of classifying. It is naturally related to the ambient isotopy and is useful to describe the embedding of orbits into the underlying manifold. Then we concentrate on the classification of simple Smale flows on $S^3$ up to isotopy by using the induced spatial graph and its generalized Kauffman's invariant.
Let $M$ be a connected compact smooth manifold. All the flows we consider here are supposed to be smooth.
\begin{definition}
Two flows $\phi_t$ and $\psi_t$ on $M$ are isotopic if one of them can be deformed to the other through smooth flows, \ie there is a smooth map
$$H\col \R \times M\times I \to M, \quad (t, x, \la)\mapsto H_\la (t, x)$$
such that $H_0 (t, x) =\phi_t (x)$, $H_1 (t, x) =\psi_t (x)$, and for any $\la \in I$, $H_\la (\cdot, \cdot)$ is a smooth flow on $M$. Then we denote
$\phi_t \sim \psi_t$, and call $H_\la$ an isotopy from $\phi_t$ to $\psi_t$.
\end{definition}
The isotopy of two flows is an equivalent relation (for the transitivity, see p. 111 of \cite{Hir}). There is a close relationship between isotopy of flows and ambient isotopy. If $\phi_t$ is a flow on $M$, $\{ h_\la \}_{\la \in I}$ is an ambient isotopy, then $\phi_t = h_0 \circ \phi_t$ and $h_1 \circ \phi_t$ are isotopic by defining
$$H_\la (t, x)= h_\la \circ \phi_t(x), \quad (t, x)\in \R \times M$$
for $\la \in I$. Conversely, we have the following proposition.
\begin{proposition}
Let $\phi_t$ and $\psi_t$ be two flows on $M$ which are isotopic. Then there exists an ambient isotopy $\{ h_\la \}_{\la \in I}$ so that
$h_1\circ \phi_t =\psi_t$.
\end{proposition}
\begin{proof}
If $H_\la$ is an isotopy from $\phi_t$ to $\psi_t$, then
$$h_\la (y)= H_\la (t, \phi_{-t} (y)), \quad y\in M$$
gives an ambient isotopy for $\la \in I$, satisfying that $h_0\circ \phi_t =\phi_t$ and $h_1\circ \phi_t =\psi_t$.
\end{proof}
Next we consider the relationship between topological equivalence and isotopy of flows in dynamics. For a flow $\phi_t\col M\to M$, denote the orbit of $x\in M$ by
\begin{equation*}
O_{\phi} (x)=\{ \phi_t (x): t\in \R\}.
\end{equation*}
\begin{proposition}
If two flows are isotopic, then they are topologically equivalent.
\end{proposition}
\begin{proof}
Suppose that $H_\la$ is an isotopy from $\phi_t$ to $\psi_t$ on $M$. As above,
$$h_\la (x)= H_\la (t, \phi_{-t} (x)), \quad x\in M$$
gives an ambient isotopy for $\la \in I$. Then $h_1$ is a diffeomorphism which takes orbits of $\phi_t$ to orbits of $\psi_t$ and preserves the time orientation, \ie for any $x\in M$,
$$h_1 (O_\phi (x))=O_\psi (h_1 (x)),$$
and for any $x\in M$, $\eps >0$, there is a $\delta > 0$ such that $\forall t\in (0,\delta)$, $\exists s\in (0,\eps)$,
$$h_1 (\phi_t (x))=\psi_s (h_1 (x)).$$
Thus $\phi_t$ and $\psi_t$ are topologically equivalent by $h_1$.
\end{proof}
Now we focus on simple Smale flows. First we give the general definition of the Smale flow.
\begin{definition} \cite{Fr}
Let $M$ be a connected orientable closed $3$-manifold. A flow $\phi_t\col M\to M$ is called a Smale flow if (1) its chain-recurrent set $\mathcal{R}(\phi_t)$ is hyperbolic, (2) the basic sets are of dimensions $0$ or $1$, and (3) for any $x, y\in \mathcal{R}(\phi_t)$, the stable manifold $W^s (x)$ of $x$ and the unstable manifold $W^u (y)$ of $y$ have transversal intersection.
In particular, a simple Smale flow (SSF) is a Smale flow whose chain-recurrent set $\mathcal{R}(\phi_t)$ is composed of exactly three basic sets- a closed orbit attractor $a$, a closed orbit repeller $r$, and a nontrivial saddle set $\Om$ \cite{Su, Y}.
\end{definition}
Next we consider SSFs on $3$-sphere by using spatial graphs they induces and the generalized Kauffman's invariant.
\begin{theorem} \label{lemg}
Let $\phi_t$ be a simple Smale flow on $S^3$ with attractor $a$ and a template $\fT$. Then the spatial graph composed of $a$ and the boundary of $\fT$
$$G\defeq a\cup \pa \fT$$
is invariant under isotopy of $\phi_t$. Furthermore, the generalized Kauffman's invariant $\tau_a (G)$ is an isotopic invariant of $\phi_t$.
\end{theorem}
\begin{proof}
Let
$\{ h_\la \}_{\la \in I}$
be an ambient isotopy. Due to isotopy preserving the chain-recurrent set decomposition all the time and $h_0\circ \phi_t =\phi_t$,
$\{ h_\la \circ \phi_t\}_{\la \in I}$
is a family of SSFs. Therefore $h_1(a)$ is the closed orbit attractor of $h_1\circ \phi_t$, and it is isotopic to $a$. By Proposition~\ref{isog}, $\pa \fT$ is isotopic to $\pa h_1(\fT)$. Since all the linking relationships are preserving under ambient isotopy,
$$h_1(G) = h_1(a)\cup h_1(\pa \fT) = h_1(a)\cup \pa h_1(\fT) \sim a\cup \pa \fT = G$$
as spatial graphs.
By Corollary~\ref{taua}, $\tau_a (G)$ is also invariant under $\{ h_\la \}_{\la \in I}$.
\end{proof}
As for Proposition~\ref{isog}, we can ask the inverse questions of Theorem~\ref{lemg} that
(1) wether or not the spatial graph $G= a\cup \pa \fT$ is a complete isotopic invariant of SSFs on $S^3$; and
(2) wether or not the generalized Kauffman's invariant $\tau_a (G)$ is complete.
For the first question, we give a general discussion on the structure of SSFs \cite{Su, Y}, which is fundamental for the classification problem. Let $\phi_t$ be a SSF on a $3$-manifold $M$ with attractor $a$, repeller $r$, and saddle set $\Om$.
For the saddle set $\Om$, choose a template $\fT$. As a topological space, $\fT$ is a connected compact branched $2$-manifold. The thicken template $\bar{\fT}$ is an isolated neighborhood of $\Om$, which is generally a knotted handlebody embedded in $M$. Assume the genus of $\bar{\fT}$ is $g$, then $\pa \bar{\fT}$ is a connected orientable closed surface of genus $g$. It is composed of two compact subsurfaces- the exit set $\pa_- \bar{\fT}$ and the entrance set
$\pa_+ \bar{\fT}$, which are in fact homeomorphic
$$\pa_- \bar{\fT} \cong \pa_+ \bar{\fT}$$
and intersect at a finite number of disjoint circles as their common boundary. That is,
$$\pa (\pa_- \bar{\fT})= \pa (\pa_+ \bar{\fT})$$
is a disjoint union of finite circles. The core of the exit set $\pa_- \bar{\fT}$ is exactly $\pa \fT$, the boundary of the template.
Let $A$ and $R$ be the tubular neighborhoods of the attractor $a$ and the repeller $r$ in the underlying $3$-manifold $M$ respectively. Since $a$ and $r$ are generally knotted, $A$ and $R$ are knotted solid tori in $M$ with knotted tori $\pa A$ and $\pa R$ as their boundaries.
The original flow $(M, \phi_t)$ can be reconstructed by attaching $A, \bar{\fT}$ and $R$ along their boundaries in the way that gluing $\pa_- \bar{\fT}$ to $\pa A$, $\pa_+ \bar{\fT}$ to $\pa R$, and the remainder part of $\pa A$ to the remainder part of $\pa R$. Both the manifold and flow obtained from the attaching can be smooth by a modification such as in \cite{M}. Let
$$N\defeq A\cup \bar{\fT}$$
be the result of attaching $\bar{\fT}$ to $A$.
The following theorem gives a partial answer of the first question in the case of SSFs on $3$-sphere. It essentially relies on Theorem~\ref{crit} and the fact that the knot complement in $S^3$ is a complete invariant.
\begin{theorem} \label{flowg}
Let $\phi_t$ and $\psi_t$ be two simple Smale flows on $S^3$ with attractors $a$ and $a'$ respectively. Suppose that $\phi_t$ and $\psi_t$ restricted on their saddle sets are topologically equivalent. Then they have topologically equivalent templates, denoted by $\fT$ and $\fT'$ respectively. Moreover, if
$G= a\cup \pa \fT$ and $G'=a' \cup \pa \fT'$ are isotopic spatial graphs, then $\phi_t$ and $\psi_t$ are isotopic flows, after reversing the directions of the attractor or the repeller if necessary.
\end{theorem}
\begin{proof}
Since $\phi_t$ and $\psi_t$ restricted on their saddle sets are topologically equivalent, from the constructing procedure of templates following \cite{BW2, GHS}, they have topologically equivalent templates clearly.
Let $\fT$ and $\fT'$ be two topologically equivalent templates of $\phi_t$ and $\psi_t$ respectively. By the above reconstruction of a SSF, it is sufficient to construct an ambient isotopy preserving the isolated neighborhoods decomposition
$$S^3= A\cup \bar{\fT}\cup R$$
of $\phi_t$ for all time and connecting the two flows.
If $G= a\cup \pa \fT$ and $G'=a' \cup \pa \fT'$ are isotopic spatial graphs, there is an ambient isotopy $\{ f_\la \}_{\la \in I}$ sending $G$ to $G'$. Then for $f_1(\fT)$ and $\fT'$, $\pa f_1(\fT) =\pa \fT'$, one can construct an isotopy $\{ g_\mu \}_{\mu \in I}$ of $S^3$ sending $f_1(\fT)$ to $\fT'$ and keeping
$\pa f_1(\fT) =\pa \fT'$
as in the proof of Theorem~\ref{crit}. The isotopy $\{ g_\mu \}_{\mu \in I}$ has compact support near $f_1(\fT)$ thus can be required to fix $f_1(a) =a'$. Gluing $\{ f_\la \}_{\la \in I}$ and $\{ g_\mu \}_{\mu \in I}$ together yields an ambient isotopy $\{ h_\nu \}_{\nu \in I}$ satisfying that
$$h_1(G) =G', \quad h_1(\fT) =\fT'.$$
Then $\{ h_\nu \}_{\nu \in I}$ sends the thicken template $\bar{\fT}$ into $\bar{\fT'}$, as well as the tubular neighborhood $A$ into a tubular neighborhood $A'$, and preserves all the linking relationships including for bands of the template and for the attractor with the template. Therefore, the complements of the repeller $r$ and $r'$ are corresponding under $\{ h_\nu \}_{\nu \in I}$. Since the complement is a complete invariant of knots \cite{GL}, the isolated neighborhoods decomposition is invariant by $\{ h_\nu \}_{\nu \in I}$.
The argument leaves out the direction of the closed orbit attractor and repeller, \ie they are under consideration only as point-sets. Nevertheless, if necessary, we can reverse the direction of $a$ or $r$ such that $h_1\circ \phi_t =\psi_t$.
\end{proof}
For the second question that wether or not the generalized Kauffman's invariant $\tau_a (G)$ is a complete invariant for SSFs on $S^3$, the answer seems to be no. A reason is that the Kauffman's invariant is not powerful enough to distinguish spatial graphs. In fact, even for $\theta$-curves, non-isotopic spatial graphs may have the same Kauffman's invariant \cite{Zh}. Thus it is quite a question.
\begin{question}
For a simple Smale flow on $S^3$ with attractor $a$ and a template $\fT$, let $G= a\cup \pa \fT$. Suppose that the homeomorphism type of $\fT$ is fixed. What is the condition of $\fT$ such that the generalized Kauffman's invariant $\tau_a (G)$ is a complete isotopic invariant of the flow?
\end{question}
\begin{remark}
The invariant $\tau_a (G)$ is complete for SSFs on $S^3$ with Lorenz-like templates \cite{Su, Y} or the universal template and its analogues \cite{HS, AS}, up to the disc sign.
\end{remark}
| 5ee3f845dbf286badc3a11fef8a4e9e735f29110 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:introduction}}
\else
\section{Introduction}
\label{sec:introduction}
\fi
\noindent Analysing the direction of influence between participants in dyadic interactions (i.e., human communication between two people) is of increasing interest in the psychological sciences with applications in psychotherapy \cite{Tschacher2014}, critical political interactions \cite{Renoust2016}, and interpersonal conflict behavior \cite{Reeck2016}, among others. In any interpersonal communication, attitudes towards the interacting partner play a crucial role for the outcome of this interaction (e.g., liking or disliking, engagement or disengagement, trust or no trust towards the interacting partner) \cite{Ambady2000,Todorov2005,Fiske2007}. Often self-report measures or verbal cues are used for getting at attitudes towards the interacting partner. However, most interacting partners are very adapted towards socially desirable behavior and thus would not express negative attitudes in an overt manner. The concept of embodiment \cite{Tschacher2014,Gallese2005,Fuchs2009} denotes the theoretical perspective that mental processes are not isolated from bodily processes. With this view, psychology is becoming increasingly sensitized to investigate the close association between mental and bodily parameters. In this paper, we are interested in the analysis of the direction of emotional influence in dyadic interactions using non-verbal facial expressions.
Facial expressions, especially basic emotion expressions, are widely described as being automatic, thus hard to overtly influence and independent of culture \cite{Ekman1976}. As such, they seem a perfect measure for a better understanding of unintentional behavioral cues about social-emotional cognitive processes. Furthermore, there is recent evidence that non-verbal emotion processing is much more relevant than verbal emotion processing \cite{Filippi2017} and that language plays a marginal role in the perception of emotion \cite{Sauter2017}.
Novel developments in computer vision yielded accurate, open-source and real-time tools to easily extract bodily parameters from images and videos. In this paper, a complete concept for analyzing the direction of emotional influence in a dyadic face to face dialogues, when starting with raw video material, is presented. We exploit computer vision capabilities for quantitative verification of hypotheses on the direction of emotional influence in experimentally recorded dyadic dialogues using facial expressions only, i.e., facial muscle activation (i.e., Action Units- AUs) \cite{ekman2002facial}.
Human nonverbal communication is a process of continual two-sided influence. In a dyadic interaction, such influence occurs over transient time intervals with intensity and direction that are variant over time. To approach the problem of analysing the direction of influence, we first present each person's facial emotional expressions in the form of time series. To investigate who is influencing whom and how, we perform a causal effect analysis on the obtained time series using Granger causality (GC) \cite{geweke1982measurement,GRANGER1980329}, the most widely used method for causal inference in diverse fields \cite{Peters2017,ClimateGC2013,Seth2015,Shadaydeh2019}. GC is based on the idea that causes both precede and help predict their effects.
The contributions of this work can be summarized as follows.
\begin{itemize}
\item[1.] The use of GC for causal inference in nonverbal communication data has been addressed by several authors using different bodily parameters. However, to the best of our knowledge, no other work has used GC to identify the direction of emotional influence when it comes to facial expressions in dyadic dialogues.
\item[2.] Using facial expressions in the upper and lower face regions, we were able to extract fine-grained facial features from the six basic emotions \cite{ekman1992argument} which can be present even when strong emotional expressions are not visible.
\item[3.] We present a relevant interval selection approach that we use prior to causal inference to identify those transient intervals where causal inference should be applied. The validation of the relevant interval selection on synthetic data for improved causal inference has been presented in \cite{Mueller2019} along with initial results of this study on small data set. Here we show based on a larger real data set the superiority of such an approach in detecting the direction of emotional influence when compared to applying GC test on the entire time-series.
\item[4.] The current work is a study of covert attitudes in interacting partners from a 2nd person perspective \cite{Schilbach2013}, meaning in a truly interactive manner. In the past studying real-time social encounters in an interactive manner was described as the “dark matter” of social (neuro-) science as it is typically explored from a 3rd person perspective \cite{Schilbach2013}. Most studies in the field of social cognition \cite{Happe2017} are conducted in a manner that interacting partners observe other interacting partners from a distance (e.g., by judging facial expressions of another person from a picture). However, our work goes a step into the the "dark matter" and is looking at interacting partners that are readily involved, thus continuously responding to the others' actions.
\end{itemize}
The remainder of this paper is organized as follows. Related work are summarized in Section \ref{sec:Related}. We introduce the experimental setup and used psychological hypotheses in Section \ref{sec:ExperimentalSetup}. The methodological details are presented in Section \ref{sec:methodology} followed by experimental results and discussion in Section \ref{sec:results}. Finally, Conclusions are given in Section \ref{sec:conclusion}.
\section{Related Work}
\label{sec:Related}
The topic of finding causal structures in nonverbal communication data is addressed by Kalimeri et al. \cite{kalimeri2012modeling}. In their paper, the authors used GC for modeling the effects that dominant people might induce when it comes to nonverbal behavior i.e., speech energy and body motion, of other people. Besides audio cues, motion vectors and residual coding bit rate features from skin colored regions were extracted. In two systems, one for body movement and another one for speaking activity, with four-time series each, a small GC based causal network was used to identify the participants with high or low causal influence. Unlike our approach, the authors did not use facial expressions and did not identify relevant intervals in a previous step, but used the entire time series instead.
Kaliouby and Robinson \cite{el2005real} provided the first classification system for agreement and disagreement as well as other mental states based on nonverbal cues only. They used head motion and facial AUs together with a dynamic Bayesian network for classification. Also, a survey on cues, databases, and tools related to the detection of spontaneous agreement and disagreement was conducted by Bousmalis et al. \cite{bousmalis2013towards}. Despite their ingenious methods, these approaches did not investigate causal effect relations in a social interaction situation.
Postma and Postma-Nilsenova \cite{Postma2016} used the convergent cross mapping method (a causal inference method originating from dynamical systems theory) to study nonverbal interactions. They concluded that there exists bidirectional causal coupling in facial dynamics. Unlike our study, their method focuses on bidirectional causality only rather than identifying all possible directions of emotional influences. Further, their approach was applied to two time-series of single AU over the entire time-series rather than a combination of two or more AUs presenting emotions in selected intervals. Beyond this, Sheerman-Chase et al. \cite{sheerman2009feature} used visual cues to distinguish between states such as thinking, understanding, agreeing, and questioning to recognize agreement. Most intriguingly, Matsuyama et al. \cite{matsuyama2016socially} developed a socially-aware robot assistant responding to visual and vocal cues. For visual features, the robot extracted facial cues (based on OpenFace \cite{baltrusaitis2018openface}) using landmarks, head pose, gaze, and facial AUs. Conversational strategies that build, maintain, or destroy connecting relationships were classified. The researchers' approach investigates the building of a social relationship between a human and a robot; however, this study does not deal with the time-variant direction of causal effect relations.
Joo et al. \cite{joo2019towards} recently presented a motion capture data set and a neural network architecture to analyze the direction of influence in a triadic interaction. Their approach focuses on predicting social signals, such as speaking status, social formation, and body gesture. The face is represented using deformable face model with five facial expression parameters. In contrast to Joo et al., we are only interested in facial expressions. This requires a more detailed and supervised representation of facial motion.
Several methods, based on temporal causality analysis, have been proposed for categorizing interactions of individuals in video sequences \cite{Prabhakar2010,Prabhakar2012,Ayazoglu2013}. These methods use Granger causality, but they are based on motion analysis of different individuals and are not concerned with facial emotional influences.
\section{Experimental Setup and Psychological Hypotheses}
\label{sec:ExperimentalSetup}
\begin{figure}[!h]
\centering
\includegraphics[width=0.4\textwidth]{images/ExperimentalSetupBlue.pdf}
\caption{Experimental setup showing an example interaction pair (i.e, the sender and receiver) sitting face to face with camera positions towards each interacting partner's face. }
\label{fig:ExperimentalSetup}
\end{figure}
We created an experimental setup (Figure \ref{fig:ExperimentalSetup}) in which two participants sat face to face while talking about their personal weaknesses. One participant was in the assigned role of the receiver (R), the other was in the assigned role of the sender (S). S is instructed to take on a certain attitude (i.e., respectful, objective, contemptuous) whereas R is unaware of this and acts spontaneously.
In total, participants were asked to talk to each other about their personal weaknesses three times, either in circumstances of a respectful, contemptuous, or objective/neutral situation. Both partners were given specific times when to speak and when to listen. In all three experimental conditions, each participant kept their initially assigned role of acting as a S or R. The experimental conditions (i.e., respectful, contemptuous, objective) were conducted in various orders for all pairs to avoid the possible confounding effect of the order of these conditions. The objective/neutral condition consisted of the instruction that S and R should behave as neutral as possible (i.e., trying not to react readily towards the interacting partner). Instead, they should take their time to consider the world through the eyes of the other in order to understand their perspective in an uninvolved manner. As only S had the active experimental interaction attitude task (i.e., to behave either respectfully, contemptuously, or objectively), we
expected S to influence R in relevant facial expressions. In order to avoid flirtatious situations, that may overwrite the instructed condition, interacting partners were always from the same gender. See materials for the experimental instructions in the supplementary materials, part A.
Consistency of emotional influence was secured by the fact that after each conversation both interacting partners filled in a questionnaire in which they indicated in a self-report manner their positive affect during the interaction (i.e., on 4 items using a 7-point Likert scale). This measure revealed a clear linear function: In the respectful condition both interacting partner indicated the highest positive affect (mean value: M= 5.61 and standard deviation: SD= 0.91), followed by the objective condition (M = 5.40, SD = .98), followed by the contemptuous condition (M= 4.15, SD= 1.36).
To capture nonverbal facial behavior, we positioned two frontal perspective cameras (Figure \ref{fig:ExperimentalSetup}), recording at 25 frames per second. Camera positions and lighting conditions were optimized during a test session before the study started. This ensured high video quality in terms of a plain frontal view of the faces and two-sided illumination. Motion blur rarely occurred, but could not be prevented entirely, especially in cases of faster movements like head turns.
Except for the label of the experimental condition no other information (e.g., expression annotation per frame) were available for image analysis. The entire data set consisted of 34 pairs (mean age = 20.72, 24 female pairs, German-speaking participants), three conditions per pair and about four minutes of video per condition for each of the interacting partners, thus about 13 hours of video material. All participants gave written informed consent. The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Friedrich Schiller University of Jena. The ethical clearance number is FSV 18/04. Further, the study was pre-registered with aspredicted.org under the title ”Social Communication” $\#$13078 (See pre-registration in the supplementary materials, part B).
The psychological research question was, whether and how S and R influence each other given different attitude situations. First, we expected more harmonic expressions (i.e., happiness) when both interacting partners were confronted with medium to high levels of respect (i.e., respectful and objective/neutral vs. contemptuous). Further, we expected the strongest activation of negative expressions (e.g., sadness) in the disrespectful condition (i.e., contemptuous vs. respectful and objective/neutral). Last, we expected, for all emotional facial actions, S to cause the effects and influence R . In terms of the different facial expressions, we expected the strongest GC causality from S to R for positive expressions (i.e., happiness), followed by negative expressions (e.g., sadness).
\section{Methodology}
\label{sec:methodology}
Figure \ref{fig:workflow} shows the workflow of the proposed method. In the following subsections, we introduce the feature extraction procedure, the relevant interval selection approach, and the causal inference method. Then, we combine all these steps to elucidate our entire approach.
\begin{figure}[!th]
\centering
\includegraphics[width=0.5\textwidth]{images/workflow_final.pdf}\caption{
The workflow of the proposed concept for analyzing the direction of emotional influence in dyadic dialogues.}
\label{fig:workflow}
\end{figure}
\subsection{Facial Feature Extraction}
\label{sec:FeatureExtraction}
\noindent According to Ekman and Rosenberg \cite{ekman1997face}, facial expressions are the most important nonverbal signals when it comes to human interaction. The Facial Action Coding System (FACS) was developed by Ekman and Friesen \cite{ekman1978facial,ekman2002facial}. It specifies facial AUs, based on facial muscle activation. Examples of AUs are the \textit{inner brow raiser}, the \textit{nose wrinkler}, or the \textit{lip corner puller}. Any facial expression is a combination of facial muscles being activated, and thus, can be described by a combination of AUs. Hence, the six basic emotions (\textit{anger}, \textit{fear}, \textit{sadness}, \textit{disgust}, \textit{surprise}, and \textit{happiness}) can also be represented via AUs. Thus, when for example AU 6 (\textit{cheek raiser}), 12 (\textit{lip corner puller}), and 25
(\textit{lips part}) are activated the facial expression \textit{happiness} is visible \cite{langner2010presentation}.
In general, all facial dynamic expressions are visual nonverbal communication cues transferable to time-series. Regarding our real experimental data, this approach is reasonable for positive emotions like \textit{happiness}, which is frequently visible throughout the dyadic interactions. Yet, it is not applicable for negatively associated emotions such as \textit{anger, disgust, fear,} or \textit{sadness}, as these emotions tend to be rare in dyadic interactions and thus are often only visible in subtle manners. This is emphasized in Table \ref{tab:emotionCounts}, where we compute the visibility of emotions as defined in \cite{langner2010presentation} in all recordings.
\begin{table}[!h]
\centering
\caption{Percentage of frames where emotions were visible across all experimental conditions.}
\begin{tabular}{|c|c|}
\hline
Emotion & Occurrence (in \%) \\
\hline
Happiness & 12.9 \\
Surprise & 1.00 \\
Anger & .15 \\
Disgust & 3.07\\
Fear & .06 \\
Sadness & 1.55 \\
\hline
\end{tabular}
\label{tab:emotionCounts}
\end{table}
Wegrzyn et al. \cite{wegrzyn2017mapping} studied the relevance of facial areas for emotion classification and found differences in the importance of the eye and mouth regions. AUs can be divided into upper and lower AUs \cite{cohn2007observer}. Upper AUs belong to the upper half of the face and cover the eye region, whereas AUs in the lower face half cover the mouth region. We decided to split emotions into upper and lower emotions,
according to the affiliation of AUs to upper and lower face regions. For example, instead of using \textit{sadness} as a combination of AU1, AU4, AU15 and AU17 we used \textit{sadness upper} (AU1 and AU4) and \textit{sadness lower} (AU15 and AU17).
All other emotions were split according to their AUs belonging to the upper or lower facial half (Table \ref{tab:expressionAndAUs}). This procedure ensured that subtle facial expressions were also detectable.
\begin{table}[!h]
\centering
\caption{Expressions and the corresponding AUs.}
\label{tab:expressionAndAUs}
\begin{tabular}{|c|c|}
\hline
Expression & Active AUs \\
\hline
Happiness upper & 6 \\
Happiness lower & 12, 25 \\
Surprise upper & 1, 2, 5 \\
Surprise lower & 26 \\
Disgust lower & 9, 10, 25 \\
Fear upper & 1, 2, 4, 5 \\
Fear lower & 20, 25 \\
Sadness upper & 1, 4 \\
Sadness lower & 15, 17 \\
Anger upper & 4, 5, 7 \\
Anger lower & 17, 23, 24 \\
\hline
\end{tabular}
\end{table}
In Table \ref{tab:emotionCountsUpperLower} the detection percentage of upper and lower
expressions is illustrated. After splitting, the emotional expressions \textit{anger lower}, \textit{sadness lower},
\textit{sadness upper}, and \textit{surprise lower} could be detected in over 7\% of the video material on average.
Figure \ref{fig:AUSActivation} illustrates participant with different facial expressions.
\begin{table}[!ht]
\centering
\caption{Percentage of emotional expressions in the upper and lower face parts visible throughout
experiment.}
\label{tab:emotionCountsUpperLower}
\begin{tabular}{|c|c|}
\hline
Emotion & Detection (in \%) \\
\hline
\textit{ Anger lower } & 8.20 \\
\textit{ Anger upper } & .67 \\
\hline
\textit{ Disgust lower } & 3.07 \\
\hline
\textit{ Fear lower } & 5.94 \\
\textit{ Fear upper } & .97 \\
\hline
\textit{ Happiness lower } & 16.24 \\
\textit{ Happiness upper } & 26.32 \\
\hline
\textit{ Sadness lower } & 9.28 \\
\textit{ Sadness upper } & 7.05 \\
\hline
\textit{ Surprise lower } & 26.91 \\
\textit{ Surprise upper } & 2.48 \\
\hline
\end{tabular}
\end{table}
\begin{figure}[!th]
\centering
\includegraphics[width=0.4\textwidth]{images/Expressions.pdf}
\caption{ Participant with different facial expressions. From left to right: \textit{happiness lower}, \textit{sadness lower}, and \textit{sadness upper}}.
\label{fig:AUSActivation}
\end{figure}
\begin{figure}[!th]
\centering
\includegraphics[width=0.4\textwidth,height=3cm,keepaspectratio]{images/Face_FacePlusLandmarks.pdf}\\
(a)\\
\includegraphics[width=0.38\textwidth]{images/AUDetectionOnly.pdf}\\
\label{fig:AUDetectionOnly}
(b)\\
\caption{ Detection of facial expression \textit{angry}. (a) Landmarks of the facial expression; (b) Detected AUs by
OpenFace: Strong activation of AU4 (\textit{brow lowerer}), 7 (\textit{lip
tightener}), 14 (\textit{dimpler}), and 17 (\textit{chin
raiser})}
\label{fig:AUDetection}
\end{figure}
We evaluated the detection accuracy by mapping AUs to emotions based on the AffectNet dataset \cite{mollahosseini2017affectnet}. We reached an accuracy of $25.8 \%$ for the six basic emotions plus contempt plus neutral. Especially happiness was detected very well, with $77\%$ accuracy, whereas fear and surprise classification seem to be more complex. For feature extraction, we used OpenFace 2.0 \cite{baltrusaitis2018openface,baltruvsaitis2015cross} which is a state-of-the-art, open-source tool for landmark detection; it estimates AUs based on landmark positions. OpenFace preserves much of the information by regressing AUs instead of only classifying them and is capable of extracting 17 different AUs (1, 2, 4, 5, 6, 7, 9, 10, 12, 14, 15, 17, 20, 23, 25, 26, 45) with an intensity scaled from 0 to 5. Figure \ref{fig:AUDetection} illustrates the detection of landmarks and AUs for an example image.
Let AUI denotes the action unit I, e.g., AU15 is action unit 15. From the OpenFace detections, we compute each participant's average face across the three conditions. That is, for or each participant we compute the mean $\text{AUI}_{mean}$ and standard deviation $\text{AUI}_{std}$ of the activation of $\text{AUI}$ over all three experimental conditions. We then count $\text{AUI}$ as activated in time frame k, if $\text{AUI}_{k} \ge \text{AUI}_{mean} + 0.5 * \text{AUI}_{std}$. Then an expression is counted as activated when all its corresponding AUs are activated. For example, \textit{sadness lower} is considered to be visible in frame $k$ if $\text{AU15}_{k} \geq \text{AU15}_{mean} + .5*\text{AU15}_{std}$ and $\text{AU17}_{k} \geq \text{AU17}_{mean} + .5*\text{AU17}_{std}$ hold. Following this step, the number of activations per person and per expression was counted for each experimental condition, and normalized by the video length and maximum count of the expression.
\subsection{Causal Inference with Granger Causality}
\label{sec:GrangerCausality}
Let $X$ and $Y$ be two stationary time series with zero mean and length $L$. It is said that a time series $X$ Granger causes a time series $Y$ if the inclusion of past observations of $X$ beside $Y$ improves the prediction of $Y$ significantly when compared to the prediction using only past values of $Y$. These two time series can be represented by the following two vector autoregressive models.
\begin{equation} \label{eq:YgcX}
X_t = \sum_{j=1}^M a_{j} X_{t-j} + \sum_{j=1}^M
b_{j} Y_{t-j} + \varepsilon_t
\end{equation}
\begin{equation} \label{eq:XgcY}
Y_t = \sum_{j=1}^M c_{j} X_{t-j} + \sum_{j=1}^M
d_{j} Y_{t-j} + \vartheta_t
\end{equation}
\noindent where $\quad t = 1 \dots L$ denotes the time index, $\varepsilon_t$ and $\vartheta_t$ being two independent noise
processes. The model order $M$ defines the maximum lag used to estimate causal interactions. It can be estimated using either Akaike \cite{Akaike1974} or Bayesian Criterion \cite{Schwarz1978}. The model parameters $a_j,b_j,c_j,d_j, j=1, \ldots, M $ can then be estimated using, for example, the method of Least Squares (LS) \cite{Haykin1996}.
To test whether $Y$ Granger causes $X$, two vector autoregressive models are compared. The first model, in which $Y$ is included for predicting $X$, as in (\ref{eq:YgcX}). The second model is
\begin{equation} \label{eq:XsimpleModel}
X_{t} = \sum_{j=1}^M a'_{j} X_{t-j} + \varepsilon'_t
\end{equation}
where $Y$ is not included. Those models are then compared against each other via a statistical
significance test, where the null hypothesis $H_0$ is tested against the
alternative $H_1$, with
\begin{eqnarray}
H_0: b_1 = \dots = b_M = 0, \\
H_1: \exists b_k \neq 0 \quad k \in \{1 \dots M\}.
\end{eqnarray}
\noindent These hypotheses are equal to the variable selection problem in linear
regression. Hence, an F-Test is applicable with
\begin{equation}
F = \frac{(|\Sigma_{\varepsilon'}| - |\Sigma_\varepsilon|)(T-2M-1)}{|\Sigma_{\varepsilon'}|~M}
\end{equation}
where $\Sigma_{\varepsilon'}$ is the covariance matrix of the residual $\varepsilon_t'$ of the simple model in (\ref{eq:XsimpleModel}), and $\Sigma_\varepsilon$ is the covariance matrix of the residual $\varepsilon_t$ of the enriched model in (\ref{eq:YgcX}) . $F$ follows an F-distribution, with $(M, T-2M-1)$ degrees of freedom. The null hypothesis ($Y$ does not Granger cause $X$) can be rejected at a level of significance $\alpha$, if $ F > F_{1-\alpha}(x;M,T-2M-1)$ where $F_{1-\alpha}(x; M,T-2M- 1)$ denotes the value $x$, where the $F(x; M,T- 2M-1) = 1-\alpha$.
\noindent To test whether $X$ Granger causes $Y$, the above steps can be applied, but with the first model as in (\ref{eq:XgcY}), the second model being $Y_t = \sum_{j=1}^M
d'_{j} Y_{t-j} + \vartheta'_t$. When testing for GC , three different cases regarding the direction of influence can occur \cite{schulze2004granger}:
\begin{enumerate}
\item[1.] If $c_k = 0$ for $k = 1 \dots M$ and $\exists b_k \neq 0$ for $1 \leq k
\leq M$ then $Y$ Granger causes $X$.
\item[2.] If $b_k = 0$ for $k = 1 \dots M$ and $\exists c_k \neq 0$ for $1 \leq k
\leq M$ then $X$ Granger causes $Y$.
\item[3.] If for both $\exists b_k \neq 0$ for $1 \leq k \leq M$ and $\exists c_k \neq
0$ for $1 \leq k \leq M$ then a bidirectional (feedback) relation exists.
\end{enumerate}
\noindent If none of the above cases holds, $X$ and $Y$ are not Granger causing each other.
\subsection{Relevant Interval Selection }
\label{sec:InervalSelection}
\noindent Considering the experimental setup, we had to expect multiple temporal scenes, further referred to as subintervals, in which the participants influenced each other. The time spans where causality is visible might range from half a second to half a minute. This may occur several times and can be interrupted by irrelevant scenes that differ in the length of time. As outlined above, the direction of influence in a subinterval can either be bidirectional or unidirectional driven by either S or R . This implies that three unwanted effects can occur if the full-time span is analyzed. First, temporal relations are not found at all; second, bidirectional relations mask temporal
unidirectional relations and; third, unidirectional relations from $X$ to
$Y$ mask temporal bidirectional influences or unidirectional influences from
$Y$ to $X$ and vice versa.
The estimation accuracy of the causal effect intensity parameters $a_j, b_j, c_j, d_j, j=1, \dots, M$ is mainly influenced by the accuracy of estimating the correlation of the two time series $X$ and $Y$ at different time shifts. When the time series contain several intervals of irrelevant information, the transient similarity in $X$ and $Y$ may not be captured. Figure \ref{fig:corsubints} illustrates that two-time series can have highly correlated subintervals within low correlated full time span interval or low correlated subintervals embedded in high correlated full time span interval. That is to say, that similarity measures that are applied to the entire time range would fail to capture transient similarities.
Our central idea is to apply GC only to time series obtained by concatenating highly coherent (e.g., in terms of Pearson correlation) subintervals of raw time series. Instead of using a brute force algorithm, we suggest using a bottom-up approach for finding the longest set of maximal, non-overlapping, correlated intervals in time-series as proposed by Atluri et al. \cite{atluri2014discovering1}. The authors applied their approach to fMRI data where they achieved good results for clustering coherent working brain regions.
\begin{figure}[!h]
\centering
\includegraphics[width=0.45 \textwidth]{images/Subinterval_Correlation_Synthetic_Data.png}
\includegraphics[width=0.45\textwidth]
{images/Subinterval_Correlation_Synthetic_Data2.png}
\caption{Synthetic example of two time series showing wrong reasoning of correlation due to surrounding intervals. Upper graph shows an example of highly correlated subintervals of these time series (Pearson correlation parameter, r=0.85 and r=1) within low correlated interval (r=0.23). While the lower graph shows an example of weakly correlated subinterval (r=0.5) within highly correlated interval (r= 0.85).}
\label{fig:corsubints}
\end{figure}
Let $X$ and $Y$ be two time series of length $L$. An interval is called \textit{correlated interval} for a threshold $\beta$ when all its subintervals up to a lower interval length $L_{min}$ are correlated as well. An
interval $I_{(a,b)}$ from $a$ to $b$ is called maximal, when $I_{(a,b)}$ is a \textit{correlated interval}, but $I_{(a-1, b)}$ and $I_{(a,b+1)}$ are not. And two intervals $I_{(a,b)}$ and $I_{(c,d)}$ are called non-overlapping, when $I_{(a,b)} \cap I_{(c,d)} = \emptyset$. From all intervals fulfilling these conditions the longest set (total length of intervals) is computed.
In the multivariate case, e.g., when multiple AUs define an expression, we propose to compute the set of relevant intervals for each AU. Using the detection
of each AU pair, the intersection over all AUs can be used as the set of selected relevant intervals. In the following, the set of selected relevant intervals between two time-series $X$ and $Y$ is called $AW_{XY}$.
\begin{figure}[!h]
\includegraphics[width=0.5\textwidth]{images/AU6LaminaDetectionMedianFilter.pdf}\\
{\small (a) Relevant interval selection for AU6.}\\%\label{fig:MultiAU6}\\
\par \par
\includegraphics[width=0.5\textwidth]{images/AU12LaminaDetectionMedianFilter.pdf}\\
{\small (b) Relevant interval selection for AU12.}\\
\par \par
\includegraphics[width=0.5\textwidth]{images/RelevantIntervalDetectionExample.pdf}\\
{\small (c) Selected intervals for a facial expression with AU6 and AU12 being activated ((a) $\cap$ (b)). }\\
\par\par
\includegraphics[width=0.5\textwidth]{images/RelevantIntervalDetectionConcatExample.pdf}
{\small (d) Concatenation of the intervals highlighted in (c).}\\
\caption{Process of relevant interval selection in multivariate case, exemplified by AU6 and AU12.}
\label{fig:RIS}
\end{figure}
Figure \ref{fig:RIS} illustrates the selection process, exemplified by AU6 and AU12. First, the relevant intervals are computed for time
shifts $s \in \{0,4,8,12\}$ and minimum interval length $L_{min}=75$, between
S and R, for AU6 and AU12 separately. After that step,
additional processing steps, such as median filtering can be applied.
Figure \ref{fig:RIS} (a) shows the relevant intervals of AU6 and
Figure \ref{fig:RIS} (b) shows the relevant intervals for AU12, where both sets of relevant intervals were median filtered with kernel size 51. Thereafter, the selected relevant intervals are obtained by computing the intersection between the relevant intervals of AU6 and AU12, as visualized in Figure
\ref{fig:RIS} (c). Finally, the selected relevant intervals can be concatenated for further processing, as illustrated in Figure
\ref{fig:RIS} (d).
\subsection{Causal Inference with Relevant Interval Selection and Granger Causality}
\label{ref:ModelCauseEffectRelations}
\noindent
There are two major challenges in the analysis of the emotional causal effect relation in dyadic dialogues. First, due to the constructed situations, strong distinct emotions, computed by using traditional AU combinations, were barely visible. Second, the time-variant and situation-dependent communication, resulted in high variety and volatility of time spans in which causal effect behavior between interacting partners is visible. To tackle these difficulties, we use a combination of the time series of the facial features described in (\ref{sec:FeatureExtraction}) and the proposed relevant interval selection approach (\ref{sec:InervalSelection}) as detailed in the following steps.
\begin{itemize}
\item[1.] We applied the relevant interval selection approach pairwise to the time series of the identified AUs of all of the relevant facial expression as illustrated in Figure \ref{fig:RIS}, with a minimum interval length of 75 frames and a threshold of 0.8 for Pearson correlation. Based on known average human reaction times (ca. 200 ms or 6 frames \cite{jain2015comparative}), we shifted one time series by 0, 4, 8, and 12 frames both, back and forth in time, and computed relevant intervals. The grid selected for shifting does cover quicker and slower reactions of participants.
Afterwards, we computed the longest set of the list of relevant intervals obtained from the different shifts. Before computing GC, we median filtered the selected intervals with a filter length of 51 (2 seconds) and extended the intervals by 12 frames on each side. We removed frames for both, S and R, when for either S or R the OpenFace confidence value was below .89. The confidence score gives a rough orientation how reliable the AU score computed by OpenFace is. A low confidence score might occur when, for example, the face is occluded or the person moved quickly, which can lead to blurred images.
\item[2.] We calculated the average GC on the set of selected intervals of the standardized time-series. The results were counted according to the possible outcomes of the GC test as described in Section \ref{sec:GrangerCausality}, as either unidirectional caused by S, unidirectional caused by R, bidirectional, or no causality.
\end{itemize}
\section{Experimental Results and Discussion}
\label{sec:results}
\subsection{Appearance of facial expressions in the different experimental conditions}
In order to compare the experimental conditions (i.e., respectfully, contemptuously, and objectively) with regards to total counts of facial expressions for all participants (i.e., S and R), we applied a Wilcoxon signed-rank test. This allowed us to compare how often a specific expression was visible over the full-time span between the different experimental conditions. For testing, we used a Benjamini-Hochberg p-value correction \cite{benjamini1995controlling} with a false discovery rate of $Q = .05$ and individual p-values of 0.05. Table \ref{tab:signed-rank} summarizes the results of the Wilcoxon signed-rank test. As expected, participants showed significantly more happiness upper and lower in the respectful condition than in the contempt and objective condition. Furthermore, as expected, we found more sadness lower and upper expressions in the contempt compared to the objective condition. This is fully in line with our psychological hypotheses and demonstrates that our instructed attitude manipulations for the sender had an effective influence on both interaction partners (i.e., sender and receiver). Note, found expressions for fear lower and disgust lower occurred against expectations (i.e., those were found significantly more in the respectful compared to the objective condition). This is highly valuable information as it suggests that facial happiness and sadness expressions are the most indicative when comparing negative and positive interaction attitudes. Fear and disgust expressions seem to rather occur for attitudes between the extremes (i.e., respectful and contemptuous).
\begin{table}[!h]
\caption{Results of the Wilcoxon signed-rank (W) test. The table shows only all expressions that were significantly (p-value $<0.05$) different in their occurrence counts between different conditions (Insignificant results are not shown). We permuted all experimental conditions and tested all expressions.}
\label{tab:signed-rank}
\begin{tabular}{|l|c|c|c|c|}
\hline
Expression & First & Second & p-value & W test \\
& condition & condition & & statistics \\
\hline
Happiness upper & respectful & contempt & .0035 & 695 \\
& respectful & objective & .0001 & 511 \\ \hline
Happiness lower & respectful & contempt & .0003 & 580 \\
& respectful & objective & .0017 & 660 \\ \hline
Sadness lower & contempt & objective & .0088 & 744 \\
& respectful & objective & .0005 & 606 \\ \hline
Sadness upper & contempt & objective & .001 & 635 \\ \hline
Fear lower & respectful & objective & .007 & 730 \\ \hline
Disgust lower & respectful & objective & .018 & 783 \\ \hline
\end{tabular}
\end{table}
\begin{table*}
\centering
\caption{Number of pairs for which the Granger causality (GC) test, with p-value =0.05, showed a specific direction of influence in the respectful condition. Average count is the average count of pairs across all expressions. Dominant causal direction (the higher value between Sender GC Receiver (S GC R) and Reciever GC Sender (R GC S)) is shown in bold font.}
\label{tab1}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline \hline
{\multirow{2}{*}{Expression}} & \multicolumn{4}{c|} {Full Time Span } & \multicolumn{4}{c|} {Relevant Interval Selection } \\ \cline{2-9}
& S GC R & R GC S & Bidirectional& No causality & S GC R & R GC S & Bidirectional& No causality\\ \hline
Happiness lower & ${\bf 4}$ & 1 & 22 & 7 & ${\bf 8}$ & 4 & 14 & 8 \\
Happiness upper & 1 & ${\bf 3}$ & 24 & 6 & ${\bf 6}$ & 4 & 18 & 6 \\
Sadness lower & 6 & 6 & 4 & 18 & ${\bf 9}$ & 3 & 8 & 14 \\
Sadness upper & ${\bf 6}$ & 0 & 7 & 19 & $ 7$ & {\bf 8} & 6 & 11 \\
\hline
Average count & ${\bf 4.25}$ & 2.5 & 14.25 & 12.5 & ${\bf 7.5}$ & 4.75 & 11.5 & 9.75 \\
\hline \hline
\end{tabular}
\end{table*}
\begin{table*}
\centering
\caption{Number of pairs for which the Granger causality (GC) test, with p-value =0.05, showed a specific direction of influence in the contempt condition. Average count is the average count of pairs across all expressions. Dominant causal direction (the higher value between Sender GC Receiver (S GC R) and Reciever GC Sender (R GC S)) is shown in bold font.}
\label{tab2}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline \hline
{\multirow{2}{*}{Expression}} & \multicolumn{4}{c|} {Full Time Span } & \multicolumn{4}{c|} {Relevant Interval Selection } \\ \cline{2-9}
& S GC R & R GC S & Bidirectional& No causality & S GC R & R GC S & Bidirectional& No causality\\ \hline
Happiness lower & ${\bf 3}$ & 2 & 17 & 12 & ${\bf 8}$ & 4 & 10 & 12 \\
Happiness upper & ${\bf 4}$ & 1 & 24 & 5 & 3 & ${\bf 7}$ & 17 & 7 \\
Sadness lower & ${\bf 5}$ & 3 & 10 & 16 & 7 & ${\bf 8}$ & 1 & 18 \\
Sadness upper & ${\bf 5}$ & 3 & 2 & 24 & 4 & ${\bf 5}$ & 6 & 19 \\
\hline \hline
Average count & ${\bf 4.25}$ & 2.25 & 13.25 & 14.25 & 5.5 & ${\bf 6}$ & 8.5 & 14 \\
\hline
\end{tabular}
\end{table*}
\begin{table*}
\centering
\caption{Number of pairs for which the Granger causality (GC) test, with p-value =0.05, showed a specific direction of influence in the neutral/objective condition. Average count is the average count of pairs across all expressions. Dominant causal direction (the higher value between Sender GC Receiver (S GC R) and Reciever GC Sender (R GC S)) is shown in bold font.}
\label{tab3}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline \hline
{\multirow{2}{*}{Expression}} & \multicolumn{4}{c|} {Full Time Span } & \multicolumn{4}{c|} {Relevant Interval Selection } \\ \cline{2-9}
& S GC R & R GC S & Bidirectional& No causality & S GC R & R GC S & Bidirectional& No causality\\ \hline
Happiness lower & 0 & ${\bf 5}$ & 23 & 4 & ${\bf 9}$ & 7 & 7 & 9 \\
Happiness upper & ${\bf 5}$ & 0 & 24 & 5 & ${\bf 9}$ & 4 & 21 & 0 \\
Sadness lower & 2 & ${\bf 3}$ & 7 & 22 & ${\bf 6}$ & 7 & 9 & 12 \\
Sadness upper & ${\bf 4}$ & 3 & 3 & 24 & ${\bf 10}$ & 5 & 2 & 17 \\
\hline \hline
Average count & 2.75 & 2.75 & 14.25 & 13.75 &${\bf 8.5}$ & 5.75 & 9.75 & 9.5 \\ \hline
\hline
\end{tabular}
\end{table*}
\subsection{Direction of emotional influence }
In order to study the direction of emotional influence, we compared results of GC test on the relevant interval selection approach versus results of GC test on the full-time span approach. The comparison is represented by the count of pairs for which the Granger causality (GC) test, with p-value =0.05, showed a specific direction of influence, under the three experimental conditions (Tables \ref{tab1} (i.e., respectful), \ref{tab2} (i.e., contemptuouse), and \ref{tab3} (i.e., objective)), for the expressions: happiness upper/lower and sadness upper/lower. These results clearly indicate that the use of the relevant interval selection approach prior to causal inference resulted in considerably more pairs showing unidirectional causation as well as less bidirectional or no causation.
Most interestingly, S influences R particularly in the respectful and objective/neutral compared to the contemptuous condition. In the contemptuous condition the pattern of dominant influence changes (i.e., on average R and S influence each other similarly) when using relevant interval selection approach. Psychologically this indicates that the receiver in this condition is actively trying to repair the overall negative interaction quality, for example by inducing emphatic concern in the sender. These result indicate that using facial expressions only holds the potential to reveal covert attitudes and behaviors that would easily be missed and overlooked when working with verbal behavioral cues. That is most likely in situations like these (i.e., an interaction partner acting ignorant and dismissive), a receiver of such contemptuous information would simply produce less speech content. Finally we can notice that inline with prediction there is significantly higher bidirectional influence for positive emotions than in negative emotions in all experimental conditions. However, while we expected considerable reduction in emotional influence from S to R for negative expression when comparing to positive expression, our results indicate that there is only a slight reduction in the influence from S to R for negative expressions, such as sadness lower and upper when using relevant interval selection. For the full time span, the influence of S on R is even higher for negative emotions than positive ones which is against prediction.
Overall these results indicate that the use of the relevant interval selection approach significantly improved the detection of the emotional influence in all experimental conditions.
\section{Conclusions}
\label{sec:conclusion}
In this paper, we have presented a complete concept for identifying the direction of emotional influence in nonverbal dyadic communication when starting with raw video materials using facial expressions only. To this end, we presented an algorithm for the extraction of emotional facial features, capable of capturing emotional expressions even when strong distinct emotions are not visible. To improve causal inference we proposed an intelligent interval selection approach for filtering relevant information in dyadic dialogues. Subsequently, we were able to apply Granger causality to the set of selected relevant intervals and compute the direction of influence. We applied our approach to real data obtained from a psychological experimental setup. The obtained results revealed that the use of the relevant interval selection approach combined with the proposed facial features significantly improved the detection of the emotional influence for dyadic communication in various instructed interaction conditions. This work also allowed a major step forward in the 2nd person social sciences as we were able to study social emotions in a truly interactive manner. Further, the results of this study indicate that using facial expressions only holds the potential to study implicit attitudes and behaviors in emotionally laden circumstances that would easily be missed when using speech-content information. Overall, we identify our contribution as an important step towards an interdisciplinary research that combines computer vision potentials, psychological observation, and theoretical knowledge of causality methods, to gain novel insights into emotions in real-time social encounters. For further research, we suggest using a learning system capable of classifying upper and lower emotional expressions based on all AUs. Also it would be interesting to tear apart the contributions of various social information channels (i.e., speech, non-verbal speech, facial expressions) towards outcomes of the interaction experience (e.g., enjoyment, engagement, and liking of interacting partner).
\section*{ACKNOWLEDGMENTS}
The authors thank the Carl Zeiss Foundation for the financial support within
the scope of the program line "Breakthroughs: Exploring Intelligent Systems"
for "Digitization — explore the basics, use applications". D. Schneider was supported by the DFG Scientific Network "Understanding Others" - SCHN 1481/2-1.
\bibliographystyle{ieeetr}
| 84d2d559d879723216597fec0d773a52f4028ef9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sect:intro}
Exoplanet transit depths in multiple bands in the light curve of a star provide clues into understanding the atmospheric composition of the exoplanet \citep{Vidal2003,sing2011,Berta2012}.
The transmission spectrum leads to the detection of certain atomic and molecular species, as well as haze and clouds in the planet atmosphere
\citep{seager2010}. For instance, the transit spectrum of HD 209458b and XO-1b provides evidence for water and sodium absorption \citep{Deming2013}, while
\citet{sing2016} found evidence for water vapor in hot Jupiters. Similarly, WASP-12b has been characterized as a prototypical hot Jupiter with a high C/O ratio and significant presence of aerosols in its atmosphere \citep{Madhusudhan2011, sing2011}. However, many of these analyses rely on {\it Hubble} Space Telescope spectroscopy. {\it Hubble}
due to its 90 min orbit, and passage through the South Atlantic Anomaly does not provide a good sampling of the light curve during ingress and egress. Those data are crucial for constraining the inclination of the orbit and thereby the relative transit depths in different bands.
The recently commissioned Transiting Exoplanet Survey Satellite \citep[TESS]{Ricker2015} offers high precision photometric measurements in a broad optical band ($0.6-1$\,$\mu$m) that is redward of the band used by {\it Kepler}. TESS aims to discover transiting exoplanets around the brightest stars in the vicinity of the Sun. It has four cameras with a total field of view 24$\times$96 degrees. The 50$\%$ ensquared-energy (within a square) half-width is 1 detector pixel or 21 arcseconds. The 90$\%$
ensquared-energy half-width is 2x2 pixels. TESS produces multi-frame images at a cadence of 30 minutes with a baseline of at least 27 days
and thus does not suffer from gaps in the light curve like {\it Hubble}. The photometric precision is about 1$\%$ sensitivity at 16th mag.
Here we present a re-analysis of the exoplanets
KELT-19Ab, WASP-156b, and WASP-121b
by combining TESS with other past observations. This paper is organized as follows. Section 2 describes the generation of TESS light curves, derivation of planet parameters from the TESS data using a Monte-Carlo Markov Chain (MCMC) method, transit parameters bias caused by binning and a comparison between our values of inclination and $R_{p}/R_{\ast}$ with previous estimates. Constraints on the exoplanet atmosphere and the predictions of $R_{p}/R_{\ast}$ in the Kepler band and a $z-$band, which is sensitive to water vapor is presented in Section 3.
In Section 4, we present a summary of our results.
\section{Data reduction}
\subsection{TESS Photometry }
The exposure time of TESS for each frame is 2 seconds. Due to
data downlink limitations, two images products are released; the individual frames in a few pixels around certain sources are integrated to 2 minute cadence, and are called Target Pixel Files (TPF). The frames are also integrated to 30 minute cadence for all the sources in the field of view, to generate a Full Frame Image (FFI) \citep{2015Ricker}. We use 2 minute cadence frames to derive the light curves and 30 minute cadence frames to calibrate and remove the blending flux from surrounding sources.
The FFI shows an offset in astrometry relative to the nominal position of the exoplanet host star, as measured by Gaia. However, the astrometric offset is similar across all the frames. So for each source, we checked the astrometry in a few frames and applied a-few-pixel-level offset if it is present. A smaller fractional offset is
determined from a 2-dimensional Gaussian fit of the target. Example images after correcting for astrometric offsets are as shown in Figure 1.
The numbers of frames for KELT-19Ab, WASP-156b, WASP-121b are 17612, 17612, and 15973, respectively. The median offsets for KELT-19Ab, WASP-156b, WASP-121b are 0.10, 0.27, 0.07 in pixels. The standard deviations are 0.04, 0.09, 0.04 in pixels. We apply the same astrometry correction when we analyze the 2 minute cadence image.
The Target Pixel File (TPF) is a 11 pixels $\times$ 11 pixels
cutout of the image around the TESS Input Catalog (TIC). The pipeline versions are spoc-3.3.57-20190215 for KELT-19Ab, spoc-3.3.51-20181221 for WASP-156b, spoc-3.3.57-20190215 for WASP-121b.
\begin{figure}[ht]
\centering
\includegraphics[width=3.5in]{threesources.pdf}
\caption{14$\times$14 pixel TESS images of KELT-19Ab, WASP-156b, WASP-121b, from top to bottom, respectively. The blue markers indicate the position of the exoplanet host star in the Gaia catalog. The red markers indicate the position of nearby sources in the Gaia catalog which contaminate the photometry of the host star i.e. within 6 pixels separation. Each pixel in the image is 21$\arcsec$. The direction is North up, East to the left.
}
\end{figure}
We measured the photometry of the host star in a circular aperture with TPF. The aperture radius was chosen to be 3 pixels, corresponding to 63$\arcsec$. The flux from pixels whose fractional area falls within the aperture is correctly accounted for.
The sky background per image was estimated as the median of the pixels which constitute the lowest fifth percentile in flux and subtracted
from the photometry in the aperture. The standard deviation of these pixels yields the background noise in the photometry. The quadrature sum of this noise and the Poisson noise of the source itself is the photometric uncertainty of the data point.
The contribution to the photometry from nearby unresolved stars was then removed based on the relationship between flux brightness profile and the distance to the Gaia centroid of the unresolved stars (details in the next section).
\subsection{Correction for Blending Sources}
We then corrected the photometry to remove the contamination from surrounding unresolved sources. We used external information from the Gaia database \citep{gaia2016,Gaia2018}, including position, brightness in Gaia G and Rp bands. The Gaia Rp band is similar to the bandpass of TESS. However, it is less sensitive than the Gaia G band. For stars which have Rp band photometry from Gaia, we used its brightness information directly.
For faint stars with no Gaia Rp band detection, we used their G band flux density and applied a median color correction which is the median flux density ratio between the Rp band and G band of stars in the vicinity.
In order to estimate the flux from these contaminating sources inside the aperture centered on the exoplanet host star, we used the TESS FFI data to derive the relationship between flux contamination and the distance between the center of the photometric aperture and a source. We chose 5 bright (brighter than 13 mag in Gaia G-band) and isolated sources from the TESS Input Catalog. It is difficult to guarantee complete isolation to within 6 TESS pixels but the bright stars were chosen to have
no detectable companions in the image and the Gaia catalog showed an absence of sources within 5 mags of the bright stars. We divided the 6 pixel radius circle into a set of annuli, each of 0.25 pixel width. We then chose a set of positions that traverse
the annulus circumference in 0.1 pixel steps. With each of these positions as center, we performed aperture photometry by
measuring the flux density within a circular aperture of radius 3 pixels as for the exoplanet host star. Sky subtraction was performed as described in the previous section.
The maximum separation between the center of the aperture and from the bright star position was allowed to be 6 pixels where the contamination is negligible. This is because
TESS has a 90\% ensquared energy within a half-width of 2 pixels.
Finally, we normalized the set of sky subtracted flux density measurements by the flux measured at zero separation. We repeated this for 5 targets and for each target we used 10 frames. The median and standard deviation of the normalized brightness profile for all these measurements is shown in Figure 2.
The relationship between distance and brightness profile was then estimated by the polynomial fitting of the data pairs (see in Figure 2). We then apply
this relation to every contaminating source whose brightness has been estimated as described above, to derive the flux density contributed by the blending source to the photometry of the exoplanet host star. This contaminating flux density, which is constant for a particular exoplanet host star, was then removed from our photometry values\footnote{
The code to apply this correction is available at \url{https://github.com/sailoryf/TESS_Deblending/
}}. The total contaminating flux percentage and uncertainty for the three stars is
6.22$\%$$\pm$0.42$\%$, 0.64$\%$$\pm$0.03$\%$ and 31.49$\%$$\pm$1.79$\%$ for KELT-19Ab, WASP-156b and WASP-121b respectively.
In addition, there is a possibility of blending from an unresolved binary companion.
In the case of WASP-156b and WASP-121b, there is no evidence of a binary companion
in past work \citep{Delrez,Demangeon}. In the case of KELT-19Ab, a nearby stellar companion was detected with the Palomar/Hale 200$''$ telescope using the near-infrared adaptive optics (AO) system \citep{KELT-19}. The stars have a measured magnitude difference of $\Delta J=2.50\pm0.06$ and $\Delta K_{s}=2.045\pm0.03$.
\citet{KELT-19} report a flux ratio of 0.0270$\pm$0.0034 at a wavelength of 5200$\pm$150$\AA$, effective temperatures of 7500$\pm$200K for the primary component, and 5200$\pm$100K for the secondary component. We search the ATLAS library \citep{Kurucz1993} grid based on the flux ratio and effective temperatures. The best fit stellar templates
are multiplied by the TESS filter response to estimate the brightness of the companion star and its contamination to
the TESS photometry. The derived contaminating flux fraction is 0.059$\pm$0.006 which we remove.
\begin{figure}[h]
\centering
\includegraphics[width=3.8in]{Distance_flux.pdf}
\caption{The normalized brightness profile of stars as seen by TESS used to estimate the amount of contamination of an exoplanet host stars' photometry by background stars seen by Gaia. The scatter points are the observed data. The red line indicates the polynomial fit to the data.}
\end{figure}
\subsection{Detrending of the Light Curve And Comparison with TESS-PDC Light Curve}
Measuring an accurate transit depth of an exoplanet requires a high precision light curve.
This requires detrending to remove the long term trends in the photometry outside of transit events \citep{GPdetrending,AGP,SFF,PDC}. Since we already know the existence of an exoplanet around
these stars, we can use prior knowledge of the transit mid-point time, period and transit duration to perform
a simple detrending.
To do this, we predicted the transit time in the TESS light curve and extract the parts of the light curve spanning 0.6 days centered at the
transit mid-point.
The data points of the light curve during the planetary transit were first masked out. We fitted a linear function to the signal in the masked, cut-off, light curve. The fit is used to estimate and remove the trending in the light curve
within a day. We applied this correction for every single data point of the extracted light curve and fold the light curve around the transit midpoint to
get the transit events shown in Figure 3. We also tried higher order polynomial functions, including orders of 2, 3, 5 and 10. The difference in the detrended light curve is negligible for these choices of polynomials.
The photometric data reduction could be influenced by factors already corrected in our pipeline, or by some factors beyond the consideration of our work, e.g., point spread function (PSF) variation, focus changes, bad pixels, and thermal variations. In order to make sure these factors do not dominate the precision of transit depth measurements, we compare
our results with light curves derived using the TESS Pre-Search Data Conditioning (PDC) modules. PDC light curve is produced by TESS collaboration with consideration of factors above. The comparison covers three targets in this work, as well as HD-219666, an isolated exoplanet host star identified by TESS which is bright compared to its surrounding stars \citep{Esposito2019}.
The deblending correction for HD-219666 is 0.212$\%\pm0.006\%$.
We compare the median flux of our light curve with the G-band flux reported by Gaia \citep{Gaia2018}. The flux density ratio of our decontaminated flux to Gaia flux for KELT-19Ab, WASP-156b, WASP-121b and HD-219666 is 97, 71, 90, and 80, respectively (in arbitrary units). The standard deviation is 9.7. The flux ratio of PDC flux to Gaia flux is 95, 66, 86, 75, respectively. The standard deviation is 11.0. This indicates that the two techniques are perfectly consistent with each other. Furthermore, we find that
the derived planetary transit depths from the two sets of light curves of these sources are within 1 $\sigma$ of each other.
\subsection{Planet Parameters Derived from TESS-Detected Transits}
\subsubsection{Fitting with Free Inclination and Semi-major Axis}
We used an MCMC fitting technique to derive the parameters of the exoplanets from the TESS transit curves \citep{Mandel_Agol2002}. We adopted a
circular orbit as in previous articles that confirmed the corresponding exoplanets with orbital periods derived therein \citep{KELT-19,Demangeon,Delrez}.
The value of T$_{eff}$, log $g$, and [Fe/H] was taken from the
discovery articles as well.
The free parameters in our fit are the radius ratio of the planet to the host star ($R_{p}/R_{\ast}$), the inclination of the planets orbit (i), the semi-major axis in unit of stellar radii ( $a/R_{\ast}$), time offset of transit center (T0) and the limb darkening parameters (a1 for linear limb-darkening coefficient, a2 for quadratic limb-darkening coefficient), as shown in table 1. Except for limb darkening, the other parameters have uniform priors.
We adopted the common inclination prior of $\mathcal{U}$[70,90] but changed it to be $\mathcal{U}$[70,110] when the posterior of inclination is within a degree of 90 degrees. This is because the distribution of inclination performs differently when the real inclination is close to 90. A prior cut-off at 90 degree severely distorts the posterior distribution to be non-Gaussian. A prior with no 90-degree cut-off yields a more reasonable posterior distribution. The difference in inclination posterior between these two scenarios of priors is systematic, but consistent within 1 $\sigma$.
The limb darkening model is the same as that adopted in the exoplanet discovery articles \citep{KELT-19,Demangeon,Evans}.
The coefficients for limb darkening have a Gaussian prior centered on the prediction for TESS data \citep{TESSLD} with an uncertainty of 0.05 \citep{YangLD}.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{KELT-19_A_B_lc.pdf}
\includegraphics[width=3.5in]{WASP-156_lc.pdf}
\includegraphics[width=3.5in]{WASP121_lc.pdf}
\caption{Fits to the TESS TPF transit light curves of KELT-19Ab (top), WASP-156b (middle), WASP-121b (bottom). The blue points are observed 2 minute data decontaminated using Gaia photometry. The green points are bins of 30 minutes cadence data. The red dashed line represents the best-fit transit model to the 2 minute data. The yellow region shows 1 $\sigma$ confidence region of the fits. The midpoint of the transit event is shifted to half the period, at the center of the x-axis. The free parameters in the fits are inclination, semi-major axis and limb darkening parameters as shown in Figure 4.}
\end{figure}
In the case of KELT-19Ab, we adopted T$_{eff}$, log \emph{g}, and [Fe/H] of 7500$\pm$110 K, 4.127$\pm$0.029 cgs, -0.12$\pm0.51$. As in \citet{KELT-19}, a quadratic limb darkening model is applied. The priors for the limb darkening coefficient are 0.27$\pm$0.05, 0.23$\pm$0.05, for the linear (linLimb) and quadratic coefficients (quadLimb), respectively.
In the case of WASP-156b, the T$_{eff}$, log \emph{g}, and [Fe/H] values are 5871$\pm$57 K, 4.35$\pm$0.03 cgs, 0.10$\pm0.10$. A quadratic limb darkening model is used \citep{Demangeon}.
The priors for limb darkening coefficients are 0.36$\pm$0.05 and 0.22$\pm$0.05 for linLimb and quadLimb respectively.
\begin{figure}[!htb]
\centering
\includegraphics[width=3in]{KELT-19_A_B_corner.pdf}
\includegraphics[width=3in]{WASP-156_corner.pdf}
\includegraphics[width=3in]{WASP121_corner.pdf}
\caption{The probability distribution of the fit parameters for KELT-19Ab (top), WASP-156b (middle), WASP-121b (bottom). The grayscale map indicate the density of samples derived from MCMC analysis for different pairs of parameters. The histograms along the diagonal are the marginalized density distribution for individual parameters. The dashed lines denote 68$\%$ of samples. The plot applies corner routine \citep{corner2016}. }
\end{figure}
In the case of WASP-121b, the T$_{eff}$, log \emph{g}, and [Fe/H] values are 6460$\pm$140 K, 4.2$\pm$0.2 cgs, 0.13$\pm0.09$. The limb darkening model used is the quadratic limb darkening model \citep{Delrez,Evans2018}. The limb darkening coefficients that we applied to our MCMC fits are 0.33$\pm$0.05, 0.21$\pm$0.05, respectively.
For each fitting, the parameters' probability distributions are derived using Python MCMC tool, PYMC \citep{pymc}. We applied 200000 iterations for the whole chain, and ignore the first 100000 steps to ensure the chain is stable. The best-fitting light curves are shown in Figure 3, respectively. The distribution of $R_{p}/R_{\ast}$, $a/R_{\ast}$, inclination and limb darkening parameters are shown in Figure 4.
The best fit results are shown in Table 1.
The inclinations derived from TESS TPF data for KELT-19Ab, WASP-156b, and WASP-121b are 3.3, 1.0 and 2.3 degrees larger than the results based on identification paper, respectively. Specifically, in case of WASP-121b, the inclination derived is more consistent with the follow-up work value 89.1 by \citet{Evans2018}.
Due to TESS's continuous sampling with 2 minutes and high precision photometry, the ingress and egress of transit are well monitored. These short phases in the transit duration are extremely sensitive to the inclination and semi-major axis. We, thereforebelieve that the TESS result as well as result from \citet{Evans2018} on inclination for WASP-121b are robust.
\subsubsection{Fitting with Fixed Inclination and Semi-major Axis}
Although our best fits are similar to other work, there are clear signs of differences, particularly in our derived inclination and transit depths.
In order to focus on differences in the radius ratio caused by wavelength dependent properties, we refit the TESS light curves, by fixing the inclination and semi-major axis to the referenced work \citep{KELT-19,Demangeon,Evans}.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{KELT-19_A_B_lc_fixed.pdf}
\includegraphics[width=3.5in]{WASP-156_lc_fixed.pdf}
\includegraphics[width=3.5in]{WASP121_lc_fix.pdf}
\caption{Results from fitting with fixed inclination and semi-major axis for KELT-19Ab (top), WASP-156b (middle), WASP-121b (bottom). The blue points are deconfused photometry. The green points indicate 30 minutes cadence binned result. The red dashed line represents best-fitting model. The black line indicates the result of previous work (referenced in the text). The yellow region shows 1 $\sigma$ confidence region of the fitting. The transit event is shifted at the center of x-axis which is half of the period.}
\end{figure}
The fitting followed the same procedure as section 2.4.1. The fitting results, as well as the derived parameters from previous work, are listed in Table 1. The differences between our fits and the result from the former work
are shown in Figure 5. The reduced chi-square in the fits are slightly larger than the best fit from previous section (Table 1). The small difference in the chi-square values indicate that the value of inclination and semi-major axis preferred in other work, is reasonably consistent with the TESS data but is not the preferred solution.
The fixed parameter MCMC fits have reduced uncertainties due to the strong priors on inclination and semi-major axis. These constraints on inclination and semi-major disturb the Gaussian distribution of other parameters. The revision for the three sigma rules for the parameters is beyond this work \citep[see in reference][]{Friedrich1994,MacKay2002,Hogg2018}. We cited the 1$\sigma$ uncertainty from the MCMC derived standard deviation for each parameter.
Our fits to the TESS data gives us a consistent planet to star radius ratio for WASP-121b and WASP-156b at the $\sim$2$\sigma$ level (Table 1). However, the planet to star radius ratio for KELT-19Ab is 7.5$\sigma$ smaller than previous work \citep{KELT-19}. The TESS radius ratio for WASP-121b is deeper than the median value including all the bands with the TESS wavelength range from \citep{Evans2018}. For WASP-156b, the radius ratio is smaller than past work \citep{Demangeon}.
\subsection{ Transit Parameters Bias Caused by Binning }
Sampling cadence results in morphological distortions to the transit light curve due to sampling of transit ingress and egress \citep{David2010}. For the TESS cadence, an assessment of the parameter bias is necessary for assessing the differences in transit depths for different bands.
We firstly investigate if our derived parameters from
the TESS 2 minute data are consistent with that from the 30 minute data by repeating the whole data reduction and fitting with 30 minute cadence image. In case of free inclination fitting, the inclinations derived for KELT-19Ab, WASP-156b, and WASP-121b are 4.6, 5.1 and 13.6 degrees smaller than the results based on 2 minutes binning data, respectively. Furthermore, the residuals from the light curve for WASP-121b when fixing the inclination shows obvious structure during the phase of ingress and egress (seen in Figure 6). The radius ratio when fixed inclination and semi-major axis is 0.1195$\pm$0.0004, about 7$\sigma$ smaller than 2 minutes result, with substantially worse chi-square values.
The inconsistency, especially for WASP-121b, motivates a simulation for a more detailed assessment of the differences.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{WASP121_lc_fix_30.pdf}
\includegraphics[width=3.5in]{residual_121.pdf}
\caption{Top: Result from fitting with a fixed inclination and semi-major axis for WASP-121b (TESS 30 minute data). The red dashed line represents the best-fitting model. The black line indicates the fit from previous work. Bottom: Residuals after fitting for the exoplanet transit around transit midpoint for WASP-121b based on 30 minutes data. The upper panel is residuals from the best fit. The lower panel is residuals assuming a fixed inclination and semi-major axis adopted
from previous work. The 30 minutes data are clearly
inconsistent with the previously derived inclination value.}
\end{figure}
We simulated 1000 light curves for each source. The input parameters were set to that in the discovery papers \citep[Table 1;][]{KELT-19, Demangeon, Delrez}. We note that for WASP-121b, the newest reported inclination is 1.5 degrees larger than initially assessed \citep{Evans2018}. The result from \citet{Evans2018} is also well matched to the TESS TPF derived result as shown in Table 1. But for the purpose of the simulations, we chose the value from the discovery paper. The time sampling for the simulated light curve is 1 second and is binned to 2 minutes and 30 minutes. The numbers of transit events are set to be the same as the event numbers observed by TESS. We add a Gaussian error to the light curves. The adopted 1$\sigma$ scatter is the standard deviation of the residuals from fitting the TESS light curves scaled by a factor of the square root of time scale.
These simulated light curves are then fit as described earlier.
Figure 7 and 8 show the distribution of derived parameters compared
to the input values.
We find that the magnitude of bias in the derived parameters especially inclination and semi-major axis, depends on the duration time of the transit, especially ingress and egress. We take WASP-121b as an illustrative example which gives the largest inclination difference. The derived median inclination value with the 30 minute cadence data is 76.4$\deg$ with a standard deviation of 0.7, 11.2 degrees smaller than the input value. The inclination decrease is because the binning smooths the rapid change of the flux decrease during ingress and egress. The planet to star radius ratio when fixing the inclination and semi-major axis is 0.1201$\pm$0.00035, also smaller than the input value at the level of $\sim5\sigma$ (seen in Figure 7).
\citet{David2010} provide a numerical solution to compensate for biases caused by binning, using Kepler’s equations.
We investigate the derived correction through a simulation. 1000 simulated light curves of 30 minute cadence
are generated with input parameters the same as described before. Using the batman routine \citep{Kreidberg2015}, we resample and integrate the simulated light curves to 2 min cadence. The derived fit inclination is 86.4$\pm$1.3 deg, indicating that resampling corrects the inclination bias to within 1$\sigma$. However, the derived R$_{p}$/R$_{\ast}$ from the resampled light curve has a median and scatter of 0.1231$\pm$0.00068. The bias in R$_{p}$/R$_{\ast}$ is 0.0011$\pm$0.00068, while the bias without resampling is -0.0019$\pm$0.00035.
Such differences of $\sim 0.8\%$ are negligible for most transit applications but need to be considered carefully for atmospheric characterization.
We use further simulations to assess if any R$_{p}$/R$_{\ast}$ bias after oversampling still exists. The resampling cadence is changed from 1 minute to 10 minutes with a step of 1 minute. The resampling cadence below 1 minute is also tested from 10 seconds to 50 seconds with a step of 10 seconds.
The R$_{p}$/R$_{\ast}$ reveals an offset at the level of 0.001 with an uncertainty $\sim$ 0.0006 when the resampling cadence is below 5 minutes. The offset is larger and increases with resampling cadence when the resampling cadence is larger than 5 minutes.
In addition, we apply simulations to check if the bias still exists when the initial sampling rate is 2 minutes. The resampling cadence is set to be 30 seconds, 1 minute, and 2 minutes. The derived R$_{p}$/R$_{\ast}$ does not reveal any evidence of offset suggesting that the batman routine itself does not induce any systematic bias.
We thus conclude that the resampling technique \citep{David2010, Kreidberg2015} significantly reduces the systematic bias caused by undersampling. However, we conclude that the very high precision (R$_{p}$/R$_{\ast}$ $\sim$ 0.5$\%$) needed for atmospheric analysis needs careful correction for sampling-induced systematics.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{30m_free_inl.pdf}
\includegraphics[width=3.5in]{30m_fix_rp.pdf}
\caption{Top: The derived inclination distribution from fitting 1000 simulated light curves sampled at 30 minutes, for the parameters of WASP-121b. Bottom: The planet to star radius ratio distribution when fixing the inclination and semi-major axis. The vertical lines for both panels give the input values to the simulations. Clearly, for WASP-121b, a light curve cadence of 30 minutes would result in a large bias in derived parameters.}
\end{figure}
In the case of light curves with 2 minute sampling, the bias is assessed in the same way. The best fit inclination for WASP-121b is found to be 87.7$\pm$1.3$\deg$. The planet to star radius ratio with a fixed inclination and semi-major axis is 0.1219$\pm$0.0004. We find the difference between the derived values and the input values to be negligible with the 2 minute data.
For all three sources, we find that
the simulations with 2 minute sampling yield similar parameters as the input value. Meanwhile, with 30 minutes cadence, the result is significantly biased when we do not correct for the undersampling \citep{David2010,Kreidberg2015}. Also, the differences in the inclination and planet-to-star radius ratio between 2 and 30 minutes cadence matches the difference between them from real TESS data. From these simulation, we therefore conclude that the biases in parameters from fitting TESS data of different cadence comes from binning, and that the fit parameters from the 2 minute cadence data for our three targets are robust.
Finally, we undertake a simulation to investigate the dependence of inclination and semi-major axis upon sampling time interval when not applying the oversampling technique (Figure 8). The sampling time interval in the simulation ranges from 1 minute to 50 minutes. For each interval, the light curve is generated as described above and fit. We duplicate 10 simulations for each sampling interval. The median fit values and standard deviations of the derived parameters are shown in Figure 8 for each sampling interval. It is clear that the inclination and semi-major axis are not significantly biased when the sampling time is less than 5 minutes. Also, the result indicates that the inclination and semi-major axis derived are distorted to be smaller when applying a larger sampling interval (as shown in Figure 8). This suggests that great care must be taken when comparing transit depths in different bands, especially when the light curve sampling intervals are likely to be quite different. The simulations were conducted without the oversampling correction \citep{David2010} which can mitigate the bias of inclination and semi-major axis as previously discussed.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{SemiAInc.pdf}
\caption{Results from fitting inclination (green squares) and semi-major axis (red diamonds) to simulated light curves of different sampling intervals when not correcting for the undersampling. The vertical red dotted line gives the sampling interval limit below which the parameters derived are not biased and corresponds to 5 minutes. The horizontal black line dotted is the input value of the semi-major axis; the dotted blue line is the input value of inclination. Clearly, TESS FFI 30 minute cadence data would result in significantly biased parameters for planets with short duration transits such as those considered here.}
\end{figure}
\begin{table*}
\setlength{\tabcolsep}{3mm}
\begin{center}
\caption{Planet Parameters}
\label{tab1}
\begin{tabular}{ccccc}
\hline
\hline
\small
{Parameters }& {Description }& KELT-19Ab & WASP-156b & WASP-121b \\
& & 68$\%$ Confidence & 68$\%$ Confidence & 68$\%$ Confidence \\
\hline
\multicolumn{5}{c} {Best Fitting Parameters} \\
\hline
R$_{p}$/R$_{\ast}$ & Planet/star radius ratio in TESS & 0.09576$\pm$ 0.00082 & 0.0670 $\pm$ 0.0012 & 0.1234$\pm$0.0004\\
i [deg] & Inclination & 88.7$\pm0.7$ & 90.1$\pm1.4$ & 89.9$\pm1.6$ \\
a/R$_{\ast}$ & Semi-major axis of planet orbit in unit of stellar radii & 9.04$\pm0.19$ & 12.98$\pm+0.73$ & 3.81$\pm0.03$ \\
std of residual & Parts per million & 1253 & 2130 & 1763 \\
reduced $ \chi^{2}$ & Reduced chi-square & 1.001 & 1.0009 & 1.0004 \\
\hline
\multicolumn{5}{c} {Fitting Results with Fixed Inclination and Semi-major Axis} \\
\hline
R$_{p}$/R$_{\ast}$ & Planet/star radius ratio in TESS& 0.09940$\pm0.00048$ & 0.0662$\pm0.0009$ & 0.1230$\pm0.00025$\\
& Planet/star radius ratio in other work &
0.10713$\pm$0.00092$^{a}$ & 0.0685$^{+0.0012}_{-0.0008}$$^{e}$ & 0.1220$^{+0.0005}_{-0.0005}$ $^{b,c,d}$ \\
difference significance & Difference between TESS and former work in $\sigma$ & 7.549 & 1.533 & 1.789 \\
i [deg] & Inclination & 85.41$^{+0.34}_{-0.31}$$^{a}$ & 89.1$^{+0.6}_{-0.9}$$^{d}$ & 87.6$\pm$0.6$^{b,c}$ \\
&&&&89.1$\pm$0.5$^{d}$ \\
a/R$_{\ast}$ & Semi-major axis of planet orbit in unit of stellar radii & 7.50$^{+0.20}_{-0.18}$$^{a}$ & 12.8$^{+0.3}_{-0.7}$$^{e}$ & 3.754$^{+0.023}_{-0.028}$$^{b,c}$ \\
&&&&3.86$\pm$0.02$^{d}$\\
std of residual & Parts per million & 1266 & 2132 & 1765 \\
reduced $\chi^{2}$ & Reduced chi-square & 1.021 & 1.001 & 1.002 \\
linLimb & Linear limb-darkening coefficient & 0.29$\pm$0.04 & 0.37 $\pm$0.04 & 0.24 $\pm$0.03 \\
quadLimb & Quadratic limb-darkening coefficient & 0.23$\pm$0.04 & 0.23 $\pm$0.04 & 0.21 $\pm$0.04 \\
\hline
\end{tabular}
\end{center}
\begin{flushleft}
Reference:
(a) \citet{KELT-19};
(b) \citet{Delrez};
(c) \citet{Evans};
(d) \citet{Evans2018};
(e) \citet{Demangeon};
\end{flushleft}
\end{table*}
\subsection{A difference in stellar activity level?}
As explained in \citet{pont2008, Agol2010}, differences in the number of starspots between past observations and the TESS observations could account for differences in the measured transit depth. It was estimated that when the flux density of the star changes by 1\% because of stellar spots, the influence on the radius ratio is of the order of 10$^{-3}$R$_{p}$/R$_{\ast}$, comparable to the uncertainty of our TESS derived value. None of the three sources studied in this paper \citep{KELT-19,Demangeon,Evans} have shown
any past evidence of stellar activity. Furthermore, if starspots were significant, the standard deviation of the light
curve outside the transit should be different from that during transit. So we conclude that the impact
of starspots on our transit depth measurement is negligible.
\section{Atmospheric Constraints}
Our primary scientific motivation for undertaking this work was to compare $R_{p}/R_{\ast}$ in the TESS band with the value from previous work to constrain the atmospheres of the exoplanets. In particular, the TESS bandpass is significantly redder than the {\it Kepler} band and encompasses the
0.95\,$\mu$m water feature. The presence of these features in the planets'
atmosphere would result in a larger value of $R_{p}/R_{\ast}$ in the TESS band. For instance, the $R_{p}/R_{\ast}$ of KELT-19Ab in TESS data is significantly smaller, at the level of 7.5$\sigma$, indicating the possible existence of atmospheric features.
Our fixed inclination fits (Table 1) allow for a more direct comparison with past work.
The significance level of differences in $R_{p}/R_{\ast}$ are 7.5, 1.5, 1.8 $\sigma$ for KELT-19Ab, WASP-156b and WASP-121b, respectively. We discuss the implications of this difference for each of the exoplanets.
\subsection{KELT-19Ab}
KELT-19Ab is a giant planet transiting a moderately bright (V $\sim$ 9.9\,mag) A8V star with a orbital period of 4.61 days \citep{KELT-19}. The host stars has $M_{\ast}=1.62^{+0.25}_{-0.20}M_{\odot}$, $R_{\ast}=1.83\pm0.10R_{\odot}$. The planet has a radius of $R_{P}=1.91\pm0.11R_{J}$.
The ground-based light curves from \citet{KELT-19} in each band ($B$, $g$, $r$, $i(7519\AA)$, $I(8317\AA)$ and $z(8992\AA)$) were obtained. They are re-fit, to compare the variation in $R_{p}/R_{\ast}$ values between the different bands to that presented for the combined band. The light curves are performed with binary component deblending and linear detrending, using the same method as described previously. The flux contamination fraction adopted is the same as \citet{KELT-19}. The ground-based light curve does not have a long and stable baseline as TESS data. The long-term trend is therefore very difficult to estimate and dominates the $R_{p}/R_{\ast}$ uncertainty.
We masked the transit phase in the light curve as described previously, and fit the baseline, using different baseline lengths. This is repeated 5 times with the baseline length reduced by a factor of 10$\%$ each time. Three detrended light curves with the best residual are used for transit fitting. The $R_{p}/R_{\ast}$ and uncertainty are taken as the median value and the standard deviation of three fitted values. We do not fit the $B$ and $r$-band data because of the short baseline and partial coverage of the transit. The $i-$band values from the different telescopes are combined. KELT-19Ab has $R_{p}/R_{\ast}$ of 0.10713 derived from a joint fit to ground-based (B, g, r, i, I, z) data. The refit of our result presents smaller values of $R_{p}/R_{\ast}$ in each band compared to what has previously been published
\citep[0.10713,][]{KELT-19}. This indicates the $R_{p}/R_{\ast}$ differences might partially come from a different process choice.
We fitted the multi-band transit data with the grid of atmospheric transmission spectra from \citet{Goyal2019, Goyal2018} which are developed for hot, H$_{2}$/He dominated Jupiters.
This model is a 1D radiative-convective-equilibrium solution based on the ATMO code \citep{Tremblin2015,Tremblin2016, Drummond2016}. ATMO solves the radiative transfer for isothermal pressure-temperature (P-T) profiles, with opacity and chemical abundance being other parameters. The grid model has a parameter space of two chemical scenarios, 22 temperatures, 4 planetary gravities, 5 atmospheric metallicities, 4 C/O ratios, 4 scattering haze, 4 cloud parameters and a scaling factor to the specific planetary radius. Modifying specific chemical abundance will lead to different chemical equilibrium which is avoided in the use of the forward grid model.
In order to fit the model, we minimized the chi-square value between the observations and model predictions at all the bands. The model prediction at each band is derived by the following equation where $S_{\lambda}$ is the
filter response.
\begin{equation}
\frac{R_{p}}{R_{\ast}}=\frac{\int \frac{R_{p}}{R_{\ast}}_{\lambda}S_{\lambda}d{\lambda} }{\int S_{\lambda}d{\lambda}}
\end{equation}
We find that the haze dominated model at 1500K has the least $\chi$$^{2}$ of 2.80 (Figure \ref{plot:keltindi}). A clear model showing weak water vapor with rainout condensation at 1500K gives the second best $\chi^{2}$ of 3.05, while a clear model showing strong water vapor at 1000K has the third best $\chi^{2}$ of 3.26. A flat model with opaque featureless is also shown with a $\chi^{2}=3.14$. Other models resulted in much larger chi-square values (more than 6) and are not consistent with the data. We also obtain the Bayesian Information Criterion (BIC) for evaluating the model evidence. BIC theory is effective when the data points are much more than the free parameters \citep{BIC}. This might not be the case for KELT-19Ab but we still present the BIC here for reference. The BIC is 4.41, 9.49, 9.69 for haze, weak water, and strong water models. The BIC is 4.74 for an opaque featureless model.
The free parameter numbers of the opaque featureless model, haze, weak water, and strong water models are 1, 1, 4, and 4 respectively. The smaller $\chi^{2}$ and BIC of the haze model suggest that it is possibly the best interpretation given the current data. However, given the relatively narrow range in $\chi^{2}$ values and
small difference in BIC,
the opaque featureless model would also be a reasonable explanation.
\begin{figure*}[!htb]
\centering
\includegraphics[width=7in]{KELT19_transmission_spectrum_individualband_detrendingerror.pdf}
\caption{$R_{p}/R_{\ast}$ of KELT-19Ab in different bandpasses based on our re-analysis of those light curves. The red point is $R_{p}/R_{\ast}$ derived from TESS. The black points are $R_{p}/R_{\ast}$ from former work. In the top panel, the lowest $\chi$$^{2}$ model (haze dominant) is shown as a blue solid line. The green line indicates a flat, opaque featureless model that is consistent with the data. Blue diamonds are best-fit model predictions at the bands. The yellow diamond is the best fit model prediction at the SDSS $z$ band, green diamond is the best-fit model prediction at the Kepler band. In the bottom panel, a clear model with strong water (green) is shown with a faint water model (orange) as a comparison. Blue diamonds are strong water model predictions in different wavelengths bands while the green diamond is for the {\it Kepler} band. Dark green and brown diamonds are the strong water and faint water model prediction in the SDSS $z-$band. Clearly, 500ppm photometry in the $z-$band can distinguish between these scenarios.}
\label{plot:keltindi}
\end{figure*}
Based on our fits, we predicted the possible $R_{p}/R_{\ast}$ value in the {\it Kepler} band and $z-$ band based on these templates (Figure \ref{plot:keltindi}). The {\it Kepler} band would show $R_{p}/R_{\ast}$ of 0.1003 for a haze dominated atmosphere and 0.0997 for a clear model. In the $z-$band,
the presence of water would introduce a significant difference in the ratio. We estimate that a clear atmosphere model with strong water contribution would show $R_{p}/R_{\ast}$=0.1002. The value is smaller, 0.0990, from a clear model with faint water, as well as from a haze dominant model.
We find that a difference of 1.2$\%$ in $R_{p}/R_{\ast}$ arises between strong water contribution and faint water contribution models.
If we scale from our TESS measurements, and assume that the signal to noise
ratio depends on the square root of the number of points in the transit, we find that with 500ppm precision for 200 data points around the transit event in the $z-$band, one can distinguish between these two scenarios. This is of course a minimum since additional systematic effects can affect the measurements.
Alternately, instead of using our derived values of $R_{p}/R_{\ast}$ in each of the optical bands, we can use the value of \citet{KELT-19} which is derived from a joint fit to all the bands.
We treated their joint value of $R_{p}/R_{\ast}$ as the value in each band. The uncertainty follows the equation:
$\sigma_{joint}=\sigma_{i}/\sqrt{N}$, according to the error propagation law. $\sigma_{joint}$ is the joint band uncertainty, N is the number of bandpasses, $\sigma_{i}$ is the uncertainty in the band $i$. We assume the uncertainty in each band is the same.
Fitting $R_{p}/R_{\ast}$, the $\chi^{2}$ values for the different atmospheric models discussed - haze, strong water, weak water, and opaque featureless model - are 47.52, 50.68, 53.24, and 55.07, respectively. The BIC is 49.47, 58.46, 61.02, and 57.02 for the same models, respectively. Other models resulted in much larger chi-square values (more than 60) and are also not consistent with the data. Again, we find that a haze dominant atmosphere is preferred, as before.
\subsection{WASP-156b}
The discovery of WASP-156b was reported in \citet{Demangeon}. It is a super-Neptune orbiting a K type star (V=11.6\,mag, $M_{\ast}=0.842\pm0.052M_{\odot}$, $R_{\ast}=0.76\pm0.03R_{\odot}$) with a period of 3.84 days. It has a mass of $0.128^{+0.010}_{-0.009}M_{J}$, a radius of $0.51^{+0.02}_{-0.02}R_{J}$.
We followed the same procedure as for KELT-19Ab, to analyze the differences in $R_{p}/R_{\ast}$ in different bands. The published value of $R_{p}/R_{\ast}$ for comparison with our TESS analysis, is taken from \citet{Demangeon} at Johnson R, Gunn r, Johnson I and Gunn z (Figure \ref{plot:156}).
We find that a clear atmosphere at 1500K with strong water contribution (best fit model) has a $\chi^{2}$ of 2.35, and a BIC of 8.80. The faint water model at the same temperature follows with a worse $\chi^{2}$ of $\sim$ 2.44, and a BIC of 8.87. The haze model has a $\chi^{2}$ of 2.44 and a BIC of 4.04. An opaque featureless model has the same $\chi^{2}$ of 2.35 as the best fit model. The smallest BIC value of 3.96 indicates that the data might prefer the opaque featureless atmosphere model. The strong water contribution causes a shift of 0.9$\%$ in $z-$band prediction of $R_{p}/R_{\ast}$ compared to the faint water model. In the Kepler band, the haze dominant model has a larger $R_{p}/R_{\ast}$ of 1.3$\%$ than the no atmosphere model. We conclude that the current measurements are consistent with an opaque featureless model.
\begin{figure*}[!htb]
\centering
\includegraphics[width=7in]{WASP156_transmission_spectrum.pdf}
\caption{$R_{p}/R_{\ast}$ of WASP-156b in different bandpasses. The top panel shows the clear model with strong water (blue solid line), the haze dominant model (green solid line) and the opaque featureless model (brown solid line).
The red point is $R_{p}/R_{\ast}$ derived from TESS. The black points are $R_{p}/R_{\ast}$ from former work with the blue diamonds the "clear model" predictions at those bands. The yellow diamond is the model prediction at the SDSS $z-$band, green diamond is the model prediction at Kepler band. In the bottom panel, a clear model with strong water (green solid line) is shown along with a faint water model (Orange solid line) for comparison.
The empty green diamond is a strong water model prediction at Kepler band. The dark green diamond is the strong water model prediction at the SDSS $z-$band. Brown diamond is faint water model prediction at $z-$band.}
\label{plot:156}
\end{figure*}
\subsection{WASP-121b}
WASP-121b was discovered by \citet{Delrez}. It is a hot Jupiter around a main sequence star (V=10.4\,mag, $1.353^{+0.080}_{-0.079}M_{\odot}$, $1.458\pm0.030R_{\odot}$) with a period of 1.28 days. The planets has a mass of $1.183^{+0.064}_{-0.062}M_{J}$ and a radius of $1.865\pm0.044R_{J}$.
We used external observations from \citet{Delrez, Evans2016, Evans2018}
in conjunction with our analysis of the TESS data (Figure \ref{plot:121}).
These measurements utilized the standard ground-based filter, STIS, and WFC3 instruments on {\it Hubble} and span 0.3 to 1.65$\mu$m.
As outlined in \citet{Evans2018}, a global fit for the whole spectrum is challenging due to the complex problem of non-equilibrium chemistry.
Our fits with the inclination fixed to past work however, yield a value of $R_{p}/R_{\ast}$ which is in very good agreement with \citet{Evans2018}. The median value of the radius ratio is 0.1225 at the wavelength arrange of TESS.
To fit the spectrum, we again use the general grid model from \citet{Goyal2019}.
\begin{figure*}[!htb]
\centering
\includegraphics[width=7in]{WASP-121_transimission_spectrum.pdf}
\caption{The transmission spectrum for WASP-121b. The solid yellow point is the best fit value of $R_{p}/R_{\ast}$ from the TESS light curve. However, if the inclination is fixed to previous work for a consistent comparison, the solid redpoint is $R_{p}/R_{\ast}$ derived from TESS data. The black points are $R_{p}/R_{\ast}$ from former measurements.
The grey solid line is the best fit from \citet{Evans2018}. The blue line shows the haze model. The brown line shows the opaque featureless model.
The black line indicates a clear atmosphere at a cool temperature (shifted for clarity).
\label{plot:121}
}
\end{figure*}
The best-fit model from \citet{Evans2018} is still the best fit model when we add the TESS constraints. The fitting has a reduced chi-square of 4.91, a BIC of 449.74, with 92 data points, 4 fixed, and 4 free parameters (in Figure \ref{plot:121}). The fixed parameters are local condensation, atmospheric metallicities at 20x, haze parameters of 1.0, and no cloud. The fitted free parameters are temperature of 1500K, gravity at 10ms$^{-2}$, C/O ratio of 0.7, and planet radius (without atmosphere).
An opaque featureless model has a reduced chi-square of 5.19, a BIC of 477.39. The haze dominant model is with a chi-square of 5.87, a BIC of 538.76. A transmission model at 1500K with features of potassium, sodium, and water vapor has a much larger reduced chi-square of 27.40, a BIC of 2429.51, and seems to be ruled out. According to BIC theory, any model having a BIC smaller than 10 is favored \citep{BICmostcited}.
\section{SUMMARY AND DISCUSSION}
We have analyzed the TESS data for KELT-19Ab, WASP-156b, WASP-121b to
measure $R_{p}/R_{\ast}$ in a broadband spanning 600nm to 1000nm. We identify the significant role of contamination by unresolved background stars in the TESS data and
use Gaia data to subtract their contribution to the transit light curves.
We provide general-purpose software to address this issue for all TESS data.
We detrend the light curves through a linear fit to the photometry near the transit events. To derive the exoplanet transit parameters, we performed two ways of MCMC model fitting: with and without inclination and semi-major axis fixed. The former in particular helps with the comparison with past work on these stars. Using simulations, we demonstrate the systematic biases in inclination and radius ratio with 30 minutes cadence data when no correction for undersampling is undertaken \citep{David2010}.
We also find the inclination and semi-major axis are distorted to be smaller with larger sampling times. However,
we find that 2 minute sampling does not cause any bias for transit parameters. The atmosphere of these three exoplanets was then constrained based on $R_{p}/R_{\ast}$ in different bands.
In the case of KELT-19Ab, we find that the TESS-measured $R_{p}/R_{\ast}$ is $\sim$7.5$\sigma$ smaller than past work. The $R_{p}/R_{\ast}$ could be due to the choice of analysis or due to the presence of atmospheric features which further observations in narrower bands can reveal. The model fitting possibly favors a haze dominant atmosphere. In the case of WASP-156b, we find that the TESS transit depths agree with previous work and prefer a no atmosphere model fit to the measurements.
In the case of WASP-121b, the fits to the TESS light curve prefer the newest reported inclination. We find that the atmospheric transmission spectrum is best fitted with a 1500K model with water vapor as reported before.
Our work shows that exoplanet atmospheric model constraints benefit from high precision continuous photometry which provides strong constraints on key parameters such as inclination.
Adding one high precision observation at certain bands improves the constraints on atmospheric composition. Since TESS
has now observed the Northern Hemisphere,
we now have Kepler \citep{Kepler2010} and TESS constraints on a few bright stars, which will help assess the role of haze in exoplanet atmospheres. Finally, we show that precision photometry with $\sim$500ppm to 1000ppm in the $z-$band can help directly constrain the presence of water vapor in the exoplanet atmosphere, when combined with precise TESS photometry.
\acknowledgments
We are very grateful to the thoughtful comments by a referee which greatly improved the quality of results presented in this manuscript.
This work made use of PyAstronomy\footnote{https://github.com/sczesla/PyAstronomy} \citep{pya} and the NASA Exoplanet Archive \citep{ExoplanetArchive}. We would like to thank Jayesh M. Goyal for useful inputs and discussion and Caltech Optical Observatories for supporting the visit by the lead author. We also thank Hilke Schlichting for useful feedback on the manuscript. We sincerely thank Karen A. Collins for providing KELT-19Ab ground-based light curves. Fan Yang and Ji-Feng Liu acknowledge funding from National Natural Science Foundation of China (NSFC.11988101), National Key Research and Development Program of China (No.2016YFA0400800) and National Science Fund for Distinguished Young Scholars (No.11425313).
\bibliographystyle{aasjournal}
\section{}
\textit{Research Notes of the \href{https://aas.org}{American Astronomical Society}}
(\href{http://rnaas.aas.org}{RNAAS}) is a publication in the AAS portfolio
(alongside ApJ, AJ, ApJ Supplements, and ApJ Letters) through which authors can
promptly and briefly share materials of interest with the astronomical community
in a form that will be searchable via ADS and permanently archived.
The astronomical community has long faced a challenge in disseminating
information that may not meet the criteria for a traditional journal article.
There have generally been few options available for sharing works in progress,
comments and clarifications, null results, and timely reports of observations
(such as the spectrum of a supernova), as well as results that wouldn’t
traditionally merit a full paper (such as the discovery of a single exoplanet
or contributions to the monitoring of variable sources).
Launched in 2017, RNAAS was developed as a supported and long-term
communication channel for results such as these that would otherwise be
difficult to broadly disseminate to the professional community and persistently
archive for future reference.
Submissions to RNAAS should be brief communications - 1,000 words or fewer
\footnote{An easy way to count the number of words in a Research Note is to use
the \texttt{texcount} utility installed with most \latex\ installations. The
call \texttt{texcount -incbib -v3 rnaas.tex}) gives 57 words in the front
matter and 493 words in the text/references/captions of this template. Another
option is by copying the words into MS/Word, and using ``Word Count'' under the
Tool tab.}, and no more than a single figure (e.g. Figure \ref{fig:1}) or table
(but not both) - and should be written in a style similar to that of a
traditional journal article, including references, where appropriate, but not
including an abstract.
Unlike the other journals in the AAS portfolio, RNAAS publications are not
peer reviewed; they are, however, reviewed by an editor for appropriateness
and format before publication. If accepted, RNAAS submissions are typically
published within 72 hours of manuscript receipt. Each RNAAS article is
issued a DOI and indexed by ADS \citep{2000A&AS..143...41K} to create a
long-term, citable record of work.
Articles can be submitted in \latex\ (preferably with the new "RNAAS"
style option in AASTeX v6.2), MS/Word, or via the direct submission in the
\href{http://www.authorea.com}{Authorea} or
\href{http://www.overleaf.com}{Overleaf} online collaborative editors.
Authors are expected to follow the AAS's ethics \citep{2006ApJ...652..847K},
including guidance on plagiarism \citep{2012AAS...21920404V}.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.85,angle=0]{aas.pdf}
\caption{Top page of the AAS Journals' website, \url{http://journals.aas.org},
on October 15, 2017. Each RNAAS manuscript is only allowed one figure or
table (but not both). Including the
\href{http://journals.aas.org//authors/data.html\#DbF}{data behind the figure}
in a Note is encouraged, and the data will be provided as a link in the
published Note.\label{fig:1}}
\end{center}
\end{figure}
\acknowledgments
Acknowledge people, facilities, and software here but remember that this counts
against your 1000 word limit.
\section{Introduction}
\label{sect:intro}
Exoplanet transit depths in multiple bands in the light curve of a star provide clues into understanding the atmospheric composition of the exoplanet \citep{Vidal2003,sing2011,Berta2012}.
The transmission spectrum leads to the detection of certain atomic and molecular species, as well as haze and clouds in the planet atmosphere
\citep{seager2010}. For instance, the transit spectrum of HD 209458b and XO-1b provides evidence for water and sodium absorption \citep{Deming2013}, while
\citet{sing2016} found evidence for water vapor in hot Jupiters. Similarly, WASP-12b has been characterized as a prototypical hot Jupiter with a high C/O ratio and significant presence of aerosols in its atmosphere \citep{Madhusudhan2011, sing2011}. However, many of these analyses rely on {\it Hubble} Space Telescope spectroscopy. {\it Hubble}
due to its 90 min orbit, and passage through the South Atlantic Anomaly does not provide a good sampling of the light curve during ingress and egress. Those data are crucial for constraining the inclination of the orbit and thereby the relative transit depths in different bands.
The recently commissioned Transiting Exoplanet Survey Satellite \citep[TESS]{Ricker2015} offers high precision photometric measurements in a broad optical band ($0.6-1$\,$\mu$m) that is redward of the band used by {\it Kepler}. TESS aims to discover transiting exoplanets around the brightest stars in the vicinity of the Sun. It has four cameras with a total field of view 24$\times$96 degrees. The 50$\%$ ensquared-energy (within a square) half-width is 1 detector pixel or 21 arcseconds. The 90$\%$
ensquared-energy half-width is 2x2 pixels. TESS produces multi-frame images at a cadence of 30 minutes with a baseline of at least 27 days
and thus does not suffer from gaps in the light curve like {\it Hubble}. The photometric precision is about 1$\%$ sensitivity at 16th mag.
Here we present a re-analysis of the exoplanets
KELT-19Ab, WASP-156b, and WASP-121b
by combining TESS with other past observations. This paper is organized as follows. Section 2 describes the generation of TESS light curves, derivation of planet parameters from the TESS data using a Monte-Carlo Markov Chain (MCMC) method, transit parameters bias caused by binning and a comparison between our values of inclination and $R_{p}/R_{\ast}$ with previous estimates. Constraints on the exoplanet atmosphere and the predictions of $R_{p}/R_{\ast}$ in the Kepler band and a $z-$band, which is sensitive to water vapor is presented in Section 3.
In Section 4, we present a summary of our results.
\section{Data reduction}
\subsection{TESS Photometry }
The exposure time of TESS for each frame is 2 seconds. Due to
data downlink limitations, two images products are released; the individual frames in a few pixels around certain sources are integrated to 2 minute cadence, and are called Target Pixel Files (TPF). The frames are also integrated to 30 minute cadence for all the sources in the field of view, to generate a Full Frame Image (FFI) \citep{2015Ricker}. We use 2 minute cadence frames to derive the light curves and 30 minute cadence frames to calibrate and remove the blending flux from surrounding sources.
The FFI shows an offset in astrometry relative to the nominal position of the exoplanet host star, as measured by Gaia. However, the astrometric offset is similar across all the frames. So for each source, we checked the astrometry in a few frames and applied a-few-pixel-level offset if it is present. A smaller fractional offset is
determined from a 2-dimensional Gaussian fit of the target. Example images after correcting for astrometric offsets are as shown in Figure 1.
The numbers of frames for KELT-19Ab, WASP-156b, WASP-121b are 17612, 17612, and 15973, respectively. The median offsets for KELT-19Ab, WASP-156b, WASP-121b are 0.10, 0.27, 0.07 in pixels. The standard deviations are 0.04, 0.09, 0.04 in pixels. We apply the same astrometry correction when we analyze the 2 minute cadence image.
The Target Pixel File (TPF) is a 11 pixels $\times$ 11 pixels
cutout of the image around the TESS Input Catalog (TIC). The pipeline versions are spoc-3.3.57-20190215 for KELT-19Ab, spoc-3.3.51-20181221 for WASP-156b, spoc-3.3.57-20190215 for WASP-121b.
\begin{figure}[ht]
\centering
\includegraphics[width=3.5in]{threesources.pdf}
\caption{14$\times$14 pixel TESS images of KELT-19Ab, WASP-156b, WASP-121b, from top to bottom, respectively. The blue markers indicate the position of the exoplanet host star in the Gaia catalog. The red markers indicate the position of nearby sources in the Gaia catalog which contaminate the photometry of the host star i.e. within 6 pixels separation. Each pixel in the image is 21$\arcsec$. The direction is North up, East to the left.
}
\end{figure}
We measured the photometry of the host star in a circular aperture with TPF. The aperture radius was chosen to be 3 pixels, corresponding to 63$\arcsec$. The flux from pixels whose fractional area falls within the aperture is correctly accounted for.
The sky background per image was estimated as the median of the pixels which constitute the lowest fifth percentile in flux and subtracted
from the photometry in the aperture. The standard deviation of these pixels yields the background noise in the photometry. The quadrature sum of this noise and the Poisson noise of the source itself is the photometric uncertainty of the data point.
The contribution to the photometry from nearby unresolved stars was then removed based on the relationship between flux brightness profile and the distance to the Gaia centroid of the unresolved stars (details in the next section).
\subsection{Correction for Blending Sources}
We then corrected the photometry to remove the contamination from surrounding unresolved sources. We used external information from the Gaia database \citep{gaia2016,Gaia2018}, including position, brightness in Gaia G and Rp bands. The Gaia Rp band is similar to the bandpass of TESS. However, it is less sensitive than the Gaia G band. For stars which have Rp band photometry from Gaia, we used its brightness information directly.
For faint stars with no Gaia Rp band detection, we used their G band flux density and applied a median color correction which is the median flux density ratio between the Rp band and G band of stars in the vicinity.
In order to estimate the flux from these contaminating sources inside the aperture centered on the exoplanet host star, we used the TESS FFI data to derive the relationship between flux contamination and the distance between the center of the photometric aperture and a source. We chose 5 bright (brighter than 13 mag in Gaia G-band) and isolated sources from the TESS Input Catalog. It is difficult to guarantee complete isolation to within 6 TESS pixels but the bright stars were chosen to have
no detectable companions in the image and the Gaia catalog showed an absence of sources within 5 mags of the bright stars. We divided the 6 pixel radius circle into a set of annuli, each of 0.25 pixel width. We then chose a set of positions that traverse
the annulus circumference in 0.1 pixel steps. With each of these positions as center, we performed aperture photometry by
measuring the flux density within a circular aperture of radius 3 pixels as for the exoplanet host star. Sky subtraction was performed as described in the previous section.
The maximum separation between the center of the aperture and from the bright star position was allowed to be 6 pixels where the contamination is negligible. This is because
TESS has a 90\% ensquared energy within a half-width of 2 pixels.
Finally, we normalized the set of sky subtracted flux density measurements by the flux measured at zero separation. We repeated this for 5 targets and for each target we used 10 frames. The median and standard deviation of the normalized brightness profile for all these measurements is shown in Figure 2.
The relationship between distance and brightness profile was then estimated by the polynomial fitting of the data pairs (see in Figure 2). We then apply
this relation to every contaminating source whose brightness has been estimated as described above, to derive the flux density contributed by the blending source to the photometry of the exoplanet host star. This contaminating flux density, which is constant for a particular exoplanet host star, was then removed from our photometry values\footnote{
The code to apply this correction is available at \url{https://github.com/sailoryf/TESS_Deblending/
}}. The total contaminating flux percentage and uncertainty for the three stars is
6.22$\%$$\pm$0.42$\%$, 0.64$\%$$\pm$0.03$\%$ and 31.49$\%$$\pm$1.79$\%$ for KELT-19Ab, WASP-156b and WASP-121b respectively.
In addition, there is a possibility of blending from an unresolved binary companion.
In the case of WASP-156b and WASP-121b, there is no evidence of a binary companion
in past work \citep{Delrez,Demangeon}. In the case of KELT-19Ab, a nearby stellar companion was detected with the Palomar/Hale 200$''$ telescope using the near-infrared adaptive optics (AO) system \citep{KELT-19}. The stars have a measured magnitude difference of $\Delta J=2.50\pm0.06$ and $\Delta K_{s}=2.045\pm0.03$.
\citet{KELT-19} report a flux ratio of 0.0270$\pm$0.0034 at a wavelength of 5200$\pm$150$\AA$, effective temperatures of 7500$\pm$200K for the primary component, and 5200$\pm$100K for the secondary component. We search the ATLAS library \citep{Kurucz1993} grid based on the flux ratio and effective temperatures. The best fit stellar templates
are multiplied by the TESS filter response to estimate the brightness of the companion star and its contamination to
the TESS photometry. The derived contaminating flux fraction is 0.059$\pm$0.006 which we remove.
\begin{figure}[h]
\centering
\includegraphics[width=3.8in]{Distance_flux.pdf}
\caption{The normalized brightness profile of stars as seen by TESS used to estimate the amount of contamination of an exoplanet host stars' photometry by background stars seen by Gaia. The scatter points are the observed data. The red line indicates the polynomial fit to the data.}
\end{figure}
\subsection{Detrending of the Light Curve And Comparison with TESS-PDC Light Curve}
Measuring an accurate transit depth of an exoplanet requires a high precision light curve.
This requires detrending to remove the long term trends in the photometry outside of transit events \citep{GPdetrending,AGP,SFF,PDC}. Since we already know the existence of an exoplanet around
these stars, we can use prior knowledge of the transit mid-point time, period and transit duration to perform
a simple detrending.
To do this, we predicted the transit time in the TESS light curve and extract the parts of the light curve spanning 0.6 days centered at the
transit mid-point.
The data points of the light curve during the planetary transit were first masked out. We fitted a linear function to the signal in the masked, cut-off, light curve. The fit is used to estimate and remove the trending in the light curve
within a day. We applied this correction for every single data point of the extracted light curve and fold the light curve around the transit midpoint to
get the transit events shown in Figure 3. We also tried higher order polynomial functions, including orders of 2, 3, 5 and 10. The difference in the detrended light curve is negligible for these choices of polynomials.
The photometric data reduction could be influenced by factors already corrected in our pipeline, or by some factors beyond the consideration of our work, e.g., point spread function (PSF) variation, focus changes, bad pixels, and thermal variations. In order to make sure these factors do not dominate the precision of transit depth measurements, we compare
our results with light curves derived using the TESS Pre-Search Data Conditioning (PDC) modules. PDC light curve is produced by TESS collaboration with consideration of factors above. The comparison covers three targets in this work, as well as HD-219666, an isolated exoplanet host star identified by TESS which is bright compared to its surrounding stars \citep{Esposito2019}.
The deblending correction for HD-219666 is 0.212$\%\pm0.006\%$.
We compare the median flux of our light curve with the G-band flux reported by Gaia \citep{Gaia2018}. The flux density ratio of our decontaminated flux to Gaia flux for KELT-19Ab, WASP-156b, WASP-121b and HD-219666 is 97, 71, 90, and 80, respectively (in arbitrary units). The standard deviation is 9.7. The flux ratio of PDC flux to Gaia flux is 95, 66, 86, 75, respectively. The standard deviation is 11.0. This indicates that the two techniques are perfectly consistent with each other. Furthermore, we find that
the derived planetary transit depths from the two sets of light curves of these sources are within 1 $\sigma$ of each other.
\subsection{Planet Parameters Derived from TESS-Detected Transits}
\subsubsection{Fitting with Free Inclination and Semi-major Axis}
We used an MCMC fitting technique to derive the parameters of the exoplanets from the TESS transit curves \citep{Mandel_Agol2002}. We adopted a
circular orbit as in previous articles that confirmed the corresponding exoplanets with orbital periods derived therein \citep{KELT-19,Demangeon,Delrez}.
The value of T$_{eff}$, log $g$, and [Fe/H] was taken from the
discovery articles as well.
The free parameters in our fit are the radius ratio of the planet to the host star ($R_{p}/R_{\ast}$), the inclination of the planets orbit (i), the semi-major axis in unit of stellar radii ( $a/R_{\ast}$), time offset of transit center (T0) and the limb darkening parameters (a1 for linear limb-darkening coefficient, a2 for quadratic limb-darkening coefficient), as shown in table 1. Except for limb darkening, the other parameters have uniform priors.
We adopted the common inclination prior of $\mathcal{U}$[70,90] but changed it to be $\mathcal{U}$[70,110] when the posterior of inclination is within a degree of 90 degrees. This is because the distribution of inclination performs differently when the real inclination is close to 90. A prior cut-off at 90 degree severely distorts the posterior distribution to be non-Gaussian. A prior with no 90-degree cut-off yields a more reasonable posterior distribution. The difference in inclination posterior between these two scenarios of priors is systematic, but consistent within 1 $\sigma$.
The limb darkening model is the same as that adopted in the exoplanet discovery articles \citep{KELT-19,Demangeon,Evans}.
The coefficients for limb darkening have a Gaussian prior centered on the prediction for TESS data \citep{TESSLD} with an uncertainty of 0.05 \citep{YangLD}.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{KELT-19_A_B_lc.pdf}
\includegraphics[width=3.5in]{WASP-156_lc.pdf}
\includegraphics[width=3.5in]{WASP121_lc.pdf}
\caption{Fits to the TESS TPF transit light curves of KELT-19Ab (top), WASP-156b (middle), WASP-121b (bottom). The blue points are observed 2 minute data decontaminated using Gaia photometry. The green points are bins of 30 minutes cadence data. The red dashed line represents the best-fit transit model to the 2 minute data. The yellow region shows 1 $\sigma$ confidence region of the fits. The midpoint of the transit event is shifted to half the period, at the center of the x-axis. The free parameters in the fits are inclination, semi-major axis and limb darkening parameters as shown in Figure 4.}
\end{figure}
In the case of KELT-19Ab, we adopted T$_{eff}$, log \emph{g}, and [Fe/H] of 7500$\pm$110 K, 4.127$\pm$0.029 cgs, -0.12$\pm0.51$. As in \citet{KELT-19}, a quadratic limb darkening model is applied. The priors for the limb darkening coefficient are 0.27$\pm$0.05, 0.23$\pm$0.05, for the linear (linLimb) and quadratic coefficients (quadLimb), respectively.
In the case of WASP-156b, the T$_{eff}$, log \emph{g}, and [Fe/H] values are 5871$\pm$57 K, 4.35$\pm$0.03 cgs, 0.10$\pm0.10$. A quadratic limb darkening model is used \citep{Demangeon}.
The priors for limb darkening coefficients are 0.36$\pm$0.05 and 0.22$\pm$0.05 for linLimb and quadLimb respectively.
\begin{figure}[!htb]
\centering
\includegraphics[width=3in]{KELT-19_A_B_corner.pdf}
\includegraphics[width=3in]{WASP-156_corner.pdf}
\includegraphics[width=3in]{WASP121_corner.pdf}
\caption{The probability distribution of the fit parameters for KELT-19Ab (top), WASP-156b (middle), WASP-121b (bottom). The grayscale map indicate the density of samples derived from MCMC analysis for different pairs of parameters. The histograms along the diagonal are the marginalized density distribution for individual parameters. The dashed lines denote 68$\%$ of samples. The plot applies corner routine \citep{corner2016}. }
\end{figure}
In the case of WASP-121b, the T$_{eff}$, log \emph{g}, and [Fe/H] values are 6460$\pm$140 K, 4.2$\pm$0.2 cgs, 0.13$\pm0.09$. The limb darkening model used is the quadratic limb darkening model \citep{Delrez,Evans2018}. The limb darkening coefficients that we applied to our MCMC fits are 0.33$\pm$0.05, 0.21$\pm$0.05, respectively.
For each fitting, the parameters' probability distributions are derived using Python MCMC tool, PYMC \citep{pymc}. We applied 200000 iterations for the whole chain, and ignore the first 100000 steps to ensure the chain is stable. The best-fitting light curves are shown in Figure 3, respectively. The distribution of $R_{p}/R_{\ast}$, $a/R_{\ast}$, inclination and limb darkening parameters are shown in Figure 4.
The best fit results are shown in Table 1.
The inclinations derived from TESS TPF data for KELT-19Ab, WASP-156b, and WASP-121b are 3.3, 1.0 and 2.3 degrees larger than the results based on identification paper, respectively. Specifically, in case of WASP-121b, the inclination derived is more consistent with the follow-up work value 89.1 by \citet{Evans2018}.
Due to TESS's continuous sampling with 2 minutes and high precision photometry, the ingress and egress of transit are well monitored. These short phases in the transit duration are extremely sensitive to the inclination and semi-major axis. We, thereforebelieve that the TESS result as well as result from \citet{Evans2018} on inclination for WASP-121b are robust.
\subsubsection{Fitting with Fixed Inclination and Semi-major Axis}
Although our best fits are similar to other work, there are clear signs of differences, particularly in our derived inclination and transit depths.
In order to focus on differences in the radius ratio caused by wavelength dependent properties, we refit the TESS light curves, by fixing the inclination and semi-major axis to the referenced work \citep{KELT-19,Demangeon,Evans}.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{KELT-19_A_B_lc_fixed.pdf}
\includegraphics[width=3.5in]{WASP-156_lc_fixed.pdf}
\includegraphics[width=3.5in]{WASP121_lc_fix.pdf}
\caption{Results from fitting with fixed inclination and semi-major axis for KELT-19Ab (top), WASP-156b (middle), WASP-121b (bottom). The blue points are deconfused photometry. The green points indicate 30 minutes cadence binned result. The red dashed line represents best-fitting model. The black line indicates the result of previous work (referenced in the text). The yellow region shows 1 $\sigma$ confidence region of the fitting. The transit event is shifted at the center of x-axis which is half of the period.}
\end{figure}
The fitting followed the same procedure as section 2.4.1. The fitting results, as well as the derived parameters from previous work, are listed in Table 1. The differences between our fits and the result from the former work
are shown in Figure 5. The reduced chi-square in the fits are slightly larger than the best fit from previous section (Table 1). The small difference in the chi-square values indicate that the value of inclination and semi-major axis preferred in other work, is reasonably consistent with the TESS data but is not the preferred solution.
The fixed parameter MCMC fits have reduced uncertainties due to the strong priors on inclination and semi-major axis. These constraints on inclination and semi-major disturb the Gaussian distribution of other parameters. The revision for the three sigma rules for the parameters is beyond this work \citep[see in reference][]{Friedrich1994,MacKay2002,Hogg2018}. We cited the 1$\sigma$ uncertainty from the MCMC derived standard deviation for each parameter.
Our fits to the TESS data gives us a consistent planet to star radius ratio for WASP-121b and WASP-156b at the $\sim$2$\sigma$ level (Table 1). However, the planet to star radius ratio for KELT-19Ab is 7.5$\sigma$ smaller than previous work \citep{KELT-19}. The TESS radius ratio for WASP-121b is deeper than the median value including all the bands with the TESS wavelength range from \citep{Evans2018}. For WASP-156b, the radius ratio is smaller than past work \citep{Demangeon}.
\subsection{ Transit Parameters Bias Caused by Binning }
Sampling cadence results in morphological distortions to the transit light curve due to sampling of transit ingress and egress \citep{David2010}. For the TESS cadence, an assessment of the parameter bias is necessary for assessing the differences in transit depths for different bands.
We firstly investigate if our derived parameters from
the TESS 2 minute data are consistent with that from the 30 minute data by repeating the whole data reduction and fitting with 30 minute cadence image. In case of free inclination fitting, the inclinations derived for KELT-19Ab, WASP-156b, and WASP-121b are 4.6, 5.1 and 13.6 degrees smaller than the results based on 2 minutes binning data, respectively. Furthermore, the residuals from the light curve for WASP-121b when fixing the inclination shows obvious structure during the phase of ingress and egress (seen in Figure 6). The radius ratio when fixed inclination and semi-major axis is 0.1195$\pm$0.0004, about 7$\sigma$ smaller than 2 minutes result, with substantially worse chi-square values.
The inconsistency, especially for WASP-121b, motivates a simulation for a more detailed assessment of the differences.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{WASP121_lc_fix_30.pdf}
\includegraphics[width=3.5in]{residual_121.pdf}
\caption{Top: Result from fitting with a fixed inclination and semi-major axis for WASP-121b (TESS 30 minute data). The red dashed line represents the best-fitting model. The black line indicates the fit from previous work. Bottom: Residuals after fitting for the exoplanet transit around transit midpoint for WASP-121b based on 30 minutes data. The upper panel is residuals from the best fit. The lower panel is residuals assuming a fixed inclination and semi-major axis adopted
from previous work. The 30 minutes data are clearly
inconsistent with the previously derived inclination value.}
\end{figure}
We simulated 1000 light curves for each source. The input parameters were set to that in the discovery papers \citep[Table 1;][]{KELT-19, Demangeon, Delrez}. We note that for WASP-121b, the newest reported inclination is 1.5 degrees larger than initially assessed \citep{Evans2018}. The result from \citet{Evans2018} is also well matched to the TESS TPF derived result as shown in Table 1. But for the purpose of the simulations, we chose the value from the discovery paper. The time sampling for the simulated light curve is 1 second and is binned to 2 minutes and 30 minutes. The numbers of transit events are set to be the same as the event numbers observed by TESS. We add a Gaussian error to the light curves. The adopted 1$\sigma$ scatter is the standard deviation of the residuals from fitting the TESS light curves scaled by a factor of the square root of time scale.
These simulated light curves are then fit as described earlier.
Figure 7 and 8 show the distribution of derived parameters compared
to the input values.
We find that the magnitude of bias in the derived parameters especially inclination and semi-major axis, depends on the duration time of the transit, especially ingress and egress. We take WASP-121b as an illustrative example which gives the largest inclination difference. The derived median inclination value with the 30 minute cadence data is 76.4$\deg$ with a standard deviation of 0.7, 11.2 degrees smaller than the input value. The inclination decrease is because the binning smooths the rapid change of the flux decrease during ingress and egress. The planet to star radius ratio when fixing the inclination and semi-major axis is 0.1201$\pm$0.00035, also smaller than the input value at the level of $\sim5\sigma$ (seen in Figure 7).
\citet{David2010} provide a numerical solution to compensate for biases caused by binning, using Kepler’s equations.
We investigate the derived correction through a simulation. 1000 simulated light curves of 30 minute cadence
are generated with input parameters the same as described before. Using the batman routine \citep{Kreidberg2015}, we resample and integrate the simulated light curves to 2 min cadence. The derived fit inclination is 86.4$\pm$1.3 deg, indicating that resampling corrects the inclination bias to within 1$\sigma$. However, the derived R$_{p}$/R$_{\ast}$ from the resampled light curve has a median and scatter of 0.1231$\pm$0.00068. The bias in R$_{p}$/R$_{\ast}$ is 0.0011$\pm$0.00068, while the bias without resampling is -0.0019$\pm$0.00035.
Such differences of $\sim 0.8\%$ are negligible for most transit applications but need to be considered carefully for atmospheric characterization.
We use further simulations to assess if any R$_{p}$/R$_{\ast}$ bias after oversampling still exists. The resampling cadence is changed from 1 minute to 10 minutes with a step of 1 minute. The resampling cadence below 1 minute is also tested from 10 seconds to 50 seconds with a step of 10 seconds.
The R$_{p}$/R$_{\ast}$ reveals an offset at the level of 0.001 with an uncertainty $\sim$ 0.0006 when the resampling cadence is below 5 minutes. The offset is larger and increases with resampling cadence when the resampling cadence is larger than 5 minutes.
In addition, we apply simulations to check if the bias still exists when the initial sampling rate is 2 minutes. The resampling cadence is set to be 30 seconds, 1 minute, and 2 minutes. The derived R$_{p}$/R$_{\ast}$ does not reveal any evidence of offset suggesting that the batman routine itself does not induce any systematic bias.
We thus conclude that the resampling technique \citep{David2010, Kreidberg2015} significantly reduces the systematic bias caused by undersampling. However, we conclude that the very high precision (R$_{p}$/R$_{\ast}$ $\sim$ 0.5$\%$) needed for atmospheric analysis needs careful correction for sampling-induced systematics.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{30m_free_inl.pdf}
\includegraphics[width=3.5in]{30m_fix_rp.pdf}
\caption{Top: The derived inclination distribution from fitting 1000 simulated light curves sampled at 30 minutes, for the parameters of WASP-121b. Bottom: The planet to star radius ratio distribution when fixing the inclination and semi-major axis. The vertical lines for both panels give the input values to the simulations. Clearly, for WASP-121b, a light curve cadence of 30 minutes would result in a large bias in derived parameters.}
\end{figure}
In the case of light curves with 2 minute sampling, the bias is assessed in the same way. The best fit inclination for WASP-121b is found to be 87.7$\pm$1.3$\deg$. The planet to star radius ratio with a fixed inclination and semi-major axis is 0.1219$\pm$0.0004. We find the difference between the derived values and the input values to be negligible with the 2 minute data.
For all three sources, we find that
the simulations with 2 minute sampling yield similar parameters as the input value. Meanwhile, with 30 minutes cadence, the result is significantly biased when we do not correct for the undersampling \citep{David2010,Kreidberg2015}. Also, the differences in the inclination and planet-to-star radius ratio between 2 and 30 minutes cadence matches the difference between them from real TESS data. From these simulation, we therefore conclude that the biases in parameters from fitting TESS data of different cadence comes from binning, and that the fit parameters from the 2 minute cadence data for our three targets are robust.
Finally, we undertake a simulation to investigate the dependence of inclination and semi-major axis upon sampling time interval when not applying the oversampling technique (Figure 8). The sampling time interval in the simulation ranges from 1 minute to 50 minutes. For each interval, the light curve is generated as described above and fit. We duplicate 10 simulations for each sampling interval. The median fit values and standard deviations of the derived parameters are shown in Figure 8 for each sampling interval. It is clear that the inclination and semi-major axis are not significantly biased when the sampling time is less than 5 minutes. Also, the result indicates that the inclination and semi-major axis derived are distorted to be smaller when applying a larger sampling interval (as shown in Figure 8). This suggests that great care must be taken when comparing transit depths in different bands, especially when the light curve sampling intervals are likely to be quite different. The simulations were conducted without the oversampling correction \citep{David2010} which can mitigate the bias of inclination and semi-major axis as previously discussed.
\begin{figure}[!htb]
\centering
\includegraphics[width=3.5in]{SemiAInc.pdf}
\caption{Results from fitting inclination (green squares) and semi-major axis (red diamonds) to simulated light curves of different sampling intervals when not correcting for the undersampling. The vertical red dotted line gives the sampling interval limit below which the parameters derived are not biased and corresponds to 5 minutes. The horizontal black line dotted is the input value of the semi-major axis; the dotted blue line is the input value of inclination. Clearly, TESS FFI 30 minute cadence data would result in significantly biased parameters for planets with short duration transits such as those considered here.}
\end{figure}
\begin{table*}
\setlength{\tabcolsep}{3mm}
\begin{center}
\caption{Planet Parameters}
\label{tab1}
\begin{tabular}{ccccc}
\hline
\hline
\small
{Parameters }& {Description }& KELT-19Ab & WASP-156b & WASP-121b \\
& & 68$\%$ Confidence & 68$\%$ Confidence & 68$\%$ Confidence \\
\hline
\multicolumn{5}{c} {Best Fitting Parameters} \\
\hline
R$_{p}$/R$_{\ast}$ & Planet/star radius ratio in TESS & 0.09576$\pm$ 0.00082 & 0.0670 $\pm$ 0.0012 & 0.1234$\pm$0.0004\\
i [deg] & Inclination & 88.7$\pm0.7$ & 90.1$\pm1.4$ & 89.9$\pm1.6$ \\
a/R$_{\ast}$ & Semi-major axis of planet orbit in unit of stellar radii & 9.04$\pm0.19$ & 12.98$\pm+0.73$ & 3.81$\pm0.03$ \\
std of residual & Parts per million & 1253 & 2130 & 1763 \\
reduced $ \chi^{2}$ & Reduced chi-square & 1.001 & 1.0009 & 1.0004 \\
\hline
\multicolumn{5}{c} {Fitting Results with Fixed Inclination and Semi-major Axis} \\
\hline
R$_{p}$/R$_{\ast}$ & Planet/star radius ratio in TESS& 0.09940$\pm0.00048$ & 0.0662$\pm0.0009$ & 0.1230$\pm0.00025$\\
& Planet/star radius ratio in other work &
0.10713$\pm$0.00092$^{a}$ & 0.0685$^{+0.0012}_{-0.0008}$$^{e}$ & 0.1220$^{+0.0005}_{-0.0005}$ $^{b,c,d}$ \\
difference significance & Difference between TESS and former work in $\sigma$ & 7.549 & 1.533 & 1.789 \\
i [deg] & Inclination & 85.41$^{+0.34}_{-0.31}$$^{a}$ & 89.1$^{+0.6}_{-0.9}$$^{d}$ & 87.6$\pm$0.6$^{b,c}$ \\
&&&&89.1$\pm$0.5$^{d}$ \\
a/R$_{\ast}$ & Semi-major axis of planet orbit in unit of stellar radii & 7.50$^{+0.20}_{-0.18}$$^{a}$ & 12.8$^{+0.3}_{-0.7}$$^{e}$ & 3.754$^{+0.023}_{-0.028}$$^{b,c}$ \\
&&&&3.86$\pm$0.02$^{d}$\\
std of residual & Parts per million & 1266 & 2132 & 1765 \\
reduced $\chi^{2}$ & Reduced chi-square & 1.021 & 1.001 & 1.002 \\
linLimb & Linear limb-darkening coefficient & 0.29$\pm$0.04 & 0.37 $\pm$0.04 & 0.24 $\pm$0.03 \\
quadLimb & Quadratic limb-darkening coefficient & 0.23$\pm$0.04 & 0.23 $\pm$0.04 & 0.21 $\pm$0.04 \\
\hline
\end{tabular}
\end{center}
\begin{flushleft}
Reference:
(a) \citet{KELT-19};
(b) \citet{Delrez};
(c) \citet{Evans};
(d) \citet{Evans2018};
(e) \citet{Demangeon};
\end{flushleft}
\end{table*}
\subsection{A difference in stellar activity level?}
As explained in \citet{pont2008, Agol2010}, differences in the number of starspots between past observations and the TESS observations could account for differences in the measured transit depth. It was estimated that when the flux density of the star changes by 1\% because of stellar spots, the influence on the radius ratio is of the order of 10$^{-3}$R$_{p}$/R$_{\ast}$, comparable to the uncertainty of our TESS derived value. None of the three sources studied in this paper \citep{KELT-19,Demangeon,Evans} have shown
any past evidence of stellar activity. Furthermore, if starspots were significant, the standard deviation of the light
curve outside the transit should be different from that during transit. So we conclude that the impact
of starspots on our transit depth measurement is negligible.
\section{Atmospheric Constraints}
Our primary scientific motivation for undertaking this work was to compare $R_{p}/R_{\ast}$ in the TESS band with the value from previous work to constrain the atmospheres of the exoplanets. In particular, the TESS bandpass is significantly redder than the {\it Kepler} band and encompasses the
0.95\,$\mu$m water feature. The presence of these features in the planets'
atmosphere would result in a larger value of $R_{p}/R_{\ast}$ in the TESS band. For instance, the $R_{p}/R_{\ast}$ of KELT-19Ab in TESS data is significantly smaller, at the level of 7.5$\sigma$, indicating the possible existence of atmospheric features.
Our fixed inclination fits (Table 1) allow for a more direct comparison with past work.
The significance level of differences in $R_{p}/R_{\ast}$ are 7.5, 1.5, 1.8 $\sigma$ for KELT-19Ab, WASP-156b and WASP-121b, respectively. We discuss the implications of this difference for each of the exoplanets.
\subsection{KELT-19Ab}
KELT-19Ab is a giant planet transiting a moderately bright (V $\sim$ 9.9\,mag) A8V star with a orbital period of 4.61 days \citep{KELT-19}. The host stars has $M_{\ast}=1.62^{+0.25}_{-0.20}M_{\odot}$, $R_{\ast}=1.83\pm0.10R_{\odot}$. The planet has a radius of $R_{P}=1.91\pm0.11R_{J}$.
The ground-based light curves from \citet{KELT-19} in each band ($B$, $g$, $r$, $i(7519\AA)$, $I(8317\AA)$ and $z(8992\AA)$) were obtained. They are re-fit, to compare the variation in $R_{p}/R_{\ast}$ values between the different bands to that presented for the combined band. The light curves are performed with binary component deblending and linear detrending, using the same method as described previously. The flux contamination fraction adopted is the same as \citet{KELT-19}. The ground-based light curve does not have a long and stable baseline as TESS data. The long-term trend is therefore very difficult to estimate and dominates the $R_{p}/R_{\ast}$ uncertainty.
We masked the transit phase in the light curve as described previously, and fit the baseline, using different baseline lengths. This is repeated 5 times with the baseline length reduced by a factor of 10$\%$ each time. Three detrended light curves with the best residual are used for transit fitting. The $R_{p}/R_{\ast}$ and uncertainty are taken as the median value and the standard deviation of three fitted values. We do not fit the $B$ and $r$-band data because of the short baseline and partial coverage of the transit. The $i-$band values from the different telescopes are combined. KELT-19Ab has $R_{p}/R_{\ast}$ of 0.10713 derived from a joint fit to ground-based (B, g, r, i, I, z) data. The refit of our result presents smaller values of $R_{p}/R_{\ast}$ in each band compared to what has previously been published
\citep[0.10713,][]{KELT-19}. This indicates the $R_{p}/R_{\ast}$ differences might partially come from a different process choice.
We fitted the multi-band transit data with the grid of atmospheric transmission spectra from \citet{Goyal2019, Goyal2018} which are developed for hot, H$_{2}$/He dominated Jupiters.
This model is a 1D radiative-convective-equilibrium solution based on the ATMO code \citep{Tremblin2015,Tremblin2016, Drummond2016}. ATMO solves the radiative transfer for isothermal pressure-temperature (P-T) profiles, with opacity and chemical abundance being other parameters. The grid model has a parameter space of two chemical scenarios, 22 temperatures, 4 planetary gravities, 5 atmospheric metallicities, 4 C/O ratios, 4 scattering haze, 4 cloud parameters and a scaling factor to the specific planetary radius. Modifying specific chemical abundance will lead to different chemical equilibrium which is avoided in the use of the forward grid model.
In order to fit the model, we minimized the chi-square value between the observations and model predictions at all the bands. The model prediction at each band is derived by the following equation where $S_{\lambda}$ is the
filter response.
\begin{equation}
\frac{R_{p}}{R_{\ast}}=\frac{\int \frac{R_{p}}{R_{\ast}}_{\lambda}S_{\lambda}d{\lambda} }{\int S_{\lambda}d{\lambda}}
\end{equation}
We find that the haze dominated model at 1500K has the least $\chi$$^{2}$ of 2.80 (Figure \ref{plot:keltindi}). A clear model showing weak water vapor with rainout condensation at 1500K gives the second best $\chi^{2}$ of 3.05, while a clear model showing strong water vapor at 1000K has the third best $\chi^{2}$ of 3.26. A flat model with opaque featureless is also shown with a $\chi^{2}=3.14$. Other models resulted in much larger chi-square values (more than 6) and are not consistent with the data. We also obtain the Bayesian Information Criterion (BIC) for evaluating the model evidence. BIC theory is effective when the data points are much more than the free parameters \citep{BIC}. This might not be the case for KELT-19Ab but we still present the BIC here for reference. The BIC is 4.41, 9.49, 9.69 for haze, weak water, and strong water models. The BIC is 4.74 for an opaque featureless model.
The free parameter numbers of the opaque featureless model, haze, weak water, and strong water models are 1, 1, 4, and 4 respectively. The smaller $\chi^{2}$ and BIC of the haze model suggest that it is possibly the best interpretation given the current data. However, given the relatively narrow range in $\chi^{2}$ values and
small difference in BIC,
the opaque featureless model would also be a reasonable explanation.
\begin{figure*}[!htb]
\centering
\includegraphics[width=7in]{KELT19_transmission_spectrum_individualband_detrendingerror.pdf}
\caption{$R_{p}/R_{\ast}$ of KELT-19Ab in different bandpasses based on our re-analysis of those light curves. The red point is $R_{p}/R_{\ast}$ derived from TESS. The black points are $R_{p}/R_{\ast}$ from former work. In the top panel, the lowest $\chi$$^{2}$ model (haze dominant) is shown as a blue solid line. The green line indicates a flat, opaque featureless model that is consistent with the data. Blue diamonds are best-fit model predictions at the bands. The yellow diamond is the best fit model prediction at the SDSS $z$ band, green diamond is the best-fit model prediction at the Kepler band. In the bottom panel, a clear model with strong water (green) is shown with a faint water model (orange) as a comparison. Blue diamonds are strong water model predictions in different wavelengths bands while the green diamond is for the {\it Kepler} band. Dark green and brown diamonds are the strong water and faint water model prediction in the SDSS $z-$band. Clearly, 500ppm photometry in the $z-$band can distinguish between these scenarios.}
\label{plot:keltindi}
\end{figure*}
Based on our fits, we predicted the possible $R_{p}/R_{\ast}$ value in the {\it Kepler} band and $z-$ band based on these templates (Figure \ref{plot:keltindi}). The {\it Kepler} band would show $R_{p}/R_{\ast}$ of 0.1003 for a haze dominated atmosphere and 0.0997 for a clear model. In the $z-$band,
the presence of water would introduce a significant difference in the ratio. We estimate that a clear atmosphere model with strong water contribution would show $R_{p}/R_{\ast}$=0.1002. The value is smaller, 0.0990, from a clear model with faint water, as well as from a haze dominant model.
We find that a difference of 1.2$\%$ in $R_{p}/R_{\ast}$ arises between strong water contribution and faint water contribution models.
If we scale from our TESS measurements, and assume that the signal to noise
ratio depends on the square root of the number of points in the transit, we find that with 500ppm precision for 200 data points around the transit event in the $z-$band, one can distinguish between these two scenarios. This is of course a minimum since additional systematic effects can affect the measurements.
Alternately, instead of using our derived values of $R_{p}/R_{\ast}$ in each of the optical bands, we can use the value of \citet{KELT-19} which is derived from a joint fit to all the bands.
We treated their joint value of $R_{p}/R_{\ast}$ as the value in each band. The uncertainty follows the equation:
$\sigma_{joint}=\sigma_{i}/\sqrt{N}$, according to the error propagation law. $\sigma_{joint}$ is the joint band uncertainty, N is the number of bandpasses, $\sigma_{i}$ is the uncertainty in the band $i$. We assume the uncertainty in each band is the same.
Fitting $R_{p}/R_{\ast}$, the $\chi^{2}$ values for the different atmospheric models discussed - haze, strong water, weak water, and opaque featureless model - are 47.52, 50.68, 53.24, and 55.07, respectively. The BIC is 49.47, 58.46, 61.02, and 57.02 for the same models, respectively. Other models resulted in much larger chi-square values (more than 60) and are also not consistent with the data. Again, we find that a haze dominant atmosphere is preferred, as before.
\subsection{WASP-156b}
The discovery of WASP-156b was reported in \citet{Demangeon}. It is a super-Neptune orbiting a K type star (V=11.6\,mag, $M_{\ast}=0.842\pm0.052M_{\odot}$, $R_{\ast}=0.76\pm0.03R_{\odot}$) with a period of 3.84 days. It has a mass of $0.128^{+0.010}_{-0.009}M_{J}$, a radius of $0.51^{+0.02}_{-0.02}R_{J}$.
We followed the same procedure as for KELT-19Ab, to analyze the differences in $R_{p}/R_{\ast}$ in different bands. The published value of $R_{p}/R_{\ast}$ for comparison with our TESS analysis, is taken from \citet{Demangeon} at Johnson R, Gunn r, Johnson I and Gunn z (Figure \ref{plot:156}).
We find that a clear atmosphere at 1500K with strong water contribution (best fit model) has a $\chi^{2}$ of 2.35, and a BIC of 8.80. The faint water model at the same temperature follows with a worse $\chi^{2}$ of $\sim$ 2.44, and a BIC of 8.87. The haze model has a $\chi^{2}$ of 2.44 and a BIC of 4.04. An opaque featureless model has the same $\chi^{2}$ of 2.35 as the best fit model. The smallest BIC value of 3.96 indicates that the data might prefer the opaque featureless atmosphere model. The strong water contribution causes a shift of 0.9$\%$ in $z-$band prediction of $R_{p}/R_{\ast}$ compared to the faint water model. In the Kepler band, the haze dominant model has a larger $R_{p}/R_{\ast}$ of 1.3$\%$ than the no atmosphere model. We conclude that the current measurements are consistent with an opaque featureless model.
\begin{figure*}[!htb]
\centering
\includegraphics[width=7in]{WASP156_transmission_spectrum.pdf}
\caption{$R_{p}/R_{\ast}$ of WASP-156b in different bandpasses. The top panel shows the clear model with strong water (blue solid line), the haze dominant model (green solid line) and the opaque featureless model (brown solid line).
The red point is $R_{p}/R_{\ast}$ derived from TESS. The black points are $R_{p}/R_{\ast}$ from former work with the blue diamonds the "clear model" predictions at those bands. The yellow diamond is the model prediction at the SDSS $z-$band, green diamond is the model prediction at Kepler band. In the bottom panel, a clear model with strong water (green solid line) is shown along with a faint water model (Orange solid line) for comparison.
The empty green diamond is a strong water model prediction at Kepler band. The dark green diamond is the strong water model prediction at the SDSS $z-$band. Brown diamond is faint water model prediction at $z-$band.}
\label{plot:156}
\end{figure*}
\subsection{WASP-121b}
WASP-121b was discovered by \citet{Delrez}. It is a hot Jupiter around a main sequence star (V=10.4\,mag, $1.353^{+0.080}_{-0.079}M_{\odot}$, $1.458\pm0.030R_{\odot}$) with a period of 1.28 days. The planets has a mass of $1.183^{+0.064}_{-0.062}M_{J}$ and a radius of $1.865\pm0.044R_{J}$.
We used external observations from \citet{Delrez, Evans2016, Evans2018}
in conjunction with our analysis of the TESS data (Figure \ref{plot:121}).
These measurements utilized the standard ground-based filter, STIS, and WFC3 instruments on {\it Hubble} and span 0.3 to 1.65$\mu$m.
As outlined in \citet{Evans2018}, a global fit for the whole spectrum is challenging due to the complex problem of non-equilibrium chemistry.
Our fits with the inclination fixed to past work however, yield a value of $R_{p}/R_{\ast}$ which is in very good agreement with \citet{Evans2018}. The median value of the radius ratio is 0.1225 at the wavelength arrange of TESS.
To fit the spectrum, we again use the general grid model from \citet{Goyal2019}.
\begin{figure*}[!htb]
\centering
\includegraphics[width=7in]{WASP-121_transimission_spectrum.pdf}
\caption{The transmission spectrum for WASP-121b. The solid yellow point is the best fit value of $R_{p}/R_{\ast}$ from the TESS light curve. However, if the inclination is fixed to previous work for a consistent comparison, the solid redpoint is $R_{p}/R_{\ast}$ derived from TESS data. The black points are $R_{p}/R_{\ast}$ from former measurements.
The grey solid line is the best fit from \citet{Evans2018}. The blue line shows the haze model. The brown line shows the opaque featureless model.
The black line indicates a clear atmosphere at a cool temperature (shifted for clarity).
\label{plot:121}
}
\end{figure*}
The best-fit model from \citet{Evans2018} is still the best fit model when we add the TESS constraints. The fitting has a reduced chi-square of 4.91, a BIC of 449.74, with 92 data points, 4 fixed, and 4 free parameters (in Figure \ref{plot:121}). The fixed parameters are local condensation, atmospheric metallicities at 20x, haze parameters of 1.0, and no cloud. The fitted free parameters are temperature of 1500K, gravity at 10ms$^{-2}$, C/O ratio of 0.7, and planet radius (without atmosphere).
An opaque featureless model has a reduced chi-square of 5.19, a BIC of 477.39. The haze dominant model is with a chi-square of 5.87, a BIC of 538.76. A transmission model at 1500K with features of potassium, sodium, and water vapor has a much larger reduced chi-square of 27.40, a BIC of 2429.51, and seems to be ruled out. According to BIC theory, any model having a BIC smaller than 10 is favored \citep{BICmostcited}.
\section{SUMMARY AND DISCUSSION}
We have analyzed the TESS data for KELT-19Ab, WASP-156b, WASP-121b to
measure $R_{p}/R_{\ast}$ in a broadband spanning 600nm to 1000nm. We identify the significant role of contamination by unresolved background stars in the TESS data and
use Gaia data to subtract their contribution to the transit light curves.
We provide general-purpose software to address this issue for all TESS data.
We detrend the light curves through a linear fit to the photometry near the transit events. To derive the exoplanet transit parameters, we performed two ways of MCMC model fitting: with and without inclination and semi-major axis fixed. The former in particular helps with the comparison with past work on these stars. Using simulations, we demonstrate the systematic biases in inclination and radius ratio with 30 minutes cadence data when no correction for undersampling is undertaken \citep{David2010}.
We also find the inclination and semi-major axis are distorted to be smaller with larger sampling times. However,
we find that 2 minute sampling does not cause any bias for transit parameters. The atmosphere of these three exoplanets was then constrained based on $R_{p}/R_{\ast}$ in different bands.
In the case of KELT-19Ab, we find that the TESS-measured $R_{p}/R_{\ast}$ is $\sim$7.5$\sigma$ smaller than past work. The $R_{p}/R_{\ast}$ could be due to the choice of analysis or due to the presence of atmospheric features which further observations in narrower bands can reveal. The model fitting possibly favors a haze dominant atmosphere. In the case of WASP-156b, we find that the TESS transit depths agree with previous work and prefer a no atmosphere model fit to the measurements.
In the case of WASP-121b, the fits to the TESS light curve prefer the newest reported inclination. We find that the atmospheric transmission spectrum is best fitted with a 1500K model with water vapor as reported before.
Our work shows that exoplanet atmospheric model constraints benefit from high precision continuous photometry which provides strong constraints on key parameters such as inclination.
Adding one high precision observation at certain bands improves the constraints on atmospheric composition. Since TESS
has now observed the Northern Hemisphere,
we now have Kepler \citep{Kepler2010} and TESS constraints on a few bright stars, which will help assess the role of haze in exoplanet atmospheres. Finally, we show that precision photometry with $\sim$500ppm to 1000ppm in the $z-$band can help directly constrain the presence of water vapor in the exoplanet atmosphere, when combined with precise TESS photometry.
\acknowledgments
We are very grateful to the thoughtful comments by a referee which greatly improved the quality of results presented in this manuscript.
This work made use of PyAstronomy\footnote{https://github.com/sczesla/PyAstronomy} \citep{pya} and the NASA Exoplanet Archive \citep{ExoplanetArchive}. We would like to thank Jayesh M. Goyal for useful inputs and discussion and Caltech Optical Observatories for supporting the visit by the lead author. We also thank Hilke Schlichting for useful feedback on the manuscript. We sincerely thank Karen A. Collins for providing KELT-19Ab ground-based light curves. Fan Yang and Ji-Feng Liu acknowledge funding from National Natural Science Foundation of China (NSFC.11988101), National Key Research and Development Program of China (No.2016YFA0400800) and National Science Fund for Distinguished Young Scholars (No.11425313).
\bibliographystyle{aasjournal}
\section{}
\textit{Research Notes of the \href{https://aas.org}{American Astronomical Society}}
(\href{http://rnaas.aas.org}{RNAAS}) is a publication in the AAS portfolio
(alongside ApJ, AJ, ApJ Supplements, and ApJ Letters) through which authors can
promptly and briefly share materials of interest with the astronomical community
in a form that will be searchable via ADS and permanently archived.
The astronomical community has long faced a challenge in disseminating
information that may not meet the criteria for a traditional journal article.
There have generally been few options available for sharing works in progress,
comments and clarifications, null results, and timely reports of observations
(such as the spectrum of a supernova), as well as results that wouldn’t
traditionally merit a full paper (such as the discovery of a single exoplanet
or contributions to the monitoring of variable sources).
Launched in 2017, RNAAS was developed as a supported and long-term
communication channel for results such as these that would otherwise be
difficult to broadly disseminate to the professional community and persistently
archive for future reference.
Submissions to RNAAS should be brief communications - 1,000 words or fewer
\footnote{An easy way to count the number of words in a Research Note is to use
the \texttt{texcount} utility installed with most \latex\ installations. The
call \texttt{texcount -incbib -v3 rnaas.tex}) gives 57 words in the front
matter and 493 words in the text/references/captions of this template. Another
option is by copying the words into MS/Word, and using ``Word Count'' under the
Tool tab.}, and no more than a single figure (e.g. Figure \ref{fig:1}) or table
(but not both) - and should be written in a style similar to that of a
traditional journal article, including references, where appropriate, but not
including an abstract.
Unlike the other journals in the AAS portfolio, RNAAS publications are not
peer reviewed; they are, however, reviewed by an editor for appropriateness
and format before publication. If accepted, RNAAS submissions are typically
published within 72 hours of manuscript receipt. Each RNAAS article is
issued a DOI and indexed by ADS \citep{2000A&AS..143...41K} to create a
long-term, citable record of work.
Articles can be submitted in \latex\ (preferably with the new "RNAAS"
style option in AASTeX v6.2), MS/Word, or via the direct submission in the
\href{http://www.authorea.com}{Authorea} or
\href{http://www.overleaf.com}{Overleaf} online collaborative editors.
Authors are expected to follow the AAS's ethics \citep{2006ApJ...652..847K},
including guidance on plagiarism \citep{2012AAS...21920404V}.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.85,angle=0]{aas.pdf}
\caption{Top page of the AAS Journals' website, \url{http://journals.aas.org},
on October 15, 2017. Each RNAAS manuscript is only allowed one figure or
table (but not both). Including the
\href{http://journals.aas.org//authors/data.html\#DbF}{data behind the figure}
in a Note is encouraged, and the data will be provided as a link in the
published Note.\label{fig:1}}
\end{center}
\end{figure}
\acknowledgments
Acknowledge people, facilities, and software here but remember that this counts
against your 1000 word limit.
| 564d5c23bc67dfa708f85f0445221fcc40b5f8d7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{intro}
The pelvis is an important structure connecting the spine and lower limbs and plays a vital role in maintaining the stability of the body and protecting the internal organs of the abdomen. The abnormality of the pelvis, like hip dysplasia~\cite{HipDysplasia} and pelvic fractures~\cite{2018pelvicFracture}, can have a serious impact on our physical health. For example, as the most severe and life-threatening bone injuries, pelvic fractures can wound other organs at the fracture site, and the mortality rate can reach 45\%~\cite{2020openpelvic} at the most severe situation, the open pelvic fractures. Medical imaging~\cite{zhou2021review,zhou2019handbook} plays an important role in the whole process of diagnosis and treatment of patients with pelvic injuries. Compared with X-Ray images, CT preserves the actual anatomic structure including depth information, providing more details about the damaged site to surgeons, so it is often used for 3D reconstruction to make follow-up \textbf{surgery planning} and evaluation of postoperative effects. In these applications, \textbf{accurate pelvic bone segmentation} is crucial for assessing the severity of pelvic injuries and helping surgeons to make correct judgments and choose the appropriate surgical approaches. In the past, surgeons segmented pelvis manually from CT using software like Mimics\footnote{\url{https://en.wikipedia.org/wiki/Mimics}}, which is time-consuming and non-reproducible. To address these clinical needs, we here present an automatic algorithm that can accurately and quickly segment pelvic bones from CT.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{Fig1.pdf}
\caption{Pelvic CT image examples with various conditions.} \label{fig-pic-3}
\end{figure}
Existing methods for pelvic bone segmentation from CT mostly use simple thresholding~\cite{gaussianThreshold_a6}, region growing~\cite{wavelet_a7}, and handcrafted models, which include deformable models~\cite{deformable_a8,activeContour_a9}, statistical shape models~\cite{shapeModel_a10,3DshapeModel_sta1}, watershed~\cite{keyframes_a11} and others~\cite{mriPelvicStruc_tra1,knowledgeOrganSpecificStrategies_tra2,femurSeg_tra3,femurXray_tra4,RFregressionVoting_rfr1,RF&HsparseShape_rfr2}. These methods focus on local gray information and have limited accuracy due to the density differences between cortical and trabecular bones. And trabecular bone is similar to that of the surrounding tissues in terms of texture and intensity. Bone fractures, if present, further lead to weak edges. Recently, deep learning-based methods~\cite{FCN_a13,unet_a12,FabianNNUnet_a4,zhao2017pyramid_dl1,atrousSeparableConvolution_dl2,3dUNet_dl3,pancreaticCyst_dl4,DenseVNet_dl5} have achieved great success in image segmentation; however, their effectiveness for CT pelvic bone segmentation is not fully known. Although there are some datasets related to pelvic bone~\cite{lee2012virtual_pelvicBoneDataset1,wu2016segmentation_pelvicBoneDataset2,hemke2020deep_pelvicBoneDataset3,chandar2016segmentation_pelvicBoneDataset4}, only a few of them are open-sourced and with small size (less than 5 images or 200 slices), far less than other organs~\cite{kits19_url3,MSD}. Although ~\cite{hemke2020deep_pelvicBoneDataset3} conducted experiments based on deep learning, the result was not very good (Dice=0.92) with the dataset only having 200 CT slices. For the robustness of the deep learning method, it is essential to have a comprehensive dataset that includes as many real scenes as possible. In this paper, we bridge this gap by curating a large-scale CT dataset and explore the use of deep learning in this task, which marks, to the best of our knowledge, \textbf{the first real attempt} in this area, with more statistical significance and reference value.
To build a comprehensive dataset, we have to deal with diverse image appearance variations due to differences in imaging resolution and field-of-view (FOV), domain shift arising from different sites, the presence of contrasted vessels, coprolith and chyme, bone fractures, low dose, metal artifacts, etc. Fig. \ref{fig-pic-3} gives some examples about these various conditions. Among the above-mentioned appearance variations, the challenge of the metal artifacts is the most difficult to handle. Further, we aim at a multi-class segmentation problem that separates the pelvis into multiple bones, including \textit{lumbar spine}, \textit{sacrum}, \textit{left hip}, and \textit{right hip}, instead of simply segmenting out the whole pelvis from CT.
The contributions of this paper are summarized as follows:
\begin{itemize}
\item A \textit{pelvic CT dataset} pooled from multiple domains and different manufacturers, including $1,184$ CT volumes (over 320K CT slices) of diverse appearance variations (including 75 CTs with metal artifacts). Their multi-bone labels are carefully annotated by experts. We open source it to benefit the whole community;
\item Learning a \textit{deep multi-class segmentation network}~\cite{FabianNNUnet_a4} to obtain more effective representations for joint lumbar spine, sacrum, left hip, and right hip segmentation from multi-domain labeled images, thereby yielding desired accuracy and robustness;
\item A \textit{fully automatic analysis pipeline} that achieves high accuracy, efficiency, and robustness, thereby enabling its potential use in clinical practices.
\end{itemize}
\section{Our Dataset}
\label{sec:1}
\begin{table}[t]
\centering
\caption{Overview of our large-scale Pelvic CT dataset. `\#' represents the number of 3D volumes. `Tr/Val/Ts' denotes training/validation/testing set. Ticks[\Checkmark] in table refer to we can access the CT images' acquisition equipment manufacturer[M] information of that sub-dataset. Due to the difficulty of labeling the CLINIC-metal, CLINIC-metal is taken off in our supervised training phase.}\label{tab0}
\newsavebox{\tablebox}
\begin{lrbox}{\tablebox}
\begin{tabular}{lrcccc}
\hline
Dataset name[M] & \# & Mean spacing(mm) & Mean size & \# of Tr/Val/Ts & Source and Year \\
\hline
ABDOMEN & 35 & (0.76, 0.76, 3.80) & (512, 512, 73) & 21/7/7 & Public 2015\\
\hline
COLONOG[\Checkmark] & 731 & (0.75, 0.75, 0.81) & (512, 512, 323) & 440/146/145& Public 2008\\
\hline
MSD\_T10 & 155 & (0.77, 0.77, 4.55) & (512, 512, 63) & 93/31/31& Public 2019\\
\hline
KITS19 & 44 & (0.82, 0.82, 1.25) & (512, 512, 240) & 26/9/9& Public 2019\\
\hline
CERVIX & 41 & (1.02, 1.02, 2.50) & (512, 512, 102) &24/8/9 & Public 2015 \\
\hline
CLINIC[\Checkmark] & 103 & (0.85, 0.85, 0.80) & (512, 512, 345) & 61/21/21& Collected 2020\\
\hline
CLINIC-metal[\Checkmark] & 75 & (0.83, 0.83, 0.80) &(512, 512, 334) & 0(61)/0/14 & Collected 2020\\
\hline\hline
Our Datasets & $1,184$ &(0.78, 0.78, 1.46) &(512, 512, 273) & 665(61)/222/236 & -\\
\hline
\end{tabular}
\end{lrbox}
\scalebox{0.8}[0.8]{\usebox{\tablebox}}
\end{table}
\subsection{Data Collection}
To build a comprehensive pelvic CT dataset that can replicate practical appearance variations, we curate a large dataset of pelvic CT images from seven sources, two of which come from a clinic and five from existing CT datasets~\cite{COLONOGRAPHY,kits19_url3,MSD,matlas}. The overview of our large dataset is shown in Table~\ref{tab0}.
These seven sub-datasets are curated separately from different sites and sources with different characteristics often encountered in the clinic. In these sources, we exclude some cases of very low quality or without pelvic region and remove the unrelated areas outside the pelvis in our current dataset.
Among them, the raw data of COLONOG, CLINIC, and CLINIC-metal are stored in a DICOM format, with more information like scanner manufacturers can be accessed. More details about our dataset are given in Online Resource 1\footnote{\url{https://drive.google.com/file/d/115kLXfdSHS9eWxQmxhMmZJBRiSfI8_4_/view}\label{em1}}.
We reformat all DICOM images to NIfTI to simplify data processing and de-identify images, meeting the institutional review board (IRB) policies of contributing sites. All existing sub-datasets are under Creative Commons license CC-BY-NC-SA at least and we will keep the license unchanged. For CLINIC and CLINIC-metal sub-datasets, we will open-source them under Creative Commons license CC-BY-NC-SA
4.0.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{Fig2.pdf}
\caption{The designed annotation pipeline based on an AID (Annotation by Iterative Deep Learning) strategy. In Step \uppercase\expandafter{\romannumeral1}, two senior experts first manually annotate 40 cases of data as the initial database. In Step \uppercase\expandafter{\romannumeral2}, we train a deep network based on the human annotated database and use it to predict new data. In Step \uppercase\expandafter{\romannumeral3}, initial annotations from the deep network are checked and modified by human annotators. Step \uppercase\expandafter{\romannumeral2} and Step \uppercase\expandafter{\romannumeral3} are repeated iteratively to refine a deep network to a more and more powerful `annotator'. This deep network `annotator' also unifies the annotation standards of different human annotators.} \label{fig-pic-10}
\end{figure}
\subsection{Data Annotation}
Considering the scale of thousands of cases in our dataset and annotation itself is truly a subjective and time-consuming task. We introduce a strategy of Annotation by Iterative Deep Learning
(AID)~\cite{AIL} to speed up our annotation process. In the AID workflow, we train a deep network with a few precisely annotated data in the beginning. Then the deep network is used to automatically annotate more data, followed by revision from human experts. The human-corrected annotations and their corresponding images are added to the training set to retrain a more powerful deep network. These steps are repeated iteratively until we finish our annotation task. In the last, only minimal modification is needed by human experts.
The annotation pipeline is shown in Fig.~\ref{fig-pic-10}. In Step \uppercase\expandafter{\romannumeral1}, two senior experts are invited to pixel-wise annotate 40 cases of CLINIC sub-dataset precisely as the initial database based on the results from simple thresholding method, using ITK Snap (Philadelphia, PA) software. All annotations are performed in the transverse plane. The sagittal and coronal planes are used to assist the judgment in the transverse plane. The reason for starting from the CLINIC sub-dataset is that the cancerous bone and surrounding tissues exhibit similar appearances at the fracture site, which needs more prior knowledge guidance from doctors. In Step \uppercase\expandafter{\romannumeral2}, we train a deep network with the updated database and make predictions on new 100 data selected randomly at a time. In Step \uppercase\expandafter{\romannumeral3}, some junior annotators refine the labels based on the prediction results, and each junior annotator is only responsible for part of 100 new data. A coordinator will check the quality of refinement by all junior annotators. For easy cases, the annotation process is over in this stage; for hard cases, senior experts are invited to make more precise annotations. Step \uppercase\expandafter{\romannumeral2} and Step \uppercase\expandafter{\romannumeral3} are repeated until we finish the annotation of all images in our dataset. Finally, we conduct another round of scrutiny for outliers and mistakes and make necessary corrections to ensure the final quality of our dataset.
In Fig.~\ref{fig-pic-10}, `Junior annotators' are graduate students in the field of medical image analysis. The `Coordinator' is a medical image analysis practitioner with many years of experience, and the `Senior experts' are cooperating doctors in the partner hospital, one of the best orthopedic hospitals in our country.
In total, we have annotations for $1,109$ metal-free CTs and 14 metal-affected CTs. The remaining 61 metal-affected CTs of image are left unannotated and planned for use in unsupervised learning.
\section{Segmentation Methodology}
\label{sect-2}
The overall pipeline of our deep approach for segmenting pelvic bones is illustrated in Fig.~\ref{fig-pic-5}. The input is a 3D CT volume. (i) First, the input is sent to our segmentation module. It is a plug and play (PnP) module that can be replaced at will.
(ii) After segmentation is done, we send the multi-class 3D prediction to a SDF post-processor, which removes some false predictions and outputs the final multi-bone segmentation result.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\textwidth]{Fig3.pdf}
\caption{Overview of our pelvic bones segmentation system, which learns from multi-domain CT images for effective and robust representations. The 3D U-Net cascade is used here to exploit more spatial information in 3D CT images. SDF is introduced to our post-processor to add distance constraint besides size constraint used in traditional MCR-based method.} \label{fig-pic-5}
\vspace{-0.0cm}
\end{figure}
\subsection{Segmentation Module}
\label{subsectionSegModule}
Based on our large-scale dataset collected from multiple sources together with annotations, we use a fully supervised method to train a deep network to learn an effective representation of the pelvic bones.
The deep learning framework we choose here is 3D U-Net cascade version of nnU-Net~\cite{FabianNNUnet_a4},
which is a robust state-of-the-art deep learning-based medical image segmentation method.
3D U-Net cascade contains two 3D U-net, where the first one is trained on downsampled images (stage 1 in Fig.~\ref{fig-pic-5}), the second one is trained on full resolution images (stage 2 in Fig.~\ref{fig-pic-5}).
A 3D network can better exploit the useful 3D spatial information in 3D CT images. Training on downsampled images first can enlarge the size of patches in relation to the image, then also enable the 3D network to learn more contextual information. Training on full resolution images second refines the segmentation results predicted from former U-Net.
\subsection{SDF Post Processor}
Post-processing is useful for a stable system in clinical use, preventing some mispredictions in some complex scenes. In the segmentation task, current segmentation systems usually determine whether to remove the outliers according to the size of the connected region to reduce mispredictions. However, in the pelvic fractures scene, broken bones may also be removed as outliers. To this end, we introduce the SDF~\cite{sdf_url1} filtering as our post-processing module to add a \textit{distance constraint} besides the \textit{size constraint}. We calculate SDF based on the maximum connected region (MCR) of
the anatomical structure in the prediction result, obtaining a 3D distance map that increases from the bone border to the image boundary to help determining whether `outlier prediction' defined by traditional MCR-based method should be removed.
\section{Experiments}\label{exp}
\subsection{Implementation Details}
We implement our method based on open source code of nnU-Net\footnote{\url{https://github.com/mic-dkfz/nnunet}}~\cite{FabianNNUnet_a4}. We also used MONAI\footnote{\url{https://monai.io/}} during our algorithm development. Details please refer to the Online Resource 1\textsuperscript{\ref {em1}}. For our metal-free dataset, we randomly select 3/5, 1/5, 1/5 cases in each sub-dataset as the training set, validation set, and testing set, respectively, and keep such a data
partition unchanged in all-dataset experiments and sub-datasets experiments.
\begin{table}[t]
\centering \small
\caption{(a) The DC and HD results for different models tested on `ALL' dataset. (b) Effect of different post-processing methods on `ALL' dataset. `ALL’ refers to the six metal-free sub-datasets. `Average' refers to the mean value of four anatomical structures' DC/HD.
`Whole' refers to treating Sacrum, Left hip, Right hip and Lumbar spine as a whole bone. The top three numbers in each part are marked in \textbf{bold}, \textcolor{red}{red} and \textcolor{blue}{blue}.}\label{tab_basic}
\begin{lrbox}{\tablebox}
\begin{tabular}{lllllllll}
\begin{tabular}{lllllllll}
\hline
Exp & Test & Model & Whole & Sacrum & Left hip & Right hip & Lumbar spine & Average \\
\cline{4-9}
& (Dataset)&(Dataset)&Dice/HD&Dice/HD&Dice/HD&Dice/HD&Dice/HD&Dice/HD\\
\hline
(a)& ALL & $\Phi_{ALL(2.5D)}$ & \textcolor{red}{.988}/\textbf{9.28}
& \textcolor{red}{.979}/\textcolor{blue}{9.34}
& \textbf{.990}/\textbf{3.58}
& \textcolor{red}{.990}/\textcolor{red}{3.44}
& \textcolor{blue}{.978}/\textcolor{blue}{8.32}
& \textcolor{blue}{.984}/\textcolor{red}{6.17} \\
& ALL & $\Phi_{ALL(3D)}$ & \textcolor{red}{.988}/\textcolor{blue}{11.38}
& \textbf{.984}/\textcolor{red}{8.13}
& \textcolor{blue}{.988}/\textcolor{blue}{4.99}
& \textcolor{red}{.990}/\textcolor{blue}{4.26}
& \textcolor{red}{.982}/\textcolor{red}{7.80}
& \textcolor{red}{.986}/\textcolor{blue}{6.30}\\
& ALL & $\Phi_{ALL(3D\_cascade)}$& \textbf{.989}/\textcolor{red}{10.23}
& \textbf{.984}/\textbf{7.24}
& \textcolor{red}{.989}/\textcolor{red}{4.24}
& \textbf{.991}/\textbf{3.03}
& \textbf{.984}/\textbf{7.49}
& \textbf{.987}/\textbf{5.50}\\
\hline
\hline
(b)& w/o Post & $\Phi_{ALL}$ & \textcolor{red}{.988}/36.27
& \textbf{.984}/\textcolor{blue}{38.36}
& \textcolor{red}{.988}/\textcolor{blue}{35.43}
& \textbf{.991}/28.70
& \textcolor{red}{.983}/11.25
& \textbf{.987}/28.43\\
& MCR & $\Phi_{ALL}$ & \textcolor{red}{.988}/12.93
& \textbf{.984}/\textcolor{red}{7.50}
& \textbf{.989}/\textbf{4.24}
& \textbf{.991}/3.72
& .978/10.46
& \textcolor{red}{.986}/6.48\\
& SDF(5)& $\Phi_{ALL}$ & \textcolor{red}{.988}/12.02
& \textbf{.984}/\textbf{7.24}
& \textbf{.989}/\textbf{4.24}
& \textbf{.991}/3.51
& \textcolor{blue}{.980}/\textcolor{blue}{9.54}
& \textcolor{red}{.986}/6.13\\
& SDF(15)& $\Phi_{ALL}$ & \textbf{.989}/\textcolor{red}{10.40}
& \textbf{.984}/\textbf{7.24}
& \textbf{.989}/\textbf{4.24}
& \textbf{.991}/\textcolor{red}{3.35}
& \textbf{.984}/\textcolor{red}{7.61}
& \textbf{.987}/\textcolor{red}{5.61}\\
& SDF(35)& $\Phi_{ALL}$ & \textbf{.989}/\textbf{10.23}
& \textbf{.984}/\textbf{7.24}
& \textbf{.989}/\textbf{4.24}
& \textbf{.991}/\textbf{3.03}
& \textbf{.984}/\textbf{7.49}
& \textbf{.987}/\textbf{5.50}\\
& SDF(55)& $\Phi_{ALL}$ & \textbf{.989}/\textcolor{blue}{10.78}
& \textbf{.984}/\textbf{7.24}
& \textbf{.989}/\textcolor{red}{4.52}
& \textbf{.991}/\textcolor{blue}{3.38}
& \textbf{.984}/\textbf{7.49}
& \textbf{.987}/\textcolor{blue}{5.66}\\
\hline
\end{tabular}
\end{tabular}
\end{lrbox}
\scalebox{0.7}[0.7]{\usebox{\tablebox}}
\end{table}
\subsection{Results and Discussion}\label{results}
\subsubsection{Segmentation Module}
To prove that learning from our large-scale pelvic bones CT dataset is helpful to improve the robustness of our segmentation system, we conduct a series of experiments in different aspects.
\textbf{Performance of baseline models.}
Firstly, we test the performance of models of different dimensions on our entire dataset. The Exp (a) in Table~\ref{tab_basic} shows the quantitative results. $\Phi_{ALL}$ denotes a deep network model trained on `ALL' dataset. Following the conventions in most literature, we use Dice coefficient(DC) and Hausdorff distance (HD) as the metrics for quantitative evaluation. All results are tested on our testing set. Same as we discussed in Sect.~\ref{subsectionSegModule}, $\Phi_{ALL(3D\_cascade)}$ shows the best performance, achieving an average DC of 0.987 and HD of 5.50 voxels, which means 3D U-Net cascade can learn the semantic features of pelvic anatomy better then 2D/3D U-Net. As the following experiments are all trained with 3D U-Net cascade, the mark $_{(3D\_cascade)}$ of $\Phi_{ALL(3D\_cascade)}$ is omitted for notational clarity.
\begin{table}[t]
\centering \small
\caption{The `Average' DC and HD results for different models tested on different datasets. Please refer to the Online Resource 1\textsuperscript{\ref {em1}} for details. The top three numbers in each part are marked in \textbf{bold}, \textcolor{red}{red} and \textcolor{blue}{blue}.}\label{tab_subdataset}
\begin{lrbox}{\tablebox}
\begin{tabular}{llllllll}
\hline
Average Dice/HD & & & Test & Datasets & & \\
\cline{2-8}
Models &ALL &ABDOMEN&COLONOG&MSD\_T10&KITS19&CERVIX&CLINIC\\
\hline
$\Phi_{ABDOMEN}$&.604/92.81
& \textbf{.979}/5.84
&.577/104.04
&\textcolor{blue}{.980}/3.74
&.360/158.02
&.305/92.07
&.342/148.12\\
$\Phi_{COLONOG}$& \textcolor{red}{.985}/\textcolor{red}{5.84}
& \textcolor{blue}{.975}/3.29
&\textbf{.989}/\textbf{5.65}
&.979/4.41
&\textcolor{blue}{.982}/8.41
&.969/5.17
&.974/9.24 \\
$\Phi_{MSD\_T10}$ & .534/96.14
& \textbf{.979}/\textcolor{blue}{2.97}
&.501/106.86
&\textbf{.987}/\textcolor{red}{3.36}
&.245/170.39
&.085/111.72
&.261/151.61 \\
$\Phi_{KITS19}$ &.704/68.29
& .255/120.31
&.746/70.75
&.267/121.57
&\textbf{.986}/\textbf{5.65}
&\textcolor{red}{.973}/5.14
& \textcolor{blue}{.977}/9.25 \\
$\Phi_{CERVIX}$ &\textcolor{blue}{.973}/\textcolor{blue}{14.75}
& .969/4.30
& \textcolor{blue}{.974}/18.74
&.967/6.55
&.979/\textcolor{blue}{7.78}
& \textcolor{red}{.973}/\textbf{4.49}
& .974/10.17 \\
$\Phi_{CLINIC}$ & .692/69.89
& .275/117.09
&.728/71.93
&.254/126.66
&\textcolor{red}{.985}/11.16
& .968/9.69
& \textbf{.983}/\textbf{7.27}\\
$\Phi_{ALL}$ & \textbf{.987}/\textbf{5.50}
& \textbf{.979}/\textcolor{red}{2.88}
&\textbf{.989}/\textcolor{red}{5.87}
&\textbf{.987}/\textbf{3.11}
&\textcolor{red}{.985}/\textcolor{red}{5.77}
& \textcolor{blue}{.972}/\textcolor{blue}{5.01}
&\textcolor{red}{.982}/\textcolor{red}{7.42} \\
$\Phi_{ex\ sub-dataset}$& -
& \textcolor{red}{.978}/\textbf{2.77}
&\textcolor{red}{.986}/\textcolor{blue}{7.37}
&\textcolor{red}{.984}/\textcolor{blue}{3.37}
&\textcolor{blue}{.982}/8.33
&\textbf{.975}/\textcolor{red}{4.92}
&.975/\textcolor{blue}{8.87} \\
\hline
\end{tabular}
\end{lrbox}
\scalebox{0.7}[0.7]{\usebox{\tablebox}}
\end{table}
\begin{figure}[t]
\centering
\subfigure[mean DC in Table~\ref{tab_subdataset}]{
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=2in]{Fig5_1.pdf}\\
\end{minipage}%
}%
\subfigure[mean HD in Table~\ref{tab_subdataset}]{
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=2in]{Fig5_2.pdf}\\
\end{minipage}%
}%
\centering
\caption{Heat map of DC \& HD results in Table~\ref{tab_subdataset}. The vertical axis represents different sub-datasets and the horizontal axis represents different models. In order to show the normal values more clearly, we clip some outliers to the boundary value, i.e., 0.95 in DC and 30 in HD. The values out of range are marked in the grid. The cross in the lower right corner indicates that there is no corresponding experiment.}
\vspace{-0.2cm}
\label{table2fig}
\end{figure}
\textbf{Generalization across sub-datasets.} Secondly, we train six deep networks, one network per single sub-dataset ($\Phi_{ABDOMEN}$, etc.). Then we test them on each sub-dataset.
Quantitative and qualitative results are shown in Table~\ref{tab_subdataset} and Fig.~\ref{fig-pic-6}, respectively. We also calculate the performance of $\Phi_{ALL}$ on each sub-dataset. For a fair comparison, cross-testing of sub-dataset networks is also conducted on each sub-dataset's testing set. We observe that the evaluation metrics of model $\Phi_{ALL}$ are generally better than those for the model trained on a single sub-dataset. These models trained on a single sub-dataset are difficult to consistently perform well in other domains, except $\Phi_{COLONOG}$, which contains the largest amount of data from various sources, originally. This observation implies that the domain gap problem does exist and the solution of collecting data directly from multi-source is effective. More intuitively, we show the `Average' values in heat map format in Fig.~\ref{table2fig}.
Furthermore, we implement \textit{leave-one-out} cross-validation of these six metal-free sub-datasets to verify the generalization ability of this solution. Models are marked as $\Phi_{ex\ ABDOMEN}$, etc. The results of $\Phi_{ex\ COLONOG}$ can fully explain that training with data from multi-sources can achieve good results on data that has not been seen before. When the models trained separately on the other five sub-datasets cannot achieve good results on COLONOG, aggregating these five sub-datasets can get a comparable result compared with $\Phi_{ALL}$, using only one third of the amount of data. More data from multi-sources can be seen as additional constraints on model learning, prompting the network to learn better feature representations of the pelvic bones and the background. In Fig.~\ref{fig-pic-6}, the above discussions can be seen intuitively through qualitative results.
\textbf{Others.} For more experimental results and discussions, e.g. `Generalization across manufacturers', `Limitations of the dataset', please refer to Online Resource 1\textsuperscript{\ref {em1}}.
\begin{figure}[t]
\includegraphics[width=\textwidth]{Fig4.pdf}
\caption{Visualization of segmentation results from six datasets tested on different models.
Among them, the white, green, blue and yellow parts of the segmentation results represent the sacrum, left hip bone, right hip bone and lumbar spine, respectively.
} \label{fig-pic-6}
\end{figure}
\subsubsection{SDF post-processor}\label{sdfpostprocessor}
The Exp (b) in Table~\ref{tab_basic} shows the effect of the post-processing module. SDF post-processor yields a decrease of 80.7\% and 15.1\% in HD compared with no post-processor and MCR post-processor. Details please refer to Online Resource 1\textsuperscript{\ref {em1}}.
The visual effects of two cases are displayed in Fig.~\ref{fig-pic-8}. Large fragments near the anatomical structure are kept with SDF post-processing but are removed by the MCR method.
\begin{figure}[t]
\includegraphics[width=\textwidth]{Fig6.pdf}
\caption{Comparison between post-processing methods: traditional MCR and the proposed SDF filtering.} \label{fig-pic-8}
\end{figure}
\section{Conclusion}
To benefit the pelvic surgery and diagnosis community, we curate and open source\textsuperscript{\ref{opensource}} a large-scale pelvic CT dataset pooled from multiple domains, including $1,184$ CT volumes (over 320K CT slices) of various appearance variations, and present a pelvic segmentation system based on deep learning, which, to the best of our knowledge, marks the first attempt in the literature. We train a multi-class network for segmentation of lumbar spine, sacrum, left hip, and right hip using the multiple-domain images to obtain more effective and robust features. SDF filtering further improves the robustness of the system. This system lays a solid foundation for our future work. We plan to test the significance of our system in real clinical practices, and explore more options based on our dataset, e.g. devising a module for metal-affected CTs and domain-independent pelvic bones segmentation algorithm.
\section*{Declarations}
\paragraph{Funding}
This research was supported in part by the Youth Innovation Promotion Association CAS (grant 2018135).
\paragraph{Conflict of interest}
The authors have no relevant financial or non-financial interests to disclose.
\paragraph{Availability of data and material} Please refer to URL\textsuperscript{\ref{opensource}}.
\paragraph{Code availability} Please refer to URL\textsuperscript{\ref{opensource}}.
\paragraph{Ethical approval} We have obtained the approval from the Ethics Committee of clinical hospital.
\paragraph{Informed consent} Not applicable.
| f703b7e99a552246ddc1c7af66961acfd91e30c7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
\label{sec:intro}
Polarimetry is a technique used in many areas of astronomy to study physical characteristics of a wide range of astronomical sources. In a number of researches, polarimetry plays a vital role in deriving interesting inferences which could not be revealed with any other technique. Understanding these advantages of the polarimetric techniques in astronomy, the Astronomy and Astrophysics Division of Physical Research Laboratory (PRL), India, has developed and been using an Optical Photo-polarimeter since the mid 1980s\cite{MRD1985BASI,UCJ1985MNRAS,UCJ1987AA,SG2009PRLTN}. In the mid 1990s an effort was initiated to build an experimental instrument for infrared photo-polarimetry\cite{UCJ1997BASI,Manian1999BASI} using a single element InSb detector. The lack of a sensitive detector effectively limited the use of this infrared instrument to bright sources. Later, with the availability of fast optical imagers with compact electronics, the opto-mechanical components were reused in a new instrument for optical imaging polarimetry by replacing the single element infrared detector Dewar with an EMCCD. This EMCCD based optical imaging polarimeter (EMPOL) has been in use since several years at the Mount Abu IR Observatory (MIRO). The polarization optics consists of a rotating half-wave plate (modulator) followed by a Foster prism (analyzer). The instrument has been used for various observations ranging such as some regions of the lunar surface, near-Earth asteroids, comets, open clusters, stars surrounding Lynds dark nebulae and active galactic nuclei such as blazars.
\noindent
Here we discuss the development of the instrument and present the details of the implementation with a few results from the characterisation observing runs. Many observations have been completed with this instrument and the results will be presented elsewhere. In the first section we list the science cases of interest for observations with EMPOL. In the next section the instrument development and implementation is described and the following section presents a few results from the calibration observations. We conclude with a summary.
\section{Science cases for EMPOL linear polarimetric observations}
Study of many astronomical sources would benefit from an examination of their linear polarimetric properties. Here we list a few of the science cases of interest for observations with EMPOL.
\subsection{Minor bodies of the solar system}
Comets and asteroids (including near-Earth asteroids)
exhibit strong linearly polarised light arising from the scattering of sunlight by dust present in them. Study of the polarisation properties as a function of wavelength helps to understand the characteristics of the dust grains.
\subsection{Open clusters}
The light from the stars get partially polarized after passing through the dusty interstellar medium. The polarization provides the information about the line of sight dust distribution and the physical properties of the dust grains. The open clusters are spread out over the sky so the evolution of the physical parameters of the dust can be analysed over the region. Linear Polarization is an excellent tool for determining the cluster membership. The procedure to identify memberships is based on the idea that the member stars are located behind common dust clouds which polarize the light, while most non-members maybe more randomly distributed in space and so their polarisation properties would be quite distinct.
\subsection{Dark molecular clouds}
Dark clouds in the Milky Way are places where new stars may be forming in deeply embedded regions. They are usually regions of dense molecular gas and dust. Observations of the linear polarisation of starlight passing through these dense clouds are used to trace the magnetic fields in the region. This can be used to trace the role of the magnetic fields in the star formation process.
\subsection{Active Galactic Nuclei}
Active galactic nuclei have supermassive black holes at their centres. These manifest their presence by emitting non-thermal highly polarised radiation. The polarisation is also found to be variable and the timescale of variability provides an indication of the size of the emitting regions. Oftentimes the energy emitted from an AGN emanates from a volume the size of the solar system and this energy is far larger than that from a whole normal galaxy. Thus optical polarisation is a very important property to be monitored for AGN in order to understand their energy generation mechanisms.
\section{Instrument design and implementation}
\noindent
The principle of the rapid modulation photo-polarimetric method\cite{Serkowski1974psns,FS1976ApOpt} is adopted in the current instrument. Normally, in instruments with rapid modulation requirement, photomultiplier tubes(PMT) are used as detectors in photon-counting mode. Rapid modulation allows to minimise or alleviate the atmospheric transmission variability affecting the polarisation measurement. In the case of integrating detectors like conventional CCDs, such a fast modulation rate is not possible. With the advent of electron multiplying CCD (EMCCD), near photon counting mode of detection has become available, with very high frame rates. The characteristics of the EMPOL instrument, further discussed in the following sections, are shown in Fig. \ref{fig:pars}.
\begin{figure} [hb]
\begin{center}
\includegraphics[width=0.4\textwidth]{parametertable.png}
\end{center}
\caption
{ \label{fig:pars}
Characteristics of EMPOL at a glance.}
\end{figure}
\subsection{Polarisation optics}
\begin{figure} [ht]
\begin{center}
\includegraphics[scale=0.7]{empol.png}
\end{center}
\caption
{ \label{fig:opticslayout}
Schematics of the optical layout for EMPOL.}
\end{figure}
Imaging (linear) polarimetry requires array detectors with images taken through an analyser at 3 or more orientations.
In our implementation of the linear imaging polarimeter, we employ a rotating half wave plate as modulator followed by a Foster prism as an analyser (see Fig. \ref{fig:opticslayout}).
\noindent
A Glan prism is used to generate a 100\% linearly polarised beam for calibration purpose. All the 3 polarisation optics are kept in the f/13 converging beam of the telescope just before the filter and EMCCD focal plane array detector. Fig. \ref{fig:empolontel} (left panel) shows the EMPOL instrument mounted at the Cassegrain focus of the 1.2 m telescope. The different parts of the instrument are shown in the annotated right panel of the same figure.
\begin{figure} [ht]
\begin{center}
\includegraphics[width=0.32\textwidth]{IMG_4355_Cropped.JPG}
\includegraphics[width=0.59\textwidth]{impol_annotated.png}
\end{center}
\caption
{ \label{fig:empolontel}
Left: PRL 1.2 m MIRO telescope with the imaging polarimeter mounted at the cassegrain focus. Right: Annotated photo of EMPOL instrument on the optical bench during development in the lab}
\end{figure}
\subsection{Modulation and acquisition trigger electronics}
The first version of the electronics used in this setup was shared with the optical photopolarimeter described\cite{SG2009PRLTN} earlier. A PC104 CPU board had been used for driving the stepper motor with the help of an additional 8254 timer/counter based PC104 board to generate the three square waves of different frequencies required for rotating the stepper motor and triggering the image acquisition. However, in a recent upgrade the entire PC104 stack was replaced with an Arduino board generating the square waves (shown in the right panel of Fig. \ref{fig:pcblayout}). We still retain the interface board (PCB layout in the left panel of Fig. \ref{fig:pcblayout}) to drive the stepper motor as well as to trigger the image acquisition. A 74LS164 serial-to-parallel shift register is used to generate the 8 square waves shifted by one clock pulse each. The clock pulse has a frequency 8 times the stepper pulse frequency. These 8 square waves are input to the ULN2803A current driver IC to drive the stepper motor.
\begin{figure} [ht]
\begin{center}
\includegraphics[width=0.38\textwidth]{nishtha_pcb_pol.png}
\hfill
\includegraphics[width=0.49\textwidth]{arduinoSqWaves.jpg}
\end{center}
\caption
{ \label{fig:pcblayout}
Left: PCB layout for the interface board used in EMPOL.
Right: Oscilloscope screenshot showing square waves of the three frequencies required to drive the 8-phase stepper motor.}
\end{figure}
\subsection{Zero point of position angle}
In order to define a reference position angle location for all observations, the image acquisition sequence is started at a fixed orientation of the half-wave plate. This position is sensed using an led + photo-resistor combination across the driving gear wheel fixed to the shaft of the stepper motor. A tiny hole in the disc of the gear wheel allows the light from the led to fall on the photo-resistor once per rotation. The resistance is monitored by a transistor amplifier and pulse shaping circuit (upper part of the interface PCB in the left panel of Fig. \ref{fig:pcblayout}). This 'index' pulse is provided as an input to the Andor SOLIS software via the auxiliary digital input available on the Andor PCI interface card which drives the EMCCD.
\subsection{EMCCD and image acquisition scheme}
EMCCD detectors have the possibility of generating several frames per second (based on the array size and on-chip binning employed).
The Andor iXon EMCCD used in EMPOL has 1K $\times$ 1K pixels of 13 $\mu$m each. With 4 $\times$ 4 binning over the full frame, we can achieve one frame per second with seeing limited nyquist sampling of the PSF. Images are taken in synchronisation with every step of the half-wave plate. The rotation of the half-wave plate is achieved with an 8-phase low power stepper motor. 48 steps complete one full rotation of the half-wave plate. The Andor EMCCD has several modes of image acquisition. In EMPOL, we use the external trigger mode for the image acquisition where the start of each image exposure is triggered externally using a square wave derived from the clock frequency described above. This trigger pulse is applied as input from the Arduino to the external trigger input pin available on the EMCCD camera.
\noindent
The AndorBASIC scripting facility available in the SOLIS software is used to complete the data acquisition. This script takes input from the user for the total effective exposure time required (per position of the half-wave plate). It then calculates the number of frames required to achieve the effective exposure time. The script monitors the auxiliary digital input for the 'index' pulse and then initiates the exposures once the 'index' pulse has been detected. Exposure times are set depending on the magnitude of the source being observed. For faint sources, the exposures are built up by stacking the images taken at the corresponding positions of the half-wave plate to obtain the required effective exposure time. Since this is currently completed offline, the observing procedure is quite data storage intensive.
\noindent
A CFW-12 filter wheel with 12 slots, manufactured by Finger Lakes Instruments, is used in the light path after the Foster prism and just before the EMCCD detector. This allows to use standard UBVRI or Sloan filters for the polarimetric measurements. We also have the Hale-Bopp comet set of narrow-band filters\cite{Farnham2000Icarus} for use with this instrument.
\section{Observations of astronomical targets}
\subsection{Standard star observations}
Polarisation standard stars are observed to calibrate the zero point of the polarisation position angle and to confirm proper operation of the instrument. Fig. \ref{fig:9gem} (Left panel) shows the modulation curve obtained for polarisation standard 9 Gem, with 3\% polarisation.
\begin{figure} [ht]
\begin{center}
\includegraphics[width=0.49\textwidth]{9gem.png}
\includegraphics[width=0.49\textwidth]{100p.png} \end{center}
\caption
{ \label{fig:9gem}
Left: Modulation curve for polarisation standard 9 Gem.
Right: Modulation curve for 100\% polarisation observed with Glan prism.}
\end{figure}
\noindent
To confirm that the instrument is behaving linearly over the entire range of modulation we also observe a star with a Glan prism in the light path. The resulting polarisation due to the presence of the Glan is 100\%. The observed modulation curve is shown in the right panel of Fig. \ref{fig:9gem}. The value derived is 97\% showing that the instrument is working in an acceptable manner.
\subsection{Blazar OJ 287}
\begin{figure} [ht]
\begin{center}
\includegraphics[width=\textwidth]{oj287_p_vs_t.png}
\end{center}
\caption
{ \label{fig:oj287}
Polarisation variability in blazar OJ 287. Open circles are for OJ 287 while the blue diamonds represent a comparison star nearby.}
\end{figure}
\noindent
The well known blazar OJ 287 shows rapid variability in flux as well as polarisation within a night with very strong variations in linear polarisation. This blazar was observed with EMPOL for linear polarisation over nearly 2 hours of observations\cite{SG2014ATel}. During the observations presented here (Fig. \ref{fig:oj287}), the polarisation has decreased from nearly 6\% to 3\% within two hours, while the observations of a comparison star in the same field remained stable below 1\% level.
\section{Summary}
\begin{itemize}
\item An EMCCD based imaging polarimeter developed at PRL has been in use at the 1.2m telescope for several years with a variety of objects being observed.
\item Linear polarisation is measured using a rotating half-wave plate as modulator and a Foster prism as analyser.
\item Polarisation measurement is linear over the entire range as demonstrated by the observations using a Glan prism to measure 100\% polarisation.
\item A 12 position filter wheel is used to obtain polarimetry in the standard UBVRI or Sloan filter set as well as with the Hale-Bopp comet filter set.
\end{itemize}
\section*{Acknowledgements}
We acknowledge the contribution of many summer trainees and project students in the various stages of the development of this instrument. Work at PRL is funded by the Dept. of Space, Govt. of India.
| 9182726a45252064e3986fc488fe60b240af2305 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sc:introduction}
The stylish Chinese font generation has attracted rising attention within recent years \cite{Lin2016,Cha2020,Chang2017,Tian2017,Kong2017,Jiang2017,Jiang2019,Chang2018,chen2019,Wu2020,Gao2020,Zhang2020}, since it has a wide range of applications including but not limited to the automatic generation of artistic Chinese calligraphy \cite{Zhao2020}, art font design \cite{Lin2014} and personalized style generation of Chinese characters \cite{Liu2012}.
The existing Chinese font generation methods can be generally divided into two categories. The first category is firstly to extract some explicit features such as strokes and radicals of Chinese characters and then utilize some traditional machine learning methods to generate new characters \cite{Xu2005,Lin2016}. The quality of feature extraction plays a central role in the first category of Chinese font generation methods. However, such feature extraction procedure is usually hand-crafted, and thus time and effort consuming.
The second category of Chinese font generation methods has been recently studied in \cite{Tian2017,Chang2018,chen2019,Gao2020,Wu2020,Zhang2020} with the development of deep learning \cite{Goodfellow2016}, particularly the generative adversarial networks (GAN) \cite{Goodfellow2014}. Due to the powerful expressivity and approximation ability of deep neural networks, feature extraction and generation procedures can be combined into one procedure, and thus, Chinese font generation methods in the second category can be usually realized in an end-to-end training way. Instead of using the stroke or radical features of Chinese characters, the methods in the second category usually regard Chinese characters directly as images, and then translate the Chinese font generation problem into certain image style translation problem \cite{Zhu2017,Isola2017}, for which GAN and its variants are the principal techniques. However, it is well-known that GAN usually suffers from the issue of mode collapse \cite{Goodfellow2014}, that is, producing the same patterns for different inputs by generator. Such issue will significantly degrade the diversity and quality of the generated results (see, Figure \ref{fig:mode-collapse} below).
When adapted to Chinese font generation problem, the mode collapse issue will happen more frequently due to there are many Chinese characters with very similar strokes.
\begin{figure}[ht]
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.25]{mode-cllopse.png}
\end{minipage}
\hfill
\caption{CycleGAN for Chinese font generation suggested in \cite{Chang2018} suffers from the mode collapse issue when translating \textit{Black} font to \textit{Shu} font, while the suggested StrokeGAN can effectively tackle this issue. The settings are presented in the later experiment section.}
\label{fig:mode-collapse}
\end{figure}
Due to the artificial nature of Chinese characters, the explicit stroke information contains amount of mode information of Chinese characters (see Figure \ref{fig:32 basic strokes}). This is very different from natural images, which are usually regarded to be generated according to some probability distributions at some latent spaces.
Inspired by this observation, in this paper, we at first introduce a one-bit stroke encoding to preserve the key mode information of a Chinese character, and then suggest certain stroke-encoding reconstruction loss to reconstruct the stroke encoding of the generated character such that the key mode information can be well preserved, and finally incorporate them into CycleGAN for Chinese font generation \cite{Chang2018}. Thus, our suggested model is called \textit{StrokeGAN}.
The contributions of this paper can be summarized as follows:
\begin{enumerate}
\item[(a)]
We propose an effective method called \textit{StrokeGAN} for the generation of Chinese fonts with unpaired data. Our main idea is firstly to introduce a one-bit stroke encoding to capture the mode information of Chinese characters and then incorporate it into the training of CycleGAN \cite{Chang2018}, in the purpose of alleviating the mode collapse issue of CycleGAN and thus improving the diversity of its generated characters.
In order to preserve the stroke encoding, we introduce a stroke-encoding reconstruction loss to the training of CycleGAN. By the use of such one-bit stroke encoding and the associated reconstruction loss, StrokeGAN can effectively alleviate the mode collapse issue for Chinese font generation, as shown by Figure \ref{fig:mode-collapse}.
\item[(b)]
The effectiveness of StrokeGAN is verified over a set of Chinese character datasets with 9 different fonts (see Figure \ref{fig:stgan-results}),
that is, a handwriting font, 3 standard printing fonts and 5 pseudo-handwriting fonts. Compared to CycleGAN for Chinese font generation \cite{Chang2018}, StrokeGAN can generate Chinese characters with higher quality and better diversity, particularly, the strokes are better preserved. Besides CycleGAN \cite{Chang2018}, our method also outperforms the other state-of-the-art methods including zi2zi \cite{Tian2017} and Chinese typography transfer (CTT) method \cite{Chang2017} using paired data, in terms of generating Chinese characters with higher quality and accuracy. Some generated characters by our method with 9 different fonts can be found in Figure \ref{fig:stgan-results}. It can be observed that these generated characters of StrokeGAN are very realistic.
\end{enumerate}
\begin{figure}[ht]
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.35]{32-basic-strokes.png}
\centerline{{\small (a) 32 basic strokes that make up Chinese characters}}
\end{minipage}
\hfill
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.35]{yi-character.png}
\centerline{{\small (b) Strokes of the Chinese character "Yi"}}
\end{minipage}
\hfill
\caption{(a) 32 basic strokes that make up Chinese characters. The first and fourth columns present these 32 basic strokes, the second and fifth columns present the names of these basic strokes, and the third and sixth columns present some typical Chinese characters involved these basic strokes. (b) The strokes of the Chinese character "Yi".
}
\label{fig:32 basic strokes}
\end{figure}
\begin{figure}[ht]
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.35]{results.png}
\end{minipage}
\hfill
\caption{Some Chinese characters with 9 different fonts generated by StrokeGAN. The first row presents the generated Chinese characters with the handwriting font, the second to the fourth rows present the generated Chinese characters with three different standard printing fonts, and the fifth to the ninth rows present the generated Chinese characters with five different pseudo-handwriting fonts, which are very different from the standard printing fonts and in general embody more personalized fonts.
}
\label{fig:stgan-results}
\end{figure}
\subsection{Related work}
In recent years, many generation methods of stylish Chinese fonts have been suggested in the literature \cite{Tian2017,Chang2017,Chang2018,chen2019,Wu2020,Jiang2017,Jiang2019,Zhang2020} with the development of deep learning. In \cite{Tian2017}, the authors adapted \textit{pix2pix} model developed in \cite{Isola2017} for the image style translation problem to Chinese font generation and then suggested \textit{zi2zi} method with paired training data, that is, there is a one-to-one correspondence between the characters in the source (input) style domain and target (output) style domain. Similar idea was extended to realize the Chinese character generation from one font to multiple fonts in \cite{chen2019}. Besides \cite{Tian2017} and \cite{chen2019}, some other paired data based Chinese font generation methods were suggested in \cite{Chang2017,Jiang2017,Wu2020}. However, it is usually human-intensive to build up the paired training data. In order to overcome this challenge, \cite{Chang2018} adapted CycleGAN developed in \cite{Zhu2017} for the image style translation to Chinese font generation based on unpaired training data. Yet, the CycleGAN based method suggested in \cite{Chang2018} (called \textit{CCG-CycleGAN}) may suffer from the mode collapse issue \cite{Goodfellow2014}. When mode collapse occurs, the generator produces fewer patterns for different inputs, and thus significantly degrades the diversity and quality of generated results.
Motivated by the observation from traditional Chinese character generation and recognition methods (see \cite{Xu2005,Kim-stroke1999}) that the explicit stroke feature can provide much mode information for a Chinese character, in this paper, we incorporate such stroke information into the training of CycleGAN for Chinese font generation \cite{Chang2018} to tackle the issue of mode collapse, via introducing a one-bit stroke encoding and certain stroke-encoding reconstruction loss.
The very recent papers \cite{Wu2020,Jiang2019,Zhang2020} also incorporated some stroke or radical information of Chinese characters into Chinese font generation. Their main idea is firstly to utilize a deep neural network to extract the strokes or radicals of Chinese characters and then merge them by another deep neural network, which is very different to our idea of the use of a very simple one-bit stroke encoding. According to our later numerical experiments, our introduced one-bit stroke encoding is very effective.
The rest of this paper is organized as follows. In Section 2\ref{sc:preliminary work}, we present some preliminary work. In Section 3\ref{sc:Stroke GAN}, we introduce the proposed method in detail. In Section 4\ref{sc:experiments}, we provide a series of experiments to demonstrate the effectiveness of the proposed method. We conclude this paper in Section 5\ref{sc:conclusion}.
\begin{figure*}[ht]
\begin{minipage}[b]{1\linewidth}
\centering
\includegraphics*[scale=0.43]{network.png}
\end{minipage}
\hfill
\caption{The diagram of StrokeGAN for Chinese font generation. There are a discriminator $D$ and a generator $G$ in StrokeGAN. The workflow of StrokeGAN can be divided into three parts: (a) we at first yield the stroke encodings of Chinese characters via certain one-bit encoding way, and then send the characters to $G$ as input; (b) $G$ at first generates a fake character according to the input character from the source font domain, and later reconstructs a character in the source font domain from the generated fake character in the target font domain such that the reconstructed character is as similar as possible to the original input character; (c) $D$ takes the fake character generated by $G$ and the real character in the source font domain as the input, and attempts to on one hand distinguish whether the generated character is real or fake and on the other hand reconstruct the stroke encoding of the generated character as accurately as possible, compare with the real stroke encoding to preserve the mode information of Chinese characters. The network architectures of generator and discriminator are stacked by the modules presented in the right column.
}
\label{fig:stgan-model}
\end{figure*}
\section{Preliminary Work}
\label{sc:preliminary work}
In this section, we introduce some preliminary work, which serves as the basis of this paper.
\subsection{Generative Adversarial Networks}
\label{sc:Generative adversarial networks}
Generative adversarial networks (GAN) \cite{Goodfellow2014} have achieved great achievements in the task of high-quality image synthesis. The classic GAN model consists of two parts: a generator $G$ and a discriminator $D$. The generator $G$ generates fake images, and the discriminator $D$ judges whether the images generated by $G$ are fake or real. Mathematically, GAN can be formulated as the following two-player minimax game between $G$ and $D$,
\begin{align*}
\min_G \max_D \ \mathbb{E}_{x\sim \mathbb{P}_{data}} [\log D(x)] + \mathbb{E}_{z\sim \mathbb{P}_{z}} [\log(1-D(G(z)))]
\end{align*}
where $\mathbb{P}_{data}$ and $\mathbb{P}_{z}$ are the distributions of data $x$ and the input noise variable $z$ for the generator, respectively. In practice, the generator $G$ and discriminator $D$ are generally represented by some deep neural networks.
\subsection{Conditional GAN}
\label{sc:Conditional-GAN-(cGAN)-and-Cycle-GAN-(CycleGAN)}
The conditional GAN (cGAN) was suggested in \cite{Mirza2014} mainly to embed some conditional information such as the category information of samples. Such idea was later exploited in \cite{Isola2017} for the image style translation. For cGAN, besides the original input $z$, the conditional information $c$ is also a part of input for $G$. Thus, the objective for GAN should be slightly modified for cGAN, shown as follows:
\begin{align*}
{\cal L}_{adv}(D,G)
&=\mathbb{E}_{x\sim \mathbb{P}_{data}} [\log D(x)] \\
&+ \mathbb{E}_{z\sim \mathbb{P}_{z}, c\sim \mathbb{P}_c} [\log(1-D(G(z,c)))], \nonumber
\end{align*}
where $\mathbb{P}_c$ represents the distribution of the referred conditional information $c$.
\subsection{CycleGAN}
The training of cGAN \cite{Isola2017} is based on the paired data, of which the collection is usually time-consuming and laborious. In order to overcome this challenge, CycleGAN was proposed in \cite{Zhu2017} for the image style translation based on the unpaired data. The main idea of CycleGAN is to preserve key attributes between the source and target domains by utilizing a cycle consistency loss. Specifically, let $x\in {\cal X}$ and $x'\in {\cal X}'$ be two images from two different style domains conditioned on respectively some conditional information domains $c\in {\cal C}$ and $c' \in {\cal C}'$. In order to realize the bidirectional translation between them, two generators $G:{\cal X} \times {\cal C} \rightarrow {\cal X}'$ and $G':{\cal X}' \times {\cal C}' \rightarrow {\cal X}$ are exploited. With these, the cycle consistency loss can be defined as follows:
\begin{align*}
{\cal L}_{cyc}(G,G')
&= \mathbb{E}_{x\sim \mathbb{P}_x, c \sim \mathbb{P}_c, c' \sim \mathbb{P}_{c'}}[\|x-G'(G(x,c),c')\|_1]\\
&+\mathbb{E}_{x'\sim \mathbb{P}_{x'}, c' \sim \mathbb{P}_{c'}, c \sim \mathbb{P}_c}[\|x'-G(G'(x',c'),c)\|_1],
\end{align*}
where $x\in {\cal X}$, $x'\in {\cal X}'$, $c\in {\cal C}$, $c'\in {\cal C}'$, and $\mathbb{P}_*$ represents the distribution of the associated domain $*$. The generators $G$ and $G'$ are trained to make the cycle consistency loss ${\cal L}_{cyc}(G,G')$ small.
\section{StrokeGAN for Chinese Font Generation}
\label{sc:Stroke GAN}
In this section, we describe the proposed \textit{StrokeGAN} for automatic generation of stylish Chinese font. The core idea of StrokeGAN is to incorporate some one-bit stroke encodings of Chinese characters into CycleGAN to alleviate the issue of mode collapse, as presented in Figure \ref{fig:stgan-model}, mainly motivated by the basic observation that the stroke information embodies amount of mode information of Chinese characters. Recent studies suggested that mode collapse in GANs might be a by-product of removing sparse outlying modes toward robust estimation \cite{Yao2019iclr,Yao2020jmlr}. So a natural strategy to alleviate mode collapse is to enforce faithful conditional distribution reconstruction, conditioning on important modes represented by stroke encoding here. From Figure \ref{fig:stgan-model}, we at first yield the stroke encodings of Chinese characters via certain one-bit encoding way, and then take the characters in the source font domain as the inputs of generator $G$. After $G$, we yield a fake character in the target domain, then send such fake character to both generator and discriminator, where generator tries to reconstruct the character in the source font domain, and discriminator attempts to distinguish whether the generated character is real or fake and also reconstruct its stroke encoding. Thus, distinguished with the original CycleGAN for Chinese font generation in \cite{Chang2018}, there are two parts in the discriminator $D$, i.e.,
\[
D : x \rightarrow (D_{src}(x), D_{st}(x)),
\]
where $D_{src}(x)$ and $D_{st}(x)$ represent respectively the probability distribution over the source font domain and its stroke encoding for a given character $x$.
\subsection{Stroke Encoding}
\label{sc:structure loss}
From Figure \ref{fig:stgan-model}, the stroke encoding of character is taken as a part of input of CycleGAN. To realize this, we introduce a simple one-bit encoding way to yield the associated stroke encoding for a given Chinese character. Specifically, according to Figure \ref{fig:32 basic strokes}, there are in total 32 kinds of strokes to make up Chinese characters. Thus, for any given Chinese character $x$, we define its stroke encoding as a 32-dimensional vector $c \in \{0,1\}^{32}$ with the $i$-th entry being $1$ if the $i$-th kind of stroke is included in $x$ and otherwise $0$ for all $i$ from $1$ to $32$. In this paper, we only use the indicator function of such kind of stroke instead of its exact number for a given Chinese character mainly in consideration of the robustness of StrokeGAN. This is in general sufficient according to our later experiments (see, Figure \ref{fig:comp-similar-character}).
\subsection{Training Loss for StrokGAN}
\label{sc:Structure GAN for CCG}
The training loss for StrokeGAN consists of three parts, that is, the general adversarial loss, cycle consistency loss and stroke-encoding reconstruction loss, where the \textit{stroke-encoding reconstruction loss} is firstly introduced in this paper for the generation of stylish Chinese fonts.
\begin{table*}
\caption{The performance of StrokeGAN in 9 generation tasks via comparing with CycleGAN \cite{Chang2018}. }
\label{tab:result-number}
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|}\hline
\multirow{2}*{Character style translation} &
\multicolumn{2}{c|}{Content accuracy (\%) $\uparrow$} & \multicolumn{2}{c|}{Recognition accuracy (\%) $\uparrow$} & \multicolumn{2}{c|}{Stroke error ($\times 10^{-2}$) $\downarrow$} \\
\cline{2-3} \cline{4-5} \cline{6-7}
& CycleGAN & StrokeGAN & CycleGAN & StrokeGAN & CycleGAN & StrokeGAN\\ \hline
\textit{Regular Script} $\rightarrow$ \textit{Shu} & 89.56 & \textbf{90.48} & 89.36 & \textbf{90.52} &6.79 & \textbf{5.69}\\\hline
\textit{Regular Script} $\rightarrow$ \textit{Huawen Amber} & 86.88 & \textbf{88.68} & 87.56 & \textbf{88.92} &8.71 &\textbf{7.20}\\\hline
\textit{Regular Script} $\rightarrow$ \textit{Hanyi Lingbo} & 87.64 & \textbf{88.24} & 87.48 & \textbf{88.32} &7.90 &\textbf{6.81}\\\hline
\textit{Regular Script} $\rightarrow$ \textit{Imitated Song}& 90.28 & \textbf{91.72} & 90.84 & \textbf{91.60} &7.33 &\textbf{5.55} \\\hline
\textit{Regular Script} $\rightarrow$ \textit{Handwriting} & 87.08 & \textbf{87.64} & 87.12 & \textbf{87.60} &7.67 &\textbf{6.50}\\\hline
\textit{Hanyi Thin Round} $\rightarrow$ \textit{Hanyi Doll} & 87.00 & \textbf{87.60} & 86.92 & \textbf{87.80} &7.69 &\textbf{6.52} \\\hline
\textit{Black} $\rightarrow$ \textit{Hanyi Thin Round} & 86.60 & \textbf{87.76} & 86.60 & \textbf{87.92} &7.79 &\textbf{6.96}\\\hline
\textit{Imitated Song} $\rightarrow$ \textit{Black} & \textbf{89.56}& 88.96 & \textbf{89.44} & 89.12 &8.06 & \textbf{6.72} \\\hline
\textit{Imitated Song} $\rightarrow$ \textit{Regular Script} & 89.96 & \textbf{90.32} & 90.04 & \textbf{90.28} &8.49 &\textbf{6.28} \\
\hline
\end{tabular}
\end{center}
\end{table*}
{\bf A. Adversarial loss.} The first part of loss is the adversarial loss defined commonly as follows,
\begin{align}
\label{Eq:ad-loss}
{\cal L}_{adv}(D,G)
&= \mathbb{E}_{x}[\log D_{src}(x)] \\
&+ \mathbb{E}_{x}[\log(1-D_{src}(G(x)))], \nonumber
\end{align}
where the generator $G$ generates the fake character $G(x)$ conditional over the input character $x$, and the first part of discriminator $D_{src}$ attempts to distinguish the generated character is real or fake.
{\bf B. Cycle consistency loss.}
The second part of loss is the cycle consistency loss, which is introduced to let the generator reconstruct the character in the source font domain from the generated fake character, and thus avoid using the paired data. Specifically, such part of loss can be defined as follows:
\begin{align}
\label{Eq:cycle-consistency-loss}
{\cal L}_{cyc}(G)
=\mathbb{E}_{x}[\|x-G(G(x))\|_1],
\end{align}
where $G(x)$ represents the generated fake character according to the input pair $x$ and $G(G(x))$ represents the reconstructed character from the character $G(x)$ generated by $G$.
{\bf C. Stroke-encoding reconstruction loss.}
Notice that in both adversarial loss and cycle consistency loss, the characters are regarded as images during the training, while the stroke information is not paid much attention. Actually, as discussed before, the stroke information embodies amount of mode information of a Chinese character, and thus is some kind of very important information that should be preserved during the training. Also, the stroke information is very special for Chinese characters and makes the generation of Chinese characters very different from the style translation of natural images. Specifically, the \textit{stroke-encoding reconstruction loss} can be defined as follows:
\begin{align}
\label{Eq:structural-information-loss}
{\cal L}_{st}(D) = \mathbb{E}_{x,c} \left[{\|D_{st}(G(x))-c\|_2} \right].
\end{align}
where $D_{st}(G(x))$ is the stroke encoding yielded by discriminator for the generated fake character $G(x)$. Such stroke-encoding reconstruction loss is used to guide the networks to reconstruct the stroke encodings as accurately as possible so that the modes of characters can be preserved much better.
{\bf D. Total training loss.}
Combining the above three parts of loss, i.e., \eqref{Eq:ad-loss}-\eqref{Eq:structural-information-loss}, the total training loss of StrokeGAN is presented as follows:
\begin{align}
\label{Eq:total-loss}
{\cal L}_{strokegan}(D,G)
&= {\cal L}_{adv}(D,G) +\lambda_{cyc} {\cal L}_{cyc}(G) \\
&+ \lambda_{st}{\cal L}_{st}(D),\nonumber
\end{align}
where $\lambda_{cyc}$ and $\lambda_{st}$ are two penalty parameters. Based on the above defined loss ${\cal L}_{strokegan}(D,G)$, the discriminator $D$ attempts to maximize it while the generator $G$ tries to minimize it, shown as follows:
\begin{align}
\label{Eq:stgan-training}
\min_G \max_D \ {\cal L}_{strokegan}(D,G).
\end{align}
\section{Numerical Experiments}
\label{sc:experiments}
In this section, we provide a series of experiments to demonstrate the effectiveness of the suggested StrokeGAN. All experiments were carried out in Pytorch environment running Linux, AMD(R) Ryzen 7 2700x eight-core processor $\times16$ CPU, GeForce RTX 2080 GPU. Our codes are available in \url{https://github.com/JinshanZeng/StrokeGAN}.
\subsection{Experiment Settings}
\label{sc:exp-settings}
\begin{figure}[ht]
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.24]{results-4style.png}
\end{minipage}
\hfill
\caption{Examples of Chinese characters generated by StrokeGAN. The characters in the left-hand side are real characters in the target font domains, while the characters in the right-hand side are generated characters in the associated target font domains. From the first row to the fourth row, the target fonts are $\{$\textit{Imitated Song}, \textit{Shu}, \textit{Hanyi Lingbo}, \textit{Huawen Amber}$\}$, respectively. It can be observed that the characters generated by StrokeGAN are very realistic.
}
\label{fig:results-4styles}
\end{figure}
\begin{figure}[ht]
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.21]{stGAN-cycleGAN.png}
\end{minipage}
\hfill
\caption{Stoke-missing phenomena in the generated characters of CycleGAN \cite{Chang2018}. The suggested StrokeGAN can preserve strokes much better. The first two columns, the third and fourth columns, and the last two columns respectively present the generated characters from \textit{Regular Script} to \textit{Shu}, from \textit{Regular Script} to \textit{Huawen Amber}, and from \textit{Hanyi Thin Round} to \textit{Hanyi Doll}.
}
\label{fig:st-cycle}
\end{figure}
{\bf A. Collection of datasets.}
The dataset used in this paper consists of 9 sub-datasets with different fonts divided into three categories, i.e., a handwriting font, three standard printing fonts $\{$\textit{Black}, \textit{Regular Script}, \textit{Imitated Song}$\}$, and five pseudo-handwriting fonts $\{$\textit{Huawen Amber}, \textit{Shu}, \textit{Hanyi Lingbo}, \textit{Hanyi Doll}, \textit{Hanyi Thin Round}$\}$, where the pseudo-handwriting fonts are personalized fonts designed by artistic font designers.
The first kind of dataset related to the handwriting Chinese characters is built up from \textit{CASIA-HWDB1.1} \footnote{\url{http://www.nlpr.ia.ac.cn/databases/handwriting/Home.html}}, which was collected by 300 people.
There are in total 3755 different commonly used Chinese characters written by everyone, and thus there are $3755 \times 300$ handwriting characters in this dataset.
To build up our handwriting dataset, for each character, we randomly selected one sample from these 300 samples. Therefore, the size of the handwriting font dataset used in this paper is 3755. Except the handwriting Chinese characters, the other font datasets were collected by ourselves from the internet \footnote{say, \url{http://fonts.mobanwang.com/}} and made automatically via TTF. Specifically, the second kind of datasets contain three printing font datasets, of which the sizes are 2560, 3757 and 2506 respectively for the \textit{Black}, \textit{Regular Script} and \textit{Imitated Song} fonts. The third kind of datasets consist of five pseudo-handwriting fonts, of which the sizes are 3596, 2595, 3673, 3213 and 2840 respectively for the \textit{Huawen Amber}, \textit{Shu}, \textit{Hanyi Lingbo}, \textit{Hanyi Doll} and \textit{Hanyi Thin Round} fonts. The size of each character was resized to $128 \times 128 \times 3$. In our experiments, we used 90\% and 10\% of the samples respectively as the training and test sets.
{\bf B. Network architectures and optimizer.}
The network structure of the generator in StrokeGAN is the same as CycleGAN \cite{Zhu2017}, including 2 convolutional layers in the down-sampling module, 9 residual modules with 2 convolutional layers of residual networks for each residual module and 2 deconvolutional layers in the up-sampling module, as presented in Table \ref{tab:generator-structure} in Appendix. The network structure of the discriminator in StrokeGAN is similar to PatchGAN \cite{Isola2017} with 6 hidden convolutional layers and 2 convolutional layers in the output module, as presented in Table \ref{tab:discriminator-structure} in Appendix. Moreover, the batch normalization \cite{Ioffe2015} was used in all layers.
In our experiments, we used the popular Adam algorithm \cite{Kingma2014} as the optimizer with the associated parameters $(0.5, 0.999)$ in both the generator and discriminator optimization subproblems. The penalty parameters of the cycle consistency loss and stroke reconstruction loss were fine-tuned at 10 and 0.18, respectively.
{\bf C. Evaluation metrics.}
To evaluate the performance of StrokeGAN, we introduced three evaluation metrics. The first one is the commonly used \textit{content accuracy} \cite{Zhu2017}, which was suggested to justify the quality of the contents of generated characters. Specifically, a pre-trained HCCG-GoogLeNet \cite{Szegedy2014} was exploited to calculate the content accuracy. Besides the \textit{content accuracy}, we also suggested the \textit{recognition accuracy} and \textit{stroke error} particularly for the Chinese character generation task. The \textit{recognition accuracy} is defined as the ratio of those generated characters that can be correctly recognized by people to all the generated characters for testing, via a crowdsourcing way. Specifically, we randomly invited five Chinese adults to recognize the generated Chinese characters, and then took their average as the \textit{recognition accuracy}. The \textit{stroke error} is defined as the ratio of the number of missing and redundant strokes in the generated Chinese characters to its true total number of strokes. Thus, the smaller stroke error means the strokes are preserved better.
\begin{figure}[ht]
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.27]{comp-similar-character.png}
\end{minipage}
\hfill
\caption{Generated results of StrokeGAN for some very similar Chinese characters with the same stroke encodings.}
\label{fig:comp-similar-character}
\end{figure}
\begin{figure*}[ht]
\begin{minipage}[b]{0.99\linewidth}
\centering
\includegraphics*[scale=0.45]{four-result.png}
\end{minipage}
\hfill
\caption{Some examples of Chinese characters generated by StrokeGAN and three existing methods in the generation task from \textit{Regular Script} font to \textit{Shu} font.}
\label{fig:four-model-result}
\end{figure*}
\subsection{Experiment Results}
\label{sc:exp-results}
Our experiments consist of three parts, where the first two parts were conducted to respectively show the effectiveness and sufficiency of the introduced one-bit stroke encoding,
and the third part was conducted to demonstrate the effectiveness of StrokeGAN via comparing with the state-of-the-art methods including CycleGAN \cite{Chang2018}, zi2zi \cite{Tian2017} and Chinese typography transfer (CTT) method \cite{Chang2017}, where the latter two methods are based on paired training data.
{\bf A. Effectiveness of one-bit stroke encoding.}
In these experiments, we verified the feasibility and effectiveness of our idea that incorporating the stroke information into CycleGAN can preserve the modes of Chinese characters better and thus alleviate the issue of mode collapse. In order to verify this, we implemented nine one-to-one Chinese character style translation experiments as listed in Table \ref{tab:result-number}. For each experiment, we selected one font domain (say \textit{Regular Script}) from these nine different font styles as the source font domain and another font domain (say \textit{Shu}) as the target font domain, and then trained StrokeGAN as well as CycleGAN as the baseline. The performance of the suggested StrokeGAN is presented in Table \ref{tab:result-number}. Some examples of the generated Chinese characters by StrokeGAN are shown in the previous Figure \ref{fig:stgan-results} and Figure \ref{fig:results-4styles}, which demonstrate that StrokeGAN can generate very realistic Chinese characters for all these nine fonts.
From Table \ref{tab:result-number}, StrokeGAN outperforms CycleGAN (without using the stroke encoding) in most of the tasks except the style translation task from \textit{Imitated Song} to \textit{Regular Script}. In terms of the stroke error presented in the last two columns, StrokeGAN consistently improves the performance of CycleGAN \cite{Chang2018}. This shows the feasibility and effectiveness of the suggested StrokeGAN.
Moreover, as demonstrated in the previous Figure \ref{fig:mode-collapse}, CycleGAN sometimes suffers from the issue of mode collapse (particularly, when applied to the generation task from \textit{Black} font to \textit{Shu} font), while the suggested StrokeGAN can significantly alleviate this issue. Besides the issue of mode collapse, as shown in Figure \ref{fig:st-cycle}, CycleGAN also sometimes suffers from another issue, that is, missing some key strokes of Chinese characters generated, which may result in the recognition issues of these characters, while the suggested StrokeGAN can preserve these strokes much better.
{\bf B. Sufficiency of one-bit stroke encoding.} Notice that there are many very similar Chinese characters having the same stroke encodings. In order to show the sufficiency of the introduced one-bit encoding, we tested the performance of StrokeGAN on some very similar Chinese characters,
as shown in Figure \ref{fig:comp-similar-character},
which shows that the generated characters can be well distinguished with well preserved strokes.
This demonstrates that such one-bit encoding way is in general enough to preserve the key modes of Chinese characters.
{\bf C. Comparison with the state-of-the-art methods.}
In this part of experiments, besides CycleGAN \cite{Chang2018}, we compared StrokeGAN with two paired data based methods, i.e., \textit{zi2zi} \cite{Tian2017} and CTT method \cite{Chang2017}.
In order to implement both zi2zi and CTT methods, we manually built up two paired datasets based on the \textit{Regular Script} font and \textit{Shu} font datasets. Since the collection of paired training data is very costly, in these experiments, we only considered the generation task from the \textit{Regular Script} font to \textit{Shu} font for all these four methods, while for the other generation tasks, it can be implemented similarly. The performance of these four methods is presented in Table \ref{4-model-result}, and some examples of generated characters are shown in Figure \ref{fig:four-model-result}. From Table \ref{4-model-result}, the suggested StrokeGAN outperforms all these existing methods in terms of the suggested three evaluation metrics, and also generates the Chinese characters with the highest quality among all these methods. These experiment results demonstrate the effectiveness of the proposed StrokeGAN.
\begin{table}
\caption{Comparison on the performance of the suggested StrokeGAN with existing methods in the generation task from \textit{Regular Script} font to \textit{Shu} font. The first, second and third rows respectively present the content accuracies (\%), recognition accuracies (\%) and stroke errors in the scale $\times 10^{-2}$ of all these four methods.}
\label{4-model-result}
\begin{center}
\footnotesize
\begin{tabular}{|c|c|c|c|c|}\hline
Methods & StrokeGAN & CycleGAN & zi2zi & CTT \\\hline
Content acc. ($\uparrow$) & \textbf{90.48} & 89.56 & 90.12 & 87.92 \\\hline
Recog. acc. ($\uparrow$) & \textbf{90.52} & 89.36 & 90.04 & 88.01 \\\hline
Stroke error ($\downarrow$) & \textbf{5.67} & 6.79 & 6.36 & 7.52 \\\hline
\end{tabular}
\end{center}
\end{table}
\section{Conclusion}
\label{sc:conclusion}
This paper proposes an effective Chinese font generation method called \textit{StrokeGAN} by incorporating a one-bit stroke encoding into CycleGAN
to tackle the mode collapse issue. The key intuition of our idea is that the stroke encodings of Chinese characters contain amount of mode information of Chinese characters, unlike the natural images. A new stroke-encoding reconstruction loss was introduced
to enforce a faithful reconstruction of the stroke encoding as accurately as possible and thus preserve the mode information of Chinese characters. Besides the commonly used content accuracy, the crowdsourcing recognition accuracy and stroke error are also introduced to evaluate the performance of our method. The effectiveness of StrokeGAN is demonstrated by a series of Chinese font generation tasks over 9 datasets with different fonts, comparing with CycleGAN and other two existing methods based on the paired data. The experiment results show that StrokeGAN helps preserve the stroke modes of Chinese characters in a better way and generates very realistic characters with higher quality.
Besides Chinese font generation, our idea of the one-bit stroke encoding can be easily adapted to other deep generative models and applied to the font generation related to other languages such as Korean and Japanese.
\section*{Acknowledgment}
The work of Jinshan Zeng is supported in part by the National Natural Science Foundation (NNSF) of China (No.61977038), and Thousand Talents Plan of Jiangxi Province (No. jxsq2019201124). The work of Mingwen Wang is supported in part by NNSF of China (No. 61876074).
The research of Yuan Yao is supported in part by HKRGC 16303817, ITF UIM/390, as well as awards from Tencent AI Lab and Si Family Foundation.
Part of Jinshan Zeng's work was done when he visited at Liu Bie Ju Centre for Mathematical Sciences, City University of Hong Kong.
| 54cfa3d500fe156af9bf55825c84089105ac787c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Studying giant spiral galaxies beyond the Local Group determines whether the assembly of the
Milky Way was typical. Having a vantage point outside the system also makes observations of
the stellar halo and circumgalactic medium (CGM) easier. The disk component can be completely separated
from the halo, for example, when such galaxies are viewed edge-on. The edge-on orientation of
the nearby galaxy, \n891, is ideal for studying the interaction between the disk, CGM, and satellite
galaxies.
\n891 resides in a spiral-rich group representative of
typical environments and has a luminosity similar to the group's brightest member, \n1023.
\n891 is similar to the Milky Way in spiral type and rotation speed \citep{Swaters1997}. In both galaxies,
a large fraction of the halo stars have been accreted from satellites \citep{Mouhcine2010,Bell2008} and recent star formation is elevated (compared to the disk as a whole) in a molecular ring and
central circumnuclear disk \citep{Scoville1993a}.
The warm ionized gas in the interstellar medium of \n891 is more extended than the Reynold's layer \citep{Reynolds1989}
in the Milky Way, and the central surface density of the warm ionized medium is roughly twice as high
in \n891 as in the Milky Way \citep{Rand1990,Dettmar1990}. The molecular
gas mass of \n891 is 2.5 times larger than that of the Milky Way,
and the current star formation rate (SFR) is proportionally higher. Star formation in
\n891 is more active than in the Milky Way disk yet not high enough to be considered a starburst.
The specific SFR and stellar mass of \n891 are similar to many COS Halos galaxies
at redshift $z \approx 0.2$ \citep{Tumlinson2011,Werk2014a}.
The proximity of \n891, however,
has made it possible to directly detect emission from the circumgalactic medium (CGM). The recent
21-cm detection of cold gas out to 90 - 120 kpc along the minor axis \citep{Das2020},
combined with the X-ray detection of the virialized halo gas \citep{Hodges-Kluck2018} and
the well-studied synchroton halo \citep{Schmidt2019}, together
provide excellent observational contraints on the properties of the inner CGM.
The vertical distribution of hot gas \citep{Bregman1994}, neutral gas \citep{Swaters1997}, molecular
gas \citep{GarciaBurillo1992}, warm dust \citep{McCormick2013a}, and stars \citep{vanderKruit1981} have been
measured in \n891 and found to contain extraplanar components.
A full 30\% of the HI resides in a halo \citep{Oosterloo2007a}. The thermal emission from the halo dust
indicates a larger dust mass than was originally estimated from absorption features
\citep{Howk1997a,Howk2000a,Popescu2000,Bianchi2008,Popescu2011}.
The first maps of the dust mass and temperature
showed a strong correlation with the distribution of stellar mass and star formation, respectively \citep{Hughes2014}.
The X-ray halo of \n891 was the first detected in a non-starburst galaxy, but it has
remained controversial as to whether the hot gas was heated by galactic accretion \citep{Hodges-Kluck2012a}
or feedback processes \citep{Strickland2004a,Strickland2004b,Tullmann2006}.
The presence of cold dust in \n891 has been known for several decades, but little was known about
its spatial distribution due to the low resolution of the far-infrared and submm observations. Observations of
\n891 with {\it Spitzer}/MIPS resolved disk and halo components at 24~${\mu}$m\ but lacked the spatial resolution
at longer wavelengths required to map the
spectral energy distribution, hereafter the SED
(GO-20528, PI C. Martin).
As part of the Very Nearby Galaxies Survey, a {\it Herschel} Guaranteed Time Key Project, the large vertical
extent of the dust in \n891 was resolved with {\it Herschel} PACS/SPIRE \citep{Hughes2014}
In this paper, we combine deeper {\it Herschel} observations (OT1\_sveilleu\_2, PI S. Veilleux)
with the {\it Spitzer}/MIPS data to better describe the amount of halo dust and its origin.
This work builds on our examination of \n4631 \citep[][hereafter Paper I]{Melendez2015a}, a lower mass galaxy with
a higher SFR surface density \citep{Tullmann2006}, and six nearby dwarf galaxies including \n1569 \citep[][hereafter Paper II]{McCormick2018}. The interstellar dust in \n891 plays an important
role in its appearance. Young, star-forming regions, for example, are much more apparent along the northeast
(approaching) side of the disk compared to the receding (southwest) half due to the accumulation of dust
along the inner edges of the trailing spiral arms. Our deep observations directly show the relative importance
of stellar feedback and satellite interactions for building a dusty halo. This new perspective should
prove generally valuable for understanding the dust-to-gas ratio in galaxy halos \citep{Menard2012a}, and the processes taking place in the outskirts of galaxies in general \citep{Veilleux2020}.
Our presentation is structured as follows. In Section~\ref{sec:observations}, we present the far-IR/submm
observations.
The morphological components of the dust distribution are defined in Section~\ref{sec:morph}
before describing our fits to the multi-band photometry
and the resulting physical properties in Section~\ref{sec:temp_mass}.
In Section~\ref{sec:discuss},
we argue that the superbubble will become a galactic wind and discuss the origin of a puzzling,
extraplanar dust spur. Our conclusions are briefly summarized in Section~\ref{sec:conclusions}.
We adopt a distance of 9.77~Mpc \citep{Ferrarese2000a}, which is the weighted mean of distances
derived from HST observations of planetary nebula and surface brightness fluctuations.
This distance gives an angular scale of 47.37~pc/\arcsec.
\section{Observation and Data Reduction} \label{sec:observations}
\subsection{Spitzer/IRAC and Spitzer/MIPS data}
\label{spitzer.sec}
{\it Spitzer} images at 3.6~${\mu}$m, 4.5~${\mu}$m, and 24~${\mu}$m\ are used in this study. Images taken with
the Infrared Array Camera (IRAC) at 3.6~${\mu}$m\ and 4.5~${\mu}$m\ cover the entire galaxy
(PID: 3, PI: G. Fazio). The observations consist of a series of
dithered images taken with exposure times from 0.4 to 96.80
seconds.
The basic calibrated
data were obtained from the {\it Spitzer} archive and combined into a
mosaic using the
{\sc montage}
package (http://montage.ipac.caltech.edu/) correcting for
variations in the overall level and rejecting outliers. The point
source flux accuracy is $\sim$2\% \citep{Reach2005} and the extended
source surface brightness accuracy is estimated at $\sim$4\%.
We observed \n891 with the {\it Multiband-Imaging Photometer
for Spitzer} \citep[MIPS,][]{Rieke2004a} at 24~${\mu}$m\
(PID: 20528, PI: C. Martin).
The images were obtained using
the medium rate, scan-map mode and the scan legs of $0.5^\circ$ with
148\arcsec\ cross-scan offsets. This observation consists of 5 scans
(Obs. id 14815488, 14815744, 14816000, 14816256, 14816512) with total
exposure times of 1600 seconds. The data were reduced with the MIPS
Data Analysis Tool v3.04 \citep[DAT][]{Gordon2005a}. Extra steps were
carried out to improve the images including readout offset correction,
array averaged background subtraction (using a low order polynomial
fit to each leg, with the region including NGC 891 excluded from this
fit), and exclusion of the first five images in each scan leg due to
boost frame transients. The point source flux accuracy is $\sim$2\%
\citep{Engelbracht2007} and the extended source surface brightness
accuracy is estimated at $\sim$4\%.
The 24~${\mu}$m\ image resolves stars and background galaxies, which
we mask out for accurate photometry of the galaxy.
We identified background galaxies using \texttt{SExtractor}. Briefly, we
set the detection threshold to 1.5$\sigma$ in surface brightness. The sources
identified by the segmentation map were then replaced by the values in the
background map at the same locations.
\subsection{Herschel/PACS and Herschel/SPIRE Data}
The Photodetector Array Camera and Spectrometer \citep[PACS,][]{Poglitsch2010a}
observation for NGC 891 was conducted as our cycle 1 open-time program
(OT1\_sveilleu\_2, PI: S. Veilleux). This program utilizes PACS blue
70~${\mu}$m\ and red 160~${\mu}$m\ channels with 6 scan position angles at 55$^{\circ}$,
70$^{\circ}$, 85$^{\circ}$, 95$^{\circ}$, 110$^{\circ}$, and 125$^{\circ}$
(Obs IDs: 1342237999, 1342238000, 1342238001, 1342238002, 1342238003, and
1342238004). The exposure time per each scan was 5596 seconds (total 9.3
hours) including overhead.
The scan map leg length and separation were
4\arcmin\ and 4\arcsec\ and the number of scan map legs was 60. Each scan map was repeated three times for redundancy reasons.
We also retrieved, reduced, and analyzed the PACS data from {\it The Herschel EDGE-on galaxy Survey}
\citep[HEDGES,][OT2\_emurph01\_3]{Murphy2011a} for the analysis of extended
structure as these data have a larger field-of-view than do our observations.
For the blue channel
(70~${\mu}$m), the scan angle was 45$^{\circ}$ and 135$^{\circ}$ with an exposure
time of 6550 seconds. The scan map leg length and separation were 14\farcm6
and 14\farcs0 and the number of scan map legs was 70. For the green channel
(100~${\mu}$m), the exposure time was 8977 seconds at scan angles of 45$^{\circ}$
and 135$^{\circ}$. The scan map leg length and separation were 14\farcm6
and 20\farcs0 and the number of scan map legs was 51.
The observation in 250~${\mu}$m, 350~${\mu}$m, and 500~${\mu}$m\ bands were performed using
the Spectral and Photometric Imaging Receiver \citep[SPIRE,][]{Griffin2010a}
onboard {\it Herschel Space Observatory}. Our SPIRE images are the part of the
KINGFISH open-time key program, KPGT\_cwilso01\_1, from {\it Herschel} guaranteed
time key project,{\it Physical Processes in the Interstellar Medium of Very
Nearby Galaxies}\footnote{http://hedam.lam.fr/VNGS/index.php} \citep[ID:
1342189430,][]{Griffin2010a}. The beam sizes are 465.39, 822.58, and 1768.66
arcsec$^2$ at 250, 350, and 500~${\mu}$m, respectively (SPIRE Data Reduction
Guide, Table 6.7) and the exposure times were 1678 seconds for each channel.
\subsection{Data Reduction}
For the PACS data, we used the {\it Herschel Interactive Processing Environment}
\citep[HIPE,][]{Ott2010a} version 10.0 for data reduction. This reduction
procedure contains the extraction of the calibration tree for the data
processing, electronic crosstalk correction, flat-fielding, and unit conversion
from volts to Janskys per pixel. The final map was created with the algorithm
implemented in {\it Scanamorphos} \citep[v24.0,][]{Roussel2013a}. The error
and weight maps were also produced with {\it Scanamorphos} and the error
map is defined as the error on the mean brightness in each pixel.
The SPIRE data were processed from level 0 up to level 1 with HIPE scripts
in the Scanamorphos distribution. The same steps as PACS processing were
conducted. The thermal drifts are subtracted by using the smoothed series
of thermistors located on the detector array as the input of the drift model (see \citet{Ott2010a} for details).
A more detailed description of the data reduction can be found in Paper I.
Both sets of PACS observations contain a
low-amplitude artifact parallel to the major axis on the eastern side of the disk midplane. It is 70\arcsec\ away to the south-east and the flux in the ghosts is less than 0.5\% of the corresponding object flux.
This region immediately east of the disk midplane was not used in our analysis of the vertical structure of the disk.
The amplitude of the feature is small enough, that it has a negligible effect on our integrated flux measurements.
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{f1new.pdf}
\caption{{\it Spitzer} MIPS 24~${\mu}$m, {\it Herschel} PACS 70 and 160~${\mu}$m, and
{\it Herschel} SPIRE 250, 350, and 500~${\mu}$m\ images of \n891 The color bars
show surface brightness on a logarithmic scale; the maximum central
surface brightness is white, and the images are grayed out below a
S/N ratio of 1.5. The black arrow indicates the dust spur to the
south-east. The red circles indicate the size of the point spread
function.
The color bar ranges (in Janskys per pixel on a logarithmic scale)
are as follows: 0.0032 to 63 (24 micron), $10^{-5}$ to 0.079 (70 micron),
$10^{-5}$ to 0.40 (160 micron), $10^{-4}$ to 0.50 (250 micron),
$10^{-4}$ to 0.25 (350 micron), and
$10^{-4}$ to 0.16 (500 micron).
}
\label{fig:6bands}
\end{center}
\end{figure*}
\subsection{Integrated Photometry of \n891 \& Global Properties}
Figure~\ref{fig:6bands} compares the new (24, 70, 160~${\mu}$m) images to
the SPIRE images at 250, 350, and 500~${\mu}$m. We define the center
of \n891 by the location of the maximum 4.5~${\mu}$m\ IRAC intensity, a proxy
for the center of the stellar bulge. In Figure~1 and subsequent images,
the coordinates (0,0) reference this position. Table~\ref{tab:ngc891} provides the
corresponding sky coordinates.
We measured fluxes from the 24, 70, 160, 250, 350, and 500~${\mu}$m~images.
In order to match the point spread function (PSF), we convolved the images with the kernels provided
by \citet{Aniano2011a}. The image pixel scale and coordinates were then
matched using the IDL code \texttt{hrebin.pro} and \texttt{hastrom.pro}
in the Goddard package. These processes conserve total image flux.
The aperture adopted for the integrated photometry is based on a growth curve measured
from photometry of the SPIRE 500~${\mu}$m~image in increasingly larger apertures. It is an
elliptical aperture of semi-major axis of 7\farcm35 and semi-minor axis of
3\farcm68 that includes most of the dust emission seen in 24~${\mu}$m\ image after
convolution to the PSF of the 500~${\mu}$m\ image. Figure~\ref{fig:clean24} shows this
aperture on the 24~${\mu}$m~image.
Within an annulus at larger radii, we masked point sources and background galaxies
based on the morphology in the 24~${\mu}$m\ data. The sky background was taken as
the mean data value following 2$\sigma$ clipping. The variance in this region
was used to compute the standard deviation of the mean. This uncertainty on
the mean background dominates the total background error for large apertures.
We also include a term for the pixel-to-pixel statistical variance in the background.
The error budget is dominated by the calibration uncertainties, which
exceed the uncertainties from both background subtraction and photon noise.
Following Paper I, we conservatively estimate the background error at 10\% in all bands.
We computed color corrections for the PACS and SPIRE fluxes assuming a
blackbody spectrum modified by the dust emissivity, $\kappa_{\nu}=\kappa_{\rm 0} (\nu/\nu_{\rm 0})^{\beta}$
with $\beta = 2$, and a blackbody temperature of $T = 30$~K.\footnote{
As the PACS and SPIRE pipelines compute fluxes assuming a monochromatic spectrum, we applied a color correction.
See section 4.3 in ``PACS Photometer Passbands and Color Correction
Factors for Various Source SEDs'', PICC-ME-TN-038, version 1.0, and
Table 5.7 in SPIRE Handbook, HERSCHEL-DOC-0798, version 2.5.}
We divided the original flux measurements by the following correction factors:
0.977, 1.037, 0.9796, 0.9697, and 0.9796 for,
respectively, 70~${\mu}$m, 160~${\mu}$m, 250~${\mu}$m, 350~${\mu}$m, and 500~${\mu}$m.
The estimated aperture corrections were small compared to the photometric errors and were therefore not applied.
The integrated fluxes listed in Table~\ref{tab:ngc891} for the 24, 70, 160, 250, 350,
and 500~${\mu}$m~bands.
Our measurements
agree with both \citet{Bendo2012a} and \citet{Hughes2014}
to within the $1\sigma$ errors.
\begin{table}
\centering
\caption{Properties of \n891}
\label{tab:ngc891}
\begin{tabular}{lr}
\hline
Property & Value \\
\hline
RA (J2000)$^{a}$ & 2:22:33.0 \\
DEC (J2000)$^{a}$ & 42:20:53 \\
d (Mpc)$^{b}$ & 9.77 \\
$V_{\rm circ}$ (\mbox {~km~s$^{-1}$})$^{c}$ & 225 \\
SFR (\mbox {$~{\rm M_\odot}$~yr$^{-1}$}) & 5.0 \\
$F_{\nu}^{ 24\mu m}({\rm Jy})$ & $ 6.2 \pm 0.6$ \\
$F_{\nu}^{70\mu m}({\rm Jy})$ & $ 100 \pm 10$ \\
$F_{\nu}^{160\mu m}({\rm Jy})$ & $ 334 \pm 33$ \\
$F_{\nu}^{250\mu m}({\rm Jy})$ & $ 167 \pm 17$ \\
$F_{\nu}^{350\mu m}({\rm Jy})$ & $ 72 \pm 7$ \\
$F_{\nu}^{500\mu m}({\rm Jy})$ & $ 26 \pm 3$ \\
\hline
\end{tabular}
\\ $^{a}${Coordinates of the maximum intensity in the 4.5~${\mu}$m\ IRAC
image.} $^{b}${Distance from \citet{Ferrarese2000a}.}
$^{c}${Rotation speed from \citet{Swaters1997}.}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth, angle=0, trim = 0 0 0 0]{f2new.pdf}
\caption{{\it Spitzer}/MIPS 24 ~${\mu}$m\ image for NGC 891 adaptively smoothed after subtraction of background sources.
The color scale for the surface brightness is logarithmic.
The scale height and scale length of the stars are indicated by the rectangle,
$\pm0\farcm5$ (1.4 kpc) by $\pm 7\farcm35$ (20.9 kpc) \citep{Ibata2009a}.
The extended emission forms a halo which we define by an ellipse
(semi-major axis of $7\farcm35$ and semi-minor axis of 3\farcm68, or 10.5~kpc).
An arrow indicates the dust spur.}
\label{fig:clean24}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[trim={0 0 0 0}]{f3new.pdf}
\caption{Adaptively smoothed maps.
We applied \texttt{ADAPTSMOOTH} \citep{Zibetti2009a} with a minimum S/N = 5 pixel$^{-1}$.
{\it Left:} The {\it Herschel} 250~${\mu}$m\ map overlaid with the \text{H\,\textsc{I}}~emission contours from \citet{Oosterloo2007a}.
The contour levels for \text{H\,\textsc{I}}~emission are 0.01, 0.03, 0.08, 0.16, 0.47,
0.78, 1.56, 3.11, 6.23, 9.34$\times \rm 10^{21}~cm^{-2}$.
{\it Middle:} The {\it Herschel} 250~${\mu}$m\ map overlaid with the radio continuum contours
from \citet{Oosterloo2007a}. The contour levels for the radio continuum are
0.1, 0.4, 2, and 10 mJy beam$^{-1}$.
{\it Right:} The {\it Spitzer} 24~${\mu}$m\ map with the \text{H\,\textsc{I}}\ contours. The discrete sources in the background
were masked prior to the smoothing.
}
\label{fig:spire_smooth}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{f4new.pdf}
\caption{The PACS 70~${\mu}$m/160~${\mu}$m\ color map.
The arrow points at the western superbubble.
{\it Left:}
The X-ray contours for the adaptively smoothed {\it Chandra} image
(0.3-2.0 keV) presented by \citet{Hodges-Kluck2012a}; the contour
levels are 9, 10, 15, 20, 30, 40, and 50 counts s$^{-1}$ deg$^{-2}$.
{\it Top-right:}
The same color maps overlaid with radio continuum contours. The spatial resolution is 17\arcsec\ $\times$ 12\arcsec\ and the rms noise is 23 $\mu$Jy beam$^{-1}$ \citep{Oosterloo2007a}.
The contour levels for the radio continuum are
0.1, 0.4, 2, and 10 mJy beam$^{-1}$.
{\it Bottom-right:}
The same color maps overlaid with \mbox {H$\alpha$}\ surface brightness contours
from an \mbox {H$\alpha$}\ image taken at Steward Observatory by C. Engelbracht
and calibrated to the total luminosity reported in \citet{Hoopes1999}.
Contours are shown at $\Sigma(H\alpha) = 9.9 \times 10^{-15},
1.0 \times 10^{-14},
1.23\times 10^{-14},
2.61 \times 10^{-15},
6.09 \times 10^{-15},
1.02 \times 10^{-13}$~ergs~s$^{-1}$~cm$^{-2}$\ arcsec$^{-2}$.
}
\label{fig:pacs_color}
\end{center}
\end{figure*}
\section{Morphological Features of the Warm Dust in \n891} \label{sec:morph}
The 24~${\mu}$m~image in Figure~\ref{fig:clean24} resolves several structures
contributing to the thermal emission. The brightest of these are thin
and thick dust components aligned with the center of the galaxy, components
we will refer to as the dusty disk and dusty halo, respectively. Modeling
these components in the longer wavelength {\it Herschel} images is
sensitive to how the PSF is modeled \citep{Bocchio2016}.
We adopt the scale heights that \citet{Bocchio2016} fitted to the
vertical surface brightness profiles. Figure 2 of their paper
shows that the shape of the vertical surface brightness profile
requires a two-component fit. The thin, unresolved dust component
significantly reduces the scale height of the fitted thick dust component.
A consequence of this interplay is that the thick dust component
has the largest 24~${\mu}$m\ scale height above the region of the stellar disk where
the thin dust component is faintest. Based on the vertical profile
perpendicular to a region of the disk with low star formation activity,
\citet{Bocchio2016} conclude that the scale height of the dusty halo
is $h_z = 1.36 \pm 0.01$~kpc.
The scale height of the dusty halo is therefore comparable to that of the thick
stellar disk, whose size we have indicated in Figure~\ref{fig:clean24}.
Our analysis describes two non-axisymmetric components of the far-infrared morphology:
(1) the large filament marked by the arrow in Figure~\ref{fig:clean24}, and (2) a
dusty superbubble erupting from the center of the galaxy. The superbubble was
detected previously at X-ray energies \citep{Hodges-Kluck2012a,Hodges-Kluck2018}, and
we present the first description of its dust content. The filament, or {\it dust spur},
is a new discovery enabled by the spatial resolution and sensitivty of the new observations.
\subsection{The Dust Spur}
The arrow in Fig.~\ref{fig:clean24} indicates a filament of 24~${\mu}$m\ emission extending
6\farcs90 (19.6 kpc) southeast of the dust halo. We discovered this filament
in the {\it Spitzer} MIPS data and confirmed it using {\it Herschel}.
Figure~\ref{fig:6bands} shows the SPIRE detections. The structure is
outside the field-of-view of our new PACS images, so we examined the
PACS images obtained previously by {\it The Herschel EDGE-on galaxy Survey}
\citep[HEDGES,][]{Murphy2011a} and found a clear detection at at 160~${\mu}$m,
a few clumps of emission at 100~${\mu}$m, and weaker emission at 70~${\mu}$m. The
far-infrared filament is therefore not an instrumental artifact.
We looked for a background cluster of galaxies at this location but found
none in either X-ray images or optical images.\footnote{The $z=0.0184$ cluster Abell~347 lies
to the southeast at 02:25:50.9,+41:52:30 (NED).} Our photometry indicates a
broad spectral energy distribution, which we show is consistent with thermal
dust emission in Section~\ref{sec:temp_mass}. The far-infrared filament may
therefore be a component of the dusty halo of \n891, and we will refer to
this morphological feature as {\it the dust spur}. Spectroscopy of an emission
line would confirm its association with \n891; alternatively, higher-resolution
mapping would be useful to definitively rule out an unresolved galaxy cluster.
The morphology of the dust spur shows no connection to the center of the galaxy,
making an association with a galactic outflow unlikely. When projected on the sky,
the long-axis of the dust spur is roughly perpendicular to the disk. The filament
emerges at a galactocentric radius $R$ = 4\farcm0 (11 kpc) and extends 6\farcm9 (19.6 kpc)
to the southeast, and has a width of $\sim$ 2\farcm0 (6 kpc).
The dust spur is {\it not} coincident with the large filament of neutral hydrogen
described by \citet{Oosterloo2007a} \citep[see also][]{Pingel2018}. In Figure~\ref{fig:spire_smooth}, we show their
\text{H\,\textsc{I}}\ 21-cm emission contours on the {\it Herschel} and {\it Spitzer} images. The
lowest contour at $N(HI)= 1\times10^{19}\rm~cm^{-2}$ extends 22~kpc
northwest of \n891. The dust spur extends to the southeast, in contrast, and is not
visible in the \text{H\,\textsc{I}}\ contours.
We will show that the \text{H\,\textsc{I}}\ filament and the dust spur have different compositions
in Section~\ref{sec:temp_mass}.
The middle panel of Fig.~\ref{fig:spire_smooth} compares the dust emission to the
radio continuum. The overall shape of the radio continuum contours is similar
to the distribution of far-infrared emission, whereas the \text{H\,\textsc{I}}\ contours show
a lower ratio of vertical height to galactocentric radius.
A close inspection of the dust spur suggests a spatial coincidence between
the brightest emission regions at 250~${\mu}$m\ and the local maxima in the
the faintest radio continuum contours. This correlation should be interpreted
cautiously. Unresolved sources in the SPIRE map could be background sources
unrelated to \n891, and the radio detections have low signal-to-noise. If
the radio sources are at the redshift of the dust spur, then
the radio flux levels of the clumps, 0.2 mJy beam$^{-1}$,
combined with the fluxes listed in Table~\ref{tab:fir_radio},
are consistent with far-IR -- radio relation defined by star-forming regions
\citep{Yun2001a,Bell2003a}. These data raise the possibility that
the dust spur is a region of active star formation.
\subsection{The Superbubble}
We smoothed the 70 ~${\mu}$m\ image to match the resolution of the 160 ~${\mu}$m\ image, and then we divided the
smoothed 70 ~${\mu}$m\ image by the 160 ~${\mu}$m\ image to produce a color map from the PACS data. This color
map resolves a broad plume of thermal dust emission, indicated by the arrow in Figure~\ref{fig:pacs_color}.
This structure emerges from a region of the disk within roughly 1\farcm8 (5.1 kpc) of the nucleus, slightly
inside the 3\farcm1 radius of the molecular ring. It extends northwest, roughly perpendicular to the disk,
reaching a height of 2\farcm7 (7.7 kpc). We refer to this structure as {\it the superbubble} but will argue
in Section~\ref{sec:discuss} that a galactic wind is actually developing here. We will show that this region
has a slightly higher dust temperature than the surrounding halo in Section~\ref{sec:temp_mass}.
Contours of the soft X-ray emission \citep{Hodges-Kluck2012a} outline the perimeter of the dusty superbubble.
In deeper X-ray observations, this hot ($kT = 0.71 \pm 0.01$ keV) gas is concentrated near the most
active region of star formation and coexists with gas near the virial temperature, $kT = 0.20 \pm 0.01$ keV
\citep{Hodges-Kluck2018}. Harder diffuse X-ray emission is detected within 5 kpc of the galactic center;
its physical origin remains unclear.
The soft X-ray emission from galactic winds is produced mostly at the interface between the hot wind and a cooler
component of the multi-phase outflow \citep{Strickland2004a,Strickland2004b}.
The smaller panels in Figure~\ref{fig:pacs_color} overlay the radio continuum contours and the \mbox {H$\alpha$}\ contours
on the PACS color map. The \mbox {H$\alpha$}\ image shows one prominent filament extending well into the dusty bubble, but the \mbox {H$\alpha$}\
image is not sensitive enough to determine the amount of correlation with the superbubble morphology at other wavelengths.
We have shown in Figure~\ref{fig:spire_smooth} that the radio contours closely follow the morphology of
the dusty halo. As would therefore be expected, the radio contours do not describe the PACS color map well.
The fate of the superbubble will be determined in large part by the CGM.
The inner CGM of \n891 contains much more cold gas than hot, virialized gas \citep{Hodges-Kluck2018,Das2020}
\subsection{The Dusty Halo} \label{sec:halo}
On the eastern side of the disk, the 24~${\mu}$m\ emission
extends 2\farcm03 (5.76 kpc) over most of the disk,
reaching a larger distance along the dust spur (Fig.\ \ref{fig:spire_smooth}).
On the western side of the \n891 disk, the dusty halo is detected out
to 3\farcm68 (10.5 kpc) in the 24~${\mu}$m\ map.
The extent of the PAH emission is similar to that of the cold dust. \citet{McCormick2013a}
estimate a symmetric extent of $\pm 5\farcm0$ with a bulk height, $H_{e,PAH} = 7.1$~kpc.
\section{Dust Temperature and Mass} \label{sec:temp_mass}
We can derive the dust mass, $M_{\rm d}$, from the opacity per unit dust mass
and the optical depth at far-infrared wavelengths. We fit a simple,
single-temperature modified blackbody (MBB) function to the photometry in the
wavelength range $100 \le\ \lambda (\mu {\rm m}) \le\ 500$. The dust
spectrum consists of a blackbody spectrum at the dust
temperature, $B_{\nu}(T_d)$, modified by the dust opacity, $\kappa_{\nu}$,
such that \begin{eqnarray}
F_{\nu} = {{M_{\rm d} \kappa_{\nu} B_{\nu} (T_{\rm d})} \over {D^2}},
\end{eqnarray}
where $D$ is the distance to the galaxy.
This simple approach yields accurate dust masses and temperatures
provided the normalization and the spectral index of the opacity
are consistent with full dust models, which include
the distribution of dust grain properties and a range of interstellar
radiation fields \citep{Bianchi2013}.
We adopt a dust opacity model, $\kappa_{\nu}=\kappa_{\rm 0} (\nu/\nu_{\rm 0})^{\beta}$.
The normalization, $\kappa_{\rm 0} = 1.92~\rm cm^2~g^{-1}$ at 350~${\mu}$m,
and spectral index, $\beta = 2.0$, are based on the Milky Way model presented in
\citet{Draine2007a}.
In the next section, we validate our approach using the spectral energy distribution for
the entire galaxy. In Section~\ref{sec:sed_components}, we then describe the dust properties of the
morphological components identified in the previous section and defined in Figure \ref{fig:region}. We begin with the dust mass and temperature
of each region as a whole, and then we explore these properties on a pixel-by-pixel basis.
We estimate dust-to-gas ratios for the various components in Section~\ref{sec:d2g}.
\subsection{Validation of the Spectral Energy Distribution Fitting}
The top panel of Figure~\ref{fig:sed_fit} shows the SED for the entire galaxy.
We fit only the 70 $-$ 500~${\mu}$m\ data because the 24~${\mu}$m\ flux comes from a
warmer dust component.
The single-temperature, MBB model with $\beta \equiv 2$
gives a dust mass of $M_{d} = 1.1 \times 10^8$\mbox {${\rm ~M_\odot}$}\ and dust
temperature $T =21.7\pm0.2$~K. We also plot the fit with $\beta$
taken as a free parameter. A comparison shows that our fitted dust
parameters are robust to minor changes in the dust emissivity; see
Table~\ref{tab:dust} for a quantitative comparison.
The dominant error term is the 10\% uncertainty in the flux calibration.
To determine the uncertainty on the fitted mass and temperture, we
scale the photometry by 10\% and refit the models. We find that
systematically increasing (or decreasing) all the fluxes can increase (decrease)
the fitted dust masses by $\sim$10-20\%, but dust temperature varies by less than a degree.
Our results confirm the dust mass and temperature obtained previously by \citet{Hughes2014,Bocchio2016}.
The total dust mass fit to the FIR observations is nearly twice
that inferred from radiative transfer modeling of the distribution of
optical and near-IR emission. For example, when scaled to our adopted distance,
\citet{Xilouris1998} found $5.3 \times 10^7$\mbox {${\rm ~M_\odot}$}\ of dust. A significant fraction
of the dust presumably goes undetected because little optical/near-IR radiation emerges
from some regions of the galaxy. Previous work has suggested
that this hidden dust could reside in a geometrically thinner dust disk \citep{Popescu2000,Popescu2011} or
in clumpy clouds associated with molecular gas \citep{Bianchi2008}.
\citet{Seon2014a} considered \n891 specifically and introduced a thin dust disk of
scale height $\approx 5\arcsec$ to simultaneously describe the distribution of GALEX UV
emission \citep{Morrissey2007} and {\it Herschel} SPIRE photometry \citep{Bianchi2011}.
For completeness, we point out that
the impact of smaller dust grains (grain radii $< 50$ \AA) on our fits
is constrained by the 24~${\mu}$m\ flux. We fit a two temperature model to the
full SED including the 24~${\mu}$m\ flux. The temperature of a warmer component is not
well constrained by this single data point, so we assumed a dust temperature of
$T=220$~K to facilitate comparison to
\n4631. In Paper I, we found an additional dust component at 220 K in \n4631.
In their single-temperature, MBB fits, they treated the 70~${\mu}$m~flux as
an upper limit because the contribution from the warmer component
was non-negligible. We find, however, that the contribution of
this warmer component to the 70~${\mu}$m~flux is only 0.3\% for this
two temperature model fitted to \n891.
Assuming that the peak wavelength of the MBB curve for the warmer component is shorter than 24~${\mu}$m, we vary the temperature of the warmer component, and the contribution to 70~${\mu}$m\ flux is 1.3\% at most. Therefore, we consider 70~${\mu}$m\ data point in the same way as other wavelength data in the fitting process.
For an assumed warm dust temperature
of 220~K, we fit a two-component model and find dust mass in the 220 K component is negligible ($\sim0.00003$\%) compared with dust mass in the $\sim22$ K component. The mass in the cool component in this two-component fitting does not deviate from the mass in the one-component, cool component only, fitting as shown in Table~\ref{tab:dust}.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{f9.pdf}
\caption{Apertures used for far-infrared photometry shown
relative to \text{H\,\textsc{I}}\ 21-cm contours from \citet{Oosterloo2007a}. The semi-major and semi-minor axises are 7\farcm4 (21 kpc) and 3\farcm7 (10.5 kpc), respectively.
We define the thickness of the disk by the thick stellar disk with the width of $\pm 0\farcm5$ (1.4 kpc). The inner disk has a width of 4\farcm1 (11.7 kpc) which includes the three star-forming clumps. The other rectangle marks the aperture we use for the superbubble photometry. The height and width of the superbubble are 2\farcm7 (7.7 kpc) and 2\farcm8 (8.0 kpc), respectively.
The ellipse encompasses the extended halo emission in the lowest
resolution band.
We define the aperture for the dust spur southeast of the galaxy
by the 250~${\mu}$m\ contour at 1.7 mJy/pix, which is 3$\sigma$ of
the background noise, and outside the extended dust halo.
We define an aperture northwest of the
galaxy in order to place upper limits on the far-infrared emission
from the \text{H\,\textsc{I}}\ spur.
Each dust clump and hot spot (star forming clump) listed in Table~\ref{tab:fir_radio} is labeled with a number.
}
\label{fig:region}
\end{center}
\end{figure}
\subsection{SED Fitting of Dust Components} \label{sec:sed_components}
Table~\ref{tab:dust} summarizes the dust properties for the regions labeled in Figure~\ref{fig:region}.
The photometry shown in Figure~\ref{fig:sed_fit} was obtained after
degrading the resolution of the 70, 160, 250, and 350~${\mu}$m\ maps to match that of the 500~${\mu}$m\ map;
see details in Section~\ref{sec:observations}.
Our analysis constrains the contribution from dust well beyond the galactic plane where
the starlight is too faint to model its attenuation.
Inspection of Table~\ref{tab:dust} indicates 87\% of the dust mass is associated with
the galactic disk (\rm log ($M_{\rm dust}/M_{\odot}$) $=7.970\pm0.006$) with 65\% of the disk dust
concentrated in the inner disk (\rm log ($M_{\rm dust}/M_{\odot}$) $=7.781\pm0.012$).
We associate roughly 13\% of the total dust mass with the halo component at
$|z|>0\farcm5 $, or $\pm\ 1.42$ kpc; the same halo aperture includes 20\%
of the total \text{H\,\textsc{I}}\ mass. This difference can be attributed to the lower dust-to-gas ratio
of the halo. Roughly 25\% of the halo dust mass (\rm log ($M_{\rm dust}/M_{\odot}$) $=7.163\pm0.053$) comes from
the superbubble region, while the dust spur, \rm log ($M_{\rm dust}/M_{\odot}$) $= 5.88\pm0.02$, makes a smaller
contribution.
The halo dust temperature ($T=19.24\pm0.62$~K) is significantly cooler
than the disk dust ($T=21.89\pm0.06$~K). The superbubble dust ($T=20.61\pm0.03$~K)
is warmer than the surrounding halo but cooler than the disk.
The emission from the dust spur is weak at 70~${\mu}$m\ compared to the SPIRE bands, and
the fitted dust temperature is considerably lower than that of the disk or halo.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\columnwidth, angle=0, trim = 100 0 100 0]{f10.pdf}
\caption{The best fit MBB model for integrated flux of each part of NGC 891 in
24~${\mu}$m, 70~${\mu}$m, 160~${\mu}$m, 250~${\mu}$m, 350~${\mu}$m, and 500~${\mu}$m\ bands.
The regions are defined in Figure \ref{fig:region}. The flux for 24~${\mu}$m\ is not used for the fitting as it traces different dust components. Gray-dashed line is the MBB model fits when $\beta$ is fixed at 2.}
\label{fig:sed_fit}
\end{center}
\end{figure}
\begin{table*}
\centering
\caption{Dust properties derived from the MBB model fitting.}
\label{tab:dust}
\begin{tabular}{ccccccccc}
\hline
Component$^{d}$ & Total & Inner Disk & Disk & Halo & Superbubble & Dust Spur & \text{H\,\textsc{I}}\ Spur \\
\hline
log ${\rm M}_{\rm dust}$ & 8.025$\pm$0.005 & 7.781$\pm$0.012 & 7.970$\pm$0.006 & 7.163$\pm$0.053 & 6.468$\pm$0.003 & 5.883$\pm$0.019 & $<3.443$\,$^{a}$ \\
$\prime\prime$ & 8.018$\pm$0.007 & 7.800$\pm$0.018 & 7.970$\pm$0.013 & 7.061$\pm$0.074 & 6.474$\pm$0.004 & 5.868$\pm$0.061 & \ldots \\
$\beta$ & 2 & 2 & 2 & 2 & 2 & 2 & 2\\
$\prime\prime$ & 1.97$\pm$0.03 & 2.09$\pm$0.07 & 2.00$\pm$0.05 & 1.66$\pm$0.22 & 2.03$\pm$0.02 & 1.96$\pm$0.16 & \ldots\\
T (K) & 21.72$\pm$0.05 & 22.49$\pm$0.12 & 21.89$\pm$0.06 & 19.25$\pm$0.62 & 20.61$\pm$0.03 & 16.57$\pm$0.17 & 19.25$^{a}$ \\
$\prime\prime$ & 21.89$\pm$0.15 & 22.02$\pm$0.37 & 21.89$\pm$0.26 & 21.30$\pm$1.45 & 20.50$\pm$0.08 & 16.79$\pm$0.81 & \ldots \\
$\chi_{\nu}^2$ & 0.02 & 0.11 & 0.03 & 1.05 & 0.01 & 0.08 & \ldots \\
$\prime\prime$ & 0.02 & 0.09 & 0.05 & 0.75 & 0.01 & 0.11 & \ldots \\
log $({\rm M}_{\rm HI}/{\rm M}_{\odot}$)\,$^{b}$ & 9.64 & 9.09 & 9.48 & 9.10 & 8.07 & $<7.22$ & 7.20 \\
${\rm M}_{\rm dust}$/${\rm M}_{\rm gas}$\,$^{c}$ & 0.0088 & 0.0082 & 0.0095 & 0.0084 & 0.025 & $>0.0046$ & $<0.00017$ \\
$\prime\prime$ & 0.0087 & 0.0086 & 0.0095 & 0.0067 & 0.025 & $>0.0044$ & \ldots \\
\hline
\end{tabular}
\\
$^{a}$
The dust mass limit is computed by rescaling the MBB
model to satisfy the FIR flux limits of the
\text{H\,\textsc{I}}\
spur assuming
$\beta=2$ and the halo dust temperature.
$^{b}$ {The \text{H\,\textsc{I}}\ masses were measured from \text{H\,\textsc{I}}\ map of \citet{Oosterloo2007a}.}
$^{c}$ {The gas mass was computed as $M_{\rm gas} = 1.36 (M(\text{H\,\textsc{I}}) +
M({\rm H_2}))$, where the coefficient accounts for the gas mass in
helium. }
$^{d}$ {Alternate rows compare fixed and free spectral indices for the dust opacity.}
\end{table*}
\subsection{Pixel by Pixel Analysis} \label{sec:pixel_by_pixel}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{f11new.pdf}
\caption{Pixel-to-pixel dust mass, dust temperature, and
$\chi_{\nu}^2$ of the model fit with $\beta \equiv 2$
are shown. Each pixel is $9\farcs0
\times 9\farcs0$. The disk boundary defined in Figure~\ref{fig:region}
is indicated by the white-dotted line in each panel.
}
\label{fig:pix2pix}
\end{center}
\end{figure}
We also examined the variation in dust properties at the resolution limit of
the 500~${\mu}$m\ image. We fit the SED of all pixels that had $S/N > 3\sigma$ in all five bands,
essentially the entire disk component and the disk -- halo interface.
\citet{Hughes2014} previously found the variation in dust temperature to be more uneven
than that of the dust mass. Figure~\ref{fig:pix2pix} shows our resulting maps of the temperature
and dust surface density, i.e., the column density of dust in mass units.
The dust surface density decreases with distance from the nucleus along the major axis.
In the vertical direction, the dust surface density
drops very quickly to values less than 0.50\mbox {${\rm ~M_\odot}$}\ pc$^{-2}$.
The temperature range in Fig.~\ref{fig:dust_property}
is similar to the $T_d = 17-24$~K range found previously by \citet{Hughes2014}.
The hotspots identified in both the
24~${\mu}$m\ map and the 70~${\mu}$m/160~${\mu}$m~color map stand out as regions with
$T_d \ge\ 22.5$~K in Figure~\ref{fig:pix2pix}. This region of the disk is
surrounded by a molecular ring \citep{Israel1999a,Scoville1993a}.
The dust cools off with increasing radius in the disk beyond the molecular ring.
We compare the dust temperature and surface density directly to the 3.6~${\mu}$m\ and 24~${\mu}$m\ emission in
Figure~\ref{fig:dust_property}. The 24~${\mu}$m\ emission is a good proxy for
the star formation rate surface density \citep{Calzetti2007}, while the
3.6~${\mu}$m\ emission traces the stellar mass density. We have color coded the
points in Figure~\ref{fig:dust_property} by their location in \n891.
The dust in the midplane of the inner disk has temperature, $T_D \approx 22-23$~K and a high
mass column, as indicated by its location in the upper right of Figure~\ref{fig:dust_property}.
With increasing vertical height at fixed
radius, the temperaure decreases by roughly one degree as the surface brightness drops by an order of magnitude.
This locus (blue points) defines a narrow band that continues smoothly into the emission from the superbubble
region where the temperature drops a few more degrees (orange and yellow points). \citet{Hughes2014} produced
color maps that identify this same spectral transition with height above the disk. Adding the halo emission
immediately above the inner disk extends this correlation to dust temperatures $T_D \lower2pt\hbox{$\buildrel {\scriptstyle < 20$~K. Based on
this correlation between dust temperature and 24~${\mu}$m\ emission, the dust in and above the inner
disk is heated primarily by the starlight from the star-forming regions.
The width of the SB(24~${\mu}$m) -- $T_D$ locus in the inner disk correlates with radial distance from
the nucleus. For example, following the midplane points in Figure~\ref{fig:dust_property} from
the inner disk to the outer disk (green points) follows the upper edge of the triangular locus. At
fixed 24~${\mu}$m\ surface brightness, dust in the midplane of the outer disk is cooler than dust above
the plane at small radii. Equivalently, at fixed dust temperature, the superbubble and halo emission
are not as bright at 24~${\mu}$m\ as the outer disk. We expect a transition in the outer disk toward
heating by an older stellar population, and it seems plausible this transition manifests as the broad
triangle of green points in Figure~\ref{fig:dust_property}. Consistent with this interpretation, the
3.6~${\mu}$m\ emission from the midplane points in the outer disk (green points with cyan dot) are strongly
correlated with the dust temperature. Dust above the midplane of the outer disk fills the interior
of the triangular locus, as might be expected for dust heated by a mixture of the two stellar populations.
We acknowledge that the temperature correlations look nearly identical against
the near-infrared (3.6~${\mu}$m) and mid-infrared (24~${\mu}$m) emission, so the temperatures have the
same correlations with the tracers of stellar mass and star formation, respectively. It seems likely
that this happens because the mass and SFR both decrease with separation from the center of
the galaxy.
The panels on the right side of Figure~\ref{fig:dust_property} show power law
correlations between infrared emission and dust surface density. For the inner disk,
the correlation is tighter in the mid-infrared (24~${\mu}$m), more closely related to SFR surface density.
In the outer disk, the 3.6~${\mu}$m\ emission shows the stronger correlation, consistent with a closer relation
to stellar mass surface density.
\begin{figure*}
\begin{center}
\includegraphics[trim={0 0 0 0}]{f12.pdf}
\caption{
Pixel-to-pixel correlations between dust properties and shorter wavelength
emission not used in the SED fitting. The regions shown correspond to the map
in Figure~\ref{fig:pix2pix}. In this plot, {\it midplane} is defined as the
region within $|z| \le 0\farcm7$ (1.99 kpc) of the disk plane.
}
\label{fig:dust_property}
\end{center}
\end{figure*}
\subsection{Dust-to-gas Ratio} \label{sec:d2g}
Table~\ref{tab:dust} compares the dust mass to the gas mass, where all
measurements taken from the literature have been scaled to an \n891
distance of 9.77~Mpc. The total \text{H\,\textsc{I}}\ mass of \n891 is then $4.4
\times 10^9$\mbox {${\rm ~M_\odot}$} \citep{Oosterloo2007a}. The total CO(1-0) intensity
\citep{Scoville1993a,GarciaBurillo1992} correspondes to a molecular gas mass
of $M({\rm H_2}) \approx 4.2 \times 10^9\mbox {${\rm ~M_\odot}$}$ for the \citet{Bolatto2013}
calibration of the CO-to-$\rm H_2$ conversion factor,
$X_{\rm CO} = 2 \times 10^{20} {\rm cm}^{-2} {\rm [K~km~s^{-1}]}^{-1}$.
The ratio of molecular-to-atomic gas is close to unity, $M({\rm H_2})/M(\text{H\,\textsc{I}}) \approx 0.95$,
which is higher than the median of 0.6 found for Sab galaxies \citep{Obreschkow2009a}.
Including the mass contribution from helium, the total mass of cold gas
is $1.2 \times 10^{10}$\mbox {${\rm ~M_\odot}$}.
Comparing our measured dust mass to the atomic and molecular gas mass, we find a global gas-to-dust ratio
$M_{\rm gas}/M_{\rm dust} \approx 110$,
where the uncertainty introduced by the $X_{\rm CO}$-factor is 30\% \citep{Bolatto2013}.
The gas-to-dust ratios for the SINGS galaxies range from 100 to 340 after scaling the values in
Table~4 of \citet{Draine2007a} to the same $X_{\rm CO}$ and including the helium mass. On this
scale, the gas-to-dust ratio of the Milky Way is 200 based on
the $M_{\rm dust}/M_{\rm H} \approx 0.007$ ratio given by \citet{Draine2007a}.
Based on its atomic and molecular gas content, \n891 is dustier than the median Sab galaxy.
The molecular gas is concentrated in the inner disk of \n891, so it is not surprising that
the inner disk is the dustiest region. The \text{H\,\textsc{I}}\ disk is larger than the molecular disk.
\citet{Oosterloo2007a} attribute slightly over 70\% of the total \text{H\,\textsc{I}}\ to the disk, and
our measurements of their \text{H\,\textsc{I}}\ map indicate the inner disk \text{H\,\textsc{I}}\ mass is $1.22 \times 10^9$\mbox {${\rm ~M_\odot}$},
slightly less than half the mass of the entire \text{H\,\textsc{I}}\ disk. The total gas mass of the inner (and
full) disk is then $7.4 \times 10^9$\mbox {${\rm ~M_\odot}$}\ ($9.8 \times 10^9$\mbox {${\rm ~M_\odot}$}). These disk masses do not
include the small contribution from warm ionized gas \citep{Rand1990,Dettmar1990}, but
they do include the mass contribution from helium, as described in the notes to
Table~\ref{tab:dust}.
The filament northwest of \n891 contains at least $1.6 \times 10^7$\mbox {${\rm ~M_\odot}$}\ of \text{H\,\textsc{I}}
\citep{Oosterloo2007a}. No thermal dust emission is detected from the \text{H\,\textsc{I}}\ filament.
The upper limit on the dust mass excludes gas-to-dust ratios similar to galactic
disks. If the \text{H\,\textsc{I}}\ filament is accreting gas as \citet{Oosterloo2007a} suggest, then
the lack of dust emission favors low metallicity gas. Alternatively, if the \text{H\,\textsc{I}}\
filament is fountain gas, the timescale to recycle the disk gas must be long enough
to destroy even large grains, which seems unlikely based on the large amount of dust
that persists in the CGM \citep{Menard2012a}.
The properties of the dust spur southeast of \n891 contrast sharply with
those of the \text{H\,\textsc{I}}\ filament. At the distance of \n891, the far-infrared filament contains nearly
$8 \times 10^5$\mbox {${\rm ~M_\odot}$}\ of dust. Based on the dust mass, this circumgalactic gas was once in a
galaxy, either \n891 or a satellite. Using the \text{H\,\textsc{I}}\ map \citep{Oosterloo2007a},
the \text{H\,\textsc{I}}\ mass in the dust spur is less than $2 \times 10^7$\mbox {${\rm ~M_\odot}$}. The exceptionally low
ratio of neutral hydrogen gas to dust is a puzzle. Perhaps a background source will be
found for the infrared emission. An alternative, however, is that most of the gas mass is
either molecular or ionized.
The mass ratio of \text{H\,\textsc{I}}\ to dust in the inner CGM of \n891 is indistinguishable from the
galactic disk. We find this surprising because the neutral fraction of halo gas is
around 1\% \citep{Popping2009,Bland-Hawthorn2017}. The dust was presumably made inside galaxies
and transported into the halo via outflows or stripping. Gas clouds are normally
destroyed by these processes unless special conditions are met, so we would not expect
the galactic gas to remain neutral. The inner CGM of \n891 contains more neutral
gas than the median COS (Cosmic Origins Spectrograph) Halos galaxy \citep{Prochaska2017,Das2020}. The {\it disk-like}
ratio of \text{H\,\textsc{I}}\ to dust certainly suggests that the excess of neutral gas in the inner
CGM of \n891 was removed from galaxies. We will discuss the extrapolated dust mass of the CGM
in Section~\ref{sec:extrapolate_halo}.
\begin{table}
\centering
\caption{FIR and Radio Fluxes within \n891}
\label{tab:fir_radio}
\begin{tabular}{lccccc}
\hline
Part &
log $L_{\rm FIR}$ &
log $L_{\rm 60\mu m}$ &
log $L_{\rm Cont}$ &
$q_{\rm IR}$ &
$q_{\rm 60\mu m}$ \\
&
(\mbox {${~\rm L_\odot}$}) &
(\mbox {${~\rm L_\odot}$}) &
($\rm W~Hz^{-1}$) &
&
\\
(1) &
(2) &
(3) &
(4) &
(5) &
(6)\\
\hline
Total & 10.44 & 9.64 & 22.00 & 2.45 & 2.19 \\
Inner Disk & 10.30 & 9.55 & 21.75 & 2.56 & 2.32 \\
Disk & 10.41 & 9.62 & 21.87 & 2.55 & 2.29 \\
Halo & 9.28 & 8.32 & 21.39 & 1.90 & 1.56 \\
Bubble & 8.75 & 7.82 & 20.81 & 1.96 & 1.64 \\
Dust Spur & 7.60 & 6.04 & 19.13 & 2.48 & 1.93 \\
Dust Spur1 & 7.23 & 6.08 & 18.68 & 2.56 & 2.13 \\
Dust Spur2 & 6.91 & 5.27 & 17.68 & 3.24 & 2.71 \\
Dust Spur3 & 7.12 & 5.50 & 18.78 & 2.35 & 1.76 \\
Dust Spur4 & 6.56 & 5.18 & 18.30 & 2.27 & 1.76 \\
Hotspot1 & 9.76 & 9.02 & 21.20 & 2.57 & 2.33 \\
Hotspot2 & 9.80 & 9.08 & 21.24 & 2.57 & 2.34 \\
Hotspot3 & 9.56 & 8.80 & 20.98 & 2.59 & 2.34\\
\hline
\end{tabular}
\\
Columns: (1): Part of NGC 891, (2): IR luminosity at
8~${\mu}$m\ - 1000~${\mu}$m\ from SED fit, (3): 60~${\mu}$m\ flux computed from the
SED fit, (4): radio continuum flux from the continuum map by
\citet{Oosterloo2007a}, (5) $q_{\rm IR} = {\rm log} \left( {\rm
IR}\over{3.75\times10^{12}~{\rm W~m^{-2}}} \right) - {\rm log}
\left( {S_{\rm 1.4~GHz}\over{\rm W~m^{-2}~Hz^{-1}}} \right) $ where
IR is the FIR flux and $S_{\rm 1.4~GHz}$ is flux density at 1.4 GHz.
(6): For $q_{\rm 60\mu m}$, the flux at 60~${\mu}$m\ is used. This
definition is adopted from \citet{Condon1991a,Yun2001a,Bell2003a}.
\end{table}
\section{Discussion} \label{sec:discuss}
\subsection{The Dusty Superbubble}
\label{superbubble_wind.sec}
We found an extraplanar, dusty region that is significantly warmer than the
dusty halo. We called it the dusty superbubble because its location is coincident
with thermal X-ray emission ($kT = 0.71$~keV) likely associated with an outflow.
In Section~\ref{sec:blowout} below, we show that the thermal energy of this
superbubble is sufficient for the superbubble to punch through the disk and
blowout into the halo. Blowout in \n891 is of particular interest because the
SFR surface density is a factor of three below the commonly assumed threshold
of 0.1 \mbox {$~{\rm M_\odot}$~yr$^{-1}$} kpc$^{-2}$. The fate of this outflow depends largely on the
halo gas density profile, which is reasonably well constrained in \n891. We present
this evolution in Section~\ref{sec:wind_fate} and discuss its relationship to
the recent detection of cosmic ray advection in the halo of \n891.
\subsubsection{A Galactic Wind at SFR Surface Density $< 0.1$ \mbox {$~{\rm M_\odot}$~yr$^{-1}$} kpc$^{-2}$}
\label{sec:blowout}
The first numerical simulations of superbubble blowout predicted that
a superbubble would grow to heights of one to two vertical scale heights before the extraplanar shell accelerated and
broke up \citep{MacLow1988,MacLow1989}.
Given the scale height measurements for \n891 \citep{Bocchio2016},
the height of the superbubble (2\farcm7) is 8 times the scale height of the dust/gas thick disk
(0\farcm34 in 24~${\mu}$m) and 68 times that of the dust/gas thin disk (0\farcm04 in 24~${\mu}$m).
We will therefore discuss whether the \n891 superbubble will develop into a wind. The star formation rate surface
density is elevated in the disk of \n891, similar to that in a typical disk galaxy several Gyr ago. Whereas the galaxy M82
has become the prototypical starburst outflow, we suggest that the galaxy \n891 provides a more typical example
of star-formation feedback in well-developed galactic disks.
To better understand how winds are launched, sophisticated simulations that include multiphase gas, cosmic
rays, and turbulence should be tested against the observed properties of \n891 as well as starburst galaxies.
The pioneering work of \citet{MacLow1988,MacLow1989} predicts that when a superbubble grows
to scales comparable to the gas scaleheight, the expansion velocity $V(r)$ of the shell determines
its fate. Shells moving faster than the sound speed of the ambient medium will accelerate
as they break through the disk; hydrodynamic instabilities then disrupt the shell, and a
free-flowing galactic wind develops \citep{CC85}. Shells with $V(H) \le\ c_s$ will continue
to decelerate and will eventually collapse due to disruption caused by differential rotation
and ISM turbulence. The critical rate of energy injection for blowout to occur depends
on the vertical gas density distribution.
Eqn.~5 of \citet{Strickland2004b} gives the critical power for an exponential gas distribution,
$\rho(z) = \rho_0 \exp(-z/H)$, in terms of the gas scale height $H$. The scale height of
the thin, molecular disk in \n891 is just 83~pc \citep{Scoville1993a}.
The \text{H\,\textsc{I}}\ disk has a thickness $H \approx 2.6$~kpc \citep{Oosterloo2007a}, and
warm ionized gas extends 4.5 kpc above the disk plane \citep{Dettmar1990}.
Scaling to the thick \text{H\,\textsc{I}}\ disk, the critical mechanical
power increases to
\begin{eqnarray}
L_{crit} = 2.3 \times 10^{39} {\rm ~ergs~s}^{-1}
\left( {1 {\rm ~cm}^{-3}}\over{n_0} \right)^{1/2} \times \\ \nonumber
\left( {P_0/k}\over{10^4 {\rm ~K~cm}^{-3}} \right)^{3/2}
\left( {H}\over{2.6 {\rm ~kpc}} \right)^2.
\end{eqnarray}
The rate of mechanical power injection from supernovae and stellar winds can be
estimated from the SFR. We adopt the solar metallicity models from Starburst~99
and the Salpeter initial mass function. For a star formation history with a
constant SFR, the steady-state feedback power is $L_{\rm w} = 10^{41.8}$~ergs~s$^{-1}$
for a SFR of 1\mbox {$~{\rm M_\odot}$~yr$^{-1}$} \citep{Leitherer1999}. This calibration includes only
stars in the mass range $1 \le\ {\rm M_*}/\mbox {${\rm ~M_\odot}$} \le\ 100$. To be consistent with
the mass range of $0.1 \le\ {\rm M_*}/\mbox {${\rm ~M_\odot}$} \le\ 100$ used to describe the SFR
\citep{Kennicutt1998a}, we divide by a factor of 2.55, obtaining
$L_{\rm w} = 10^{41.4}$~ergs~s$^{-1}$ for a SFR of 1\mbox {$~{\rm M_\odot}$~yr$^{-1}$}.
We have listed far-infrared luminosities for the entire galaxy and the inner disk
in Table~\ref{tab:fir_radio}. Applying $SFR (\mbox {$~{\rm M_\odot}$~yr$^{-1}$}) = 4.5 \times 10^{-44} L_{\rm FIR}
({\rm ergs~s}^{-1}~{\rm Hz}^{-1})$ \citep{Kennicutt1998a}, the rate of obscured star formation
for the entire galaxy is 4.8\mbox {$~{\rm M_\odot}$~yr$^{-1}$}. This is technically a lower limit on the SFR because
not all of the UV light from \n891 is obscured by dust. However, we obtained the FUV and NUV magnitudes
of \n891 from the GALEX nearby galaxy catalog and found that the contribution from unobscured SFR is a
relatively small, roughly $0.2$\mbox {$~{\rm M_\odot}$~yr$^{-1}$}.
The mechanical energy produced by massive stars in the inner disk is most relevant
to the evolution of the superbubble. We measure an inner disk SFR of 3.5\mbox {$~{\rm M_\odot}$~yr$^{-1}$}\ from
the 24~${\mu}$m\ emission. The available mechanical power is then $L_{\rm w} = 8.7 \times 10^{41}$~ergs~s$^{-1}$,
and $L_{\rm w} / L_{\rm crit} \approx 380$. Even a few percent of the mechanical
energy produced in the inner disk is sufficient to power the expansion of a superbubble that
will blowout of the disk of \n891.
Theoretical studies have tried to predict the critical SFR surface density
for the formation of a galactic wind. Overcoming radiative losses has been
a major hurdle when star-forming regions are embedded in realistic gas clouds.
\citet{Murray2011} considered the role of radiation pressure from massive star
clusters in launching winds and suggested 0.1\mbox {$~{\rm M_\odot}$~yr$^{-1}$}\ per kpc$^2$ as a threshold. \citet{Scannapieco2013} and \citet{Hayward2017} both connected blowout directly to the turbulence in the ISM.
Empirically, \citet{Heckman2002} suggested 0.1\mbox {$~{\rm M_\odot}$~yr$^{-1}$}\ per kpc$^2$, based largely on
nearby starburst galaxies. When they were younger, however, essentially all disk
galaxies drove massive winds \citep{Martin2012,Rubin2014}. \citet{Kornei2012}
measured SFR surface densities
for redshift $z\sim 1$ star-forming galaxies. Roughly 30\% of the outflow galaxies
has SFR surface densities below the suggested threshold of 0.1\mbox {$~{\rm M_\odot}$~yr$^{-1}$}\ per kpc$^2$.
We suggest that \n891 is an excellent nearby analog, where a
wind is developing at a
SFR surface density (SFRSD)
a factor of a few lower than the threshold
suggested by \citet{Heckman2002}. Normalizing the 24~${\mu}$m\ luminosity of the inner
disk by the area of a disk of diameter 4\farcm1, we obtain a SFRSD within the molecular
ring of $\Sigma = 0.031$\mbox {$~{\rm M_\odot}$~yr$^{-1}$}~kpc$^{-2}$ for \n891.
In comparison, for NGC 4631 (Paper I), the SFRSD is estimated to be
$\Sigma = 0.15$\mbox {$~{\rm M_\odot}$~yr$^{-1}$}~kpc$^{-2}$ as the SFR is 2.9\mbox {$~{\rm M_\odot}$~yr$^{-1}$}\ within a radius of 2.5 kpc
where almost all of the star formation is taking place, or a SFR surface density
about 5 times higher than that for NGC 891.
\citet{Hayward2017} describe the condition for blowout by a critical gas fraction,
which they find is 30\%. We showed in Section~\ref{sec:d2g} that the total gas mass
of the inner disk is $7.4 \times 10^9$\mbox {${\rm ~M_\odot}$}. We now place bounds on the stellar mass.
The IRAC 3.6~${\mu}$m\ flux of the inner disk requires $M_* \lower2pt\hbox{$\buildrel {\scriptstyle > 4 \times 10^{10}$ \mbox {${\rm ~M_\odot}$},
a lower limit due to optical depth effects. The rotation speed requires a dynamical
mass $M_{\rm Dyn} \approx 6.9 \times 10^{10}$\mbox {${\rm ~M_\odot}$}\ within a radius of $R=5.83$~kpc.
The gas fraction in the inner disk of \n891 is in the 10-16\% range, a factor of two
below the \citet{Hayward2017} threshold.
We found that the dusty superbubble is spatially coincident with hot gas ($kT = 0.71 \pm 0.04$~kev) that
is concentrated above the star-forming regions in \n891 \citep{Hodges-Kluck2018}. This soft X-ray emission
very likely comes from the interaction of a hotter wind with cold gas. The cold gas mass is roughly $3 \times 10^8$\mbox {${\rm ~M_\odot}$}\
based on the outflowing dust mass, $3 \times 10^6$\mbox {${\rm ~M_\odot}$}, and the disk gas-to-dust ratio. Hot winds have very
low emission measure and have only been directly detected in a couple of starburst galaxies \citep{Strickland2004a,Strickland2004b}. It
is therefore of significant interest that the central star-forming region in \n891 is surrounded by
diffuse hard X-ray emission \citep{Hodges-Kluck2018}, possibly related to the hot wind. Regardless of the origin
of this hard component, however, \n891 shows all the signatures of a thermally driven wind.
\subsubsection{Interaction of Wind-Driven Bubble with CGM}
\label{sec:wind_fate}
As the thermal wind plows into the CGM, it does work on the halo gas and also cools. The classic wind model
widely applied to starburst galaxies \citet{CC85} does not adequately capture this wind -- halo interaction
because it neglects gravity, radiative losses, and the density profile of the CGM. These processes
ultimately determine whether the wind cools, whether a galactic fountain forms, and whether
outflowing gas escapes the gravitational potential.
Hydrodynamic simulations have not yet spatially resolved the production and evolution of cool gas
throughout a galactic disk and the surrounding CGM, but the wind -- halo interaction
has been explored over a broad parameter range, however, using semi-analytic models
\citep{Scannapieco2002,Furlanetto2003,Samui2008,Lochhaas2018}.
We applied the \citet{Lochhaas2018} model to \n891 to gain further insight about the outflow.
The \citet{Lochhaas2018} model
extends the classic structure of interstellar bubbles \citep{Weaver1977} to circumgalactic scales.
The hot wind drives an expanding shock front that compresses the surrounding halo gas into a thin shell.
A contact discontinuity separates this shell from the wind. The wind fluid is
described by the shocked wind immediately behind the shell, an interior region of unshocked
cool wind, and the innermost region of unshocked hot wind. The results suggest that
whether adiabatic cooling of the hot wind is accompanied by radiative losses depends on
the star formation rate, halo density profile, and wind density in a non-monotonic manner.
Figure 2 of \citet{Lochhaas2018} shows the evolution of the contact discontinuity and indicates
when the shocked wind begins to cool. The figure cannot be directly applied to NGC 891, however,
because the fiducial launch radius, $r_1 = 1$~kpc, corresponds to a SFRSD $\ge\ 0.3$ \mbox {$~{\rm M_\odot}$~yr$^{-1}$} ~kpc$^{-2}$,
considerably higher than the 0.031 \mbox {$~{\rm M_\odot}$~yr$^{-1}$} ~kpc$^{-2}$ in the central disk of NGC 891. Repeating the
calculation with $r_1 = 7$~kpc and SFR=4.5\mbox {$~{\rm M_\odot}$~yr$^{-1}$}\ shows that the NGC 891 outflow lies in a sweet spot
where the shocked wind cools radiatively; there is enough mass going into the shocked wind, yet the
expansion rate is not fast enough to dropping the density rapidly (C. Lochhaas, private communication).
After 20 Myr, the contact discontinuity has reached a height of 14 kpc and decelerated to $\approx 200$\mbox {~km~s$^{-1}$}.
The mean advection speed is therefore larger than the 150\mbox {~km~s$^{-1}$} fit to the spectral index profile of the
halo radio emission \citep{Schmidt2019}. Whether this discrepancy is significant is unclear.
The semi-analytic model does not specify the physical mechanism launching the wind; it simply predicts how the
superbubble evolves for specified $\alpha$ and $\beta$. The parameter $\alpha$ describes the efficiency at
which supernova energy is transferred to the wind; the $\beta$ parameter describes the mass loading. Neither
parameter is well constrained, and both are assumed to be unity in the \citet{Lochhaas2018} calculations.
Values of $\alpha < 1$ would lower the launch velocity, however, so it is not obvious that small adjustments
could produce a better match between the velocities.
We have argued that the thermal pressure of the star-forming region is sufficient
for blowout, but cosmic ray electrons may also be important for driving the \n891 outflow. The
non-thermal radio spectral index steepens nearly linearly with height just above the
disk of \n891, a signature that the cosmic ray electrons are transported by advection
rather than diffusion \citep{Mulcahy2018,Schmidt2019}.
The non-thermal spectral index flattens at heights above 2~kpc, consistent with
adiabatic expansion dominating synchrotron losses across the region we call the superbubble.
\citet{Schmidt2019} argue that the cosmic ray advection plausibly reaches the
halo escape velocity 9 to 17 kpc above the disk.
\subsection{The Origin of the Dust spur and the \text{H\,\textsc{I}}\ Filament} \label{sec:explain_spur}
\subsubsection{The Dust Spur}
The {\it Herschel} imaging with PACS and SPIRE
confirms the presence of a dust spur discovered in the {\it Spitzer}/MIPS 24~${\mu}$m\ image.
The shape of the far-infrared/sub-mm SED (see Figure~\ref{fig:sed_fit}) is consistent with emission from a
structure associated with \n891.
The relative locations of the dust spur to the southeast and the \text{H\,\textsc{I}}\ spur to the northwest is perplexing
in part because these two streams have distinctly different composition. The dust-to-gas ratio
(see Table~\ref{tab:dust}) is high in the dust spur but very low in the \text{H\,\textsc{I}}\ spur. The dust spur is
therefore likely composed of higher metallicity gas than the \text{H\,\textsc{I}}\ spur.
The dust spur contains roughly $7.6 \times 10^5$\mbox {${\rm ~M_\odot}$}\ of dust. If the dust-to-gas ratio is similar
to the disk of \n891, then the gas mass is $\approx 9.5 \times 10^7$\mbox {${\rm ~M_\odot}$}. This mass exceeds the upper
limit on the mass of neutral hydrogen, $M_{\text{H\,\textsc{I}}} < 1.7 \times 10^7$\mbox {${\rm ~M_\odot}$}, in this region of the halo.
This apparent contradiction may be explained by additional gas mass in another phase, a higher
dust-to-gas ratio in the spur (compared to the \n891 disk), or some combination of these properties.
Within the dust spur, the local maxima in the far-infrared maps coincide with several
knots in the radio continuum map. From inspection of Figure~\ref{fig:spire_smooth}, we
defined four clumps (see Fig.~\ref{fig:region}) for which we provide photometry in Table~\ref{tab:fir_radio}.
The radio and 60~${\mu}$m\ fluxes of these clumps lie on or just above the radio-IR relation defined by
star-forming galaxies \citep{deJong1985,Condon1992,Yun2001a}. We therefore suggest that the dust
spur contains young, possibly obscured, star clusters. Since radio and 60~${\mu}$m\ flux both
underestimate SFR at low luminosities \citep{Bell2003a}, however, it is difficult to make an accurate
estimate of their SFR.
Could tidal forces exerted on the disk of NGC 891 by a satellite galaxy produce the dust spur?
We estimated the required size of a companion using the Dahari parameter \citep{Dahari1984} and
the projected separation between the tip of the dust spur and the center of NGC 891,
$ S =7\farcm88$. The tidal force on NGC 891 scales as the the mass of the companion, $F_t \propto
M_c / R^3$, where $R$ is the distance between the galaxies, but the companion mass is unknown.
\citet{Dahari1984} introduced a dimensionless parameter $Q \propto F_t$ which uses the diameter of
the primary galaxy, $D_p = 14\farcm46$
\citep{rc3}, as a scaling parameter, obtaining
\begin{equation}
F_t \propto Q \equiv\ \frac{(D_c D_p)^{1.5}}{ S^{3}},
\end{equation}
and empirically showing that a strong tidal interaction requires $Q \ge\ 1$. This argument
indicates that a satellite galaxy large enough to pull the dust spur out of the NGC 891 disk
would have a major axis $D_c \approx 4\farcm30$ or larger. This size is remarkably similar
to the extent of the dust spur, which is roughly 2\arcmin\ by 4\arcmin\ in Figure 2. Yet
the old stellar population of this hypothetical satellite is not detected.
The dust spur may be the {\it smoking gun} of a dark satellite passing through the disk of
\n891.
The Milky Way's population of high velocity clouds (HVCs) is thought to include a subpopulation that
is confined by dark matter minihalos \citep{Blitz1999}. The properties of compact, isolated HVCs are
similar in many respects to those of dwarf irreglular galaxies, but they lack a high surface brightness
stellar population \citep{Braun2000}. They have \text{H\,\textsc{I}}\ masses of $10^5 - 10^6$\mbox {${\rm ~M_\odot}$}, i.e., well below the
upper limits on the \text{H\,\textsc{I}}\ mass of the dust spur, but total masses of $10^7 - 10^8$\mbox {${\rm ~M_\odot}$}\ which are as large
as the masses of some dwarf galaxies \citep{Adams2013}. Hydrodynamical simulations of a minihalo
colliding with a disk show that the interaction pulls a coherent gas cloud out of the far side of the
disk \citep{Nichols2014,Galyardt2016,Tepper2018},
a qualitatively different result than the collision of a pure
baryonic cloud with a disk.
These structures persist for $\sim 60$~Myr \citep{Galyardt2016} or perhaps even longer
\citep{Tepper2018}, trigger
star-formation in the minihalo, but eventually deposit the majority of the HVC's gas mass to the
disk of the primary galaxy.
\subsubsection{The H I Filament}
\citet{Oosterloo2007a} considered several interpretations of the \text{H\,\textsc{I}}\ spur. They considered gas
previously ejected from the disk and now returning to the disk in a galactic fountain and found
the recycled gas had to lose significant angular momentum, perhaps to interaction from
infall from the intergalactic medium (IGM), to be consistent with the \text{H\,\textsc{I}}\ kinematics. Alternatively,
they suggested the stream could be cold gas directly accreted from the IGM or condensing out of a hot, virialized
halo. No star formation has been associated with the \text{H\,\textsc{I}}\ spur.
In our opinion, the substantial mass of the \text{H\,\textsc{I}}\ filament, $M(\text{H\,\textsc{I}}) \approx 1.6 \times 10^7$\mbox {${\rm ~M_\odot}$},
suggests a discrete event. We suggest the ram pressure from the gas halo in \n891 pushed the
gas out of a high-velocity cloud or satellite galaxy as it approached the disk.
For purposes of illustration, we have estimated the halo density $n_0 \approx 4 \times 10^{-4}$~cm$^{-3}$
at $z = 12.5$~kpc and $R= 22.7$~kpc from the fitted \text{H\,\textsc{I}}\ surface density model \citep{Oosterloo2007a}.
Models connecting the pressure to the
ratio of molecular-to-atomic hydrogen, for example, suggest $P_0 / k \lower2pt\hbox{$\buildrel {\scriptstyle <\ 3 \times 10^3$~cm$^{-3}$~K
for $\Sigma({\rm H_2}) / \Sigma(\text{H\,\textsc{I}}) \lower2pt\hbox{$\buildrel {\scriptstyle <\ 0.5$ \citep{Yim2011}.
To strip the ISM from the satellite, the satellite must approach the
disk at a velocity
\begin{eqnarray}
V_{sat} > 210 {\rm ~km~s}^{-1}
\left( {P_{sat}} \over 3 \times 10^3 {{\rm ~cm}^{-3} {\rm ~K}} \right)^{1/2} \\ \nonumber
\left( {4 \times 10^{-4} {\rm ~cm}^{-3}} \over {n_0} \right)^{1/2},
\end{eqnarray}
which is comparable to the circular velocity of \n891\ (Table
\ref{tab:ngc891})
and therefore a tenable explanation.
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{f13.pdf}
\caption{The color map shown in Figure~\ref{fig:pacs_color} is shown with contours of 24~${\mu}$m\ (left) and 160~${\mu}$m\ (middle and right) after image sharpening (see text). }
\label{fig:sharp}
\end{center}
\end{figure*}
\subsubsection{Recent Accretion}
Our analysis suggests that recent accretion events produced both the dust spur and the \text{H\,\textsc{I}}\
filament, but the nature of these events were different based on the distinct
compositions of the two structures.
The interaction of an infalling gas-rich satellite or gas stream with the CGM plausibly generated
the \text{H\,\textsc{I}}\ spur, but a collisionless component such as a dark matter minihalo was required to
pull the dust spur out the back side of the disk. These processes deliver gas to the
disk of NGC 891, thereby fueling further growth of the stellar disk.
Studies of the stellar halo of NGC 891
show evidence that accretion of satellite galaxies has formed giant streams much like those mapped
around the Milky Way \citep{Ibata2003,Yanny2003,Belokurov2007}.
\citet{Mouhcine2010} resolve structures that loop around the galaxy reaching heights of
$\approx 30$~kpc much like the rosette-shaped pattern generated by a tidally disrupting dwarf galaxy.
We note that the \text{H\,\textsc{I}}\ spur follows
one of the stellar streams where it plunges toward the disk. This alignment is most easily seen in
Figure~4 of \citet{Schulz2014} who identify seven faint satellites beyond this stellar stream.
This satellite population provides more frequent interactions with the disk than the more massive
group member UGC~1807 (1:10 mass ratio) over $ 80$~kpc northwest of \n891.
Interaction with a satellite may also explain the lopsidedness of the disk of
\n891 \citep{Oosterloo2007a}.
\subsection{Origins of Halo Dust} \label{sec:extrapolate_halo}
We have detected thermal emission from $1.5 \times 10^7$\mbox {${\rm ~M_\odot}$}\ of cool dust in the inner halo of \n891, i.e.,
the region within the ellipse drawn in Figure~\ref{fig:clean24}. We estimate the total halo dust mass
from an extrapolation of the halo surface brightness profile.
We use the 24~${\mu}$m\ profile on the west side of the disk and
assume the halo extends vertically from a base of radius 20.9~kpc. We equate the central surface brightness of
6.0 MJy~sr$^{-1}$ with a mass surface density of 0.20\mbox {${\rm ~M_\odot}$}\ pc$^{-2}$ based on Figure~\ref{fig:pix2pix},
arriving at a total halo dust mass of $4 \times 10^7$\mbox {${\rm ~M_\odot}$}.
With this extrapolation, the halo dust mass increases from 15\% to roughly 43\% of the dust mass in the thick
disk. The estimated dust mass in the halo therefore approaches the cosmic ratio of halo dust to disk dust.
\citet{Menard2012a} measured the reddening of background quasars caused by dust in \text{Mg\,\textsc{II}}~absorbers and
found that these absorbers account for most of the dust inside the virial radii of galaxies and that
the galaxy halos contain at least 52\% as much dust as do galaxy disks. While we cannot rule out the
possibility of substantially more halo dust in a cooler component at $T\sim7~\rm K$, the similarity of these
numbers suggests that the {\it Herschel} imaging has detected the main component of halo dust in \n891.
We have identified two structures, a dusty superbubble and a dust spur, that may pollute the CGM with dust.
To add further insight into exactly how dust is transported into the CGM,
we applied unsharp masking techniques to highlight the high spatial frequency structure
in the dust emission. Regardless of the exact procedure used, the results in Figure~\ref{fig:sharp} show
dust filaments emanating from the hotspots associated with the molecular ring
as well as from the central hotspot. These filaments are detected over the full diameter of the molecular ring
and extend roughly 4~kpc from the midplane. These dust filaments appear to be associated with the feedback
processes that are mixing dust into the halo; they are found over a larger region of the disk than
the base of the superbubble. It should be possible (in future work) to compare the properties of such
filaments directly to wind simulations that take the interactions of the bubbles driven by multiple
star-forming regions into account \citep{Tanner2016}.
The halo dust clearly traces a fluid that is not pristine gas. While this material most likely originated in
galaxies, the question remains whether it was primarily blown out of the disk by feedback processes or whether
tidal streams and ram pressure stripping of satellites have contributed a substantial fraction of the mass.
The newly discovered dust spur clearly shows that interactions with satellites do play some role.
The mass in the dust spur is about 25\% of that in the superbubble, however; and the filaments are also
lifting dust about 4 kpc off the disk midplane. Hence our results favor a dominant role for stellar feedback
in enriching the halo with dust.
To determine whether it is energetically plausible for feedback to lift most of the gas associated
with the halo dust out of the disk, we conservatively estimate the total mass of gas lifted into the halo
along with the dust. The dust-to-gas mass ratio of the galaxy as a whole is $M(dust)/M(gas) \approx 1 / 110$.
We estimate $M(gas) \approx 1.7 \times 10^9$\mbox {${\rm ~M_\odot}$}\ just above the disk and, very roughly for the entire halo,
$M(gas) \approx 4 \times 10^9$\mbox {${\rm ~M_\odot}$}. A lower dust-to-gas ratio would yield more halo gas associated with the dust,
but the average galactic value is most appropriate for testing the hypothesis that the dust was lifted out of the
disk by feedback. We note that $4 \times 10^9$\mbox {${\rm ~M_\odot}$}\ is a small fraction of the total
mass found in the cool CGM of galaxies \citep[$7-12\times10^{10}~\mbox {${\rm ~M_\odot}$}$]{Werk2014a}.
Following \citet{Howk1997a}, we assume an isothermal
sheet model for the vertical distribution of light and mass in \n891. Based on the similarity
of the rotation curves for \n891 and the Milky Way, we adopt the Milky Way midplane mass density $\rho_0$
\citep{Bahcall1984} as a proxy for \n891; the mass scale height $z_0$ is taken from infrared imaging
of \n891 \citep{Aoki1991}.
The potential energy, $W$, of a cloud of mass $M_{\rm cl}$ a height $z$ above the midplane is then
\begin{eqnarray}
W = 10^{52} {\rm~ergs} \left( {M_{\rm cl}}\over{10^5~\mbox {${\rm ~M_\odot}$}} \right)
\left( {z_{\rm 0}}\over{700\rm~pc}\right)^2 \\ \nonumber
\left( {\rho_{\rm 0}}\over{0.185~\mbox {${\rm ~M_\odot}$}~\rm pc^{-3}} \right)
{\rm ln} \left[ {\rm cosh} \left({z}\over{z_{\rm 0}} \right) \right].
\end{eqnarray}
The energy required to lift the $4 \times 10^9$\mbox {${\rm ~M_\odot}$}\ of \text{H\,\textsc{I}}, $\rm H_2$, and dust in the superbubble
to a height of 4.3~kpc is roughly
\begin{eqnarray}
W \approx 1.2 \times 10^{58} {\rm ~ergs}
\left( {M_{\rm cl}}\over{4 \times 10^9~\mbox {${\rm ~M_\odot}$}} \right)
\left( {z_{\rm 0}}\over{700\rm~pc}\right)^2 \\ \nonumber
\left( {\rho_{\rm 0}}\over{0.185~\mbox {${\rm ~M_\odot}$}~\rm pc^{-3}} \right).
\end{eqnarray}
For a constant SFR of 4.8\mbox {$~{\rm M_\odot}$~yr$^{-1}$}, supernovae and stellar winds
provide a mechanical power, $L_{\rm w} \approx 1.2 \times 10^{42}$~ergs~s$^{-1}$.
Assigning an efficiency $\epsilon$ for transferring this energy to the cool gas and dust,
the time required to enrich the halo with dust is roughly
\begin{eqnarray}
t \approx 3.1 \times 10^8 \epsilon^{-1} {\rm ~yr}.
\end{eqnarray}
The current rate of star formation in the disk provides enough feedback energy to lift
the dust and associated gas into the halo in just a few rotational periods only if the
feedback efficiency is very high. However, if we picture the CGM as continually recycling
disk material over a longer timescale, then feedback can easily account for most of halo dust in \n891.
For example, if the disk of \n891 was assembled at $z \lower2pt\hbox{$\buildrel {\scriptstyle <\ 2$, and the SFR has steadily declined
over the last 10.3~Gyr, then the dust can be lifted into the halo with inefficient feedback characterized by
$\epsilon \approx\ 0.03$.
\subsection{Comparison to \n4631}
It is interesting to compare the feedback in \n891 to that in \n4631, another edge-on galaxy which we
recently studied with {\it Herschel} in Paper I. The galaxies \n4631 and \n891 have
comparable \mbox {H$\alpha$}, X-ray, and radio luminosities. Linearly polarized radio continuum emission has been
detected in the halos of both galaxies and shows that magnetic field lines reach into the halos of both galaxies
\citep{Hummel1991}. The FIR to radio flux ratio for \n891 is $q_{60\mu m}=2.19$ in Table~\ref{tab:fir_radio}. The FIR luminosity at 60~${\mu}$m\ is $9.64-9.70\mbox {${~\rm L_\odot}$}$ \citep{Sanders2003,Surace2004} and radio continuum luminosity is $21.85-21.76~\rm W~m^{-2}~Hz^{-1}$ \citep{Condon2002,Murphy2009}. Thus, we estimate $q_{60\mu m}=2.03-2.19$ for \n4631. This ratio is a sensitive probe of age
during a starburst phase \citep{Bressan2002}, and the comparable values for \n4631 and \n891 indicate
both galaxies are in a post-starburst phase.
In spite of these similarities, the halos of the two galaxies show some significant
differences. The density and scale height of the warm ionized gas are a bit larger
in \n4631, and the magnetic field extends further into the halo of \n4631 \citep{Hummel1991}.
The direction of the magnetic field lines is not well constrained for either galaxy due to
poorly constrained models for Faraday rotation, but after correction the foreground
Faraday rotation the field lines are quite different in the two galaxies.
The magnetic field lines point radially outward from the disk of \n4631 while
no overall orientation is visible in \n891 where the field appears to be drawn
into the halo by more local events (especially on the south side of the disk).\footnote{
The radio continuum emission in Fig.~\ref{fig:pacs_color}
also shows an extension in total intensity away from the disk plane to the southwest (opposite the dust spur).
The linear polarization of this southwest extension is larger than the polarization in any other region
of \n891, and the polarized emission extends further into the halo here than it does above the superbubble
\citep{Hummel1991}.
Bulk motion has convected the magnetic field outwards, and the adiabatic losses of the electrons
may explain the changes in spectral index reported by \citet{Hummel1991}. Whether there is a direct connection
between this inferred velocity gradient and the dust spur remains unclear.}
Overall, these results suggest the galactic wind is stronger in \n4631 than in \n891.
Stronger feedback in \n4631 would be expected based on its higher SFR surface
density and gas fraction.
{ The SFRSD of \n4631 is about 5 times larger than that of \n891 as estimated in Section~\ref{superbubble_wind.sec}.}
It is also clear that \n4631 has a higher gas fraction than \n891 based on its higher \text{H\,\textsc{I}}\ mass, lower dynamical mass, and later spiral type
\citep{rc3}.
It is therefore quite interesting to compare the dust properties of these two halos.
We find the { measured} dust mass of \n891 is 2.6 times higher than \n4631 in Paper I.
We attribute this result to the larger size and mass of \n891.
The superbubble in \n4631 is slightly smaller and contains a factor of two less dust mass
than the \n891 superbubble.
The temperature of the dust, however, differs significantly between the two superbubbles.
The temperature of the dusty superbubble is significantly higher in \n4631, $T_{\rm dust} = 24.46 \pm 0.77$~K.
When we include the 100~${\mu}$m\ data point from the HEDGES data, the superbubble temperature for \n891
increases by $\sim 1$ K which is still 3~K below the the temperature of the superbubble in \n4631.
When we take the 70~${\mu}$m\ data point as an upper limit, the temperature decreases slightly.
Therefore, the temperature difference of superbubble for \n4631 and \n891 cannot be alleviated by
different SED fitting processes.\footnote{
For the SED modeling, we used 5 bands, 70~${\mu}$m, 160~${\mu}$m, 250~${\mu}$m, 350~${\mu}$m,
and 500~${\mu}$m\ while Paper I adopted 100~${\mu}$m\ in addition to
the 5 bands. Also, in Paper I,
the 70~${\mu}$m\ data point was considered as an upper limit.
The higher temperature of the SED fit for \n4631 is driven
by the data point at 100~${\mu}$m.
When we include the 100~${\mu}$m\ data point from
the HEDGES data, the temperature of all components of \n891 increases by $\sim 1$~K.
}
Dust in galaxy halos can be heated by the evolved stellar population
or processes associated with the stellar feedback. Using B and V band photometry,
\citet{Ann2011} measured a scale height for the thick stellar disk of \n4631 of
$1.33 \pm 0.25$~kpc and $1.25 \pm 0.28$~kpc, and \citet{Morrison1997} report an
R band scale height of $1.5-2.5$~kpc for \n891. It therefore seems unlikely
that the stellar halo of \n891 is any less effective at heating dust than
that in \n4631.
The SFR surface density of \n4631 is only slightly higher than that of \n891,
so an attempt to explain the difference in superbubble temperatures would need to appeal to
the lower dust mass and smaller size of the superbubble in \n4631. It would be interesting
to explore whether these properties of the \n4631 superbubble reflect a younger age or
the more ordered magnetic field of the halo. The increased temperature of the superbubble might
reflect the larger pressure required to push these field lines out of the superbubble's path.
\section{Conclusions} \label{sec:conclusions}
We present the results of our {\it Spitzer}/MIPS and {\it Herschel}/PACS
GO observations of \n891. The angular resolution and
sensitivity allow us to resolve the distribution of dust in the disk and halo.
We place these components in the context of the stellar component and multi-phase ISM
using supplementary observations from the radio to the X-ray. We obtain the
following insight into the disk - halo interaction.
\begin{itemize}
\item The MIPS image at 24~${\mu}$m\ shows a dust spur extending 20 kpc southeast of the galaxy.
The PACS and SPIRE images confirm the presence of this feature. Fitting a MBB model to the
SED indicates a dust mass of $7.6 \times 10^5$\mbox {${\rm ~M_\odot}$}, which is surprisingly large compared
to the upper limits on the \text{H\,\textsc{I}}\ gas mass. The high dust-to-gas ratio is inconsistent with a
primordial origin (e.g., a cold stream of infalling material) and indicates the dust was
likely pulled out of \n891 by a satellite interaction. Its composition differs distinctly
from the \text{H\,\textsc{I}}\ spur to the northwest where we do not detect thermal dust emission.
The local maxima in the far-infrared surface brightness coincide with knots of radio
continuum emission, perhaps a sign that they are young star-forming regions. The absence
of optical or near-infrared emission from the dust spur suggests the colliding satellite
had little stellar mass, suggesting that the dust spur may be the remnant of a minihalo
collision with the disk.
\item
The PACS 70/160~${\mu}$m\ ratio map draws attention to a superbubble extending over 7 kpc above
the midplane on its western side. We find a close correspondence between the superbubble and
the thermal X-ray emission suggesting that the thermal pressure of hot gas drives the expansion.
We argue that the superbubble is likely accelerating as it breaks
through the disk because its height exceeds the pressure scale height of not only the thin
disk but also the gaseous halo. The feedback associated with the star formation activity
within the inner disk appears to be sufficient to launch this galactic wind. The SFR surface
density of 0.03\mbox {$~{\rm M_\odot}$~yr$^{-1}$}\ kpc$^{-2}$ and gas fraction ($\lower2pt\hbox{$\buildrel {\scriptstyle <\ 10\% $) suggest somewhat lower
threshold values for driving winds than do the values measured previously in starburst galaxies.
\item
This disk contains regions of hot dust associated with the galactic center and the molecular ring.
Image sharpening techniques reveal dust filaments rising roughly 4 kpc above these regions. The
dust temperature is higher in the inner disk than the halo, and the mass surface density of the
dust is centrally concentrated.
\item
{ The halo dust component is detected 10~kpc or more above the entire stellar disk with the measured halo dust mass of $M_{\rm dust,halo} = 1.5\times10^{7}$\mbox {${\rm ~M_\odot}$}.} A rough extrapolation of the dust mass surface
density suggests $\approx 4 \times 10^7$\mbox {${\rm ~M_\odot}$}\ of dust has been lifted into the halo carrying
with it a gas mass roughly 100 times larger. This mass flux accounts for roughly 10\%
of the gas mass estimated to reside in the CGM of similar galaxies.
\end{itemize}
Because \n891 is a very close analog of the Milky Way in respect to many of its properties,
we believe these results provide a viable picture of how dust is deposited in the halos of
typical galaxies. The elevated SFR and molecular gas fraction with respect to the Milky
Way would perhaps most closely describe the Milky Way in its recent past when its SFR
was
higher. In this sense, \n891 provides a direct image of the
emission from dust and multi-phase gas in the inner halo of a galaxy where the baryon
budget of the CGM is well constrained observationally \citep{Hodges-Kluck2018,Das2020}.
\section*{Acknowledgements}
We thank Cassi Lochhaas for discussions about the circumgalactic bubble which
clarified how its structure depends on SFR surface density,
and we thank an anonymous referee
for suggesting we apply the Dahari parameter.
This work was supported in part by the National Science Foundation (NSF) under
AST-1817125 (CLM) and ASTR1009583 (SV), JPL Awards 1276783 and
1434779, and NASA grants NHSC/JPL RSA 1427277, 1454738
(SV and MM), and ADAP NNX16AF24G (SV). Some of the work was completed
at the Aspen Center for Physics, and accompanying support from the NSF through
PHY-1066293 is gratefully acknowledged.
We deeply regret that the co-author (CE) who provided the
\mbox {H$\alpha$}\ image and made a significant contribution to the analysis of the
{\it Spitzer} data did not live to see the article published.
{\it Herschel} is an ESA
space observatory with science instruments provided by
European-led Principal Investigator consortia and with important
participation from NASA.
\section*{Data Availability}
The data underlying this article will be shared on reasonable request
to Dr. Joo Heon Yoon.
\bibliographystyle{mnras}
| e4a16c0b509bff1cc73b1f7980105ec8ac59aef2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:intro}
\subsection{Overview}
This article presents a method for using computations with quaternion algebras over totally real fields to determine explicit upper bounds for the systolic lengths of certain Riemann surfaces, namely the Galois-Belyi curves constructed by Clark and Voight~\cite{CV/19}. These are also referred to as triangular modular curves. The most celebrated members of this family are the congruence Hurwitz curves. We generalize the treatment in~\cite{KKSV/16} of a family of Riemann surfaces related to the Bolza surface and extend the methods of~\cite{KKSV/16} to some non-arithmetic surfaces.
Let $X$ be a compact Riemann surface. Our main object of interest, the systolic length of $X$, is denoted $\mathrm{sys}(X)$ and is the minimal length of a non-contractible closed geodesic, or systole, of $X$.
Let $\mathcal{H}$ be the upper half plane with the usual hyperbolic metric. The group $\mathrm{SL}_2(\mathbb{R})$, via its quotient $\mathrm{PSL}_2(\mathbb{R})$, acts on $\mathcal{H}$ by M\"{o}bius transformations. Every hyperbolic Riemann surface arises as a quotient $X_\Gamma = \Gamma \backslash \mathcal{H}$, where $\Gamma \leq \mathrm{PSL}_2(\mathbb{R})$ is a discrete subgroup (i.e.~a Fuchsian group) acting on $\mathcal{H}$ without fixed points. As we will see in Section~\ref{sec:traces} below, bounding the systolic length $\mathrm{sys}(X_\Gamma)$ amounts to finding an upper bound for the minimal trace of a non-trivial element of $\Gamma$.
If $\Gamma$ is a non-arithmetic Fuchsian group, there is generally no natural construction of congruence subgroups of $\Gamma$. However, if $\Gamma = \overline{\Delta}(a,b,c) = \langle x, y \rangle / (x^a = y^b = (xy)^c = 1)$ is a hyperbolic triangle group, then Clark and Voight~\cite{CV/19} define congruence subgroups $\Gamma(I) \trianglelefteq \Gamma$ of finite index, where $I$ runs over ideals of a Dedekind domain associated to $\Gamma$, and describe the quotient $\Gamma / \Gamma(I)$. If $\Gamma = \overline{\Delta}(2,3,7)$, then the Riemann surfaces $X_{\Gamma(I)}$ are congruence Hurwitz curves. For all but finitely many triples $(a,b,c)$, the group $\overline{\Delta}(a,b,c)$ is non-arithmetic.
To estimate $\mathrm{sys}(X_{\Gamma(I)})$, where $\Gamma$ is a hyperbolic triangle group, we compute a finite set of Schreier generators of $\Gamma(I)$. We then embed $\Gamma(I)$ into $\mathrm{PSL}_2(\mathbb{R})$ explicitly and consider the minimum among the traces of the images of the Schreier generators.
A brief overview of the paper follows.
Section~\ref{sec:traces} describes the connection between the systolic length of $X_\Gamma$ and the traces of elements of $\Gamma$.
Section~\ref{sec:triangle.groups} is an exposition of work by previous authors, whose aim is to present the necessary background material in a form useful for our applications. Some properties of triangle groups and of their congruence subgroups are recalled. Proposition~\ref{pro:alpha.beta.fulldelta}, which restates a result of Takeuchi~\cite{Takeuchi/77JFS}, provides an embedding $\Gamma \hookrightarrow \mathcal{O}^1 / \{ \pm 1 \}$, where $\mathcal{O}^1$ is the group of elements of reduced norm one in an explicit order $\mathcal{O}$ of an explicit quaternion algebra over a totally real field. The explicit quaternion order of Proposition~\ref{pro:presentation.a} realizes $\Gamma$ as a Fuchsian group when it is arithmetic. This construction generalizes ones used, in special cases, in~\cite{KKSV/16, KSV/11} and is useful to us in non-arithmetic cases as well. Proposition~\ref{pro:counting.normal.subgroups}, using Macbeath's classification~\cite{Macbeath/69} of subgroups of $\mathrm{PSL}_2(\mathbb{F}_q)$ and work of Clark and Voight~\cite{CV/19}, bounds the number of normal subgroups of $\Gamma$ admitting certain finite quotients. These bounds simplify the computations of Section~\ref{sec:computations} in some cases.
Section~\ref{sec:semiarithmetic} uses recent work of Cosac and D\'{o}ria~\cite{CosacDoria/20} to prove (Theorem~\ref{thm:semiarithmetic}) that $\mathrm{sys}(X_{\Gamma(I)})$ grows at least logarithmically with respect to the genus of $X_{\Gamma(I)}$ when at least two of $a,b,c$ are odd.
Section~\ref{sec:computations} presents our method of computation and some examples. Directions to a Github repository of our code are provided. In particular, we bound the systolic lengths of curves associated to congruence subgroups for ideals of small norm in the arithmetic triangle groups $\overline{\Delta}(2,3,7)$, $\overline{\Delta}(2,3,8)$, $\overline{\Delta}(2,3,12)$, $\overline{\Delta}(2,7,7)$, as well as in the non-arithmetic $\overline{\Delta}(3,3,10)$. The first two of these examples were previously considered by Katz, Schaps, and Vishne~\cite{KSV/07, KSV/11} and by Katz, Katz, the first author, and Vishne~\cite{KKSV/16}, respectively; we recover constructions from those papers. Our upper bound is equal to the exact value of $\mathrm{sys}(X_{\Gamma(I)})$ in all cases, among the examples considered, where this value is known.
\subsection{Traces and geodesic lengths} \label{sec:traces}
If $\gamma \in \mathrm{PSL}_2 (\mathbb{R})$, then the trace $\tr \gamma$ is defined up to sign, so the absolute value $| \tr \gamma |$ makes sense; the same is true for the eigenvalues of $\gamma$. If $ | \tr \gamma | > 2$, then $\gamma$ is called hyperbolic; in this case, a simple calculation shows that $\gamma$ has no fixed points in the interior of $\mathcal{H}$ and two distinct fixed points on the boundary of $\mathcal{H}$. If the Fuchsian group $\Gamma \leq \mathrm{PSL}_2(\mathbb{R})$ is torsion-free, then there is a bijection between conjugacy classes of hyperbolic elements of $\Gamma$ and closed geodesics of $X_\Gamma$. Indeed, if $\gamma \in \Gamma$ is hyperbolic, the unique geodesic of $\mathcal{H}$ connecting the two fixed points of $\gamma$ is called the axis of $\gamma$; clearly it is preserved by the action of $\gamma$. Fix a point $P$ on the axis of $\gamma$. The geodesic of $\mathcal{H}$ passing through $P$ and $\gamma P$ descends to a closed geodesic $\delta_\gamma$ of $X_\Gamma$. One checks that this construction is independent of the choice of $P$ and depends only on the conjugacy class of $\gamma$. Clearly, any closed geodesic of $X_\Gamma$ arises in this way. Moreover, the length $\ell(\delta_\gamma)$ of the geodesic $\delta_\gamma$ satisfies $e^{\ell(\delta_\gamma)/2} = | \lambda_\gamma |$, where $\lambda_\gamma$ is the unique eigenvalue of $\gamma$ such that $ | \lambda_\gamma | > 1$. It follows that
$$ 2 \cosh \frac{\ell(\delta_\gamma)}{2} = e^{\ell(\delta_\gamma)/2} + e^{- \ell (\delta_\gamma)/2} = | \lambda_\gamma | + | \lambda_\gamma |^{-1} = | \tr \gamma |.$$
Hence, letting $\mathrm{Hyp}(\Gamma)$ denote the set of hyperbolic elements of $\Gamma$, we have
\begin{equation*}
\mathrm{sys}(X_\Gamma) = \min_{\gamma \in \mathrm{Hyp}(\Gamma)} \ell(\delta_\gamma) = \min_{\gamma \in \mathrm{Hyp}(\Gamma)} 2 \arcosh \frac{| \tr \gamma |}{2}.
\end{equation*}
Bounding the systolic length of $X_\Gamma$ thus amounts to bounding the traces of hyperbolic elements of the Fuchsian group $\Gamma$. In particular, if $\Gamma^\prime \leq \Gamma$, then $\mathrm{sys}(X_{\Gamma^\prime}) \geq \mathrm{sys}(X_\Gamma)$.
\section{Triangle groups} \label{sec:triangle.groups}
\subsection{Definition}
Suppose that $a,b,c \in \mathbb{Z} \cup \{ \infty \}$ satisfy $2 \leq a,b,c$. The triangle group $\overline{\Delta}(a,b,c)$ is given by the presentation
\begin{equation} \label{equ:triangle.group}
\overline{\Delta}(a,b,c) = \langle x, y \rangle / (x^a = y^b = (xy)^c = 1).
\end{equation}
The extended triangle group $\Delta(a,b,c)$ is defined as
\begin{equation} \label{equ:delta.presentation}
\Delta(a,b,c) = \langle x, y, z, -1 \rangle / \left({x^a = y^b = z^c = xyz = -1, (-1)^2 = 1 \atop -1 \in Z(\Delta(a,b,c))}\right).
\end{equation}
Any permutation of $a,b,c$ in either presentation gives rise to an isomorphic group, so we make the convention that $2 \leq a \leq b \leq c$.
The triple $(a,b,c)$ is classified as spherical, Euclidean, or hyperbolic if the quantity $\frac{1}{a} + \frac{1}{b} + \frac{1}{c}$ is greater than, equal to, or less than $1$, respectively. If $(a,b,c)$ is a hyperbolic triple, then it has been known at least since the 1930's~\cite[\S1]{Petersson/37} that there is an embedding $\iota: \Delta(a,b,c) \hookrightarrow \mathrm{SL}_2(\mathbb{R})$ determined by
\begin{eqnarray} \label{equ:iota.embedding}
\iota (x) & = & \left( \begin{array}{cc} \cos \frac{\pi}{a} & \sin \frac{\pi}{a} \\ - \sin \frac{\pi}{a} & \cos \frac{\pi}{a} \end{array} \right), \\ \nonumber
\iota (y) & = & \left( \begin{array}{cc} 1 & 0 \\ 0 & t \end{array} \right)^{-1}
\left( \begin{array}{cc} \cos \frac{\pi}{b} & \sin \frac{\pi}{b} \\ - \sin \frac{\pi}{b} & \cos \frac{\pi}{b} \end{array} \right)
\left( \begin{array}{cc} 1 & 0 \\ 0 & t \end{array} \right),
\end{eqnarray}
where $t$ is a root of the quadratic equation
$$ t^2 - 2 \left( \frac{\cos \frac{\pi}{a} \cos \frac{\pi}{b} + \cos \frac{\pi}{c}}{\sin \frac{\pi}{a} \sin \frac{\pi}{b}} \right) t + 1 = 0;$$
the roots are real when $(a,b,c)$ is hyperbolic. The induced embedding $\overline{\Delta}(a,b,c) \hookrightarrow \mathrm{PSL}_2(\mathbb{R})$ realizes $\overline{\Delta}(a,b,c)$ as a Fuchsian group. If $b < \infty$, then Takeuchi~\cite[Proposition~1]{Takeuchi/77} showed that any embedding of $\Delta(a,b,c)$ into $\mathrm{SL}_2(\mathbb{R})$ is conjugate to $\iota$. In particular, if $\Gamma \leq \mathrm{SL}_2(\mathbb{R})$ satisfies $\Gamma \simeq \Delta(a,b,c)$, then $\mathrm{sys}(X_\Gamma)$ depends only on the triple $(a,b,c)$.
Consider the full triangle group $\widetilde{\Delta}(a,b,c)$ generated by reflections across the sides of a triangle with angles $\frac{\pi}{a}$, $\frac{\pi}{b}$, $\frac{\pi}{c}$. It is well-known to have the presentation
\begin{equation} \label{equ:full.triangle.group}
\widetilde{\Delta}(a,b,c) = \langle \widetilde{x}, \widetilde{y}, \widetilde{z} \rangle / (\widetilde{x}^2 = \widetilde{y}^2 = \widetilde{z}^2 = (\widetilde{x} \widetilde{y})^a = (\widetilde{y} \widetilde{z})^b = (\widetilde{z} \widetilde{x})^c = 1 ).
\end{equation}
It is easy to see that $\overline{\Delta}(a,b,c)$ embeds in $\widetilde{\Delta}(a,b,c)$ as the subgroup of index two of orientation-preserving transformations. An embedding is given by $(x,y) \mapsto (\widetilde{x} \widetilde{y}, \widetilde{y} \widetilde{z}, \widetilde{z} \widetilde{x})$.
\subsection{Quaternion algebras}
In a series of papers in the 1970's~\cite{Takeuchi/69, Takeuchi/75, Takeuchi/77, Takeuchi/77JFS}, Takeuchi determined~\cite[Theorem~3]{Takeuchi/77} with computer assistance that there are precisely $85$ triples $(a,b,c)$ for which $\overline{\Delta}(a,b,c)$ is arithmetic. Moreover, he provided an explicit construction of orders in quaternion algebras that realize this arithmeticity. In this section, we give a brief overview of Takeuchi's work, stated in the explicit form needed for our computations. As is standard, for a field $F$ we denote by $\left\langle \frac{x,y}{F} \right\rangle$ the $F$-algebra spanned by $\{1, i, j, ij \}$, with multiplication determined by the relations $i^2 = x, j^2 = y, ij = - ji$. If $F$ is a totally real number field, then this algebra ramifies at an embedding $v : F \hookrightarrow \mathbb{R}$ if and only if $v(x) < 0$ and $v(y) < 0$.
Let $\Gamma \leq \mathrm{SL}_2(\mathbb{R})$ be any discrete subgroup whose image in $\mathrm{PSL}_2(\mathbb{R})$ has finite covolume. Let $F_\Gamma$ be the field generated over $\mathbb{Q}$ by the traces of the elements of $\Gamma$. Under the hypotheses above, $F_\Gamma$ is a number field; by construction, it is equipped with a distinguished embedding $v_0: F_\tau \hookrightarrow \mathbb{R}$. By Propositions~2 and~3 of~\cite{Takeuchi/69}, the $F_\Gamma$-vector subspace $F_\Gamma[\Gamma] \subseteq M_2(\mathbb{R})$ is a quaternion algebra over $F_\Gamma$; if the traces of all elements of $\Gamma$ are algebraic integers, then $\mathcal{O}_{F_\Gamma} [\Gamma]$ is an order of $F_\Gamma[\Gamma]$.
We assume henceforth that $\tau = (a,b,c)$ is a hyperbolic triple with $c < \infty$, since this hypothesis is needed for the explicit presentations of Propositions~\ref{pro:alpha.beta.fulldelta} and~\ref{pro:presentation.a} below. See~\cite[\S5]{CV/19} for some treatment of the case $c = \infty$. If we take $\Gamma = \iota(\Delta(\tau)) \leq \mathrm{SL}_2(\mathbb{R})$, where $\iota$ is the embedding of~\eqref{equ:iota.embedding}, then
\begin{equation*}
F_\Gamma = F_{\iota(\Delta(\tau))} = \mathbb{Q} \left( \cos \frac{\pi}{a}, \cos \frac{\pi}{b}, \cos \frac{\pi}{c} \right).
\end{equation*}
We denote this field by $F_\tau$.
It is clear from~\cite[Remark~5.24]{CV/19} that the quaternion algebra $B_\tau = F_\Gamma [\Gamma] = F_{\iota(\Delta(\tau))}[\iota(\Delta(\tau))]$ has the presentations
\begin{equation} \label{equ:b.presentation}
B_\tau = \left\langle \frac{4 \cos^2 \left( \frac{\pi}{s} \right) - 4, \delta}{F_\tau} \right\rangle,
\end{equation}
where $s$ is any element of $\{a,b,c \}$ and
\begin{equation*}
\delta = 4 \left( \cos^2 \frac{\pi}{a} + \cos^2 \frac{\pi}{b} + \cos^2 \frac{\pi}{c} + 2 \cos \frac{\pi}{a} \cos \frac{\pi}{b} \cos \frac{\pi}{c} - 1 \right).
\end{equation*}
\begin{rem} \label{rmk:chebyshev}
Observe that $2 \cos \frac{\pi}{n} = \zeta_{2n} + \zeta_{2n}^{-1}$, where $\zeta_{2n}$ is a suitable primitive $2n$-th root of unity. Hence $2 \cos \frac{\pi}{n}$ is an algebraic integer. Consequently, $ \delta \in \mathcal{O}_{F_{\tau}}$; this is the reason for the factor of $4$ in the definition of $\delta$. Note that $v_0(\delta) > 0$ whenever $\tau$ is hyperbolic.
Moreover, for any $m \in \mathbb{N}$ observe that $\cos \frac{m \pi}{n} = T_m \left( \cos \frac{\pi}{n} \right)$, where $T_m(x) \in \mathbb{Z}[x]$ is the Chebyshev polynomial of degree $m$. Similarly, if $(m,n) = 1$, then $\cos \frac{\pi}{n}$ can be expressed as a polynomial in $\cos \frac{m \pi}{n}$. It follows that $F_\tau = \mathbb{Q} \left( \cos \frac{m_a \pi}{a}, \cos \frac{m_b \pi}{b}, \cos \frac{m_c \pi}{c} \right)$ for any $m_a, m_b, m_c$ satisfying $(m_a, a) = (m_b, b) = (m_c, c) = 1$.
\end{rem}
\begin{pro} \label{pro:alpha.beta.fulldelta}
Consider the elements $\alpha = 1 \cdot \iota(x)$ and $\beta = 1 \cdot \iota(y)$ of $B_\tau$, where $x,y \in \Delta(\tau)$ are as in~\eqref{equ:delta.presentation}. Then $\{1, \alpha, \beta, \alpha \beta \}$ is an $\mathcal{O}_{F_\tau}$-basis of the order $\mathcal{O}_\tau = \mathcal{O}_{F_\tau} [\iota(\Delta(\tau))] \subset B_\tau$. Moreover, there exist elements $i,j \in B_\tau$ providing
the presentation $B_\tau = \left\langle \frac{4 \cos^2 \left( \frac{\pi}{a} \right) - 4, \delta}{F_\tau} \right\rangle$, for which
\begin{eqnarray*}
\alpha & = & \cos \frac{\pi}{a} + \frac{1}{2} i \\
\beta & = & \cos \frac{\pi}{b} - \frac{\cos \frac{\pi}{a} \cos \frac{\pi}{b} + \cos \frac{\pi}{c}}{2 \left( \cos^2 \left( \frac{\pi}{a} \right) - 1 \right)} i + \frac{1}{4 \left( \cos^2 \left( \frac{\pi}{a} \right) - 1 \right)} ij.
\end{eqnarray*}
\end{pro}
\begin{proof}
The first claim is established in the discussion preceding~\cite[Lemma~5.4]{CV/19}. The second follows easily from~\cite[Remark~5.24]{CV/19}, noting that the third element in the list given there of elements of $B_\tau$
realizing the presentation above should read $(\lambda_{2a}^2 - 4)\delta_b + (\lambda_{2a} \lambda_{2b} + 2 \lambda_{2c})\delta_a - (\lambda_{2a}^2 \lambda_{2b} + \lambda_{2a} \lambda_{2c} - 2 \lambda_{2b})$ in the notation of~\cite{CV/19}.
\end{proof}
\begin{rem} \label{rmk:embedding}
By construction, the elements $\alpha, \beta \in B_\tau$ have reduced norm $1$, inducing embeddings $\Delta(\tau) \hookrightarrow (\mathcal{O}_\tau)^1$ and $\overline{\Delta}(\tau) \hookrightarrow (\mathcal{O}_{\tau})^1 / \{ \pm 1 \}$; here $\mathcal{O}_\tau^1$ is the group of elements of $\mathcal{O}_\tau$ of reduced norm $1$. Since $v_0(\delta) > 0$, the quaternion algebra $B_\tau$ splits at the distinguished embedding $v_0$. Thus induces an embedding $\Delta(\tau) \hookrightarrow (\mathcal{O}_\tau)^1 \subseteq (B_\tau)^1 \stackrel{v_0}{\hookrightarrow} \mathrm{SL}_2(\mathbb{R})$, which we denote $\varepsilon_0$.
\end{rem}
Since the reduced traces of $\alpha$, $\beta$, and $\alpha \beta$ are $2 \cos \frac{\pi}{a}$, $2 \cos \frac{\pi}{b}$, and $-2 \cos \frac{\pi}{c}$, respectively, we verify that $\mathcal{O}_\tau$ is integral. Indeed,
$$ \alpha \beta = - \cos \frac{\pi}{c} - \frac{\cos \frac{\pi}{b} + \cos \frac{\pi}{a} \cos \frac{\pi}{c}}{2 \left( \cos^2 \frac{\pi}{a} - 1 \right)} i + \frac{1}{2} j + \frac{\cos \frac{\pi}{a}}{4 \left( \cos^2 \frac{\pi}{a} - 1 \right)} ij.$$
We collect some facts about the order $\mathcal{O}_{\tau}$.
\begin{pro}[{\cite[Lemmas~5.4 and~5.5]{CV/19}}] \label{pro:discriminant}
The discriminant of $\mathcal{O}_{\tau}$ is the principal $\mathcal{O}_{F_\tau}$-ideal generated by $ \delta$. Let $\mathfrak{P} \vartriangleleft \mathcal{O}_{F_\tau}$ be a prime ideal such that $\mathfrak{P} | \delta$. Then $\mathfrak{P} | 2abc$. Moreover, if the triple $\tau = (a,b,c)$ is not of the form $(mk, m(k+1), mk(k+1))$ for any $m,k \in \mathbb{N}$, then $\mathfrak{P} | abc$.
\end{pro}
Even if $\overline{\Delta}(\tau)$ is arithmetic, the quaternion algebra $B_\tau$ need not be split at exactly one infinite place of the totally real field $F_\tau$. Define the subgroup $\Delta^{(2)}(\tau) \leq \Delta(\tau)$ generated by $-1$ and by the set $\{ \gamma^2 : \gamma \in \Delta(\tau) \}$. This is a normal subgroup satisfying~\cite[(5.9)]{CV/19}
$$ \Delta(\tau) / \Delta^{(2)}(\tau) \simeq \begin{cases}
1 &: \, \text{at least two of} \, a,b,c \, \text{are odd,} \\
\mathbb{Z} / 2\mathbb{Z} &: \text{one of} \, a,b,c \, \text{is odd,} \\
\mathbb{Z} / 2\mathbb{Z} \times \mathbb{Z} / 2\mathbb{Z} &: a,b,c \, \text{are all even or $\infty$.}
\end{cases}
$$
Now repeat the previous construction for the group $\Gamma = \iota(\Delta^{(2)}(\tau)) \leq \mathrm{SL}_2(\mathbb{R})$. By the union of~\cite[Proposition~4]{Takeuchi/75} and~\cite[Proposition~5]{Takeuchi/77}, the trace field is then
\begin{equation} \label{equ:def.etau}
E_\tau := \mathbb{Q} \left( \cos^2 \frac{\pi}{a}, \cos^2 \frac{\pi}{b}, \cos^2 \frac{\pi}{c}, \cos \frac{\pi}{a} \cos \frac{\pi}{b} \cos \frac{\pi}{c} \right),
\end{equation}
which we view as a subfield of $\mathbb{R}$.
Let $A_\tau$ denote the quaternion algebra $E_\tau[\iota(\Delta^{(2)}(\tau))]$. Write $\mathcal{Q}_\tau$ for the quaternion order $\mathcal{O}_{E_\tau}[\iota(\Delta^{(2)}(\tau))] \subset A_\tau$. An explicit description of $A_\tau$ and $\mathcal{Q}_\tau$ is given by
the following analogue of Proposition~\ref{pro:alpha.beta.fulldelta}.
\begin{pro} \label{pro:presentation.a}
Consider the elements $\gamma_1 = 1 \cdot \iota(y^2)$ and $\gamma_2 = 1 \cdot \iota(z^2)$ of $A_\tau$, where $y,z \in \Delta(\tau)$ are as in~\eqref{equ:delta.presentation}. Then $\{1, \gamma_1, \gamma_2, \gamma_1 \gamma_2 \}$ is an $\mathcal{O}_{E_\tau}$-basis of the order $\mathcal{Q}_\tau = \mathcal{O}_{E_\tau}[\iota(\Delta^{(2)}(\tau))] \subset A_\tau$. Moreover, there exist elements $i,j \in A_\tau$ realizing the presentation
\begin{equation*}
A_\tau \simeq \left \langle \frac{16 \cos^2 \frac{\pi}{b} \left( \cos^2 \left( \frac{\pi}{b} \right) - 1 \right),
16 \delta \cos^2 \frac{\pi}{b} \cos^2 \frac{\pi}{c}}{E_\tau} \right \rangle
\end{equation*}
and satisfying
\begin{eqnarray*}
\gamma_1 & = & \cos \frac{2 \pi}{b} + \frac{i}{2} \\
\gamma_2 & = & \cos \frac{2 \pi}{c} + \frac{\cos \frac{2 \pi}{b} + \cos \frac{2 \pi}{b} \cos \frac{2 \pi}{c} + \cos \frac{2 \pi}{c} + 4 \cos \frac{\pi}{a} \cos \frac{\pi}{b} \cos \frac{\pi}{c} + 1}{2 \left( 1 - \cos^2 \frac{2 \pi}{b} \right)} i + \frac{1}{4 \left( 1 - \cos^2 \frac{2 \pi}{b} \right)} ij.
\end{eqnarray*}
\end{pro}
\begin{proof}
All this is stated explicitly in~\cite[Proposition~2]{Takeuchi/77JFS} (where $\gamma_1$ and $\gamma_2$ are denoted $\gamma_2^2$ and $\gamma_3^2$, respectively) except for the expressions for $\gamma_1$ and $\gamma_2$ as linear combinations of $1, i, j, ij$. These are obtained by inverting the explicit transformation given in the proof of~\cite[Proposition~2]{Takeuchi/77JFS} and observing the following identities, where $c_1$, $c_2$, $c_3$ are as in~\cite{Takeuchi/77JFS}:
\begin{alignat*}{3}
c_1 & = & &\frac{1}{2} \mathrm{tr} \, \iota(y^2) & = &\cos \frac{2 \pi}{b} \\
c_2 & = & &\frac{1}{2} \mathrm{tr} \, \iota(z^2) & = &\cos \frac{2 \pi}{c} \\
c_3 & = & &\frac{1}{2} \mathrm{tr} \, \iota(y^2 z^2) & = &1 - 2 \cos^2 \frac{\pi}{b} - 4 \cos \frac{\pi}{a} \cos \frac{\pi}{b} \cos \frac{\pi}{c} - 2 \cos^2 \frac{\pi}{c}. \qedhere
\end{alignat*}
\end{proof}
It is immediate that $\mathrm{tr}(\mathcal{Q}_\tau) \subseteq \mathcal{O}_{E_\tau}$. Therefore, $\overline{\Delta}(\tau)$ is \emph{semi-arithmetic}~\cite[Definition~3]{SchallerWolfart/00}.
Takeuchi shows~\cite[Theorem~1]{Takeuchi/77} that $\overline{\Delta}(\tau)$ is arithmetic if and only if the finite-index subgroup $\overline{\Delta}^{(2)}(\tau)$ is isomorphic to a finite-index subgroup of $G(A_\tau, \mathcal{Q}_\tau)$, if and only if $A_\tau$ ramifies at all infinite places of $E_\tau$ other than the distinguished one.
\subsection{Congruence subgroups of triangle groups} \label{sec:congruence.subgroups}
As before, let $\tau = (a,b,c)$ be a hyperbolic triple with $c < \infty$. Let $J \vartriangleleft \mathcal{O}_{E_\tau}$ be an ideal, and let $\mathfrak{J} \vartriangleleft \mathcal{O}_{F_\tau}$ be an ideal such that $\mathfrak{J} \cap \mathcal{O}_{E_\tau} = J$. Suppose that $J$ is coprime to $abc$. Moreover suppose that $\mathfrak{J}$ is coprime to the discriminant of $\mathcal{O}_\tau$; by Proposition~\ref{pro:discriminant}, this is only an additional assumption if $(a,b,c) = (mk, m(k+1), mk(k+1)$ for some $n,k \in \mathbb{N}$. In this setup, Clark and Voight define a finite-index congruence subgroup $\overline{\Delta}(\tau;J) \trianglelefteq \overline{\Delta}(\tau)$ as the kernel of an explicit homomorphism $\Psi_J : \overline{\Delta}(\tau) \to \mathrm{PSL}_2(\mathcal{O}_{F_\tau} / \mathfrak{J})$ constructed in~\cite[\S5]{CV/19}. If $\overline{\Delta}(\tau) \simeq G(B_\tau, \mathcal{O}_\tau)$, then $\Psi_J$ is surjective. The significance of these groups is that if $J = \ensuremath{\mathfrak{p}}$ is a prime of $E_\tau$ dividing a rational prime $p$, then the Riemann surface $X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})}$ admits a Belyi map $X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})} \to \mathbb{P}^1(\mathbb{C})$, namely a non-constant morphism ramified at exactly three points, which is also a Galois cover.
If $k$ is a finite field and $k^\prime / k$ is the extension of degree two, then the homomorphism $\mathrm{GL}_2(k) \to \mathrm{PSL}_2(k^\prime)$ given by $g \mapsto \pm g / \sqrt{\det g}$ induces an embedding $\mathrm{PGL}_2(k) \hookrightarrow \mathrm{PSL}_2(k^\prime)$; note that every element of $k$ is a square in $k^\prime$.
The finite quotient $\overline{\Delta}(\tau) / \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$ is described explicitly in~\cite[Theorem~9.1]{CV/19}:
\begin{pro}[Clark-Voight] \label{pro:CV.congruence.quotient}
Suppose that $\tau = (a,b,c)$ and the prime $p$ are as above. Let $\mathfrak{p}$ be a prime of $E_\tau$ dividing $p$, and let $k_\ensuremath{\mathfrak{p}} = \mathcal{O}_{E_\tau} / \ensuremath{\mathfrak{p}}$ be its residue field. Then
$$ \overline{\Delta}(\tau) / \overline{\Delta}(\tau; \ensuremath{\mathfrak{p}}) \simeq
\begin{cases}
\mathrm{PSL}_2 (k_\ensuremath{\mathfrak{p}}) &: \ensuremath{\mathfrak{p}} \text{ splits completely in the extension } F_\tau / E_\tau \\
\mathrm{PGL}_2 (k_\ensuremath{\mathfrak{p}}) &: \text{otherwise.}
\end{cases}
$$
\end{pro}
It is easy to see that $F_\tau$ is contained in the cyclotomic field $\mathbb{Q}(\zeta_{2abc})$, where $\zeta_{2abc}$ is a primitive $2abc$-th root of unity. Hence $F_\tau / \mathbb{Q}$ is an abelian Galois extension. Therefore $E_\tau / \mathbb{Q}$ is also Galois, and the quotient $\overline{\Delta}(\tau) / \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$ is independent, up to isomorphism, of the choice of prime ideal $\ensuremath{\mathfrak{p}} \vartriangleleft \mathcal{O}_{E_\tau}$ dividing a rational prime $p$. Set $K(\tau;p) = \overline{\Delta}(\tau) / \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$.
The area of a fundamental domain of $\overline{\Delta}(\tau;J)$ is $2\pi \left( 1 - \frac{1}{a} - \frac{1}{b} - \frac{1}{c} \right) [\overline{\Delta}(\tau) : \overline{\Delta}(\tau;J)]$. By the Gauss-Bonnet theorem, the genus of the Riemann surface $X_{\overline{\Delta}(\tau;J)}$ is thus
\begin{equation} \label{equ:genus}
g(X_{\overline{\Delta}(\tau;J)}) = \frac{[\overline{\Delta}(\tau) : \overline{\Delta}(\tau;J)]}{2} \left( 1 - \frac{1}{a} - \frac{1}{b} - \frac{1}{c} \right) + 1.
\end{equation}
\begin{rem} \label{rmk:pgl.psl}
We will view $K(\tau;p) = \Psi_\ensuremath{\mathfrak{p}}(\overline{\Delta}(\tau))$ as a subgroup of $\mathrm{PSL}_2(\mathcal{O}_{F_\tau} / \mathfrak{P})$, where $\mathfrak{P}$ is a prime of $F_\tau$ dividing $\ensuremath{\mathfrak{p}}$. From now on we will omit $\tau$ from the notation and write $\overline{\Delta}$ and $\overline{\Delta}(\ensuremath{\mathfrak{p}})$ when this should not cause confusion.
\end{rem}
\subsection{Normal subgroups of $\overline{\Delta}(\tau)$} \label{sec:normal.subgroups}
Let $\tau = (a,b,c)$ as before. We will be interested below in classifying the normal subgroups $H \trianglelefteq \overline{\Delta}(\tau)$ satisfying
\begin{equation} \label{equ:right.quotient}
\overline{\Delta}(\tau)/H \simeq K(\tau;p).
\end{equation}
Let $T(\tau; p)$ be the set of triples $(\gamma_1, \gamma_2, \gamma_3) \in K(\tau; p)^3$ such that $\gamma_1^a = \gamma_2^b = \gamma_3^c = \gamma_1 \gamma_2 \gamma_3 = e$ and such that $\langle \gamma_1, \gamma_2, \gamma_3 \rangle = K(\tau;p)$. The automorphism group $\mathrm{Aut}(K(\tau;p))$ acts on $T(\tau;p)$ in the obvious way. Each element $(\gamma_1,\gamma_2,\gamma_3) \in T(\tau;p)$ gives rise to a normal subgroup $H \trianglelefteq \overline{\Delta}(\tau)$ satisfying~\eqref{equ:right.quotient}, namely the kernel of the epimorphism $\psi : \overline{\Delta}(\tau) \twoheadrightarrow K(\tau;p)$ determined by $\psi(x) = \gamma_1, \psi(y) = \gamma_2$, where $x,y \in \overline{\Delta}(\tau)$ are generators as in~\eqref{equ:triangle.group}. Clearly two elements in the same orbit of the $\mathrm{Aut}(K(\tau;p))$-action produce the same normal subgroup, and every normal subgroup satisfying~\eqref{equ:right.quotient} arises in this way. Thus the number of normal subgroups satisfying~\eqref{equ:right.quotient} is bounded by $| T(\tau;p) / \mathrm{Aut}(K(\tau;p))|$.
This quantity can be studied by means of the theory of Macbeath~\cite{Macbeath/69}, as extended by Clark and Voight~\cite{CV/19}. We present here just enough to state the results we will use and refer the reader to~\cite{Macbeath/69} and~\cite[\S6-8]{CV/19} for details.
For any field $k$, let $\pi : \mathrm{SL}_2(k) \to \mathrm{PSL}_2(k)$ denote the natural projection. For any $n \in \mathbb{N}$ and any prime power $q$, define $\kappa(n,q)$ to be the number of conjugacy classes of $\mathrm{PSL}_2(\mathbb{F}_q)$ whose elements have order $n$.
\begin{lem} \label{lem:eigenvalues}
Let $\mathbb{F}_q$ be a finite field of odd characteristic $p$, and let $n \in \mathbb{N}$ be coprime to $p$. Then $\kappa(n,q) = 0$ if $q^2 \not\equiv 1 \, \mathrm{mod} \, 2n$. Otherwise,
$$ \kappa(n,q) = \begin{cases}
1 &: n \in \{ 1, 2 \} \\
\frac{\varphi(n)}{2} &: n \geq 3 \, \mathrm{odd} \\
\frac{\varphi(2n)}{4} &: n \geq 4 \, \mathrm{even},
\end{cases}
$$
where $\varphi$ is Euler's totient function. The common trace of the elements in each conjugacy class of $\mathrm{PSL}_2(\mathbb{F}_q)$ with elements of order $n$ has the form $\pm (\zeta + \zeta^{-1})$, where $\zeta$ is a primitive $n$-th root of unity if $n$ is odd, and a primitive $2n$-th root of unity if $n$ is even.
\end{lem}
\begin{proof}
Observe first that if $q^2 \not\equiv 1 \, \mathrm{mod} \, 2n$, then, under our hypotheses, $n \nmid | \mathrm{PSL}_2(\mathbb{F}_q)|$ and hence $\kappa(n,q) = 0$. Now let $\gamma \in \mathrm{PSL}_2(\mathbb{F}_q)$ have order $n$, and let $M \in \pi^{-1}(\gamma) \subset \mathrm{SL}_2(\mathbb{F}_q)$. Then $M$ is semisimple, and its eigenvalues are reciprocal primitive $2n$-th roots of unity if $n$ is even and reciprocal primitive $n$-th or $2n$-th roots of unity if $n$ is odd. It is a simple exercise to count the possible traces, up to sign. Since a semisimple element of $\mathrm{PSL}_2(\mathbb{F}_q)$ is determined up to conjugacy by its trace, this completes the proof.
\end{proof}
Consider a triple $\underline{t} = (t_1, t_2, t_3) \in \mathbb{F}_q^3$. We will refer to such objects as trace triples. Let $T(\underline{t})$ be the set of triples $(g_1, g_2, g_3) \in \mathrm{SL}_2(\mathbb{F}_q)$ such that $g_1 g_2 g_3 = e$ and $\tr g_i = t_i$ for all $i \in \{1,2,3 \}$; note that we do not make any hypothesis here about the orders of the $g_i$ or the subgroup of $\mathrm{SL}_2(\mathbb{F}_q)$ generated by them. A fundamental result of Macbeath~\cite[Theorem~1]{Macbeath/69} is that $T(\underline{t}) \neq \varnothing$ for all trace triples $\underline{t}$. Moreover, Macbeath classified trace triples as follows. We say that $\underline{t}$ is commutative if there exists $(g_1, g_2, g_3) \in T(\underline{t})$ such that $\pi(\langle g_1, g_2, g_3 \rangle)$ is an abelian subgroup of $\mathrm{PSL}_2(\mathbb{F}_q)$. Commutative trace triples are easily identified in practice thanks to the following consequence of~\cite[Corollary~1]{Macbeath/69}.
\begin{lem}[Macbeath] \label{lem:commutative}
Let $\underline{t} = (t_1, t_2, t_3) \in \mathbb{F}_q^3$ be a trace triple. Then $\underline{t}$ is commutative if and only if $t_1^2 + t_2^2 + t_3^2 - t_1 t_2 t_3 - 4 = 0$.
\end{lem}
Let $\underline{g} = (g_1, g_2, g_3) \in \mathrm{SL}_2(\mathbb{F}_q)^3$. We associate to it the order triple $o(\underline{g}) \in \mathbb{N}^3$ consisting of the orders of the elements $\pi(g_1), \pi(g_2), \pi(g_3) \in \mathrm{PSL}_2(\mathbb{F}_q)$, arranged in non-descending order. A trace triple $\underline{t}$ is called exceptional if there exists $\underline{g} \in T(\underline{t})$ such that
\begin{multline*}
o(\underline{g}) \in \{ (2,2,c) : c \geq 2 \} \cup \\ \{ (2,3,3),(2,3,4),(2,3,5),(2,5,5),(3,3,3),(3,3,5),(3,4,4),(3,5,5),(5,5,5) \}.
\end{multline*}
If $\underline{t}$ is not commutative, then $o(\underline{g})$ is constant on all $\underline{g} \in T(\underline{t})$. We say that $\underline{t}$ is projective if, for all $\underline{g} \in T(\underline{t})$, the subgroup $\pi(\langle g_1, g_2, g_3 \rangle) \leq \mathrm{PSL}_2(\mathbb{F}_q)$ is isomorphic to $\mathrm{PSL}_2(k)$ or $\mathrm{PGL}_2(k)$ for some subfield $k \subseteq \mathbb{F}_q$. Macbeath~\cite[Theorem~4]{Macbeath/69} proved that every trace triple $\underline{t}$ is commutative, exceptional, or projective.
Recall that the trace of an element of $\mathrm{PSL}_2(\mathbb{F}_q)$ is only defined up to sign. Reflecting this and following~\cite[\S8]{CV/19}, we define a trace triple up to sign to be a triple $\pm \underline{t} = (\pm t_1, \pm t_2, \pm t_3)$, where $t_1, t_2, t_3 \in \mathbb{F}_q$. A triple $(t_1^\prime, t_2^\prime, t_3^\prime) \in \mathbb{F}_q^3$ is called a lift of $\pm \underline{t}$ if $t_i^\prime = \pm t_i$ for all $i \in \{ 1, 2, 3 \}$, where the signs in each component may be taken independently. We say that a trace triple up to sign is commutative or exceptional if it has a lift with the corresponding property. Following the terminology of~\cite[\S8]{CV/19}, we say that $\pm \underline{t}$ is partly projective if it has a lift which is projective. Every trace triple up to signs is commutative, exceptional, or partly projective~\cite[Lemma~8.9]{CV/19}.
Let $\underline{C} = (C_1, C_2, C_3)$ be a triple of conjugacy classes of $\mathrm{PSL}_2(\mathbb{F}_q)$. Again following~\cite{CV/19}, define $\Sigma(\underline{C})$ to be the set of triples $\underline{\gamma} = (\gamma_1, \gamma_2, \gamma_3) \in \mathrm{PSL}_2(\mathbb{F}_q)^3$ such that $\gamma_1 \gamma_2 \gamma_3 = e$, that $\gamma_i \in C_i$ for all $i \in \{ 1,2,3 \}$, and that $\langle \gamma_1, \gamma_2, \gamma_3 \rangle = K(\tau;\ensuremath{\mathfrak{p}})$.
The natural action of $\mathrm{Aut}(K(\tau;p))$ on triples of elements need not preserve $\Sigma(\underline{C})$. However, we may still define an equivalence relation on $\Sigma(\underline{C})$ by $\underline{\gamma} \sim \underline{\gamma}^\prime$ if there exists $\sigma \in \mathrm{Aut}(K(\tau; p))$ such that $\sigma(\gamma_i) = \gamma_i^\prime$ for all $i$. Denote the set of equivalence classes by $\Sigma(\underline{C})/\mathrm{Aut}(K(\tau;p))$. Let $o(\underline{C})$ be the triple of the common orders of the elements of $C_1$, $C_2$, and $C_3$, arranged in non-descending order, and let $\tr \underline{C}$ be the trace triple up to signs associated to $\underline{C}$. Let $\mathbb{F}_p(\tr \underline{C})$ be the subextension of $\mathbb{F}_q / \mathbb{F}_p$ generated by the components of $\tr \underline{C}$.
We are now ready to state a bound for the number of normal subgroups $H \trianglelefteq \overline{\Delta}(\tau)$ such that $\overline{\Delta}(\tau) / H \simeq K(\tau;p)$; see~\cite[Theorem~1.6]{Sah/69} for an overlapping result. Suppose that $K(\tau;p) \leq \mathrm{PSL}_2(\mathbb{F}_q)$ as in Remark~\ref{rmk:pgl.psl}. Let $\mathcal{C}(\tau;p)$ be the set of triples $\underline{C} = (C_1,C_2,C_3)$ of conjugacy classes of $\mathrm{PSL}_2(\mathbb{F}_q)$ such that $\Sigma(\underline{C}) \cap T(\tau;p) \neq \varnothing$.
\begin{pro} \label{pro:counting.normal.subgroups}
Let $\tau = (a,b,c)$, and let $p > 2$ be a rational prime.
Suppose that for every $\underline{C} \in \mathcal{C}(\tau;p)$ the associated trace triple up to signs $\tr \underline{C}$ is partly projective and not exceptional, and that it satisfies $\mathbb{F}_q = \mathbb{F}_p(\tr \underline{C})$. Set $\Omega(\tau;p) = \{ o(\underline{C}) : \underline{C} \in \mathcal{C}(\tau;p) \}$ and $\mathcal{N}(\tau;p) = \left\{ H \trianglelefteq \overline{\Delta}(\tau) : \overline{\Delta}(\tau) / H \simeq K(\tau;p) \right\}$. Then
$$
| \mathcal{N}(\tau;p) | \leq \sum_{(a^\prime, b^\prime, c^\prime) \in \Omega(\tau;p) \atop a^\prime = 2} \kappa(a^\prime,q) \, \kappa(b^\prime,q) \, \kappa(c^\prime,q) + \sum_{(a^\prime, b^\prime, c^\prime) \in \Omega(\tau;p) \atop a^\prime > 2} 2 \, \kappa(a^\prime,q) \, \kappa(b^\prime,q) \, \kappa(c^\prime,q).
$$
\end{pro}
\begin{proof}
It is clear that
$$ | \mathcal{N}(\tau;p) | \leq |T(\tau;p)/\mathrm{Aut}(K(\tau;p))| \leq \sum_{\underline{C} \in \mathcal{C}(\tau;p)} |\Sigma(\underline{C})/\mathrm{Aut}(K(\tau;p))|,$$
where the first inequality follows from the discussion at the beginning of this section. The second inequality need not be an equality since the action of $\mathrm{Aut}(K(\tau;p))$ on $T(\tau;p)$ need not preserve the sets $\Sigma(\underline{C}) \cap T(\tau;p)$. Our hypotheses ensure that~\cite[Proposition~8.10]{CV/19} applies; from it we conclude that $| \Sigma(\underline{C})/\mathrm{Aut}(K(\tau;p))| = 1$ if $o(\underline{C})$ has a component equal to $2$ and $| \Sigma(\underline{C})/\mathrm{Aut}(K(\tau;p))| \leq 2$ otherwise. This completes the proof.
\end{proof}
\begin{rem}
Since $a^\prime b^\prime c^\prime | abc$, it is clear that if $p \nmid abc$, then the integers $\kappa(a^\prime,q)$, $\kappa(b^\prime,q)$, and $\kappa(c^\prime,q)$ appearing on the right-hand side of the conclusion of Proposition~\ref{pro:counting.normal.subgroups} may be computed from the formula in Lemma~\ref{lem:eigenvalues}. We also note that the prime $p = 2$ has been excluded from the discussion in this section only to streamline the exposition; analogous results may be obtained in this case by analogous arguments.
\end{rem}
\section{Semi-arithmeticity} \label{sec:semiarithmetic}
As mentioned above, only finitely many of the hyperbolic triangle groups $\overline{\Delta}(\tau)$ are arithmetic. However, $\overline{\Delta}(\tau)$ is semi-arithmetic for all hyperbolic triples $\tau$ in the sense of~\cite[Definition~3]{SchallerWolfart/00}. In this section we apply a recent argument of Cosac and D\'{o}ria~\cite[Theorem~1.5]{CosacDoria/20} to obtain, in some cases, a lower bound on $\mathrm{sys}(X_{\overline{\Delta}(\tau;J)})$ in terms of the genus $g(X_{\overline{\Delta}(\tau;J)})$, which is given explicitly by~\eqref{equ:genus}, and an invariant of an explicit quaternion algebra. We then discuss earlier work in the arithmetic case, where more precise bounds are sometimes available.
Recall the quaternion algebra $A_\tau$ of Proposition~\ref{pro:presentation.a}. It is defined over the totally real field $E_\tau$ of~\eqref{equ:def.etau}, which is the invariant trace field of $\overline{\Delta}(\tau)$. Let $r_\tau$ be the number of infinite places of $E_\tau$ at which $A_\tau$ splits. Takeuchi~\cite[\S3]{Takeuchi/75} proved that $\overline{\Delta}(\tau)$ is arithmetic if and only if $r_\tau = 1$ and that this condition holds for exactly $85$ triples $\tau$. Nugent and Voight~\cite{NV/17} later showed that there are finitely many hyperbolic triples $\tau$ satisfying $r_\tau = N$ for any $N \in \mathbb{N}$ and provided an algorithm for computing them.
Cohen and Wolfart~\cite[\S2]{CohenWolfart/90} observed that the triangle group $\overline{\Delta}(\tau)$ admits a modular embedding. This means (cf.~\cite[Definition~4]{SchallerWolfart/00}) that there is an arithmetic lattice $\Lambda < \mathrm{PSL}_2(\mathbb{R})^{r_\tau}$, defined over $E_\tau$, a holomorphic map $F : \mathcal{H} \to \mathcal{H}^{r_\tau}$, and embeddings $\Phi_1, \dots, \Phi_{r_\tau} : \overline{\Delta}(\tau) \hookrightarrow \mathrm{PSL}_2(\mathbb{R})$, extending each of the embeddings $E_\tau \hookrightarrow \mathbb{R}$ at which $A_\tau$ splits, such that the map $\Phi = (\Phi_1, \dots, \Phi_{r_\tau}) : \overline{\Delta}(\tau) \to \mathrm{PSL}_2(\mathbb{R})^{r_\tau}$ satisfies $\Phi(\overline{\Delta}(\tau)) \subseteq \Lambda$ and $F(\gamma(z)) = \Phi(\gamma)(F(z))$ for all $\gamma \in \overline{\Delta}(\tau)$ and $z \in \mathcal{H}$. The lattice $\Lambda$ is commensurate with a subgroup of an arithmetic lattice arising from a quaternion algebra, but it need not in general be contained in such a lattice.
More explicitly, Clark and Voight~\cite[Proposition~5.13]{CV/19} construct an embedding $\overline{\Delta}(\tau) \hookrightarrow N_{A_\tau}(\mathcal{Q}_\tau)/E_\tau^\times$, where $\mathcal{Q}_\tau \subset A_\tau$ is the order defined in Proposition~\ref{pro:presentation.a} and $N_{A_\tau}(\mathcal{Q}_\tau)$ is its normalizer in $A_\tau$; see also~\cite[Remark~3(ii)]{SchallerWolfart/00}.
However, if at least two of the components of the triple $\tau$ are odd, then the fields $E_\tau = F_\tau$ coincide, as do the quaternion algebras $A_\tau = B_\tau$ and the orders $\mathcal{Q}_\tau = \mathcal{O}_\tau$; see~\eqref{equ:b.presentation} and Proposition~\ref{pro:alpha.beta.fulldelta} for the definitions. In this case there is an embedding $\overline{\Delta}(\tau) \hookrightarrow \mathcal{Q}^1_\tau / \{ \pm 1 \}$ as in Remark~\ref{rmk:embedding}, and the congruence subgroups are $\overline{\Delta}(\tau;J) = \overline{\Delta}(\tau) \cap \mathcal{Q}_\tau^1 (J)$ for all ideals $J \vartriangleleft \mathcal{O}_{E_\tau}$ coprime to $2abc$, where $\mathcal{Q}_\tau^1 (J) = \{ x \in \mathcal{Q}_\tau^1 : x \equiv 1 \, \mathrm{mod} \, J \mathcal{Q}_\tau \}$. Note that $\mathcal{Q}_\tau^1 (J)$ embeds in $\mathcal{Q}_\tau^1 / \{ \pm 1 \}$ by our assumptions on $J$. Then an argument of Cosac and D\'{o}ria obtains the following systolic bound.
\begin{thm} \label{thm:semiarithmetic}
Let $\tau = (a,b,c)$ be a hyperbolic triple with $c < \infty$. Suppose that at least two of $a,b,c$ are odd. There exists a constant $c_\tau$ such that, for every ideal $J \vartriangleleft \mathcal{O}_{E_\tau}$ coprime to $2abc$, the following holds:
\begin{equation} \label{equ:semiarithmetic}
\mathrm{sys}(X_{\overline{\Delta}(\tau;J)}) \geq \frac{4}{3r_\tau} \log g(X_{\overline{\Delta}(\tau;J)}) - c_\tau.
\end{equation}
\end{thm}
\begin{proof}
We follow the proof of~\cite[Theorem~1.5]{CosacDoria/20}.
Let $J = \prod_{i = 1}^r \ensuremath{\mathfrak{p}}_i^{e_i}$, where $\ensuremath{\mathfrak{p}}_1, \dots, \ensuremath{\mathfrak{p}}_r$ are distinct prime ideals of $\mathcal{O}_{E_\tau}$ that are coprime to $2abc$. We write $N(J)$ for the norm of $J$. By the proof of~\cite[Proposition~9.7]{CV/19}, we have
\begin{equation} \label{equ:general.quotient}
\overline{\Delta}(\tau) / \overline{\Delta}(\tau;J) \simeq \prod_{i = 1}^r \mathrm{PSL}_2(\mathcal{O}_{E_\tau} / \ensuremath{\mathfrak{p}}_i^{e_i}).
\end{equation}
Setting $q_i = N(\ensuremath{\mathfrak{p}}_i) = | \mathcal{O}_{E_\tau} / \ensuremath{\mathfrak{p}}_i |$, we obtain
\begin{equation} \label{equ:norm.bound}
[\overline{\Delta}(\tau) : \overline{\Delta}(\tau;J)] = \prod_{i = 1}^r |\mathrm{PSL}_2(\mathcal{O}_{E_\tau} / \ensuremath{\mathfrak{p}}_i^{e_i})| = \prod_{i = 1}^r \frac{(q_i^3 - q_i) q_i^{3(e_i - 1)}}{2} < \prod_{i = 1}^r q_i^{3e_i} = N(J)^3.
\end{equation}
Arguing as in the proof of~\cite[Theorem~1.5]{CosacDoria/20} and observing that we may take $\mathcal{F}$ to be the family of all ideals $J \vartriangleleft \mathcal{O}_{E_\tau}$ coprime to $2abc$, we conclude that if $N({J}) \geq 2^{[E_\tau:\mathbb{Q}]}$ then
$$ \mathrm{sys}(X_{\overline{\Delta}(\tau;J)}) \geq \frac{4}{{r_\tau}} \log \mathrm{vol}(X_{\mathcal{Q}_\tau^1({J})}) - c_\tau^\prime = \frac{4}{{r_\tau}} \log N({J}) - c_\tau^{\prime \prime},$$
for constants $c_\tau^\prime, c_\tau^{\prime \prime}$ independent of $J$. Hence it follows from~\eqref{equ:genus} and~\eqref{equ:norm.bound} that
$$ \mathrm{sys}(X_{\overline{\Delta}(\tau;J)}) \geq \frac{4}{3 r_\tau} \log g(X_{\overline{\Delta}(\tau;J)}) - c_\tau $$
for all $J \vartriangleleft \mathcal{O}_{E_\tau}$ coprime to $2abc$ such that $N(J) \geq 2^{[E_\tau: \mathbb{Q}]}$ and a constant $c_\tau$ independent of $J$. Since there are only finitely many ideals $J$ with $N(J) < 2^{[E_\tau: \mathbb{Q}]}$, we obtain the claim after possibly increasing $c_\tau$.
\end{proof}
\begin{rem}
If $\tau$ has at most one odd component, then the image of the embedding $j: \overline{\Delta}(\tau) \hookrightarrow N_{A_\tau}(\mathcal{Q}_\tau)/E_\tau^\times$ is not in general contained in $\mathcal{Q}_\tau^1 / \{ \pm 1 \}$. In this case a straightforward modification of the proof of Theorem~\ref{thm:semiarithmetic} establishes~\eqref{equ:semiarithmetic} for all ideals $J \vartriangleleft \mathcal{O}_{E_\tau}$ such that $J$ is coprime to $2abc$ and $j(\overline{\Delta}(\tau;J)) \subseteq \mathcal{Q}_\tau^1 / \{ \pm 1 \}$. It would be interesting to resolve whether Theorem~\ref{thm:semiarithmetic} is valid for all hyperbolic triples.
\end{rem}
\begin{rem}
In the case of arithmetic Fuchsian groups $\Gamma$, the result of Cosac and D\'{o}ria~\cite[Theorem~1.5]{CosacDoria/20} cited above recovers an earlier theorem of Katz, Schaps, and Vishne~\cite[Theorem~1.5]{KSV/07}, which gave the bound $\mathrm{sys}(X_{\Gamma(J)}) \geq \frac{4}{3} g(X_{\Gamma(J)}) - c_{\Gamma}$. Here $\Gamma$ is commensurate with $\mathcal{O}^1 / \{ \pm 1 \}$ for an order $\mathcal{O}$ in a quaternion algebra defined over a totally real field $F$ and split at exactly one infinite place, and $\Gamma(J) = \Gamma \cap \mathcal{O}^1(J)$.
This generalized earlier work of Buser and Sarnak~\cite[(4.7)]{BS/94} in the case $F = \mathbb{Q}$. Makisumi~\cite[Theorem~1.6]{Makisumi/13} has shown that the coefficient $\frac{4}{3}$ is the best possible in a result of this form.
See, for instance,~\cite{MaclachlanReid/03} for details on arithmetic Fuchsian groups. Indeed, Theorem~\ref{thm:semiarithmetic} in the case $r_\tau = 1$ was known to us before~\cite{CosacDoria/20} appeared.
\end{rem}
In general Theorem~\ref{thm:semiarithmetic} does not provide an effective lower bound on $\mathrm{sys}(X_{\overline{\Delta}(\tau;J)})$, since no effective constraint on the constant $c_\tau$ is known. In some arithmetic cases, it is known that one may take $c_\tau = 0$; see~\cite[Theorem~1.10]{KSV/07} and~\cite[Proposition~9.1]{KKSV/16} for sufficient conditions. These results rely on the trace bounds of~\cite[Theorem~2.3]{KSV/07} and do not readily extend to non-arithmetic situations. The upper bounds on $\mathrm{sys}(X_{\overline{\Delta}(\tau;J)})$ obtained by the method of Section~\ref{sec:procedure} below show that one may {\emph{not}} take $c_\tau = 0$ for some specific triples $\tau$. Finally, we note that the bounds given by Theorem~\ref{thm:semiarithmetic} are, in general, transcendental. However, it follows from Remark~\ref{rmk:embedding} that traces of elements of $\overline{\Delta}(\tau;J)$, and hence $\mathrm{sys}(X_{\overline{\Delta}(\tau;J)})$ and our upper bound, are contained in the totally real field $F_\tau$ and are thus algebraic numbers. Hence one cannot expect to determine the exact values of the systolic lengths of triangular modular curves by matching the upper bounds of Section~\ref{sec:procedure} with lower bounds arising from Theorem~\ref{thm:semiarithmetic}.
\section{Computations} \label{sec:computations}
\subsection{Procedure} \label{sec:procedure}
Let $\tau = (a,b,c)$ be a hyperbolic triple, let $\overline{\Delta} = \overline{\Delta}(\tau)$, and let $\ensuremath{\mathfrak{p}} \vartriangleleft \mathcal{O}_{E_\tau}$ be a prime ideal such that the congruence subgroup $\overline{\Delta}(\ensuremath{\mathfrak{p}})$ is defined. Our aim is to bound the systolic length of the compact Riemann surface $X_{\overline{\Delta}(\ensuremath{\mathfrak{p}})}$. In fact, for any finite real number $M > 0$, the set of numbers at most $M$ that occur as lengths of closed geodesics of $X_{\overline{\Delta}(\ensuremath{\mathfrak{p}})}$ could likely be computed by adapting the method used by Vogeler~\cite{Vogeler/03} in the Hurwitz case $\tau = (2,3,7)$; see also earlier work~\cite{DT/00, LW/02, Woods/01} by undergraduate participants of an REU at Rose-Hulman directed by S.~A.~Broughton.
We take a simpler but less precise approach that was introduced in~\cite{KKSV/16} in the case of $\tau = (3,3,4)$. Recall from Section~\ref{sec:traces} that determining $\mathrm{sys}(X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})})$ amounts to finding
$$ \min \left\{ \left\{ | \mathrm{tr} \, \varepsilon(\tilde{\gamma}) | : \gamma \in \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}}) \right\} \cap (2, \infty) \right\},$$
where $\gamma$ runs over all elements of the infinite group $\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$, where $\tilde{\gamma} \in \Delta(\tau)$ is a lift of $\gamma$, and $\varepsilon: \Delta(\tau) \hookrightarrow \mathrm{SL}_2(\mathbb{R})$ is some embedding; by work of Takeuchi mentioned above, the result is independent of the choice of $\varepsilon$.
We restrict to a set of Schreier generators of $\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$. This is a {\emph{finite}} generating set with some pleasant properties; details may be found in classical references on combinatorial group theory such as~\cite[Theorem~2.9]{MKS/66}. Among the generating sets of a subgroup of a finitely presented group obtainable by available algorithms, the elements of a set of Schreier generators are expressible as relatively short words in the generators $\alpha$ and $\beta$ of $\overline{\Delta}(\tau)$. Thus we may hope that the hyperbolic Schreier generators will have relatively low traces. The minimal trace of a hyperbolic Schreier generator is our upper bound on $\mathrm{sys}(X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})})$. Of course, it is possible that some other hyperbolic element of $\overline{\Delta}(\tau)$ has a lower trace than that of any of our Schreier generators. However, as mentioned in the introduction, in all the examples that we have computed where $\mathrm{sys}(X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})})$ is known, it matches our bound.
The procedure behind our computations is presented below. We assume that $\ensuremath{\mathfrak{p}} \vartriangleleft \mathcal{O}_{E_\tau}$ is prime as this streamlines the discussion of possible simplifications of the various steps of the computation. However, the same procedure may be used to find upper bounds on $\mathrm{sys}(X_{\overline{\Delta}(\tau;J)})$ for composite ideals $J$; in this case the description of the quotient $\overline{\Delta}(\tau)/ \overline{\Delta}(\tau;J)$ given in~\eqref{equ:general.quotient} should be used in Step 1 instead of Proposition~\ref{pro:CV.congruence.quotient}.
Our code is available at {\texttt{github.com/mmschein/systole}}.
{\bf{Step 1:}} {\emph{Identify subgroups $H \trianglelefteq \overline{\Delta}(\tau)$ such that $\overline{\Delta}(\tau)/H \simeq K(\tau;p)$.}}
If $|K(\tau;p)|$ is reasonably small, then Magma's {\texttt{LowIndexNormalSubgroups}} routine may be used. When $K(\tau;p) = \mathrm{PSL}_2(\mathbb{F}_q)$ with $q > 3$, it follows from the classification of finite simple groups that $K(\tau;p)$ is the unique simple group of order $\frac{q}{(2,q-1)} (q^2 - 1)$ and so it is particularly easy to identify relevant subgroups $H \trianglelefteq \overline{\Delta}(\tau)$ from the output of {\texttt{LowIndexNormalSubgroups}}.
If $|K(\tau;p)|$ is large, it is more efficient to generate random pairs of elements $(z_1, z_2) \in K(\tau;p)^2$ and select those for which $(z_1, z_2, (z_1 z_2)^{-1}) \in T(\tau;p)$; recall that $T(\tau;p)$ was defined in Section~\ref{sec:normal.subgroups}. Such a pair gives rise to an epimorphism $\psi: \overline{\Delta}(\tau) \twoheadrightarrow K(\tau;p)$ determined by $\psi(x) = z_1$ and $\psi(y) = z_2$. Then we consider the kernel $H = \ker \psi$.
{\bf{Step 2:}} {\emph{Compute a set of Schreier generators for $H$.}}
See~\cite[\S4]{Neubuser/82} and~\cite[\S6]{Sims/94} for algorithms computing Schreier generators. We have used the implementation in Magma. This is by far the most time-consuming step of our computations.
{\bf{Step 3:}} {\emph{Determine whether there is an ideal $\ensuremath{\mathfrak{p}} \vartriangleleft \mathcal{O}_{E_\tau}$ dividing $p$ such that $H = \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$, and identify $\ensuremath{\mathfrak{p}}$ if so.}}
The most general method for doing this is as follows. Recall from Remark~\ref{rmk:embedding} that we have an explicit embedding $\varepsilon_0: {\Delta}(\tau) \hookrightarrow \mathcal{O}_\tau^1$. Let $S_H$ be the set of Schreier generators of $H$ computed in the previous step. For every $s \in S_H$, we compute $\varepsilon_0(s) = u_0 + u_1 i + u_2 j + u_3 ij$, in terms of the presentation of $B_\tau$ given in Proposition~\ref{pro:alpha.beta.fulldelta}. Set
$$
\left( \begin{array}{c} v_0 \\ v_1 \\ v_2 \\ v_3 \end{array} \right) =
\left( \begin{array}{cccc}
1 & \cos \frac{\pi}{a} & \cos \frac{\pi}{b} & - \cos \frac{\pi}{c} \\
0 & \frac{1}{2} & - \frac{\cos \frac{\pi}{a} \cos \frac{\pi}{b} + \cos \frac{\pi}{c}}{2 \left( \cos^2 \frac{\pi}{a} - 1 \right)} & - \frac{\cos \frac{\pi}{b} + \cos \frac{\pi}{a} \cos \frac{\pi}{c}}{2 \left( \cos^2 \frac{\pi}{a} - 1 \right)} \\
0 & 0 & 0 & \frac{1}{2} \\
0 & 0 & \frac{1}{4 \left( \cos^2 \frac{\pi}{a} - 1 \right)} & \frac{\cos \frac{\pi}{a}}{4 \left( \cos^2 \frac{\pi}{a} - 1 \right)}
\end{array} \right)^{-1}
\left( \begin{array}{c} u_0 \\ u_1 \\ u_2 \\ u_3 \end{array} \right).
$$
Then the $v_i$ are the coefficients of $\varepsilon_0(s)$ with respect to the $F_\tau$-basis $(1, \alpha, \beta, \alpha \beta)$ of $B_\tau$. If, for all $s \in S_H$, we have $v_0 \equiv \pm 1 \, \mathrm{mod} \, \mathfrak{P}$ and $v_1, v_2, v_3 \in \mathfrak{P}$ for a prime ideal $\mathfrak{P} \vartriangleleft \mathcal{O}_{F_\tau}$ dividing $\ensuremath{\mathfrak{p}}$, then $H \leq \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$ and hence $H = \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$ as they are subgroups of $\overline{\Delta}(\tau)$ of the same index.
This step of the computation can often be simplified. For instance, the upper bound on $\mathcal{N}(\tau;p) = \left\{ H \trianglelefteq \overline{\Delta}(\tau) : \overline{\Delta}(\tau) / H \simeq K(\tau;p) \right\}$ obtained from Proposition~\ref{pro:counting.normal.subgroups} is sometimes equal to the number of prime ideals of $\mathcal{O}_{E_\tau}$ dividing $p$. In this case, any subgroup $H$ obtained in Step 1 is necessarily a congruence subgroup, and testing just a few Schreier generators will suffice to eliminate all $\ensuremath{\mathfrak{p}} | p$ but one.
If, futhermore, $\mathcal{O}_{F_\tau} / \mathfrak{P} = \mathcal{O}_{E_\tau} / \ensuremath{\mathfrak{p}}$ (or, equivalently, $K(\tau;p) = \mathrm{PSL}_2(k_\ensuremath{\mathfrak{p}})$), then we may dispense with Schreier generators entirely in this step. Indeed, the explicit homomorphism $\Psi_\ensuremath{\mathfrak{p}}: \overline{\Delta}(\tau) \to \mathrm{PSL}_2(\mathcal{O}_{F_\tau} / \mathfrak{P})$, whose kernel is $\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$, is given in~\cite[\S5]{CV/19}. Here $\mathfrak{P}$ is a prime of $F_\tau$ dividing $\ensuremath{\mathfrak{p}}$.
We know by~\cite[Proposition~5.23]{CV/19} that
$\Psi_\ensuremath{\mathfrak{p}}(x)$, $\Psi_\ensuremath{\mathfrak{p}}(y)$, and $\Psi_\ensuremath{\mathfrak{p}}((xy)^{-1})$ have traces $\pm (2 \cos \frac{\pi}{a}) \, \mathrm{mod} \, \mathfrak{P}$, $\pm (2 \cos \frac{\pi}{b}) \, \mathrm{mod} \, \mathfrak{P}$, and $\pm (2 \cos \frac{\pi}{c}) \, \mathrm{mod} \, \mathfrak{P}$, respectively. Then in Step 1 we search for an epimorphism $\psi: \overline{\Delta}(\tau) \twoheadrightarrow \mathrm{PSL}_2(k_\ensuremath{\mathfrak{p}})$ satisfying these trace conditions and can conclude immediately that $\ker \psi = \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$. It is still necessary to determine the Schreier generators of $\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$ for the following step.
{\bf{Step 4:}} {\emph{Determine the minimal trace of a Schreier generator.}}
For every Schreier generator $s \in S_H$, we compute $| \mathrm{tr} \, \varepsilon(\tilde{s}) |$, where $\tilde{s} \in \Delta(\tau)$ is a lifting of $s$ and $\epsilon: \Delta(\tau) \hookrightarrow \mathrm{SL}_2(\mathbb{R})$ is some embedding. It is usually convenient to take $\varepsilon$ to be the explicit embedding of Remark~\ref{rmk:embedding}. Our upper bound on $\mathrm{sys}(X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})})$ is then
$$ \min \left\{ \left\{ | \mathrm{tr} \, \varepsilon(\tilde{s}) | : s \in S_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})} \right\} \cap (2, \infty) \right\}.$$
\begin{rem}
Let $g$ be the genus of $X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})}$; recall that an explicit formula for $g$ was given in~\eqref{equ:genus}. In all cases where we have computed it, Reidemeister-Schreier rewriting as implemented in Magma indeed produces the simplest presentation of $\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$, with $2g$ generators and a single relation of length $4g$ in which each generator and its inverse appear exactly once.
\end{rem}
In the remainder of this section we consider some specific examples of the computations outlined above.
\subsection{Hurwitz surfaces} \label{sec:hurwitz}
A Hurwitz surface is a compact Riemann surface $X$ of genus $g(X)$ with exactly $84(g(X) - 1)$ automorphisms, which is the maximal number possible by a classical theorem of Hurwitz. These have been studied since 1879, when Klein discovered his celebrated quadric of genus three, with $168$ automorphisms. Together with the theorem mentioned above, Hurwitz showed~\cite[\S7]{Hurwitz/93} that a group of order $84(g-1)$ is the automorphism group of some Riemann surface of genus $g$ if and only if it is a homomorphic image of the triangle group $\overline{\Delta}(2,3,7)$. Thus the new Hurwitz surfaces that have been discovered over the past century have amounted to constructions of normal subgroups of $\overline{\Delta}(2,3,7)$ of finite index; as a very incomplete sample of these results, we mention the work of Sinkov~\cite{Sinkov/37}, Macbeath~\cite{Macbeath/61, Macbeath/65}, Leech~\cite{Leech/65}, Lehner and Newman~\cite{LN/67}, and Cohen~\cite{Cohen/79}. The construction of congruence subgroups of $\overline{\Delta}(\tau)$ discussed above, in the special case $\tau = (2,3,7)$, recovers some, but not all, of this previous work.
Set $\mu = 2 \cos \frac{\pi}{7}$ for brevity. If $\tau = (2,3,7)$, then $F_\tau = E_\tau = \mathbb{Q}(\mu)$. This is a cubic field, where $\mu$ has minimal polynomial $x^3 - x^2 - 2x + 1$ over $\mathbb{Q}$. Let $\zeta_7$ be a primitive seventh root of unity such that $- (\zeta_7 + \zeta_7^{-1}) = \mu$. Then $\mathbb{Q}(\mu)$ is contained in the cyclotomic field $\mathbb{Q}(\zeta_7)$. In particular, $\mathcal{O}_{F_\tau} = \mathbb{Z} [\mu]$. We determine the decomposition of primes in $\mathbb{Q}(\mu)$.
\begin{lemma} \label{lem:hurwitz.decomposition}
The prime $7$ is totally ramified in $\mathbb{Q}(\mu)$. If $p \neq 7$ is a rational prime, then $p$ splits completely in $\mathbb{Q}(\mu)$ if $p \equiv \pm 1 \, \mathrm{mod} \, 7$, and $p$ is inert otherwise.
\end{lemma}
\begin{proof}
We rely on basic facts about cyclotomic fields; see, for instance,~\cite[{\S}I.10]{Neukirch/92}. The only prime ramifying in $\mathbb{Q}(\zeta_7)$ is $7$, and it is totally ramified. Therefore the same is true for the subfield $\mathbb{Q}(\mu)$. Now assume $p \neq 7$, let $\ensuremath{\mathfrak{p}}$ be a prime of $\mathbb{Q}(\mu)$ dividing $p$, and let $\mathfrak{P}$ be a prime of $\mathbb{Q}(\zeta_7)$ dividing $\ensuremath{\mathfrak{p}}$. Since $\mathbb{Q}(\mu)/\mathbb{Q}$ is a cubic Galois subextension of the sextic Galois extension $\mathbb{Q}(\zeta_7) / \mathbb{Q}$, it is clear that $p$ is completely split in $\mathbb{Q}(\mu)$ if and only if the inertia degree is $f(\ensuremath{\mathfrak{p}} / p) = 1$, which in turn is equivalent to $f(\mathfrak{P} / p) \in \{1, 2 \}$. Otherwise, $f(\ensuremath{\mathfrak{p}} / p) = 3$ and $p$ is inert in $\mathbb{Q}(\mu)$. Now $f(\mathfrak{P} / p)$ is the minimal $f \in \mathbb{N}$ such that $p^f \equiv 1 \, \mathrm{mod} \, 7$. Thus $f(\mathfrak{P}/p) \in \{1, 2 \}$ if and only if $7 | (p^2 - 1)$, which is equivalent to $p \equiv \pm 1 \, \mathrm{mod} \, 7$.
\end{proof}
Since $\overline{\Delta}(2,3,7) = \overline{\Delta}^{(2)} (2,3,7)$, we have $\overline{\Delta}(2,3,7) \leq \mathcal{O}_\tau^1 / \{ \pm 1 \}$ as a subgroup of finite index, where by Proposition~\ref{pro:alpha.beta.fulldelta} the order $\mathcal{O}_\tau \subset B_\tau$ of the quaternion algebra $B_\tau = \left \langle \frac{-4, \mu^2 - 3}{F_\tau} \right \rangle$ has $\mathcal{O}_{F_\tau}$-basis $\{ 1, \alpha, \beta, \alpha \beta \}$, with $\alpha = \frac{i}{2}$ and $\beta = \frac{1}{2} + \frac{\mu}{4} i - \frac{1}{4}ij$. In fact, $\overline{\Delta}(2,3,7)$ is arithmetic~\cite[Theorem~3]{Takeuchi/77}, so $B_\tau$ splits only at the distinguished infinite place of $F_\tau$.
\begin{rem}
The construction given here as a special case of Takeuchi's general theory is equivalent to the ``Hurwitz order'' $\mathcal{Q}_{\mathrm{Hur}}$ studied by Katz, Schaps, and Vishne~\cite[(2.8)]{KSV/11}. Indeed, they denote $\eta = 2 \cos \frac{2 \pi}{7} = \mu^2 - 2$. Observe that $\mathbb{Q}(\mu) = \mathbb{Q}(\eta)$ by Remark~\ref{rmk:chebyshev} and that
\begin{eqnarray*}
\varphi : \left \langle \frac{-4, \mu^2 - 3}{\mathbb{Q}(\mu)} \right \rangle & \to & \left \langle \frac{\eta, \eta}{\mathbb{Q}(\eta)} \right \rangle \\
\varphi(i) & = & (2 \mu - 2) ij \\
\varphi(j) & = & -(\mu^2 - \mu - 1) i
\end{eqnarray*}
is an isomorphism of quaternion algebras between our $B_\tau$ and the algebra denoted $D$ in~\cite{KSV/11}. Moreover, $\varphi(\alpha) = g_2$ and $\varphi(\beta) = g_3$, where $g_2, g_3 \in D$ are as defined in~\cite[\S4]{KSV/11}; they arise from work of Elkies~\cite{Elkies/99}. Since $\mathcal{Q}_{\mathrm{Hur}}$ is spanned by $\{1, g_2, g_3, g_2 g_3 \}$ over $\mathbb{Z}[\mu]$ by~\cite[Theorem~4.2]{KSV/11}, we have $\varphi(\mathcal{O}_\tau) = \mathcal{Q}_{\mathrm{Hur}}$.
\end{rem}
We have $\mathcal{O}_\tau^1 / \{ \pm 1 \} = \simeq \overline{\Delta} = \overline{\Delta}(2,3,7)$~\cite[\S4.4]{Elkies/99}. The order $\mathcal{O}_\tau \subset B_\tau$ is maximal and splits at all finite places of $\mathbb{Q}(\mu)$ since the quaternion algebra $B_\tau$ does so. Hence for any ideal $I \vartriangleleft \mathcal{O}_{F_{\tau}} = \mathbb{Z} [\mu]$ we get a congruence subgroup $\overline{\Delta}(I) \vartriangleleft \overline{\Delta}$ such that $\overline{\Delta} / \overline{\Delta}(I) \simeq \mathrm{PSL}_2 (\mathcal{O}_{F_\tau} / I)$. By~\cite[Proposition~9.1]{KKSV/16} we conclude that $\mathrm{sys} ( X_{\overline{\Delta}(I)}) > \frac{4}{3} g(X_{\overline{\Delta}(I)})$ for all but finitely many ideals $I \vartriangleleft \mathbb{Z} [\mu]$; indeed, $T_1 = T_2 = \varnothing$ in the notation of~\cite{KKSV/16} and $2^6 = 64 < 84 = \frac{4 \pi}{2\pi \left( 1 - \frac{1}{2} - \frac{1}{3} - \frac{1}{7} \right)}$. By~\cite[Remark~9.2]{KKSV/16} this inequality holds for all $I$ such that $g(X_{\overline{\Delta}(I)}) \geq 85$.
\begin{pro} \label{pro:hurwitz.normal.subgroups}
Let $p$ be prime, and let $H \trianglelefteq \overline{\Delta} = \overline{\Delta}(2,3,7)$ be a normal subgroup. Then
$$ \overline{\Delta} / H \simeq \begin{cases}
\mathrm{PSL}_2(\mathbb{F}_p) &: p \equiv 0, \pm 1 \, \mathrm{mod} \, 7 \\
\mathrm{PSL}_2(\mathbb{F}_{p^3}) &: p \not\equiv 0, \pm 1 \, \mathrm{mod} \, 7
\end{cases}
$$
if and only if $H = \overline{\Delta}(\ensuremath{\mathfrak{p}})$, where $\ensuremath{\mathfrak{p}} \vartriangleleft \mathbb{Z}[\mu]$ is a prime dividing $p$.
\end{pro}
\begin{proof}
One direction is clear by the previous paragraph and Lemma~\ref{lem:hurwitz.decomposition}.
The other direction is contained in~\cite[Theorem~8]{Macbeath/69} and historically was the first application of Macbeath's classification. We give the proof in some detail, as an illustration of the machinery of Section~\ref{sec:normal.subgroups}, although it is essentially the same as the one in~\cite{Macbeath/69}.
We first compute, using Magma's \texttt{LowIndexNormalSubgroups} function, that there is a unique normal subgroup $H \trianglelefteq \overline{\Delta}$ for which the quotient $\overline{\Delta}/H$ is isomorphic to each of $\mathrm{PSL}_2(\mathbb{F}_7)$, $\mathrm{PSL}_2(\mathbb{F}_8)$, and $\mathrm{PSL}_2(\mathbb{F}_{27})$. These facts can also be proven without reliance on a computer by a suitable modification of the arguments below, but we take the easier path. Thus we may henceforth assume $p \not\in \{2,3,7 \}$. It suffices to show that there are at most three normal subgroups $H \trianglelefteq \overline{\Delta}$ satisfying $\overline{\Delta} / H \simeq \mathrm{PSL}_2(\mathbb{F}_p)$ if $p$ splits in $\mathbb{Q}(\mu)$, and at most one normal subgroup satisfying $\overline{\Delta} / H \simeq \mathrm{PSL}_2(\mathbb{F}_{p^3})$ if $p$ is inert.
Let $\ensuremath{\mathfrak{p}}$ be a place of $\mathbb{Q}(\mu)$ dividing $p$. Then $K(\tau;p) = \mathrm{PSL}_2(\mathbb{F}_q)$, where $q = p$ if $p$ splits in $\mathbb{Q}(\mu)$ and $q = p^3$ otherwise. Let $(\gamma_1, \gamma_2, \gamma_3) \in T(\tau;p)$, in the notation of Section~\ref{sec:normal.subgroups}. We claim that $\gamma_1$, $\gamma_2$, and $\gamma_3$ have order $2$, $3$, and $7$, respectively, as elements of $\mathrm{PSL}_2(\mathbb{F}_q)$. Indeed, otherwise one of the $\gamma_i$ would be trivial, and consequently the subgroup $\langle \gamma_1, \gamma_2, \gamma_3 \rangle$ would be trivial, contradicting the definition of $T(\tau;p)$. Thus $\Omega(\tau;p) = \{ (2,3,7) \}$, in the notation of Proposition~\ref{pro:counting.normal.subgroups}. Moreover, it follows from Lemma~\ref{lem:eigenvalues} that $\tr \gamma_1 = 0$ and $\tr \gamma_2 = \pm 1$, whereas $\tr \gamma_3$ is one of $\pm(\zeta_7 + \zeta_7^{-1})$, $\pm(\zeta_7^2 + \zeta_7^{-2})$, or $\pm(\zeta_7^3 + \zeta_7^{-3})$, where $\zeta_7$ is a primitive seventh root of unity. Thus there are three potential trace triples up to sign associated to elements of $T(\tau;p)$. By Lemma~\ref{lem:commutative}, none of them are commutative. Moreover, they are not exceptional, since for any lift $\underline{t}$ of one of them and any $\underline{g} \in T(\underline{t})$, we have $o(\underline{g}) = (2,3,7)$. Hence all three trace triples up to signs are partly projective. It now follows from Proposition~\ref{pro:counting.normal.subgroups} and Lemma~\ref{lem:eigenvalues} that
\begin{equation} \label{equ:hurwitz.count}
| \mathcal{N}(\tau;p) | \leq \kappa(2,q) \, \kappa(3,q) \, \kappa(7,q) = \frac{\varphi(3) \, \varphi(7)}{4} = 3.
\end{equation}
Hence there are at most three subgroups $H \trianglelefteq \overline{\Delta}(\tau)$ satisfying $\overline{\Delta}(\tau) / H \simeq \mathrm{PSL}_2(\mathbb{F}_q)$.
It remains to show that there is only one such normal subgroup if $p$ is inert in $\mathbb{Q}(\mu)$. In this case, let $\sigma(\alpha) = \alpha^p$ be the Frobenius automorphism of $\mathbb{F}_q = \mathbb{F}_{p^3}$. The action of $\sigma$ on matrix elements induces an automorphism of $\mathrm{PSL}_2(\mathbb{F}_q)$. Observe that
$$\{ \sigma(\pm(\zeta_7 + \zeta_7^{-1})), \sigma^2(\pm(\zeta_7 + \zeta_7^{-1})) \} = \{ \pm(\zeta_7^2 + \zeta_7^{-2}), \pm (\zeta_7^3 + \zeta_7^{-3}) \}.$$
Thus, if the epimorphism $\psi : \overline{\Delta}(\tau) \twoheadrightarrow \mathrm{PSL}_2(\mathbb{F}_q)$ corresponds to an element of $T(\tau;p)$ having one of the three possible trace triples up to signs, the epimorphisms $\sigma \circ \psi$ and $\sigma^2 \circ \psi$, which have the same kernel, correspond to the other two trace triples up to signs. Equivalently, a single orbit of the action of $\langle \sigma \rangle \leq \mathrm{Aut}(K(\tau;p))$ intersects $\Sigma(\underline{C}) \cap T(\tau;p)$ for all three elements $\underline{C} \in \mathcal{C}(\tau;p)$. In either case, we see from the proof of Proposition~\ref{pro:counting.normal.subgroups} that $\mathcal{N}(\tau;p)| = 1$.
\end{proof}
We tabulate below our bounds on $\mathrm{sys}(X_{\overline{\Delta}(I)})$ for all proper ideals $I \vartriangleleft \mathcal{O}_{F_\tau}$ of norm $N(I) \leq 100$, to an accuracy of three decimal places. For all the prime ideals in this table, Vogeler~\cite[Appendix~C]{Vogeler/03} has computed length spectra of the corresponding Riemann surfaces. In all cases, the lowest trace identified by us corresponds to the smallest geodesic length found by Vogeler. His tables are complete for lengths smaller than $14.49$, so they determine the systolic length for surfaces $X$ with $\mathrm{sys}(X) < 14.49$. This is evidence that our method provides a reasonable estimate for the systolic length $\mathrm{sys}(X_\Gamma)$.
We see as in~\eqref{equ:genus} that $g(X_{\overline{\Delta}(I)}) = \frac{| \mathrm{PSL}_2(\mathbb{Z}[\mu]/I) |}{84} + 1$ for all $I \vartriangleleft \mathbb{Z}[\mu]$.
The Riemann surface $X_{\overline{\Delta}(I)}$ of genus $3$ corresponding to the ideal $I$ of norm $7$ in the first line of the table is, of course, none other than the Klein quadric. The surface $X_{\overline{\Delta}((2))}$ of genus $7$ was studied by Fricke~\cite{Fricke/99} and more than sixty years later by Macbeath~\cite{Macbeath/65}, who was apparently unaware of Fricke's work; see Serre's letter~\cite{Serre/90} to Abhyankar for a more modern perspective on this curve. If $N$ is a natural number with a unique prime factor $p$ such that $p$ splits or ramifies in $\mathbb{Q}(\mu)$ (i.e. $p \equiv 0, \pm 1 \, \mathrm{mod} \, 7$) and if $(p^\prime)^3 | N$ for all prime factors $p^\prime \neq p$ of $N$, then there exist three ideals $I \vartriangleleft \mathbb{Z}[\mu]$ of norm $N(I) = N$, giving rise to three Hurwitz surfaces of the same genus; these are the Hurwitz triplets.
Not all normal subgroups of $\overline{\Delta}$ arise as $\overline{\Delta}(I)$ for some ideal $I \vartriangleleft \mathbb{Z}[\mu]$. The example of lowest index is a pair of normal subgroups of index $1344$ found by Sinkov~\cite{Sinkov/37}. They are conjugate in the full triangle group $\widetilde{\Delta}(2,3,7)$ of~\eqref{equ:full.triangle.group} and correspond to a single chiral Riemann surface of genus $\frac{1344}{84} + 1 = 17$. In this case, the lowest trace we obtain is $11 \mu^2 + 9 \mu - 7$, corresponding to the systolic length found by Vogeler. Cohen~\cite[Theorem~2]{Cohen/79} described an infinite family of normal subgroups of $\overline{\Delta}$ whose quotients are not isomorphic to $\mathrm{PSL}_2(\mathbb{Z}[\mu] / I)$ for any ideal $I \vartriangleleft \mathbb{Z}[\mu]$.
\begin{center}
\begin{longtable}{| r | r || r | r || r | r |} \hline
$I$ & $N(I)$ & $x =$ lowest observed trace & $2 \arcosh \frac{x}{2}$ & $g(X_{\overline{\Delta}(I)})$ & $\frac{4}{3} \log g$ \\ \hline \hline
$(\mu + 2)$ & $7$ & $ 2 \mu^2 + \mu - 1$ & $3.936$ & $3$ & $1.465$ \\ \hline
$(2)$ & $8$ & $4 \mu^2 + 4 \mu - 2$ & $5.796$ & $7$ & $2.595$ \\ \hline
$(2 \mu + 1)$ & $13$ & $4 \mu^2 + 4 \mu - 1$ & $5.904$ & $14$ & $3.519$ \\ \hline
$(2 \mu + 3)$ & $13$ & $6 \mu^2 + 5 \mu - 4$ & $6.393$ & $14$ & $3.519$ \\ \hline
$(\mu - 3)$ & $13$ & $7 \mu^2 + 7 \mu - 4$ & $6.888$ & $14$ & $3.519$ \\ \hline
$(3)$ & $27$ & $45 \mu^2 + 36 \mu - 25$ & $10.451$ & $118$ & $6.361$ \\ \hline
$(4 \mu - 1)$ & $29$ & $19 \mu^2 + 15 \mu - 12$ & $8.680$ & $146$ & $6.645$ \\ \hline
$(3 \mu - 4)$ & $29$ & $49 \mu^2 + 41 \mu - 27$ & $10.656$ & $146$ & $6.645$ \\ \hline
$(\mu + 3)$ & $29$ & $73 \mu^2 + 60 \mu - 40$ & $11.442$ & $146$ & $6.645$ \\ \hline
$(4 \mu - 3)$ & $41$ & $33\mu^2 + 26 \mu - 17$ & $9.340$ & $411$ & $8.025$ \\ \hline
$(3 \mu + 1)$ & $41$ & $49 \mu^2 + 40 \mu - 26$ & $10.648$ & $411$ & $8.025$ \\ \hline
$(\mu - 4)$ & $41$ & $121 \mu^2 + 97 \mu - 67$ & $12.432$ & $411$ & $8.025$ \\ \hline
$(2 \mu - 5)$ & $43$ & $49 \mu^2 + 40 \mu - 28$ & $10.628$ & $474$ & $8.215$ \\ \hline
$(3 \mu + 2)$ & $43$ & $55 \mu^2 + 43 \mu - 32$ & $10.824$ & $474$ & $8.215$ \\ \hline
$(5 \mu - 3)$ & $43$ & $86 \mu^2 + 69 \mu - 48$ & $11.747$ & $474$ & $8.215$ \\ \hline
$(\mu + 2)^2$ & $49$ & $105 \mu^2 + 84 \mu - 58$ & $12.147$ & $687$ & $8.710$ \\ \hline
$(2)(\mu + 2)$ & $56$ & $184 \mu^2 + 148 \mu - 102$ & $13.272$ & $1009$ & $9.222$ \\ \hline
$(2)^2$ & $64$ & $80 \mu^2 + 64 \mu - 46$ & $11.593$ & $1537$ & $9.783$ \\ \hline
$(5 \mu - 1)$ & $71$ & $121 \mu^2 + 97 \mu - 66$ & $12.436$ & $2131$ & $10.219$ \\ \hline
$(\mu + 4)$ & $71$ & $133 \mu^2 + 106 \mu - 73$ & $12.619$ & $2131$ & $10.219$ \\ \hline
$(4 \mu - 5)$ & $71$ & $151 \mu^2 + 121 \mu - 83$ & $12.877$ & $2131$ & $10.219$ \\ \hline
$(3 \mu + 5)$ & $83$ & $276 \mu^2 + 220 \mu - 153$ & $14.077$ & $3404$ & $10.844$ \\ \hline
$(8 \mu - 3)$ & $83$ & $384 \mu^2 + 308 \mu - 213$ & $14.742$ & $3404$ & $10.844$ \\ \hline
$(5 \mu - 8)$ & $83$ & $520 \mu^2 + 416 \mu - 289$ & $15.346$ & $3404$ & $10.844$ \\ \hline
$(\mu + 2)(\mu - 3)$ & $91$ & $44 \mu^2 + 36 \mu - 25$ & $10.416$ & $2185$ & $10.252$ \\ \hline
$(\mu + 2)(2 \mu + 1)$ & $91$ & $88 \mu^2 + 72 \mu - 49$ & $11.808$ & $2185$ & $10.252$ \\ \hline
$(\mu + 2)(2 \mu + 3)$ & $91$ & $256 \mu^2 + 205 \mu - 143$ & $13.928$ & $2185$ & $10.252$ \\ \hline
$(3\mu - 7)$ & $97$ & $321 \mu^2 + 257 \mu - 179$ & $14.380$ & $5433$ & $11.467$ \\ \hline
$(4 \mu + 3)$ & $97$ & $412 \mu^2 + 331 \mu - 228$ & $14.884$ & $5433$ & $11.467$ \\ \hline
$(7 \mu - 4)$ & $97$ & $711 \mu^2 + 569 \mu - 395$ & $15.972$ & $5433$ & $11.467$ \\ \hline
\end{longtable}
\end{center}
\subsection{$(2,3,12)$ triangle surfaces} \label{sec:2312}
We next consider the case $\tau = (a,b,c) = (2,3,12)$. A Riemann surface $X$ of genus $g$ is called maximal if it is a local maximum for the function $\mathrm{sys}(X)$ on the Teichm\"{u}ller space $T_g$. The $(2,3,12)$ triangle surfaces of genus $3$ and $4$ were shown to be maximal by Schmutz Schaller~\cite[\S8]{Schmutz/93}, so this is a natural family in which to search for surfaces with large systolic length.
In this section, write $\overline{\Delta}$ for $\overline{\Delta}(2,3,12)$. Observe that $E_\tau = \mathbb{Q}(\sqrt{3})$ and $F_\tau = \mathbb{Q}(\sqrt{2}, \sqrt{3})$. The complication of this case, in comparison with the Hurwitz case considered in the previous section, is that now $\overline{\Delta}$ is not contained in $\mathcal{O}^1 / \{ \pm 1 \}$ for any quaternion order $\mathcal{O}$ defined over the invariant trace field $E_\tau$; only the subgroup $\overline{\Delta}^{(2)}$ is.
\begin{lem} \label{lem:quotient2312}
Let $\overline{\Delta} = \overline{\Delta}(2,3,12)$, and let
$p \geq 5$ be a rational prime. Then
$$
K(\tau;p) \simeq
\begin{cases}
\mathrm{PSL}_2(\mathbb{F}_p) &: p \equiv 1,23 \, \mathrm{mod} \, 24 \\
\mathrm{PSL}_2(\mathbb{F}_{p^2}) &: p \equiv 5,7,17,19 \, \mathrm{mod} \, 24 \\
\mathrm{PGL}_2(\mathbb{F}_p) &: p \equiv 11, 13 \, \mathrm{mod} \, 24.
\end{cases}
$$
\end{lem}
\begin{proof}
The claim follows from Proposition~\ref{pro:CV.congruence.quotient} and basic algebraic number theory.
\end{proof}
\begin{pro} \label{pro:2312.normal.subgroups}
Let $p \geq 5$ be a rational prime and suppose that $H \trianglelefteq \overline{\Delta} = \overline{\Delta}(2,3,12)$ is a normal subgroup such that $\overline{\Delta} / H \simeq K(\tau;p)$.
Then there exists a prime ideal $\ensuremath{\mathfrak{p}}$ of $\mathbb{Z} [ \sqrt{3} ]$ dividing $p$ such that $H = \overline{\Delta}(\ensuremath{\mathfrak{p}})$.
\end{pro}
\begin{proof}
We will show that there exist at most two normal subgroups $H \vartriangleleft \overline{\Delta}$ satisfying $\overline{\Delta} / H \simeq K(\tau;p)$ if $p$ splits in $\mathbb{Q}(\sqrt{3})$ and at most one such subgroup $H$ if $p$ is inert in $\mathbb{Q}(\sqrt{3})$. Since the congruence subgroups $\overline{\Delta}(\ensuremath{\mathfrak{p}})$ satisfy $\overline{\Delta} / \overline{\Delta}(\ensuremath{\mathfrak{p}}) \simeq K(\tau;p)$ by definition, it follows that no other subgroups satisfy this condition.
Following the strategy of the proof of Proposition~\ref{pro:hurwitz.normal.subgroups}, consider $(\gamma_1, \gamma_2, \gamma_3) \in T(\tau;p)$.
Clearly the orders of $\gamma_1, \gamma_2, \gamma_3$ divide $2$, $3$, and $12$, respectively. If one of these elements were trivial,
the three of them would generate a cyclic group, contradicting the assumption that they generate $K(\tau;p)$. Hence $o(\gamma_1) = 2$ and $o(\gamma_2) = 3$. If $o(\gamma_3) < 12$, then the non-solvable group $K(\tau;p)$ would be a homomorphic image of $\overline{\Delta}(2,2,3)$, $\overline{\Delta}(2,3,4)$, or $\overline{\Delta}(2,3,6)$, which is absurd since these three groups are solvable; see~\cite[Example~2.4]{CV/19} and the paragraph following it. Hence $o(\gamma_3) = 12$. We have shown that $\Omega(\tau;p) = \{ (2,3,12) \}$.
By Lemma~\ref{lem:eigenvalues}, there are two conjugacy class triples $\underline{C}$ satisfying $o(\underline{C}) = (2,3,12)$; the corresponding trace triples up to signs are $(0, \pm 1, \pm(\zeta_{24} + \zeta_{24}^{-1}))$ and $(0, \pm 1, \pm(\zeta_{24}^5 + \zeta_{24}^{-5}))$, where $\zeta_{24}$ is a primitive $24$-th root of unity. These are not commutative, by Lemma~\ref{lem:commutative}, and are not exceptional. It now follows from Proposition~\ref{pro:counting.normal.subgroups} that $| \mathcal{N}(\tau;p) | \leq 2$.
It remains to show that if $p$ is inert in $\mathbb{Q}(\sqrt{3})$, i.e.~if $p \equiv \pm 5 \, \mathrm{mod} \, 12$, then $| \mathcal{N}(\tau;p)| = 1$. In this case, the Frobenius automorphism $\sigma \in \mathrm{Gal}(\mathbb{F}_{p^2} / \mathbb{F}_p)$ interchanges $\pm (\zeta_{24} + \zeta_{24}^{-1})$ and $\pm (\zeta_{24}^5 + \zeta_{24}^{-5})$, and we proceed by the same argument as for the analogous case of Proposition~\ref{pro:hurwitz.normal.subgroups}.
\end{proof}
Observe that $\Delta(2,3,12)^{(2)} \simeq \Delta(3,3,6)$.
We read off from Proposition~\ref{pro:presentation.a} that $A_\tau \simeq \left \langle \frac{-3, 1 + \sqrt{3}}{\mathbb{Q}(\sqrt{3})} \right \rangle$ and that a $\mathbb{Z}[\sqrt{3}]$-basis of the order $\mathcal{Q}_\tau$ is given by $\{ 1, \gamma_1, \gamma_2, \gamma_1 \gamma_2 \}$, where
\begin{eqnarray*}
\gamma_1 & = & \frac{-1 + i}{2} \\
\gamma_2 & = & \frac{\sqrt{3}}{2} + \frac{2 + \sqrt{3}}{6} i + \frac{ij}{3}.
\end{eqnarray*}
Moreover, $\mathcal{Q}_\tau^1 / \{ \pm 1 \} \simeq \overline{\Delta}(3,3,6)$; see the table on p.~208 of~\cite{Takeuchi/77JFS}. However, $\overline{\Delta}(2,3,12)$ is isomorphic to the larger group $\mathcal{Q}_\tau^+ / (E_\tau^\times \cap \mathcal{Q}_\tau^+)$, where $\mathcal{Q}_\tau^+ \leq \mathcal{Q}_\tau^\times$ is the subgroup of elements with totally positive reduced norm. Then $\mathcal{Q}_\tau^+ / (E_\tau^\times \cap \mathcal{Q}_\tau^+)$ naturally contains $\mathcal{Q}_\tau^1 / \{ \pm 1 \}$ as a subgroup of index two.
We tabulate the results of our computations for all prime ideals $\ensuremath{\mathfrak{p}} \vartriangleleft \mathbb{Z}[\sqrt{3}]$ such that $N(\ensuremath{\mathfrak{p}}) < 60$. Note that if $p \in \{ 2, 3 \}$, then $p$ ramifies in $\mathbb{Q}(\sqrt{3})$ and also satisfies $p | abc$, so that Proposition~\ref{pro:2312.normal.subgroups} is inapplicable to primes dividing $p$. However, it is still possible to define congruence subgroups; see~\cite[Remark~5.24]{CV/19}. If $\ensuremath{\mathfrak{p}}_2$ is the prime ideal dividing $2$, one finds explicitly that $\{ g \in \mathcal{Q}_\tau^+ : g \equiv 1 \, \mathrm{mod} \, \ensuremath{\mathfrak{p}}_2 \mathcal{Q}_\tau \} \trianglelefteq \mathcal{Q}_\tau^+$ corresponds to a normal subgroup $H_2 \trianglelefteq \overline{\Delta}$ such that the quotient $\overline{\Delta}/H_2$ is a group of order $48$ isomorphic to the unique non-trivial extension of $A_4$ by $\mathbb{Z} / 4\mathbb{Z}$. This $H$ is isomorphic to a surface group of genus $3$. Similarly, for the prime ideal dividing $3$ one obtains a normal subgroup $H_3 \trianglelefteq \overline{\Delta}$ such that $\overline{\Delta} / H_3 \simeq S_4 \times \mathbb{Z} / 3\mathbb{Z}$ and $H_3$ is isomorphic to a surface group of genus $4$.
The resulting Riemann surfaces of genera $3$ and $4$ are the maximal surfaces mentioned at the beginning of this section. Their systolic lengths were computed by Schmutz Schaller~\cite{Schmutz/93} and match our estimates.
One sees by explicit computation that $\mathcal{Q}_\tau$ is a maximal order in $A_\tau$, and that the only finite place at which $A_\tau$ ramifies is $\ensuremath{\mathfrak{p}}_2$. It follows that both sides of the inequality in the hypotheses of~\cite[Proposition~9.1]{KKSV/16} are equal to $12$, and we are unable to conclude that $\mathrm{sys}(X_{\overline{\Delta}(\ensuremath{\mathfrak{p}})}) > \frac{4}{3} g(X_{\overline{\Delta}(\ensuremath{\mathfrak{p}})})$ for any $\ensuremath{\mathfrak{p}}$; this depends on the coefficient of the second-order term of the series appearing in the proof of~\cite[Proposition~9.1]{KKSV/16}. Nevertheless, from our computational data this appears plausible.
\begin{center}
\begin{longtable}{| r | r || r | r || r | r |} \hline
$\ensuremath{\mathfrak{p}}$ & $N(\ensuremath{\mathfrak{p}})$ & $x =$ lowest observed trace & $2 \arcosh \frac{x}{2}$ & $g(X_{\overline{\Delta}(\ensuremath{\mathfrak{p}})})$ & $\frac{4}{3} \log g$ \\ \hline \hline
$(1 + \sqrt{3})$ & $2$ & $ 4 + 2\sqrt{3}$ & $3.983$ & $3$ & $1.465$ \\ \hline
$(\sqrt{3})$ & $3$ & $5 + 3 \sqrt{3}$ & $4.624$ & $4$ & $1.848$ \\ \hline
$(1 - 2 \sqrt{3})$ & $11$ & $31 + 19 \sqrt{3}$ & $8.314$ & $56$ & $5.367$ \\ \hline
$(1 + 2 \sqrt{3})$ & $11$ & $36 + 21 \sqrt{3}$ & $8.563$ & $56$ & $5.367$ \\ \hline
$(4 - \sqrt{3})$ & $13$ & $35 + 20 \sqrt{3}$ & $8.486$ & $92$ & $6.029$ \\ \hline
$(4 + \sqrt{3})$ & $13$ & $45 + 27 \sqrt{3}$ & $9.038$ & $92$ & $6.029$ \\ \hline
$(2 - 3 \sqrt{3})$ & $23$ & $71 + 40 \sqrt{3}$ & $9.887$ & $254$ & $7.383$ \\ \hline
$(2 + 3 \sqrt{3})$ & $23$ & $96 + 55 \sqrt{3}$ & $10.507$ & $254$ & $7.383$ \\ \hline
$(5)$ & $25$ & $(168 + 94\sqrt{3})\cos \frac{\pi}{12}$ & $11.534$ & $326$ & $7.716$ \\ \hline
$(7 + 2 \sqrt{3})$ & $37$ & $204 + 117 \sqrt{3}$ & $12.016$ & $2110$ & $10.206$ \\ \hline
$(7 - 2 \sqrt{3})$ & $37$ & $324 + 187 \sqrt{3}$ & $12.947$ & $2110$ & $10.206$ \\ \hline
$(1 + 4\sqrt{3})$ & $47$ & $(282 + 164 \sqrt{3})\cos \frac{\pi}{12}$ & $12.608$ & $2163$ & $10.239$ \\ \hline
$(1 - 4 \sqrt{3})$ & $47$ & $(378 + 222 \sqrt{3}) \cos \frac{\pi}{12}$ & $13.204$ & $2163$ & $10.239$ \\ \hline
$(7)$ & $49$ & $341 + 196 \sqrt{3}$ & $13.046$ & $2451$ & $10.406$ \\ \hline
$(4 - 5\sqrt{3})$ & $59$ & $564 + 325 \sqrt{3}$ & $14.054$ & $8556$ & $12.073$ \\ \hline
$(4 + 5\sqrt{3})$ & $59$ & $1115 + 644 \sqrt{3}$ & $15.420$ & $8556$ & $12.073$ \\ \hline
\end{longtable}
\end{center}
\subsection{Bolza twins} \label{sec:bolza}
Consider $\tau = (2,3,8)$. Then $F_\tau = \mathbb{Q} \left( \cos \frac{\pi}{8} \right)$ and $E_\tau = \mathbb{Q}(\sqrt{2})$.
From Proposition~\ref{pro:presentation.a} we find that $A_\tau = \left\langle \frac{-3, \sqrt{2}}{\mathbb{Q}(\sqrt{2})} \right\rangle$ and that the order $\mathcal{Q}_\tau \subset A_\tau$ is spanned over $\mathbb{Z}[\sqrt{2}]$ by $\{1, \gamma_1, \gamma_2, \gamma_1 \gamma_2 \}$, where
$\gamma_1 = \frac{1}{2} (-1 + i)$ and $\gamma_2 = \frac{\sqrt{2}}{2} + \frac{2 + \sqrt{2}}{6} i + \frac{1}{3} ij$. It is easy to see that $\mathcal{Q}_\tau$ is the ``Bolza order'' defined in~\cite[\S7]{KKSV/16}; this follows from noting that $\gamma_1 = \alpha - 1$ and $\gamma_2 = (1 + \sqrt{2})\alpha - \beta$, where $\alpha$ and $\beta$ are as in~\cite{KKSV/16}.
Moreover $\mathcal{Q}_\tau^1 / \{ \pm 1 \} \simeq \overline{\Delta}^{(2)}(\tau) \simeq \overline{\Delta}(3,3,4)$~\cite[\S8]{KKSV/16}. Katz, Katz, the first author, and Vishne~\cite[\S14-16]{KKSV/16} previously computed bounds on $\mathrm{sys}(X_{\overline{\Delta}(\tau^{(2)};\ensuremath{\mathfrak{p}})})$, using the method of Section~\ref{sec:procedure}, for prime ideals $\ensuremath{\mathfrak{p}} \vartriangleleft \mathbb{Z}[\sqrt{2}]$ dividing a rational prime $p$ that splits in $\mathbb{Q}(\sqrt{2})$.
It is an exercise in algebraic number theory to deduce the following from Proposition~\ref{pro:CV.congruence.quotient}.
\begin{pro} \label{pro:bolza}
Let $\tau = (2,3,8)$ and $\tau^{(2)} = (3,3,4)$. Let $\ensuremath{\mathfrak{p}}$ be a prime ideal of $\mathbb{Z}[\sqrt{2}]$ dividing the rational prime $p \geq 5$. Then
\begin{eqnarray*}
\overline{\Delta}(\tau)/\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}}) & \simeq & \begin{cases}
\mathrm{PSL}_2(\mathbb{F}_p) &: p \equiv \pm 1, \, \mathrm{mod} \, 16 \\
\mathrm{PGL}_2(\mathbb{F}_p) &: p \equiv \pm 7, \, \mathrm{mod} \, 16 \\
\mathrm{PGL}_2(\mathbb{F}_{p^2}) &: p \equiv \pm 3, \pm 5 \, \mathrm{mod} \, 16
\end{cases} \\
\overline{\Delta}(\tau^{(2)}) / \overline{\Delta}(\tau^{(2)};\ensuremath{\mathfrak{p}}) & \simeq & \begin{cases}
\mathrm{PSL}_2(\mathbb{F}_p) &: p \equiv \pm 1, \mathrm{mod} \, 8 \\
\mathrm{PSL}_2(\mathbb{F}_{p^2}) &: p \equiv \pm 3, \mathrm{mod} \, 8.
\end{cases}
\end{eqnarray*}
\end{pro}
Let $\ensuremath{\mathfrak{p}} \vartriangleleft \mathbb{Z}[\sqrt{2}]$ be a prime ideal dividing $p \geq 5$.
Since $\overline{\Delta}(\tau^{(2)}; \ensuremath{\mathfrak{p}}) = \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}}) \cap \overline{\Delta}^{(2)}(\tau)$, it follows by Proposition~\ref{pro:bolza} that $\overline{\Delta}(\tau^{(2)}; \ensuremath{\mathfrak{p}}) = \overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})$ exactly when $p \not\equiv \pm 1 \, \mathrm{mod} \, 16$. Thus for half of the primes $p$ splitting in $\mathbb{Q}(\sqrt{2})$ we have $X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})} = X_{\overline{\Delta}(\tau^{(2)};\ensuremath{\mathfrak{p}})}$, whereas for half of them $X_{\overline{\Delta}(\tau^{(2)};\ensuremath{\mathfrak{p}})}$ is a non-trivial cover of $X_{\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}})}$ and may have larger systolic length.
The prime $\ensuremath{\mathfrak{p}}_2 = (\sqrt{2})$ does not fit into the framework of this paper. However, analogously to the previous section, it gives rise to a normal subgroup $H_2 \trianglelefteq \overline{\Delta}(\tau^{(2)})$ of index $24$ such that $\overline{\Delta}(\tau^{(2)}) / H_2 \simeq \mathrm{SL}_2(\mathbb{F}_3)$ and such that $H_2$ is isomorphic to a surface group of genus two. The Riemann surface $X_{H_2}$ is the Bolza surface, which is known to be maximal. The four generators in our presentation of $H_2$ all correspond to elements of $\mathcal{Q}_\tau^1 / \{ \pm 1 \}$ of trace $\pm 2 (1 + \sqrt{2})$, which give the true systolic length of the Bolza surface~\cite[\S5]{Schmutz/93}. We observe in passing two other interesting normal subgroups $H_3, H_5 \trianglelefteq \overline{\Delta}(\tau^{(2)})$ of indices $48$ and $96$. They are isomorphic to surface groups of genera $3$ and $5$, respectively, and neither appears to be a congruence subgroup. We have $\overline{\Delta}(\tau^{(2)})/H_3 \simeq (\mathbb{Z} / 4\mathbb{Z})^2 \rtimes \mathbb{Z} / 3\mathbb{Z}$, with $\mathbb{Z} / 3\mathbb{Z}$ acting faithfully, and the generators of $H_3$ correspond to traces $\pm 2 (2 + \sqrt{2})$. Finally, $\overline{\Delta}(\tau^{(2)})/H_5 \simeq (\mathbb{Z} / 2\mathbb{Z})^3 \rtimes A_4$, with $A_4$ acting via its quotient $\mathbb{Z} / 3\mathbb{Z}$. The generators of $H_5$ correspond to traces $\pm 2(3 + 2 \sqrt{2})$. Note that Schmutz Schaller~\cite[\S7]{Schmutz/93} discusses a maximal surface of genus $5$ and systolic length $2 \arcosh (3 + 2 \sqrt{2})$.
\subsection{$(2,7,7)$ triangle surfaces} \label{sec:277}
The case of $\tau = (2,7,7)$ is interesting for at least two reasons. First of all, the analogue of Propositions~\ref{pro:hurwitz.normal.subgroups} and~\ref{pro:2312.normal.subgroups} fails. Secondly, $\overline{\Delta}(2,7,7)$ is not maximal among triangle groups, since it embeds in $\overline{\Delta}(2,3,7)$ as a self-normalizing subgroup of index nine~\cite[Theorem~2]{Singerman/72}. An explicit embedding is given by
\begin{eqnarray*}
f: \overline{\Delta}(2,7,7) = \langle x^\prime ,y^\prime \rangle / \langle (x^\prime)^2, (y^\prime)^7, (x^\prime y^\prime)^7 = 1 \rangle & \to & \langle x, y \rangle / \langle x^2, y^3, (xy)^7 \rangle = \overline{\Delta}(2,3,7)\\
x^\prime & \mapsto & x \\
y^\prime & \mapsto & (xy)^4 y,
\end{eqnarray*}
and its image is unique up to conjugation.
In this situation, $E_\tau = F_\tau = \mathbb{Q}(\cos \frac{\pi}{7})$ is the same field as in Section~\ref{sec:hurwitz}. Recall that we set $\mu = 2 \cos \frac{\pi}{7}$. The presentation of the quaternion algebra $B_\tau = A_\tau$ given by Proposition~\ref{pro:alpha.beta.fulldelta} is $\left \langle \frac{-4, 2 \mu^2 - 4}{\mathbb{Q}(\mu)} \right\rangle$. This is actually the same algebra as in Section~\ref{sec:hurwitz}, as evidenced by the isomorphism
\begin{eqnarray*}
\varphi: \left \langle \frac{-4, 2 \mu^2 - 4}{\mathbb{Q}(\mu)} \right\rangle & \to & \left \langle \frac{-4, \mu^2 - 3}{\mathbb{Q}(\mu)} \right\rangle \\
\varphi(i) & = & i \\
\varphi(j) & = & (\mu^2 - 1) \left( -j + \frac{ij}{2} \right).
\end{eqnarray*}
We work with the presentation of Section~\ref{sec:hurwitz}. The (image under $\varphi$ of) the order $\mathcal{O}_{(2,7,7)}$ is the $\mathbb{Z}[\mu]$-span of $\{1, \alpha, \beta^\prime, \alpha \beta^\prime \}$, for $\alpha = \frac{i}{2}$ and $\beta^\prime = \frac{1}{4} \left( 2 \mu + \mu i + 2(\mu^2 - 1)j + (\mu^2 - 1)ij \right)$. It is easy to check that $\mathcal{O}_{(2,7,7)} \subset \mathcal{O}_{(2,3,7)}$ and that $\mathcal{O}_{(2,7,7)}$ has discriminant $(8)$. We have $\mathcal{O}^1_{(2,7,7)} / \{ \pm 1 \} \simeq \overline{\Delta}(2,7,7)$. If $\tau$ is either $(2,3,7)$ or $(2,7,7)$, it follows from the construction in~\cite[\S5]{CV/19} that $\overline{\Delta}(\tau;\ensuremath{\mathfrak{p}}) = \{ g \in \mathcal{O}_\tau^1 : g \equiv \pm 1 \, \mathrm{mod} \, \ensuremath{\mathfrak{p}} \mathcal{O}_\tau \} / \{ \pm 1 \}$. Hence
$$\overline{\Delta}(2,7,7;\ensuremath{\mathfrak{p}}) = \overline{\Delta}(2,3,7;\ensuremath{\mathfrak{p}}) \cap \overline{\Delta}(2,7,7)$$
for all prime ideals $\ensuremath{\mathfrak{p}} \vartriangleleft \mathbb{Z}[\mu]$ not dividing $2$. In this case, $X_{\overline{\Delta}(2,7,7;\ensuremath{\mathfrak{p}})}$ is a cover of degree nine of $X_{\overline{\Delta}(2,3,7,\ensuremath{\mathfrak{p}})}$, and thus the length spectrum of $X_{\overline{\Delta}(2,7,7;\ensuremath{\mathfrak{p}})}$ is a subset of that of $X_{\overline{\Delta}(2,3,7;\ensuremath{\mathfrak{p}})}$. The following table collects our estimates for $\mathrm{sys}(X_{\overline{\Delta}(2,7,7;\ensuremath{\mathfrak{p}})})$ for some primes $\ensuremath{\mathfrak{p}}$. The last column gives the rank of our candidate for $\mathrm{sys}(X_{\overline{\Delta}(2,7,7;\ensuremath{\mathfrak{p}})})$ in the length spectrum of $X_{\overline{\Delta}(2,3,7;\ensuremath{\mathfrak{p}})}$, as computed by Vogeler~\cite{Vogeler/03}. Where the entry in this column is $1$, we have certainly obtained the true value of $\mathrm{sys}(X_{\overline{\Delta}(2,7,7;\ensuremath{\mathfrak{p}})})$.
\begin{center}
\begin{longtable}{| r | r || r | r || r | r |} \hline
$I$ & $N(I)$ & $x =$ lowest observed trace & $2 \arcosh \frac{x}{2}$ & $g(X_{\overline{\Delta}(2,7,7;I)})$ & \\ \hline \hline
$(\mu + 2)$ & $7$ & $ 9 \mu^2 + 8 \mu - 4$ & $7.358$ & $19$ & $3$ \\ \hline
$(2)$ & $8$ & $4 \mu^2 + 4 \mu - 2$ & $5.796$ & $49$ & $1$ \\ \hline
$(2 \mu + 1)$ & $13$ & $17 \mu^2 + 12 \mu - 10$ & $8.404$ & $118$ & $2$ \\ \hline
$(2 \mu + 3)$ & $13$ & $6 \mu^2 + 5 \mu - 4$ & $6.393$ & $118$ & $1$ \\ \hline
$(\mu - 3)$ & $13$ & $8 \mu^2 + 7 \mu - 4$ & $7.085$ & $118$ & $2$ \\ \hline
$(3)$ & $27$ & $135 \mu^2 + 108 \mu - 74$ & $12.652$ & $1054$ & $3$ \\ \hline
$(4 \mu - 1)$ & $29$ & $223 \mu^2 + 180 \mu - 124$ & $13.658$ & $1306$ & $4$ \\ \hline
$(3 \mu - 4)$ & $29$ & $49 \mu^2 + 41 \mu - 27$ & $10.656$ & $1306$ & $1$ \\ \hline
$(\mu + 3)$ & $29$ & $73 \mu^2 + 60 \mu - 40$ & $11.442$ & $1306$ & $1$ \\ \hline
\end{longtable}
\end{center}
The case of $\ensuremath{\mathfrak{p}} = (2)$, which divides the discriminant of $\mathcal{O}_{(2,7,7)}$, is exceptional. The congruence subgroup $\overline{\Delta}(2,3,7;(2))$ corresponding to the Fricke-Macbeath curve of genus $7$ is contained in (any embedding of) $\overline{\Delta}(2,7,7)$ into $\overline{\Delta}(2,3,7)$. Thus the Fricke-Macbeath curve is a $(2,7,7)$-triangle surface. However,
$$\overline{\Delta}(2,7,7;(2)) = \{ g \in \mathcal{O}^1_{(2,7,7)} : g \equiv \pm 1 \, \mathrm{mod} \, 2 \mathcal{O}_{(2,7,7)} \}$$
is a subgroup of index eight in $\overline{\Delta}(2,3,7;(2))$. Thus $X_{\overline{\Delta}(2,7,7;(2))}$ is a cover of degree $8$ of the Fricke-Macbeath curve, although it has the same systolic length.
It follows from Proposition~\ref{pro:counting.normal.subgroups}, by an argument that we have already seen in several previous examples, that $| \mathcal{N}(\tau;p) | \leq 9$ for all primes $p \not\in \{ 2, 7 \}$. The smallest primes that split in $\mathbb{Q}(\mu)$ are $13$ and $29$. For either $p \in \{13, 29 \}$, we find nine normal subgroups $H \trianglelefteq \overline{\Delta}(2,7,7)$ such that $\overline{\Delta}(2,7,7)/H \simeq \mathrm{PSL}_2(\mathbb{F}_p)$, corresponding to the nine triples $\underline{C} = (C_1, C_2, C_3)$ of conjugacy classes of $\mathrm{PSL}_2(\mathbb{F}_p)$ with $o(\underline{C}) = (2,7,7)$; note that there is only one possibility for $C_1$, but three for $C_2$ and $C_3$. In each case, the three subgroups associated to triples $\underline{C}$ with $C_2 = C_3$ are the congruence subgroups arising from prime ideals dividing $p$. The six remaining normal subgroups in each case are also isomorphic to surface groups of genera $118$ or $1306$. They are not all conjugate in $\mathrm{PSL}_2(\mathbb{R})$; note that Theorems~5 and~6 of~\cite{GW/05} tell us where to look for conjugating elements. However, for all twelve of these groups, the lowest observed trace is $11 \mu^2 + 9 \mu - 7$, corresponding to a systolic length of $7.609\dots$ This is precisely $\mathrm{sys}(X_S)$, where $S \trianglelefteq \overline{\Delta}(2,3,7)$ is either of Sinkov's non-congruence normal subgroups of index $1344$. Similarly, there are three normal subgroups of $\overline{\Delta}(2,7,7)$ with quotient isomorphic to $\mathrm{PSL}_2(\mathbb{F}_{27})$. One of them is the congruence subgroup $\overline{\Delta}(2,7,7;(3))$. The other two are surface groups of genus $1054$, and the lowest observed trace is again $11 \mu^2 + 9 \mu - 7$. There are two non-congruence normal subgroups $H \trianglelefteq \overline{\Delta}(2,7,7)$ with $\overline{\Delta}(2,7,7)/H \simeq \mathrm{PSL}(\mathbb{F}_8)$, whose lowest observed trace is the same mysterious $11 \mu^2 + 9 \mu - 7$. While the pairwise intersections of all these groups with the same lowest observed trace are small, a very disproportionate number of the generators we use lie in these intersections. This may account for our computations. We hope to return to this interesting phenomenon in future work.
\subsection{$(3,3,10)$ triangle surfaces}
Finally, consider the triple $\tau = (3,3,10)$. By~\cite[Theorem~3]{Takeuchi/77}, the triangle group $\overline{\Delta} = \overline{\Delta}(\tau)$ is not arithmetic. We have $F_\tau = E_\tau = \mathbb{Q} (\nu)$, where $\nu = 2 \cos \frac{\pi}{10}$ has minimal polynomial $x^4 - 5x^2 + 5$ over $\mathbb{Q}$. This number field has class number $1$. The quaternion algebra $B_\tau = A_\tau = \left\langle \frac{-3, \nu^2 + \nu - 2}{F_\tau} \right\rangle$ splits at two of the four infinite places of $F_\tau$ and by Proposition~\ref{pro:alpha.beta.fulldelta} the order $\mathcal{O}_\tau \subset B_\tau$ is spanned over $\mathcal{O}_{F_\tau}$ by $\{ 1, \alpha, \beta, \alpha \beta \}$, where $\alpha = \frac{1}{2} + \frac{i}{2}$ and $\beta = \frac{1}{2} + \frac{1 + 2 \nu}{6} i - \frac{1}{3} ij$. By Theorem~\ref{thm:semiarithmetic} we have $\mathrm{sys}(X_{\overline{\Delta}(I)}) > \frac{2}{3} \log g(X_{\overline{\Delta}(I)}) - c$ for all ideals $I \triangleleft \mathcal{O}_{F_\tau}$ coprime to $30$.
\begin{lemma}
Let $\tau = (3,3,10)$ and let $p \geq 7$ be a rational prime. Then
$$ K(\tau;p) \simeq \begin{cases}
\mathrm{PSL}_2(\mathbb{F}_p) &: p \equiv 1, 19 \, \mathrm{mod} \, 20 \\
\mathrm{PSL}_2(\mathbb{F}_{p^2}) &: p \equiv 9, 11 \, \mathrm{mod} \, 20 \\
\mathrm{PSL}_2(\mathbb{F}_{p^4}) &: p \equiv 3, 7, 13, 17 \, \mathrm{mod} \, 20.
\end{cases}$$
\end{lemma}
\begin{proof}
The claim follows from Proposition~\ref{pro:CV.congruence.quotient} and reasoning similar to the proof of Lemma~\ref{lem:hurwitz.decomposition}.
\end{proof}
We tabulate our bounds on $\mathrm{sys}(X_{\overline{\Delta}(I)})$ for some prime ideals of $\mathcal{O}_{F_\tau}$. It is evident from the table that there exist ideals for which $\mathrm{sys}(X_{\overline{\Delta}(I)}) < \frac{4}{3} \log g(X_{\overline{\Delta}(I)})$.
\begin{center}
\begin{longtable}{| r | r || r | r || r | r |} \hline
$\ensuremath{\mathfrak{p}}$ & $N(\ensuremath{\mathfrak{p}})$ & $x =$ lowest observed trace & $2 \arcosh \frac{x}{2}$ & $g(X_{\overline{\Delta}(\ensuremath{\mathfrak{p}})})$ & $\frac{2}{3} \log g$ \\ \hline \hline
$(\nu^2 + \nu - 4)$ & $19$ & $21 + 12 \nu - 18 \nu^2 - 10 \nu^3$ & $9.002$ & $400$ & $3.994$ \\ \hline
$(\nu^2 - \nu - 4)$ & $19$ & $24 + 10 \nu - 20 \nu^2 - 10 \nu^3$ & $9.173$ & $400$ & $3.994$ \\ \hline
$(\nu^3 - \nu^2 - 3 \nu + 1)$ & $19$ & $40 + 18 \nu - 32 \nu^2 - 16 \nu^3$ & $10.043$ & $400$ & $3.994$ \\ \hline
$(\nu^3 + \nu^2 - 3 \nu - 1)$ & $19$ & $86 + 46 \nu - 64 \nu^2 - 34 \nu^3$ & $11.354$ & $400$ & $3.994$ \\ \hline
$(\nu^3 - 3 \nu - 3)$ & $41$ & $6 + 4 \nu - 7 \nu^2 - 4 \nu^3$ & $7.338$ & $4019$ & $5.533$ \\ \hline
$(\nu - 3)$ & $41$ & $36 + 18 \nu - 27 \nu^2 - 14 \nu^3$ & $9.637$ & $4019$ & $5.533$ \\ \hline
$(\nu^3 - 3 \nu + 3)$ & $41$ & $43 + 20 \nu - 32 \nu^2 - 16 \nu^3$ & $9.951$ & $4019$ & $5.533$ \\ \hline
$(\nu + 3)$ & $41$ & $243 + 125 \nu - 178 \nu^2 - 93 \nu^3$ & $13.377$ & $4019$ & $5.533$ \\ \hline
$(\nu^3 + \nu^2 - 2\nu - 4)$ & $59$ & $74 + 36\nu - 57 \nu^2 - 29 \nu^3$ & $11.147$ & $11978$ & $6.261$ \\ \hline
$(\nu^3 - \nu^2 - 4 \nu + 1)$ & $59$ & $90 + 45\nu - 66 \nu^2 - 34 \nu^3$ & $11.389$ & $11978$ & $6.261$ \\ \hline
$(\nu^3 + \nu^2 - 4\nu - 1)$ & $59$ & $301 + 156 \nu - 218 \nu^2 - 114 \nu^3$ & $13.766$ & $11978$ & $6.261$ \\ \hline
$(\nu^3 - \nu^2 - 2\nu + 4)$ & $59$ & $383 + 200 \nu - 278 \nu^2 - 146 \nu^3$ & $14.257$ & $11978$ & $6.261$ \\ \hline
$(3)$ & $81$ & $259 + 135 \nu - 189 \nu^2 - 99 \nu^3$ & $13.489$ & $30997$ & $6.894$ \\ \hline
\end{longtable}
\end{center}
\begin{acknowledgements}
We are grateful to Mikhail Katz and John Voight for helpful conversations and correspondence and to the anonymous referee for comments that improved the exposition.
\end{acknowledgements}
\bibliographystyle{amsplain}
| 3c88f24e645f50fecc5b0b88cf9a6aa2f84e159c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction: understanding red giant dynamics}
\label{sec:1}
Amongst key observations of red-giant (RG) interiors that stands out are their interestingly low core rotation rate \citep[e.g.][]{beck2011, Mosser2012, Gehan2018} that is only about 10 times higher than the surface rate. However, state-of-the-art stellar evolution models predict that the core should rotate about 100 times faster than the stellar surface according to current transport models \citep[e.g.][]{Eggenberger2012b, Ceillier2013, Marques2013a, Cantiello2014}. This highlights an strong bottleneck concerning the understanding of the strenght of the angular momentum transport inside evolved Solar-like stars. This is complemented by the surprising discovery of two populations of RG where the character of their oscillations depends upon their masses \citep[e.g.][]{garcia2014b, stello2016, Mosser2017a}. One the one hand, the lower mass (1 M$_\odot$ < M$_\star$ < 1.3 M$_\odot$) range of stars exhibit solar-like oscillations with normal amplitude. On the other hand, nearly 50\% of the more massive group (1.5 M$_\odot$ < M$_\star$ < 2.5 M$_\odot$) have disturbed power spectrum densities where some mixed modes have very low amplitudes. It emphasises that some oscillation modes are weakened and sometimes even damped in the core of a large fraction of RG. Once again, the current simplified theory needs to be corrected to account for the origin of these weakened modes.\\
The idea of a strong stable buried magnetic field modifying the oscillation modes of red giants (RG) came from \cite{Fuller2015a} who invoked a conversion from gravity waves towards magneto-gravity waves that would remain trapped inside the core of the star. This theory is controversial \citep{Mosser2017a}, but has been largely studied and improved \citep{Loi2017a, Lecoanet2017b}: magnetism is still a very good candidate for both angular momentum transport \citep{Fuller2019, Eggenberger2019} and mode supression. It was shown in \cite{Braithwaite2004} that the relaxation of a stochastic field in a radiative zone leads to the formation of a stable fossil mixed poloidal and toroidal magnetic field. The energy minimization principle employed in \cite{Duez2010} yields analytical approximations to such a mixed field solution. We use this configuration as provided in \cite{Duez2010a} to model a fossil magnetic field inside RG being the remnant of the field generated by the convective core dynamo during the main sequence (\textsc{ms}) for M$_\star$ > 1.3 M$_\odot$ (see Fig.~\ref{fig:1}).
We then theoretically investigate the effect of this stable fossil mixed poloidal and toroidal field buried inside the core of RG on their mixed-mode frequencies. In this study, we focus on the axisymmetric case, where the magnetic axis is aligned with the rotation axis of the star. The convective envelope is also represented by the vortex symbols.\\
\begin{figure}[t]
\sidecaption
\includegraphics[scale=.36]{Bugnet_Fig1BW.png}
\caption{Representation of a RG with a trapped magnetic field inside its core. The magnetic field has a poloidal (white lines) and a toroidal (grey scale) components normalised by the strenght $B_0$ of the field and is computed by using the semi-analytic description of \cite{Duez2010, Duez2010a}.}
\label{fig:1}
\end{figure}
\section{Perturbative methods}
\label{sec:3}
We use the linearised equations for stellar oscillations \cite{Unno1989}, perturbed by both rotation and magnetism to evaluate the first-order frequency perturbation $\delta \omega$ through
\begin{equation}
\delta \omega=-\frac{\langle \fvec{\xi_0},\delta \fvec{F_\ell}(\fvec{\xi_0}) \rangle + \langle \fvec{\xi_0},\fvec{F_c}(\fvec{\xi_0}) \rangle}{2 \omega_0 \langle \fvec{\xi_0}, \fvec{\xi_0} \rangle},
\label{eq:1}
\end{equation}
with $\delta \fvec{F_\ell}$ being the perturbed volumetric Lorentz force, $\fvec{F_c}$ the Coriolis acceleration, $\fvec{\xi_0}$ the unperturbed eigenfunctions of mixed modes, and $\omega_0$ the unperturbed mixed-mode frequencies. For each of the terms composing Eq.~\ref{eq:1} (see \cite{Prat2019} for more details), we numerically estimate the mass mode two following dominant components
\begin{equation}
\langle \fvec{\xi_0},\fvec{F_c}(\fvec{\xi_0}) \rangle \simeq 8 \pi m \Omega \int_0^R \rho r^2 |\xi_h|^2\textrm{d}r \int_0^\pi Y_\ell^m \partial_\theta Y_\ell^{m*}\frac{\cos{\theta}}{\sin{\theta}} \textrm{d}\theta,
\label{eq:3}
\end{equation}
\begin{multline}
\langle \fvec{\xi_0},\delta \fvec{F_\ell}(\fvec{\xi_0}) \rangle \simeq 2 \pi B_0^2 \int_0^R r \xi_h^* b_r (r\xi_h b_r)'' \textrm{d}r \\
\times \int_0^\pi\left(\left(\frac{mY_\ell^m}{\sin{\theta}} \right)^2 + (\partial_\theta Y_\ell^m)^2 \right)\cos^2{\theta} \sin{\theta}\textrm{d}\theta,
\label{eq:2}
\end{multline}
and the mass mode
\begin{equation}
\langle \fvec{\xi_0},\fvec{\xi_0} \rangle \simeq 2 \pi \int_0^R \rho r^2 |\xi_h|^2 \textrm{d}r \int_0^\pi\left(\left(\frac{mY_\ell^m}{\sin{\theta}} \right)^2 + (\partial_\theta Y_\ell^m)^2 \right) {\sin{\theta}} \textrm{d}\theta,
\label{eq:4}
\end{equation}
\noindent
with $B_0$ the magnetic field strength, $\xi_h(r)$ the radial eigenfunction describing the horizontal displacement on the spherical harmonics ($Y_\ell^m(\theta,\varphis)$ where $\ell$ is the degree, m the azimuthal degree, and $\left(r, \theta, \varphi\right)$ are the usual spherical coordinates), $b_r(r)$ the radial function describing the radial component of the magnetic field on the $\ell=1$ dipolar spherical harmonics, $\Omega$ the angular velocity, and $\rho$ the density profile of the star. The unperturbed eigenfunctions of mixed modes ($\fvec{\xi_0}$) are evaluated by using the stellar pulsation code GYRE \citep{Townsend2013} along with the stellar evolution model MESA \citep{Paxton2011}.
\section{Mixed-mode splittings in RG}
\label{sec:3}
\begin{figure}[t]
\includegraphics[scale=0.32]{Bugnet_Fig2new.png}
\caption{\textsl{Left:} Mode splittings ($\delta {\nu}$) induced by magnetism only for a M$_\star$ = 1.5 M$_\odot$ early RG with $\nu_\textnormal{max}=$ $425\upmu$Hz as a function of mode frequencies. The light gray area represents the range of frequencies $[\nu_\textnormal{max}-5\Delta \nu: \nu_\textnormal{max} +5\Delta \nu]$ for which mixed modes should be visible in real data. Each of them represents a simulated mixed mode with the lower line representing $m=$0 and the upper line $m=$1 and $m=$-1 modes, which overlap as a result of the m dependency in Eq.~\ref{eq:2}. \textsl{Right:} Same as left panel with rotation splitting corresponding to a uniform rotation of 0.7 $\upmu$Hz added.}
\label{fig:2}
\end{figure}
Figure~\ref{fig:2} shows the strength of the frequency perturbation of mixed modes due to the magnetic field only (left panel) for a field strength of $B_0=1$ MG. We observe that one individual $\ell=1$ mixed mode splits into two modes: the $m=$0 component and the $m=$1 and $m=$-1 components (overlapping as described by Eq.~\ref{eq:2}). In the frequency region where we expect to see mixed modes in the power spectrum density (light gray area), magnetic splittings are of the order of 0.1 $\upmu$Hz. This is a small effect by comparison with the rotational frequency splitting due to a uniform rotation of the star of $0.7 \upmu$Hz as shown on the right panel. Nevertheless, the effect of such a magnetic field can be large enough to be observable with current instruments.\\
The dependence of the magnetic splittings on the evolution of the star and on the magnetic field strength is represented in Fig.~\ref{fig:3}. The right panel indicates typical data frequency boundaries for the observation of magnetic splittings to be compared to the left panel on which values of magnetic splitting are represented: if the combination of magnetic field strength and of the evolution leads to a magnetic splitting below the \textit{Kepler} resolution, then the effect of magnetism on mixed-mode frequencies would not be visible in \textit{Kepler} data. Above the typical $\ell =0$ linewidth the magnetic effect is considered as being too large for our perturbative analysis to be valid (see Bugnet et al., \textsl{in prep}). The region in between the two black lines approximatively shows the ideal conditions for the detection of the magnetic effect in \textit{Kepler} data. This region would be narrowed to the area between the TESS 1 year white line and the $\ell$=1 linewidth black line if considering TESS 1 year data, and we do not expect to be able to detect any magnetic effect on mixed-mode frequencies when looking at only 1 month TESS data.
\begin{figure}[t]
\sidecaption[t]
\includegraphics[scale=.72]{Bugnet_Fig3BW.png}
\caption{\textsl{Left:} Mode splitting induced by magnetism only for a M$_\star$ = 1.5 M$_\odot$ giant as a function of the frequency of maximum power $\nu_\textnormal{max}$ and the magnetic field strength. The mode splitting is calculated at each evolutionary stage by taking the splitting corresponding to the central mixed mode (the closest to $\nu_\textnormal{max}$). \textsl{Right:} Same diagram as the left panel simplified to compare with typical frequencies: the "\textit{Kepler} resolution" line corresponds to the 7.9 nHz resolution for the 4 years \textit{Kepler} data, the "TESS 1 year" line corresponds to the 32 nHz resolution for 1 year of TESS data, the "TESS 1 month" line corresponds to the 380 nHz resolution for 1 month of TESS data, and the "$\ell=0$" line corresponds to the typical radial mode linewidths \citep{Vrard2017} for sub-giants, as being an upper limit for mixed-mode linewidth.}
\label{fig:3}
\end{figure}
\section{Conclusion \& Perspectives}
We study the axisymmetric case for which a buried stable fossil magnetic field, with poloidal and toroidal components, is aligned with the rotation axis of the star. We find that a typical field strength of about 1MG results in a detectable frequency shift of the mixed modes inside RG's power spectrum density in addition to the already measured rotational splittings. The study will be extended towards fossil dipolar fields inclined with respect to the rotational axis of the star, and towards non-fossil field topologies such as those derived in \cite{Spruit1999, Jouve2014, Fuller2019}, in order to explore the variety of possible seismic signatures that would allow us to probe the deep magnetism of evolved low-mass stars and its consequences for the transport of angular momentum.
\bibliographystyle{spphys}
\section{Introduction: understanding red giant dynamics}
\label{sec:1}
Amongst key observations of red-giant (RG) interiors that stands out are their interestingly low core rotation rate \citep[e.g.][]{beck2011, Mosser2012, Gehan2018} that is only about 10 times higher than the surface rate. However, state-of-the-art stellar evolution models predict that the core should rotate about 100 times faster than the stellar surface according to current transport models \citep[e.g.][]{Eggenberger2012b, Ceillier2013, Marques2013a, Cantiello2014}. This highlights an strong bottleneck concerning the understanding of the strength of the angular momentum transport inside evolved Solar-like stars. This is complemented by the surprising discovery of two populations of RG where the character of their oscillations depends upon their masses \citep[e.g.][]{garcia2014b, Stello2016a, Mosser2017a}. On the one hand, the lower mass (1 M$_\odot$ < M$_\star$ < 1.3 M$_\odot$) range of stars exhibit solar-like oscillations with normal amplitude. On the other hand, nearly 50\% of the more massive group (1.5 M$_\odot$ < M$_\star$ < 2.5 M$_\odot$) exhibit some mixed modes with very low amplitudes. This emphasises that some oscillation modes are weakened and sometimes they disappear in the core of a large fraction of RG. Once again, the current simplified theory needs to be modified to account for the origin of these weakened modes.\\
The idea of a strong stable buried magnetic field altering the oscillation modes of red giants (RG) came from \cite{Fuller2015a} who invoked a conversion from gravity waves towards magneto-gravity waves that would remain trapped inside the core of the star. This theory is controversial \citep{Mosser2017a}, but has been systematically studied and improved \citep{Cantiello2016, Loi2017a, Lecoanet2017b}, revealing that magnetism is still a very good candidate for both angular momentum transport \citep{Fuller2019, Eggenberger2019} and mode supression. It was shown in \cite{Braithwaite2004} that the relaxation of a stochastic field in a radiative zone leads to the formation of a stable, fossil, mixed poloidal and toroidal magnetic field. The energy minimization principle employed in \cite{Duez2010} yields analytical approximations to such a mixed field solution. We use this configuration as provided in \cite{Duez2010a} to model a fossil magnetic field inside RG being the remnant of the field generated by the convective core dynamo during the main sequence (\textsc{ms}) for M$_\star$ > 1.3 M$_\odot$ (see Fig.~\ref{fig:1}).
We then theoretically investigate the effect of this stable fossil mixed poloidal and toroidal field buried inside the core of RG on their mixed-mode frequencies. In this study, we focus on the axisymmetric case, where the magnetic axis is aligned with the rotation axis of the star. The convective envelope is also represented by the vortex symbols.\\
\begin{figure}[t]
\sidecaption
\includegraphics[scale=.35]{Bugnet_Fig1BW.png}
\caption{Representation of a trapped magnetic field inside an RG core. The magnetic field has a poloidal (white lines) and a toroidal (grey scale) components normalised by the strength $B_0$ of the field computed by using the semi-analytic description of \cite{Duez2010, Duez2010a}.}
\label{fig:1}
\end{figure}
\section{Perturbative methods}
\label{sec:3}
We use the linearised equations for stellar oscillations \cite{Unno1989}, perturbed by both rotation and magnetism to evaluate the first-order frequency perturbation $\delta \omega$ through
\begin{equation}
\delta \omega=-\frac{\langle \fvec{\xi_0},\delta \fvec{F_\ell}(\fvec{\xi_0}) \rangle + \langle \fvec{\xi_0},\fvec{F_c}(\fvec{\xi_0}) \rangle}{2 \omega_0 \langle \fvec{\xi_0}, \fvec{\xi_0} \rangle},
\label{eq:1}
\end{equation}
with $\delta \fvec{F_\ell}$ being the perturbed volumetric Lorentz force, $\fvec{F_c}$ the Coriolis acceleration, $\fvec{\xi_0}$ the unperturbed eigenfunctions of mixed modes, and $\omega_0$ the unperturbed mixed-mode frequencies. For each of the terms composing Eq.~\ref{eq:1} (see \cite{Prat2019} for more details), we numerically estimate the mass mode two following dominant components
\begin{equation}
\langle \fvec{\xi_0},\fvec{F_c}(\fvec{\xi_0}) \rangle \simeq 8 \pi m \Omega \int_0^R \rho r^2 |\xi_h|^2\textrm{d}r \int_0^\pi Y_\ell^m \partial_\theta Y_\ell^{m*}\frac{\cos{\theta}}{\sin{\theta}} \textrm{d}\theta,
\label{eq:3}
\end{equation}
\begin{multline}
\langle \fvec{\xi_0},\delta \fvec{F_\ell}(\fvec{\xi_0}) \rangle \simeq 2 \pi B_0^2 \int_0^R r \xi_h^* b_r (r\xi_h b_r)'' \textrm{d}r \\
\times \int_0^\pi\left(\left(\frac{mY_\ell^m}{\sin{\theta}} \right)^2 + (\partial_\theta Y_\ell^m)^2 \right)\cos^2{\theta} \sin{\theta}\textrm{d}\theta,
\label{eq:2}
\end{multline}
and the mode mass
\begin{equation}
\langle \fvec{\xi_0},\fvec{\xi_0} \rangle \simeq 2 \pi \int_0^R \rho r^2 |\xi_h|^2 \textrm{d}r \int_0^\pi\left(\left(\frac{mY_\ell^m}{\sin{\theta}} \right)^2 + (\partial_\theta Y_\ell^m)^2 \right) {\sin{\theta}} \textrm{d}\theta,
\label{eq:4}
\end{equation}
\noindent
with $B_0$ the magnetic field strength, $\xi_h(r)$ the radial eigenfunction describing the horizontal displacement on the spherical harmonics ($Y_\ell^m(\theta,\varphi)$ where $\ell$ is the degree, m the azimuthal degree, and $\left(r, \theta, \varphi\right)$ are the usual spherical coordinates), $b_r(r)$ the radial function describing the radial component of the magnetic field on the $\ell=1$ dipolar spherical harmonics, $\Omega$ the angular velocity, and $\rho$ the density profile of the star. The unperturbed eigenfunctions of mixed modes ($\fvec{\xi_0}$) are evaluated by using the stellar pulsation code GYRE \citep{Townsend2013} along with the stellar evolution model MESA \citep{Paxton2011}.
\section{Mixed-mode splittings in RG}
\label{sec:3}
\begin{figure}[t]
\includegraphics[scale=0.31]{Bugnet_Fig2new.png}
\caption{\textsl{Left:} Mode splittings ($\delta {\nu}$) induced by magnetism only for a M$_\star$ = 1.5 M$_\odot$ early RG with $\nu_\textnormal{max}=$ $425\upmu$Hz as a function of mode frequencies. The light gray area represents the range of frequencies $[\nu_\textnormal{max}-5\Delta \nu: \nu_\textnormal{max} +5\Delta \nu]$ for which mixed modes should be visible in real data. Each of them represents a simulated mixed mode with the lower line representing $m=$0 and the upper line $m=$1 and $m=$-1 modes, which overlap as a result of the m dependency in Eq.~\ref{eq:2}. \textsl{Right:} Same as left panel with rotation splitting corresponding to a uniform rotation of 0.7 $\upmu$Hz added.
}
\label{fig:2}
\end{figure}
Figure~\ref{fig:2} shows the strength of the frequency perturbation of mixed modes due to the magnetic field only (left panel) for a field strength of $B_0=1$ MG. We observe that one individual $\ell=1$ mixed mode splits into two modes: the $m=$0 component and the $m=$1 and $m=$-1 components (overlaping as described by Eq.~\ref{eq:2}). In the frequency region where we expect to see mixed modes in the power spectrum density (light gray area), magnetic splittings are of the order of 0.1 $\upmu$Hz. This is a small effect by comparison with the rotational frequency splitting due to a uniform rotation of the star of $0.7 \upmu$Hz as shown on the right panel. Nevertheless, the effect of such a magnetic field can be large enough to be observable with current instruments.\\
The dependence of the magnetic splittings on the evolution of the star and on the magnetic field strength is represented in Fig.~\ref{fig:3}. The right panel indicates typical data frequency boundaries for the observation of magnetic splittings to be compared to the left panel on which values of magnetic splitting are represented: if the combination of magnetic field strength and of the evolution leads to a magnetic splitting below the \textit{Kepler} resolution, then the effect of magnetism on mixed-mode frequencies would not be visible in \textit{Kepler} data. Above the typical $\ell =0$ linewidth the magnetic effect is considered as being too large for our perturbative analysis to be valid (see Bugnet et al., \textsl{in prep}). The region in between the two black lines approximatively shows the ideal conditions for the detection of the magnetic effect in \textit{Kepler} data. This region would be narrowed to the area between the TESS 1 year white line and the $\ell$=0 linewidth black line if considering TESS 1 year data, and we do not expect to be able to detect any magnetic effect on mixed-mode frequencies when looking at only 1 month TESS data.
\begin{figure}[t]
\sidecaption[t]
\includegraphics[scale=.72]{Bugnet_Fig3BW.png}
\caption{\textsl{Left:} Mode splitting induced by magnetism only for a M$_\star$ = 1.5 M$_\odot$ giant as a function of the frequency of maximum power $\nu_\textnormal{max}$ and the magnetic field strength. The mode splitting is calculated at each evolutionary stage by taking the splitting corresponding to the central mixed mode (the closest to $\nu_\textnormal{max}$). \textsl{Right:} Same diagram as the left panel simplified to compare with typical frequencies: the "\textit{Kepler} resolution" line corresponds to the 7.9 nHz resolution for the 4 years \textit{Kepler} data, the "TESS 1 year" line corresponds to the 32 nHz resolution for 1 year of TESS data, the "TESS 1 month" line corresponds to the 380 nHz resolution for 1 month of TESS data, and the "$\ell=0$" line corresponds to the typical radial mode linewidths \citep{Vrard2017} for sub-giants, as being an upper limit for mixed-mode linewidth.}
\label{fig:3}
\end{figure}
\section{Conclusion \& Perspectives}
We study the axisymmetric case for which a buried stable fossil magnetic field, with poloidal and toroidal components, is aligned with the rotation axis of the star. We find that a typical field strength of about 1MG results in a detectable frequency shift of the mixed modes inside RG's power spectrum density in addition to the already measured rotational splittings. The study will be extended towards fossil dipolar fields inclined with respect to the rotational axis of the star, and towards non-fossil field topologies such as those derived in \cite{Spruit1999, Jouve2014, Fuller2019}, in order to explore the variety of possible seismic signatures that would allow us to probe the deep magnetism of evolved low-mass stars and its consequences for the transport of angular momentum.\\
\emph{Acknowledgements:} The authors of this work acknowledge the support received from the PLATO CNES grant, the National Aeronautics and Space Administration under Grant NNX15AF13G, by the National Science Foundation grant AST-1411685, the Ramon y Cajal fellowship number RYC-2015-17697, the ERC SPIRE grant (647383), and the Fundation L'Oréal-Unesco-Académie des sciences.
\bibliographystyle{unsrt_7}
| 1d78a747df9964e471437a574f9b25563a9de1b2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Primary Hardware Platform and Experimental Settings}
The primary hardware platform on which the majority of experiments have been conducted is a D-Wave 2000Q quantum annealer at Los Alamos National Laboratory, referred to as \texttt{DW\_2000Q\_LANL}. The \texttt{DW\_2000Q\_LANL} QPU chip has a so-called chimera graph structure with $C_{16}$ topology, i.e. it is composed of two dimensional lattice of 16-by-16 unit cells. Each unit cell is composed of 8 qubits connected through a complete bipartite graph structure. A very small number of faulty qubits are disabled and not available for programming. In total, this QPU has 2032 operational qubits and 5924 operational couplers. A complete topology of the \texttt{DW\_2000Q\_LANL} hardware graph is depicted in Fig.~1a of the Main Text.
Most of experiments in this paper deal with a single unit cell with 8 qubits and 16 couplers; the specific identifiers of this cell is given in Table \ref{tbl:l2k-qpu-used}. We denote the set of qubits as $V$ with $\vert V \vert = N$, and the set of couplers as $E$. This specific set of qubits was selected as it is characteristic of a typical complete unit-cell in a hardware chip.
Unless specified otherwise, in this work we set the following additional solver parameters when submitting jobs to the D-Wave hardware: \texttt{auto\_scale = False}, which ensures that the input parameters are not automatically rescaled to utilize the maximal operating range (a feature sometimes used for optimization applications); \texttt{flux\_drift\_compensation = False}, which prevents automatic corrections to input fields based on calibration procedure that is run a few times each hour; \texttt{annealing\_time = 5}, which corresponds to a single-run annealing time of 5$\mu$s; and \texttt{num\_reads = 10000}, which specifies the number of samples collected for a single programming of the chip. The impact of the specific choice of the annealing time in the regime of parameters considered in this work is negligible, as discussed in Section \ref{sec:impact_of_annealing_time}. The motivations for disabling the flux drift compensation and the impact of spin reversal transformation are thoroughly discussed in Sections \ref{sec:effect_of_spin_reversals} and \ref{sec:local_field_variability}.
\begin{table}[!b]
\centering
\begin{tabular}{|p{0.16\columnwidth}|p{0.84\columnwidth}|}
\hline
\multicolumn{2}{|c|}{Unit Cell Considered} \\
\hline
qubits & $V$ = \{304, 305, 306, 307, 308, 309, 310, 311\} \\
\hline
couplers & $E$ =
\{(304, 308), (304, 309), (304, 310), (304, 311),
(305, 308), (305, 309), (305, 310), (305, 311),
(306, 308), (306, 309), (306, 310), (306, 311),
(307, 308), (307, 309), (307, 310), (307, 311)\} \\
\hline
\hline
\multicolumn{2}{|c|}{Unit Cell Subset for High Throughput Experiments} \\
\hline
qubits & $V'$ = \{304, 305, 308, 309\} \\
\hline
couplers & $E'$ =
\{(304, 308), (304, 309), (305, 308), (305, 309)\} \\
\hline
\end{tabular}
\caption{The 8 qubits and 16 couplers used on the \texttt{DW\_2000Q\_LANL} QPU to conduct the primary experiments of this work. A subset of 4 qubits and 4 couplers is specified for use in high-throughput experiments.}
\label{tbl:l2k-qpu-used}
\end{table}
\section{Single-qubit Experiments}
\label{sec:single_spin_exp}
As demonstrated by the results in this paper, in the regime of couplings that are interesting for sampling applications, the output statistics are perfectly described by a certain classical Boltzmann distribution. Here, we investigate at which coupling strength this classical description breaks down, and one needs to introduce a different statistics, such as quantum Boltzmann distribution, for an adequate description of the output data. We show that this regime corresponds to the couplings strengths that are order of magnitude above the intensities that we consider throughout our study.
The experiments consist in looking, for isolated spins, at the relationship between the input magnetic field $h^{\rm{in}}$ and the outcome statistic described by the effective field $h^{\rm{out}}$. The outcome statistic of a single spin being always fully expressible by a probability distribution taking the following form,
\begin{align}
\mu_{\rm{effective}}(\sigma) = \frac{\exp(h^{\rm{out}}\sigma)}{2\cosh{(h^{\rm{out}})}} = \frac{1+\sigma \tanh(h^{\rm{out}})}{2}.\label{eq:classical_single_spin_distribution}
\end{align}
Our process to estimate $h^{\rm{out}}$ for a given value of the input magnetic field consists in the following steps. We start by collecting $M$ samples from the D-wave annealer embodied as a list of single spin realizations $\sigma^{(k)}\in \{-1,1\}$ for $k=1,\ldots,M$. The statistic $S$ that we extract from these samples is the count of positive spin realization $S=\sum_{k=1}^M \delta_{\sigma^{(k)},1}$, where $\delta$ is the Kronecker delta. Assuming that each sample is effectively independent and identically distributed from Eq.~\eqref{eq:classical_single_spin_distribution}, we observe that the statistic $S$ is a Bernoulli process with $M$ trials and probability of success $p = \frac{1+ \tanh(h^{\rm{out}})}{2}$. We estimate the probability of success $p$ using the standard unbiased estimator $\widehat{p} = S/M$ for Bernoulli processes. We compute confidence intervals $I_\alpha = \left[ \underline{p}, \overline{p}\right]$ with confidence level $\alpha$ around our estimator $\widehat{p}$ using the exact method of Crow \cite{Crow1956Confidence} leading to minimal length intervals. Finally, we invert the relationship between $\widehat{p}$ and $h^{\rm{out}}$ to find an estimate of the output effective field. Confidence intervals on $h^{\rm{out}}$ are found using the same relation since it is a monotonic mapping. In the experiments, we have collected $M=5 \times 10^6$ samples using the D-wave spin reversal transform for values of $h^{\rm{in}}$ between $-1$ and $1$. We have chosen the confidence to be $\alpha = 0.997$ corresponding to a ``3$\sigma$" confidence level.
The classical Boltzmann distribution for a single spin at thermal equilibrium and exposed to a magnetic field $h^{\rm{in}}$ is given by the celebrated formula,
\begin{align}
\mu_{\rm{classical}}(\sigma) = \frac{\exp\left(\beta (h^{\rm{in}}+h^{\rm{res}})\sigma\right)}{2\cosh{\left(\beta (h^{\rm{in}}+h^{\rm{res}})\right)}},
\label{eq:classical_single_spin_gibbs}
\end{align}
where $h^{\rm{res}}$ is a residual magnetic field independent of $h^{\rm{in}}$ and $\beta$ is the inverse temperature times the Boltzmann constant $\beta = 1/k_{B}T$. By comparing Eq.~\eqref{eq:classical_single_spin_distribution} with Eq.~\eqref{eq:classical_single_spin_gibbs}, we find that the relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ predicted by a classical Boltzmann distribution is linear
\begin{align}
h^{\rm{out}}_{\rm{classical}} := \beta (h^{\rm{in}}+h^{\rm{res}}).
\label{eq:classical_h_classic}
\end{align}
The classical parameters $\beta$ and $h^{\rm{res}}$ are estimated using the maximum log-likelihood approach for which the log-likelihood function takes the following form,
\begin{align}
\mathcal{L} = \sum_{h^{\rm{in}}} \tanh(h^{\rm{out}}_{\rm{measured}})
h^{\rm{out}}_{\rm{model}} + \frac{1}{2}\ln\left(1-\tanh\left(h^{\rm{out}}_{\rm{model}})\right)^2\right),
\label{eq:classical_likelihood_c}
\end{align}
where $h^{\rm{out}}_{\rm{measured}}$ are measurements of $h^{\rm{out}}$ for different values of $h^{\rm{in}}$ and $h^{\rm{out}}_{\rm{model}}$ describes the modeled relationship between the input and output fields, e.g. $h^{\rm{out}}_{\rm{classical}}$ from Eq.~\eqref{eq:classical_h_classic}.
Measurements of $h^{\rm{out}}$ for different values of $h^{\rm{in}}$ using the aforementioned procedure on spin \#$309$ and the classical relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ are depicted in Fig.~\ref{fig:classical_raw_spin}.
\begin{figure}
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_classic.pdf}
\caption{Effective output magnetic fields $h^{\rm{out}}$ measured for different input magnetic fields $h^{\rm{in}}$ for the spin \#$309$. Each point is estimated using $M=5\times10^6$ samples with a confidence level of $\alpha=0.997$. Notice that the confidence intervals for large value of $h^{\rm{out}}$ are significantly bigger than for small value of $h^{\rm{out}}$ as the probability distribution scales exponentially with the magnitude of $h^{\rm{out}}$. The classical relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ found with the maximum log-likelihood approach is displayed in brown. The effective inverse temperature for this model is $\beta=10.5$ and the residual field is $h^{\rm{res}}=0.004$.
}
\label{fig:classical_raw_spin}
\end{figure}
We clearly see that if the classical relationship holds for small values of $h^{\rm{in}}$ it fails to explain the flat tails of the measurements for which $|h^{\rm{in}}| \gtrapprox 0.5$. In fact, we starts to already see a deviation from the linear curve for magnitudes of the input field between $0.5 \gtrapprox |h^{\rm{in}}| \gtrapprox 0.2$. In the forthcoming analysis, we will show that the first behavior can be explained by quantum effects, and the latter by classical noise in the input field.
A quantum statistical description of a one spin system is realized through the so-called density matrix formalism. We consider a quantum spin exposed on the $z$-axis to a magnetic field $h^{\rm{in}}$ and residual field $h^{\rm{res}}$ and exposed on the $x$-axis to a transverse field $h^{\rm{trans}}$. This is a natural assumption as the Hamiltonian realized during the considered quantum annealing process is composed of these two terms \cite{johnson2011quantum}. The density matrix describing this one spin system at thermal equilibrium is the following object,
\begin{align}
\rho &= \frac{\exp\left( \beta( h^{\rm{in}} + h^{\rm{res}}) \widehat{\sigma}_z + \beta h^{\rm{trans}} \widehat{\sigma}_x\right)}{\Tr{\exp\left( \beta( h^{\rm{in}} + h^{\rm{res}}) \widehat{\sigma}_z + \beta h^{\rm{trans}} \widehat{\sigma}_x\right)}} \nonumber\\
&= \frac{1}{2}\left(I + \frac{\tanh\left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(h^{\rm{trans}})^2}\right)}{\sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(h^{\rm{trans}})^2}}\left((h^{\rm{in}}+ h^{\rm{res}})\widehat{\sigma}_z + h^{\rm{trans}}\widehat{\sigma}_x\right)\right),
\label{eq:classical_density_matrix}
\end{align}
where $\widehat{\sigma}_z$ and $\widehat{\sigma}_x$ are the usual Pauli matrices for the $z$ and $x$ axis respectively. The transverse field $h^{\rm{trans}}$ that appears in Eq.~\eqref{eq:classical_density_matrix} is in general a function of $h^{\rm{in}}$. Due to an observation that the experimental points seem to flatten out for large values of $h^{\rm{in}}$ in Fig.~\ref{fig:classical_raw_spin}, we choose to parametrize the transverse fields as linear transformations of the input field, namely $h^{\rm{trans}} = \xi h^{\rm{in}}$, a dependence which is consistent with a saturation in the $h^{\rm{out}}$ response.
The mean value of observing the quantum spin along the $z$-axis is given by $\Tr{\rho \widehat{\sigma}_z}$. From this relationship we deduce that the probability of observing the system taking the value $\sigma \in \{-1,1\}$ is given by the following probability distribution,
\begin{align}
\mu_{\rm{quantum}}(\sigma) = \frac{1}{2}\left(1 + \sigma (h^{\rm{in}}+ h^{\rm{res}})\frac{\tanh\left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}\right)}{\sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}}\right).
\label{eq:classical_single_spin_quantum}
\end{align}
With these assumptions we find that that the relationship between $h^{\rm{out}}$ and $h^{\rm{in}}$ in the quantum case can be described as follows,
\begin{align}
h^{\rm{out}}_{\rm{quantum}} := \arctanh \left((h^{\rm{in}}+ h^{\rm{res}})\frac{\tanh\left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}\right)}{\sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}}\right).
\label{eq:classical_h_quantum}
\end{align}
Notice that in the limit where $\xi=0$, the quantum predictions from Eq.~\eqref{eq:classical_h_quantum} converges to its classical counterpart from Eq.~\eqref{eq:classical_h_classic}.
As we will see later in Section~\ref{sec:local_field_variability}, the variability in the residual field $h^{\rm{res}}$ is non-negligible and in fact plays a key role in the explanation of the spurious link behaviors, see Section~\ref{sec:noise_and_spurious_links} for a detailed explanation.
Assuming that $h^{\rm{res}}$ is a random variable with probability density function $f(h^{\rm{res}})$, we can obtain a noisy quantum description of the measurement outcomes by applying Bayes's rule to Eq.~\ref{eq:classical_h_quantum},
\begin{align}
\mu_{\rm{qnoise}} (\sigma) = \int dh^{\rm{res}} f(h^{\rm{res}}) \mu_{\rm{quantum}}(\sigma).
\label{eq:classical_qnoise_measure}
\end{align}
We will see later that in the regime of noise relevant to our experiments, the precise knowledge about the shape of the distribution $f(h^{\rm{res}})$ plays little role. The important quantities are the mean $\mathbb{E}_f[h^{\rm{res}}] = h^{\rm{res}}_0$ and the standard deviation $\sqrt{\mathbb{E}_f[\left( h^{\rm{res}} - h^{\rm{res}}_0\right)]} = h^{\rm{res}}_{\rm{sd}}$. For simplicity, we choose $f$ to be a binomial distribution $f(h^{\rm{res}}) = \frac{1}{2}\delta(h^{\rm{res}} - h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}}) + \frac{1}{2}\delta(h^{\rm{res}} - h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}})$, where $\delta$ denotes the Dirac distribution. Combining Eq.~\eqref{eq:classical_single_spin_quantum} with Eq.~\eqref{eq:classical_qnoise_measure}, we can write down the following relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ in the noisy quantum case,
\begin{align}
h^{\rm{out}}_{\rm{qnoise}} := \arctanh &\left( (h^{\rm{in}} + h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}}) \frac{\tanh \left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}\right)}{2\sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}} \right.\nonumber\\
& \left. + (h^{\rm{in}} + h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}}) \frac{\tanh \left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}\right)}{2\sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}}\right).
\label{eq:classical_h_quantum_noise}
\end{align}
Note that for $h^{\rm{res}}_0 = 0$, the first order expansion of Eq.~\eqref{eq:classical_h_quantum_noise} for small values of $h^{\rm{in}}$ yields the simple relationship $h^{\rm{out}}_{\rm{qnoise}} \approx \beta h^{\rm{in}} \cosh(\beta h^{\rm{res}}_{\rm{sd}})^{-2}$. This shows that one effect of noise on the system consists in reducing the effective inverse temperature that one could infer from a classical (i.e. linear) regression for small values of $h^{\rm{in}}$.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_noise_quantum.pdf}
\caption{Effective output magnetic fields $h^{\rm{out}}$ measured for different input magnetic fields $h^{\rm{in}}$ for the spin \#$309$. Each point is estimated using $M=5\times10^6$ samples with a confidence level of $\alpha=0.997$. The noiseless quantum and the noisy quantum relationships between $h^{\rm{in}}$ and $h^{\rm{out}}$ found with the maximum log-likelihood approach are displayed in red and blue respectively. The effective inverse temperature for the noisy quantum model is higher than the effective inverse temperature found with the noiseless quantum model. The noiseless quantum model and the classical models find similar temperatures. }
\label{fig:classical_noise_quantum}
\end{figure}
We infer the parameters of the proposed quantum models, i.e. $h^{\rm{res}}_0$, $\xi $ and $h^{\rm{res}}_{\rm{sd}}$, with the maximum log-likelihood approach described by Eq.~\eqref{eq:classical_likelihood_c}, replacing the quantity $h^{\rm{out}}_{\rm{model}}$ by either the functional relationship $h^{\rm{out}}_{\rm{quantum}}$ from Eq.~\eqref{eq:classical_h_quantum} or $h^{\rm{out}}_{\rm{qnoise}}$ from Eq.~\eqref{eq:classical_h_quantum_noise}.
In Fig.~\ref{fig:classical_noise_quantum}, we show the measurements acquired for spin \#$309$ along with the curves predicted by the noiseless and noisy quantum models. The flattening of the output field response for $|h^{\rm{in}}|\gtrapprox 0.5$ is correctly accounted for by the addition of the quantum transverse field. The initial deviation from the linear response for values of the input field between $0.5 \gtrapprox|h^{\rm{in}}|\gtrapprox 0.2$ are predicted accurately only by the noisy quantum model. An important feature highlighted by this study is the effect of the noise on the effective inverse temperature (or linear response) that one may infer from the data. We see from our models that the noise in the residual field effectively lowers the inverse temperature for small values of the input field of order $|h^{\rm{in}}|\lessapprox 0.2$. Indeed, the effects due to noise will be more prominent when the input field is small. In this particular illustration, the noise on the residual field accounts for more than $20\%$ of the input field for magnitudes of $|h^{\rm{in}}|\leq 0.2$.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_noiseless_quantum.pdf}
\caption{Effective output magnetic fields $h^{\rm{out}}$ measured for different input magnetic fields $h^{\rm{in}}$ for the spin \#$307$. Each point is estimated using $M=5\times10^6$ samples with a confidence level of $\alpha=0.997$. The noiseless quantum and the noisy quantum relationships between $h^{\rm{in}}$ and $h^{\rm{out}}$ found with the maximum log-likelihood approach are displayed in red and blue respectively. This particular spin displays a low-level of residual field noise. As a consequence, the noiseless and noisy quantum models find the same inverse temperature.}
\label{fig:classical_noiseless_quantum}
\end{figure}
In Fig.~\ref{fig:classical_noiseless_quantum}, we show similar measurements collected for spin \#$307$ as long as the predictions of the proposed quantum models. Among all spins that we tested, we found that spin \#$307$ has the lowest level of residual field noise. In this particular case, we see that both noiseless and noisy quantum model predictions agree and we find an inverse temperatures consistent with the inverse temperature found for spin \#$309$ with the noisy quantum model.
For the unit cell of interest to this work, we show the regression coefficients obtained with the classical model, the noiseless quantum model and the noisy quantum model in Table~\ref{tbl:classical_reg_coeff}.
\begin{table*}
\centering
\begin{tabular}{|c||c|c|c|}
\hline
Spin & Classical & Quantum & Noisy + Quantum \\
\# & $(\beta, h^{\rm{res}}_0)$ & $(\beta, h^{\rm{res}}_0, \xi)$ & $(\beta, h^{\rm{res}}_0, \xi, h^{\rm{res}}_{\rm{sd}})$ \\
\hline
\hline
304 & (11.3, 0.014) & (11.4, 0.014, 0.017) & (12.3, 0.014, 0.018, 0.029) \\
\hline
305 & (11.1, 0.003) & (11.2, 0.003, 0.016) & (13.1, 0.003, 0.017, 0.041) \\
\hline
306 & (11.4, -0.004) & (11.5, -0.004, 0.016) & (13.3, -0.004, 0.018, 0.039) \\
\hline
307 & (12.1, -0.009) & (12.2, -0.009, 0.016) & (12.4, -0.009, 0.016, 0.012) \\
\hline
308 & (11.6, -0.005) & (11.7, -0.005, 0.012) & (12.9, -0.005, 0.013, 0.032) \\
\hline
309 & (10.5, 0.004) & (10.5, 0.004, 0.011) & (12.7, 0.004, 0.013, 0.048) \\
\hline
310 & (11.2, -0.006) & (11.3, -0.006, 0.012) & (12.5, -0.006, 0.013, 0.035) \\
\hline
311 & (11.4, 0.010) & (11.5, 0.010, 0.012) & (12.5, 0.010, 0.013, 0.031) \\
\hline
\end{tabular}
\caption{List of regression coefficients for eight different spins found with the classical model, the noiseless quantum model and the noisy quantum model. The coefficients are obtained through the maximization of the log-likelihood function respective to each model. The inverse temperature found by the classical and noiseless quantum models differs from the inverse temperature obtained with the noisy quantum model for significant noise magnitudes. The other parameters are consistent between models which shows that these models form a hierarchy of increasing descriptive complexity. The transverse field responses for spins \#304 to \#307 differs significantly from the transverse field responses of spins \#308 to \#311. This suggest a difference in the hardware implementation of these two groups of spins for they are physically located on the two different sides of a chimera cell.}
\label{tbl:classical_reg_coeff}
\end{table*}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_noise_comp.pdf}
\caption{Comparison of the effect of different noise probability distributions for noisy quantum models on spin \#$309$. Predictions obtained from noisy quantum models derived from a binary distribution and a uniform distribution are displayed in blue and purple respectively. The parameters for both models are identical and corresponds to the parameters inferred for the binary noise model. The two models are indistinguishable for this level of noise magnitude, indicating that the main contributing factors in the probabilistic description of the noise are the first and the second moments.}
\label{fig:classical_noise_comparison}
\end{figure}
Finally, we study the effect of the noise probability distribution on the model. We compare our noisy quantum model derived from a binary probability distribution with the noisy quantum model that one obtains with a uniform noise distribution, i.e. $f(h^{\rm{res}}) = (2\sqrt{3} h^{\rm{res}}_{\rm{sd}})^{-1}$ for $h^{\rm{res}} \in \left[h^{\rm{res}}_0 - \sqrt{3} h^{\rm{res}}_{\rm{sd}}, h^{\rm{res}}_0 + \sqrt{3} h^{\rm{res}}_{\rm{sd}}\right]$ and zero otherwise. Predictions for the spin \#$309$ made by these two models of noise are displayed in Fig.~\ref{fig:classical_noise_comparison}. The parameters for both models are identical and correspond to the parameters inferred with maximum log-likelihood for the binary probability distribution. We see that for noise magnitudes of $h^{\rm{res}}_{\rm{sd}} = 0.048$ the predictions are practically indistinguishable.
\section{Learning of General Classical Distributions on Binary Variables}\label{sec:learning_general_distributions}
An arbitrary positive probability distribution on $N$ classical binary variables $\underline{\sigma} \in \{ -1, +1 \}^N$ can be represented in the the form of a Gibbs distribution with different interaction orders:
\begin{align}
\mu(\underline{\sigma}) & = \frac{1}{Z} \exp{ \left( \underbrace{\sum_{i} h_i \sigma_i}_{\text{1st order}} + \underbrace{\sum_{ij} J_{ij} \sigma_i \sigma_j}_{\text{2d order}} + \underbrace{\sum_{ijk} J_{ijk} \sigma_i \sigma_j \sigma_k}_{\text{3d order}} + \underbrace{\sum_{ijkl} J_{ijkl} \sigma_i \sigma_j \sigma_k \sigma_l}_{\text{4th order}} + \cdots \right)},
\label{eq:Gibbs_distribution}
\end{align}
where $Z$ denotes the normalization factor known as parition function. In general, the number of terms can be exponential. However, physical systems are typically characterized by a finite number of short-ranged multi-body interactions. For instance, the renowned Ising model corresponds to the case where only first and second orders are present. In this section, we develop an exact learning method that allows one to reconstruct an arbitrary probability distribution on binary variables, i.e. recover the parameters $\{h_i, J_{ij}, J_{ijk}, J_{ijkl}, \ldots \}$ from a number of independent spin configurations sampled from the distribution \eqref{eq:Gibbs_distribution}. For the case of Ising models with pairwise interactions only
\begin{align}
\mu(\underline{\sigma}) = \frac{1}{Z} \exp{ \left( \sum_{i} h_i \sigma_i + \sum_{ij} J_{ij} \sigma_i \sigma_j \right)},
\label{eq:Ising_distribution}
\end{align}
this reconstruction problem is known as \emph{inverse Ising problem}, and has been extensively studied in the past with a number of heuristic methods \cite{nguyen2017inverse}. However, the inverse Ising problem has been solved only recently, with the appearance of exact algorithms that showed that parameters of an arbitrary Ising model, including low-temperature and spin-glass models, can be recovered to an arbitrary precision with an appropriate number of samples \cite{bresler2015efficiently, vuffray2016interaction}. The state-of-the-art near-optimal performance for inverse Ising problem has been recently achieved with the estimator based on the Interaction Screening Objective (ISO) \cite{lokhov2018optimal}. For Ising model, the ISO reads
\begin{equation}
S_i(\mathbf{J_i},h_i) = \left\langle \exp{ \left( - h_i \sigma_i - \sum_{j} J_{ij} \sigma_i \sigma_j \right)} \right\rangle_M,
\label{eq:ISO_Ising}
\end{equation}
where $\langle f(\sigma) \rangle_M = M^{-1} \sum_{m=1}^M f(\sigma^{(m)})$ denotes the empirical average over $M$ independent samples, and $\mathbf{J_i}$ denotes the set of couplings adjacent to node $i$, i.e. $\mathbf{J_i} = \{ J_{ij} \}_{j \neq i}$ for Ising models. It is easy to see that the ISO in \eqref{eq:ISO_Ising} is a convex function of parameters $\mathbf{J_i},h_i$. Furthermore, the unique minimizer of the ISO
\begin{equation}
(\widehat{\mathbf{J}_i},\widehat{h_i}) = \operatornamewithlimits{argmin} S_i(\mathbf{J_i},h_i)
\label{eq:ISO_estimation}
\end{equation}
converges to the true parameters of the distribution in the limit of a large number of samples, and yields an $O(1/\sqrt{M})$ error on the recovered model parameters for finite $M$ \cite{vuffray2016interaction}. The ISO is a local estimator, i.e. it is defined for each spin, and only involves couplings adjacent to this spin, so for reconstructing the entire model one needs to run $N$ parallel reconstruction problems \eqref{eq:ISO_estimation}. Notice that this procedure yields two estimations for the same coupling, $\widehat{J_{ij}}$ and $\widehat{J_{ji}}$, and we use their arithmetic mean $(\widehat{J_{ij}} + \widehat{J_{ji}})/2$ as the final estimate of the coupling $J_{ij}$. This estimator is used throughout the paper for reconstructing parameters of Ising models, where the required number of samples for a given precision and the associated expected confidence interval are obtained through synthetic numerical experiments, as explained below.
Here, we generalize this objective function to the case of general Gibbs distributions with multi-body interactions of the type \eqref{eq:Gibbs_distribution}. The corresponding ISO reads
\begin{equation}
S_i(\mathbf{J_i},h_i) = \left\langle \exp{ \left( - h_i \sigma_i - \sum_{j} J_{ij} \sigma_i \sigma_j - \sum_{jk} J_{jkl} \sigma_i \sigma_j \sigma_k - \sum_{jkl} J_{jkl} \sigma_i \sigma_j \sigma_k \sigma_l - \cdots \right)} \right\rangle_M.
\label{eq:ISO_General}
\end{equation}
Similarly to the Interaction Screening estimator for the inverse Ising problem \eqref{eq:ISO_Ising}, it is easy to see that the ISO for general models is a convex function of parameters $\mathbf{J_i},h_i$. Let us present a simple argument that illustrates the fact that in the limit of large number of samples the unique minimizer \eqref{eq:ISO_estimation} of the convex ISO objective \eqref{eq:ISO_General} is achieved at $(\widehat{\mathbf{J}_i},\widehat{h_i}) = (\mathbf{J}_i, h_i)$, meaning that the true interactions present in the model are fully ``screened''. Indeed, the ISO is an empirical average of the inverse of the factors in the Gibbs measure; if
\begin{equation}
\mathcal{F}_i(\mathbf{J}_i, h_i) = \exp( h_i \sigma_i + \sum_{j} J_{ij} \sigma_i \sigma_j + \sum_{jk} J_{jkl} \sigma_i \sigma_j \sigma_k + \sum_{jkl} J_{jkl} \sigma_i \sigma_j \sigma_k \sigma_l + \cdots),
\end{equation}
then $S_i(\mathbf{J}_i,h_i) = \langle \mathcal{F}_i^{-1}(\mathbf{J}_i,h_i) \rangle_{M}$. In the limit of large number of samples $S_i(\mathbf{J}_i,h_i) \rightarrow S_i^*(\mathbf{J}_i,h_i) = \langle 1/\mathcal{F}_i(\mathbf{J}_i,h_i) \rangle$, where $\langle \cdots \rangle$ denotes the average over the measure \eqref{eq:Gibbs_distribution}. Let us look at the derivative of the ISO with respect to a given coupling, say $J_{ijk}$. This derivative corresponds to weighted three-body correlation, for instance $\partial S_i^{*} / \partial J_{ijk} = \langle\sigma_i \sigma_j \sigma_k / \mathcal{F}_i(\mathbf{J}_i,h_i)\rangle$, and this reflect the key property of the Interaction Screening based estimator. When $(\widehat{\mathbf{J}_i},\widehat{h_i}) = (\mathbf{J}_i, h_i)$, $\partial S_i^{*} / \partial J_{ij} \vert_{\mathbf{J}_i,h_i} = 0$, meaning that the minimum of ISO is achieved at $(\widehat{\mathbf{J}_i},\widehat{h_i}) = (\mathbf{J}_i, h_i)$ as $M \to \infty$.
Again, similarly to the case of Ising models, after running $N$ parallel local reconstructions, the resulting couplings $\mathbf{J_i} = \{ J_{ij}, J_{ijk}, J_{ijkl}, \ldots \}$ are symmetrized using all permutations of tuples $(i, j, k, l, \ldots)$. This estimator is used for probing general distributions with multi-body interactions in the next section.
It is important to highlight that the estimator given by the generalized ISO \eqref{eq:ISO_General} is exact: With a given number of samples $M$, the deviation between reconstructed $(\widehat{\mathbf{J}_i},\widehat{h_i})$ and true $(\mathbf{J}_i, h_i)$ model parameters decay as $\sim 1/\sqrt{M}$. Next, we describe an empirical procedure that we have developed to estimate the reconstruction fidelity in practice, which dictate how much data we needed in all experiments to statistically exclude the finite-sample considerations from all conclusions that we draw throughout the work.
\section{Empirical Estimation of Reconstruction Errors}
\label{sec:empirical_error_estimation}
For finite number of samples $M$, couplings estimated via \eqref{eq:ISO_estimation} are accurate up to a certain error that decays with $M$. This error can be estimated theoretically, however the resulting worst-case bounds can be loose for a given model. To the best of our knowledge, there is no standard approach to tightly quantify the finite sampling error of learning models of the form \eqref{eq:Gibbs_distribution} in practice. To address this challenge, we propose the following empirical error estimation procedure for a given model with specific parameters and a fixed number of samples. Specifically, given a black-box sample generator $B$, a finite sample set $M$, and a replicate parameter $R$, we conduct the following procedure:
\begin{enumerate}
\item collect $M$ samples of the black-box $B$;
\item reconstruct model $m$ from the collected samples;
\item for $r$ from $1$ to $R$ independent replicates do the following:
\begin{enumerate}
\item collect $M$ synthetic samples from $m$ using an auxiliary sampling algorithm;
\item reconstruct model $m_r$ from the collected samples;
\end{enumerate}
\item compute statistics over the parameters reconstructed across the reconstructed $\{m_r\}_{r \in 1,\ldots,R}$ models.
\end{enumerate}
In this work the black-box sampler $B$ is given by our \texttt{DW\_2000Q\_LANL} QPU, and the sampling algorithm is a brute-force approach that enumerates every possible state and computes the exact probably of each state. The this brute-force approach is feasible for small number of spins which will be the focus of our targeted experiments. For larger problems one could utilize more scalable sampling algorithms, including those based on Markov-Chain Monte-Carlo techniques, or Belief Propagation with decimation.
The last step will provide the information on the typical variability of reconstruction accuracy due the effect of finite samples. Given a set of $R$ replicates of the models of the type \eqref{eq:Gibbs_distribution} reconstructed from $R$ independent sets of samples. Let us define $(\mathbf{h}^r, \mathbf{J}^r) \;\; \forall r \in R$ as parameters of the models of the type \eqref{eq:Gibbs_distribution} learned from each of the $R$ sets of samples. We also define $(\delta \mathbf{h}, \delta \mathbf{J})$ representing deviations from the parameters $(\mathbf{h}, \mathbf{J})$ reference model that was used to produce synthetic samples:
\begin{align}
\delta \mathbf{h}^r = \mathbf{h} - \mathbf{h}^r \;\; \forall r \in R \\
\delta \mathbf{J}^r = \mathbf{J} - \mathbf{J}^r \;\; \forall r \in R
\end{align}
In the step 4 of the procedure above, we estimate the empirical mean and standard deviation on each of the parameters in the set $h_i, \mathbf{J}_i$ from $R$ values. For a sufficiently large $R$, these quantities indicate an expected scale of error coming from model recovery for a given number of samples.
\section{Probing of Multi-Body Interactions}
\label{sec:multi-body}
The objective of experiment in this section is to leverage the ISO for general distributions on binary variables \eqref{eq:ISO_General} to determine the class of models that adequately describes the output distribution produced by the \texttt{DW\_2000Q\_LANL} QPU. Although the target Hamiltonian has a form of a classical Ising model, \emph{a priori} this distribution could be more general than the Ising Gibbs distribution \eqref{eq:Ising_distribution}. In this section, we probe the existence of multi-body interactions beyond pairwise in the energy function of the output distribution. We conduct an experiment on a single chimera graph cell involving 8 qubits (see Table \ref{tbl:l2k-qpu-used}). As explained in the previous section, for an 8-spin model, the most general positive probability distribution can be written in the form \eqref{eq:Gibbs_distribution} with an energy function being a polynomial of order eight. Specifically, for $\underline{\sigma} \in \{-1, +1\}^8$, the possible distribution reads
\begin{align}
\mu(\underline{\sigma}) = \frac{1}{Z} \exp{ \left( \sum_{i} h_i \sigma_i + \sum_{ij} J_{ij} \sigma_i \sigma_j + \sum_{ijk} J_{ijk} \sigma_i \sigma_j \sigma_k + \cdots + J_{ijklmnop} \sigma_i \sigma_j \sigma_k \sigma_l \sigma_m \sigma_n \sigma_o \sigma_p \right) }.
\label{eq:ho_boltzmann}
\end{align}
This model has a total of 8 $h$ parameters 247 $J$ parameters.
Our goal will be to practically show existence or absence of multi-body interactions in the output distribution. Presence of interactions can be established if the reconstructed couplings are statistically significant, i.e. they are larger in absolute value than the reconstruction error resulting from a finite-sample reconstruction. We show that a few million D-Wave samples will be sufficient to provide an accurate reconstruction of model parameters.
\begin{table}
\centering
\begin{tabular}{|r||r|r|r|r|r|}
\hline
Model & Qubits & Couplers & $h^{\text{in}}$ & $J^{\text{in}}$ & Samples \\
\hline
\hline
ferromagnet & $V$ & $E$ & 0.0 & 0.025 & 10,000,000 \\
\hline
anti-ferromagnet & $V$ & $E$ & 0.0 & -0.025 & 10,000,000 \\
\hline
\end{tabular}
\caption{The input parameters sent to the \texttt{DW\_2000Q\_LANL} QPU to collect data for the 8-th order distribution reconstruction experiment.}
\label{tbl:8o-models}
\end{table}
The experiment we conduct here focuses on learning 8-th order models from the samples output by the D-Wave hardware on two canonical models, a ferromagnet and an anti-ferromagnet. The parameter details of these two models are presented in Table \ref{tbl:8o-models}. The primary objective of this experiment is to determine what model parameters are statistically significant. Specifically, what learned parameters can-and-cannot be attributed to artifacts from to finite sampling and the model reconstruction algorithm. Through a procedure that was detailed in section \ref{sec:empirical_error_estimation}, a number of supporting simulations are conducted to determine error bounds on the reconstructed model parameters. Leveraging the obtained error value, we determine the three standard deviations threshold that determine the statistical significance of the reconstructed values. Recovered couplings with absolute values above this threshold are very unlikely to be due to a reconstruction error, while values below can be explained by the finite sample noise in the reconstruction process.
Figs.~\ref{fig:fm-8o-parameter-hist} and \ref{fig:afm-8o-parameter-hist} present the absolute values of the 255 model parameters broken down by the interaction-order for the ferromagnetic and the anti-ferromagnetic cases, respectively. We find that in both cases, a second-order model provides an accurate representation of the output distribution of \texttt{DW\_2000Q\_LANL}. This experiment thus provides a convincing evidence that a second-order model is sufficient for modeling the distribution that the quantum annealer samples from for a range of input parameters of interest to this work.
\begin{figure}
\centering
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=7]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=8]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=9]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=10]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=11]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=12]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=13]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\caption{Histograms of reconstruction parameters magnitudes from the 8-th order reconstruction experiment on the Ferromagnet model. Values above the $3\sigma$ line are considered to be statically significant, while the values below are artifacts of finite sampling. These results indicate that the distribution that the \texttt{DW\_2000Q\_LANL} hardware is sampling from is well approximated by a second-order model.}
\label{fig:fm-8o-parameter-hist}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=7]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=8]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=9]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=10]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=11]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=12]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=13]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\caption{Histograms of reconstruction parameters magnitudes from the 8-th order reconstruction experiment on the Anti-Ferromagnet model. Values above the $3\sigma$ line are considered to be statically significant, while the values below can be artifacts of finite sampling. These results indicate that the distribution that the \texttt{DW\_2000Q\_LANL} hardware is sampling from is well approximated by a second-order model.}
\label{fig:afm-8o-parameter-hist}
\end{figure}
We further validate these results by quantifying the typucal variation of model parameters used in the 8-th order reconstruction based on finite sampling error. To that end, the validation protocol described in section \ref{sec:empirical_error_estimation} is executed with with $R=50$ replicates and $M=10^7$, to replicate the number of samples used in the data collection for most of the key experiments in this work. Fig.~ \ref{fig:8o-parameter-variance} shows data for the mean and the standard deviation for each parameter deviation $\delta h$ or $\delta J$ from the reference model, estimated from running the variance measuring procedure defined in the section \ref{sec:empirical_error_estimation} on the reconstruction models. We see that estimated variance of reconstructed values is very similar across different couplings. The average-case variance across all model parameters using three standard deviations are 0.0034 and 0.0021 for the Ferromagnet and Anti-Ferromagnet models, respectively. These values have been used for computing the threshold values that appear in Figs.~ \ref{fig:fm-8o-parameter-hist} and \ref{fig:afm-8o-parameter-hist} in the previous section.
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_resample_delta_hist.pdf}
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_resample_delta_hist.pdf}
\caption{Multi-body reconstruction variability results for the Ferromagnet model (left) and Anti-Ferromagnet (right) used in the multi-body probing experiment, using the validation protocol from section \ref{sec:empirical_error_estimation} with $R=50$ replicates. The mean values of the all 255 parameters are presented with whiskers shown at three standard deviations from the mean. The average-case deviations are 0.0034 and 0.0021, respectively.}
\label{fig:8o-parameter-variance}
\end{figure}
\section{Reconstruction and Validation of Two-Body Models}
\label{sec:two-body}
The previous section conducted a 8-th order reconstruction with 10,000,000 samples and argued that a 2-nd order model provides a sufficient approximation of the distribution that the \texttt{DW\_2000Q\_LANL} hardware samples from. To that end, the remaining experiments in this work focuses on learning 2-nd order models of the hardware's output distribution.
Leveraging the knowledge that this 2-nd order model is sufficient has a significant advantages in that the number of model parameters reduces from 255 (8-th order) to only 36 (2-nd order), which in turn reduces the amount of data required to accurately learn the associated 2-nd order model. To further lessen the data requirements, we reduce the number of qubits considered from 8 to 4, focusing on the upper-half of the cell. This effectively decreases the number of 2-nd order parameters from 36 to 16. It will later become evident that these reductions are necessary to make the experiments viable on reasonable time scales.
After conducting these model reductions, we replicate here the validation experiments from the section \ref{sec:multi-body} to establish an optimal number of samples required for reconstruction of parameters in the two-body model. Table \ref{tbl:2o-models} specifies the input parameters for two additional models, {\em strong ferromagnet} and {\em strong anti-ferromagnet}, using coupling sign convention consistent with \eqref{eq:Gibbs_distribution}.
\begin{table}
\centering
\begin{tabular}{|r||r|r|r|r|r|}
\hline
Model & Qubits & Couplers & $h^{in}$ & $J^{in}$ & Samples \\
\hline
\hline
strong ferromagnet & $V'$ & $E'$ & 0.0 & 0.05 & $4\cdot10^6$ \\
\hline
strong anti-ferromagnet & $V'$ & $E'$ & 0.0 & -0.05 & $4\cdot10^6$ \\
\hline
\end{tabular}
\caption{The input parameters sent to the \texttt{DW\_2000Q\_LANL} QPU to collect data for the 2-nd order distribution reconstruction experiment.}
\label{tbl:2o-models}
\end{table}
Fig.~\ref{fig:2o-strong-parameter-variance} presents the results of the validation experiment for the 2-nd order model reconstruction. 50 reconstruction replicates are used in the validation experiments conducted in this section. The results indicate that reconstruction accuracy is approximately 0.0025 and 0.0022, which is comparable to the accuracy used in the 8-th order reconstruction experiment, using a smaller number of samples.
Fig.~\ref{fig:2o-strong-parameter-values} presents the strength of the second order terms that are recovered from the hardware data. These absolute values are well above the recovery accuracy threshold, indicating that $4 \cdot 10^6$ samples are sufficient for accurately recovering the two-body model. We use this number of samples in the remainder of experiments in this paper.
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/fms_1x1_l2k/00500_00000_base_reconst_resample_delta_hist.pdf}
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/afms_1x1_l2k/00500_00000_base_reconst_resample_delta_hist.pdf}
\caption{Two-body reconstruction variability results for the Strong Ferromagnet model (left) and Strong Anti-Ferromagnet (right), using the proposed validation protocol with 50 replicates. The mean values of the all 16 parameters are presented with whiskers shown at three standard deviations from the mean. The average-case deviations are 0.0025 and 0.0022, respectively.}
\label{fig:2o-strong-parameter-variance}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=6]{results/multibody/fms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\includegraphics[width=0.41\linewidth,page=6]{results/multibody/afms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\caption{Second-Order terms of the two-body reconstruction of the for the Strong Ferromagnet model (left) and Strong Anti-Ferromagnet (right). The recovered values are greater than 0.03, well above the accuracy threshold of 0.0025, suggesting an accurate model.}
\label{fig:2o-strong-parameter-values}
\end{figure}
In Fig.~2 of the Main Text, we summarize the results of these experiments and show the structure of the output distribution. In particular, we find that among statistically significant couplings in the output distribution, the strongest ones are in correspondence with the input couplings, while the weakest ones are the \emph{spurious couplings} that are not present in the input problem and, moreover, are absent in the chip topology. In what follows, we construct additional experiments aimed at clarifying the nature of these spurious couplings.
\section{Impact of the Annealing Time}
\label{sec:impact_of_annealing_time}
In all presented experiments thus far, we used the annealing time of 5$\mu s$ per each sample. It is important to understand how the statistics of these reconstruction experiments might differ as the annealing time varies. Here, we replicate the model reconstruction experiment for the Strong Ferromagnet and Strong Anti-Ferromagnet models, using the following varying annealing time parameters, \texttt{annealing\_time = 1, 5, 25, 125, 625}, which corresponds to single-run annealing time of 1$\mu s$, 5$\mu s$, 25$\mu s$, 125$\mu s$, 625$\mu s$ respectively. Additionally, in this experiment, the \texttt{num\_reads} parameter was reduced from 10,000 (this work's default) to 4,000 in all cases, to adhere to the maximum job run-time limit of the \texttt{DW\_2000Q\_LANL} QPU.
\begin{figure*}
\centering
\includegraphics[width=0.88\linewidth,page=1]{results/multibody/fms_1x1_l2k/00500_00000_base_annealing_time_series.pdf}
\includegraphics[width=0.88\linewidth,page=1]{results/multibody/afms_1x1_l2k/00500_00000_base_annealing_time_series.pdf}
\caption{A comparison of the second-order model reconstructions for different annealing times on the Strong Ferromagnet model (top) and Strong Anti-Ferromagnet model (bottom). The results indicate that large increases in the annealing time results in slight reductions in the effective temperature of the reconstructed models.}
\label{fig:2o-strong-annealing-time}
\end{figure*}
Fig.~\ref{fig:2o-strong-annealing-time} presents the results of this experiment. We find that an increase of the annealing time by two orders of magnitudes results only in a slight increase of the absolute coupling values in the reconstructed model. At the same time, this minor change comes at a significant increase in data collection time. Due to this observation, in this work we chose to standardize around the annealing time of 5$\mu s$, which is essential for the high-throughput data collection, and represents the fastest available annealing time across several generations of quantum annealers.
\section{Quadratic Response}
\label{sec:quadratic_response}
In this section, we provide additional details on the study of the quadratic response experiment described in the Main Text. We assume the most general parametrization of the quadratic input-output relationship between the input parameters and the parameters of the reconstructed output distribution. Specifically, for each output model parameter we learn the quadratic $\chi^{hh}, \chi^{hJ}$, linear $\beta^h, \beta^J$ and offset $c$ coefficients of the following quadratic function,
\begin{align}
h_i &= [h^{in} \;\; J^{in}] \begin{bmatrix} \chi^{hh}_i & \chi^{hJ}_i \\ \chi^{Jh}_i & \chi^{JJ}_i \end{bmatrix} \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + [\beta^h_i \;\; \beta^J_i] \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + c_i \;\; \forall i \in V' \label{eq:h_quad_model}\\
J_{ij} &= [h^{in} \;\; J^{in}] \begin{bmatrix} \chi^{hh}_i & \chi^{hJ}_{ij} \\ \chi^{Jh}_{ij} & \chi^{JJ}_{ij} \end{bmatrix} \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + [\beta^h_{ij} \;\; \beta^J_{ij}] \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + c_{ij} \;\; \forall i \in V', j \in V' \label{eq:j_quad_model}
\end{align}
Given sufficient data, these quadratic functions can be recovered using a least squares regression. For example, the function form of the output parameter $h_i$ can be learned from a collection of $S$ different input-output realizations by solving the following convex optimization problem,
\begin{align}
\operatornamewithlimits{argmin}_{{\chi^{hh}_i}, {\chi^{hJ}_i}, {\chi^{Jh}_i}, {\chi^{JJ}_i} {\beta^h_i}, {\beta^J_i}, c} \; & \sum_{s \in S} \left( [h^{in}_s \;\; J^{in}_s] \begin{bmatrix} \chi^{hh}_i & \chi^{hJ}_i \\ \chi^{Jh}_i & \chi^{JJ}_i \end{bmatrix} \begin{bmatrix} h^{in}_s \\ J^{in}_s \end{bmatrix} + [\beta^h_i \;\; \beta^J_i] \begin{bmatrix} h^{in}_s \\ J^{in}_s \end{bmatrix} + c_i - h_{is} \right)^2 \label{eq:quad-opt}\\
& {\chi^{hh}_i}, {\chi^{hJ}_i}, {\chi^{Jh}_i}, {\chi^{JJ}_i} {\beta^h_i}, {\beta^J_i}, c \in \mathbb{R} \nonumber
\end{align}
Note that similar optimization problems can be solved to learn the quadratic relation of all output parameters $h,J$. The number of data points (i.e. $|S|$) should be at least $n \ln n$ to accurately recover the quadratic function, where $n$ is the number of unknown values in \eqref{eq:quad-opt}. The core experiment of this section consists in performing a series of 2-nd order reconstruction experiments over random input models, and then in using these pairs of input-output models to recover the coefficients in the quadratic response function.
The primary challenge of this experiment is the time required to collect a sufficient amount of data to fit the quadratic response function. To minimize the data requirements, we focus on the 4-spin model defined as $N',E'$ in Table \ref{tbl:l2k-qpu-used}. For this specific model, the quadratic functions \eqref{eq:h_quad_model}, \eqref{eq:j_quad_model} have 57 parameters; we consider 250 input-output model pairs to recover these parameters. Each of the 250 input models is selected i.i.d. from the following input parameter distribution,
\begin{align}
h^{in}_i & \in \{-0.05:0.01:0.05\} \;\; \forall i \in N' \label{eq:quadresp_field_pertubation_set},\\
J^{in}_{ij} & \in \{-0.05:0.01:0.05\} \;\; \forall i,j \in E'.\label{eq:quadresp_coupling_pertubation_set}
\end{align}
Following the validation study for the {\em strong ferromagnet} and {\em strong anti-ferromagnet} models, $4 \cdot 10^6$ samples are sufficient to accurately reconstruct 2-nd order models with parameters as large as 0.05. Altogether, this experiment reconstructs the outputs for 250 input models using 4,000,000 samples for each model, which results in a total of a \emph{billion} samples collected.
Characteristic dominant terms in the recovered quadratic response function is presented in the Fig.~4 of the Main Text (data on all measured quadratic response terms is given below, in Sections~\ref{sec:noise_and_spurious_links} and \ref{sec:three_generations}.). The zero order terms in the response functions are interpreted as residual fields and couplings; the first-order terms are related to the native couplings present in the chip; and finally, the second-order terms are responsible for the spurious couplings. In previous work, the primary hypothesis behind the response function was formulated in terms of a particular case of a linear assumption \cite{bian2010ising, benedetti2016estimation, perdomo2016determination, raymond2016global, marshall2017thermalization, li2020limitations}, where each parameter would be multiplied by a single effective temperature. The non-linearity of the general response function that we construct here may explain why this effective temperature was found to be instance-dependent: This corresponds to a linear approximation of a non-linear function.
Spurious couplings identified under a careful statistical analysis indicate that a simple linear model is not sufficient for an accurate characterization of the D-Wave's input-output relationship. It is important to note that the second-order response that we find here is different from the previously observed next-nearest-neighbour couplings in the strong input regime, where a quadratic cross-talk relation with an \emph{opposite sign} susceptibility has been suggested (see the section ``Compensation of qubit nonidealities'' in the \emph{Methods} of \cite{king2018observation}). We conjecture that the emergence of the next-nearest-neighbour couplings observed in the strong regime has a quantum nature and is due to the induced effects of the transverse field; a detailed exploration of this phenomenon is beyond the scope of our study that focuses on the classical regime of the output distribution in a multi-qubit setting.
The discovery of strong and structured quadratic response functions for the output distribution that our D-Wave \texttt{DW\_2000Q\_LANL} QPU samples is invaluable to applications such as hardware calibration, problem embedding and accurate sampling. However, identifying the root cause of these unexpected output parameters can provide valuable information about how to design better quantum annealers and can provide novel analytics for evaluating the performance of quantum annealers. To that end, the next section provides the theoretical grounds to explain the quadratic response as side effects of instantaneous qubit noise.
\section{Characterisation of the Local Field Variably}
\label{sec:local_field_variability}
In preparation for constructing a model that would explain the form of the quadratic response function, in this section we investigate possible drifts of the reconstructed model parameters. To this end, we perform reconstructions over several days, and monitor the stability of the recovered model.
The foundation of this variability study is the reconstruction of the output distribution of the zero value problem, that is $h^{in}, J^{in} = 0$ as shown in Table \ref{tbl:2o-zero-model}. Furthermore, we would like to perform this reconstruction accurately but with a minimum number of samples, so that the possible flux drift dynamics can be observed in the time between multiple reconstructions. We begin by calibrating a 2-body reconstruction specifically for the zero-value problem by proposing that only 200,000 samples are required for an accurate reconstruction, which is 20 times less data than what is used for a typical 2-body reconstruction. Repeating the previous reconstruction variability analysis, Fig.~\ref{fig:2o-zero-parameters} presents both the variance and recovery accuracy of this zero-valued model. The reconstruction accuracy is approximately 0.007, which is about two times less accurate than the previously considered reconstruction experiments. However, we find this accuracy is still acceptable as our primary interest in this experiment are model values that are above 0.100. Indeed, Fig.~\ref{fig:2o-zero-parameters} indicates that the \texttt{DW\_2000Q\_LANL} QPU has a number of biases that are on the order of magnitude larger than reconstruction accuracy threshold of 0.008.
\begin{table}
\centering
\begin{tabular}{|r||r|r|r|r|r|}
\hline
Model & Qubits & Couplers & $h^{in}$ & $J^{in}$ & Samples \\
\hline
\hline
Zero-value & $V'$ & $E'$ & 0.0 & 0.000 & 200,000 \\
\hline
\end{tabular}
\caption{The input parameters sent to the \texttt{DW\_2000Q\_LANL} QPU to collect data for the 2-nd order reconstruction experiment.}
\label{tbl:2o-zero-model}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/zeros_1x1_l2k/00000_00000_base_reconst_resample_delta_hist.pdf}
\includegraphics[width=0.41\linewidth,page=4]{results/multibody/zeros_1x1_l2k/00000_00000_base_reconst_hist.pdf}
\caption{The reconstruction variability (left) and recovered interaction magnitudes (right) of the zero problem. The results suggest indicate that 200,000 samples is sufficient to accurately measure the qubit bias occurring in the \texttt{DW\_2000Q\_LANL} QPU.}
\label{fig:2o-zero-parameters}
\end{figure}
Next, we investigate how the reconstructed values of the zero problem change over time. The objective is to understand how the low frequency noise in the hardware changes over time and how that can impact data collection over the span of minutes to hours. In this experiment, the data for the proposed zero problem is collected at 10 minute intervals over a period of 48 hours and then the 2-body reconstruction is used to recovery a model from the observed samples. Fig.~\ref{fig:zero-overtime} shows the reconstructed values over time and whiskers around the points indicate the error bounds on the model reconstruction values. Considering the mean values of these time series one can observe that there is a persistent bias on both the reconstructed fields and couplers, which is most likely an artifact of the initial hardware's calibration. Looking at the variance of the time series highlights a high variance in the fields terms and a much lower variance in the coupler terms. Overall, the results of these experiments suggest that all parameters of the output distribution at the exception of local fields remain stable over time.
\begin{figure}
\centering
\includegraphics[width=0.66\linewidth]{results/recalibration/zero_long_l2k_timeseries.pdf}
\caption{Reconstruction of the zero problem repeated over several days. A clear persistent bias in the local fields is evident from offset mean values. The fluctuations in the reconstructed field are significantly larger than the reconstruction error, suggesting fluctuation in the local fields over time.}
\label{fig:zero-overtime}
\end{figure}
\section{Explaining Quadratic Response via Instantaneous Qubit Noise}
\label{sec:noise_and_spurious_links}
In the previous section~\ref{sec:local_field_variability}, we saw that magnetic field is affected by comparably large fluctuations. However, this analysis was conducted by performing the reconstruction over a certain time window, which may average out the fluctuations. On the other hand, the analysis in section~\ref{sec:single_spin_exp} allowed us to estimate \emph{instanteneous} fluctuations of the residual random magnetic field for each qubit individually. Here, we show that these rapid fluctuations in the individual qubit fields can be responsible for spurious effective interactions with non-trivial quadratic-type responses in the input quantities. We start by showing theoretically the type of spurious interactions one may expect to reconstruct on toy models for which we can derive close form formulas. Then, using numerical simulations, we will quantify the quadratic response caused by noise on a four qubit system and compare these predictions with the quadratic response measured on the D-Wave quantum annealer.
\subsection{Spurious Magnetic Field Response}\label{subsec:theory_spurious_fields}
We first consider a simple system consisting of two classical spins $\sigma_1,\sigma_2 \in \{-1,1\}$ linked by a coupling $J\in \mathbb{R}$. The first spin is subject to a noisy magnetic field $h^{\rm{sd}}_{1} s_1$ whose direction varies according to the uniform random variable $s_1\in \{-1,1\}$, whereas the second spin is subject to a constant magnetic field $h_2$. The Bolzmann distribution of this two spins system at inverse temperature $\beta$ for a particular noise realization $s_1$ is given by the following expression,
\begin{equation}
\mu(\sigma_1,\sigma_2 \mid s_1) = \frac{\exp{\left( \beta (J \sigma_1 \sigma_2 + h^{\rm{sd}}_{1} s_1\sigma_1 +h_2 \sigma_2 )\right)}}{Z(s_1)}, \label{eq:quadth_one_coupling_dist}
\end{equation}
where the partition function $Z(s_1)$ depends on the noise realization $s_1$. Suppose now that we want to perform an Ising model reconstruction using a collection of iid. samples from the distribution in Eq.~\eqref{eq:quadth_one_coupling_dist} where the noise realization changes randomly on a sample to sample basis. Our collection of samples end up arising from a mixture of models as around half of the configurations comes from $\mu(\sigma_1,\sigma_2 \mid s_1 = +1)$ and the other half comes from $\mu(\sigma_1,\sigma_2 \mid s_1 = -1)$. Therefore, the effective model that we can reconstruct with this heterogeneous collection of samples is the following mixture of Ising models,
\begin{equation}
\mu_{\rm{effective}}(\sigma_1,\sigma_2) = \sum_{s_1\in \{-1,1\}} \frac{1}{2} \mu(\sigma_1,\sigma_2 \mid s_1).
\end{equation}
This effective model is also an Ising model and, after some algebra, it can be explicitly formulated with respect to the initial coupling and fields,
\begin{equation}
\mu_{\rm{effective}}(\sigma_1,\sigma_2) = \frac{\exp{\left( \beta (J \sigma_1 \sigma_2 + h^{\rm{effective}}_{1} \sigma_1 +h_2 \sigma_2 )\right)}}{Z_{\rm{effective}}}, \label{eq:quadth_one_coupling_mixture}
\end{equation}
where
\begin{align}
h^{\rm{effective}}_{1}
= -\frac{1}{\beta}\arctanh{\left(\tanh{(\beta J)} \tanh{(\beta h_2)} \tanh{(\beta h^{\rm{sd}}_{1})}^2\right)}. \label{eq:quadth_one_coupling_response}
\end{align}
For small coupling and field magnitudes $J$ and $h_2$, the expression in Eq.~\eqref{eq:quadth_one_coupling_response} reduces to $h^{\rm{effective}}_{1} \approx - \beta J h_2 \tanh{(\beta h^{\rm{sd}}_{1})}^2$. We see with this toy model that fast fluctuating magnetic field noise induces an effective magnetic response. The samples coming from a mixture of models with field noise are indistinguishable from samples coming from the single model with an effective response. The key qualitative features of this response is 1) its intensity is roughly proportional to the product of the coupling and opposite (constant) field intensity, and 2) the sign of the response is negative. This effect being roughly proportional to the square of the standard deviation of the noise, it is negligible for low noise values but becomes much more pronounced when the noise becomes large.
\subsection{Spurious Coupling Response}\label{subsec:theory_spurious_couplings}
To illustrate how spurious couplings can occur from a mixture of noisy Ising models, we look at a chain of three spins $\sigma_1$, $\sigma_2$ and $\sigma_3$ connected via couplings $J_{12}$ and $J_{23}$. We assume that the spin at the extremity of the chain are subject to a noisy magnetic field $h^{\rm{sd}}_{1} s_1$ and $h^{\rm{sd}}_{3} s_3$ where the random variable $s_1, s_3 \in \{-1,1\}$ are independent and uniformly distributed. The magnetic field on the middle spin is assumed to be zero. For a particular noise realization, the Boltzmann distribution of this chain of spins is given by the following conditional distribution,
\begin{align}
\mu(\sigma_1,\sigma_2, \sigma_3 \mid s_1, s_3)
= \frac{\exp{\left( \beta (J_{12} \sigma_1 \sigma_2 + J_{23} \sigma_2 \sigma_3 + h^{\rm{sd}}_{1} s_1\sigma_1 + h^{\rm{sd}}_{3} s_3\sigma_3\right)}}{Z(s_1,s_3)}, \label{eq:quadth_two_coupling_dist}
\end{align}
where the partition function $Z(s_1,s_3)$ is a function of the noise realization. Similarly to the previous subsection, we consider fast fluctuating noise that changes randomly on a sample to sample basis. In this case, a collection of samples coming from the mixture of noisy Ising models described by Eq.~\eqref{eq:quadth_two_coupling_dist} becomes indistinguishable from iid. samples coming from the effective model,
\begin{equation}
\mu_{\rm{effective}}(\sigma_1,\sigma_2,\sigma_3) = \sum_{s_1,s_3\in \{-1,1\}} \frac{1}{4} \mu(\sigma_1,\sigma_2,\sigma_3 \mid s_1,s_3).
\end{equation}
This effective model appears to be an Ising model as well with no magnetic fields and with an additional coupling between $\sigma_1$ and $\sigma_3$,
\begin{align}
\mu_{\rm{effective}}(\sigma_1,\sigma_2, \sigma_3)
= \frac{\exp{\left( \beta (J_{12} \sigma_1 \sigma_2 + J_{23} \sigma_2 \sigma_3 + J^{\rm{effective}}_{13} \sigma_1 \sigma_3) \right)}}{Z_{\rm{effective}}}, \label{eq:quadth_two_coupling_mixture}
\end{align}
where the effective coupling $J^{\rm{effective}}_{13}$ can be explicitly written with respect to the mixture parameters,
\begin{equation}
J^{\rm{effective}}_{13} = -\frac{1}{\beta}\arctanh{\left(\tanh{(\beta J_{12})} \tanh{(\beta J_{23})} \tanh{(\beta h^{\rm{sd}}_{1})}^2 \tanh{(\beta h^{\rm{sd}}_{3})}^2\right)}. \label{eq:quadth_two_coupling_response}
\end{equation}
For small coupling values, Eq.~\eqref{eq:quadth_two_coupling_response} reduces to $J^{\rm{effective}}_{13} \approx -\beta J_{12} J_{23} \tanh{(\beta h^{\rm{sd}}_{1})}^2 \tanh{(\beta h^{\rm{sd}}_{3})}^2$. We immediately see that this coupling response induced by field noise retains the main qualitative features observed in the previous subsection. The intensity of the response is quadratic in the couplings $J_{13}$ and $J_{23}$ for it is proportional to their product and the sign of the response is negative. Note that this coupling response, which involves three spins, is predicted to be weaker than the magnetic field response discussed in the previous subsection as it is roughly proportional to the square of both noise standard deviations $h^{\rm{sd}}_{1}$ and $h^{\rm{sd}}_{3}$.
\subsection{Simulations and Predictions Using Single Spin Measurements}
In the previous subsections, we have seen on simple toy models that field noise leads to effective field and spurious coupling responses. For small coupling and field magnitudes, these ``spurious" responses were mainly quadratic in the input parameters. We now want to quantify the quadratic responses cause by field noise on a realistic four spin system and compare them to the type of quadratic responses found experimentally in Section~\ref{sec:quadratic_response}. This four spin system being already too complex to obtain a closed form formula, we resort to using numerical simulations to extract the quadratic response coefficients. We model the system by a classical Boltzmann distribution conditioned on a noise realization $\underline{s}\in \{-1,1\}^4$ of the field noise parameters. The probability to obtain a configuration $\underline{\sigma}\in\{-1,1\}^4$ given $\underline{s}$ reads as follows,
\begin{equation}
\mu(\underline{\sigma} \mid \underline{s}) = \frac{\exp{\left(H_{\rm{fields}}\left(\underline{\sigma} \mid \underline{s}\right) +H_{\rm{couplings}}\left(\underline{\sigma}\right)\right)}}{Z(\underline{s})},\label{eq:quadth_four_spins}
\end{equation}
where the partition functions $Z(\underline{s})$ is noise dependant. The Hamiltonian describing the magnetic field interaction contains terms $h_i$ for input fields, $h^{\rm{bias}}$ for permanent biases and $h^{\rm{sd}}_{i}$ for the standard deviation of the noise as described in Section~\ref{sec:single_spin_exp}. An individual temperature $\beta_i$ is also assigned for each spin,
\begin{align}
H_{\rm{fields}}\left(\underline{\sigma} \mid \underline{s}\right) &= \sum_{i=1,\ldots,4}\beta_{i}(h^{\rm{sd}}_{i} s_i + h^{\rm{bias}}_{i} + h_{i})\sigma_{i}.
\end{align}
The coupling Hamiltonian contains terms $J_{ij}$ for the input coupling strengths that are only along physical couplers and possessing their individual temperatures $\beta_{ij}$. Motivated by considerations from Section~\ref{sec:local_field_variability}, we assume that the interactions are noiseless and without biases,
\begin{align}
H_{\rm{couplings}}\left(\underline{\sigma}\right)
= \beta_{12} J_{12} \sigma_1 \sigma_2 + \beta_{14}J_{14} \sigma_1 \sigma_4 + \beta_{23}J_{23} \sigma_2 \sigma_3 +\beta_{34}J_{34} \sigma_3 \sigma_4.
\end{align}
The effective model describing the probability distribution of the four spin system is obtained after averaging Eq.~\eqref{eq:quadth_four_spins} over the uniform and independant noise realizations,
\begin{align}
\mu_{\rm{effective}}\left(\underline{\sigma}\right) = \frac{1}{4}\sum_{\underline{s}\in\{-1,1\}^4} \mu(\underline{\sigma} \mid \underline{s}).
\label{eq:quadth_four_effective_model}
\end{align}
The numerical procedure to reconstruct from Eq.~\eqref{eq:quadth_four_effective_model} a quadratic response as a function of the input couplings $J_{ij}$ and fields $h_{i}$ is reminiscent of the experimental protocol described in Sections~\ref{sec:quadratic_response}.
We start by randomly selecting 20000 input coupling and field configurations whose values lies in the set $\{-0.05, -0.04, \cdots, 0.05\}$, see Eq.~\eqref{eq:quadresp_field_pertubation_set} and Eq.~\eqref{eq:quadresp_coupling_pertubation_set}. Then for each of these configurations, we compute numerically the effective frequencies of the $2^4=16$ spin configurations using Eq.~\eqref{eq:quadth_four_effective_model} and summing over the $2^4=16$ possible noise realizations. These frequencies are used in our reconstruction procedure, described in Section~\ref{sec:learning_general_distributions}, to infer an effective Ising model with Hamiltonian,
\begin{align}
H_{\rm{effective}}(\underline{\sigma}) = \hspace{-0.39cm} \sum_{i,j\in \{1,2,3,4\}}J^{\rm{effective}}_{ij} \sigma_i \sigma_j + \hspace{-0.39cm} \sum_{i\in\{1,2,3,4\}} h^{\rm{effective}}_i \sigma_i.\label{eq:quadth_effective_hamiltonian}
\end{align}
Note that the Hamiltonian in Eq.~\eqref{eq:quadth_effective_hamiltonian} contains all pairwise interactions between four spins and the spurious interactions are represented by the effective couplings $J^{\rm{effective}}_{13}$ and ${J^{\rm{effective}}_{24}}$. Finally, we fit a quadratic response model between the inputs configurations and their corresponding inferred effective couplings, as described by Eq.~\eqref{eq:h_quad_model} and Eq.~\eqref{eq:j_quad_model}, following the optimization procedure in Eq.~\eqref{eq:quad-opt}.
The spins $\sigma_1, \sigma_2, \sigma_3, \sigma_4$ in our model are identified with the hardware spins $\#304, \#308, \#305, \#309$ respectively. The values of field temperatures $\beta_i$, field biases $h^{\rm{bias}}_i$ and field noise standard deviations $h^{\rm{sd}}_i$ are chosen to be those measured using the single spin quantum experiments described in Section~\ref{sec:single_spin_exp}. These values can be found in the last column of Table~\ref{tbl:classical_reg_coeff}. The values of the coupling temperatures have been adjusted such that the simulated and measured effective temperatures coincide, i.e. $\beta_{12}=12.1, \beta_{14}=12.2, \beta_{23}=12.5$ and $\beta_{24}=12.6$.
\begin{figure*}
\centering
\includegraphics{figures/Supp_mat_full_response_simulated_vs_true.pdf}
\caption{General quadratic response of effective output parameters. A comparison of the simulated and true responses at the quadratic level are given for the effective parameters $J^{\rm{out}}_{304,305}$, $J^{\rm{out}}_{308,309}$, and $h^{\rm{out}}_{309}$. The dominant terms in the general quadratic response presented here are discussed in the Fig.~4 of the Main Text. Notice that here and below, we use a symmetric matrix representation for the quadratic response, which results in a factor 2 difference for matrix elements compared to the results presented in the Fig.~4 of the Main Text.}
\label{fig:full_response_simulated_vs_true}
\end{figure*}
The typical simulated and measured response for existing couplings, fields, and spurious couplings are depicted in
the Fig.~\ref{fig:full_response_simulated_vs_true}. We see that the patterns are qualitatively following the predictions from the simple theoretical models: The response for existing couplings is a linear self-transform or effective temperature, the response for fields consists in an effective temperature and a negative quadratic response from adjacent couplings and connecting neighboring fields, and finally the spurious couplings are formed by a negative quadratic response from adjacent couplings that span a triangle with the spurious coupling. The main noticeable difference with the theoretical predictions is the lowering of the coupling effective temperatures from its model temperature due to the presence of field noise. The comparison between the measured and simulated effective temperature can be found in Table~\ref{tbl:quadth_lin_response}. The leading off-diagonal coefficients of the quadratic response, both simulated and measured, are displayed in Table~\ref{tbl:quadth_quad_response}. The effective temperatures found through the simulation remarkably matches the effective temperatures measured in the hardware with a maximum difference of at most $7\%$. The negative sign of the quadratic response and the type of interactions involved is in perfect agreement with the theoretical model. There exists a discrepancy in the magnitudes predicted by the simulation and those found experimentally. The predictions are up to two times weaker for the susceptibility of the fields and up to four times weaker for the spurious couplings. This can be explained by the strength of the noise induced response as our theoretical models predicts that the field response is a second order effect in the noise parameters, see Subsection~\ref{subsec:theory_spurious_fields}, and the spurious links response is a fourth order effect in the field noise intensity, see Subsection~\ref{subsec:theory_spurious_couplings}. Therefore, only a $40\%$ difference in the single spin noise standard deviation can explain such differences. Note that the the quadratic response of the spurious coupling $J_{304,305}$ being weaker that the spurious coupling $J_{308,309}$ is correctly predicted by the simulation and with a similar ratio.
\begin{table}
\centering
\begin{tabular}{|c||c|c|}
\hline
\multicolumn{3}{|c|}{Linear Self-Response}\\
\hline
Output/Input & Simulated & Measured\\
\hline
\hline
$h_{304}$ & $11.1$ & $10.5$ \\
\hline
$h_{305}$ & $10.4$& $10.1$ \\
\hline
$h_{308}$ & $11.2$& $10.5$ \\
\hline
$h_{309}$ & $9.4$& $9.4$ \\
\hline
$J_{304,308}$ & $12.0$ & $12.0$ \\
\hline
$J_{304,309}$ & $11.9$ & $11.9$ \\
\hline
$J_{305,308}$ & $12.2$ & $12.2$ \\
\hline
$J_{305,309}$ & $12.0$ & $12.0$ \\
\hline
\end{tabular}
\caption{Effective simulated and measured temperatures for existing couplings and magnetic fields. The individual coupling temperatures have been adjusted in the model such that the effective simulated temperatures coincide with the effective measure temperatures.}
\label{tbl:quadth_lin_response}
\end{table}
\begin{table}
\centering
\begin{tabular}{ |c||c|c|c| }
\hline
\multicolumn{4}{|c|}{Main Quadratic Response}\\
\hline
Output& Input & Simulated & Measured \\
\hline
\hline
$h_{304}$ & $h_{308}, J_{304,308}$ & -4.9 & -10.5\\
& $h_{309}, J_{304,309}$ & -3.8 & -7.4 \\
\hline
$h_{305}$ & $h_{308}, J_{305,308}$ & -10.7 & -9.2\\
& $h_{309}, J_{305,309}$ & -8.2 & -11.6\\
\hline
$h_{308}$ & $h_{304}, J_{304,308}$ & -6.5 & -12.6\\
& $h_{305}, J_{305,308}$ & -5.9 & -11.1\\
\hline
$h_{309}$ & $h_{304}, J_{304,309}$ & -13.0 & -11.0\\
& $h_{305}, J_{305,309}$ & -11.7 & -11.9\\
\hline
$J_{304,305}$ & $J_{304,308}, J_{305,308}$ & -0.9 & -4.1\\
& $J_{304,309}, J_{305,309}$ & -0.7 & -4.0\\
\hline
$J_{308,309}$ & $J_{304,308}, J_{304,309}$ & -1.5 & -5.7\\
& $J_{305,308}, J_{305,309}$ & -1.3 & -6.8\\
\hline
\end{tabular}
\caption{Main simulated and measured components of the quadratic response. The leading terms are off-diagonal and correspond to an adjacent couplings and connected field for magnetic fields and correspond to two adjacent couplings forming a triangle for spurious links.}
\label{tbl:quadth_quad_response}
\end{table}
\section{Impacts of Spin Reversal Transformations}
\label{sec:effect_of_spin_reversals}
\subsection{Theoretical Considerations}
We are looking at distributions of Ising models on $N$ spins $\underline{\sigma} = \{\sigma_1,\ldots\,\sigma_{N}\}$ depending on input couplings $\underline{J}$ and magnetic fields $\underline{h}$. If we consider that the fields are potentially noisy and there exists individual residual fields $b_i$, random or deterministic, the Boltzmann distribution takes the following form,
\begin{align}
\mu(\underline{\sigma}\mid \underline{J}, \underline{h}) = \frac{\exp{\left(\sum_{ij} \beta_{ij} J_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i (h_i + b_i) \sigma_i\right)}}{Z},\label{eq:srt_ising_model}
\end{align}
where $\beta_{ij}$ and $\beta_i$ are effective individual temperatures for couplings and fields respectively. The residual fields may have potentially strong undesirable effects such as favoring particular spin configurations among others that were initially designed to be equiprobable. There exists a heuristical method that aims at mitigating this problem called the spin reversal transform (SRT). This method consists at looking at $2^N$ possible remapping of the model, each of them indexed by a ``gauge" which is a binary configuration $\tau = \{-1,1\}^N$. For a given configuration $\tau$, this gauge transform maps a spin configuration, input couplings and input fields to the values $\underline{\sigma}^\tau$, $\underline{h}^\tau$ and $\underline{J}^\tau$ in the following way,
\begin{align}
\sigma^{\tau}_i = \sigma_i \tau_i, \quad h^{\tau}_{i} = h_i \tau_i, \quad J^{\tau}_{ij}= J_{ij} \tau_i \tau_j.\label{eq:srt_gauge_transform}
\end{align}
The particularity of the transformation~\eqref{eq:srt_gauge_transform} is that it creates an equivalence relationship between Hamiltonians \emph{without} residual fields as for any gauge $\tau$,
\begin{align}
\sum_{ij} \beta_{ij} J^{\tau}_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i h^{\tau}_i \sigma_i = \sum_{ij} \beta_{ij} J_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i h_i \sigma_i.
\end{align}
For system without biases, it implies that samples generated from any set of gauge transformed inputs $\underline{h}^\tau$ and $\underline{J}^\tau$ are identical after a remapping of the samples using the same gauge $\underline{\sigma}^\tau$. The SRT method consists in generating samples from a mixtures of randomly selected gauge transformed models with residual fields, and which are therefore no longer equivalent, in order to empirically average over the residual field values. The effective model describing this mixture is given by the average of the Bolztmann distribution \eqref{eq:srt_ising_model} over all possible gauge transforms and reads,
\begin{align}
\mu^{\rm{effective}}(\underline{\sigma}\mid \underline{J}, \underline{h}) = \frac{1}{2^{N}} \sum_{\tau \in \{-1,1\}^N} \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}),
\end{align}
where the gauge transformed models is explicitly expressed as follows,
\begin{align}
\mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau})
= \frac{\exp{\left(\sum_{ij} \beta_{ij} J_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i (h_i + \tau_i b_i) \sigma_i\right)}}{Z^{\tau}}.\label{eq:srt_gauge_model}
\end{align}
Note that the partition function in Eq.~\eqref{eq:srt_gauge_model} depends on the gauge transforms through the residual field values. We see with Eq.~\eqref{eq:srt_gauge_model} that the sign of the residual fields are effectively randomly flipped with the SRT method. Thus, the SRT removes the undesirable effects of permanent residual fields but transforms it into magnetic field noise with other potentially unwanted effects such as lower effective temperature, spurious links and field quadratic response described in Section~\ref{sec:noise_and_spurious_links}. For the D-wave hardware where the fields noise is more important compared to permanent biases, see Section~\ref{sec:single_spin_exp}, the SRT removes permanent biases for only a limited increase in the noise. Therefore, the Ising model reconstructions appears less fluctuating over time with SRT than without.
To illustrate this last point, consider an Ising model with permanent residual fields $b_i$ and no input fields i.e. $\underline{h} \equiv 0$. In this model, the average of value of the spins are non-zero in general and are a non trivial function of the residual fields and couplings. However, in the effective model produced by the SRT $\mu^{\rm{effective}}(\underline{\sigma}\mid \underline{J}, 0)$, the average value of each spin is identically zero. This implies that the effective model has zero effective magnetic field regardless of the value of the residual biases. To see this, we first note that when $\underline{h} = 0$, the partition function of a gauge transformation is invariant under a global sign change $Z^{\tau} = Z^{ - \tau}$. This further implies the following equivalence between two probabilities of gauge transformed spin configurations $\mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau},0) = \mu(\underline{\sigma}^{-\tau}\mid \underline{J}^{-\tau},0)$ for all $\tau$. Since $\underline{\sigma}^{-\tau} = \underline{\sigma}^{\tau}$, it shows that the average value of any spin $\sigma_u$ vanishes as,
\begin{align}
\mathbb{E}_{\rm{SRT}}\left[\sigma_u\right] &= \sum_{\underline{\sigma}\in\{-1,1\}^N} \sigma_u \mu^{\rm{effective}}(\underline{\sigma}\mid \underline{J}, \underline{h}) \nonumber\\
&= \frac{1}{2^{N}} \sum_{\underline{\sigma}\in\{-1,1\}^N} \sigma_u \sum_{\tau \in \{-1,1\}^N} \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}),\nonumber\\
&= \frac{1}{2^{N}} \sum_{\underline{\sigma}\in\{-1,1\}^N} \sigma_u \sum_{\tau \in \{-1,1\}^N} \frac{1}{2}\left(\mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}) + \mu(\underline{\sigma}^{-\tau}\mid \underline{J}^{-\tau}, \underline{h}^{-\tau})\right),\nonumber\\
&= \frac{1}{2^{N}} \sum_{\tau \in \{-1,1\}^N} \sum_{\underline{\sigma}\in\{-1,1\}^N}\frac{1}{2}\left(\sigma_u \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}) + \sigma_u \mu(\underline{\sigma}^{-\tau}\mid \underline{J}^{-\tau}, \underline{h}^{-\tau})\right),\nonumber\\
&= \frac{1}{2^{N}} \sum_{\tau \in \{-1,1\}^N} \sum_{\underline{\sigma}\in\{-1,1\}^N}\frac{1}{2}\left(\sigma_u \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}) - \sigma_u \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau})\right) =0.
\end{align}
\subsection{Mitigating Persistent Bias and Flux Drift}
Section \ref{sec:local_field_variability} highlighted persistent biases that the hardware exhibits.
A useful feature of the spin reversal symmetry group is that combining data collected from symmetric models has the effect of averaging out persistent biases.
\begin{figure*}
\centering
\includegraphics{figures/Supp_mat_biases_SRT.pdf}
\caption{Illustration of the impact of spin reversal transforms on the persistent biases and flux drifts. Reconstruction of the parameters of the zero input problem was repeated over several hours. Reconstructed output parameters are not exactly centered around zero, and instead are fluctuating around values that can be interpreted as persistent biases. Local fields present much larger fluctuations compared to both native couplers and spurious couplings, that are significantly larger than the reconstruction error, a signature of the instantaneous qubit noise averaged over the reconstruction period. The left plot features normal raw data, and the right plot utilizes data obtained with random spin-reversal transformations applied to the input problems, which to a large extent eliminates the impact of persistent bias and reduces the variance of fluctuations.}
\label{fig:biases_SRT}
\end{figure*}
This property is highlighted by replicating the field variability experiment with and without D-Wave's spin reversal transform feature.
In this revised experiment, the data for the zero problem is collected at 1.5 minute intervals over a period of 2 hours and then the 2-body reconstruction is used to recovery a model from the observed samples.
This experiment is conducted with two settings, {\em raw data} using this work's standard setting of \texttt{num\_spin\_reversal\_transforms = 0} and spin reversal transform data using \texttt{num\_spin\_reversal\_transforms = 10} (this setting results in a total of 200 transforms across the 200,000 samples collected for the zero-problem).
Fig.~\ref{fig:biases_SRT} shows the reconstructed values over time.
Table \ref{tbl:slow-field-variation} presents the mean and variance of each of the field values in these time series.
These results highlight how the spin reversal transforms can provide a drastic mitigation of the hardware's persistent bias.
An unexpected result from conducing these spin reversal transforms is a notable reduction in the variance of the field values.
Although the root cause for this reduction is not clear, we hypothesize that it is a side effect of an increased number of QPU programming cycles, which present another source of biases during the data collection process.
In any case, the notable bias mitigating impacts of spin reversal transforms can have a significant positive impact on applications where the user would like an unbiased output distribution.
\begin{table}
\centering
\begin{tabular}{|r||r|r||r|r|}
\hline
& \multicolumn{2}{c||}{Raw Data} & \multicolumn{2}{c|}{Spin Reversals} \\
\hline
Spin & $\mu(h)$ & $\sigma(h)$ & $\mu(h)$ & $\sigma(h)$ \\
\hline
\hline
304 & -0.241 & 0.025 & -0.001 & 0.019 \\
\hline
305 & -0.060 & 0.024 & 0.000 & 0.006 \\
\hline
308 & 0.121 & 0.025 & 0.000 & 0.010 \\
\hline
309 & -0.033 & 0.025 & 0.001 & 0.007 \\
\hline
\end{tabular}
%
\begin{tabular}{|r||r|r||r|r|}
\hline
& \multicolumn{2}{c||}{Raw Data} & \multicolumn{2}{c|}{Spin Reversals} \\
\hline
Coupler & $\mu(J)$ & $\sigma(J)$ & $\mu(J)$ & $\sigma(J)$ \\
\hline
\hline
304-308 & 0.0215 & 0.0024 & -0.0004 & 0.0030 \\
\hline
304-309 & -0.0076 & 0.0028 & -0.0007 & 0.0029 \\
\hline
305-308 & 0.0357 & 0.0027 & 0.0002 & 0.0035 \\
\hline
305-309 & -0.0005 & 0.0026 & 0.0000 & 0.0026 \\
\hline
\end{tabular}
%
\caption{A comparison of the mean and variance of reconstructed field values over time with and without spin reversal transforms during data collection. The use of spin reversal transforms corrects for persistent bias and also reduces the variance of reconstructed values.}
\label{tbl:slow-field-variation}
\end{table}
\subsection{Model Reconstruction with Spin Reversal Transforms}
Given the potential bias mitigating benefits of spin reversal transforms, it is natural to inquire how this feature impacts the results presented thus far.
We begin by reviewing the two-body reconstruction results presented in Fig. ~\ref{fig:2o-strong-parameter-srt-values}, which provides a side-by-side comparison of the results of Strong Ferromagnet model from Table \ref{tbl:2o-models} with and without spin reversal transforms. As the first observation, we notice that there is a notable change in the results of the zero-order terms. In the case with spin reversal transforms the zero-order terms are near zero, while statically significant non-zero values are exhibited in the raw data. The second observation is that the first-order terms do not show a notable change; in fact, these two reconstructions are remarkably consistent with and without spin reversal transforms. Both of these results indicate an absence of detrimental artifacts from utilizing this feature during data collection at the time scales that these experiments require.
\begin{figure}
\centering
\includegraphics[width=0.39\linewidth,page=5]{results/multibody/fms_1x1_l2k/00500_00000_base_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=5]{results/multibody/fms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/fms_1x1_l2k/00500_00000_base_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/fms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\caption{The first-order (top) and second-order (bottom) terms of the two-body reconstruction without (left column) and with (right column) spin reversal transforms for the Strong Ferromagnet model.}
\label{fig:2o-strong-parameter-srt-values}
\end{figure}
\subsection{Quadratic Response with Spin Reversal Transforms}
Figs.~\ref{fig:quad_comparison_part1} and \ref{fig:quad_comparison_part2} replicate the quadratic response experiment from Section \ref{sec:quadratic_response} with and without spin reversal transforms. This comparison shows that the overall consistency of the quadratic response picture. In accordance with the theoretical predictions outlined in the beginning of this section, we observe that the persistent bias essentially disappears under the SRT setting, and the overall response becomes much cleaner.
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_srt_part1.pdf}
\caption{A comparison of the key motifs of the quadratic response function (here, for the chimera couplings) without (left) and with (right) spin reversal transforms. The motifs are largely similar, however one can notice a considerable reduction in apparent noise in the quadratic response of the first-order terms. We hypothesize this is due mitigation of the flux qubit drift that occurs throughout the many hours of data collection required by this analysis.}
\label{fig:quad_comparison_part1}
\end{figure*}
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_srt_part2.pdf}
\caption{A comparison of the key motifs of the quadratic response function (here, for the local fields and spurious couplings) without (left) and with (right) spin reversal transforms. The motifs are largely similar, however one can notice a considerable reduction in apparent noise in the quadratic response of the first-order terms. We hypothesize this is due mitigation of the flux qubit drift that occurs throughout the many hours of data collection required by this analysis.}
\label{fig:quad_comparison_part2}
\end{figure*}
\section{Tests Across Three Generations of Quantum Annealers and Validation of the Instantaneous Noise Model}
\label{sec:three_generations}
The previous sections have argued that the quadratic response model is a good valuable tool for characterizing the input-output behavior of a quantum annealer. Furthermore, the strength of spurious links can provide an indirect measurement of the instanteneous qubit noise that is occurring on a specific hardware device. We had the opportunity to collect data for the response analysis on 7 distinct QPUs spanning three generations of quantum annealing hardware. The specific device names and components used in the experiment are presented in Table \ref{tbl:other-qpus-used}.
\begin{table*}
\centering
\begin{tabular}{|l||l|l|}
\hline
QPU & $V'$ & $E'$ \\
\hline
\hline
\multicolumn{3}{|c|}{2X} \\
\hline
LANL's \texttt{DW2X} & \{880, 881, 884, 885\} & \{(880, 884), (880, 885), (881, 884), (881, 885)\} \\
\hline
\hline
\multicolumn{3}{|c|}{2000Q} \\
\hline
\hline
\texttt{DW\_2000Q\_LANL} & \{304, 305, 308, 309\} & \{(304, 308), (304, 309), (305, 308), (305, 309)\} \\
\hline
NASA's \texttt{C16} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\texttt{DW\_2000Q\_1} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\texttt{DW\_2000Q\_2} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\texttt{DW\_2000Q\_3} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\hline
\multicolumn{3}{|c|}{Lower-Noise 2000Q} \\
\hline
\texttt{DW\_2000Q\_5} & \{1688, 1689, 1692, 1693\} & \{(1688, 1692), (1688, 1693), (1689, 1692), (1689, 1693)\} \\
\hline
\texttt{DW\_2000Q\_6} & \{1688, 1689, 1692, 1693\} & \{(1688, 1692), (1688, 1693), (1689, 1692), (1689, 1693)\} \\
\hline
\end{tabular}
\caption{The 4 qubits and 4 couplers used in the quadratic response experiments on a variety of QPUs. The experiments span three generations of D-Wave hardware, the 2X (2015), 2000Q (2017), and Lower-Noise 2000Q (2019).}
\label{tbl:other-qpus-used}
\end{table*}
Figs.~\ref{fig:quad_response_2x}-\ref{fig:quad_response_2k6} present the quadratic response motifs from all of the QPUs that have been tested. Overall, the results are remarkably similar, which suggests the universality of the quadratic response characterization that is capturing fundamental properties of D-Wave's quantum annealing implementation, like effective temperature, persistent biases, and instantaneous noise in the local fields. With the exception of the new {\em lower-noise} QPU, the existence and strength of the spurious links is consistent across hardware realizations. We have observed that some QPU implementations feature asymmetric spurious links while others are symmetric. Identifying the root-cause of this distinction is an ongoing point of investigation.
A significant difference in the response function has been obtained for the \emph{lower-noise} version of the D-Wave 2000Q annealer \cite{D-Wave_low_noise}. We observe a drastic reduction of the susceptibility responsible for the strength of the spurious couplings, while the linear scale terms remain on par with other 2000Q implementations, see Fig.~\ref{fig:quad_spurious_terms_low_noise}. This result therefore provides strong evidence in support of the noise-based model introduced in this work. We anticipate that the quantitative measurement of the susceptibility associated with the spurious couplings using methods developed in this work will provide a valuable characterization of the qubit noise in the next generations \cite{D-Wave_pegasus} of quantum annealers and other analog machines with binary output statistics.
\section{Open-sourced Tools}
\label{sec:Algortihmic_Tools}
The experiments conducted in this work require the collection of billions of samples from D-Wave's quantum annealer and reconstruction of graphical models with multi-body interactions.
However, neither of these tasks is readily supported by established software and the following software was developed and released as open-source to support this work.
The first software is the D-Wave Ising Sample Collector (DWISC, \url{github.com/lanl-ansi/dwisc}), which enables the collection of millions runs on D-Wave hardware by orchestrating a series of jobs that conform to D-Wave's single-job run time limit of three seconds.
The second software is GraphicalModelLearning (GML, \url{github.com/lanl-ansi/GraphicalModelLearning.jl}), which takes empirical state distributions and reconstructs effective multi-body graphical models in a factor graph representation, leveraging the Interaction Screening method described in Section~\ref{sec:learning_general_distributions} and state-of-the-art second-order nonlinear optimization algorithms to provide model reconstructions that require the least amount of data. The notable improvement of reconstruction accuracy of interaction screening framework over established approaches, such as mean-field, is discussed at length in \cite{lokhov2018optimal}. DWISC and GML form the foundation of the experiments in this work by providing the data and algorithms required to reconstruct high-accuracy multi-body models of the output from D-Wave's quantum annealer.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2x.pdf}
\caption{Quadratic response experiment results for the LANL's \texttt{DW2X} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2x}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_l2k.pdf}
\caption{Quadratic response experiment results for \texttt{DW\_2000Q\_LANL} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_l2k}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_n2k.pdf}
\caption{Quadratic response experiment results for the NASA's \texttt{C16} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_n2k}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k.pdf}
\caption{Quadratic response experiment results for the \texttt{DW\_2000Q\_1} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k2.pdf}
\caption{Quadratic response experiment results for the \texttt{DW\_2000Q\_2} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k3.pdf}
\caption{Quadratic response experiment results for the \texttt{DW\_2000Q\_3} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k3}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k5.pdf}
\caption{Quadratic response experiment results for the lower-noise \texttt{DW\_2000Q\_5} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k5}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k6.pdf}
\caption{Quadratic response experiment results for the lower-noise \texttt{DW\_2000Q\_6} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k6}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.88\linewidth]{fig_quad_final/fig_quad_low_noise.pdf}
\caption{Heat maps representing the quadratic terms of the quadratic response function for the spurious link output parameters on a regular (left) and lower-noise (right) 2000Q quantum annealers. The significant reduction in the link strength in the lower-noise response (from -4.1 to -0.9 and -6.8 to -1.9) confirms the theoretical model of high-frequency qubit noise.}
\label{fig:quad_spurious_terms_low_noise}
\end{figure*}
\end{document}
\section{Primary Hardware Platform and Experimental Settings}
The primary hardware platform on which the majority of experiments have been conducted is a D-Wave 2000Q quantum annealer at Los Alamos National Laboratory, referred to as \texttt{DW\_2000Q\_LANL}. The \texttt{DW\_2000Q\_LANL} QPU chip has a so-called chimera graph structure with $C_{16}$ topology, i.e. it is composed of two dimensional lattice of 16-by-16 unit cells. Each unit cell is composed of 8 qubits connected through a complete bipartite graph structure. A very small number of faulty qubits are disabled and not available for programming. In total, this QPU has 2032 operational qubits and 5924 operational couplers. A complete topology of the \texttt{DW\_2000Q\_LANL} hardware graph is depicted in Fig.~1a of the Main Text.
Most of experiments in this paper deal with a single unit cell with 8 qubits and 16 couplers; the specific identifiers of this cell is given in Table \ref{tbl:l2k-qpu-used}. We denote the set of qubits as $V$ with $\vert V \vert = N$, and the set of couplers as $E$. This specific set of qubits was selected as it is characteristic of a typical complete unit-cell in a hardware chip.
Unless specified otherwise, in this work we set the following additional solver parameters when submitting jobs to the D-Wave hardware: \texttt{auto\_scale = False}, which ensures that the input parameters are not automatically rescaled to utilize the maximal operating range (a feature sometimes used for optimization applications); \texttt{flux\_drift\_compensation = False}, which prevents automatic corrections to input fields based on calibration procedure that is run a few times each hour; \texttt{annealing\_time = 5}, which corresponds to a single-run annealing time of 5$\mu$s; and \texttt{num\_reads = 10000}, which specifies the number of samples collected for a single programming of the chip. The impact of the specific choice of the annealing time in the regime of parameters considered in this work is negligible, as discussed in Section \ref{sec:impact_of_annealing_time}. The motivations for disabling the flux drift compensation and the impact of spin reversal transformation are thoroughly discussed in Sections \ref{sec:effect_of_spin_reversals} and \ref{sec:local_field_variability}.
\begin{table}[!b]
\centering
\begin{tabular}{|p{0.16\columnwidth}|p{0.84\columnwidth}|}
\hline
\multicolumn{2}{|c|}{Unit Cell Considered} \\
\hline
qubits & $V$ = \{304, 305, 306, 307, 308, 309, 310, 311\} \\
\hline
couplers & $E$ =
\{(304, 308), (304, 309), (304, 310), (304, 311),
(305, 308), (305, 309), (305, 310), (305, 311),
(306, 308), (306, 309), (306, 310), (306, 311),
(307, 308), (307, 309), (307, 310), (307, 311)\} \\
\hline
\hline
\multicolumn{2}{|c|}{Unit Cell Subset for High Throughput Experiments} \\
\hline
qubits & $V'$ = \{304, 305, 308, 309\} \\
\hline
couplers & $E'$ =
\{(304, 308), (304, 309), (305, 308), (305, 309)\} \\
\hline
\end{tabular}
\caption{The 8 qubits and 16 couplers used on the \texttt{DW\_2000Q\_LANL} QPU to conduct the primary experiments of this work. A subset of 4 qubits and 4 couplers is specified for use in high-throughput experiments.}
\label{tbl:l2k-qpu-used}
\end{table}
\section{Single-qubit Experiments}
\label{sec:single_spin_exp}
As demonstrated by the results in this paper, in the regime of couplings that are interesting for sampling applications, the output statistics are perfectly described by a certain classical Boltzmann distribution. Here, we investigate at which coupling strength this classical description breaks down, and one needs to introduce a different statistics, such as quantum Boltzmann distribution, for an adequate description of the output data. We show that this regime corresponds to the couplings strengths that are order of magnitude above the intensities that we consider throughout our study.
The experiments consist in looking, for isolated spins, at the relationship between the input magnetic field $h^{\rm{in}}$ and the outcome statistic described by the effective field $h^{\rm{out}}$. The outcome statistic of a single spin being always fully expressible by a probability distribution taking the following form,
\begin{align}
\mu_{\rm{effective}}(\sigma) = \frac{\exp(h^{\rm{out}}\sigma)}{2\cosh{(h^{\rm{out}})}} = \frac{1+\sigma \tanh(h^{\rm{out}})}{2}.\label{eq:classical_single_spin_distribution}
\end{align}
Our process to estimate $h^{\rm{out}}$ for a given value of the input magnetic field consists in the following steps. We start by collecting $M$ samples from the D-wave annealer embodied as a list of single spin realizations $\sigma^{(k)}\in \{-1,1\}$ for $k=1,\ldots,M$. The statistic $S$ that we extract from these samples is the count of positive spin realization $S=\sum_{k=1}^M \delta_{\sigma^{(k)},1}$, where $\delta$ is the Kronecker delta. Assuming that each sample is effectively independent and identically distributed from Eq.~\eqref{eq:classical_single_spin_distribution}, we observe that the statistic $S$ is a Bernoulli process with $M$ trials and probability of success $p = \frac{1+ \tanh(h^{\rm{out}})}{2}$. We estimate the probability of success $p$ using the standard unbiased estimator $\widehat{p} = S/M$ for Bernoulli processes. We compute confidence intervals $I_\alpha = \left[ \underline{p}, \overline{p}\right]$ with confidence level $\alpha$ around our estimator $\widehat{p}$ using the exact method of Crow \cite{Crow1956Confidence} leading to minimal length intervals. Finally, we invert the relationship between $\widehat{p}$ and $h^{\rm{out}}$ to find an estimate of the output effective field. Confidence intervals on $h^{\rm{out}}$ are found using the same relation since it is a monotonic mapping. In the experiments, we have collected $M=5 \times 10^6$ samples using the D-wave spin reversal transform for values of $h^{\rm{in}}$ between $-1$ and $1$. We have chosen the confidence to be $\alpha = 0.997$ corresponding to a ``3$\sigma$" confidence level.
The classical Boltzmann distribution for a single spin at thermal equilibrium and exposed to a magnetic field $h^{\rm{in}}$ is given by the celebrated formula,
\begin{align}
\mu_{\rm{classical}}(\sigma) = \frac{\exp\left(\beta (h^{\rm{in}}+h^{\rm{res}})\sigma\right)}{2\cosh{\left(\beta (h^{\rm{in}}+h^{\rm{res}})\right)}},
\label{eq:classical_single_spin_gibbs}
\end{align}
where $h^{\rm{res}}$ is a residual magnetic field independent of $h^{\rm{in}}$ and $\beta$ is the inverse temperature times the Boltzmann constant $\beta = 1/k_{B}T$. By comparing Eq.~\eqref{eq:classical_single_spin_distribution} with Eq.~\eqref{eq:classical_single_spin_gibbs}, we find that the relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ predicted by a classical Boltzmann distribution is linear
\begin{align}
h^{\rm{out}}_{\rm{classical}} := \beta (h^{\rm{in}}+h^{\rm{res}}).
\label{eq:classical_h_classic}
\end{align}
The classical parameters $\beta$ and $h^{\rm{res}}$ are estimated using the maximum log-likelihood approach for which the log-likelihood function takes the following form,
\begin{align}
\mathcal{L} = \sum_{h^{\rm{in}}} \tanh(h^{\rm{out}}_{\rm{measured}})
h^{\rm{out}}_{\rm{model}} + \frac{1}{2}\ln\left(1-\tanh\left(h^{\rm{out}}_{\rm{model}})\right)^2\right),
\label{eq:classical_likelihood_c}
\end{align}
where $h^{\rm{out}}_{\rm{measured}}$ are measurements of $h^{\rm{out}}$ for different values of $h^{\rm{in}}$ and $h^{\rm{out}}_{\rm{model}}$ describes the modeled relationship between the input and output fields, e.g. $h^{\rm{out}}_{\rm{classical}}$ from Eq.~\eqref{eq:classical_h_classic}.
Measurements of $h^{\rm{out}}$ for different values of $h^{\rm{in}}$ using the aforementioned procedure on spin \#$309$ and the classical relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ are depicted in Fig.~\ref{fig:classical_raw_spin}.
\begin{figure}
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_classic.pdf}
\caption{Effective output magnetic fields $h^{\rm{out}}$ measured for different input magnetic fields $h^{\rm{in}}$ for the spin \#$309$. Each point is estimated using $M=5\times10^6$ samples with a confidence level of $\alpha=0.997$. Notice that the confidence intervals for large value of $h^{\rm{out}}$ are significantly bigger than for small value of $h^{\rm{out}}$ as the probability distribution scales exponentially with the magnitude of $h^{\rm{out}}$. The classical relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ found with the maximum log-likelihood approach is displayed in brown. The effective inverse temperature for this model is $\beta=10.5$ and the residual field is $h^{\rm{res}}=0.004$.
}
\label{fig:classical_raw_spin}
\end{figure}
We clearly see that if the classical relationship holds for small values of $h^{\rm{in}}$ it fails to explain the flat tails of the measurements for which $|h^{\rm{in}}| \gtrapprox 0.5$. In fact, we starts to already see a deviation from the linear curve for magnitudes of the input field between $0.5 \gtrapprox |h^{\rm{in}}| \gtrapprox 0.2$. In the forthcoming analysis, we will show that the first behavior can be explained by quantum effects, and the latter by classical noise in the input field.
A quantum statistical description of a one spin system is realized through the so-called density matrix formalism. We consider a quantum spin exposed on the $z$-axis to a magnetic field $h^{\rm{in}}$ and residual field $h^{\rm{res}}$ and exposed on the $x$-axis to a transverse field $h^{\rm{trans}}$. This is a natural assumption as the Hamiltonian realized during the considered quantum annealing process is composed of these two terms \cite{johnson2011quantum}. The density matrix describing this one spin system at thermal equilibrium is the following object,
\begin{align}
\rho &= \frac{\exp\left( \beta( h^{\rm{in}} + h^{\rm{res}}) \widehat{\sigma}_z + \beta h^{\rm{trans}} \widehat{\sigma}_x\right)}{\Tr{\exp\left( \beta( h^{\rm{in}} + h^{\rm{res}}) \widehat{\sigma}_z + \beta h^{\rm{trans}} \widehat{\sigma}_x\right)}} \nonumber\\
&= \frac{1}{2}\left(I + \frac{\tanh\left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(h^{\rm{trans}})^2}\right)}{\sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(h^{\rm{trans}})^2}}\left((h^{\rm{in}}+ h^{\rm{res}})\widehat{\sigma}_z + h^{\rm{trans}}\widehat{\sigma}_x\right)\right),
\label{eq:classical_density_matrix}
\end{align}
where $\widehat{\sigma}_z$ and $\widehat{\sigma}_x$ are the usual Pauli matrices for the $z$ and $x$ axis respectively. The transverse field $h^{\rm{trans}}$ that appears in Eq.~\eqref{eq:classical_density_matrix} is in general a function of $h^{\rm{in}}$. Due to an observation that the experimental points seem to flatten out for large values of $h^{\rm{in}}$ in Fig.~\ref{fig:classical_raw_spin}, we choose to parametrize the transverse fields as linear transformations of the input field, namely $h^{\rm{trans}} = \xi h^{\rm{in}}$, a dependence which is consistent with a saturation in the $h^{\rm{out}}$ response.
The mean value of observing the quantum spin along the $z$-axis is given by $\Tr{\rho \widehat{\sigma}_z}$. From this relationship we deduce that the probability of observing the system taking the value $\sigma \in \{-1,1\}$ is given by the following probability distribution,
\begin{align}
\mu_{\rm{quantum}}(\sigma) = \frac{1}{2}\left(1 + \sigma (h^{\rm{in}}+ h^{\rm{res}})\frac{\tanh\left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}\right)}{\sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}}\right).
\label{eq:classical_single_spin_quantum}
\end{align}
With these assumptions we find that that the relationship between $h^{\rm{out}}$ and $h^{\rm{in}}$ in the quantum case can be described as follows,
\begin{align}
h^{\rm{out}}_{\rm{quantum}} := \arctanh \left((h^{\rm{in}}+ h^{\rm{res}})\frac{\tanh\left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}\right)}{\sqrt{(h^{\rm{in}}+ h^{\rm{res}})^2 +(\xi h^{\rm{in}})^2}}\right).
\label{eq:classical_h_quantum}
\end{align}
Notice that in the limit where $\xi=0$, the quantum predictions from Eq.~\eqref{eq:classical_h_quantum} converges to its classical counterpart from Eq.~\eqref{eq:classical_h_classic}.
As we will see later in Section~\ref{sec:local_field_variability}, the variability in the residual field $h^{\rm{res}}$ is non-negligible and in fact plays a key role in the explanation of the spurious link behaviors, see Section~\ref{sec:noise_and_spurious_links} for a detailed explanation.
Assuming that $h^{\rm{res}}$ is a random variable with probability density function $f(h^{\rm{res}})$, we can obtain a noisy quantum description of the measurement outcomes by applying Bayes's rule to Eq.~\ref{eq:classical_h_quantum},
\begin{align}
\mu_{\rm{qnoise}} (\sigma) = \int dh^{\rm{res}} f(h^{\rm{res}}) \mu_{\rm{quantum}}(\sigma).
\label{eq:classical_qnoise_measure}
\end{align}
We will see later that in the regime of noise relevant to our experiments, the precise knowledge about the shape of the distribution $f(h^{\rm{res}})$ plays little role. The important quantities are the mean $\mathbb{E}_f[h^{\rm{res}}] = h^{\rm{res}}_0$ and the standard deviation $\sqrt{\mathbb{E}_f[\left( h^{\rm{res}} - h^{\rm{res}}_0\right)]} = h^{\rm{res}}_{\rm{sd}}$. For simplicity, we choose $f$ to be a binomial distribution $f(h^{\rm{res}}) = \frac{1}{2}\delta(h^{\rm{res}} - h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}}) + \frac{1}{2}\delta(h^{\rm{res}} - h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}})$, where $\delta$ denotes the Dirac distribution. Combining Eq.~\eqref{eq:classical_single_spin_quantum} with Eq.~\eqref{eq:classical_qnoise_measure}, we can write down the following relationship between $h^{\rm{in}}$ and $h^{\rm{out}}$ in the noisy quantum case,
\begin{align}
h^{\rm{out}}_{\rm{qnoise}} := \arctanh &\left( (h^{\rm{in}} + h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}}) \frac{\tanh \left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}\right)}{2\sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 + h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}} \right.\nonumber\\
& \left. + (h^{\rm{in}} + h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}}) \frac{\tanh \left( \beta \sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}\right)}{2\sqrt{(h^{\rm{in}}+ h^{\rm{res}}_0 - h^{\rm{res}}_{\rm{sd}})^2 +(\xi h^{\rm{in}})^2}}\right).
\label{eq:classical_h_quantum_noise}
\end{align}
Note that for $h^{\rm{res}}_0 = 0$, the first order expansion of Eq.~\eqref{eq:classical_h_quantum_noise} for small values of $h^{\rm{in}}$ yields the simple relationship $h^{\rm{out}}_{\rm{qnoise}} \approx \beta h^{\rm{in}} \cosh(\beta h^{\rm{res}}_{\rm{sd}})^{-2}$. This shows that one effect of noise on the system consists in reducing the effective inverse temperature that one could infer from a classical (i.e. linear) regression for small values of $h^{\rm{in}}$.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_noise_quantum.pdf}
\caption{Effective output magnetic fields $h^{\rm{out}}$ measured for different input magnetic fields $h^{\rm{in}}$ for the spin \#$309$. Each point is estimated using $M=5\times10^6$ samples with a confidence level of $\alpha=0.997$. The noiseless quantum and the noisy quantum relationships between $h^{\rm{in}}$ and $h^{\rm{out}}$ found with the maximum log-likelihood approach are displayed in red and blue respectively. The effective inverse temperature for the noisy quantum model is higher than the effective inverse temperature found with the noiseless quantum model. The noiseless quantum model and the classical models find similar temperatures. }
\label{fig:classical_noise_quantum}
\end{figure}
We infer the parameters of the proposed quantum models, i.e. $h^{\rm{res}}_0$, $\xi $ and $h^{\rm{res}}_{\rm{sd}}$, with the maximum log-likelihood approach described by Eq.~\eqref{eq:classical_likelihood_c}, replacing the quantity $h^{\rm{out}}_{\rm{model}}$ by either the functional relationship $h^{\rm{out}}_{\rm{quantum}}$ from Eq.~\eqref{eq:classical_h_quantum} or $h^{\rm{out}}_{\rm{qnoise}}$ from Eq.~\eqref{eq:classical_h_quantum_noise}.
In Fig.~\ref{fig:classical_noise_quantum}, we show the measurements acquired for spin \#$309$ along with the curves predicted by the noiseless and noisy quantum models. The flattening of the output field response for $|h^{\rm{in}}|\gtrapprox 0.5$ is correctly accounted for by the addition of the quantum transverse field. The initial deviation from the linear response for values of the input field between $0.5 \gtrapprox|h^{\rm{in}}|\gtrapprox 0.2$ are predicted accurately only by the noisy quantum model. An important feature highlighted by this study is the effect of the noise on the effective inverse temperature (or linear response) that one may infer from the data. We see from our models that the noise in the residual field effectively lowers the inverse temperature for small values of the input field of order $|h^{\rm{in}}|\lessapprox 0.2$. Indeed, the effects due to noise will be more prominent when the input field is small. In this particular illustration, the noise on the residual field accounts for more than $20\%$ of the input field for magnitudes of $|h^{\rm{in}}|\leq 0.2$.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_noiseless_quantum.pdf}
\caption{Effective output magnetic fields $h^{\rm{out}}$ measured for different input magnetic fields $h^{\rm{in}}$ for the spin \#$307$. Each point is estimated using $M=5\times10^6$ samples with a confidence level of $\alpha=0.997$. The noiseless quantum and the noisy quantum relationships between $h^{\rm{in}}$ and $h^{\rm{out}}$ found with the maximum log-likelihood approach are displayed in red and blue respectively. This particular spin displays a low-level of residual field noise. As a consequence, the noiseless and noisy quantum models find the same inverse temperature.}
\label{fig:classical_noiseless_quantum}
\end{figure}
In Fig.~\ref{fig:classical_noiseless_quantum}, we show similar measurements collected for spin \#$307$ as long as the predictions of the proposed quantum models. Among all spins that we tested, we found that spin \#$307$ has the lowest level of residual field noise. In this particular case, we see that both noiseless and noisy quantum model predictions agree and we find an inverse temperatures consistent with the inverse temperature found for spin \#$309$ with the noisy quantum model.
For the unit cell of interest to this work, we show the regression coefficients obtained with the classical model, the noiseless quantum model and the noisy quantum model in Table~\ref{tbl:classical_reg_coeff}.
\begin{table*}
\centering
\begin{tabular}{|c||c|c|c|}
\hline
Spin & Classical & Quantum & Noisy + Quantum \\
\# & $(\beta, h^{\rm{res}}_0)$ & $(\beta, h^{\rm{res}}_0, \xi)$ & $(\beta, h^{\rm{res}}_0, \xi, h^{\rm{res}}_{\rm{sd}})$ \\
\hline
\hline
304 & (11.3, 0.014) & (11.4, 0.014, 0.017) & (12.3, 0.014, 0.018, 0.029) \\
\hline
305 & (11.1, 0.003) & (11.2, 0.003, 0.016) & (13.1, 0.003, 0.017, 0.041) \\
\hline
306 & (11.4, -0.004) & (11.5, -0.004, 0.016) & (13.3, -0.004, 0.018, 0.039) \\
\hline
307 & (12.1, -0.009) & (12.2, -0.009, 0.016) & (12.4, -0.009, 0.016, 0.012) \\
\hline
308 & (11.6, -0.005) & (11.7, -0.005, 0.012) & (12.9, -0.005, 0.013, 0.032) \\
\hline
309 & (10.5, 0.004) & (10.5, 0.004, 0.011) & (12.7, 0.004, 0.013, 0.048) \\
\hline
310 & (11.2, -0.006) & (11.3, -0.006, 0.012) & (12.5, -0.006, 0.013, 0.035) \\
\hline
311 & (11.4, 0.010) & (11.5, 0.010, 0.012) & (12.5, 0.010, 0.013, 0.031) \\
\hline
\end{tabular}
\caption{List of regression coefficients for eight different spins found with the classical model, the noiseless quantum model and the noisy quantum model. The coefficients are obtained through the maximization of the log-likelihood function respective to each model. The inverse temperature found by the classical and noiseless quantum models differs from the inverse temperature obtained with the noisy quantum model for significant noise magnitudes. The other parameters are consistent between models which shows that these models form a hierarchy of increasing descriptive complexity. The transverse field responses for spins \#304 to \#307 differs significantly from the transverse field responses of spins \#308 to \#311. This suggest a difference in the hardware implementation of these two groups of spins for they are physically located on the two different sides of a chimera cell.}
\label{tbl:classical_reg_coeff}
\end{table*}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.84\linewidth]{results/single_spin/single_spin_noise_comp.pdf}
\caption{Comparison of the effect of different noise probability distributions for noisy quantum models on spin \#$309$. Predictions obtained from noisy quantum models derived from a binary distribution and a uniform distribution are displayed in blue and purple respectively. The parameters for both models are identical and corresponds to the parameters inferred for the binary noise model. The two models are indistinguishable for this level of noise magnitude, indicating that the main contributing factors in the probabilistic description of the noise are the first and the second moments.}
\label{fig:classical_noise_comparison}
\end{figure}
Finally, we study the effect of the noise probability distribution on the model. We compare our noisy quantum model derived from a binary probability distribution with the noisy quantum model that one obtains with a uniform noise distribution, i.e. $f(h^{\rm{res}}) = (2\sqrt{3} h^{\rm{res}}_{\rm{sd}})^{-1}$ for $h^{\rm{res}} \in \left[h^{\rm{res}}_0 - \sqrt{3} h^{\rm{res}}_{\rm{sd}}, h^{\rm{res}}_0 + \sqrt{3} h^{\rm{res}}_{\rm{sd}}\right]$ and zero otherwise. Predictions for the spin \#$309$ made by these two models of noise are displayed in Fig.~\ref{fig:classical_noise_comparison}. The parameters for both models are identical and correspond to the parameters inferred with maximum log-likelihood for the binary probability distribution. We see that for noise magnitudes of $h^{\rm{res}}_{\rm{sd}} = 0.048$ the predictions are practically indistinguishable.
\section{Learning of General Classical Distributions on Binary Variables}\label{sec:learning_general_distributions}
An arbitrary positive probability distribution on $N$ classical binary variables $\underline{\sigma} \in \{ -1, +1 \}^N$ can be represented in the the form of a Gibbs distribution with different interaction orders:
\begin{align}
\mu(\underline{\sigma}) & = \frac{1}{Z} \exp{ \left( \underbrace{\sum_{i} h_i \sigma_i}_{\text{1st order}} + \underbrace{\sum_{ij} J_{ij} \sigma_i \sigma_j}_{\text{2d order}} + \underbrace{\sum_{ijk} J_{ijk} \sigma_i \sigma_j \sigma_k}_{\text{3d order}} + \underbrace{\sum_{ijkl} J_{ijkl} \sigma_i \sigma_j \sigma_k \sigma_l}_{\text{4th order}} + \cdots \right)},
\label{eq:Gibbs_distribution}
\end{align}
where $Z$ denotes the normalization factor known as parition function. In general, the number of terms can be exponential. However, physical systems are typically characterized by a finite number of short-ranged multi-body interactions. For instance, the renowned Ising model corresponds to the case where only first and second orders are present. In this section, we develop an exact learning method that allows one to reconstruct an arbitrary probability distribution on binary variables, i.e. recover the parameters $\{h_i, J_{ij}, J_{ijk}, J_{ijkl}, \ldots \}$ from a number of independent spin configurations sampled from the distribution \eqref{eq:Gibbs_distribution}. For the case of Ising models with pairwise interactions only
\begin{align}
\mu(\underline{\sigma}) = \frac{1}{Z} \exp{ \left( \sum_{i} h_i \sigma_i + \sum_{ij} J_{ij} \sigma_i \sigma_j \right)},
\label{eq:Ising_distribution}
\end{align}
this reconstruction problem is known as \emph{inverse Ising problem}, and has been extensively studied in the past with a number of heuristic methods \cite{nguyen2017inverse}. However, the inverse Ising problem has been solved only recently, with the appearance of exact algorithms that showed that parameters of an arbitrary Ising model, including low-temperature and spin-glass models, can be recovered to an arbitrary precision with an appropriate number of samples \cite{bresler2015efficiently, vuffray2016interaction}. The state-of-the-art near-optimal performance for inverse Ising problem has been recently achieved with the estimator based on the Interaction Screening Objective (ISO) \cite{lokhov2018optimal}. For Ising model, the ISO reads
\begin{equation}
S_i(\mathbf{J_i},h_i) = \left\langle \exp{ \left( - h_i \sigma_i - \sum_{j} J_{ij} \sigma_i \sigma_j \right)} \right\rangle_M,
\label{eq:ISO_Ising}
\end{equation}
where $\langle f(\sigma) \rangle_M = M^{-1} \sum_{m=1}^M f(\sigma^{(m)})$ denotes the empirical average over $M$ independent samples, and $\mathbf{J_i}$ denotes the set of couplings adjacent to node $i$, i.e. $\mathbf{J_i} = \{ J_{ij} \}_{j \neq i}$ for Ising models. It is easy to see that the ISO in \eqref{eq:ISO_Ising} is a convex function of parameters $\mathbf{J_i},h_i$. Furthermore, the unique minimizer of the ISO
\begin{equation}
(\widehat{\mathbf{J}_i},\widehat{h_i}) = \operatornamewithlimits{argmin} S_i(\mathbf{J_i},h_i)
\label{eq:ISO_estimation}
\end{equation}
converges to the true parameters of the distribution in the limit of a large number of samples, and yields an $O(1/\sqrt{M})$ error on the recovered model parameters for finite $M$ \cite{vuffray2016interaction}. The ISO is a local estimator, i.e. it is defined for each spin, and only involves couplings adjacent to this spin, so for reconstructing the entire model one needs to run $N$ parallel reconstruction problems \eqref{eq:ISO_estimation}. Notice that this procedure yields two estimations for the same coupling, $\widehat{J_{ij}}$ and $\widehat{J_{ji}}$, and we use their arithmetic mean $(\widehat{J_{ij}} + \widehat{J_{ji}})/2$ as the final estimate of the coupling $J_{ij}$. This estimator is used throughout the paper for reconstructing parameters of Ising models, where the required number of samples for a given precision and the associated expected confidence interval are obtained through synthetic numerical experiments, as explained below.
Here, we generalize this objective function to the case of general Gibbs distributions with multi-body interactions of the type \eqref{eq:Gibbs_distribution}. The corresponding ISO reads
\begin{equation}
S_i(\mathbf{J_i},h_i) = \left\langle \exp{ \left( - h_i \sigma_i - \sum_{j} J_{ij} \sigma_i \sigma_j - \sum_{jk} J_{jkl} \sigma_i \sigma_j \sigma_k - \sum_{jkl} J_{jkl} \sigma_i \sigma_j \sigma_k \sigma_l - \cdots \right)} \right\rangle_M.
\label{eq:ISO_General}
\end{equation}
Similarly to the Interaction Screening estimator for the inverse Ising problem \eqref{eq:ISO_Ising}, it is easy to see that the ISO for general models is a convex function of parameters $\mathbf{J_i},h_i$. Let us present a simple argument that illustrates the fact that in the limit of large number of samples the unique minimizer \eqref{eq:ISO_estimation} of the convex ISO objective \eqref{eq:ISO_General} is achieved at $(\widehat{\mathbf{J}_i},\widehat{h_i}) = (\mathbf{J}_i, h_i)$, meaning that the true interactions present in the model are fully ``screened''. Indeed, the ISO is an empirical average of the inverse of the factors in the Gibbs measure; if
\begin{equation}
\mathcal{F}_i(\mathbf{J}_i, h_i) = \exp( h_i \sigma_i + \sum_{j} J_{ij} \sigma_i \sigma_j + \sum_{jk} J_{jkl} \sigma_i \sigma_j \sigma_k + \sum_{jkl} J_{jkl} \sigma_i \sigma_j \sigma_k \sigma_l + \cdots),
\end{equation}
then $S_i(\mathbf{J}_i,h_i) = \langle \mathcal{F}_i^{-1}(\mathbf{J}_i,h_i) \rangle_{M}$. In the limit of large number of samples $S_i(\mathbf{J}_i,h_i) \rightarrow S_i^*(\mathbf{J}_i,h_i) = \langle 1/\mathcal{F}_i(\mathbf{J}_i,h_i) \rangle$, where $\langle \cdots \rangle$ denotes the average over the measure \eqref{eq:Gibbs_distribution}. Let us look at the derivative of the ISO with respect to a given coupling, say $J_{ijk}$. This derivative corresponds to weighted three-body correlation, for instance $\partial S_i^{*} / \partial J_{ijk} = \langle\sigma_i \sigma_j \sigma_k / \mathcal{F}_i(\mathbf{J}_i,h_i)\rangle$, and this reflect the key property of the Interaction Screening based estimator. When $(\widehat{\mathbf{J}_i},\widehat{h_i}) = (\mathbf{J}_i, h_i)$, $\partial S_i^{*} / \partial J_{ij} \vert_{\mathbf{J}_i,h_i} = 0$, meaning that the minimum of ISO is achieved at $(\widehat{\mathbf{J}_i},\widehat{h_i}) = (\mathbf{J}_i, h_i)$ as $M \to \infty$.
Again, similarly to the case of Ising models, after running $N$ parallel local reconstructions, the resulting couplings $\mathbf{J_i} = \{ J_{ij}, J_{ijk}, J_{ijkl}, \ldots \}$ are symmetrized using all permutations of tuples $(i, j, k, l, \ldots)$. This estimator is used for probing general distributions with multi-body interactions in the next section.
It is important to highlight that the estimator given by the generalized ISO \eqref{eq:ISO_General} is exact: With a given number of samples $M$, the deviation between reconstructed $(\widehat{\mathbf{J}_i},\widehat{h_i})$ and true $(\mathbf{J}_i, h_i)$ model parameters decay as $\sim 1/\sqrt{M}$. Next, we describe an empirical procedure that we have developed to estimate the reconstruction fidelity in practice, which dictate how much data we needed in all experiments to statistically exclude the finite-sample considerations from all conclusions that we draw throughout the work.
\section{Empirical Estimation of Reconstruction Errors}
\label{sec:empirical_error_estimation}
For finite number of samples $M$, couplings estimated via \eqref{eq:ISO_estimation} are accurate up to a certain error that decays with $M$. This error can be estimated theoretically, however the resulting worst-case bounds can be loose for a given model. To the best of our knowledge, there is no standard approach to tightly quantify the finite sampling error of learning models of the form \eqref{eq:Gibbs_distribution} in practice. To address this challenge, we propose the following empirical error estimation procedure for a given model with specific parameters and a fixed number of samples. Specifically, given a black-box sample generator $B$, a finite sample set $M$, and a replicate parameter $R$, we conduct the following procedure:
\begin{enumerate}
\item collect $M$ samples of the black-box $B$;
\item reconstruct model $m$ from the collected samples;
\item for $r$ from $1$ to $R$ independent replicates do the following:
\begin{enumerate}
\item collect $M$ synthetic samples from $m$ using an auxiliary sampling algorithm;
\item reconstruct model $m_r$ from the collected samples;
\end{enumerate}
\item compute statistics over the parameters reconstructed across the reconstructed $\{m_r\}_{r \in 1,\ldots,R}$ models.
\end{enumerate}
In this work the black-box sampler $B$ is given by our \texttt{DW\_2000Q\_LANL} QPU, and the sampling algorithm is a brute-force approach that enumerates every possible state and computes the exact probably of each state. The this brute-force approach is feasible for small number of spins which will be the focus of our targeted experiments. For larger problems one could utilize more scalable sampling algorithms, including those based on Markov-Chain Monte-Carlo techniques, or Belief Propagation with decimation.
The last step will provide the information on the typical variability of reconstruction accuracy due the effect of finite samples. Given a set of $R$ replicates of the models of the type \eqref{eq:Gibbs_distribution} reconstructed from $R$ independent sets of samples. Let us define $(\mathbf{h}^r, \mathbf{J}^r) \;\; \forall r \in R$ as parameters of the models of the type \eqref{eq:Gibbs_distribution} learned from each of the $R$ sets of samples. We also define $(\delta \mathbf{h}, \delta \mathbf{J})$ representing deviations from the parameters $(\mathbf{h}, \mathbf{J})$ reference model that was used to produce synthetic samples:
\begin{align}
\delta \mathbf{h}^r = \mathbf{h} - \mathbf{h}^r \;\; \forall r \in R \\
\delta \mathbf{J}^r = \mathbf{J} - \mathbf{J}^r \;\; \forall r \in R
\end{align}
In the step 4 of the procedure above, we estimate the empirical mean and standard deviation on each of the parameters in the set $h_i, \mathbf{J}_i$ from $R$ values. For a sufficiently large $R$, these quantities indicate an expected scale of error coming from model recovery for a given number of samples.
\section{Probing of Multi-Body Interactions}
\label{sec:multi-body}
The objective of experiment in this section is to leverage the ISO for general distributions on binary variables \eqref{eq:ISO_General} to determine the class of models that adequately describes the output distribution produced by the \texttt{DW\_2000Q\_LANL} QPU. Although the target Hamiltonian has a form of a classical Ising model, \emph{a priori} this distribution could be more general than the Ising Gibbs distribution \eqref{eq:Ising_distribution}. In this section, we probe the existence of multi-body interactions beyond pairwise in the energy function of the output distribution. We conduct an experiment on a single chimera graph cell involving 8 qubits (see Table \ref{tbl:l2k-qpu-used}). As explained in the previous section, for an 8-spin model, the most general positive probability distribution can be written in the form \eqref{eq:Gibbs_distribution} with an energy function being a polynomial of order eight. Specifically, for $\underline{\sigma} \in \{-1, +1\}^8$, the possible distribution reads
\begin{align}
\mu(\underline{\sigma}) = \frac{1}{Z} \exp{ \left( \sum_{i} h_i \sigma_i + \sum_{ij} J_{ij} \sigma_i \sigma_j + \sum_{ijk} J_{ijk} \sigma_i \sigma_j \sigma_k + \cdots + J_{ijklmnop} \sigma_i \sigma_j \sigma_k \sigma_l \sigma_m \sigma_n \sigma_o \sigma_p \right) }.
\label{eq:ho_boltzmann}
\end{align}
This model has a total of 8 $h$ parameters 247 $J$ parameters.
Our goal will be to practically show existence or absence of multi-body interactions in the output distribution. Presence of interactions can be established if the reconstructed couplings are statistically significant, i.e. they are larger in absolute value than the reconstruction error resulting from a finite-sample reconstruction. We show that a few million D-Wave samples will be sufficient to provide an accurate reconstruction of model parameters.
\begin{table}
\centering
\begin{tabular}{|r||r|r|r|r|r|}
\hline
Model & Qubits & Couplers & $h^{\text{in}}$ & $J^{\text{in}}$ & Samples \\
\hline
\hline
ferromagnet & $V$ & $E$ & 0.0 & 0.025 & 10,000,000 \\
\hline
anti-ferromagnet & $V$ & $E$ & 0.0 & -0.025 & 10,000,000 \\
\hline
\end{tabular}
\caption{The input parameters sent to the \texttt{DW\_2000Q\_LANL} QPU to collect data for the 8-th order distribution reconstruction experiment.}
\label{tbl:8o-models}
\end{table}
The experiment we conduct here focuses on learning 8-th order models from the samples output by the D-Wave hardware on two canonical models, a ferromagnet and an anti-ferromagnet. The parameter details of these two models are presented in Table \ref{tbl:8o-models}. The primary objective of this experiment is to determine what model parameters are statistically significant. Specifically, what learned parameters can-and-cannot be attributed to artifacts from to finite sampling and the model reconstruction algorithm. Through a procedure that was detailed in section \ref{sec:empirical_error_estimation}, a number of supporting simulations are conducted to determine error bounds on the reconstructed model parameters. Leveraging the obtained error value, we determine the three standard deviations threshold that determine the statistical significance of the reconstructed values. Recovered couplings with absolute values above this threshold are very unlikely to be due to a reconstruction error, while values below can be explained by the finite sample noise in the reconstruction process.
Figs.~\ref{fig:fm-8o-parameter-hist} and \ref{fig:afm-8o-parameter-hist} present the absolute values of the 255 model parameters broken down by the interaction-order for the ferromagnetic and the anti-ferromagnetic cases, respectively. We find that in both cases, a second-order model provides an accurate representation of the output distribution of \texttt{DW\_2000Q\_LANL}. This experiment thus provides a convincing evidence that a second-order model is sufficient for modeling the distribution that the quantum annealer samples from for a range of input parameters of interest to this work.
\begin{figure}
\centering
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=7]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=8]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=9]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=10]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=11]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=12]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=13]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\caption{Histograms of reconstruction parameters magnitudes from the 8-th order reconstruction experiment on the Ferromagnet model. Values above the $3\sigma$ line are considered to be statically significant, while the values below are artifacts of finite sampling. These results indicate that the distribution that the \texttt{DW\_2000Q\_LANL} hardware is sampling from is well approximated by a second-order model.}
\label{fig:fm-8o-parameter-hist}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=7]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=8]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=9]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=10]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=11]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf} \\
\vspace{-0.3cm}
\includegraphics[width=0.39\linewidth,page=12]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=13]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_hist.pdf}
\caption{Histograms of reconstruction parameters magnitudes from the 8-th order reconstruction experiment on the Anti-Ferromagnet model. Values above the $3\sigma$ line are considered to be statically significant, while the values below can be artifacts of finite sampling. These results indicate that the distribution that the \texttt{DW\_2000Q\_LANL} hardware is sampling from is well approximated by a second-order model.}
\label{fig:afm-8o-parameter-hist}
\end{figure}
We further validate these results by quantifying the typucal variation of model parameters used in the 8-th order reconstruction based on finite sampling error. To that end, the validation protocol described in section \ref{sec:empirical_error_estimation} is executed with with $R=50$ replicates and $M=10^7$, to replicate the number of samples used in the data collection for most of the key experiments in this work. Fig.~ \ref{fig:8o-parameter-variance} shows data for the mean and the standard deviation for each parameter deviation $\delta h$ or $\delta J$ from the reference model, estimated from running the variance measuring procedure defined in the section \ref{sec:empirical_error_estimation} on the reconstruction models. We see that estimated variance of reconstructed values is very similar across different couplings. The average-case variance across all model parameters using three standard deviations are 0.0034 and 0.0021 for the Ferromagnet and Anti-Ferromagnet models, respectively. These values have been used for computing the threshold values that appear in Figs.~ \ref{fig:fm-8o-parameter-hist} and \ref{fig:afm-8o-parameter-hist} in the previous section.
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/fmh_1x1_l2k/00250_00000_base_8o_reconst_resample_delta_hist.pdf}
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/afmh_1x1_l2k/00250_00000_base_8o_reconst_resample_delta_hist.pdf}
\caption{Multi-body reconstruction variability results for the Ferromagnet model (left) and Anti-Ferromagnet (right) used in the multi-body probing experiment, using the validation protocol from section \ref{sec:empirical_error_estimation} with $R=50$ replicates. The mean values of the all 255 parameters are presented with whiskers shown at three standard deviations from the mean. The average-case deviations are 0.0034 and 0.0021, respectively.}
\label{fig:8o-parameter-variance}
\end{figure}
\section{Reconstruction and Validation of Two-Body Models}
\label{sec:two-body}
The previous section conducted a 8-th order reconstruction with 10,000,000 samples and argued that a 2-nd order model provides a sufficient approximation of the distribution that the \texttt{DW\_2000Q\_LANL} hardware samples from. To that end, the remaining experiments in this work focuses on learning 2-nd order models of the hardware's output distribution.
Leveraging the knowledge that this 2-nd order model is sufficient has a significant advantages in that the number of model parameters reduces from 255 (8-th order) to only 36 (2-nd order), which in turn reduces the amount of data required to accurately learn the associated 2-nd order model. To further lessen the data requirements, we reduce the number of qubits considered from 8 to 4, focusing on the upper-half of the cell. This effectively decreases the number of 2-nd order parameters from 36 to 16. It will later become evident that these reductions are necessary to make the experiments viable on reasonable time scales.
After conducting these model reductions, we replicate here the validation experiments from the section \ref{sec:multi-body} to establish an optimal number of samples required for reconstruction of parameters in the two-body model. Table \ref{tbl:2o-models} specifies the input parameters for two additional models, {\em strong ferromagnet} and {\em strong anti-ferromagnet}, using coupling sign convention consistent with \eqref{eq:Gibbs_distribution}.
\begin{table}
\centering
\begin{tabular}{|r||r|r|r|r|r|}
\hline
Model & Qubits & Couplers & $h^{in}$ & $J^{in}$ & Samples \\
\hline
\hline
strong ferromagnet & $V'$ & $E'$ & 0.0 & 0.05 & $4\cdot10^6$ \\
\hline
strong anti-ferromagnet & $V'$ & $E'$ & 0.0 & -0.05 & $4\cdot10^6$ \\
\hline
\end{tabular}
\caption{The input parameters sent to the \texttt{DW\_2000Q\_LANL} QPU to collect data for the 2-nd order distribution reconstruction experiment.}
\label{tbl:2o-models}
\end{table}
Fig.~\ref{fig:2o-strong-parameter-variance} presents the results of the validation experiment for the 2-nd order model reconstruction. 50 reconstruction replicates are used in the validation experiments conducted in this section. The results indicate that reconstruction accuracy is approximately 0.0025 and 0.0022, which is comparable to the accuracy used in the 8-th order reconstruction experiment, using a smaller number of samples.
Fig.~\ref{fig:2o-strong-parameter-values} presents the strength of the second order terms that are recovered from the hardware data. These absolute values are well above the recovery accuracy threshold, indicating that $4 \cdot 10^6$ samples are sufficient for accurately recovering the two-body model. We use this number of samples in the remainder of experiments in this paper.
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/fms_1x1_l2k/00500_00000_base_reconst_resample_delta_hist.pdf}
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/afms_1x1_l2k/00500_00000_base_reconst_resample_delta_hist.pdf}
\caption{Two-body reconstruction variability results for the Strong Ferromagnet model (left) and Strong Anti-Ferromagnet (right), using the proposed validation protocol with 50 replicates. The mean values of the all 16 parameters are presented with whiskers shown at three standard deviations from the mean. The average-case deviations are 0.0025 and 0.0022, respectively.}
\label{fig:2o-strong-parameter-variance}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=6]{results/multibody/fms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\includegraphics[width=0.41\linewidth,page=6]{results/multibody/afms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\caption{Second-Order terms of the two-body reconstruction of the for the Strong Ferromagnet model (left) and Strong Anti-Ferromagnet (right). The recovered values are greater than 0.03, well above the accuracy threshold of 0.0025, suggesting an accurate model.}
\label{fig:2o-strong-parameter-values}
\end{figure}
In Fig.~2 of the Main Text, we summarize the results of these experiments and show the structure of the output distribution. In particular, we find that among statistically significant couplings in the output distribution, the strongest ones are in correspondence with the input couplings, while the weakest ones are the \emph{spurious couplings} that are not present in the input problem and, moreover, are absent in the chip topology. In what follows, we construct additional experiments aimed at clarifying the nature of these spurious couplings.
\section{Impact of the Annealing Time}
\label{sec:impact_of_annealing_time}
In all presented experiments thus far, we used the annealing time of 5$\mu s$ per each sample. It is important to understand how the statistics of these reconstruction experiments might differ as the annealing time varies. Here, we replicate the model reconstruction experiment for the Strong Ferromagnet and Strong Anti-Ferromagnet models, using the following varying annealing time parameters, \texttt{annealing\_time = 1, 5, 25, 125, 625}, which corresponds to single-run annealing time of 1$\mu s$, 5$\mu s$, 25$\mu s$, 125$\mu s$, 625$\mu s$ respectively. Additionally, in this experiment, the \texttt{num\_reads} parameter was reduced from 10,000 (this work's default) to 4,000 in all cases, to adhere to the maximum job run-time limit of the \texttt{DW\_2000Q\_LANL} QPU.
\begin{figure*}
\centering
\includegraphics[width=0.88\linewidth,page=1]{results/multibody/fms_1x1_l2k/00500_00000_base_annealing_time_series.pdf}
\includegraphics[width=0.88\linewidth,page=1]{results/multibody/afms_1x1_l2k/00500_00000_base_annealing_time_series.pdf}
\caption{A comparison of the second-order model reconstructions for different annealing times on the Strong Ferromagnet model (top) and Strong Anti-Ferromagnet model (bottom). The results indicate that large increases in the annealing time results in slight reductions in the effective temperature of the reconstructed models.}
\label{fig:2o-strong-annealing-time}
\end{figure*}
Fig.~\ref{fig:2o-strong-annealing-time} presents the results of this experiment. We find that an increase of the annealing time by two orders of magnitudes results only in a slight increase of the absolute coupling values in the reconstructed model. At the same time, this minor change comes at a significant increase in data collection time. Due to this observation, in this work we chose to standardize around the annealing time of 5$\mu s$, which is essential for the high-throughput data collection, and represents the fastest available annealing time across several generations of quantum annealers.
\section{Quadratic Response}
\label{sec:quadratic_response}
In this section, we provide additional details on the study of the quadratic response experiment described in the Main Text. We assume the most general parametrization of the quadratic input-output relationship between the input parameters and the parameters of the reconstructed output distribution. Specifically, for each output model parameter we learn the quadratic $\chi^{hh}, \chi^{hJ}$, linear $\beta^h, \beta^J$ and offset $c$ coefficients of the following quadratic function,
\begin{align}
h_i &= [h^{in} \;\; J^{in}] \begin{bmatrix} \chi^{hh}_i & \chi^{hJ}_i \\ \chi^{Jh}_i & \chi^{JJ}_i \end{bmatrix} \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + [\beta^h_i \;\; \beta^J_i] \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + c_i \;\; \forall i \in V' \label{eq:h_quad_model}\\
J_{ij} &= [h^{in} \;\; J^{in}] \begin{bmatrix} \chi^{hh}_i & \chi^{hJ}_{ij} \\ \chi^{Jh}_{ij} & \chi^{JJ}_{ij} \end{bmatrix} \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + [\beta^h_{ij} \;\; \beta^J_{ij}] \begin{bmatrix} h^{in} \\ J^{in} \end{bmatrix} + c_{ij} \;\; \forall i \in V', j \in V' \label{eq:j_quad_model}
\end{align}
Given sufficient data, these quadratic functions can be recovered using a least squares regression. For example, the function form of the output parameter $h_i$ can be learned from a collection of $S$ different input-output realizations by solving the following convex optimization problem,
\begin{align}
\operatornamewithlimits{argmin}_{{\chi^{hh}_i}, {\chi^{hJ}_i}, {\chi^{Jh}_i}, {\chi^{JJ}_i} {\beta^h_i}, {\beta^J_i}, c} \; & \sum_{s \in S} \left( [h^{in}_s \;\; J^{in}_s] \begin{bmatrix} \chi^{hh}_i & \chi^{hJ}_i \\ \chi^{Jh}_i & \chi^{JJ}_i \end{bmatrix} \begin{bmatrix} h^{in}_s \\ J^{in}_s \end{bmatrix} + [\beta^h_i \;\; \beta^J_i] \begin{bmatrix} h^{in}_s \\ J^{in}_s \end{bmatrix} + c_i - h_{is} \right)^2 \label{eq:quad-opt}\\
& {\chi^{hh}_i}, {\chi^{hJ}_i}, {\chi^{Jh}_i}, {\chi^{JJ}_i} {\beta^h_i}, {\beta^J_i}, c \in \mathbb{R} \nonumber
\end{align}
Note that similar optimization problems can be solved to learn the quadratic relation of all output parameters $h,J$. The number of data points (i.e. $|S|$) should be at least $n \ln n$ to accurately recover the quadratic function, where $n$ is the number of unknown values in \eqref{eq:quad-opt}. The core experiment of this section consists in performing a series of 2-nd order reconstruction experiments over random input models, and then in using these pairs of input-output models to recover the coefficients in the quadratic response function.
The primary challenge of this experiment is the time required to collect a sufficient amount of data to fit the quadratic response function. To minimize the data requirements, we focus on the 4-spin model defined as $N',E'$ in Table \ref{tbl:l2k-qpu-used}. For this specific model, the quadratic functions \eqref{eq:h_quad_model}, \eqref{eq:j_quad_model} have 57 parameters; we consider 250 input-output model pairs to recover these parameters. Each of the 250 input models is selected i.i.d. from the following input parameter distribution,
\begin{align}
h^{in}_i & \in \{-0.05:0.01:0.05\} \;\; \forall i \in N' \label{eq:quadresp_field_pertubation_set},\\
J^{in}_{ij} & \in \{-0.05:0.01:0.05\} \;\; \forall i,j \in E'.\label{eq:quadresp_coupling_pertubation_set}
\end{align}
Following the validation study for the {\em strong ferromagnet} and {\em strong anti-ferromagnet} models, $4 \cdot 10^6$ samples are sufficient to accurately reconstruct 2-nd order models with parameters as large as 0.05. Altogether, this experiment reconstructs the outputs for 250 input models using 4,000,000 samples for each model, which results in a total of a \emph{billion} samples collected.
Characteristic dominant terms in the recovered quadratic response function is presented in the Fig.~4 of the Main Text (data on all measured quadratic response terms is given below, in Sections~\ref{sec:noise_and_spurious_links} and \ref{sec:three_generations}.). The zero order terms in the response functions are interpreted as residual fields and couplings; the first-order terms are related to the native couplings present in the chip; and finally, the second-order terms are responsible for the spurious couplings. In previous work, the primary hypothesis behind the response function was formulated in terms of a particular case of a linear assumption \cite{bian2010ising, benedetti2016estimation, perdomo2016determination, raymond2016global, marshall2017thermalization, li2020limitations}, where each parameter would be multiplied by a single effective temperature. The non-linearity of the general response function that we construct here may explain why this effective temperature was found to be instance-dependent: This corresponds to a linear approximation of a non-linear function.
Spurious couplings identified under a careful statistical analysis indicate that a simple linear model is not sufficient for an accurate characterization of the D-Wave's input-output relationship. It is important to note that the second-order response that we find here is different from the previously observed next-nearest-neighbour couplings in the strong input regime, where a quadratic cross-talk relation with an \emph{opposite sign} susceptibility has been suggested (see the section ``Compensation of qubit nonidealities'' in the \emph{Methods} of \cite{king2018observation}). We conjecture that the emergence of the next-nearest-neighbour couplings observed in the strong regime has a quantum nature and is due to the induced effects of the transverse field; a detailed exploration of this phenomenon is beyond the scope of our study that focuses on the classical regime of the output distribution in a multi-qubit setting.
The discovery of strong and structured quadratic response functions for the output distribution that our D-Wave \texttt{DW\_2000Q\_LANL} QPU samples is invaluable to applications such as hardware calibration, problem embedding and accurate sampling. However, identifying the root cause of these unexpected output parameters can provide valuable information about how to design better quantum annealers and can provide novel analytics for evaluating the performance of quantum annealers. To that end, the next section provides the theoretical grounds to explain the quadratic response as side effects of instantaneous qubit noise.
\section{Characterisation of the Local Field Variably}
\label{sec:local_field_variability}
In preparation for constructing a model that would explain the form of the quadratic response function, in this section we investigate possible drifts of the reconstructed model parameters. To this end, we perform reconstructions over several days, and monitor the stability of the recovered model.
The foundation of this variability study is the reconstruction of the output distribution of the zero value problem, that is $h^{in}, J^{in} = 0$ as shown in Table \ref{tbl:2o-zero-model}. Furthermore, we would like to perform this reconstruction accurately but with a minimum number of samples, so that the possible flux drift dynamics can be observed in the time between multiple reconstructions. We begin by calibrating a 2-body reconstruction specifically for the zero-value problem by proposing that only 200,000 samples are required for an accurate reconstruction, which is 20 times less data than what is used for a typical 2-body reconstruction. Repeating the previous reconstruction variability analysis, Fig.~\ref{fig:2o-zero-parameters} presents both the variance and recovery accuracy of this zero-valued model. The reconstruction accuracy is approximately 0.007, which is about two times less accurate than the previously considered reconstruction experiments. However, we find this accuracy is still acceptable as our primary interest in this experiment are model values that are above 0.100. Indeed, Fig.~\ref{fig:2o-zero-parameters} indicates that the \texttt{DW\_2000Q\_LANL} QPU has a number of biases that are on the order of magnitude larger than reconstruction accuracy threshold of 0.008.
\begin{table}
\centering
\begin{tabular}{|r||r|r|r|r|r|}
\hline
Model & Qubits & Couplers & $h^{in}$ & $J^{in}$ & Samples \\
\hline
\hline
Zero-value & $V'$ & $E'$ & 0.0 & 0.000 & 200,000 \\
\hline
\end{tabular}
\caption{The input parameters sent to the \texttt{DW\_2000Q\_LANL} QPU to collect data for the 2-nd order reconstruction experiment.}
\label{tbl:2o-zero-model}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.41\linewidth,page=2]{results/multibody/zeros_1x1_l2k/00000_00000_base_reconst_resample_delta_hist.pdf}
\includegraphics[width=0.41\linewidth,page=4]{results/multibody/zeros_1x1_l2k/00000_00000_base_reconst_hist.pdf}
\caption{The reconstruction variability (left) and recovered interaction magnitudes (right) of the zero problem. The results suggest indicate that 200,000 samples is sufficient to accurately measure the qubit bias occurring in the \texttt{DW\_2000Q\_LANL} QPU.}
\label{fig:2o-zero-parameters}
\end{figure}
Next, we investigate how the reconstructed values of the zero problem change over time. The objective is to understand how the low frequency noise in the hardware changes over time and how that can impact data collection over the span of minutes to hours. In this experiment, the data for the proposed zero problem is collected at 10 minute intervals over a period of 48 hours and then the 2-body reconstruction is used to recovery a model from the observed samples. Fig.~\ref{fig:zero-overtime} shows the reconstructed values over time and whiskers around the points indicate the error bounds on the model reconstruction values. Considering the mean values of these time series one can observe that there is a persistent bias on both the reconstructed fields and couplers, which is most likely an artifact of the initial hardware's calibration. Looking at the variance of the time series highlights a high variance in the fields terms and a much lower variance in the coupler terms. Overall, the results of these experiments suggest that all parameters of the output distribution at the exception of local fields remain stable over time.
\begin{figure}
\centering
\includegraphics[width=0.66\linewidth]{results/recalibration/zero_long_l2k_timeseries.pdf}
\caption{Reconstruction of the zero problem repeated over several days. A clear persistent bias in the local fields is evident from offset mean values. The fluctuations in the reconstructed field are significantly larger than the reconstruction error, suggesting fluctuation in the local fields over time.}
\label{fig:zero-overtime}
\end{figure}
\section{Explaining Quadratic Response via Instantaneous Qubit Noise}
\label{sec:noise_and_spurious_links}
In the previous section~\ref{sec:local_field_variability}, we saw that magnetic field is affected by comparably large fluctuations. However, this analysis was conducted by performing the reconstruction over a certain time window, which may average out the fluctuations. On the other hand, the analysis in section~\ref{sec:single_spin_exp} allowed us to estimate \emph{instanteneous} fluctuations of the residual random magnetic field for each qubit individually. Here, we show that these rapid fluctuations in the individual qubit fields can be responsible for spurious effective interactions with non-trivial quadratic-type responses in the input quantities. We start by showing theoretically the type of spurious interactions one may expect to reconstruct on toy models for which we can derive close form formulas. Then, using numerical simulations, we will quantify the quadratic response caused by noise on a four qubit system and compare these predictions with the quadratic response measured on the D-Wave quantum annealer.
\subsection{Spurious Magnetic Field Response}\label{subsec:theory_spurious_fields}
We first consider a simple system consisting of two classical spins $\sigma_1,\sigma_2 \in \{-1,1\}$ linked by a coupling $J\in \mathbb{R}$. The first spin is subject to a noisy magnetic field $h^{\rm{sd}}_{1} s_1$ whose direction varies according to the uniform random variable $s_1\in \{-1,1\}$, whereas the second spin is subject to a constant magnetic field $h_2$. The Bolzmann distribution of this two spins system at inverse temperature $\beta$ for a particular noise realization $s_1$ is given by the following expression,
\begin{equation}
\mu(\sigma_1,\sigma_2 \mid s_1) = \frac{\exp{\left( \beta (J \sigma_1 \sigma_2 + h^{\rm{sd}}_{1} s_1\sigma_1 +h_2 \sigma_2 )\right)}}{Z(s_1)}, \label{eq:quadth_one_coupling_dist}
\end{equation}
where the partition function $Z(s_1)$ depends on the noise realization $s_1$. Suppose now that we want to perform an Ising model reconstruction using a collection of iid. samples from the distribution in Eq.~\eqref{eq:quadth_one_coupling_dist} where the noise realization changes randomly on a sample to sample basis. Our collection of samples end up arising from a mixture of models as around half of the configurations comes from $\mu(\sigma_1,\sigma_2 \mid s_1 = +1)$ and the other half comes from $\mu(\sigma_1,\sigma_2 \mid s_1 = -1)$. Therefore, the effective model that we can reconstruct with this heterogeneous collection of samples is the following mixture of Ising models,
\begin{equation}
\mu_{\rm{effective}}(\sigma_1,\sigma_2) = \sum_{s_1\in \{-1,1\}} \frac{1}{2} \mu(\sigma_1,\sigma_2 \mid s_1).
\end{equation}
This effective model is also an Ising model and, after some algebra, it can be explicitly formulated with respect to the initial coupling and fields,
\begin{equation}
\mu_{\rm{effective}}(\sigma_1,\sigma_2) = \frac{\exp{\left( \beta (J \sigma_1 \sigma_2 + h^{\rm{effective}}_{1} \sigma_1 +h_2 \sigma_2 )\right)}}{Z_{\rm{effective}}}, \label{eq:quadth_one_coupling_mixture}
\end{equation}
where
\begin{align}
h^{\rm{effective}}_{1}
= -\frac{1}{\beta}\arctanh{\left(\tanh{(\beta J)} \tanh{(\beta h_2)} \tanh{(\beta h^{\rm{sd}}_{1})}^2\right)}. \label{eq:quadth_one_coupling_response}
\end{align}
For small coupling and field magnitudes $J$ and $h_2$, the expression in Eq.~\eqref{eq:quadth_one_coupling_response} reduces to $h^{\rm{effective}}_{1} \approx - \beta J h_2 \tanh{(\beta h^{\rm{sd}}_{1})}^2$. We see with this toy model that fast fluctuating magnetic field noise induces an effective magnetic response. The samples coming from a mixture of models with field noise are indistinguishable from samples coming from the single model with an effective response. The key qualitative features of this response is 1) its intensity is roughly proportional to the product of the coupling and opposite (constant) field intensity, and 2) the sign of the response is negative. This effect being roughly proportional to the square of the standard deviation of the noise, it is negligible for low noise values but becomes much more pronounced when the noise becomes large.
\subsection{Spurious Coupling Response}\label{subsec:theory_spurious_couplings}
To illustrate how spurious couplings can occur from a mixture of noisy Ising models, we look at a chain of three spins $\sigma_1$, $\sigma_2$ and $\sigma_3$ connected via couplings $J_{12}$ and $J_{23}$. We assume that the spin at the extremity of the chain are subject to a noisy magnetic field $h^{\rm{sd}}_{1} s_1$ and $h^{\rm{sd}}_{3} s_3$ where the random variable $s_1, s_3 \in \{-1,1\}$ are independent and uniformly distributed. The magnetic field on the middle spin is assumed to be zero. For a particular noise realization, the Boltzmann distribution of this chain of spins is given by the following conditional distribution,
\begin{align}
\mu(\sigma_1,\sigma_2, \sigma_3 \mid s_1, s_3)
= \frac{\exp{\left( \beta (J_{12} \sigma_1 \sigma_2 + J_{23} \sigma_2 \sigma_3 + h^{\rm{sd}}_{1} s_1\sigma_1 + h^{\rm{sd}}_{3} s_3\sigma_3\right)}}{Z(s_1,s_3)}, \label{eq:quadth_two_coupling_dist}
\end{align}
where the partition function $Z(s_1,s_3)$ is a function of the noise realization. Similarly to the previous subsection, we consider fast fluctuating noise that changes randomly on a sample to sample basis. In this case, a collection of samples coming from the mixture of noisy Ising models described by Eq.~\eqref{eq:quadth_two_coupling_dist} becomes indistinguishable from iid. samples coming from the effective model,
\begin{equation}
\mu_{\rm{effective}}(\sigma_1,\sigma_2,\sigma_3) = \sum_{s_1,s_3\in \{-1,1\}} \frac{1}{4} \mu(\sigma_1,\sigma_2,\sigma_3 \mid s_1,s_3).
\end{equation}
This effective model appears to be an Ising model as well with no magnetic fields and with an additional coupling between $\sigma_1$ and $\sigma_3$,
\begin{align}
\mu_{\rm{effective}}(\sigma_1,\sigma_2, \sigma_3)
= \frac{\exp{\left( \beta (J_{12} \sigma_1 \sigma_2 + J_{23} \sigma_2 \sigma_3 + J^{\rm{effective}}_{13} \sigma_1 \sigma_3) \right)}}{Z_{\rm{effective}}}, \label{eq:quadth_two_coupling_mixture}
\end{align}
where the effective coupling $J^{\rm{effective}}_{13}$ can be explicitly written with respect to the mixture parameters,
\begin{equation}
J^{\rm{effective}}_{13} = -\frac{1}{\beta}\arctanh{\left(\tanh{(\beta J_{12})} \tanh{(\beta J_{23})} \tanh{(\beta h^{\rm{sd}}_{1})}^2 \tanh{(\beta h^{\rm{sd}}_{3})}^2\right)}. \label{eq:quadth_two_coupling_response}
\end{equation}
For small coupling values, Eq.~\eqref{eq:quadth_two_coupling_response} reduces to $J^{\rm{effective}}_{13} \approx -\beta J_{12} J_{23} \tanh{(\beta h^{\rm{sd}}_{1})}^2 \tanh{(\beta h^{\rm{sd}}_{3})}^2$. We immediately see that this coupling response induced by field noise retains the main qualitative features observed in the previous subsection. The intensity of the response is quadratic in the couplings $J_{13}$ and $J_{23}$ for it is proportional to their product and the sign of the response is negative. Note that this coupling response, which involves three spins, is predicted to be weaker than the magnetic field response discussed in the previous subsection as it is roughly proportional to the square of both noise standard deviations $h^{\rm{sd}}_{1}$ and $h^{\rm{sd}}_{3}$.
\subsection{Simulations and Predictions Using Single Spin Measurements}
In the previous subsections, we have seen on simple toy models that field noise leads to effective field and spurious coupling responses. For small coupling and field magnitudes, these ``spurious" responses were mainly quadratic in the input parameters. We now want to quantify the quadratic responses cause by field noise on a realistic four spin system and compare them to the type of quadratic responses found experimentally in Section~\ref{sec:quadratic_response}. This four spin system being already too complex to obtain a closed form formula, we resort to using numerical simulations to extract the quadratic response coefficients. We model the system by a classical Boltzmann distribution conditioned on a noise realization $\underline{s}\in \{-1,1\}^4$ of the field noise parameters. The probability to obtain a configuration $\underline{\sigma}\in\{-1,1\}^4$ given $\underline{s}$ reads as follows,
\begin{equation}
\mu(\underline{\sigma} \mid \underline{s}) = \frac{\exp{\left(H_{\rm{fields}}\left(\underline{\sigma} \mid \underline{s}\right) +H_{\rm{couplings}}\left(\underline{\sigma}\right)\right)}}{Z(\underline{s})},\label{eq:quadth_four_spins}
\end{equation}
where the partition functions $Z(\underline{s})$ is noise dependant. The Hamiltonian describing the magnetic field interaction contains terms $h_i$ for input fields, $h^{\rm{bias}}$ for permanent biases and $h^{\rm{sd}}_{i}$ for the standard deviation of the noise as described in Section~\ref{sec:single_spin_exp}. An individual temperature $\beta_i$ is also assigned for each spin,
\begin{align}
H_{\rm{fields}}\left(\underline{\sigma} \mid \underline{s}\right) &= \sum_{i=1,\ldots,4}\beta_{i}(h^{\rm{sd}}_{i} s_i + h^{\rm{bias}}_{i} + h_{i})\sigma_{i}.
\end{align}
The coupling Hamiltonian contains terms $J_{ij}$ for the input coupling strengths that are only along physical couplers and possessing their individual temperatures $\beta_{ij}$. Motivated by considerations from Section~\ref{sec:local_field_variability}, we assume that the interactions are noiseless and without biases,
\begin{align}
H_{\rm{couplings}}\left(\underline{\sigma}\right)
= \beta_{12} J_{12} \sigma_1 \sigma_2 + \beta_{14}J_{14} \sigma_1 \sigma_4 + \beta_{23}J_{23} \sigma_2 \sigma_3 +\beta_{34}J_{34} \sigma_3 \sigma_4.
\end{align}
The effective model describing the probability distribution of the four spin system is obtained after averaging Eq.~\eqref{eq:quadth_four_spins} over the uniform and independant noise realizations,
\begin{align}
\mu_{\rm{effective}}\left(\underline{\sigma}\right) = \frac{1}{4}\sum_{\underline{s}\in\{-1,1\}^4} \mu(\underline{\sigma} \mid \underline{s}).
\label{eq:quadth_four_effective_model}
\end{align}
The numerical procedure to reconstruct from Eq.~\eqref{eq:quadth_four_effective_model} a quadratic response as a function of the input couplings $J_{ij}$ and fields $h_{i}$ is reminiscent of the experimental protocol described in Sections~\ref{sec:quadratic_response}.
We start by randomly selecting 20000 input coupling and field configurations whose values lies in the set $\{-0.05, -0.04, \cdots, 0.05\}$, see Eq.~\eqref{eq:quadresp_field_pertubation_set} and Eq.~\eqref{eq:quadresp_coupling_pertubation_set}. Then for each of these configurations, we compute numerically the effective frequencies of the $2^4=16$ spin configurations using Eq.~\eqref{eq:quadth_four_effective_model} and summing over the $2^4=16$ possible noise realizations. These frequencies are used in our reconstruction procedure, described in Section~\ref{sec:learning_general_distributions}, to infer an effective Ising model with Hamiltonian,
\begin{align}
H_{\rm{effective}}(\underline{\sigma}) = \hspace{-0.39cm} \sum_{i,j\in \{1,2,3,4\}}J^{\rm{effective}}_{ij} \sigma_i \sigma_j + \hspace{-0.39cm} \sum_{i\in\{1,2,3,4\}} h^{\rm{effective}}_i \sigma_i.\label{eq:quadth_effective_hamiltonian}
\end{align}
Note that the Hamiltonian in Eq.~\eqref{eq:quadth_effective_hamiltonian} contains all pairwise interactions between four spins and the spurious interactions are represented by the effective couplings $J^{\rm{effective}}_{13}$ and ${J^{\rm{effective}}_{24}}$. Finally, we fit a quadratic response model between the inputs configurations and their corresponding inferred effective couplings, as described by Eq.~\eqref{eq:h_quad_model} and Eq.~\eqref{eq:j_quad_model}, following the optimization procedure in Eq.~\eqref{eq:quad-opt}.
The spins $\sigma_1, \sigma_2, \sigma_3, \sigma_4$ in our model are identified with the hardware spins $\#304, \#308, \#305, \#309$ respectively. The values of field temperatures $\beta_i$, field biases $h^{\rm{bias}}_i$ and field noise standard deviations $h^{\rm{sd}}_i$ are chosen to be those measured using the single spin quantum experiments described in Section~\ref{sec:single_spin_exp}. These values can be found in the last column of Table~\ref{tbl:classical_reg_coeff}. The values of the coupling temperatures have been adjusted such that the simulated and measured effective temperatures coincide, i.e. $\beta_{12}=12.1, \beta_{14}=12.2, \beta_{23}=12.5$ and $\beta_{24}=12.6$.
\begin{figure*}
\centering
\includegraphics{figures/Supp_mat_full_response_simulated_vs_true.pdf}
\caption{General quadratic response of effective output parameters. A comparison of the simulated and true responses at the quadratic level are given for the effective parameters $J^{\rm{out}}_{304,305}$, $J^{\rm{out}}_{308,309}$, and $h^{\rm{out}}_{309}$. The dominant terms in the general quadratic response presented here are discussed in the Fig.~4 of the Main Text. Notice that here and below, we use a symmetric matrix representation for the quadratic response, which results in a factor 2 difference for matrix elements compared to the results presented in the Fig.~4 of the Main Text.}
\label{fig:full_response_simulated_vs_true}
\end{figure*}
The typical simulated and measured response for existing couplings, fields, and spurious couplings are depicted in
the Fig.~\ref{fig:full_response_simulated_vs_true}. We see that the patterns are qualitatively following the predictions from the simple theoretical models: The response for existing couplings is a linear self-transform or effective temperature, the response for fields consists in an effective temperature and a negative quadratic response from adjacent couplings and connecting neighboring fields, and finally the spurious couplings are formed by a negative quadratic response from adjacent couplings that span a triangle with the spurious coupling. The main noticeable difference with the theoretical predictions is the lowering of the coupling effective temperatures from its model temperature due to the presence of field noise. The comparison between the measured and simulated effective temperature can be found in Table~\ref{tbl:quadth_lin_response}. The leading off-diagonal coefficients of the quadratic response, both simulated and measured, are displayed in Table~\ref{tbl:quadth_quad_response}. The effective temperatures found through the simulation remarkably matches the effective temperatures measured in the hardware with a maximum difference of at most $7\%$. The negative sign of the quadratic response and the type of interactions involved is in perfect agreement with the theoretical model. There exists a discrepancy in the magnitudes predicted by the simulation and those found experimentally. The predictions are up to two times weaker for the susceptibility of the fields and up to four times weaker for the spurious couplings. This can be explained by the strength of the noise induced response as our theoretical models predicts that the field response is a second order effect in the noise parameters, see Subsection~\ref{subsec:theory_spurious_fields}, and the spurious links response is a fourth order effect in the field noise intensity, see Subsection~\ref{subsec:theory_spurious_couplings}. Therefore, only a $40\%$ difference in the single spin noise standard deviation can explain such differences. Note that the the quadratic response of the spurious coupling $J_{304,305}$ being weaker that the spurious coupling $J_{308,309}$ is correctly predicted by the simulation and with a similar ratio.
\begin{table}
\centering
\begin{tabular}{|c||c|c|}
\hline
\multicolumn{3}{|c|}{Linear Self-Response}\\
\hline
Output/Input & Simulated & Measured\\
\hline
\hline
$h_{304}$ & $11.1$ & $10.5$ \\
\hline
$h_{305}$ & $10.4$& $10.1$ \\
\hline
$h_{308}$ & $11.2$& $10.5$ \\
\hline
$h_{309}$ & $9.4$& $9.4$ \\
\hline
$J_{304,308}$ & $12.0$ & $12.0$ \\
\hline
$J_{304,309}$ & $11.9$ & $11.9$ \\
\hline
$J_{305,308}$ & $12.2$ & $12.2$ \\
\hline
$J_{305,309}$ & $12.0$ & $12.0$ \\
\hline
\end{tabular}
\caption{Effective simulated and measured temperatures for existing couplings and magnetic fields. The individual coupling temperatures have been adjusted in the model such that the effective simulated temperatures coincide with the effective measure temperatures.}
\label{tbl:quadth_lin_response}
\end{table}
\begin{table}
\centering
\begin{tabular}{ |c||c|c|c| }
\hline
\multicolumn{4}{|c|}{Main Quadratic Response}\\
\hline
Output& Input & Simulated & Measured \\
\hline
\hline
$h_{304}$ & $h_{308}, J_{304,308}$ & -4.9 & -10.5\\
& $h_{309}, J_{304,309}$ & -3.8 & -7.4 \\
\hline
$h_{305}$ & $h_{308}, J_{305,308}$ & -10.7 & -9.2\\
& $h_{309}, J_{305,309}$ & -8.2 & -11.6\\
\hline
$h_{308}$ & $h_{304}, J_{304,308}$ & -6.5 & -12.6\\
& $h_{305}, J_{305,308}$ & -5.9 & -11.1\\
\hline
$h_{309}$ & $h_{304}, J_{304,309}$ & -13.0 & -11.0\\
& $h_{305}, J_{305,309}$ & -11.7 & -11.9\\
\hline
$J_{304,305}$ & $J_{304,308}, J_{305,308}$ & -0.9 & -4.1\\
& $J_{304,309}, J_{305,309}$ & -0.7 & -4.0\\
\hline
$J_{308,309}$ & $J_{304,308}, J_{304,309}$ & -1.5 & -5.7\\
& $J_{305,308}, J_{305,309}$ & -1.3 & -6.8\\
\hline
\end{tabular}
\caption{Main simulated and measured components of the quadratic response. The leading terms are off-diagonal and correspond to an adjacent couplings and connected field for magnetic fields and correspond to two adjacent couplings forming a triangle for spurious links.}
\label{tbl:quadth_quad_response}
\end{table}
\section{Impacts of Spin Reversal Transformations}
\label{sec:effect_of_spin_reversals}
\subsection{Theoretical Considerations}
We are looking at distributions of Ising models on $N$ spins $\underline{\sigma} = \{\sigma_1,\ldots\,\sigma_{N}\}$ depending on input couplings $\underline{J}$ and magnetic fields $\underline{h}$. If we consider that the fields are potentially noisy and there exists individual residual fields $b_i$, random or deterministic, the Boltzmann distribution takes the following form,
\begin{align}
\mu(\underline{\sigma}\mid \underline{J}, \underline{h}) = \frac{\exp{\left(\sum_{ij} \beta_{ij} J_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i (h_i + b_i) \sigma_i\right)}}{Z},\label{eq:srt_ising_model}
\end{align}
where $\beta_{ij}$ and $\beta_i$ are effective individual temperatures for couplings and fields respectively. The residual fields may have potentially strong undesirable effects such as favoring particular spin configurations among others that were initially designed to be equiprobable. There exists a heuristical method that aims at mitigating this problem called the spin reversal transform (SRT). This method consists at looking at $2^N$ possible remapping of the model, each of them indexed by a ``gauge" which is a binary configuration $\tau = \{-1,1\}^N$. For a given configuration $\tau$, this gauge transform maps a spin configuration, input couplings and input fields to the values $\underline{\sigma}^\tau$, $\underline{h}^\tau$ and $\underline{J}^\tau$ in the following way,
\begin{align}
\sigma^{\tau}_i = \sigma_i \tau_i, \quad h^{\tau}_{i} = h_i \tau_i, \quad J^{\tau}_{ij}= J_{ij} \tau_i \tau_j.\label{eq:srt_gauge_transform}
\end{align}
The particularity of the transformation~\eqref{eq:srt_gauge_transform} is that it creates an equivalence relationship between Hamiltonians \emph{without} residual fields as for any gauge $\tau$,
\begin{align}
\sum_{ij} \beta_{ij} J^{\tau}_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i h^{\tau}_i \sigma_i = \sum_{ij} \beta_{ij} J_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i h_i \sigma_i.
\end{align}
For system without biases, it implies that samples generated from any set of gauge transformed inputs $\underline{h}^\tau$ and $\underline{J}^\tau$ are identical after a remapping of the samples using the same gauge $\underline{\sigma}^\tau$. The SRT method consists in generating samples from a mixtures of randomly selected gauge transformed models with residual fields, and which are therefore no longer equivalent, in order to empirically average over the residual field values. The effective model describing this mixture is given by the average of the Bolztmann distribution \eqref{eq:srt_ising_model} over all possible gauge transforms and reads,
\begin{align}
\mu^{\rm{effective}}(\underline{\sigma}\mid \underline{J}, \underline{h}) = \frac{1}{2^{N}} \sum_{\tau \in \{-1,1\}^N} \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}),
\end{align}
where the gauge transformed models is explicitly expressed as follows,
\begin{align}
\mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau})
= \frac{\exp{\left(\sum_{ij} \beta_{ij} J_{ij} \sigma_i \sigma_j + \sum_{i} \beta_i (h_i + \tau_i b_i) \sigma_i\right)}}{Z^{\tau}}.\label{eq:srt_gauge_model}
\end{align}
Note that the partition function in Eq.~\eqref{eq:srt_gauge_model} depends on the gauge transforms through the residual field values. We see with Eq.~\eqref{eq:srt_gauge_model} that the sign of the residual fields are effectively randomly flipped with the SRT method. Thus, the SRT removes the undesirable effects of permanent residual fields but transforms it into magnetic field noise with other potentially unwanted effects such as lower effective temperature, spurious links and field quadratic response described in Section~\ref{sec:noise_and_spurious_links}. For the D-wave hardware where the fields noise is more important compared to permanent biases, see Section~\ref{sec:single_spin_exp}, the SRT removes permanent biases for only a limited increase in the noise. Therefore, the Ising model reconstructions appears less fluctuating over time with SRT than without.
To illustrate this last point, consider an Ising model with permanent residual fields $b_i$ and no input fields i.e. $\underline{h} \equiv 0$. In this model, the average of value of the spins are non-zero in general and are a non trivial function of the residual fields and couplings. However, in the effective model produced by the SRT $\mu^{\rm{effective}}(\underline{\sigma}\mid \underline{J}, 0)$, the average value of each spin is identically zero. This implies that the effective model has zero effective magnetic field regardless of the value of the residual biases. To see this, we first note that when $\underline{h} = 0$, the partition function of a gauge transformation is invariant under a global sign change $Z^{\tau} = Z^{ - \tau}$. This further implies the following equivalence between two probabilities of gauge transformed spin configurations $\mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau},0) = \mu(\underline{\sigma}^{-\tau}\mid \underline{J}^{-\tau},0)$ for all $\tau$. Since $\underline{\sigma}^{-\tau} = \underline{\sigma}^{\tau}$, it shows that the average value of any spin $\sigma_u$ vanishes as,
\begin{align}
\mathbb{E}_{\rm{SRT}}\left[\sigma_u\right] &= \sum_{\underline{\sigma}\in\{-1,1\}^N} \sigma_u \mu^{\rm{effective}}(\underline{\sigma}\mid \underline{J}, \underline{h}) \nonumber\\
&= \frac{1}{2^{N}} \sum_{\underline{\sigma}\in\{-1,1\}^N} \sigma_u \sum_{\tau \in \{-1,1\}^N} \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}),\nonumber\\
&= \frac{1}{2^{N}} \sum_{\underline{\sigma}\in\{-1,1\}^N} \sigma_u \sum_{\tau \in \{-1,1\}^N} \frac{1}{2}\left(\mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}) + \mu(\underline{\sigma}^{-\tau}\mid \underline{J}^{-\tau}, \underline{h}^{-\tau})\right),\nonumber\\
&= \frac{1}{2^{N}} \sum_{\tau \in \{-1,1\}^N} \sum_{\underline{\sigma}\in\{-1,1\}^N}\frac{1}{2}\left(\sigma_u \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}) + \sigma_u \mu(\underline{\sigma}^{-\tau}\mid \underline{J}^{-\tau}, \underline{h}^{-\tau})\right),\nonumber\\
&= \frac{1}{2^{N}} \sum_{\tau \in \{-1,1\}^N} \sum_{\underline{\sigma}\in\{-1,1\}^N}\frac{1}{2}\left(\sigma_u \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau}) - \sigma_u \mu(\underline{\sigma}^{\tau}\mid \underline{J}^{\tau}, \underline{h}^{\tau})\right) =0.
\end{align}
\subsection{Mitigating Persistent Bias and Flux Drift}
Section \ref{sec:local_field_variability} highlighted persistent biases that the hardware exhibits.
A useful feature of the spin reversal symmetry group is that combining data collected from symmetric models has the effect of averaging out persistent biases.
\begin{figure*}
\centering
\includegraphics{figures/Supp_mat_biases_SRT.pdf}
\caption{Illustration of the impact of spin reversal transforms on the persistent biases and flux drifts. Reconstruction of the parameters of the zero input problem was repeated over several hours. Reconstructed output parameters are not exactly centered around zero, and instead are fluctuating around values that can be interpreted as persistent biases. Local fields present much larger fluctuations compared to both native couplers and spurious couplings, that are significantly larger than the reconstruction error, a signature of the instantaneous qubit noise averaged over the reconstruction period. The left plot features normal raw data, and the right plot utilizes data obtained with random spin-reversal transformations applied to the input problems, which to a large extent eliminates the impact of persistent bias and reduces the variance of fluctuations.}
\label{fig:biases_SRT}
\end{figure*}
This property is highlighted by replicating the field variability experiment with and without D-Wave's spin reversal transform feature.
In this revised experiment, the data for the zero problem is collected at 1.5 minute intervals over a period of 2 hours and then the 2-body reconstruction is used to recovery a model from the observed samples.
This experiment is conducted with two settings, {\em raw data} using this work's standard setting of \texttt{num\_spin\_reversal\_transforms = 0} and spin reversal transform data using \texttt{num\_spin\_reversal\_transforms = 10} (this setting results in a total of 200 transforms across the 200,000 samples collected for the zero-problem).
Fig.~\ref{fig:biases_SRT} shows the reconstructed values over time.
Table \ref{tbl:slow-field-variation} presents the mean and variance of each of the field values in these time series.
These results highlight how the spin reversal transforms can provide a drastic mitigation of the hardware's persistent bias.
An unexpected result from conducing these spin reversal transforms is a notable reduction in the variance of the field values.
Although the root cause for this reduction is not clear, we hypothesize that it is a side effect of an increased number of QPU programming cycles, which present another source of biases during the data collection process.
In any case, the notable bias mitigating impacts of spin reversal transforms can have a significant positive impact on applications where the user would like an unbiased output distribution.
\begin{table}
\centering
\begin{tabular}{|r||r|r||r|r|}
\hline
& \multicolumn{2}{c||}{Raw Data} & \multicolumn{2}{c|}{Spin Reversals} \\
\hline
Spin & $\mu(h)$ & $\sigma(h)$ & $\mu(h)$ & $\sigma(h)$ \\
\hline
\hline
304 & -0.241 & 0.025 & -0.001 & 0.019 \\
\hline
305 & -0.060 & 0.024 & 0.000 & 0.006 \\
\hline
308 & 0.121 & 0.025 & 0.000 & 0.010 \\
\hline
309 & -0.033 & 0.025 & 0.001 & 0.007 \\
\hline
\end{tabular}
%
\begin{tabular}{|r||r|r||r|r|}
\hline
& \multicolumn{2}{c||}{Raw Data} & \multicolumn{2}{c|}{Spin Reversals} \\
\hline
Coupler & $\mu(J)$ & $\sigma(J)$ & $\mu(J)$ & $\sigma(J)$ \\
\hline
\hline
304-308 & 0.0215 & 0.0024 & -0.0004 & 0.0030 \\
\hline
304-309 & -0.0076 & 0.0028 & -0.0007 & 0.0029 \\
\hline
305-308 & 0.0357 & 0.0027 & 0.0002 & 0.0035 \\
\hline
305-309 & -0.0005 & 0.0026 & 0.0000 & 0.0026 \\
\hline
\end{tabular}
%
\caption{A comparison of the mean and variance of reconstructed field values over time with and without spin reversal transforms during data collection. The use of spin reversal transforms corrects for persistent bias and also reduces the variance of reconstructed values.}
\label{tbl:slow-field-variation}
\end{table}
\subsection{Model Reconstruction with Spin Reversal Transforms}
Given the potential bias mitigating benefits of spin reversal transforms, it is natural to inquire how this feature impacts the results presented thus far.
We begin by reviewing the two-body reconstruction results presented in Fig. ~\ref{fig:2o-strong-parameter-srt-values}, which provides a side-by-side comparison of the results of Strong Ferromagnet model from Table \ref{tbl:2o-models} with and without spin reversal transforms. As the first observation, we notice that there is a notable change in the results of the zero-order terms. In the case with spin reversal transforms the zero-order terms are near zero, while statically significant non-zero values are exhibited in the raw data. The second observation is that the first-order terms do not show a notable change; in fact, these two reconstructions are remarkably consistent with and without spin reversal transforms. Both of these results indicate an absence of detrimental artifacts from utilizing this feature during data collection at the time scales that these experiments require.
\begin{figure}
\centering
\includegraphics[width=0.39\linewidth,page=5]{results/multibody/fms_1x1_l2k/00500_00000_base_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=5]{results/multibody/fms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/fms_1x1_l2k/00500_00000_base_reconst_hist.pdf}
\includegraphics[width=0.39\linewidth,page=6]{results/multibody/fms_1x1_l2k/00500_00000_srt_reconst_hist.pdf}
\caption{The first-order (top) and second-order (bottom) terms of the two-body reconstruction without (left column) and with (right column) spin reversal transforms for the Strong Ferromagnet model.}
\label{fig:2o-strong-parameter-srt-values}
\end{figure}
\subsection{Quadratic Response with Spin Reversal Transforms}
Figs.~\ref{fig:quad_comparison_part1} and \ref{fig:quad_comparison_part2} replicate the quadratic response experiment from Section \ref{sec:quadratic_response} with and without spin reversal transforms. This comparison shows that the overall consistency of the quadratic response picture. In accordance with the theoretical predictions outlined in the beginning of this section, we observe that the persistent bias essentially disappears under the SRT setting, and the overall response becomes much cleaner.
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_srt_part1.pdf}
\caption{A comparison of the key motifs of the quadratic response function (here, for the chimera couplings) without (left) and with (right) spin reversal transforms. The motifs are largely similar, however one can notice a considerable reduction in apparent noise in the quadratic response of the first-order terms. We hypothesize this is due mitigation of the flux qubit drift that occurs throughout the many hours of data collection required by this analysis.}
\label{fig:quad_comparison_part1}
\end{figure*}
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_srt_part2.pdf}
\caption{A comparison of the key motifs of the quadratic response function (here, for the local fields and spurious couplings) without (left) and with (right) spin reversal transforms. The motifs are largely similar, however one can notice a considerable reduction in apparent noise in the quadratic response of the first-order terms. We hypothesize this is due mitigation of the flux qubit drift that occurs throughout the many hours of data collection required by this analysis.}
\label{fig:quad_comparison_part2}
\end{figure*}
\section{Tests Across Three Generations of Quantum Annealers and Validation of the Instantaneous Noise Model}
\label{sec:three_generations}
The previous sections have argued that the quadratic response model is a good valuable tool for characterizing the input-output behavior of a quantum annealer. Furthermore, the strength of spurious links can provide an indirect measurement of the instanteneous qubit noise that is occurring on a specific hardware device. We had the opportunity to collect data for the response analysis on 7 distinct QPUs spanning three generations of quantum annealing hardware. The specific device names and components used in the experiment are presented in Table \ref{tbl:other-qpus-used}.
\begin{table*}
\centering
\begin{tabular}{|l||l|l|}
\hline
QPU & $V'$ & $E'$ \\
\hline
\hline
\multicolumn{3}{|c|}{2X} \\
\hline
LANL's \texttt{DW2X} & \{880, 881, 884, 885\} & \{(880, 884), (880, 885), (881, 884), (881, 885)\} \\
\hline
\hline
\multicolumn{3}{|c|}{2000Q} \\
\hline
\hline
\texttt{DW\_2000Q\_LANL} & \{304, 305, 308, 309\} & \{(304, 308), (304, 309), (305, 308), (305, 309)\} \\
\hline
NASA's \texttt{C16} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\texttt{DW\_2000Q\_1} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\texttt{DW\_2000Q\_2} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\texttt{DW\_2000Q\_3} & \{1800, 1801, 1804, 1805\} & \{(1800, 1804), (1800, 1805), (1801, 1804), (1801, 1805)\} \\
\hline
\hline
\multicolumn{3}{|c|}{Lower-Noise 2000Q} \\
\hline
\texttt{DW\_2000Q\_5} & \{1688, 1689, 1692, 1693\} & \{(1688, 1692), (1688, 1693), (1689, 1692), (1689, 1693)\} \\
\hline
\texttt{DW\_2000Q\_6} & \{1688, 1689, 1692, 1693\} & \{(1688, 1692), (1688, 1693), (1689, 1692), (1689, 1693)\} \\
\hline
\end{tabular}
\caption{The 4 qubits and 4 couplers used in the quadratic response experiments on a variety of QPUs. The experiments span three generations of D-Wave hardware, the 2X (2015), 2000Q (2017), and Lower-Noise 2000Q (2019).}
\label{tbl:other-qpus-used}
\end{table*}
Figs.~\ref{fig:quad_response_2x}-\ref{fig:quad_response_2k6} present the quadratic response motifs from all of the QPUs that have been tested. Overall, the results are remarkably similar, which suggests the universality of the quadratic response characterization that is capturing fundamental properties of D-Wave's quantum annealing implementation, like effective temperature, persistent biases, and instantaneous noise in the local fields. With the exception of the new {\em lower-noise} QPU, the existence and strength of the spurious links is consistent across hardware realizations. We have observed that some QPU implementations feature asymmetric spurious links while others are symmetric. Identifying the root-cause of this distinction is an ongoing point of investigation.
A significant difference in the response function has been obtained for the \emph{lower-noise} version of the D-Wave 2000Q annealer \cite{D-Wave_low_noise}. We observe a drastic reduction of the susceptibility responsible for the strength of the spurious couplings, while the linear scale terms remain on par with other 2000Q implementations, see Fig.~\ref{fig:quad_spurious_terms_low_noise}. This result therefore provides strong evidence in support of the noise-based model introduced in this work. We anticipate that the quantitative measurement of the susceptibility associated with the spurious couplings using methods developed in this work will provide a valuable characterization of the qubit noise in the next generations \cite{D-Wave_pegasus} of quantum annealers and other analog machines with binary output statistics.
\section{Open-sourced Tools}
\label{sec:Algortihmic_Tools}
The experiments conducted in this work require the collection of billions of samples from D-Wave's quantum annealer and reconstruction of graphical models with multi-body interactions.
However, neither of these tasks is readily supported by established software and the following software was developed and released as open-source to support this work.
The first software is the D-Wave Ising Sample Collector (DWISC, \url{github.com/lanl-ansi/dwisc}), which enables the collection of millions runs on D-Wave hardware by orchestrating a series of jobs that conform to D-Wave's single-job run time limit of three seconds.
The second software is GraphicalModelLearning (GML, \url{github.com/lanl-ansi/GraphicalModelLearning.jl}), which takes empirical state distributions and reconstructs effective multi-body graphical models in a factor graph representation, leveraging the Interaction Screening method described in Section~\ref{sec:learning_general_distributions} and state-of-the-art second-order nonlinear optimization algorithms to provide model reconstructions that require the least amount of data. The notable improvement of reconstruction accuracy of interaction screening framework over established approaches, such as mean-field, is discussed at length in \cite{lokhov2018optimal}. DWISC and GML form the foundation of the experiments in this work by providing the data and algorithms required to reconstruct high-accuracy multi-body models of the output from D-Wave's quantum annealer.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2x.pdf}
\caption{Quadratic response experiment results for the LANL's \texttt{DW2X} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2x}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_l2k.pdf}
\caption{Quadratic response experiment results for \texttt{DW\_2000Q\_LANL} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_l2k}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_n2k.pdf}
\caption{Quadratic response experiment results for the NASA's \texttt{C16} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_n2k}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k.pdf}
\caption{Quadratic response experiment results for the \texttt{DW\_2000Q\_1} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k2.pdf}
\caption{Quadratic response experiment results for the \texttt{DW\_2000Q\_2} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k3.pdf}
\caption{Quadratic response experiment results for the \texttt{DW\_2000Q\_3} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k3}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k5.pdf}
\caption{Quadratic response experiment results for the lower-noise \texttt{DW\_2000Q\_5} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k5}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{fig_quad_final/fig_quad_2k6.pdf}
\caption{Quadratic response experiment results for the lower-noise \texttt{DW\_2000Q\_6} chip: (a) Chimera couplings; (b) Local Fields; (c) Spurious couplings.}
\label{fig:quad_response_2k6}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.88\linewidth]{fig_quad_final/fig_quad_low_noise.pdf}
\caption{Heat maps representing the quadratic terms of the quadratic response function for the spurious link output parameters on a regular (left) and lower-noise (right) 2000Q quantum annealers. The significant reduction in the link strength in the lower-noise response (from -4.1 to -0.9 and -6.8 to -1.9) confirms the theoretical model of high-frequency qubit noise.}
\label{fig:quad_spurious_terms_low_noise}
\end{figure*}
\end{document} | e90e33fd5ebbfb1df58a36363b67a09b01ef5a73 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Gluons are Bosons. Can gluons be stacked in a same quantum state like Bose-Einstein condensate (BEC)?
How to observe this gluon condensation (GC) phenomenon? Gluons, confined in proton/neutron,
are the fundamental components of the Universe. Previous works concerning the gluon distribution in proton based on QCD have shown the possible existence of
GC [1-3]. The correlations among gluons at very high energy (VHE) may excite chaos in proton, which arises
strong shadowing and antishadowing effects, resulting in the squeeze of gluons to a state with critical
momentum. Accordingly, we found a series of likely GC signals in the VHE cosmic ray (gamma-ray,
electron/positron, proton and nuclei) spectra, the sources include supernova remnants (SNRs), active galactic nuclei (AGN) and
gamma ray bursts (GRBs) [4-7]. The present work focuses on the GC effect in gamma-ray spectra from pulsars.
Pulsars as rapidly rotating and magnetized neutron stars, can generate bubbles of relativistic
particles when their ultra-relativistic
winds interact with the surrounding medium, and the particles can be accelerated by pulsar wind nebulae (PWNe) up to PeV and beyond.
It is widely accepted that the electromagnetic spectrum of pulsars originates from leptons.
Although protons (or ions more generally) may exist in pulsars and further be accelerated,
the hadronic scenario in pulsars is negligible since their fluxes are very weak in the pulsar surroundings [8].
However, the GC effect may
largely increase the $pp$ cross section by several orders of magnitude [3], which may compensate for the weak proton flux and produce the
observed gamma rays. Therefore, we try to look for the GC signature in the
gamma-rays from pulsars.
In section 2, we give a brief introduction of the GC model and present a complete VHE gamma-ray spectrum predicted by this
model in figure 2, which shows the typical power law on both sides of the sharp turning point
at $E_{\pi}^{GC}$. On the right side of this spectrum there is an exponential cutoff from $E>E_{\pi}^{cut}$.
In sections 3, we try to identify the GC-signature in the gamma-ray spectra of pulsars, and some examples are discussed. Due to the complex structure
of the source, the recorded spectrum may come from different emission mechanisms,
we will only focus on the part that related to the GC-spectrum.
Several VHE sources are tagged as unidentified and located near other bright ones.
Although many works have speculated the possible origins of their spectra, the nature of them is still unknown.
The simple GC-spectrum with a few free parameters allows us to infer the GC signals hidden in other mixed
spectra. We give such examples in section 4. After the analysis of the above examples, the discussions and summary are presented in section 5.
\section{The GC model}
The flux of high energy gamma ray in hadronic processes $p+p\rightarrow \pi^0\rightarrow
2\gamma$ in the laboratory frame reads [4-7]
$$\Phi_{\gamma}(E)=C_{\gamma}\left(\frac{E}{1GeV}\right)^{-\beta_{\gamma}}
\int_{E_{\pi}^{min}}^{E_{\pi}^{max}}dE_{\pi}
\left(\frac{E_p}{1GeV}\right)^{-\beta_p}$$
$$\times N_{\pi}(E_p,E_{\pi})
\frac{d\omega_{\pi-\gamma}(E_{\pi},E)}{dE},
\eqno(2.1)$$ where the spectral index $\beta_{\gamma}$
denotes the propagating loss of gamma-rays near the source. The accelerated protons obey a power law
$N_p\sim E_p^{-\beta_p}$ with the index $\beta_p$.
$C_{\gamma}$ incorporates the kinematic factor with the flux
dimension and percentage of $\pi^0\rightarrow 2\gamma$. The
normalized spectrum for $\pi^0\rightarrow 2\gamma$ is
$$\frac{d\omega_{\pi-\gamma}(E_{\pi},E)}{dE}
=\frac{2}{\beta_{\pi}
E_{\pi}}H[E;\frac{1}{2}E_{\pi}(1-\beta_{\pi}),
\frac{1}{2}E_{\pi}(1+\beta_{\pi})], \eqno(2.2)$$where $\beta_{\pi}\sim 1$, and $H(x;a,b)=1$ ( $a\leq x\leq b$) or $H(x;a,b)=0$
(otherwise). Usually, the relations among
$N_{\pi}, E_{\pi}$ and $E_p$ are very complicated and determined by using the limited experimental data.
However, we show that the GC effect simplifies the above relations and gives a special energy spectrum.
Since the more gluons, the more secondary pions in the inelastic $pp$ collision, we assume
that a huge number of gluons at the central region due to
the GC effect may create the maximum number $N_{\pi}$ of pions, which take
up all available kinetic energy if we neglect the other secondary
particles. One can get the following characteristic distributions in the GeV-unit
$$\ln N_{\pi}=0.5\ln E_p+a, ~~\ln N_{\pi}=\ln E_{\pi}+b, \eqno(2.3)$$
$$~~ where~E_{\pi}
\in [E_{\pi}^{GC},E_{\pi}^{max}], $$and
$$a\equiv 0.5\ln (2m_p)-\ln m_{\pi}+\ln K, ~~b\equiv \ln (2m_p)-2\ln m_{\pi}+\ln K, \eqno(2.4)$$ $K\simeq 0.5$ is the inelasticity.
Substituting equations (2.2)-(2.4) into equation (2.1), we directly get the gamma ray spectrum,
$$E^2\Phi^{GC}_{\gamma}(E)=\left\{
\begin{array}{ll}
\frac{2e^bC_{\gamma}}{2\beta_p-1}(E_{\pi}^{GC})^3\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\beta_{\gamma}+2} \\ {\rm ~~~~~~~~~~~~~~~~~~~~~~~~if~}E\leq E_{\pi}^{GC},\\\\
\frac{2e^bC_{\gamma}}{2\beta_p-1}(E_{\pi}^{GC})^3\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\beta_{\gamma}-2\beta_p+3}
\\ {\rm~~~~~~~~~~~~~~~~~~~~~~~~ if~} E>E_{\pi}^{GC}.
\end{array} \right. .\eqno(2.5)$$Surprisingly, it shows a typical broken power law.
Figure 1 is a schematic diagram for the gluon
rapidity distribution with the GC effect at the $pp$ collision, which is taken from [3].
One can find that the GC effect
begins to work if it's peak locates at $y_{max}=\ln(\sqrt{s_{p-p}^{GC}}/\underline{k}_c)$, thus
$$x_c=\frac{\underline{k}_c}{\sqrt{s_{p-p}^{GC}}}e^{-y_{max}}\simeq \frac{\underline{k}_c^2}
{s_{p-p}^{GC}}.\eqno(2.6)$$Applying $\sqrt{s_{p-p}}=\sqrt{2m_pE_p}$ into equation (2.3), one gets
$$E_{\pi}^{GC}=\exp\left(0.5\ln\frac{\underline{k}^2_c}{2m_px_c}-(b-a)\right).\eqno(2.7)$$
On the other hand, the cross section of the $pp$ collision disappears if the peak of gluon distribution moves
to the rapidity center $y=0$, i.e., at
$$x_c=\frac{\underline{k}_c}{\sqrt{s_{p-p}^{max}}}e^{y=0}, \eqno(2.8)$$or at
$$E_{\pi}^{max}=\exp\left(0.5\ln\frac{\underline{k}^2_c}{2m_px_c^2}-(b-a)\right)=
\frac{E_{\pi}^{GC}}{\sqrt{x_c}}$$
$$=e^{b-a}\sqrt{\frac{2m_p}{\underline{k}_c^2}}(E_{\pi}^{GC})^2=
14(E_{\pi}^{GC})^2.\eqno(2.9)$$ However, figure 1 shows that the plateau is shrinking as
energy increasing rather than expanding, since lots of soft gluons enter the interaction range early
due to the condensation, which implying that the suppression of $\Phi_{\gamma}$ occurs before $E_{\pi}^{max}$. Therefore we add a cut-off factor in equation (2.5), i.e.,
$$E^2\Phi^{GC}_{\gamma}(E)=\left\{
\begin{array}{ll}
\frac{2e^bC_{\gamma}}{2\beta_p-1}(E_{\pi}^{GC})^3\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\beta_{\gamma}+2} \\ {\rm ~~~~~~~~~~~~~~~~~~~~~~~~if~}E\leq E_{\pi}^{GC},\\\\
\frac{2e^bC_{\gamma}}{2\beta_p-1}(E_{\pi}^{GC})^3\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\beta_{\gamma}-2\beta_p+3}
\\ {\rm~~~~~~~~~~~~~~~~~~~~~~~~ if~} E_{\pi}^{GC}<E<E_{\pi}^{cut},\\\\
\frac{2e^bC_{\gamma}}{2\beta_p-1}(E_{\pi}^{GC})^3\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\beta_{\gamma}-2\beta_p+3}
\exp\left(-\frac{E}{E_{\pi}^{cut}}+1\right).
\\ {\rm~~~~~~~~~~~~~~~~~~~~~~~~ if~} E\geq E_{\pi}^{cut},
\end{array} \right. \eqno(2.10)$$or
$$\Phi^{GC}_{\gamma}(E)\equiv\left\{
\begin{array}{ll}
\Phi_0\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\Gamma_1} \\ {\rm ~~~~~~~~~~~~~~~~~~~~~~~~if~}E_{\gamma}\leq E_{\pi}^{GC},\\\\
\Phi_0\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\Gamma_2}
\\ {\rm~~~~~~~~~~~~~~~~~~~~~~~~ if~} E_{\pi}^{GC}<E_{\gamma}<E_{\pi}^{cut},\\\\
\Phi_0\left(\frac{E}{E_{\pi}^{GC}}\right)^{-\Gamma_2}
\exp\left(-\frac{E}{E_{\pi}^{cut}}+1\right),\\
\\ {\rm~~~~~~~~~~~~~~~~~~~~~~~~ if~} E\geq E_{\pi}^{cut},
\end{array} \right. \eqno(2.11)$$
where
$$E_{\pi}^{cut}=\alpha E_{\pi}^{max}, \eqno(2.12)$$ we assume $\alpha=0.1$ in this work.
Note that the energy is the GeV-unit. A schematic GC-spectrum is presented in figure 2.
Although the breaks in spectrum with an exponential cutoff are ubiquitous either for the leptonic or hadronic scenarios,
the GC-spectrum (2.11) has its own features. (i) The shape of gamma-ray spectra in leptonic and traditional (without the GC effect) hadronic models sensitively depend
on the distribution of parent particles (electrons or protons), while the GC-spectrum is characterized by a few free
parameters ($\Phi_0, \Gamma_1, \Gamma_2, \alpha$ and $E_{\pi}^{GC}$), where the incalculable $pp$ interactions are
simplified by the saturation condition in equation (2.3). (ii) On both sides of the sharp breaking
point $E_{\pi}^{GC}$ the spectrum appears as the single power-law.
(iii) The GC-threshold $E_{\pi}^{GC}$ in equation (2.10) is nuclear target-dependent, which may be from $100~GeV$
to $TeV$ and beyond (see discussions in section 5). (iv) Equation (2.10)
is an analytical solution of the GC model, where the parameters have their
physical meaning and we will use them to predict the undetected
spectra in section 4.
\section{HESSs J1825-137, J1640-465, J1809-193, J1813-178 and PSR J0205+6449/3C 58}
Although a lot of data about the VHE gamma ray spectra of pulsars have been reported, however,
leptonic emission scenarios are usually favored
over hadronic emission scenarios.
Besides, astronomical radiations are complex and
different emission mechanisms may exist in a same source simultaneously. We found the following five examples in the existing data showing a complete
power-law as shown in figure 2.
HESS J1825-137 is one of the most efficient TeV gamma ray emitting PWNes [9].
It was detected by the Fermi Large Area Telescope(Fermi-LAT) in the GeV energy band,
and usually discussed in connection with leptonic acceleration scenarios [10] since it is not a prime candidate for pion decay according to the traditional hadronic scenario. We found that HESS J1825-137 with the
Fermi-LAT data in $10 ~GeV-10 ~TeV$ can also be fitted using the
GC-spectrum (see figure 3), where the
$E_{\pi}^{cut}=14~TeV$ is fixed by equation (2.12). For comparison, we present a leptonic spectrum in figure 3 (dashed curve).
HESS J1640-465 is one of the TeV sources discovered by the HESS survey in inner Galaxy [11]. It's spectrum, explained in the traditional hadronic model [12,13],
is connected to a hard Fermi-LAT gamma-ray spectrum and it forms a long flat spectrum from $100~MeV$ to $1~TeV$. However, Y.L. Xin et al
have suspected that there could not be such a strong proton flux in pulsars [14]. They
reanalyzed the relating Fermi-LAT data and found that an extended GeV gamma-ray source was coincident with J1640-465. Its photon spectrum was described by a power-law with an index of 1.42$\pm$0.19 in the energy range of $10-500~GeV$, and smoothly connected
with the TeV spectrum of HESS J1640-465. Using the same Fermi-LAT data we describe the expanded spectrum of HESS J1640-465 in the GC model, and the result is presented in figure 4 (solid curve).
HESS J1809-193 was initially discovered during the systematic search for VHE emission from pulsars in
the Galactic Plane Survey performed with HESS [15].
The environment of HESS J1809-193 is complicated, containing several different sources.
Combining with the Fermi LAT data, the spectrum presents a broad energy plateau in $300~GeV-30~TeV$, the nature of which with such a shape has been regarded as the traditional hadronic model but with a harder proton index. We take an alternate way, i.e,
assuming this broadly extended gamma-ray spectrum might be produced by the leptonic radiation and the hadronic model
with the GC effect. To begin with, we use the precise data for $dN/dE\sim E^{-\Gamma_2}$ [8] and equation (2.11) to obtain $\Gamma_2=2.22$ (see figure 5). Figure 6 is our result using the GC spectrum (2.11), the data are taken from [14].
A high energy pulsar PSR J1813-178 is classified as a PWN candidate, which has a similar spectrum of figure 2.
A new
detection of extended emission in the
range $0.5-500~ GeV$ finds that it is difficult to explain the spectrum
with inverse Compton (IC) emission from high-energy electrons alone [16].
Our predictions of GC (mixing with leptonic models)
are presented in figure 7.
3C 58, classified as a pulsar wind nebula [17], is a complex radio source with an extended flat spectrum.
PSR J0205+6449 is a pulsar near 3C 58,
it's flux is divided into two parts: off-peak and on-peak.
We connect the off-peak part of PSR J0205+6449 with the flux of 3C 58, the result can be fitted by the GC model
as shown in figure 8, and the on-peak part can be described by other leptonic model.
\section{HESSs J1826-130, J1641-463 and J1741-302}
There are several VHE sources presenting a typical power law with the hard photon index $\Gamma_{\gamma}\sim 2$
in a broad energy range from hundreds of
GeV to beyond 10 TeV. Besides, all of them are tagged as unidentified and located near other bright VHE sources.
Although many works have speculated the possible origins of these VHE gamma-ray spectra, their nature is still open.
Especially, where and how cosmic rays are accelerated to reach the necessary PeV energies and beyond in our Galaxy, and in which mechanism
the kinetic energy of the initial particles (protons or electrons) converts into photons
are still unknown. An important obstacle is that we lack the extension of
the above spectra in the GeV-energy region, and an incomplete spectrum cannot provide the correct information of these sources.
The GC-model predicts the typical broken power law, it reduces the number of free parameters
and accordingly the uncertainty. We will use the GC-model
to predict the incomplete spectra of the following three examples.
J1826-130 is an unidentified hard spectrum source discovered by HESS along the Galactic plane [18],
and was previously hidden in the extended tail of emission from the nearby bright source HESS J1825-137.
HESS J1826-130 shows a relatively hard gamma-ray spectrum, which obviously incomplete. However we can perfect
it by using the GC model. Our logical reasoning is:
(i) Using equation (2.11) at $E_{\gamma}>E_{\pi}^{GC}$ to fit the HESS J1826-130 data, one can find that
a single power law with $\Gamma_{\gamma}\sim 2$
extends up to $14~TeV$, and then appears a cut-off factor (see figure 9). According to equations (2.9) and (2.12), the
breaking energy is $E_{\pi}^{GC}=100~GeV$.
(ii) We turn to HESS J1825-137 for information related to HESS J1826-130.
Note that the parameter $\beta_p$ in the GC model is determined by the accelerate mechanism of protons. Therefore, we assume that the two sources are close to each other and
have a same value $\beta_p$. Thus, we have
$\Gamma_2=-\beta_{\gamma}+2\times 0.91-1=2 $ for HESS J1826-130. As a result a complete spectrum of
J1826-130 can be shown in figure 10 since we already know $\beta_{\gamma}=1.18$.
Why there is no recorded spectrum of HESS J1826-130 at $E_{\gamma}<100~GeV$? We noticed that a gamma-ray pulsar J1826-1256
was discovered by Fermi-LAT [19], and
it's spectrum at the GeV energies is shown
in figure 10, which may shadow the spectrum of HESS J1826-130 in the observation and leads to an incomplete spectrum.
A similar example is HESS J1641-463 [20], whose spectrum is also unclear
due to the confusion with the bright nearby source HESS J1640-465 [11]. No $X$-ray candidate stands out as a clear association. Usually
it's spectrum is explained by hadronic mechanism, where the emission
is produced by cosmic ray protons colliding with the ambient gas [21].
The data in figure 11 tells us that $\Gamma_2=2.07$, $E_{\pi}^{cut}=14~TeV$
in the GC model, and accordingly $E_{\pi}^{GC}=0.1~TeV$ by equation (2.9). We take $\beta_p=1.02$ in HESS J1641-463 as in HESS J1640-465
assuming their proton flux are originated from a same accelerator, consequently $\beta_{\gamma}=0.99$. Similarly, a complete spectrum of J1641-463 is presented in figure 12.
The hollow points are the detected pulsed emission of SNR G338.5+0.1
by Fermi-LAT [22], we add it in figure 12 for they
may cover the extension of the HESS spectrum in the GeV energies.
A preliminary detection of HESS J1741-302 was early announced by HESS [23]
and thanks to the increased amount of high quality VHE data and improved analysis techniques,
the spectrum of it can now be analyzed in detail [21]. HESS J1741-302 is not only unidentified but also has no plausible counterpart at the high-energy
($0.1 ~GeV-100~ GeV$) range,
the region around it is rather complex, as a compact radio source and a variable star
are spatially coincident with the position of HESS J1741-302.
There is only one incomplete spectrum as shown in figure 13, again using
the GC model we try to unveil its nature. From the gamma ray distribution in figure 13,
one can find that $\beta_{\gamma}+2\beta_p-1=\Gamma_2=2.3$, and $E_{\pi}^{cut}\simeq 14~TeV$ corresponds to $E_{\pi}^{GC}=0.1~TeV$. Thus, we have got $\beta_p=0.91$ and $1.02$ for J1826-130 and J1641-463, differed by 0.11. Besides,
the measured value of $\Gamma_2$ for these two sources has an error of $0.11\pm 0.20$.
Therefore, we assume that the above three VHE gamma
rays originate from the same proton accelerate mechanism in our Galaxy and we take $\beta_p\simeq 1$ for J1741-302.
Thus, a complete VHE gamma ray spectrum can be predicted in figure 14.
An investigation of Fermi LAT data has revealed a new high energy source Fermi J1740.1-3013 [24], which is $\sim 0.3^o$ offset from
the best fit position of HESS J1741-302. We presented it in figure 14 considering that the spectrum of HESS J1741-302 at GeV energies may be
covered or mixed by this new source.
\section{Discussions and Summary}
It is generally believed that the gamma rays in pulsar environment are mainly generated by leptonic mechanism
since the proton flux is weak. However, with the GC model, the
above point of view could be modified. The increment of the $pp$ cross section due to
the GC effect may compensate for the lack of proton flux and emit
gamma-rays with the GC-characteristics. On the other hand, not all the hadronic accelerators in the Galaxy can generate the GC effect, because comparing with the traditional hadronic model, the GC effect needs
higher energy of parent protons to convert into more mesons.
Therefore, the hadronic scenario with the GC effect is not as common as leptonic scenario in pulsars.
The inverse Compton emission (or curvature radiation) accompanied by synchrotron X-ray spectrum is an important
feature of leptonic model for VHE gamma-ray radiation. The GC model has no such correlation but has a
much deeper connection with other processes. In fact, the GC spectrum (2.11) can be used to explain the VHE gamma ray spectra
in supernova remnant (SNR) [5], active galactic nuclei (AGN) [6] and gamma ray burst(GRB) [6]. Besides, adding the process $\gamma
\rightarrow e^-+e^+$ or a hadronization mechanism, the same GC model can investigate anomalous excess in electron-positron spectra [5]
or hadronic fluxes [7].
A question needs to be explained is that different processes may give different parameters in
equation (2.11), which are related to the structure and nature of the source, especially, the important parameter $E_{\pi}^{GC}\simeq
400~GeV$ for Tycho's SNR while $E_{\pi}^{GC}\simeq 100~GeV$ for the sources in this work.
Why these $E_{\pi}^{GC}$ are different is because that
they are target nuclei number $A$-dependent: $E_{\pi}^{GC}(p-A)>E_{\pi}^{GC}(p-A')$ (if $A<A'$)
since the nonlinear corrections enhance as $A$ increases.
A roughly simulation of the GC model shows that
GC-threshold $E_{\pi}^{GC}$ drops rapidly from $A=1$ and then decrease slowly from the intermediate nucleus [3,4].
We
divide the targets into a few of categories and assume that on average, target nuclei near pulsar is
heavier than that in Tycho's SNR. As discussed above, the broken points $E_{\pi}^{GC}$ in the GC model
are concentrated around a few different values.
A detailed estimation requires more precise examples to be collected. The spectrum
of the investigated pulsars is compatible with being produced through the
GC effect, compensating for a weak flux of hadrons.
Therefore, as shown in figures 4, 6-8, 10, 12 and 14, the GC-spectrum is always accompanied by a leptonic spectrum, although the leptonic fit
is beyond the scope of this work. The parameters used/fitted for each source are presented in table 1.
An exception in our examples is HESS J1825-137, it extends to a long distance from the central pulsar PSR J1826-1334, where the electrons may escape from the nebula and diffuse in the interstellar medium [25].
In summary, the gluon condensation originating from a nonlinear QCD evolution equation
may greatly increase the cross sections of
proton-target interactions and arise a characteristic broken power law in the gamma-ray spectra. The result is used to explore the gluon condensation signature in the observed gamma-ray spectra from pulsars.
{\it Acknowledgments:} We thank Zhiyi Cui and Lihong Wan for useful comments.
This work is supported by the National
Natural Science of China (No.11851303). \\\\
| 2d1dcdeb61882323096c0b1d94ea41f3089aa497 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the last decades, harmonic interactions have played a crucial role in quantum dynamics to study the behavior of the moving particle subject to the electromagnetic field and scalar potential. Hence, implementing harmonic oscillators in quantum field theory in flat and curved space-time has been a prevalent scenario. Therefore, to conceive and describe harmonic oscillator in the relativistic regime, Moshinsky and Szczepaniak \cite{re1,re2} presented the well-known coupling model between a relativistic spin-half particle and the harmonic quantum oscillator in the Dirac equation. This model came up according to modifying the momentum operator by considering a nonminimal coupling. The relevant relativistic oscillator was called a Dirac oscillator \cite{CarvalhoEPJC2016,AndradeEPJC2014,re3,re17.2,re4}.\\
We know that the physical system associated with the Dirac oscillator can explained as the electric field interact with anomalous magnetic moment\cite{re4}.
In 2013, a very important experiment for the Dirac oscillator model was confirmed by Franco-Villafa\~{n}e et al. \cite{re5}. Subsequently, the relativistic fermions with spin one-half is widely employed especially in topological defect space-time. Now, let us mention some of the work done in this area, for instance, examining the influence of the Dirac oscillator in presence of the spinning cosmic string space-time \cite{re6,re7}, the fermions described by the Dirac equation interacting with the scalar and vector potentials in the topological defect background \cite{re8,re9,re10}, the Dirac oscillator in the most simple topological defect described by the cosmic string space-time under the background of the rainbow gravity \cite{re11,re12}, the non-inertial effect on the Dirac field with topological defects background \cite{re13,re14}, and the Dirac oscillator under Lorentz symmetry breaking \cite{rer1}. Meanwhile, the Dirac oscillator, generated by the $\kappa$-Poincar\'{e}-algebra, in the context of quantum deformation, is studied in Refs. \cite{AndradePLB2014-1,AndradePLB2014-2}.
Our work is motivated in examining a modified Dirac oscillator interaction related to a neutral particle interacting with the external electromagnetic field under the influences of the Aharonov-Casher (A-C) and curvature effects in $(1+2)$-dimensional cosmic string space-time. By the way, the external electromagnetic fields can generate the Aharonov-Bohm (A-B) \cite{re15} and A-C \cite{re16} effects. In particular, the effects are often non-negligible in dealing with the Dirac equation problem interacting with an external electromagnetic field in topological defects background. Accordingly, some interesting examples related to the A-B effect had been studied in connection with the fermions interacting with the localized magnetic field in topological defects \cite{re7,re17,re17.1}, and the influence of the A-C effect on Dirac theory in the different scenarios of curved space-time \cite{re18,re19}. Meanwhile, some significant works had also been done based on the extended Dirac oscillator defined by replacing a potential function $f(r)$ instead of the radial coordinate $r$ in cosmic string \cite{re20} and cosmic dislocation space-time \cite{chen1}. Other similar works include the Duffin-Kemmer-Petiau \cite{re21,re23} and Klein-Gordon oscillators \cite{re24,FF2}. Therefore, it seems interesting to investigate the generalized Dirac oscillator associated with a neutral particle having a permanent magnetic dipole moment interacting with the external electromagnetic fields in $(2+1)$-dimensional cosmic string space-time. Our main goal is to consider the Dirac oscillator with a Coulomb-type potential function.
The outline of this paper is:
In Sect. \eqref{sec2}, we investigate the influences of the A-C and curvature effects on the Dirac oscillator containing coulomb-type potential function and obtain the corresponding relativistic radial wave equation. In Sect. \eqref{sec3}, after replacing the Coulomb-type potential function in the resulting wave equation, we try to solve it and find the relativistic exact analytical solutions by applying the Nikiforov-Uvarov (NU) method. In Sect. \eqref{Conc}, we present the conclusions
\section{The generalized Dirac oscillator with the Aharono-Casher effect in cosmic string space-time \label{sec2}}
The influence of topological defects on quantum effect has been widely studied in several works \cite{re28,re29}, and the cosmic string is known as a simple example of a defect that contains curvature. Accordingly, we are motivated in investigating the influence of the A-C effect on the interaction of a Coulomb-type nonminimal coupling with the spinor field related to a neutral particle interacting with the external electric field in $(1+2)$-dimensional curved space-time. Accordingly, we begin by presenting the relevant line element as follows \cite{re30,re32} (in natural units $\hbar=c=G=1$)
\begin{equation}\label{LECosmicSt}
ds^{2}= d t^{2}-d r^{2}-\alpha^{2} r^{2} d \phi^{2},
\end{equation}
with $t\in(-\infty,\infty)$, $r\in[0,\infty)$ and $\phi\in[0,2\pi]$. The parameter $\alpha$ is associated with the linear mass density, $\tilde{m}$, of the string through the relation $\alpha=1-4\tilde{m}$ and it runs in the range $(0,1]$ as well as it related to a deficit angle via $\gamma=2\pi(1-\alpha)$ \cite{re10}. If the $\alpha$ parameter run in the range $(1,\infty)$, it corresponds to an anti-conial space-time with no-positive curvature.
Nevertheless, the cosmic string characterized by equation 1 has a conical singularity that can be demonstrated by the curvature tensor $R_{r, \phi}^{r, \phi}=\frac{1-\alpha}{4 \alpha} \delta_{2}(\mathbf{r})$, with two-dimensional Dirac delta $\delta_{2}(\mathbf{r})$. Based on the line element in Eq. \eqref{LECosmicSt}, the metric tensor $g_{\mu\nu}$ and inverse metric tensor $g^{\mu\nu}$ indicate
\begin{equation}
g_{\mu\nu}=\left(\begin{array}{cccc}
1 & 0 & 0 \\
0 & -1 & 0 \\
0 & 0 & -\alpha^{2}r^{2}
\end{array}\right), g^{\mu\nu}=\left(\begin{array}{cccc}
1 & 0 & 0 \\
0 & -1 & 0 \\
0 & 0 & -\frac{1}{\alpha^{2}r^{2}}
\end{array}\right).
\end{equation}
The covariant Dirac equation related to dynamics of a neutral spin-half particle, with the permanent magnetic dipole moment, $\tilde{\mu}$, reads \cite{re33}
\begin{equation}\label{GDirac1}
\left[i \gamma^{\mu}(x)\left(\partial_{\mu}+\Gamma_{\mu}(x)\right)+\frac{\tilde{\mu}}{2} \sigma^{\mu \nu}(x) F_{\mu \nu}-M\right] \Psi(t, \mathbf{r})=0,
\end{equation}
where $\Gamma_{\mu}(x)=\frac{1}{8} \omega_{\mu a b}(x) [\gamma^{a},\gamma^{b}]$ is named as spinor affine connection, $M$ denote mass of the particle, the generalized Dirac matrices, $\gamma^{\mu}=e^{\mu}_{\,\,\,a}\gamma^{a}$, are associated with the ordinary Dirac matrices, $\gamma^{0}=\sigma^{3}$, $\gamma^{1}=i\sigma^{1}$, $\gamma^{2}=i\sigma^{2}$, in $(1+2)$-dimensional flat space-time. They satisfy the anticommutation relation $\left\{\gamma^{a}, \gamma^{b}\right\}=2 \eta^{ab}$. In addition, the electromagnetic tensor is denoted by $F_{\mu \nu}$ in such a way that one can demonstrate its components as
$F_{0i}=-F_{i0}=E_{i}$ and $F_{ij}=\epsilon_{ijk}B^{k}$. The Dirac oscillator with the oscillator frequency $\omega$ can be given by non-minimal substitution\ $\mathrm{p}_{\mu} \rightarrow\mathrm{p}_{\mu}+i M \omega \beta \mathrm{x}_{\mu}$ with $\beta=\gamma^{0}$\cite{re1,re2}. Besides, the generalized Dirac oscillator can be written according to the following modified momentum operator \cite{re20}
\begin{equation}\mathrm{p}_{\mu} \longrightarrow \mathrm{p}_{\mu}+i M \omega \beta \mathrm{X}_{\mu}\delta^{r}_{\mu}, \qquad \mathrm{X}_{\mu} \rightarrow \mathrm{X}_{r}\equiv f(r).
\end{equation}
In this case, by considering the non-minimal coupling, the generalized Dirac oscillator with arbitrary potential function in this system reads
\begin{equation}\label{GDirac2}
\begin{aligned}
&\left[i \gamma^{\mu}(x)\left(\partial_{\mu}+\Gamma_{\mu}(x)+M \omega f(r) \gamma^{0} \delta_{\mu}^{r}\right) \right]\Psi(t,{\bf r})\\
&+\left[\frac{\tilde{\mu}}{2} \sigma^{\mu \nu}(x) F_{\mu \nu}-M\right]\Psi(t,{\bf r})=0.
\end{aligned}
\end{equation}
We choose the basis tetrad in the cosmic string background as follow:
\begin{equation}
e_{\,\,\,a}^{\mu}=\left(\begin{array}{cccc}
1 & 0 & 0 \\
0 & \cos \phi & \sin \phi \\
0 & -\frac{\sin \phi}{\alpha r} & \frac{\cos \phi}{\alpha r}
\end{array}\right), e_{\,\,\,\mu}^{a}=\left(\begin{array}{cccc}
1 & 0 & 0 \\
0 & \cos \phi & -\alpha r \sin \phi \\
0 & \sin \phi & \alpha r \cos \phi
\end{array}\right).
\end{equation}
Based on properties of the basis tetrad, the $\gamma^{\mu}(x)$ matrices can be expressed as
\begin{eqnarray}\label{Ggamma}
&&\gamma^{t}=\gamma^{0},\quad \gamma^{r}(x)=\tilde{\gamma}^{r}(x)=\cos \phi \gamma^{1}+\sin \phi \gamma^{2}, \nonumber\\ &&\gamma^{\phi}(x)=\frac{\tilde{\gamma}^{\phi}}{\alpha r}=\frac{(-\sin \phi \gamma^{1}+\cos \phi \gamma^{2})}{\alpha r},\nonumber\\ && \gamma^{\phi}(x) \Gamma_{\phi}(x)=\frac{(1-\alpha)}{2 \alpha r} \gamma^{\,r}.
\end{eqnarray}
On the other hand, the electric field consists of two distinct parts, one of which is denoted by $\textbf{E}_{1}$, which comes from a uniformly infinite charge filament located along the $z$-axis perpendicular to the polar plane, and the other indicated by $\textbf{E}_{2}$ originated of a uniformly charged non-conducting cylinder whose length and radius are $L$ and $R$, respectively \cite{re34,re35}.
It is worth mentioning that
the following configuration of the electric field can lead to the appearance of the A-C effect in this framework,
\begin{subequations}
\begin{align}
&\textbf{E}_{1}=\frac{2 \lambda_{1}}{r} \widehat{e}_{r},\quad \nabla \cdot \textbf{E}_{1}=2 \lambda_{1} \delta(r), \quad\left(r=\sqrt{x^{2}+y^{2}}\right),\label{E1}\\
&\textbf{E}_{2}=\frac{\lambda_{2} r}{2} \widehat{e}_{r}, \quad \nabla \cdot \textbf{E}_{2}=\lambda_{2}, \frac{\partial \textbf{E}_{2}}{\partial t}=0, \nabla \times \textbf{E}_{2}=0,\label{E2}
\end{align}
\end{subequations}
with $\lambda_{1}=\lambda_{0} / 4 \pi \epsilon_{0}, \quad \lambda_{2}=\chi / \epsilon_{0}$, the electric charge linear density $\lambda_{0}$ satisfies $\lambda_{0}>0$ , $r>0$ is the radial coordinate, and the electric charge volumetric density of cylinder is $\chi=\left(Q / \pi R^{2} L\right)>0$. In the polar coordinates system $(t,r,\phi)$, by considering the non-minimal coupling of the generalized Dirac oscillator, obtaining the term corresponding to the interaction of a spin-half particle having a permanent magnetic dipole moment with the electric field arising from $\frac{\tilde{\mu}}{2} \sigma^{\mu \nu} F_{\mu \nu}=-\tilde{\mu} \tilde{\gamma}^{\phi}E_{r}$, and substituting Eq. \eqref{Ggamma} in Eq. \eqref{GDirac2}, we get
\begin{equation}\label{GDirac3}
\begin{aligned}
&\left[i\gamma^{t}\partial_{t}+i\tilde{\gamma}^{r}\partial_{r}+\frac{i}{\alpha r} \tilde{\gamma}^{\phi}\partial_{\phi}+\frac{1-\alpha}{2\alpha r}\tilde{\gamma}^{r}
+M\omega f(r)\tilde{\gamma}^{\phi} \right]\Psi(t,{\bf r})\\
&-\left(\tilde{\mu} \tilde{\gamma}^{\phi}E_{r}+M\right)\Psi(t,{\bf r})=0.
\end{aligned}
\end{equation}
Equation \eqref{GDirac3} is obtained in the absence of the magnetic field.
By applying the similarity transformation matrix $S(\phi)$ on the matrices $\tilde{\gamma}^{r}$ and $\tilde{\gamma}^{\phi}$, they can be reduced to the constant Dirac matrices as follows:
\begin{equation}\label{similmatrix}
S(\phi)\tilde{\gamma}^{r}S^{-1}(\phi)=\gamma^{1}\equiv i\sigma^{1}, S(\phi)\tilde{\gamma}^{\phi}S^{-1}(\phi)=\gamma^{2}\equiv i\sigma^{2}.
\end{equation}
With respect to Eqs. \eqref{Ggamma}, \eqref{E1}, \eqref{E2} and \eqref{similmatrix}, we can rewrite Eq. \eqref{GDirac3} in the following form
\begin{equation}\label{GDriac4}
\begin{aligned}
&\left[\sigma^{2}\left(-\frac{\partial_{\phi}}{\alpha r}+iM\left(\omega f(r)-\frac{\omega_{AC}}{2}\right)-\frac{i \Phi_{AC}}{2\pi \alpha r}\right) \right]\Psi(t,{\bf r})\\
&+\left[i \sigma^{3} \partial_{t}-\sigma^{1}\left(\partial_{r}+\frac{1-\alpha}{2\alpha r}\right)-M I\right)]\Psi(t,{\bf r})=0,
\end{aligned}
\end{equation}
where $I$ is a $2 \times 2$ identity matrix and
the parameters $\omega_{AC}\equiv \tilde{\mu}\lambda_{2}/M$ denotes the A-C frequency related to the spin-half particle. In this way, $\Phi_{AC}=4 \pi \alpha \tilde{\mu} \lambda_{1}$ indicates the A-C phase. Note that we deal with the modified electric field $\mathbf{E} \rightarrow \mathbf{E}/\alpha$
impressed by the cosmic string space-time. In this case, we assume the Dirac spinor reads \cite{re37,re38}
\begin{equation}\label{spinorDirac}
\Psi(t, r, \theta)=\frac{e^{i\left(m_{l} \phi-E t\right)}}{\sqrt{2 \pi}}\begin{pmatrix}
\psi_{+}(r) \\
i \psi_{-}(r)
\end{pmatrix}, \left(m_{l}=\pm 1 / 2, \ldots\right),
\end{equation}
Then, we can obtain the two coupled Dirac equations by substituting Eq. \eqref{spinorDirac} into Eq. \eqref{GDriac4}
\begin{equation} \label{hh1}
\begin{aligned}
&i\left[\partial_{r}+\frac{1-\alpha}{2 \alpha r}+\frac{m_{l}}{\alpha r}-M\left(\omega f(r)-\frac{\omega_{A C}}{2}\right)+\frac{\Phi_{A C}}{2\pi \alpha r}\right]\psi_{-}(r)\\
&=\left(E-M\right)\psi_{+}(r),
\end{aligned}
\end{equation}
\begin{equation}\label{hh2}
\begin{aligned}
&\left[-\partial_{r}-\frac{1-\alpha}{2 \alpha r}+\frac{m_{l}}{\alpha r}-M\left(\omega f(r)-\frac{\omega_{A C}}{2}\right)+\frac{\Phi_{A C}}{2\pi \alpha r}\right]\psi_{+}(r)\\
&=i\left(E-M\right)\psi_{-}(r).
\end{aligned}
\end{equation}
Afterward, to combine the two coupled Dirac equations to get the following second-order differential equation, we need to remove $\psi_{-}(r)$ from Eqs. \eqref{hh1} and \eqref{hh2} and arrive at
\begin{equation} \label{GDriac6}
\begin{split}
&\frac{d^2\psi_{+}(r)}{dr^2}+\frac{1-\alpha}{\alpha r}\frac{d\psi_{+}(r)}{dr}+\left[E^2-\xi+M \omega \frac{df(r)}{dr}\right.\\
&\left.+2M\left(\frac{m_{l}}{\alpha r}+\frac{\Phi_{AC}}{2\pi \alpha r}\right) \left(\omega f(r)-\frac{\omega_{AC}}{2}\right)+\left(\left(\frac{1}{2\alpha}-1\right)^2\right.\right.\\
&\left.\left.+\left(\frac{m_{l}}{\alpha }+\frac{\Phi_{AC}}{2\pi \alpha }\right)-\frac{1}{4}-\left(\frac{m_{l}}{\alpha }+\frac{\Phi_{AC}}{2\pi \alpha }\right)^2\right)\frac{1}{r^2}
\right]\psi_{+}(r)=0
\end{split}
\end{equation}
with $\xi=M^2+M^2\left(\omega f(r)-\frac{\omega_{AC}}{2}\right)^2$.
In this case, the generalized Dirac oscillator with the external electric field in the cosmic string background breaks the degeneracy \cite{re19} if the potential function $f(r)=r$ is embedded in the Dirac oscillator.
Our next goal is mainly to study the energy eigenvalues
related to the Dirac oscillator with the Coulomb-type potential function under the background of cosmic string space-time.
\section{Exact solutions under the Coulomb-type potential function f(r)} \label{sec3}
In recent years, the Coulomb potential has received much attention for its importance in many subjects of physics \cite{re25,QuesneJPAMG2004,HassanabadiGRG2018,re21,ZareIJMPA2020,MontignyEPJP2022,re44,re45}. From now on, by selecting the Coulomb form for the potential function, that is,
\begin{equation}\label{PotFunc}
f(r) = \frac{\mathrm{N}_{1}}{r},
\end{equation}
where the constant potential parameter $\mathrm{N}_{1}$ is related to the Coulomb strength, we are able to solve the second-order differential equation, given by Eq. \eqref{GDriac6}, by employing the NU method. Now, let us place the potential function given by Eq. \eqref{PotFunc} in the Eq. \eqref{GDriac6} and get to the following expression
\begin{equation} \label{GDriac7}
\begin{split}
&\frac{d^2\psi_{+}(r)}{dr^2}+\frac{1-\alpha}{\alpha r}\frac{d\psi_{+}(r)}{dr}
+\frac{1}{r^2}\left[-\tau_{1}\,r^2+\tau_{2}\,r-\tau_{3}\right]
\psi_{+}=0.\\
&\tau_{1}=-E^2+M^2\left(1+\frac{\omega^{2}_{AC}}{4}\right),\\
&\tau_{2}=M^2 \omega \omega_{AC} N_{1}-\left(\frac{m_{l}}{\alpha }+\frac{\Phi_{AC}}{2\pi \alpha }\right)M\omega_{AC},\\
&\tau_{3}=\frac{1}{4}-\left(\frac{1}{2\alpha}-1\right)^2-\left(\frac{m_{l}}{\alpha }+\frac{\Phi_{AC}}{2\pi \alpha }\right)+\left(\frac{m_{l}}{\alpha }+\frac{\Phi_{AC}}{2\pi \alpha }\right)^2\\
&\quad+M\omega N_{1}\left(1+M\omega N_{1}-2\left(\frac{m_{l}}{\alpha}+\frac{\Phi_{AC}}{2\alpha \pi}\right)\right).
\end{split}
\end{equation}
Obviously, Eq. \eqref{GDriac7} is a particular case of the NU equation \cite{HassanabadiGRG2018,re40,re41, MontignyEPJP2022,TezcanIJTP2009,MontignyGRG2018,MontignyEPJP2021}. It is well-known that the NU method is an efficient model for finding eigenfunctions and eigenvalues of the Schr\"odinger-like equation \cite{MontignyEPJP2022}. Therefore, the second-order differential equation that is suitable for implementing the NU method can be presented as
\begin{equation}\label{NUEq}
\begin{aligned}
&\left[\frac{d^{2}}{d r^{2}}+\frac{1}{\left[r\left(1-\beta_{3} r\right)\right]^{2}}\left(-\zeta_{1} r^{2}+\zeta_{2} r-\zeta_{3}\right)\right]\psi(r)\\
&+\frac{\beta_{1}-\beta_{2} r}{r\left(1-\beta_{3} r\right)} \frac{d}{d r}\psi(r)=0,
\end{aligned}
\end{equation}
whose eigenfunctions can be written as
\begin{equation}
\psi(r)=r^{\beta_{12}}\left(1-\beta_{3}r\right)^{-\beta_{12}-\frac{\beta_{13}}{\beta_{3}}}P_{n}^{\left(\beta_{10}-1,\frac{\beta_{11}}{\beta_{3}}-\beta_{10}-1\right)}\left(1-2\beta_{3}r\right),
\end{equation}
also its eigenvalues can be presented by the following equation
\begin{equation}
\begin{split}
&n \beta_{2}-(2 n+1) \beta_{5}+(2 n+1)\left(\sqrt{\beta_{9}}-\beta_{3} \sqrt{\beta_{8}}\right) \nonumber\\
&+\beta_{7}+2 \beta_{3} \beta_{8}+2 \sqrt{\beta_{8} \beta_{9}}+n(n-1) \beta_{3}=0.
\end{split}
\end{equation}
Note that the parameters $\beta_{4}\dots, \beta_{13}$ can be found from the six parameters $\beta_{1}, \beta_{2}, \beta_{3}$, $\zeta_{1}, \zeta_{2}$ and $\zeta_{3}$ in Eq. \eqref{NUEq} according to the following relation (see Refs. \cite{MontignyEPJP2022,re40,TezcanIJTP2009,MontignyGRG2018,MontignyEPJP2021} for more details.)
\begin{equation}
\begin{split}
&\beta_{4}=\frac{1}{2}\left(1-\beta_{1}\right), \beta_{5}=\frac{1}{2}\left(\beta_{2}-2 \beta_{3}\right),
\beta_{6}=\beta_{5}^{2}+\zeta_{1}, \\
&\beta_{7}=2 \beta_{4} \beta_{5}-\zeta_{2}, \beta_{8}=\beta_{4}^{2}+\zeta_{3},
\beta_{9}=\beta_{3}\beta_{7}+\beta_{3}^{2}\beta_{8}+\beta_{6},\\
&\beta_{10}=\beta_{1}+2\beta_{4}+2\sqrt{\beta_{8}}, \\
&\beta_{11}=\beta_{2}-2\beta_{5}+2\left(\sqrt{\beta_{9}}+\beta_{3}\sqrt{\beta_{8}}\right),\\
&\beta_{12}=\beta_{4}+\sqrt{\beta_{8}}, \beta_{13}=\beta_{5}-\sqrt{\beta_{9}}-\beta_{3} \sqrt{\beta_{8}}.
\end{split}
\end{equation}
If we focus on Eq. \eqref{GDriac7}, we see that it is a particular case of the differential equation compatible with the NU method, where $\beta_{3}=0$. So, it makes sense that
\begin{equation}\label{limitedpsi}
\begin{split}
\lim\limits_{\beta_{3} \rightarrow 0} \psi(r)=r^{\beta_{12}} e^{\left(\beta_{13} r\right)} L_{n}^{\left(\beta_{10}-1\right)}\left(\beta_{11} r\right).
\end{split}
\end{equation}
Given Eq. \eqref{limitedpsi}, it can be seen that the $\psi(r)$ function is related to the generalized Laguerre polynomial, $L_{n}^{\beta} (x)$. With regard to these descriptions, the eigenfunctions related to the wave equation given by Eq. \eqref{GDriac7} can be written as follows
\begin{equation}\label{WF1}
\psi_{+}(r)=r^{1-\frac{1}{2\alpha}+\sqrt{\tau_{3}+\left(\frac{1}{2\alpha}-1\right)^{2}}}\,e^{-\sqrt{\tau_{1}}}\,\,L_{n}^{2\sqrt{\tau_{3}+\left(\frac{1}{2\alpha}-1\right)^{2}}}\left(2\sqrt{\tau_{1}}\, r\right),
\end{equation}
and the corresponding energy eigenvalues can be determined as
\begin{equation}\label{energy1}
\begin{split}
E_{n\,m_{l}}=\pm \sqrt{M^{2}\left(1+\frac{\omega^{2}_{AC}}{4}\right)-\frac{\tau^{2}}{\left[1+2n+2\sqrt{\tau_{3}+\left(\frac{1}{2\alpha}-1\right)^2}\right]^{2}}}\,.
\end{split}
\end{equation}
\begin{figure}
\includegraphics[width=0.46\textwidth]{WFN1.pdf}
\caption{The eigenfunctions in Eq. \eqref{WF1} as a function of $r$ in terms of four different values of $N_{1}$ and other parameters $n=3$, $\omega=M=m_{l}=\lambda_{2}=1$, $\lambda_{1}=-1$, $\tilde{\mu}=2$ and $\alpha=0.8$.}
\label{FIG1}
\end{figure}
\begin{figure*}
\resizebox{\linewidth}{!}{\includegraphics{WFN3.pdf},\includegraphics{WFN4.pdf}}
\caption{ Left panel: The energy eigenvalues in Eq. \eqref{energy1} as a function of $\alpha$ in terms of five different values of the quantum number $n$ and other parameters $\omega=1$, $M=0.2$, $m_{l}=2$, $N_{1}=-3$, $\lambda_{1}= -0.01$, $\lambda_{2}=5$ and $\tilde{\mu}=2$. Right
panel:The energy eigenvalues in Eq. \eqref{energy1} as a function of $n$ in terms of five different values of the A-C frequency parameter $\omega_{AC}$ and the permanent magnetic dipole moment parameter $\tilde{\mu}$, and other parameters $\omega=1$, $M=0.2$, $m_{l}=2$, $N_{1}=-3$, $\lambda_{1}= -0.01$, $\lambda_{2}=0.5$ and $\alpha=0.2$.}\label{fig:GreyBodyfromA0}
\end{figure*}
\begin{figure*}
\resizebox{\linewidth}{!}{\includegraphics{WFN5.pdf},\includegraphics{WFN2.pdf}}
\caption{Left panel: The energy eigenvalues in Eq. \eqref{energy1} as a function of he A-C frequency parameter $\omega_{AC}$ in terms of five different values of the A-C geometric phase $\Phi_{AC}$ and the permanent magnetic dipole moment parameter $\tilde{\mu}$, and other parameters $\omega=1$, $M=0.2$, $m_{l}=2$, $N_{1}=-3$, $\lambda_{1}= 0.1$, $\lambda_{2}=0.5$, $\alpha=0.2$ and $n=3$. Right
panel: The energy eigenvalues in Eq. \eqref{energy1} as a function of the Coulomb strength $N_{1}$ in terms of five different values of the quantum number $n$ and other parameters $\omega=1$, $M=0.2$, $m_{l}=2$, $\lambda_{1}= -0.01$, $\lambda_{2}=5$ and $\alpha=0.2$.}\label{fig:GreyBodyfromA0}
\end{figure*}
In Eq. \eqref{energy1}, one can see the influence of the Aharonov-Casher effect on the relativistic energy spectrum of the generalized Dirac oscillator containing the coulomb-type potential function associated with a neutral fermion with the permanent magnetic dipole interacting with the external electric field in $(1+2)$-dimensional cosmic string spacetime. Therefore it is clear that the relevant energy spectrum depend on the the A-C frequency parameter $\omega_{AC}$, the A-C geometric phase $\Phi_{AC}$, the Coulomb strength $N_{1}$, the defect parameter $\alpha$ related to the deficit angle, the oscillator frequency $\omega$, the mass of the fermion $M$, and quantum numbers $n$ and $m_{l}$.
Similarly, it can be seen that, under the influences of the A-C and curvature effects on such a Dirac oscillator in the present external electric field, the corresponding eigenfunctions depend on defined parameters $\omega_{AC}$, $\Phi_{AC}$, $\alpha$, $N_{1}$, $\omega$, $n$ and $m_{l}$.
Now, to observe the influence of the Coulomb strength $N_{1}$ on the radial wave function given by Eq. \eqref{WF1}, we draw $\psi_{+}(r)$ in terms of the coordinates $r$, according to four different values of the Coulomb strength $N_{1}$ in FIG.1.
It is work worth emphasizing that the energy eigenvalues of the Dirac oscillator is obtained for the ordinary potential function, that is, $f(r)=r$ in Ref. \cite{re19}.
To visually observe the influence of various parameters on the energy spectrum in the system, we consider the positive energies states. Therefore, in FIG.2(Left panel), we show that the relevant energy eigenvalues considered as the function of $\alpha$ in terms of five different values of $n$ can increase with a decreasing slope. Similarly, as the values of the A-C frequency $\omega_{AC}$ increases, the energy spectrum also gets smaller in FIG.2(Right panel).
Furthermore, we can intuitively understand the influence of the external electric fields on the energy spectrum of the generalized Dirac oscillator containing the Coulomb-type potential function in FIG.3(Left panel). Thus, we indicate that the corresponding energy eigenvalues, considered as the function of $\omega_{AC}$ in terms of five different values of $\tilde{\mu}$ and $\Phi_{AC}$, decrease as the values of the permanent magnetic dipole moment parameter $\tilde{\mu}$
and the A-C geometric phase $\Phi_{AC}$ increase.
It should be noted that the A-C phase $\Phi_{AC}$ implies a
periodicity on the relevant energy eigenvalues so that
$E_{n m_{l}}\left(\Phi_{A C} \pm 2 \pi\right)=E_{n m_{l}+1}\left(\Phi_{A C}\right)$ \cite{re16}.
Besides, one can see that the relevant relativistic energy eigenvalues considered as a function of $N_{1}$ increase as the quantum number $n$ increases in FIG.3(Right panel).
It is clear that the energy eigenvalues of the relevant generalized Dirac oscillator are modified by the Coulomb strength parameter $N_{1}$ related to the Coulomb-type potential function.
\section{Conclusions}\label{Conc}
In the framework of quantum field theory in curved space-time, we try to investigate the influence of the curvature and the A-C effects on the interaction of a Coulomb-type nonminimal coupling with the spinor field related to a relativistic neutral fermion interacting with the external electric field in $(1+2)$-dimensional cosmic string space-time. Meanwhile, the associated particle possesses a permanent magnetic dipole moment in the direction of the z-axis, which interacts with the external electric field constructed by two distinct radial field configurations; the first comes from a uniformly infinite charge filament presented by a Coulomb-type electric field, and the other originated of a uniformly charged non-conducting cylinder denoted by a linear-type electric field. Then, in the middle of solving the generalized Dirac oscillator containing the Coulomb-type potential function to find the exact analytical solutions related to the particle, we see that the Coulomb-type electric field leads to appears the A-C geometric quantum phase, and the linear-type electric field provides the A-C frequency in the associated wave equation.
In this way, by applying the NU method, we solve the corresponding generalized Dirac oscillator and obtain the eigenfunction and relevant energy eigenvalues.
Then, we see that the resulting solutions are shifted because of the Coulomb-potential function, A-C, and curvature effects, as can be seen in dependence of the A-C frequency parameter $\omega_{AC}$, the A-C geometric phase $\Phi_{AC}$, the Coulomb strength $N_{1}$, the defect parameter $\alpha$ related to the deficit angle and the oscillator frequency $\omega$. By the way, we can see that the degeneracy of the relevant relativistic energy eigenvalues is broken by depending on the Coulomb strength parameter under the influence of the curvature effect and the Aharonov-Casher effect. In addition, we plot some figures of the resulting solutions for further examination.
\vspace{5mm}
\begin{acknowledgments}
This work is supported by the National Natural Science Foundation of China (Grant Nos. 11465006 and 11565009) and the Major Research Project of innovative Group of Guizhou province (2018-013).
\end{acknowledgments}
| d5729daa853119e880b0a6111315dd2964652bda | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Recently, it has been very popular to utilize cloud servers to carry out machine learning algorithms instead of using local servers. However, since cloud servers are semi-trusted, private data, such as personal information and medical records, may be revealed in cloud computing. For the reason, privacy-preserving machine learning has become an urgent challenge\cite{homomorphic1, 2007.08775, sirichotedumrong2019privacy, sirichotedumrong2019pixel, 2006.01342}.
In this paper, we focus on dimensionality reduction methods in terms of two issues: difficulty in estimating visual information on original images from dimensionally reduced ones, and performance that reduced data can maintain in an image classification experiment.
In machine learning, dimensionality reduction is used for not only reducing the number of random variables, but also protecting visual information for privacy-preserving machine learning. However, dimensionality reduction methods have never been evaluated in terms of above the two issues at same time.
For such a reason, difficulty in estimating visual information is discussed. In particular, the random sampling method that was proposed for privacy-preserving machine learning\cite{aprilpyone2020encryption}, is compared with typical dimensionality reduction methods such as random projection and PCA\cite{bingham2001random, wold1987principal}.
In an image classification experiment, the random sampling method is demonstrated not only to maintain high difficulty, but also to have close machine learning performance to that of the random projection method.
\section{Linear Dimensionality Reductions}
Let us consider a projection from a vector $x\in\mathbb{R}^D$ to a low-dimensional vector $y\in\mathbb{R}^K (K<D)$.
If the projection can be represented by using a matrix $P\in\mathbb{R}^{K\times D}$ as
\begin{align}
y=Px\quad, \label{linear dim reduction}
\end{align}
it is a linear dimensionality reduction and $P$ is called a projection matrix.
In machine learning, $P$ is used for reducing the number of random variables for avoiding negative effects of high-dimensional data.
The random projection method\cite{bingham2001random} and principal component analysis (PCA) are typical linear dimensionality reduction methods.
The random projection is a method that does not use any statistics of dataset, but PCA is not.
For the random projection, elements of $P$ have a normal distribution with an average value of 0 and a variance of $\sqrt{1/K}$. Therefore, the random projection is not required to calculate any statistics of dataset for designing a projection matrix $P$.
\section{Random Sampling}
The random sampling method was proposed as a dimensionality reduction method for privacy preserving machine learning\cite{randomsampling}.
It is also expected to be applied to deep convolutional neural network, due to the property of spatial information invariant\cite{aprilpyone2020encryption}.
Let us consicer applying the random sampling method to a pixel vector $x\in\mathbb{R}^D$ of an image to create $y\in\mathbb{R}^K (K<D)$.
Next, let $\{\phi(i)\, |\, i=1,\ldots,K\}$ denote $K$ indexes selected from $D$ pixel indexes,
where $\phi(i)\neq\phi(i^\prime)$ if $i\neq i^\prime$, randomly generated with a seed.
By using $\phi(i)$, the random sampling operation can be written as
\footnotesize
\begin{align}
\label{random sampling}
y=(x_{\phi(1)},x_{\phi(2)},\ldots,x_{\phi(K)})^{\rm T}\quad,
\end{align}
\normalsize
where $x_{\phi(i)}$ is the $\phi(i)$-th element of $x$.
Here, if we define a matrix $P\in\mathbb{R}^{K\times D}$ with elements $p_{ij} (i=1,\ldots,K,j=1,\ldots,D)$ defined by
\footnotesize
\begin{align}
& p_{i,j}=\left\{
\begin{array}{ll}
1 & (j=\phi(i)) \\
0 & ({\rm otherwise})
\end{array}
\right. \quad, \label{random sampling 2}
\end{align}
\normalsize
the random sampling is reduced to the form of Eq.(\ref{linear dim reduction}).
That is, the random sampling is a linear dimensionality reduction, and is a method that does not use the statistics of dataset as well as the random projection.
\section{Visual Information Estimation}
\label{attacks}
Assuming that an attacker knows a projection matrix $P$ used in dimensionality reduction, difficulty in estimating visual information on plain images is discussed.
The attacker's goal is to create $Q$ to approximatelly reconstruct the target image $x$ from the low dimensional vector $y$ as
\begin{align}
x^\prime=Qy\quad. \label{reconstruction}
\end{align}
In this paper, two attacks are considered to estimate $Q$.
\subsection{Attack With Pseudo-Inverse Matrix}
An attacker can use a pseudo-inverse matrix ($Q_{\rm pinv}$) of projection matrix $P$ to estimate visual information on original images,
where $Q_{\rm pinv}$ is designed by using an algorithm with the singular-value decomposition of $P$\cite{harville1998matrix}.
\subsection{Regression Attack With Attacker's Dataset}
\label{regression attack}
An attacker first prepares his own dataset ($X_{\rm attack}$) and a dataset ($Y_{\rm attack}$) projected from $X_{\rm attack}$ by using $P$, and then designs a linear reconstruction matrix ($Q_{\rm reg}$) that regresses $X_{\rm attack}$ from $Y_{\rm attack}$ in accordance with the least squares method.
In general, the effectiveness of this attack depends on the relation between the distribution of $X_{\rm attack}$ and that of target images.
Therefore, in this paper, we classify $X_{\rm attack}$ into two types in accordance with the distribution of $X_{\rm attack}$.
\begin{itemize}
\item type 1: $X_{\rm attack}$ consists of images with the same class-labels and distribution as those of the target images.
\item type 2: $X_{\rm attack}$ consists of images with class-labels and a distribution that are different from those of the target images.
\end{itemize}
\begin{comment}
\begin{itemize}
\item type1 ({\it Same Labels}): $X_{\rm attack}$ exactly has the same conditions and class labels as the target images.
\item type2 ({\it Different Labels}): $X_{\rm attack}$ has similar conditions to the target images but has different class labels.
\item type3 ({\it Totally Different}): $X_{\rm attack}$ has different conditions from the target images.
\end{itemize}
The {\it type 3} is practical when, for example, the attacker estimates facial images in a face recognition dataset.
In this case, the attacker knows the conditions under the faces were taken, but the face samples of the target people are available.
Here, we show concrete examples of the above three conditions, taking the case of an attacker attempting to reconstruct facial images in a face recognition dataset.
\begin{itemize}
\item {\it Same Labels} condition: The attacker knows the training images in the face recognition dataset.
\item {\it Different Labels} condition: The attacker knows the conditions under which the faces were taken, but the face samples of the the target people are not available.
\item {\it Totally Different} condition: The attacker does not know any condition about the dataset.
\end{itemize}
\end{comment}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.47]{./accs_linsvm.png}
\caption{Classification accuracy with various dimensionality reduction methods (Linear SVM). \label{accs linsvm}}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.47]{./accs_rf.png}
\caption{Classification accuracy with various dimensionality reduction methods (Random Forest). \label{accs rf}}
\end{center}
\end{figure}
\section{Experiment}
Face-image classification experiments were carried out for evaluating the random sampling method in terms of both classification accuracy and difficulty in estimating visual information.
The dataset was Extended Yale Database B\cite{Yale}, which contains 38 individuals and 64 frontal facial images with 168$\times$192 pixels per each person.
Each image was normalized to a size of 28$\times$28, so it had $D=784$ dimension as a vector,
we splitted the dataset into two datasets: $X_{\rm main}$ and $X_{\rm sub}$, each of which had 19 classes and 1216 images, without duplication of classes.
Moreover, $X_{\rm main}$ was divided into $X_{\rm train}$ (912 images) for training and $X_{\rm test}$ (304 images) for testing.
We also used the CIFAR-10\cite{krizhevsky2009learning} dataset: $X_{\rm CIFAR-10}$ for evaluating difficulty in estimating visual information on $X_{\rm test}$.
This dataset consists of 60k images with 10 classes such as dogs and ships,
whose distribution is different from $X_{\rm train}$, $X_{\rm test}$ and $X_{\rm sub}$.
Finally, each vector $x\in\{X_{\rm train}$, $X_{\rm test}$, $X_{\rm sub}$, $X_{\rm CIFAR-10}\}$ was projected to $y\in\{Y_{\rm train}$, $Y_{\rm test}$, $Y_{\rm sub}$, $Y_{\rm CIFAR-10}\}$ with a target dimension ($K$) by using the random sampling and three dimensionality reduction methods: the random projection, PCA, and a feature selection algorithm proposed by
Ono\cite{featureselector}.
PCA and Ono's method require calculating the statistics of $X_{\rm train}$, but the random sampling and random projection do not.
\begin{figure*}[t]
\begin{center}
\includegraphics[scale=0.77]{mse.png}
\caption{Mean squared error between original image and reconstructed image. \label{mse}}
\end{center}
\end{figure*}
\subsection{Machine Learning Performance}
We trained a random forest classifier and SVM with the linear kernel by using $Y_{\rm train}$, and tested by using $Y_{\rm test}$.
Figures \ref{accs linsvm} and \ref{accs rf} show the comparison of the dimensionality reduction methods in term of classification accuracy.
Under the use of SVM, the random sampling had a similar performance to Ono's method.
For the random forest, the random sampling also has almost the same accuracy as that of Ono's method.
As a result, the random sampling was demonstrated to be comparable with other dimensionality reduction methods, while maintaining the property of spatial-information invariant.
\subsection{Robustness Against Visual Information Estimation}
Assuming that an attacker knows only projection matrix $P$ used in dimensionality reduction, difficulty in estimating visual information on $X_{\rm test}$ was evaluated.
We assumed the following four attacks to estimate $X_{\rm test}$ from $Y_{\rm test}$.
\begin{itemize}
\item Attack 1: Attack using a pseudo-inverse matrix of $P$.
\item Attack 2: Regression attack with $X_{\rm train}$ (type 1 in section \ref{regression attack}).
\item Attack 3: Regression attack with $X_{\rm CIFAR-10}$ (type 2 in section \ref{regression attack}).
\item Attack 4: Regression attack with $X_{\rm sub}$.
\end{itemize}
In attack 4, the attacker does not have facial images of the people included in $X_{\rm test}$, but knows the conditions under which $X_{\rm test}$ was taken.
Figure \ref{mse} shows MSE values between $X_{\rm test}$ and $X_{\rm test}^\prime$.
From the figure, the random sampling was relatively robust compared with the other dimensionality reduction methods.
The absolute values of MSE of the random sampling were large for attacks 1 and 3,
but were small for attacks 2 and 4 as well as the other methods.
We defined accuracy reduction ratio (ARR) as another criterion.
First, we trained a logistic regression classifier ($\theta$) by using $X_{\rm train}$, and then the ARR is defined as
\begin{align}
{\rm ARR}=\frac{{\rm ACC}_\theta(X_{\rm test})-{\rm ACC}_\theta(X_{\rm test}^\prime)}{{\rm ACC}_\theta(X_{\rm test})}\quad, \label{definied arr}
\end{align}
where ${\rm ACC}_\theta(X)$ is the function which returns the accuracy when dataset $X$ is applied to $\theta$.
A high ARR value indicates that $X_{\rm test}^\prime$ has low class-specific visual information, i.e., the dimensionality reduction is robust against the attack.
Figure \ref{arr} shows the comparison of ARR values.
From the figure, the random sampling was demonstrated to be robust against attack 4.
It indicates that attack 4 did not effectively recover the class-specific information of the target images.
Figure \ref{reconst images} shows an original images and examples of reconstructed images.
Although the images reconstructed by attack 4 can be easily interpreted as human faces,
the facial features were different from the original person.
\onecolumn
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.68]{arr.png}
\caption{Comparison of accuracy reduction ratio values. \label{arr}}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\includegraphics[scale=0.7]{PinvReconstImgs3.png}
\caption{Examples of reconstructed images ($D$=784) with MSE (left) and ARR (right) values. \label{reconst images}}
\end{center}
\end{figure}
\twocolumn
\section{conclusion}
In this paper, we compared the random sampling method with typical linear dimensionality reduction methods in terms of both machine learning performance and difficulty in estimating visual information in face classification experiments. The random sampling was demonstrated not only to have the property of spatial position invariant which is useful for privacy-preserving learning, but also to maintain comparable performance to other dimensionality reduction methods and difficulty in visual information estimation under practical conditions.
\bibliographystyle{IEEEbib}
| db19791034ac4a7620c4ddbfbe27e0e94540dd38 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{1. Introduction}
Deep learning is developing rapidly in many areas of Natural Language Understanding. While the deep learning requires a large amount of data to ensure the generalization of the model. For example, in the famous pre-trained language model ''BERT'' \cite{devlin-etal-2019-bert} where Google provides 400 million data to train Bert model, which requires a huge amount of data to guarantee the model performance. However, in many cases we cannot obtain a large amount of data. For example, training data is scarce in the task of word emotional classification.\\
Previous research \cite{DBLP:conf/acl/HatzivassiloglouM97} tried to get the "semantic orientation" of the sentence. For example, positive and negative tendencies. through obtaining the predictions of emotion tendencies and comparing the predictions with the real sentences, the accuracy of the prediction is obtained.\\
Further, the researchers found that, based on psychological research, multiple factors will lead to different emotions rather than solely determined by emotional polarity. For example, in the study of Bradley and Lang's research \cite{Measuring-emotion}, Valence-Arousal-Dominance model to determine the emotional tendency from three different factors.\\
Word embedding is an effective way to represent language features, such as similarity\cite{DBLP:conf/emnlp/PenningtonSM14}, emotional orientation \cite{DBLP:conf/naacl/RotheES16}, and sentiment\cite{DBLP:journals/ci/CalvoK13}. Word embedding is the method to victories strings, different dimensions of vectors may contain different information, but embedding representation is "black box", it is hard to know what exactly mean the dimension present.\\
In this paper, we proposed that emotional information can be compressed as a one dimension embedding called ultradense subspace, we hypotheses the ultradence subspace as the space where the sentiment be quantified as numerical values and be classified by trained ultradense word embedding. For Chinese and English, we predict that the emotional words of two languages can correctly classify but not restricted by language types and grammars.\\
The advantage of this method is that even if the data is lacking, the model has good performance. And the data does not need to be labeled, so there is a large amount of data that can be used to train. Another advantage is that, compared to the origin word embedding, the ultradense subspace has higher performance and quality. For example, in classification tasks, unrelated information is considered as noise, which is filtered through orthogonal transformation. As a result, the model has fewer parameters, which means that the training speed is improved and the possibility of overfitting is reduced.\\
In this paper, our goal is to produce a decent word embedding in Chinese. The output of model will be a lexicon, which is converted from the original embedding, the embedding of lexicon presented as one-dimensional ultradense embedding table, this table produce a clearer, easier-to-explain word representation for word emotions.
\section{2. motivation}
In most of the previous studies, the models were trained using English datasets, and rarely use Chinese datasets, especially web review data, the web review contain stronger emotional tendencies than the written context. I would like to do Chinese emotional classification by using this method, which transforms embedding into a low-dimensional, interpretable embedding. Based on the output of the embedding, we can know whether the model is well in Chinese emotion classification. At the same time, I am also looking for factors that decrease the performance of the model. For example, according to research \cite{DBLP:conf/naacl/RotheES16}, they find that some factors of embedding such as embedding size, stop words, embedding dimensions, and seed words may effect performance of model. Thus, we decide to evaluate what kind of factors
influence the mode results.\\
\section{3. Background}
In this section, I will introduce some relevant background about emotional expression and word embedding, and some of techniques will use in our methodological work.
\subsection{3.1 emotion representation}
Emotional expression is not only distinguished by positive and negative, but also by multiple factors. The VAD model (Figure 1) \cite{Measuring-emotion} considers that emotions to be composed of three different dimensions, named Valence (positive-negative), Arousal (clam-excited), and Dominance (perceived degree of control situation). There are also studies \cite{DBLP:conf/hci/ZhongQZ19} which suggest that Dominance does not affect the expression of emotion (only VA model works). As the most common system for assessing emotional tendencies, the VAD model quantifies Valence, Arousal and Dominance, and then evaluates them.
\begin{figure}
\centering
\includegraphics[scale=0.6]{VAE_model.png}
\caption{The emotion space covered by the Valence-Arousal-Dominance (VAD) model}~\label{fig:figure1}
\end{figure}
\subsection{3.2 emotion lexicon}
In some psychological studies, words with emotional tendencies illustrated as specific numerical values. For example, if the value of a word is greater than a certain threshold, the word is considered to be a positive emotional word, whereas if a word is less than the threshold, it is Negative emotion word. In this paper, the emotion lexicon that we build only use positive and negative emotions as a one-dimensional evaluation. For illustration, we will visualize the result as a 2d figures to build the lexicon that the emotional values of word distributed on the figures.\\
\subsection{3.3 word embedding}
Word embedding is a vector representation, which encodes word into a vector form. In this way, word is converted to the numeric values, it can use into unsupervised learning. Here are some very popular embedding algorithms today: WORD2VEC is a popular embedding algorithm, which provides a trimmed down neural network \cite{DBLP:conf/nips/MikolovSCCD13}. FASTTEXT is based on WORD2VEC, which is a derivative of WORD2VEC and combining the n-gram of characters \cite{DBLP:journals/tacl/BojanowskiGJM17}. Unlike the previous embedding algorithms, GLOVE directly trains word vectors on the word co-occurrence matrix, which improves the training efficiency.
\subsection{3.4 word-level prediction}
In previous research \cite{DBLP:journals/tois/TurneyL03}, researchers distinguished the emotional tendency of words by using the polarity of words. The specific approach is to divide words into two kind of training data, the original word embedding and the seed words. The seed words are usually positive or negative The word, the distance between the pair of seed words is obtained point-wise, so as to obtain the emotional tendencies of words. Afterwards, Rothe \cite{DBLP:conf/naacl/RotheES16} proposed to transform the word embedding from the original space to ultradense subspace by training the orthogonal matrix, which will get embedding of word's emotion.\\
\section{4. Research Method}
In order to analyze the sentiment of TikTok comment data and build TikTok sentiment lexicon, I need to collect and process reviews from the TikTok app. My plan is to use web crawlers to crawl data from the TikTok app and preprocess and victories the data, which need to implement a web crawler framework to collect data for the TikTok app, then use python data analysis tool Pandas to pre-process review data and use WORD2VEC training tool Gensim to victories review sentences, the processed word embedding table saved as the VEC files. For seed words, I labeled 5, 10, 15 for seed word groups. In the end, After trained model, I visualized and analysed results.\\
\section{5. Research Questions}
First, "Is there a relationship between the seed words and the emotions?" When the size of word reaches a certain amount, the generalization ability of the model improved, that is, the model may learn more prior information. While the model learned the distance between seed words, we can verify that whether the distinguishing ability of model be affected when the size of seed words change. One hypothesis is that the model's ability to distinguish emotions grows if we expands the size of the seed word. To verify the hypothesis, I set the size of seeds to 5, 10, 15 respectively, and each seed word group join the training processes. Finally, 3 results will be visualized to verify the hypothesis.\\
The second question is "Does a specific video guide the commentator's emotions?" When we collected data, we found that the commentator's emotional inclination would be consistent with the emotional tendencies that the video wanted to express. For example, most Comments are sad in videos of funerals, but most of the comments are pleasing in funny videos. Therefore, we made a hypothesis that the sentiment of a particular video will guide the comments emotional tendencies. To verify the hypothesis, we collected two different data, comment data from random videos and comment data from special videos, I use different datasets to train the model, and finally compare the distribution of words.\\
The third question is "Compared with the previous method, is there any improvement in compressing embedding to ultra dense subspace?" In the previous research, the commonly used approach of feature extraction is Principle Compose Analysis, PCA is the feature selection method, which is a commonly used tool in data analysis and model prediction. It is used to visualize the distance and relationship between data. PCA uses Singular Value Decomposition to select appropriate dimensions to achieve the goal of dimensionality reduction of data. In this paper, after processing data by PCA and the ultradense compression, we will compare the results to find that which method can better show the emotion of words. One hypothesis is than the data processed by ultradense compression has the better performance on emotional classification task, because ultradense compression uses seed words, which reduces the distance between the same emotion words and increases the distance between different emotion words. To verify this hypothesis, we will use the same training data to get results from PCA and ultradense compression, and then we will visualize and analyze the results to verify hypothesis.
\section{6. Data Collection}
\subsection{6.1 collection}
The data we need to collect are user's comments from TikTok random videos and user's comments from particular videos. The comment data from random videos can analyse the user's sentiment for making sentiment lexicon, and the comment data of specific videos can analyze the user's sentiment expression under specific videos. People's sentiment often be reflected in words, and words with sentiment appeal are often commented by people. As a result, after the vectorization of comment data in vector space, the vector that have same sentiment are often collected together. Through research on RQ1 We can get the distribution of the vector of random video comments in vector space, which will tell us what sentiments expression will show what vector distribution.\\
We randomly crawl TikTok comments data through the web crawler, and crawl specific comment data under some very high like videos. These comment data can be found at \url{https://github.com/h2222/douyin_comment_dataset}
the data contains 3 type of files. The dataset.csv file represents the original data crawled from TikTok where data preprocessing is required. The fixdata.py file is a data processing file for processing raw data. The processed file saved in dataupdate.xlsx file For further use, the data collects the user's age, gender, nickname, comments, number of comments liked, etc.
\subsection{6.2 preprocess}
The data will be saved as a csv file, and a new repository will be created to save the data on GitHub. The data set to two categories: random video comment data and specific video comment data. For the dataset, the related heads includes the user's nickname, age, gender, number of likes, and comment details, where gender head can be binarized, such as 0 for female and 1 for male. In addition, comments can be grouped with users of different ages to research if age affects sentiment expression. The Python Pandas module use to analyse and preprocess the data and Gensim module use to train a original embedding table by pre-processed comment data, we will also labeled emotional words as seed words.
\section{7. Data Analysis}
First of all, I want to evaluate the performance of the model When using a different number of seed words. I decided to plot the distribution of a word in the case of using different seed words. the figure 2 is the Chinese TikTok words distribution, which is based on the training results of the model, the emotional distribution can be shown as the distance between words.\\
\begin{figure}
\centering
\includegraphics[scale=0.25]{differences.png}
\caption{the words distribution based on different seed words, red part is 5 seeds result, blue part is 10 seeds result, green part is 15 seeds result }~\label{fig:figure1}
\end{figure}
According to the word distribution, the X axis is the number of all words, which comes from the TikTok comment data after the word segmentation. The Y axis is a random number between 100 and -100, the reason we add the Y axis is that the word distribution is 1D distribution, which means that the experiment only obtains positive and negative emotions, and the distribution results will be "crowded" on a line. In order to better visualize the distribution results, we add a random number as the second dimension, so The words are mapped on the 2D plane.\\
Looking at the image, we find that the model's ability to distinguish positive emotions is stronger than that of negative emotions. For example, if we think that the position where index equals 0 is the distribution of neutral words, the minimum value of negative words are close to -600. Instead, The maximum value of positive words are close to +2000. Evaluating from the semantics, the model are also better distinguishes positive sentiment words. I think the reason for that is because there are more positive words in the TikTok dataset than negative words, so it will Let the model produce better training results to positive words.\\
Further, we found that the size of the seed words may affects the ability of emotional classification. For example, under the condition of training the model with same dataset, we found that when the seed word is 15 (green distribution in the figure), the positive word The maximum value is +1000 and the minimum value of the negative word is -250, and the span is approximately 1250. When the seed word is 5 (the red part in the figure), the maximum value of the word is +2000 and the minimum value is -600, and the span is 2600. Therefore , the smaller the seed words, the stronger the model's ability to distinguish sentiment words. In my opinion, during the training process, the model distinguishes word sentiment by optimizing the distance between seed words and embedding words. Excessive addition of seed words may increase the model's Parameters, which leads to overfitting of the model. Based on the experiments, when the size of the seed words equal to 5, the model's result is the best.\\
\begin{figure}
\centering
\includegraphics[scale=0.25]{line_graph_nosorted.png}
\caption{the result embedding value against the set of words, red line is 5 seed result, blue line is 10 seed result, green line is 15 seed result }~\label{fig:figure1}
\end{figure}
Next, we draw a line graph to show the model's distinction between sentiment words. As shown in figure.3, the X axis is the number of all words, and the Y axis represents the output value of each words. According to the figure, we found that the model can be more distinguish the first 2000 words to a large extent of which 1 to 2000 words have a larger interval on the Y axis. When the word is after 2000, the model gradually loses its ability to distinguish the emotion of the words.\\
Next, we sort the results. The model converts words embedding table (shape is [vocabulary size, embedding size]) in the original space to words representation (shape is [vocabulary size, 1]) in ultradense subspace. Therefore, we sort all the words according to the values in the new 1D space. The result of sorting is shown in figure 4.\\
\begin{figure}
\centering
\includegraphics[scale=0.25]{line_graph_sorted.png}
\caption{the line graph of sorted words, red line is 5 seed result, blue line is 10 seed result, green line is 15 seed result }~\label{fig:figure1}
\end{figure}
We found two obvious upward and downward trends on both sides of Figure 4, which shows that the word representation can distinguish the sentiment of the word in the ultradense subspace. When the seed words are 15 (the green part in the figure), the turning point is more obvious, I guess the increase of seed words may increase the model's ability to classify fuzzy words. For example, some nouns words with emotional tendencies such as 'cake', 'firework' can be correctly classified.\\
\begin{figure}
\centering
\includegraphics[scale=0.25]{PCA_curve.png}
\caption{the result of PCA, the X axis is words index sorted by embedding value of PCA, the Y axis is the word embedding of PCA}~\label{fig:figure1}
\end{figure}
In order to prove that the emotion can be better represented by using the ultradense compression method, we use PCA for comparative experiments. The original word embedding table (shape is [vocabulary size, embedding size]) is compressed into a new embedding ( shape is [vocabulary size, 1], shape is consistent with that of result of ultradense compression), as shown in Figure 5.\\
As shown in Figure 5, the index of the sorted words on the X axis is based on the word embedding value of PCA, and the Y axis is the word embedding value. Compared with the orthogonal transformation method (Figure 4), the curve of the PCA-based method is more gentle, This shows that PCA cannot distinguish the emotion of words well. Moreover, the PCA method has insufficient predictive ability for negative emotion words.
\section{8. Findings and discussion}
\subsection{8.1 Is there a relationship between the seed words and the emotion?}
A result shows that there is indeed a relationship between seed words and emotional tendencies, but the word emotion may more related with training data. An obvious conclusion is that most of the comments use positive words but negative words have not been widely appeared in the comments, which can be verified by figure 6, The interesting result is that words with a large ultradense embedding value appear very frequently in comments. These words usually have a positive meaning. table 1 are some of the most frequent words and the number of occurrences.\\
\begin{table}
\centering
\begin{tabular}{l r r r}
{\small\textit{word}} & {\small \textit{numbers}}\\
\midrule
feel & 3710 \\
Awesome & 3523 \\
heart & 3365\\
praise & 2351 \\
...\\
\end{tabular}
\caption{Table captions should be placed below the table. We
recommend table lines be 1 point, 25\% black. Minimize use of
table grid lines.}~\label{tab:table1}
\end{table}
In the comment dataset, the most of high frequent words have positive meanings, even when we set the seed words, we found that 30\% of the seed word appeared in the top 200 words of comment dataset. Combined with the previous model result, we guess that if words with certain type appear frequently, the word is easier to be distinguished by the model. In addition, we found that in the dataset, in the comments that contain high-frequency words, the comment with 2 or more The high-frequency words are 44\%. In other words, nearly half of the high-frequency words appear in pairs in the comments, which will cause the associated relationship when the word is converted into embedding. As the result, the model is affected.\\
\subsection{8.2 Does a specific video guide the commentator's emotions?}
Special model videos will impact on the comments. the five different categories of video comment are collected and the categories are health, star, funny, news and entertainment. Through model training to get emotional lexicon, we found that the emotional tendency is biased to positive in funny videos, but the emotional tendencies of comments are more neutral in news videos. \\
I guess the reason for this result is that the content of the video will guide the emotional tendency of the comment, so that the frequency of the words that match the emotional tendency of the video will increase, and the model can better learn the emotional information of those words, thereby improving the model The ability to discern specific emotions. For example, when processing the comment data of funny videos, we found that some words with a clear positive tendency have a high Term Frequency (other words divided by the most frequent words), such as Awesome, funny , happy face, the Term Frequency of these words can be close to 0.9. Finally, we use the model to classify the emotions of the words, and find that the model has a more obvious effect on the classification of words with high frequency.
\subsection{8.3 Compared with the previous method, is there any improvement in compressing embedding to ultra dense subspace?}
After testing, we found that the performance of ultra dense word embedding is better than that of PCA method in sentiment classification tasks. Comparing Figure 5 and Figure 6, we observed a clear distinction between words using the ultra dense embedding method (the turning point in Figure 5). On the contrary, although the PCA method can distinguish words (the curve in Figure 6), it cannot clearly distinguish words, and the curve is relatively smooth, so PCA cannot distinguish the emotion of words well.\\
I think that ultra dense word embedding performs better because the model learns emotional information from seed words. The model reduces the distance of the same emotional words in the seed word and increases the distance of the opposite emotional words, so that the transformed representation of the model learns the emotion Information, which improves the word emotional representation. The PCA method does not analyze the emotion of the word, so ultra dense word embedding is better in the emotion classification.\\
\section{9 Threats to Validity}
My paper may still have some disadvantages. First, my data collection volume is small and limited. Since English review data cannot be collected, we only build sentiment dictionaries on Chinese data, because both Chinese and English are different in grammar and expression, I expect differences in different language data sets, so a comparative test is necessary To evaluate the performance difference of the model when dealing with different languages. \\
secondly, the amount of data we collected did not meet expectations. We spent a long time on collecting TikTok's review data. As a very popular application, TikTok's company designed a complex encryption algorithm to protect user privacy, This is a very good thing, but for researchers who want to perform data analysis, they have to spend a lot of time users cracking the anti-crawl mechanism of TikTok. When collecting data with comments, it took us a long time to crack TikTok ’s x-gergon encryption algorithm, which is a hexadecimal encryption based on MD5, and it is difficult to crack. As a result, the data I collected did not reach the expected amount. In the future research, we will strengthen the data collect.\\
Finally, I think that in the experimental stage, the performance of the model can be further optimized. In the training process of the model, because we need to reduce the distance between the same emotion words and increase the distance between different emotion words, we need to optimize the same and different emotions at the same time. The loss of the word. Therefore, we use a loss function with hyper-parameter alpha, which can combine the two losses in a certain proportion, so that the global loss can be optimized. The formula is as follows:\\
\begin{align*}
Loss = (1 - \alpha) \times SLoss + \alpha \times DLoss
\end{align*}
where $\alpha$ is the hyperparameter of the model, SLoss is the loss of the same sentiment words, and DLoss is the loss of different sentiment words.\\
During training, we found that when the iteration of the model reaches the 4th epoch (we set the model to iterate 10epoch), the loss of the model hardly drops. Regarding the cause of this problem, we guess that it is caused by the bias of the model data. In find and In the discussion section, we mentioned that positive emotion words have a higher frequency and higher in the data set, which leads to uneven distribution of positive emotion words and negative emotion words in the data set. Therefore, in the training stage of the model,
the positive words are better optimized by the model compared to negative words, so the classification ability of the model for positive words is better than the classification ability of negative words.
\section{10 Related Work}
Faruqui \cite{DBLP:conf/acl/FaruquiD15} use word post-processing based on word similarity, so Faruqui does not use processing based on orthogonal transformation and does not need to consider word distance, which makes it worse for other applications such as syntax detection.The method of Faruqui does not benefit that ultra-dense embedding is more efficient.\\
In the tensor framework, Rothe and SchUtz \cite{DBLP:conf/acl/RotheS15} converts word embedding to perceptual (synonym) embedding. In their work, all the embeddings exist in the same subspace. However, we want to keep only specific information embedded in the subspace to create ultra-dense subspaces with a specific role (sentiment embedding).\\
The method we used is also related to directed PCA \cite{PCA}. However, compared to PCA, the directional PCA solution is not orthogonal.\\
The creation of the sentiment lexicon usually needs to label the training corpus. For example, people manually label the sentiment tendency of the words in the training set or form words with the same meaning into the training set, or add words with similar meanings to the seed set (\cite{DBLP:conf/acl/Turney02} and \cite{DBLP:journals/jair/KiritchenkoZM14}). Heerschop team \cite{DBLP:conf/bis/HeerschopHF11} used wordNet and pageRank-based algorithms to propagate the emotional information of seed words into unknown words. Scheible team \cite{DBLP:conf/acl/Scheible10} proposed a semi-automatic method for machine translation Based on sentiment lexicon. Hamdan team \cite{DBLP:conf/semeval/HamdanBB15} evaluated the sentiment level based on the average of six sentiment lexicons. Those method cannot apply less-resource languages. Our experiments show that the orthogonal transformation method can train the sentiment lexicon with a small amount of training resources, and has better performance than the lexicon created by other semi-automatic methods.\\
\section{11 Conclusions}
We use the web crawler method to collect TikTok's Chinese comment data, and apply the method that can embed the original vector of data into the ultra-dense subspace. In this study, two experiments were used to verify that the ultra-dense subspace can classify the emotion of words. 1. By setting the number of different seed words, we get that the model can distinguish emotions through model training. 2. By comparing with the PCA method, ultra-dense word embedding can better distinguish the emotion of words. we only need to use 5-15 seed words as a training example to learn the subspace, and get the emotional information about the words through the subspace.
\section{12 Future Work}
In future work, we will find the probabilities that other information embedded in one or more orthogonal subspaces rather than just emotional embedding. This decomposition will not change the content of the information embedding, but it will make them more compact, meaningful and easier to interpret for any given application. In addition, we will use the larger dataset for training, because the size of the data determines the improvement of the model's generalization ability.
\bibliographystyle{plain}
| 41761f3cf85f8cdd8d4f25b3a07cceab1106f6cf | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In particle and nuclear physics experiments, large-scale liquid-based detectors have been widely used, particularly in neutrino detectors and in experiments searching for dark matter or neutrino-less double-beta decay. For example, the Super-Kamiokande detector used 50,000 tons of ultra-pure water to detect neutrinos\cite{abe2013evidence}. In addition, Sudbury Neutrino Observatory used 1000 tons of heavy water to detect solar neutrinos\cite{adam2015juno}. In the Jiangmen Underground Neutrino Observatory (JUNO), 20,000 tons of linear alkylbenzene liquid scintillator will be used to detect the reactor neutrinos\cite{adam2015juno}. For the two future flagship neutrino experiments, Hyper-Kamiokande\cite{Abe2018uyc} and the Deep Underground Neutrino Experiment (DUNE)\cite{acciarri2016long} will use 1 million tons of ultrapure water and 40,000 tons of liquid argon, respectively. In addition to neutrino experiments, liquid-based detectors have been used in dark matter studies, such as the Particle and Astrophysical Xenon (PandaX) experiment\cite{cao2014liquid,zhang2019dark} and the Large Underground Xenon (LUX) experiment \cite{akerib2017results}, which use liquid xenon in detecting dark matter candidates as well as searching for neutrino-less nuclear double-beta decay signals.
Liquid has the advantages of a high transparency and flowability, and is easily purified and filled into the detectors, which makes it a good candidate for the design of particle and nuclear physics experiments\cite{juyal2020proportional,huang2020simulation}, particularly for largescale experiments searching for rare signals\cite{yan2020study} with optical processes and requiring low background environments\cite{giboni2020ln2}.
With liquid-based detectors designed to be increasingly larger, as a growing problem, it is becoming more difficult to keep the uniformity of the liquid throughout an extremely large liquid detector. Moreover, owing to the flowability of the liquid, the liquid flow in the detector may change the uniformity of the optical properties of the liquid. For example, the change in refractive index owing to the temperature and pressure difference may lead to a deviation of the optical photon transport in the medium from ideal uniform conditions. The impact of such deviation caused by a detector with a non-uniformity of the physical signal measurements also needs to be studied, particularly in large-scale experiments searching for rare signal events.
To study the liquid flow and non-uniformity in the detector, it is necessary to have a technique to exchange the detector geometry information between different software programs, such as computational fluid dynamics (CFD) software and detector simulation software. A uniform detector description with a detailed geometry at different parts of the liquid detector needs to be provided, and such information needs to be shared between different software with a common interface. Because the geometry information is dynamic owing to a liquid flow, the transformation of the detector description with the interface should be automatically realized.
We propose the development of a method based on the geometry description markup language (GDML)\cite{chytracek2006geometry} and tessellated detector description\cite{agostinelli2003geant4,allison2006geant4} for sharing the dynamic geometry of liquid-based detectors between CFD simulation software and detector simulation software, making it possible to further study the extent of the nonuniformity of the geometry in the CFD and study its impact on the detector performance.
The rest of this paper is structured as follows. In Sect. 2, we introduce the method of sharing the detector geometry between software and the workflow. In Sect. 3, the method is applied to a toy spherical detector model to study the nonuniformity of the refractive index caused by a liquid flow and its influence on the detector simulation and event reconstruction. In Sects. 4 and 5, the performance of the method and its further application are discussed. Finally, a summary and some concluding remarks are provided in Sect. 6.
\section{Methodologies}
In this section, we describe the method of sharing the detector geometry information between different types of software. With a consistent detector description and automatic geometry data conversion through an interface, the change in detector geometry can be studied in CFD simulation software, and its influence on the detector performance can also be considered using detector simulation software and reconstruction software.
A typical application of the method is in an experiment using optical photons to reconstruct the event vertex and energy of the physical signals, in which the density and refractive index of the liquid medium may change in different parts of the detector owing to changes in the temperature and pressure. Consequently, transportation of the optical photons in the medium will also change, leading to deviations in the detector simulation, event reconstruction, and detector performance.
The structure of the detector geometry data flow used in the simulation and a reconstruction applying the method is shown in Fig.~\ref{data_flow}. The entire process can be divided into the following parts: a CFD simulation, geometry data conversion, detector simulation, reconstruction, and a detector performance study.
\begin{figure}
\includegraphics[width=0.8\hsize]{data_flow.pdf}
\caption{Data flow of the detector geometry information from simulation to reconstruction.}
\label{data_flow}
\end{figure}
In this method, a CFD simulation is first conducted to study the flow of fluids in the detector, with the detector geometry model constructed and its initial conditions set. After a CFD simulation, the information of the geometry mesh and fluid properties (temperature, pressure, and density distribution) are exported. Meanwhile, the related physics properties of the fluid and the refractive index distribution can also be calculated and exported. The conversion of geometric information from the CFD to the detector simulation is realized using GDML or the text format interface, including the geometric mesh and fluid property information. Next, a geometry with non-uniform medium properties can be built in the detector simulation software, and the propagation process of the particles in the detector with a nonuniform geometry can be simulated. Finally, by comparing the reconstruction results of the uniform detector simulation, the influence of the geometry change on the detector performance is evaluated.
\subsection{CFD simulation}
CFD uses numerical analysis and data structures to analyze and solve problems related to fluid flows. COMSOL Multiphysics\cite{multiphysics1998introduction} and Open-source Field Operation And Manipulation (OpenFOAM) \cite{jasak2007openfoam} are two popular CFD simulation software programs that have been widely used in scientific computations.
COMSOL uses a finite element analysis to decompose objects into a series of small-volume grids for calculation and approximate real-world physical phenomena, particularly to solve the problem of multiphysics field coupling. In our study, the detector is first modeled and meshed, the appropriate fluid and thermal fields are then constructed, and the initial boundary conditions are set. After a finite element analysis and evolution over time, the temperature, density, and flow velocity distribution in each part of the detector become stable and a steady state can be obtained.
Constructing a mesh of the geometric model is one of the most important steps in a CFD simulation and later studies. The geometric object grid directly determines the solution method of the model and affects the calculation of the problem, including the solution time, the amount of memory required, the interpolation method between the solution nodes, and the accuracy of the solution. The automatic mesh construction function provided by COMSOL can simplify this process; otherwise, a customized mesh needs to be built manually. Fig.~\ref{meshing} shows the description of a sphere with a mesh and a constructive solid geometry (CSG).
\begin{figure}[!htb]
\includegraphics[width=1\hsize]{meshing.pdf}
\caption{Mesh geometry (left) and CSG shape (right) of a sphere.}
\label{meshing}
\end{figure}
OpenFOAMis another popular CFD simulation software. By default, OpenFOAM defines a mesh of arbitrary polyhedral cells in 3D, bounded by arbitrary polygonal faces. The cells can have an unlimited number of faces, and there is no limit to the number of edges or any restriction on its alignment. This type of mesh offers significant freedom in a mesh generation and manipulation, particularly if the geometry of the domain is complex or changes over time. Users can also generate meshes using other packages and convert them into a format that OpenFOAM can use.
\subsection{Geometry data conversion}
In COMSOL, to construct a mesh of the geometry, the entire detector is divided into many tetrahedrons with a certain precision. The fluid property of the center point of each tetrahedron is taken as the nature of the liquid in the entire tetrahedral area. The mesh and fluid property information of each tetrahedron, including temperature, density, and refractive index, must be converted into a format that can be automatically imported into detector simulation software, such as Geant4\cite{agostinelli2003geant4}, to construct the detector geometry.
In Geant4, the shape is implemented as a classG4Tessellated Solid, which can be used to generate a generic solid defined by several facets. With such a definition, a complex geometrical shape bounded with surfaces can be converted into an approximate description with facets of a defined dimension. Two types of facets can be used for the construction of a G4TessellatedSolid: a triangular facet and a quadrangular facet. The meshed geometry can be constructed by defining every face of the tetrahedron. Finally, all tessellated solids form the shape of a complete detector.
There are two different ways to realize this process. One is to first generate a GDML file based on the geometric information exported by COMSOL, and then import the GDML file into Geant4 to construct the geometry. GDML is an XML-based\cite{bray2000extensible} geometry description language used to describe the geometry of detectors in physical simulations, including their positions, rotations, shapes, and materials. It is designed as an application-independent persistent format. As pure XML, GDML can be universally used as a format to interchange the geometry between different applications\cite{wang2020cad}. The GDML geometry can be imported into Geant4 using the GDML parser function to construct the detector in Geant4.
GDML has been successfully applied in many HEP experiments\cite{you2008gdml}, such as JUNO\cite{li2018gdml} and BESIII\cite{liang2009uniform}. A GDML-based geometry management system in JUNO offline software is designed to provide a consistent detector description for different applications\cite{you2018root, zhu2019method}. In BESIII, the detector is described using GDML and then applied in a Geant4-based simulation and ROOT-based reconstruction and visualization\cite{liang2009uniform}.
Another method is to directly read the text format output by COMSOL in a Geant4-based simulation code. A complete set of geometry information-sharing interfaces, from COMSOL to Geant4, has been developed. By using volume classes to construct solids for each tetrahedron according to its vertex coordinates and using the G4Material class to customize the material properties of each tetrahedron, the detector geometry with a non-uniform medium is constructed in Geant4.
\subsection{Detector simulation}
Geant4 is the most popular detector simulation software used in particle and nuclear physics experiments. The simulation process in Geant4 includes the detector construction, particle transport, and interaction with the materials. For a single detector component made up of a uniform material, it is usually described as a single solid volume represented by a CSG shape\cite{dong2020study}. However, for liquid-based detectors, such an approximation will lead to deviation from reality. The level of deviation owing to non-uniformity in large-scale detectors can be too large to be ignored, and thus the difference between a detailed detector description and an approximate uniform description needs to be studied.
To compare the impact of non-uniformity of a refractive index on a detector simulation, only two types of physical processes are included in the detector simulation: photon propagation in a uniform medium and Fresnel refraction at the medium boundary. In the former case, the photon is regarded as a free particle and travels in a straight line throughout the detector. For the latter case, owing to the nonuniform nature of the medium in the detector, the detector has many interfaces. The refractive index on each side of the interface is different, and thus the photon refracts when passing through these interfaces, and the photon path in the detector is composed of a series of steps. As an example, Fig.~\ref{optical_path} shows the refraction of photons in a non-uniform medium.
\begin{figure}[!htb]
\includegraphics[width=0.9\hsize]{optical_path_w.pdf}
\caption{(Color figure online) Schematic of photon transportation in a spherical detector. The bold green line shows the path of a photon with multiple refractions in a non-uniform medium.}
\label{optical_path}
\end{figure}
\subsection{Detector performance study}
With the output from the detector simulation, the next step is to study the impact of the difference in detector description on the detector performance. The detector performance, for example, the resolution of a certain physical measurement, can be determined by analyzing the detector response during the simulation. The description of the detector geometry plays a key role in the simulation. In this study, we test different descriptions of the detector geometry for their detector performance while leaving the other conditions unchanged.
\section{Application}
The idea of the detector geometry sharing method was originally conceived to study the optical photon transportation in liquid-based detectors. Such processes are important in many large-scale liquid-based detector experiments, particularly those searching for rare signal events. For example, in neutrino experiments such as the Daya Bay reactor neutrino experiment\cite{an2013improved}, JUNO, the Kamioka Liquid Scintillator anti-neutrino Detector (KamLAND)\cite{abe2010production}, Hyper-K, and DUNE, optical photons from Cherenkov radiation or scintillation are used to reconstruct the signal events. In a dark matter search or double beta-decay experiments such as PandaX, LUX, Dark matter Experiment with Argon and Pulseshape discrimination 3600 (DEAP-3600)\cite{amaudruz2019design},XENON dark matter search (XENON)\cite{aprile2019light}, and Depleted Argon cryogenic Scintillation and Ionization Detection (DarkSide)\cite{agnes2018low}, photons or electrons are also transported in liquid detectors.
In this section, we describe the use of a toy detector model to demonstrate the feasibility of applying the method in a photon transportation study. A simple spherical volume filled with water is constructed as the detector. Assuming that the temperature distribution is not uniform in the spherical detector, the density of the water in the detector will be different, which makes the water flow. This step is calculated using CFD simulation software COMSOL with a meshed geometry. The refractive index of the water is determined based on the temperature and difference in density. The fluid properties and mesh geometry information are then automatically converted to set up the geometry in Geant4 with tessellated solids.
In a Geant4-based detector simulation, a virtual physical event is generated at a fixed vertex inside the sphere, and multiple photons are emitted from the vertex. The photons propagate in the water detector and finally reach the surface of the sphere. The detector simulation outputs are the hit positions of the photons on the spherical surface, which will be used to reconstruct the position of the event vertex.
To study how a non-uniformity of the liquid can affect the detector performance, the detector is constructed using two methods. One is to construct a single non-uniform water sphere, and the other is to construct the same water sphere with a tessellated geometry, with the fluid property of each tetrahedron set from the CFD simulation result. The simulation and reconstruction output will be different by using the two detector descriptions, from which the deviation and difference in performance can be studied.
\subsection{Detector mesh geometry}
In the CFD simulation, the radius of the sphere is set to 15 m, and the mesh is configured to divide the sphere evenly into approximately 10,000 tetrahedrons. The distribution of the side length and volume size of every tetrahedron is shown in Fig.~\ref{mesh}, respectively. No tetrahedron has a particularly long side length or particularly large volume, which guarantees that the detector is evenly segmented; otherwise, the properties of the medium in the central tetrahedral area cannot be used to approximate the properties of the liquid within the whole tetrahedral area.
\begin{figure}
\includegraphics[width=1\hsize]{mesh.pdf}
\caption{Distribution of the tetrahedron side length and volume size in a meshed sphere.
}
\label{mesh}
\end{figure}
The precision of the detector description depends on the granularity of the tessellated geometry. The more tetrahedrons the sphere is divided into, the more precise the detector is described, and the more computing resources are required in the CFD and detector simulations. Fig.~\ref{mesh_diff_tet} shows the division of the sphere into 1000 tetrahedrons and 100,000 tetrahedrons. In Sect. 4, the dependence of the deviation of the vertex reconstruction on the number of tetrahedrons is studied. The correlations between the program running speed and the number of tetrahedrons are also provided.
\begin{figure}
\centering
\subfigure{
\begin{minipage}[t]{0.5\hsize}
\centering
\includegraphics[width=\hsize]{sphere_small.pdf}
\end{minipage}
}%
\subfigure{
\begin{minipage}[t]{0.5\hsize}
\centering
\includegraphics[width=\hsize]{sphere_big.pdf}
\end{minipage}
}%
\caption{Mesh geometry of the sphere with different granularities of 1000 tetrahedrons (left) and 100,000 tetrahedrons (right).}
\label{mesh_diff_tet}
\end{figure}
\subsection{Fluid properties in CFD simulation}
In the CFD simulation based on COMSOL, a fixed temperature difference is set between the top and bottom of the detector as the boundary conditions. To make the difference caused by temperature more significant and to test the feasibility of the method, a temperature difference of 35$^{\circ}$C is set in a preliminary study. The correlation of the performance difference with temperature will be provided in Sect. 4.
The temperature at the top is set to higher than that at the bottom. A constant heat flux is provided in the top and bottom areas to simulate the temperature difference from the environment.
The simulation results show that when there is a temperature difference between the upper and lower surfaces of the detector, a stable temperature gradient is formed in the medium. Fig.~\ref{simulation_sphere} shows the distribution of fluid temperature and flow velocity at a steady state. The temperature gradient direction is along the Z-axis, and the temperature is approximately the same on the X-Y plane. The medium in the sphere only slowly flows horizontally on the X-Y plane, except for the area near the north and south poles.
\begin{figure}[!htb]
\includegraphics[width=\hsize]{sphere_VandT.pdf}
\caption{Fluid temperature (left) and flow velocity (right) distribution in the spherical detector. }
\label{simulation_sphere}
\end{figure}
The refractive index of a liquid is highly dependent on the temperature and density. According to\cite{schiebener1990refractive}, the refractive index of water in a different grid of the mesh can be calculated using the following equation:
\begin{equation}
\begin{aligned}
\frac{n^2-1}{(n^2+2)\bar\rho} &= a_0 + a_1\bar\rho + a_2\bar T + a_3{\bar \lambda}^2\bar T + a_4/{\bar \lambda}^2 \\ &+ \frac{a_5}{{\bar \lambda}^2-{{\bar \lambda}_{UV}}^2} + \frac{a_6}{{\bar \lambda}^2-{{\bar \lambda}_{IR}}^2} + a_7{\bar\rho}^2
\\
\bar\rho &= \rho/\rho^* , \bar T = T/T^* , \bar \lambda = \lambda/\lambda^*
\end{aligned}
\end{equation}
where $n$ is the refractive index of the liquid, $\rho$ is the density of the liquid, $T$ is the temperature of the liquid, and $\lambda$ is the wavelength of the photon. The other parameters used in the formula are listed in Table ~\ref{parameters}.
\begin{table}[!htb]
\center
\caption{Parameters of the refractive index of water with wavelength, density, and temperature}
\setlength{\tabcolsep}{3mm}{%
\begin{tabular}{@{}lll@{}}
\toprule
\midrule
$a_0=0.244257733$ & & $a_4 = 1.58920570\cdot 10^{-3}$\\
$a_1=9.74634476\cdot 10^{-3}$ & & $a_5=2.45934259\cdot 10^{-3}$ \\
$a_2=-3.73234996\cdot 10^{-3}$ & & $a_6=0.900704920$ \\
$a_3=2.68678472\cdot 10^{-4}$ & & $a_7=-1.66626219\cdot 10^{-2}$ \\
${\bar \lambda}_{UV}=0.2292020$ & & ${\bar \lambda}_{IR}=5.432937$ \\
$\rho^*=1000 \rm kg\cdot m^{-3}$ & & $T^*=273.15 \rm K$ \\
$\lambda^*=0.589 \rm \mu m$ & & \\
\bottomrule
\end{tabular}}%
\label{parameters}
\end{table}
\subsection{Deviation in detector simulation}
The mesh of the spherical detector and its fluid properties in different grids of the mesh, including the density and refractive index, will be converted into a GDML format and text format that can be read automatically by the Geant4 simulation program to construct the detector geometry.
To simulate a physics event at a fixed vertex inside the sphere, 1000 optical photons are generated from the vertex isotropically at the $4\pi$ solid angle. Physically, the number of photons generated corresponds to the amplitude of the physical event, such as the incident particle energy. In the Geant4 simulation, the optical photons are transported in the water sphere through optical processes and finally reach the surface of the sphere. Ideally, we assume that all photons can be recorded as hits at the sphere surface with 100\% detection efficiency. These signals will be used to reconstruct the original vertex of the physics event.
To obtain the simulation deviation, we first simulate using a uniform medium detector to obtain the nominal simulation outputs. Then,we run a second simulation with all parameters unchanged, except for a replacement of a uniform sphere with a tessellated geometry describing a non-uniform medium.
Fig.~\ref{refraction} describes the scenario using the above two geometries in a detector simulation. One uses a uniform material in the detector construction, and the other uses a tessellated geometry for a more precise detector description. For the optical photons crossing the neighboring tessellated volumes with different refractive indices, only the Fresnel refraction is processed.
\begin{figure}[!htb]
\includegraphics[width=1\hsize]{two_simulation.pdf}
\caption{Two scenarios with refraction. The left shows that the photon travels in a straight line in the detector described with a uniform medium. The right shows that the photon refracts multiple times in the detector with a non-uniform medium.}
\label{refraction}
\end{figure}
The magnitude of the photon momentum at the sphere surface is the same as the incident momentum, but the direction of the photon changes in each Fresnel refraction, which is described by the following equations:
\begin{equation}
\begin{aligned}
\boldsymbol{p}_2 = \boldsymbol{p}_1 + | \boldsymbol{p}_1|(\cos\theta_1-\frac{n_2}{n_1}\cos\theta_2)\hat{\boldsymbol{n}}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\cos\theta_1 = \frac{|{\boldsymbol{p}_1} \cdot {\hat{\boldsymbol{n}}}|}{|\boldsymbol{p}_1|} , \sin\theta_2 = \frac{n_1}{n_2}\sin\theta_1
\end{aligned}
\end{equation}
where $n_1$ and $n_2$ are the refractive indices on different sides of the interface, respectively.
$\theta_1$ is the incident angle, and $\theta_2$ is the refraction angle.
$\boldsymbol{p}_1$ is the incident momentum, $\boldsymbol{p}_2$ is the momentum after refraction, and $\hat{\boldsymbol{n}}$ is the normalized normal vector of the refraction plane.
To obtain the relation of simulation deviation with the event vertex, 10,000 events are simulated with vertices randomly distributed in the sphere. Each event is simulated twice with two different detector descriptions to obtain the deviation. We can then compare the photon hit positions on the detector surface owing to the refraction caused by the nonuniformity of the medium.
A comparison of the results from two simulations shows that the hit position on the detector surface is significantly shifted under the preliminary conditions. The deviation is at the millimeter level. Fig.~\ref{arrow_and_proj} shows the distribution of the magnitude and direction of the deviation on the surface of the sphere. The magnitude of deviation has no obvious relationship with the hit position on the sphere. Most of the photon positions on the surface of the sphere deflect toward the negative Z-axis in the simulation with the non-uniform detector, regardless of the original upward or downward direction of the photon from the vertex points. This is because the temperature increases along the positive Z-axis direction and the refractive index decreases along this direction, as shown in Fig.~\ref{simulation_sphere}.
\begin{figure}[!htb]
\centering
\subfigure{
\includegraphics[width=1\hsize]{arrow.pdf}
}
\quad
\subfigure{
\includegraphics[width=1\hsize]{proj.pdf}
}
\caption{(Color figure online) Distribution of the magnitude and direction of the deviation on the detector surface (top).
The blue and red arrows indicate the offset direction along the positive Z-axis and negative Z-axis directions, respectively.
The length of the arrow indicates the relative magnitude of the deviation.
A 2D projection of the magnitude of the offset on the sphere surface is also shown (bottom).}
\label{arrow_and_proj}
\end{figure}
Fig.~\ref{theta_and_alpha} shows the distribution of the deviation angle offset in the longitude ($\theta$) and latitude ($\alpha$) directions, which are defined in the schematic view of the sphere in Fig.~\ref{angle}. Because the temperature on the same $Z$-plane is the same, there is no significant trend in which the photon hits with the same $Z$ on the surface deflect systematically toward the left or right.
\begin{figure}[!htb]
\centering
\subfigure{
\includegraphics[width=1\hsize]{theta.pdf}
}
\quad
\subfigure{
\includegraphics[width=1\hsize]{alpha.pdf}
}
\caption{Distribution of the deviation angle $\theta$ (top) and $\alpha$ (bottom) with the Z-coordinate.}
\label{theta_and_alpha}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.6\hsize]{angle.pdf}
\caption{Schematic view of the deviation angle on sphere surface. Here, $\theta$ and $\alpha$ represent the deviation angle component along the longitude and latitude directions, respectively.}
\label{angle}
\end{figure}
\subsection{Deviation in vertex reconstruction}
In most experiments, the event vertex position, which shows where the physics event is generated, is an important physics quantity to measure. The resolution of the vertex position is a key indicator for evaluating the detector performance. To study the effect of the non-uniformity of the medium properties on the resolution of the detector vertex position, a charge-weighted algorithm is used to reconstruct the event vertex from the hit position of the photons on the sphere surface.
Fig.~\ref{cc_principle} shows the principle of the charge-weighted algorithm. Assuming that the true event vertex is located at the position $Z=z_0$ and that there are $n$ photons emitted from the vertex, whose momentum directions are isotropically distributed in the $4\pi$ solid angle of the vertex, the average value of the $z$ component of the hit position of all photons on the sphere is
\begin{equation}
\begin{aligned}
\bar{z} &= \frac{1}{4\pi}\int z\,d\Omega
\\ &= \frac{1}{4\pi}\int_0^{2\pi}\,d\phi \int_0^{\pi} (z_0 + r\cos\theta)\sin\theta\,d\theta
\\ &= \frac{2}{3}z_0
\end{aligned}
\end{equation}
The reconstructed event vertex $\bar{\boldsymbol{V}}$ can then be obtained as follows:
\begin{equation}
\bar{\boldsymbol{V}} = \frac{3}{2n}\sum_i \boldsymbol{V}_i
\end{equation}
where $n$ is the total number of photons and $\boldsymbol{V}_i$ is the hit position of the $i$-th photon on the sphere.
\begin{figure}[!htb]
\includegraphics[width=0.6\hsize]{ccMethod.pdf}
\caption{Sketch of the charge-weighted algorithm for vertex reconstruction.}
\label{cc_principle}
\end{figure}
To compare the event vertex reconstruction deviation, we simulate 10,000 events. Each event was simulated twice, with a uniform and non-uniform detector geometry, respectively. The two simulation results are reconstructed with the same charge-weighted algorithm. We then compare the two reconstructed vertices to obtain the deviation in this event owing to the difference in geometry. Fig.~\ref{bias_v} shows the distribution of reconstructed vertex deviation for the 10,000 events. The deviations owing to the difference in geometry are a few millimeters in length with a long tail of up to 80 mm. The average deviation is approximately 10 mm. Although the charge-weighted algorithm has a vertex reconstruction resolution itself, in the two simulations of every event, every original condition is the same except for the detector geometry. Therefore, the comparison is between the two reconstructed vertices rather than their respective bias from the true vertex.
\begin{figure}[!htb]
\includegraphics[width=0.8\hsize]{diff_V.pdf}
\caption{The vertex reconstruction deviation owing to difference in geometry.}
\label{bias_v}
\end{figure}
In this section, we use a set of fixed-parameters (sphere radius = 15 m, temperature difference = 35$^{\circ}$C, and sphere mesh number = 10,000) to obtain preliminary results showing the feasibility of applying the geometry sharing method to a liquid-based detector simulation and reconstruction, and to obtain the deviation distribution.
\section{Comparison}
To study the influence of different conditional parameters on the detector performance, we change the size of the detector, the temperature difference, and the mesh granularity to check the consequential changes in the simulation and vertex reconstruction results.
\begin{figure}[!htb]
\includegraphics[width=0.8\hsize]{diff_vs_diff_T.pdf}
\caption{Average deviation of the reconstructed vertex changes with the temperature difference.}
\label{diff_dT}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.8\hsize]{diff_vs_R.pdf}
\caption{Average deviation of the reconstructed vertex changes with the radius of the sphere.}
\label{diff_R}
\end{figure}
Under real experiment conditions, the environment temperature could be different. Usually, the temperature difference is at a few degrees Celsius. The higher it is, the more significant its impact on the detector performance. Fig.~\ref{diff_dT} shows that the reconstructed vertex deviation is linearly dependent on the temperature difference. Fig.~\ref{diff_R} shows that the reconstructed vertex deviation is linear with the size of the detector. These conclusions are expected because the method itself is size independent.
Fig.~\ref{diff_tetnum} shows the dependency of the vertex reconstruction deviation on the mesh granularity, \emph{i.e.}, the number of tetrahedrons that a sphere is segmented into. When the segmentation is extremely coarse, the deviations of the reconstructed vertex gradually increase with the improvement in the segmentation accuracy. After the number of tetrahedrons reaches 5000, the deviation becomes stable. To determine the granularity of the detector segmentation, it is necessary to comprehensively consider the available computing resources and resulting accuracy requirements.
\begin{figure}[!htb]
\includegraphics[width=0.8\hsize]{diff_vs_tetnum.pdf}
\caption{The average deviation of the reconstructed vertex changes with the segmentation accuracy.}
\label{diff_tetnum}
\end{figure}
Fig.~\ref{diff_angle} shows the deviation dependency on the heat flux area angle. When the heat flux area is close to the top or bottom of the sphere, the deviation of the reconstruction vertex is slightly affected by the heat flux location change, but when the location deviates greatly from the $Z$-axis by nearly $90^{\circ}$, the deviation of the reconstructed vertex is significantly reduced.
\begin{figure}[!htb]
\includegraphics[width=0.8\hsize]{diff_vs_angle.pdf}
\caption{The average deviation of the reconstructed vertex changes with the heat flux area angle.}
\label{diff_angle}
\end{figure}
As the reason for the influence of the location of the heat flux area, the formation of the temperature gradient has a strong correlation with the location of the heat flux area. Fig.~\ref{diff_angle_v} shows the temperature distribution that reaches stability when the heat flux region deviates from the $Z$-axis by $30^{\circ}$ and $150^{\circ}$, respectively, in which the red and blue areas on the surface represent hot and cold areas where a heat flux exists. When the heat flux area deviates from the $Z$-axis by $30^{\circ}$, the medium heated in the hot area increases along the ball wall to the top and decelerates, and then flows back to the hot area along the horizontal ball wall. A similar case occurs in the cold area. Therefore, the liquid in the sphere has almost no vertical flow, but mainly a horizontal flow.
\begin{figure}[!htb]
\includegraphics[width=1\hsize]{diff_deg.pdf}
\caption{The distributions of temperature and liquid flow in the sphere when the CFD simulation reaches stability for the two cases when the heat flux area deviates from the $Z$-axis by $30^{\circ}$ (left) and by $150^{\circ}$ (right), respectively.}
\label{diff_angle_v}
\end{figure}
Because there is only local circulation, when the heat flux area deviates from the $Z$-axis by an angle that is not too large, a temperature gradient similar to the case in which the heat flux area is on the $Z$-axis can still be formed. When the heat flux area deviates far from the $Z$-axis, the temperature difference between the upper and lower endpoints along the temperature gradient direction under a stable state becomes smaller.
Fig.~\ref{t_vs_z} shows the temperature distribution along the $Z$-axis at a steady state. When the temperature difference of the heat flux is 35 K and the heat flux location deviates from the $Z$-axis by $60^{\circ}$, the temperature difference is comparable to the case in which the temperature difference of the heat flux is 15 K and the position is along the $Z$-axis, although the temperature gradient of the former is greater. However, when the heat flux area deviates from the $Z$-axis by more than $90^{\circ}$ (Fig.~\ref{diff_angle_v}, right), the heated liquid flows upward from the hot area, and when it reaches the cold area on the top, the liquid is quickly cooled and sinks. Therefore, it forms an annular flow in the sphere, and no obvious temperature gradient can be formed owing to the circulation.
\begin{figure}[!htb]
\includegraphics[width=0.9\hsize]{t_vs_z.pdf}
\caption{The temperature distribution in the sphere along the $Z$-axis with different ${\Delta}T$ and different heat flux angle.}
\label{t_vs_z}
\end{figure}
The consumption of computing resources is also compared in this study. When the number of tetrahedrons increases with the improvement in the segmentation accuracy, it takes more CPU time in the COMSOL and Geant4 simulations. The simulation times with different numbers of tetrahedrons are listed in Table ~\ref{cpu_time}, which is the CPU time required for COMSOL to simulate a fluid flow for a 5 h clock time and for Geant4 to simulate 100,000 particles. The amount of CPU time used by the two geometry data conversion methods based on the text format and GDML is also compared. In the Geant4 simulation, it takes more time to import the geometry through a GDML file than directly reading the text file to construct the geometry. With an increase in the segmentation accuracy, the difference in the computing speed between the two methods becomes more obvious.
\begin{table}[!htb]
\caption{CPU time consumption for each part of the simulation}
\begin{tabular}{ccccc}
\hline
\multirow{3}{*}{tetrahedron numbers} & \multicolumn{3}{c}{Geant4} & \multirow{3}{*}{COMSOL} \\ \cline{2-4}
& \multicolumn{2}{l}{Geometry initialization} & \multirow{2}{*}{simulation} & \\ \cline{2-3}
& Text & GDML & & \\ \hline
1,000 & \textless{}1s & 8s & 27s & 21min \\
10,000 & 2s & 13s & 58s & 137min \\
30,000 & 7s & 44s & 85s & 321min \\ \hline
\end{tabular}
\label{cpu_time}
\end{table}
\section{Discussion}
With this method, the generation of a correct geometry mesh is an important step. The entire detector must be fully segmented; otherwise, the properties of the tetrahedral central medium cannot be used to approximate the properties of the entire tetrahedron. In this study, only the method of a global automatic segmentation is adopted, and only a tetrahedron segmentation is considered.
At the same time, the influence of medium non-uniformity will reach saturation with an improvement in the segmentation accuracy. An excessively high segmentation accuracy does not significantly improve the simulation and reconstruction precision but will greatly increase the CPU time consumption in the CFD and Geant4 simulations. In more realistic detector simulations, more complicated physical processes, more particles, and more complicated environments are usually introduced, which may further increase the simulation time.
The GDML and text format are both used for geometry data sharing from COMSOL to Geant4. Creating a geometry by directly reading the text files is much faster than importing GDML files into Geant4 to initialize the geometry. However, GDML is widely used in a Geant4 simulation and reconstruction based on the ROOT\cite{brun1997root} geometry package, which can provide consistent geometric information for a subsequent data analysis.
In our research, it is assumed that the hot area is on the top and the cold area is at the bottom, and thus a temperature gradient can be formed. When the location of the hot area is lower than that of the cold area because of the formation of convection and circulation in the sphere, the temperature in the sphere is almost the same and no temperature gradient can be formed. The entire spherical detector may be considered to be uniform, and no obvious deviations in simulation and reconstruction will be observed.
\section{Conclusion}
A method of dynamic geometry information sharing between a CFD simulation and a detector simulation is proposed for liquid-based detectors. Its feasibility is demonstrated by applying the method to a simulation with a non-uniform medium to study the photon transport and a deviation in the event of a vertex reconstruction. This method can also be used to study other dynamic geometry-related problems in particle and nuclear physics experiments, such as the expansion and contraction of the detector volumes owing to a temperature change, detector alignment at different running periods of the experiments, and geometry-related changes to the magnetic field.
\subparagraph{Author contributions} All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Shu Zhang, Jing-Shu Li, Yang-Jie Su, and Zi-Yuan Li. The first draft of the manuscript was written by Shu Zhang, Yu-Mei Zhang, and Zheng-Yun You, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.
| 9c205dbc540128078f630c0ed8483cff25f20edd | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Graph, as a complex data structure, consists of nodes (or vertices) and edges (or links). It can be used to model lots of complex systems in real world, e.g. social networks, protein-protein interaction networks,
brain networks, road networks, physical interaction networks and knowledge graph etc. Thus, Analyzing the complex networks becomes an intriguing research frontier.
With the rapid development of deep learning techniques, many scholars employ the deep learning architectures to tackle the graphs.
Graph Neural Networks (GNNs) emerge under these circumstances. Up to now, the GNNs have evolved into a prevalent and powerful computational framework for tackling irregular data such as graphs and manifolds.
The GNNs can learn task-specific node/edge/graph representations via hierarchical iterative operators so that the traditional machine learning methods can be employed
to perform graph-related learning tasks, e.g. node classification, graph classification, link prediction and clustering etc. Although the GNNs has attained substantial success over the graph-related learning tasks,
they still face great challenges. Firstly, the structural complexity of graphs incurs expensive computational cost on large graphs.
Secondly, perturbing the graph structure and/or initial features incurs sharp performance decay. Thirdly, the Wesfeiler-Leman (WL) graph isomorphism test impedes the performance improvement of the GNNs.
At last, the blackbox work mechanism of the GNNs hinders safely deploying them to real-world applications.
In this paper, we generalize the conventional deep architectures to the non-Euclidean domains,
and summarize the architectures, extensions and applications, benchmarks and evaluation pitfalls and future research directions of the graph neural networks.
Up to now, there have been several surveys on the GNNs. However, they usually discuss the GNN models from different angles and with different emphasises.
To the best of our knowledge, the first survey on the GNNs was conducted by Michael M. Bronstein et al\cite{1}.
Peng Cui et al\cite{2} reviewed different kinds of deep learning models applied to graphs from three aspects: semi-supervised learning methods including
graph convolutional neural networks, unsupervised learning methods including graph auto-encoders, and recent advancements including graph recurrent neural networks and graph reinforcement learning.
This survey laid emphasis on semi-supervised learning models, i.e. the spatial and spectral graph convolutional neural networks, yet comparatively less emphasis on the other two aspects.
Due to the space limit, this survey only listed a few of key applications of the GNNs, but ignored the diversity of the applications.
Maosong Sun et al\cite{3} provided a detailed review of the spectral and spatial graph convolutional neural networks from three aspects: graph types, propagation step and training method,
and divided its applications into three scenarios: structural scenarios, non-structural scenarios and other scenarios. However, this article did not involve the other GNN architectures
such as graph auto-encoders, graph recurrent neural networks and graph generative networks.
Philip S. Yu et al\cite{4} conducted a comprehensive survey on the graph neural networks, and investigated available datasets, open-source implementations and practical applications.
However, they only listed a few of core literatures on each research topic.
Davide Bacciu et al\cite{367} gives a gentle introduction to the field of deep learning for graph data. The goal of this article is to introduce the main concepts
and building blocks to construct neural networks for graph data, and therefore it falls short of an exposition of recent works on graph neural networks.
\begin{figure}[htb]
\centering
\includegraphics[width=0.9\textwidth]{GNN-Panorama.pdf}
\caption{The architecture of this paper.}
\label{GNN Panorama}
\end{figure}
It is noted that all of the aforementioned surveys do not concern capability and interpretability of GNNs, combinations of the probabilistic inference and GNNs,
and adversarial attacks on graphs. In this article, we provide a panorama of GNNs for readers from 4 perspectives:
architectures, extensions and applications, benchmarks and evaluations pitfalls, future research directions,
as shown in Fig. \ref{GNN Panorama}. For the architectures of GNNs, we investigate the studies on graph convolutional neural networks (GCNNs), graph pooling operators, graph attention mechanisms and
graph recurrent neural networks (GRNNs). The extensions and applications demonstrate some notable research topics on the GNNs through integrating the above architectures.
Specifically, this perspective includes the capabilities and interpretability, deep graph representation learning, deep graph generative models, combinations of the Probabilistic Inference (PI) and the GNNs,
adversarial attacks for GNNs, Graph Neural Architecture Search and graph reinforcement learning and applications. In summary, our article provides a complete taxonomy for GNNs,
and comprehensively review the current advances and trends of the GNNs. These are our main differences from the aforementioned surveys.
\textbf{Contributions.} Our main contributions boils down to the following three-fold aspects.
\begin{enumerate}
\item We propose a novel taxonomy for the GNNs, which has three levels. The first includes architectures, benchmarks and evaluation pitfalls, and applications. The architectures are classified into 9 categories, the benchmarks and evaluation pitfalls into 2 categories, and the applications into 10 categories. Furthermore, the graph convolutional neural networks, as a classic GNN architecture, are again classified into 6 categories.
\item We provide a comprehensive review of the GNNs. All of the literatures fall into the corresponding categories. It is expected that the readers not only understand the panorama of the GNNs, but also comprehend the basic principles and various computation modules of the GNNs through reading this survey.
\item We summarize four future research directions for the GNNs according to the current facing challenges, most of which are not mentioned the other surveys. It is expected that the research on the GNNs can progress into a new stage by overcoming these challenges.
\end{enumerate}
\textbf{Roadmap.} The remainder of this paper is organized as follows. First of all, we provide some basic notations and definitions that will be often used in the following sections.
Then, we start reviewing the GNNs from 4 aspects: architectures in section 3, extensions and applications in section 4, benchmarks and evaluation pitfalls in section 5
and future research directions in section 6. Finally, we conclude our paper.
\section{Preliminaries}
In this section, we introduce relevant notations so as to conveniently describe the graph neural network models.
A simple graph can be denoted by $G=(V,E)$ where $V$ and $E$ respectively denote the set of $N$ nodes (or vertices) and $M$ edges.
Without loss of generality, let $V=\left\{v_1,\cdots,v_N\right\}$ and $E=\left\{e_1,\cdots,e_M\right\}$. Each edge $e_j\in E$ can be denoted by $e_j=\left(v_{s_j},v_{r_j}\right)$ where $v_{s_j}, v_{r_j}\in V$.
Let $A_G$ denote the adjacency matrix of $G$ where $A_G(s,r)=1$ iff there is an edge between $v_s$ and $v_r$. If $G$ is edge-weighted, $A_G(s,r)$ equals the weight value of the edge $\left(v_s,v_r\right)$.
If $G$ is directed, $\left(v_{s_j},v_{r_j}\right)\neq\left(v_{r_j}, v_{s_j}\right)$ and therefore $A_G$ is asymmetric.
A directed edge $e_j=\left(v_{s_j},v_{r_j}\right)$ is also called an arch, i.e. $e_j=\left\langle v_{s_j},v_{s_j}\right\rangle$.
Otherwise $\left(v_{s_j},v_{r_j}\right)=\left(v_{r_j}, v_{s_j}\right)$ and $A_G$ is symmetric. For a node $v_s\in V$, let $N_G(v_s)$ denote the set of neighbors of $v_s$, and $d_G(v_s)$ denote the degree of $v_s$.
If $G$ is directed, let $N_G^+(v_s)$ and $N_G^-(v_s)$ respectively denote the incoming and outgoing neighbors of $v_s$, and $d_G^+(v_s)$ and $d_G^-(v_s)$ respectively denote the incoming and outgoing degree of $v_s$.
Given a vector $a=(a_1,\cdots,a_N)\in\mathbb{R}^N$, $\text{diag}(a)$ (or $\text{diag}(a_1,\cdots,a_N)$) denotes a diagonal matrix consisting of the elements $a_n, n=1,\cdots,N$.
A vector $x\in\mathbb{R}^{N}$ is called a 1-dimensional graph signal on $G$. Similarly, $X\in\mathbb{R}^{N\times d}$ is called a $d\text{-dimensiaonl}$ graph signal on $G$.
In fact, $X$ is also called a feature matrix of nodes on $G$. Without loss of generality, let $X[j,k]$ denote the $(j,k)\text{-th}$ entry of the matrix $X\in\mathbb{R}^{N\times d}$,
$X[j,:]\in\mathbb{R}^d$ denote the feature vector of the node $v_j$ and $X[:,j]$ denote the $1\text{-dimensional}$ graph signal on $G$.
Let $\mathbb{I}_N$ denote a $N\times N$ identity matrix. For undirected graphs, $L_G=D_G-A_G$ is called the Laplacian matrix of $G$,
where $D_G[r,r]=\sum_{c=1}^N A_G[r,c]$. For a 1-dimensional graph signal $x$, its smoothness $s(x)$ is defined as
\begin{equation}\label{SmoothnessOfGraphSignal}
s(x)=x^TL_Gx=\frac{1}{2}\sum_{r,c=1}^NA_G(r,c)\left(x[r]-x[c]\right)^2.
\end{equation}
The normalization of $L_G$ is defined by $\overline{L}_G=\mathbb{I}_N-D_G^{-\frac{1}{2}}A_GD_G^{-\frac{1}{2}}$.
$\overline{L}_G$ is a real symmetric semi-positive definite matrix. So, it has $N$ ordered real non-negative eigenvalues $\left\{\lambda_n:n=1,\cdots,N\right\}$
and corresponding orthonormal eigenvectors $\left\{u_n:n=1,\cdots,N\right\}$, namely $\overline{L}_G=U\Lambda U^T$ where $\Lambda=\text{diag}(\lambda_1,\cdots,\lambda_N)$
and $U=\left(u_1,\cdots,u_N\right)$ denotes a orthonormomal matrix. Without loss of generality, $0=\lambda_1\leq\lambda_2\leq\cdots\leq\lambda_N=\lambda_{\text{max}}$.
The eigenvectors $u_n, n=1,\cdots,N$ are also called the graph Fourier bases of $G$. Obviously, the graph Fourier basis are also the 1-dimensional graph signal on $G$.
The graph Fourier transform\cite{5} for a given graph signal $x$ can be denoted by
\begin{equation}\label{GraphFourierTransform}
\hat{x}\triangleq\mathcal{F}(x)=U^Tx.
\end{equation}
The inverse graph Fourier transform can be correspondingly denoted by
\begin{equation}\label{InverseGraphFourierTransform}
x\triangleq\mathcal{F}^{-1}(\hat{x})=U\hat{x}.
\end{equation}
Note that the eigenvalue $\lambda_n$ actually measures the smoothness of the graph Fourier mode $u_n$.
Throughout this paper, let $\rho(\cdot)$ denote an activation function, $\bowtie$ denote the concatenation of at least two vectors, and $\left<\right>$ denote the inner product of two vectors/matrices.
We somewhere use the function $\text{Concat}(\cdot)$ to denote the concatenation of two vectors as well.
\section{Architectures}
\subsection{Graph Convolutional Neural Networks (GCNNs)}
The GCNNs play pivotal roles on tackling the irregular data (e.g. graph and manifold). They are motivated by the Convolutional Neural Networks (CNNs) to learn hierarchical representations of irregular data.
There have been some efforts to generalize the CNN to graphs \cite{27,31,48}. However, they are usually computationally expensive and cannot capture spectral or spatial features.
Below, we introduce the GCNNs from the next 6 aspects: spectral GCNNs, spatial GCNNs, Graph wavelet neural networks and GCNNs on special graphs.
\subsubsection{Spectral Graph Convolution Operators}
\begin{figure}[htb]
\centering
\includegraphics[width=1.0\textwidth]{Spectral-GCNN.pdf}
\caption{Computational framework of the spectral GCNN.}
\label{Spectral GCNN}
\end{figure}
The spectral graph convolution operator is defined via the graph Fourier transform. For two graph signals $x$ and $y$ on $G$,
their spectral graph convolution $x\ast_Gy$ is defined by
\begin{equation}\label{Spectral Graph Convolution}
\begin{array}{rcl}
x\ast_G y & = & \mathcal{F}^{-1}(\mathcal{F}(x)\circledast\mathcal{F}(y)) \\ [2mm]
& = & U\left(U^Tx\circledast U^Ty\right) \\[2mm]
& = & U\text{diag}(U^Ty)U^Tx,
\end{array}
\end{equation}
where $\circledast$ denotes the element-wise Hadamard product \cite{7,18,27}.
The spectral graph convolution can be rewritten as
\begin{displaymath}
x\ast_G f_{\theta}=Uf_{\theta}U^Tx,
\end{displaymath}
where $f_{\theta}$ is a diagonal matrix consisting of the learnable parameters.
That is, the signal $x$ is filtered by the spectral graph filter (or graph convolution kernel) $f_{\theta}$.
For a $d^{(l)}\text{-dimensional}$ graph signal $X^{(l)}$ on $G$,
the output $X^{(l+1)}$ yielded by a graph convolution layer, namely $d^{(l+1)}\text{-dimensional}$ graph signal on $G$,
can be written as
\begin{equation}
\label{Spectral GCNN Formula}
X^{(l+1)}[:,k]=\rho\left(\sum_{j=1}^{d^{(l)}}Uf_{\theta,j,k}^{(l)}U^TX^{(l)}[:,j]\right),
\end{equation}
where $f_{\theta,j,k}^{(l)}$ is a spectral graph filter, i.e. a $N\times N$ diagonal matrix consisting of learnable parameters corresponding to the $j\text{-th}$ graph signal
at $l\text{-th}$ layer and the $k\text{-th}$ graph signal at $(l+1)\text{-th}$ layer.
The computational framework of the spectral GCNN in Eq. (\ref{Spectral GCNN Formula}) is demonstrated in Fig. \ref{Spectral GCNN}.
It is worth noting that the calculation of the above graph convolution layer takes $O(N^3)$ time and $O(N^2)$ space
to perform the eigendecomposition of $\overline{L}_G$ especially for large graphs.
The article \cite{179} proposes a regularization technique, namely GraphMix, to augment the vanilla GCNN with a parameter-sharing Fully Connected Network (FCN).
\textbf{Spectral Graph Filter.} Many studies \cite{27} focus on designing different spectral graph filters.
In order to circumvent the eigendecomposition, the spectral graph filter $f_{\theta}$ can formulated as a $K\text{-localized}$ polynomial of the eigenvalues
of the normalized graph Laplacian $\overline{L}_G$ \cite{6,8,78},
i.e.
\begin{equation}\label{Chebyshev Spectral Filter}
f_{\theta}=f_{\theta}(\Lambda)\triangleq\sum_{k=0}^{K-1}\theta_k\Lambda^k.
\end{equation}
In practice, the $K\text{-localized}$ Chebyshev polynomial \cite{78} is a favorable choice of formulating the spectral graph filter, i.e.
\begin{displaymath}
f_{\theta}(\Lambda)=\sum_{k=0}^{K-1}\theta_kT_k(\widetilde{\Lambda}),
\end{displaymath}
where the Chebyshev polynomial is defined as
\begin{equation}\label{Chebeshev Polynomial}
T_0(x)=1,\quad T_1(x)=x,\quad T_k(x)=2xT_{k-1}(x)-T_{k-2}(x)
\end{equation}
and $\widetilde{\Lambda}=\frac{2}{\lambda_{\text{max}}}\Lambda-\mathbb{I}_N$.
The reason why $\widetilde{\Lambda}=\frac{2}{\lambda_{\text{max}}}\Lambda-\mathbb{I}_N$ is because it can map eigenvalues $\lambda\in\left[0,\lambda_{\text{max}}\right]$ into $[-1,1]$.
This filter is $K\text{-localized}$ in the sense that it leverages information from nodes which are at most $K\text{-hops}$ away.
In order to further decrease the computational cost, the $1\text{st-order}$ Chebyshev polynomial is used to define the spectral graph filter. Specifically, it lets $\lambda_{\text{max}}\approx 2$
(because the largest eigenvalue of $\overline{L}_G$ is less than or equal to 2 \cite{374}) and $\theta=\theta_0=-\theta_1$.
Moreover, the renormalization trick is used here to mitigate the limitations of the vanishing/exploding gradient,
namely substituting $\widetilde{D}_G^{-\frac{1}{2}}\widetilde{A}_G\widetilde{D}_G^{-\frac{1}{2}}$ for $\mathbb{I}_N+D_G^{-\frac{1}{2}}A_GD_G^{-\frac{1}{2}}$ where $\widetilde{A}_G=A_G+\mathbb{I}_N$
and $\widetilde{D}_G=\text{diag}\left(\sum_{k=1}^{N}\widetilde{A}[1,k],\cdots,\sum_{k=1}^{N}\widetilde{A}[N,k]\right)$.
As a result, the Graph Convolutional Network (GCN) \cite{6,375} can be defined as
\begin{equation}\label{GCN-Layer}
X^{(l+1)}=\rho\left(\widetilde{D}_G^{-\frac{1}{2}}\widetilde{A}_G\widetilde{D}_G^{-\frac{1}{2}}X^{(l)}\Theta^{(l)}\right).
\end{equation}
The Chebyshev spectral graph filter suffers from a drawback that the spectrum of $\overline{L}_G$ is linearly mapped into $\left[-1,1\right]$.
This drawback makes it hard to specialize in the low frequency bands.
In order to mitigate this problem, Michael M. Bronstein et al \cite{12} proposes the Cayley spectral graph filter
via the $\text{order-}r$ Cayley polynomial $f_{c,h}(\lambda)=c_0+2\text{Re}\left(\sum_{j=0}^{r}c_h\mathcal{C}(\lambda)^j\right)$ with
the Cayley transform $\mathcal{C}(\lambda)=\frac{\lambda-i}{\lambda+i}$.
Moreover, there are many other spectral graph filters, e.g. \cite{8,14,19,22,30,37,38,39,44,90}.
In addition, some studies employ the capsule network \cite{400} to construct capsule-inspired GNNs \cite{52,53,54}.
\textbf{Overcoming Time and Memory Challenges.} A chief challenge for GCNNs is that their training cost is strikingly expensive, especially on huge and sparse graphs.
The reason is that the GCNNs require full expansion of neighborhoods for the feed-forward computation of each node,
and large memory space for storing intermediate results and outputs.
In general, two approaches, namely sampling \cite{9,10,11,25} and decomposition \cite{23,24}, can be employed to mitigate the time and memory challenges for the spectral GCNNs.
\textbf{Depth Trap of Spectral GCNNs.} A bottleneck of GCNNs is that their performance maybe decease with ever-increasing number of layer.
This decay is often attributed to three factors: (1) overfitting resulting from the ever-increasing number of parameters;
(2) gradient vanishing/explosion during training; (3) oversmoothing making vertices from different clusters more and more indistinguishable.
The reason for oversmoothing is that performing the Laplacian smoothing many times forces the features of vertices
within the same connected component to stuck in stationary points \cite{16}.
There are some available approaches, e.g. \cite{33,373,379,398,399}, to circumvent the depth trap of the spectral GCNNs.
\subsubsection{Spatial Graph Convolution Operators}
Original spatial GCNNs \cite{47,76,77,380} constitutes a transition function, which must be a contraction map in order to ensure the uniqueness of states, and an update function.
In the following, we firstly introduce a generic framework of the spatial GCNN, and then investigate its variants.
Graph networks (GNs) as generic architectures with relational inductive bias \cite{85} provide an elegant interface
for learning entities, relations and structured knowledge.
Specifically, GNs are composed of GN blocks in a sequential, encode-process-decode or recurrent manner. GN blocks contain three kinds of update functions,
namely $\phi^{e}(\cdot),\phi^{v}(\cdot),\phi^{u}(\cdot)$, and three kinds of aggregation functions, namely $\psi^{e\rightarrow v}(\cdot), \psi^{e\rightarrow u}(\cdot), \psi^{v\rightarrow u}(\cdot)$.
The iterations are described as follows.
\begin{equation}\label{GN-BLOCK}
\begin{array}{lll}
e'_k=\phi^{e}\left(e_k,v_{r_k},v_{s_k},u\right), & \bar{e}'_i=\psi^{e\rightarrow v}\left(E'_i\right), & v'_i=\phi^{v}\left(v_i,\bar{e}'_i,u\right), \\[2mm]
\bar{e}'=\psi^{e\rightarrow u}\left(E'\right), & u'=\phi^{u}\left(u,\bar{e}',\bar{v}'\right), & \bar{v}'=\psi^{v\rightarrow u}\left(V'\right)
\end{array}
\end{equation}
where $e_k$ is an arch from $v_{s_k}$ to $v_{r_k}$, $E'_i=\left\{(e'_k,s_k,r_k):r_k=i,k=1,\cdots,M\right\}$, $V'=\left\{v'_i:i=1,\cdots,N\right\}$ and $E'=\left\{(e'_k,s_k,r_k):k=1,\cdots,M\right\}$,
see Fig. \ref{Spatial GCNN}. It is noted that the aggregation functions should be invariant to any permutations of nodes or edges.
In practice, the GN framework can be used to implement a wide variety of architectures in accordance with three key design principles,
namely flexible representations, configuable within-block structure and flexible multi-block architectures.
Below, we introduce three prevalent variants of the GNs, namely Message Passing Neural Networks (MPNNs) \cite{93}, Non-local Neural Networks (NLNNs) \cite{102} and GraphSAGE \cite{104}.
\begin{figure}[htb]
\centering
\includegraphics[width=1.0\textwidth]{Spatial-GCNN.pdf}
\caption{Computational framework of the spatial GCNN.}
\label{Spatial GCNN}
\end{figure}
\textbf{Variants of GNs------MPNNs.} MPNNs \cite{93} have two phases, a message passing phase and a readout phase. The message passing phase is defined by a message function $M_l$
(playing the role of the composition of the update function $\psi^{e\rightarrow v}(\cdot)$ and the update function $\phi^{e}(\cdot)$)
and a vertex update function $U_l$ (playing the role of the update function $\phi^{v}(\cdot)$). Specifically,
\begin{displaymath}
m_v^{(l+1)}=\displaystyle\sum_{u\in N_G(v)}M_l\left(x_v^{(l)},x_u^{(l)},e_{v,u}\right),\quad x_v^{(l+1)}=U_l\left(x_v^{(l)},m_v^{(l+1)}\right)
\end{displaymath}
where $e_{v,u}$ denotes the feature vector of the edge with two endpoints $v$ and $u$.
The readout phase computes a universal feature vector for the whole graph using a readout function $R(\cdot)$, i.e. $u=R\left(\left\{x_v^{(L)}:v\in V\right\}\right)$.
The readout function $R(\cdot)$ should be invariant to permutations of nodes.
A lot of GCNNs can be regarded as special forms of the MPNN, e.g. \cite{103,122,336,195}.
\textbf{Variants of GNs------NLNNs.} NLNNs \cite{102} give a general definition of non-local operations \cite{381} which is a flexible building block and can be easily integrated into convolutional/recurrent layers.
Specifically, the generic non-local operation is defined as
\begin{equation}\label{NLNN}
y_s=\frac{1}{\mathcal{C}(x_s)}\sum_{t}f(x_s,x_t)g(x_t),
\end{equation}
where $f(\cdot,\cdot)$ denotes the affinity between $x_s$ and $x_t$, and $\mathcal{C}(x_s)=\sum_{t}f(x_s,x_t)$ is a normalization factor.
The affinity function $f(\cdot,\cdot)$ is of the following form
\begin{enumerate}
\item Gaussian: $f(x_s,x_t)=e^{x_s^Tx_t}$;
\item Embedded Gaussian: $f(x_s,x_t)=e^{\theta(x_s)^T\eta(x_t)}$, where $\theta(x_s)=W_{\theta}x_s$ and $\eta(x_t)=W_{\eta}x_t$;
\item Dot Product: $f(x_s,x_t)=\theta(x_s)^T\eta(x_t)$;
\item Concatenation: $f(x_s,x_t)=\text{ReLU}(w_f^T\left[\theta(x_s),\eta(x_t)\right])$.
\end{enumerate}
The non-local building block is defined as $z_s=W_zy_s+x_s$ where "$+x_s$" denotes a residual connection.
It is noted that $f(\cdot,\cdot)$ and $g(\cdot)$ play the role of $\phi^e(e_k,v_{r_k},v_{s_k},u)$, and the summation in Eq. (\ref{NLNN}) plays the role of $\psi^{e\rightarrow v}(E'_i)$.
\textbf{Variants of GNs------GraphSAGE.} GraphSAGE (\textsc{SAmple} and \textsc{aggreGatE}) \cite{104} is a general inductive framework capitalizing on node feature information to efficiently generate node embedding vectors for previously unseen nodes.
Specifically, GraphSAGE is composed of an aggregation function $\text{\textsc{Aggregate}}^{(l)}(\cdot)$ and an update function $\text{\textsc{Update}}^{(l)}$, i.e.
\begin{displaymath}
\begin{array}{l}
x_{N_G(v)}^{(l)}=\text{\textsc{Aggregate}}^{(l)}\left(\left\{x_u^{(l-1)}: u\in N_G(v)\right\}\right) \\[5mm]
x_v^{(l)}=\text{\textsc{Update}}^{(l)}\left(\left\{x_v^{(l-1)},x_{N_G(v)}^{(l)}\right\}\right)
\end{array}
\end{displaymath}
where $N_G(v)$ denotes a fixed-size set of neighbors of $v$ uniformly sampling from its whole neighbors.
The aggregation function is of the following form
\begin{enumerate}
\item Mean Aggregator: $x_v^{(l)}=\sigma\left(W\cdot\text{\textsc{Mean}}\left(\left\{x_v^{(l-1)}\right\}\cup\left\{x_u^{(l-1)}:u\in N_G(v)\right\}\right)\right)$;
\item LSTM Aggregator: applying the LSTM \cite{160} to aggregate the neighbors of $v$;
\item Pooling Aggregator: $\text{\textsc{Aggregate}}^{(l)}=\max\left(\left\{\sigma\left(Wx_u^{(l)}+b\right): u\in N_G(v)\right\}\right)$.
\end{enumerate}
Note that the aggregation function and update function play the role of $\psi^{e\rightarrow v}(E'_i)$ and $\phi^{v}(v_i,\overline{e}'_i,u)$ in formula (\ref{GN-BLOCK}) respectively.
\textbf{Variants of GNs------Hyperbolic GCNNs.} The Euclidean GCNNs aim to embed nodes in a graph into a Euclidean space. This will incur a large distortion
especially when embedding real-world graphs with scale-free and hierarchical structure. Hyperbolic GCNNs pave an alternative way of embedding with little distortion.
The $n\text{-dimensional}$ hyperbolic space \cite{69,193}, denoted as $\mathbb{H}_K^n$, is a unique, complete, simply connected $d\text{-dimensional}$ Riemannian manifold
with constant negative sectional curvature $-\frac{1}{K}$, i.e.
\begin{displaymath}
\mathbb{H}_K^n=\left\{x\in\mathbb{R}^{n+1}:\langle x,x\rangle_{\mathcal{M}}=-K,x_0>0\right\},
\end{displaymath}
where the Minkowski inner produce $\langle x,y\rangle_{\mathcal{M}}=-x_0y_0+\sum_{j=1}^dx_jy_j,\forall x,y\in\mathbb{R}^{n+1}$.
Its tangent space centered at point $x$ is denoted as $\mathcal{T}_x\mathbb{H}_K^n=\left\{v\in\mathbb{R}^{n+1}:\langle x,v\rangle_{\mathcal{M}}=0\right\}$.
Given $x\in\mathbb{H}_K^n$, let $u\in\mathcal{T}_x\mathbb{H}_K^n$ be unit-speed. The unique unit-speed geodesic $\gamma_{x\rightarrow u}(\cdot)$ such that $\gamma_{x\rightarrow u}(0)=x$
and $\dot{\gamma}_{x\rightarrow u}(0)=u$ is denoted as $\gamma_{x\rightarrow u}(t)=\cosh\left(\frac{t}{\sqrt{K}}\right)x+\sqrt{K}\sinh\left(\frac{t}{\sqrt{K}}\right), u, t>0.$
The intrinsic distance between two points $x,y\in\mathbb{H}_K^n$ is then equal to
\begin{displaymath}
d_{\mathcal{M}}^K\left(x,y\right)=\sqrt{K}\text{arcosh}\left(-\frac{\langle x,y\rangle_{\mathcal{M}}}{K}\right).
\end{displaymath}
Therefore, the above $n\text{-dimensional}$ hyperbolic space with constant negative sectional curvature $-\frac{1}{K}$ is usually denoted as $\left(\mathbb{H}_K^n,d_{\mathcal{M}}^K(\cdot,\cdot)\right)$.
In particular, $\mathbb{H}_1^n$, i.e. $K=1$, is called the hyperboloid model of the hyperbolic space.
Hyperbolic Graph Convolutional Networks (HGCN) \cite{15} benefit from the expressiveness of both GCNNs and hyperbolic embedding.
It employs the exponential and logarithmic maps of the hyperboloid model, respectively denoted as $\exp_x^K(\cdot)$ and $\log_x^K(\cdot)$, to realize the mutual transformation
between Euclidean features and hyperbolic ones. Let $\|v\|_{\mathcal{M}}=\langle v,v\rangle_{\mathcal{M}},v\in\mathcal{T}_x\mathbb{H}_K^n$. The $\exp_x^K(\cdot)$ and $\log_x^K(\cdot)$ are respectively defined to be
\begin{displaymath}
\begin{array}{l}
\exp_x^K(v)=\cosh\left(\frac{\|v\|_{\mathcal{M}}}{\sqrt{K}}\right)x+\sqrt{K}\sinh\left(\frac{\|v\|_{\mathcal{M}}}{\sqrt{K}}\right)\frac{v}{\|v\|_{\mathcal{M}}} \\[2mm]
\log_x^{K}(y)=d_{\mathcal{M}}^K(x,y)\dfrac{y+\frac{1}{K}\langle x,y\rangle_{\mathcal{M}}x}{\left\|y+\frac{1}{K}\langle x,y\rangle_{\mathcal{M}}x\right\|_{\mathcal{M}}},
\end{array}
\end{displaymath}
where $x\in\mathbb{H}_K^n$, $v\in\mathcal{T}_x\mathbb{H}_K^n$ and $y\in\mathbb{H}_K^n$ such that $y\neq 0$ and $y\neq x$.
The HGCN architecture is composed of three components: a Hyperbolic Feature Transform (HFT), an Attention-Based Aggregation (ABA) and a Non-Linear Activation with Different Curvatures (NLADC).
They are respectively defined as
\begin{displaymath}
\begin{array}{lr}
h_j^{H,l}=\left(W^{(l)}\otimes^{K_{l-1}}x_j^{H,l-1}\right)\oplus^{K_{l-1}}b^{(l)} &\quad \text{(HFT)}, \\[2mm]
y_j^{H,l}=\text{\textsc{Aggregate}}^{K_{l-1}}(h^{H,l})_j &\quad \text{(ABA)}, \\[2mm]
x_j^{H,l}=\exp_o^{K_l}\left(\rho\left(\log_o^{K_{l-1}}\left(y_j^{H,l}\right)\right)\right) &\quad \text{(NLADC)},
\end{array}
\end{displaymath}
where $o=\left(\sqrt{K},0,\cdots,0\right)\in\mathbb{H}_K^n$, the subscript $j$ denotes the indices of nodes, the superscript $l$ denotes the layer of the HGCN.
The linear transform in hyperboloid manifold is defined to be $W\otimes^K x^H=\exp_o^K\left(W\log_o^K(x^H)\right)$ and $x^H\oplus^K b=\exp_{x^H}^K\left(P_{o\rightarrow x^H}^K(b)\right)$,
where $P_{o\rightarrow x^H}^K(b)$ is the parallel transport from $\mathcal{T}_o\mathbb{H}_K^n$ to $\mathcal{T}_{x^H}\mathbb{H}_K^{n}$.
The attention-based aggregation is defined to be $\text{\textsc{Aggregate}}^K(x^H)_j=\exp_{x_j^H}^K\left(\sum_{k\in N_G(j)}\omega_{j,k}\log_{x_j^H}^K\left(x_k^H\right)\right)$,
where the attention weight $\omega_{j,k}=\text{Softmax}_{k\in N_G(j)}\left(\text{MLP}\left(\log_o^K\left(x_j^H\right)\bowtie\log_o^K\left(x_k^H\right)\right)\right)$.
\textbf{Higher-Order Spatial GCNNs.} the aforementioned GCNN architectures are constructed from the microscopic perspective.
They only consider nodes and edges, yet overlook the higher-order substructures and their connections, i.e subgraphs consisting of at least 3 nodes.
Here, we introduce the studies on the $k\text{-dimensional}$ GCNNs \cite{88}. Specifically, they take higher-order graph structures at multiple scales into consideration by leveraging the $k\text{-Weisfeiler-Leman}$
($k\text{-WL}$) graph isomorphism test so that the message passing is performed directly between subgraph structures rather than individual nodes. Let $\{\hspace{-1.2mm}\{\cdots\}\hspace{-1.2mm}\}$
denote a multiset, $\text{\textsc{Hash}}(\cdot)$ a hashing function and $C_{l,k}^{(l)}(s)$ the node coloring (label) of $s=(s_1,\cdots,s_k)\in V^k$ at the $l\text{-th}$ time. Moreover, let $N_G^j(s)=\left\{(s_1,\cdots,s_{j-1},r,s_{j+1},\cdots,s_k):r\in V\right\}$.
The $k\text{-WL}$ is computed by
\begin{displaymath}
C_{l,k}^{(l+1)}(s)=\text{\textsc{Hash}}\left(C_{l,k}^{(l)}(s),\left(c_1^{(l+1)}(s),\cdots,c_k^{(l+1)}(s)\right)\right),
\end{displaymath}
where $c_{j}^{(l+1)}=\text{\textsc{Hash}}\left(\left\{\hspace{-1.2mm}\left\{C_{l,k}^{(l)}(s'):s'\in N_G^j(s)\right\}\hspace{-1.2mm}\right\}\right)$.
The $k\text{-GCNN}$ computes new features of $s\in V^k$ by multiple computational layers. Each layer is computed by
\begin{displaymath}
X_k^{(l+1)}[s,:]=\rho\left(X_k^{(l)}[s,:]W_1^{(l)}+\sum_{t\in N_G(s)}X_k^{(l)}[t,:]W_2^{(l)}\right).
\end{displaymath}
In practice, the local $k\text{-GCNNs}$ is often employed to learn the hierarchical representations of nodes in order to scale to larger graphs and mitigate the overfitting problem.
\textbf{Other Variants of GNs.} In addition to the aforementioned GNs and its variants, there are still many other spatial GCNNs which is defined from other perspectives, e.g.
Diffusion-Convolutional Neural Network (DCNN) \cite{20}, Position-aware Graph Neural Network (P-GNN) \cite{46}, Memory-based Graph Neural Network (MemGNN) and Graph Memory Network (GMN) \cite{84},
Graph Partition Neural Network (GPNN) \cite{35}, Edge-Conditioned Convolution (ECC) \cite{43}, DEMO-Net \cite{386}, Column network \cite{29}, Graph-CNN \cite{32}.
\textbf{Invariance and Equivariance.} Permutation-invariance refers to that a function $f:\mathbb{R}^{n^k}\rightarrow\mathbb{R}$ (e.g. the aggregation function)
is independent of any permutations of node/edge indices \cite{50,389}, i.e. $f(P^TA_GP)=f(A_G)$ where $P$ is a permutation matrix and $A_G\in\mathbb{R}^{n^k}$ is
a $k\text{-order}$ tensor of edges or multi-edges in the (hyper-)graph $G$. Permutation-equivariance refers to that a function $f:\mathbb{R}^{n^k}\rightarrow\mathbb{R}^{n^l}$
coincides with permutations of node/edge indices \cite{50,389}, i.e. $f(P^TA_GP)=P^Tf(A_G)P$ where $P$ and $A_G$ are defined as similarly as the permutation-invariance.
For permutation-invariant aggregation functions, a straightforward choice is to take $\text{sum}/\max/\text{average}/\text{concatenation}$ as heuristic aggregation schemes \cite{50}. Nevertheless,
these aggregation functions treat all the neighbors of a vertex equivalently so that they cannot precisely distinguish the structural effects of different neighbors to the target vertex.
That is, the aggregation functions should extract and filter graph signals aggregated from neighbors of different hops away and different importance.
GeniePath \cite{34} proposes a scalable approach for learning adaptive receptive fields of GCNNs. It is composed of two complementary functions,
namely adaptive breadth function and adaptive depth function. The former learns the importance of different sized neighborhoods,
whereas the latter extracts and filters graph signals aggregated from neighbors of different hops away. More specifically, the adaptive breadth function is defined as follows.
\begin{displaymath}
h_{v_j}^{\text{temp}}=\tanh\left((W^{(t)})^T\sum_{v_k\in N_G(v_j)\cup\{v_j\}}\alpha(h_{v_j}^{(t)},h_{v_k}^{(t)})\cdot h_{v_k}^{(t)}\right),
\end{displaymath}
where $\alpha(x,y)=\text{Softmax}_y\left(\alpha^T\tanh\left(W_x^Tx+W_y^Ty\right)\right)$.
The adaptive depth function is defined as a LSTM \cite{160}, i.e.
\begin{equation}\label{LSTM Formulation}
\begin{array}{ll}
i_{v_j}=\sigma\left(\left(W_i^{(t)}\right)^Th_{v_j}^{\text{temp}}\right) & \qquad f_{v_j}=\sigma\left(\left(W_f^{(t)}\right)^Th_{v_j}^{\text{temp}}\right) \\
o_{v_j}=\sigma\left(\left(W_o^{(t)}\right)^Th_{v_j}^{temp}\right) & \qquad \widetilde{C}_{v_j}=\tanh\left(\left(W_c^{(t)}\right)^Th_{v_j}^{(temp)}\right) \\
C_{v_j}^{(t+1)}=f_{v_j}\circledast C_{v_j}^{(t)}+i_{v_j}\circledast\widetilde{C}_{v_j} & \qquad h_{v_j}^{(t+1)}=o_{v_j}\circledast\tanh\left(C_{v_j}^{(t+1)}\right).
\end{array}
\end{equation}
\textsc{Geom-GCN} \cite{17} proposes a novel permutation-invariant geometric aggregation scheme consisting of three modules,
namely node embedding, structural neighborhood, and bi-level aggregation. This aggregation scheme does not lose structural information of nodes and fail to capture long-range dependencies in disassortative graphs.
For the permutation-invariant graph representations, PiNet \cite{41} proposes an end-to-end spatial GCNN architecture that utilizes the permutation equivariance of graph convolutions.
It is composed of a pair of double-stacked message passing layers, namely attention-oriented message passing layers and feature-oriented message passing layers.
\textbf{Depth Trap of Spatial GCNNs.} Similar to the spectral GCNNs, the spatial GCNNs is also confronted with the depth trap.
As stated previously, the depth trap results from oversmoothing, overfitting and gradient vanishing/explosion.
In order to escape from the depth trap, some studies propose some available strategies, e.g. DeepGCN \cite{13} and Jumping Knowledge Network \cite{89}.
The jumping knowledge networks \cite{89} adopt neighborhood aggregation with skip connections to integrate information from different layers.
The DeepGCN \cite{13} apply the residual/dense connections \cite{382,383} and dilated aggreagation \cite{384} in the CNNs to construct the spatial GCNN architecture.
They has three instantiations, namely ResGCN, DenseGCN and dilated graph convolution.
ResGCN is inspired by the ResNet \cite{382}, which is defined to be
\begin{displaymath}
\begin{array}{lcl}
G^{(l+1)} & \triangleq & \mathcal{H}(G^{(l)},\Theta^{(l)}) \\ [2mm]
& = & \mathcal{F}(G^{(l)},\Theta^{(l)})+G^{(l)},
\end{array}
\end{displaymath}
where $\mathcal{F}(\cdot,\cdot)$ can be computed by spectral or spatial GCNNs.
DenseGCN collectively exploit information from different GCNN layers like the DenseNet \cite{383}, which is defined to be
\begin{displaymath}
\begin{array}{lcl}
G^{(l+1)} & \triangleq & \mathcal{H}(G^{(l)},\Theta^{(l)}) \\ [2mm]
& = & \text{Concat}\left(\mathcal{F}(G^{(l)},\Theta^{(l)}),G^{(l)}\right) \\ [2mm]
& = & \text{Concat}\left(\mathcal{F}(G^{(l)},\Theta^{(l)}),\cdots,\mathcal{F}(G^{(0)},\Theta^{(0)}),G^{(0)}\right).
\end{array}
\end{displaymath}
The dilated aggregation \cite{384} can magnify the receptive field of spatial GCNNs by a dilation rate $d$. More specifically, let $N_G^{(k,d)}(v)$ denote the set of $k$ $d\text{-dilated}$ neighbors of vertex $v$ in $G$.
If $\left(u_1,u_2,\cdots,u_{k\times d}\right)$ are the first sorted $k\times d$ nearest neighbors, then $N_G^{(k,d)}(v)=\left\{u_1,u_{1+d},\cdots,u_{1+(k-1)d}\right\}$.
Thereby, we can construct a new graph $G^{(k,d)}=\left(V^{(k,d)},E^{(k,d)}\right)$ where $V^{(k,d)}=V$ and $E^{(k,d)}=\left\{\left\langle v,u\right\rangle:v\in V,u\in N_G^{(k,d)}\right\}$.
The dilated graph convolution layer can be obtained by running the spatial GCNNs over $G^{(k,d)}$.
\subsubsection{Graph Wavelet Neural Networks}
As stated previously, the spectral and spatial GCNNs are respectively inspired by the graph Fourier transform and message-passing mechanism.
Here, we introduce a new GCNN architecture from the perspective of the Spectral Graph Wavelet Transform (SGWT) \cite{189}.
First of all, the SGWT is determined by a graph wavelet generating kernel $g:\mathbb{R}^+\rightarrow\mathbb{R}^+$ with the property $g(0)=0,g(+\infty)=\lim_{x\rightarrow\infty}g(x)=0$.
A feasible instance of $g(\cdot)$ is parameterized by two integers $\alpha$ and $\beta$, and two positive real numbers $x_1$ and $x_2$ determining the transition regions, i.e.
\begin{displaymath}
g(x;\alpha,\beta,x_1,x_2)=\left\{
\begin{array}{ll}
x_1^{-\alpha}x^{\alpha} & x<x_1 \\
s(x) & x_1\leq x\leq x_2 \\
x^{-\beta}x_2^{\beta} & x>x_2,
\end{array}
\right.
\end{displaymath}
where $s(x)$ is a cubic polynomial whose coefficients can be determined by the continuity constraints $s(x_1)=s(x_2)=1$, $s'(x_1)=\frac{\alpha}{x_1}$ and $s'(x_2)=-\frac{\beta}{x_2}$.
Given the graph wavelet generating kernel $g(\cdot)$ and a scaling parameter $s\in\mathbb{R}^+$, the spectral graph wavelet operator $\Psi_g^s$ is defined to be $\Psi_g^s=Ug(s\Lambda)U^T$
where $g(s\Lambda)=g\left(\text{diag}\left(s\lambda_1,\cdots,s\lambda_N\right)\right)$.
A graph signal $x\in\mathbb{R}^N$ on $G$ can thereby be filtered by the spectral graph wavelet operator, i.e. $\mathcal{W}_{g,s}^x=\Psi_g^sx\in\mathbb{R}^N$.
The literature \cite{190} utilizes a special instance of the graph wavelet operator $\Psi_g^s$ to construct a graph scattering network,
and proves its covariance and approximate invariance to permutations and stability to graph operations.
\textbf{Graph Wavelet Neural Networks.} The above spectral graph wavelet operator $\Psi_g^s$ can be employed to construct a Graph Wavelet Neural Network (GWNN) \cite{191}.
Let $\Psi_g^{-s}\triangleq\left(\Psi_g^s\right)^{-1}$. The graph wavelet based convolution is defined to be
\begin{displaymath}
x*_Gy=\Psi_g^{-s}\left(\Psi_g^sx\circledast\Psi_g^sy\right).
\end{displaymath}
The GWNN is composed of multiple layers of the graph wavelet based convolution. The structure of the $l\text{-th}$ layer is defined as
\begin{equation}\label{Graph Wavelet based Convolution in vector form}
X^{(l+1)}[:,j]=\rho\left(\sum_{k=1}^{d^{(l)}}\Psi_g^{-s}\Theta_{j,k}^{(l)}\Psi_g^sX^{(l)}[:,k]\right),
\end{equation}
where $\Theta_{j,k}^{(l)}$ is a diagonal filter matrix learned in spectral domain.
Eq. (\ref{Graph Wavelet based Convolution in vector form}) can be rewritten as a matrix form, i.e. $X^{(l+1)}=\rho\left(\Psi_g^{-s}\Theta\Psi_g^{s}X^{(l)}W^{(l)}\right)$.
The learnable filter matrix $\Theta$ can be replaced with the $K\text{-localized}$ Chebyshev Polynomial so as to eschew the time-consuming eigendecomposition of $\overline{L}_G$.
\subsubsection{Summary}
The aforementioned GCNN architectures provide available ingredients of constructing the GNNs.
In practice, we can construct our own GCNNs by assembling different modules introduced above.
Additionally, some scholars also study the GCNNs from some novel perspectives,
e.g. the parallel computing framework of the GCNNs \cite{40}, the hierarchical covariant compositional networks \cite{213},
the transfer active learning for GCNNs \cite{45} and quantum walk based subgraph convolutional neural network \cite{51}.
They are closely related to the GCNNs, yet fairly different from the ones introduced above.
\subsection{Graph Pooling Operators}
Graph pooling operators are very important and useful modules of the GCNNs, especially for graph-level tasks such as the graph classification.
There are two kinds of graph pooling operators, namely global graph pooling operators and hierarchical graph pooling operators.
The former aims to obtain the universal representations of input graphs, and the latter aims to capture adequate structural information for node representations.
\subsubsection{Global Graph Pooling Operators.}
Global graph pooling operators pool all of representations of nodes into a universal graph representation.
Many literatures \cite{8,21,41} apply some simple global graph pooling operators, e.g. max/average/concatenate graph pooling, to performing graph-level classification tasks.
Here, we introduce some more sophisticated global graph pooling operators in contrast to the simple ones.
Relational pooling (RP) \cite{98} provides a novel framework for graph representation with maximal representation power.
Specifically, all node embeddings can be aggregated via a learnable function to form a global embedding of $G$.
Let $X^{(v)}\in\mathbb{R}^{N\times d_v}$ and $\underline{X}^{(e)}\in\mathbb{N}^{N\times N\times d_e}$ respectively denote node feature matrix and edge feature tensor.
Tensor $\underline{A}_G\in\mathbb{R}^{N\times N\times (1+d_e)}$ combines the adjacency matrix $A_G$ of $G$ with its edge feature tensor $\underline{X}^{(e)}$,
i.e. $\underline{A}_G[u,v,:]=\mathbb{I}_{(u,v)\in E_G}\bowtie\underline{X}^{(e)}[u,v,:]$.
After performing a permutation on $V_G$, the edge feature tensor $\underline{A}_G^{(\pi,\pi)}\left[\pi(r),\pi(c),d\right]=\underline{A}_G[r,c,d]$ and the node feature matrix $X_{\pi}^{(v)}[\pi(r),c]=X^{(v)}[r,c]$.
The joint RP permutation-invariant function for directed or undirected graphs is defined as
\begin{displaymath}
\bar{\bar{f}}(G)=\frac{1}{N!}\sum_{\pi\in\Pi_{|V|}}\overrightarrow{f}(\underline{A}_G^{\pi,\pi},X_{\pi}^{(v)}),
\end{displaymath}
where $\Pi_{|V|}$ is the set of all distinct permutations on $V_G$ and $\overrightarrow{f}(\cdot,\cdot)$ is an arbitrary (possibly permutation-sensitive) vector-valued function.
Specifically, $\overrightarrow{f}(\cdot,\cdot)$ can be denoted as Multi-Layer Perceptrons (MLPs), Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs) or Graph Neural Networks (GNNs).
The literature \cite{98} proves that $\bar{\bar{f}}(G)$ has the most expressive representation of $G$ under some mild conditions, and provides approximation approaches to making RP computationally tractable.
In addition, there are some other available global graph pooling operators, e.g. SortPooling \cite{92} and function space pooling \cite{97}.
\subsubsection{Hierarchical Graph Pooling Operators.}
Hierarchical graph pooling operators group a set of proximal nodes into a super-node via graph clustering methods.
Consequently, the original graph is coarsened to a new graph with coarse granularity. In practice, the hierarchical graph pooling operators are interleaved with the vanilla GCNN layers.
In general, there are three kinds of approaches to performing the graph coarsening operations, namely invoking the existing graph clustering algorithms (e.g. spectral clustering \cite{7} and Graclus \cite{99}),
learning a soft cluster assignment and selecting the first $k$ top-rank nodes.
\textbf{Invoking existing graph clustering algorithms.}
The graph clustering aims to assign proximal nodes to the same cluster and in-proximal nodes to different clusters.
The coarsened graph regards the resulting clusters as super-nodes and connections between two clusters as super-edges.
The hierarchical graph pooling operators aggregate the representations of nodes in super-nodes via aggregation functions such as max pooling and average pooling \cite{91}
to compute the representations of the super-nodes. The literature \cite{82} proposes the EigenPooling method and presents the relationship between the original and coarsened graph.
In order to construct a coarsened graph of $G$, a graph clustering method is employed to partition $G$ into $K$ disjoint clusters, namely $\left\{G_k:k=1,\cdots,K\right\}$.
Suppose each cluster $G_k$ has $N_k$ nodes, namely $\left\{v_{k,1},\cdots,v_{k,N_k}\right\}$, and its adjacency matrix is denoted as $A_{G_k}$.
The coarsened graph $G_{\text{coar}}$ of $G$ can be constructed by regarding the clusters $G_k,k=1,\cdots,K$ as super-nodes and connections between two super-nodes as edges.
For $G_k$, its sampling matrix $C_k$ of size $\left(N\times N_k\right)$ is defined by
\begin{displaymath}
C_k(s,t)=\left\{
\begin{array}{lll}
1 && \text{if node $v_{k,s}$ in $G_k$ is identical to vertex $v_{t}$ in $G$} \\ [2mm]
0 && \text{otherwise}
\end{array}
\right.
\end{displaymath}
On one hand, $C_k$ can be used to down-sample a $1\text{-dimensional}$ graph signal $x$ on $G$ to obtain an contracted graph signal $x_{G_k}$ on $G_k$, i.e. $x_{G_k}=C_k^Tx$.
On the other hand, $C_k$ can also be used to up-sample a graph signal $x_{G_k}$ on $G_k$ to obtain a dilated graph $G$, i.e. $x=C_kx_{G_k}$.
Furthermore, the adjacency matrix $A_{G_k}$ of $G_k$ can be computed by
\begin{displaymath}
A_{G_k}=C_k^TA_GC_k.
\end{displaymath}
The intra-subgraph adjacency matrix of $G$ is computed by $A_{\text{intra}}=\sum_{k=1}^KC_kA_{G_k}C_k^T$.
Thereby, the inter-subgraph adjacency matrix of $G$ can be computed by $A_{\text{inter}}=A_G-A_{\text{intra}}$.
Let $M_{\text{coar}}\in\mathbb{R}^{N\times K}$ denote the assignment matrix from $G$ to $G_{\text{coar}}$. Its $(j,k)\text{-th}$ entry is defined as
\begin{displaymath}
M_{\text{coar}}[j,k]=\left\{
\begin{array}{lll}
1 && \text{if $v_j$ in $G$ is grouped into $G_k$ in $G_{\text{coar}}$} \\ [2mm]
0 && \text{otherwise} \\
\end{array}
\right.
\end{displaymath}
As a result, the adjacency matrix $A_{\text{coar}}$ of the coarsened graph $G_{\text{coar}}$ is computed by $A_{\text{coar}}=M_{\text{coar}}^TA_{\text{inter}}M_{\text{coar}}$.
In fact, $A_{\text{coar}}$ can be written as $A_{\text{coar}}=f\left(M_{\text{coar}}^TA_GM_{\text{coar}}\right)$ as well,
where $f(\widetilde{a}_{i,j})=1$ if $\widetilde{a}_{i,j}>0$ and $f(\widetilde{a}_{i,j})=0$ otherwise.
As stated previously, $X$ is a $d\text{-dimensional}$ graph signal on $G$.
Then, a $d\text{-dimensional}$ graph signal $X_{\text{coar}}$ on $G_{\text{coar}}$ can be computed by $X_{\text{coar}}=M_{\text{coar}}^TX$.
EigenPooling \cite{82} employs spectral clustering to obtain the coarsened graph, and then up-sample
the Fourier basis of subgraphs $G_k,k=1,\cdots,K$. These Fourier basis are then organized into pooling operators with regard to ascending eigenvalues.
Consequently, the pooled node feature matrix is obtained via concatenating the pooled results.
The literature \cite{42} proposes a novel Hierarchical Graph Convolutional Network (H-GCN) consisting of graph coarsening layers and graph refining layers.
The former employs structural equivalence grouping and structural similarity grouping to construct the coarsened graph,
and the latter restores the original topological structure of the corresponding graph.
\textbf{Learning a soft cluster assignment.}
\textsc{StructPool} \cite{390}, as a structured graph pooling technique, regards the graph pooling as a graph clustering problem so as to learn a
cluster assignment matrix
via the feature matrix $X$ and adjacency matrix $A_G$. Learning the cluster assignment matrix can formulated as a Conditional Random Field (CRF) \cite{391} based probabilistic inference.
Specifically, the input feature matrix $X$ is treated as global observation, and $Y=\left\{Y_1,\cdots,Y_N\right\}$ is a random field where $Y_i\in\{1,\cdots,K\}$ is a random variable
indicating which clusters the node $v_i$ is assigned to. As a result, $\left(Y,X\right)$ can be characterized by a CRF model, i.e.
\begin{displaymath}
\begin{array}{lll}
\mathbb{P}(Y|X) & = & \dfrac{1}{Z(X)}\exp\left(-\mathcal{E}(Y|X)\right) \\[2mm]
& = & \dfrac{1}{Z(X)}\exp\left(\displaystyle\sum_{C\in\mathcal{C}_G}\psi_C(Y_C|X)\right)
\end{array}
\end{displaymath}
where $\mathcal{E}(Y|X)=-\sum_{C\in\mathcal{C}_G}\psi_C(Y_C|X)$ is called an energy function, $\mathcal{C}_G$ is a set of cliques, $\psi_C(Y_C|X)$ is a potential function and $Z(X)$ is a partition function.
The energy function $\mathcal{E}(Y|X)$ can be characterized by an unary energy $\psi_u(\cdot)$ and a pairwise energy $\psi_p(\cdot,\cdot)$, i.e.
\begin{displaymath}
\mathcal{E}(Y|X)=\sum_{s=1}^N\psi_u(y_s|X)+\sum_{s\neq t}\psi_p(y_s,y_t|X)a_{s,t}^l,
\end{displaymath}
where $a_{s,t}^l$ denotes the $(s,t)\text{-th}$ entry of the $l\text{-hop}$ adjacency matrix $A_G^l$.
The unary energy matrix $\Psi_u=\left(\psi_u(y_s|X)\right)_{N\times K}$ can be obtained by a GCNN taking the global observation $X$ and the adjacency $A_G$ as input.
The pairwise energy matrix $\Psi_p=\left(\psi_p(y_s,y_t|X)\right)_{K\times K}$ can be obtained by
\begin{displaymath}
\psi_p(y_s,y_t|X)=\mu(y_s,y_t)\frac{x_s^Tx_t}{\sum_{j\neq s}x_s^Ts_j},
\end{displaymath}
where $\mu(y_s,y_t)$ is a learnable compatibility function. Minimizing the energy function $\mathcal{E}(Y|X)$ via mean-field approximation
results in the most probable cluster assignment matrix $M$ for a give graph $G$. As a result, we obtain a new graph $A_{\text{coar}}=f\left(M^TA_GM\right)$ and $X_{\text{coar}}=M^TX$.
\textsc{DiffPool} \cite{94} is a differentiable graph pooling operator
which can generate hierarchical representations of graphs and can be incorporated into various GCNNs in an end-to-end fashion.
It maps an adjacency matrix $A_{G^{(l)}}$ and embedding matrix $Z^{(l)}$ at the $l\text{-th}$ layer
to a new adjacency matrix $A_{G^{(l+1)}}$ and a coarsened feature matrix $X^{(l+1)}$, i.e. $\left(A_{G^{(l+1)}},X^{(l+1)}\right)=\text{\textsc{DiffPool}}\left(A_{G^{(l)}},Z^{(l)}\right)$.
More specifically, $X^{(l+1)}=\left(M^{(l)}\right)^TZ^{(l)}$, $A_{G^{(l+1)}}=\left(M^{(l)}\right)^TA_{G^{(l)}}M^{(l)}$.
Note that the assignment matrix $M^{(l)}$ and embedding matrix $Z^{(l)}$ are respectively computed by two separate GCNNs, namely embedding GCNN and pooling GCNN, i.e.
$Z^{(l)}=\text{GCNN}_{\text{embed}}\left(A_{G^{(l)}},X^{(l)}\right)$, $M^{(l)}=\text{Softmax}\left(\text{GCNN}_{\text{pool}}\left(A_{G^{(l)}},X^{(l)}\right)\right)$.
\textbf{Selecting the first $k$ top-rank nodes.}
The literature \cite{86} proposes a novel Self-Attention Graph Pooling operator (abbreviated as SAGPool).
Specifically, SAGPool firstly employs the GCN \cite{6} to calculate the self-attention scores, and then invokes the top-rank function to select the top $\lceil kN\rceil$ node indices, i.e.
\begin{displaymath}
\begin{array}{lll}
Z=\rho\left(\widetilde{D}_G^{-\frac{1}{2}}\widetilde{A}_G\widetilde{D}_G^{-\frac{1}{2}}X\Theta\right), & \text{idx}=\text{top-rank}\left(Z,\lceil kN\rceil\right), & Z_{\text{mask}}=Z_{\text{idx}} \\
X'=X_{\text{idx}}, & X_{\text{out}}=X'\circledast Z_{\text{mask}}, & A_{\text{out}}=A_{\text{idx},\text{idx}}.
\end{array}
\end{displaymath}
As a result, the selected $\text{top-}\lceil kN\rceil$ node indices are employed to extract the output adjacency matrix $A_{\text{out}}$ and feature matrix $X_{\text{out}}$.
In order to exploit the expressive power of an encoder-decoder architecture like U-Net \cite{72}, the literature \cite{95} proposes a novel graph pooling (gPool) layer and a graph unpooling (gUnpool) layer.
The gPool adaptively selects $\text{top-}k$ ranked node indices by the down-sampling technique to form a coarsened graph ($A_{\text{coar}}\in\mathbb{R}^{N\times N}$ and $X_{\text{coar}}\in\mathbb{R}^{N\times d}$)
based on scalar projection values on a learnable projection vector, i.e.
\begin{displaymath}
\begin{array}{lll}
y=\frac{Xp}{\|p\|}, & \text{idx}=\text{top-rank}(y,k), & \widetilde{y}=\tanh(y_{\text{idx}}) \\
\widetilde{X}_{\text{coar}}=X_{\text{idx},:}, & A_{\text{coar}}=A_{\text{idx},\text{idx}}, & X_{\text{coar}}=\widetilde{X}_{\text{coar}}\circledast(\widetilde{y}\mathbf{1}_C^T),
\end{array}
\end{displaymath}
where $y\in\mathbb{R}^{d}$. The gUnpool performs the inverse operation of the gPool layer so as to restore the coarsened graph into its original structure.
To this end, gUnpool records the locations of nodes selected in the corresponding gPool layer, and then restores the selected nodes to their original positions in the graph.
Specifically, let $X_{\text{refine}}=\text{Distribute}(0_{N\times d},X_{\text{coar}},\text{idx})$,
where the function $\text{Distribute}(\cdot,\cdot,\cdot)$ distributes row vectors in $X_{\text{coar}}$ into $0_{N\times d}$ feature matrix according to the indices $\text{idx}$.
Note that row vectors of $X_{\text{refine}}$ with indices in $idx$ are updated by the ones in $X_{\text{coar}}$, whereas other row vectors remain zero.
It is worth noting that the literature \cite{96} adopts the similar pooling strategy as gPool to learn the hierarchical representations of nodes.
\subsection{Graph Attention Mechanisms}
Attention mechanisms, firstly introduced in the deep learning community, guide deep learning models to focus on the task-relevant part of its inputs so as to make precise predictions or inferences \cite{73,306,393}.
Recently, applying the attention mechanisms to GCNNs has gained considerable attentions so that various attention techniques have been proposed.
Below, we summarize the graph attention mechanisms on graphs from the next 4 perspectives \cite{394}, namely softmax-based graph attention, similarity-based graph attention,
spectral graph attention and attention-guided walk. Without loss of generality, the neighbors of a given node $v_0$ in $G$ are denoted as $v_1,\cdots,v_{d_0}$,
and their current feature vectors are respectively denoted as $x_0,x_1,\cdots,x_{d_0}$, where $d_0=d_{G}(v_0)$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{GAT.pdf}
\caption{Two kinds of graph attention mechanisms.}
\label{GAT}
\end{figure}
\subsubsection{Concatenation-based Graph Attention.}
The softmax-based graph attention is typically implemented by employing a softmax with learnable weights \cite{66,110}
to measure the relevance of $v_j,j=1,\cdots,d_G(v_0)$ to $v_0$. More specifically, the softmax-based attention weights between $v_0$ and $v_j$ can be defined as
\begin{equation}\label{Softmax-based Attention}
\begin{array}{lll}
\omega_{0,j} & = & \text{Softmax}\left(\left[e_{0,1},\cdots,e_{0,d_G(v_0)}\right]\right) \\[2mm]
& = & \dfrac{\exp\left(\rho\left(a^T(Wx_0\bowtie Wx_j)\right)\right)}{\sum_{k=1}^{d_G(v_0)}\exp\left(\rho\left(a^T(Wx_0\bowtie W_k)\right)\right)},
\end{array}
\end{equation}
where $e_{0,j}=\exp\left(\rho\left(a^T(Wx_0\bowtie Wx_j)\right)\right)$, $a$ is a learnable attention vector and $W$ is a learnable weight matrix, see Fig. \ref{GAT}(a).
As a result, the new feature vector of $v_0$ can be updated by
\begin{equation} \label{Single_Graph_Attention}
x'_{0}=\rho\left(\sum_{j=1}^{d_G(v_0)}\omega_{0,j}Wx_j\right).
\end{equation}
In practice, multi-head attention mechanisms are usually employed to stabilize the learning process of the single-head attention \cite{66}.
For the multi-head attention, assume that the feature vector of each head is $x_0^{(h)}=\rho\left(\sum_{j=1}^{d_G(v_0)}\omega_{0,j}^{(h)}W^{(h)}x_j\right)$.
The concatenation based multi-head attention is computed by
$x'_0=\bowtie_{h=1}^H x_0^{(h)}=\bowtie_{h=1}^H \rho\left(\sum_{j=1}^{d_G(v_0)}\omega_{0,j}^{(h)}W^{(h)}x_j\right)$.
The average based multi-head attention is computed by $x'_0=\rho\left(\frac{1}{H}\sum_{h=1}^H\sum_{j=1}^{d_G(v_0)}\omega_{0,j}^{(h)}W^{(h)}x_j\right)$.
The conventional multi-head attention mechanism treats all the attention heads equally so that feeding the output of an attention that captures a useless representation maybe mislead the final prediction of the model.
The literature \cite{74} computes an additional soft gate to assign different weights to heads, and gets the formulation of the gated multi-head attention mechanism.
The Graph Transformer (GTR) \cite{110} can capture long-range dependencies of dynamic graphs with softmax-based attention mechanism by propagating features within the same graph structure via an intra-graph message passing.
The Graph-BERT \cite{367} is essentially a pre-training method only based on the graph attention mechanism without any graph convolution or aggregation operators.
Its key component is called a graph transformer based encoder, i.e. $X^{(l+1)}=\text{Softmax}\left(\frac{QK^T}{\sqrt{d_h}}\right)V$,
where $Q=X^{(l)}W_Q^{(l+1)}$, $K=X^{(l)}W_K^{(l+1)}$ and $V=X^{(l)}W_V^{(l+1)}$.
The Graph2Seq \cite{106} is a general end-to-end graph-to-sequence neural encoder-decoder model converting an input graph to a sequence of vectors with the attention based LSTM model.
It is composed of a graph encoder, a sequence decoder and a node attention mechanism. The sequence decoder takes outputs (node and graph representations) of the graph encoder as input, and employs the softmax-based
attention to compute the context vector sequence.
\subsubsection{Similarity-based Graph Attention.}
The similarity-based graph attention depends on the cosine similarities of the given node $v_0$ and its neighbors $v_j,j=1,\cdots,d_G(v_0)$.
More specifically, the similarity-based attention weights are computed by
\begin{equation}\label{Similarity-based Attention}
\omega_{0,j}=\dfrac{\exp\left(\beta\cdot\cos\left(Wx_0,Wx_j\right)\right)}{\sum_{k=1}^{d_G(v_0)}\exp\left(\beta\cdot\cos\left(Wx_0,Wx_k\right)\right)},
\end{equation}
where $\beta$ is learnable bias and $W$ is a learnable weight matrix, see Fig. \ref{GAT}(b). It is well known that $\cos\left(x,y\right)=\frac{\langle x,y\rangle}{\|x\|_2\|y\|_2}$.
Attention-based Graph Neural Network (AGNN) \cite{67} adopts the similarity-based attention to construct the propagation matrix $P^{(l)}$ capturing the relevance of $v_j$ to $v_i$.
As a result, the output hidden representation $X^{(l+1)}$ at the $(l+1)\text{-th}$ layer is computed by
\begin{displaymath}
X^{(l+1)}=\rho\left(P^{(l)}X^{(l)}W^{(l)}\right),
\end{displaymath}
where $P^{(l)}(i,j)\triangleq\omega_{i,j}$ is defined in formula (\ref{Similarity-based Attention}).
\subsubsection{Spectral Graph Attention.}
The Spectral Graph Attention Networks (SpGAT) aims to learn representations for different frequency components \cite{362}.
The eigenvalues of the normalized graph Laplacian $\overline{L}_G$ can be treated as frequencies on the graph $G$. As stated in the Preliminary section, $0=\lambda_1\leq\lambda_2\leq\cdots\leq\lambda_N=\lambda_{\max}$.
The SpGAT firstly extracts the low-frequency component $B_L=\left\{u_1,\cdots,u_{n}\right\}$ and the high-frequency component $B_H=\left\{u_{N-n+1},\cdots,u_N\right\}$
from the graph Fourier bases $\left\{u_1,u_2,\cdots,u_N\right\}$.
So, we have
\begin{equation}\label{Spectral Graph Attention Networks}
\begin{array}{l}
X_L=X^{(l)}\Theta_L,\quad X_H=X^{(l)}\Theta_H \\
X^{(l+1)}=\rho\left(\text{\textsc{Aggregate}}\left(B_LF_LB_L^TX_L,B_HF_HB_H^TX_H\right)\right),
\end{array}
\end{equation}
where $F_L$ and $F_H$ respectively measure the importance of the low- and high-frequency.
In practice, we exploit a re-parameterization trick to accelerate the training. More specifically, we replace $F_L$ and $F_H$ respectively with the learnable attention weights
$\Omega_L=\text{diag}\left(\omega_L,\cdots,\omega_L\right)$ and $\Omega_H=\text{diag}\left(\omega_H,\cdots,\omega_H\right)$ so as to reduce the number of learnable parameters.
To ensure that $\omega_L$ and $\omega_H$ are positive and comparable, we normalize them by the softmax function, i.e.
$\omega_L=\frac{\exp\left(\omega_L\right)}{\exp\left(\omega_L\right)+\exp\left(\omega_H\right)},\quad \omega_H=\frac{\exp\left(\omega_H\right)}{\exp\left(\omega_L\right)+\exp\left(\omega_H\right)}$.
In addition to the attention weights, another important issue is how to choose the low- and high-frequency components $B_L$ and $B_H$.
A natural choice is to use the graph Fourier bases, yet the literatures \cite{191,194} conclude that utilizing the spectral graph wavelet operators can achieve better embedding results than the graph Fourier bases.
Therefore, we substitute $B_L$ and $B_H$ in Eq. (\ref{Spectral Graph Attention Networks}) for the spectral graph wavelet operator $\Psi_{L,g}^s$ and $\Psi_{H,g}^s$, i.e.
\begin{displaymath}
X^{(l+1)}=\rho\left(\text{\textsc{Aggregate}}\left(\left(\Psi_{L,g}^s\right)F_L\left(\Psi_{L,g}^s\right)^{-1}X_L,\left(\Psi_{H,g}^s\right)F_H\left(\Psi_{H,g}^s\right)^{-1}X_H\right)\right).
\end{displaymath}
\subsubsection{Attention-guided Walk.}
The two aforementioned kinds of attention mechanisms focus on incorporating task-relevant information from the neighbors of a given node into the updated representations of the pivot.
Here, we introduce a new attention mechanism, namely attention-guided walk \cite{81}, which has different purpose from the softmax- and similarity-based attention mechanisms.
Suppose a walker walks along the edges of the graph $G$ and he currently locates at the node $v_t$.
The hidden representation $x^{(t)}$ of $v_t$ is computed by a recurrent neural network $f_x\left(\cdot\right)$
taking the step embedding $s^{(t)}$ and internal representation of the historical information from the previous step $x^{(t-1)}$ as input, i.e.
\begin{displaymath}
x^{(t)}=f_x\left(s^{(t)},x^{(t-1)};\Theta_x\right).
\end{displaymath}
The step embedding $s^{(t)}$ is computed by a step network $f_s\left(r^{(t-1)},x_{v_t};\Theta_s\right)$ taking the ranking vector $r^{(t-1)}$
and the input feature vector $x_{v_t}$ of the top-priority node $v_t$ as input, i.e.
\begin{displaymath}
s^{(t)}=f_s\left(r^{(t-1)},x_{v_t};\Theta_s\right).
\end{displaymath}
The hidden representation $x^{(t)}$ is then feeded into a ranking network $f_r\left(x^{(t)};\Theta_r\right)$ and a predicting network $f_p\left(x^{(t)};\Theta_p\right)$, i.e.
\begin{displaymath}
r^{(t)}=f_r\left(x^{(t)};\Theta_r\right), \quad\hat{l}^{(t)}=f_p\left(x^{(t)};\Theta_p\right).
\end{displaymath}
The ranking network $f_r\left(x^{(t)};\Theta_r\right)$ determines which neighbors of $v_t$ should be prioritized in the next step, and the predicting network $f_p\left(x^{(t)};\Theta_p\right)$
makes a prediction on graph labels. Now, $x^{(t)}$ and $r^{(t)}$ are feeded into the next node to compute its hidden representations.
Fig. \ref{attention-guided-walk} shows the computational framework of the attention-guided walk.
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{attention-guided-walk.pdf}
\caption{The computational framework of the attention-guided walk.}
\label{attention-guided-walk}
\end{figure}
\subsection{Graph Recurrent Neural Networks}
The Graph Recurrent Neural Networks (GRNNs) generalize the Recurrent Neural Networks (RNNs) to process the graph-structured data.
In general, the GRNN can be formulated as $h_{v_j}^{(l+1)}=\text{GRNN}\left(x_{v_j}^{(l)},\left\{h_{v_k}^{(l)}:v_k\in N_G(v_j)\cup\left\{v_j\right\}\right\}\right)$.
Below, we introduce some available GRNN architectures.
\subsubsection{Graph LSTM}
The Graph Long Short Term Memroy (Graph LSTM) \cite{117,118,119,120,121,123,143} generalizes the vanilla LSTM for the sequential data to the ones for general graph-structured data.
Specifically, the graph LSTM updates the hidden states and cell states of nodes by the following formula,
\begin{displaymath}
\begin{array}{ll}
i_{v_j}^{(l+1)}=\sigma\left(W_ix_{v_j}^{(l)}+\sum_{v_k\in N_G(v_j)\cup\left\{v_j\right\}}U_ih_{v_k}^{(l)}+b_i\right), \\ [2mm]
o_{v_j}^{(l+1)}=\sigma\left(W_ox_{v_j}^{(l)}+\sum_{v_k\in N_G(v_j)\cup\left\{v_j\right\}}U_oh_{v_k}^{(l)}+b_o\right) \\ [2mm]
\widetilde{c}_{v_j}^{(l+1)}=\tanh\left(W_cx_{v_j}^{(l)}+\sum_{v_k\in N_G(v_j)\cup\left\{v_j\right\}}U_ch_{v_k}^{(l)}+b_c\right), \\ [2mm]
f_{v_j,v_k}^{(l+1)}=\sigma\left(W_fx_{v_j}^{(l)}+U_fh_{v_k}^{(l)}+b_f\right) \\ [2mm]
c_{v_j}^{(l+1)}=i_{v_j}^{(l+1)}\circledast\widetilde{c}_{v_j}^{(l+1)}+\sum_{v_k\in N_G{v_j}\cup\left\{v_j\right\}}f_{v_j,v_k}^{(l+1)}\circledast c_{v_k}^{(l)}, \\ [2mm] h_{v_j}^{(l+1)}=o_{v_j}^{(l+1)}\circledast\tanh(c_{v_j}^{(l+1)}).
\end{array}
\end{displaymath}
see the Fig. \ref{Graph LSTM}.
The literature \cite{267} develops a general framework, named structure-evolving LSTM, for learning interpretable data representations via the graph LSTM.
It progressively evolves the multi-level node representations by stochastically merging two adjacent nodes with high compatibilities estimated by the adaptive forget gate of the graph LSTM.
As a result, the new graph is produced with a Metropolis-Hastings sampling method.
The Gated Graph Sequence Neural Networks (GGS-NNs) \cite{122} employs the Gated Recurrent Unit (GRU) \cite{115} to modify the vanilla GCNNs so that it can be extended to process the sequential data.
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{Graph-lstm.pdf}
\caption{Computational framework of the Graph LSTM.}
\label{Graph LSTM}
\end{figure}
\subsubsection{GRNNs for Dynamic Graphs}
A dynamic graph is the one whose structure (i.e. adding a node, removing a node, adding an edge, removing an edge) or features on nodes/edges evolve over time.
The GRNNs are a straightforward approach to tackling dynamic graphs. Below, we introduce some studies on the GRNNs for dynamic graphs.
The literature \cite{107} proposes a Dynamic Graph Neural Network (DGNN) concerning the dynamic graph only with nodes or edges adding.
More specifically, the DGNN is composed of two key components: an update component and a propagation component. Suppose an edge $(v_s,v_g,t)$ is added to the input dynamic graph at time $t$.
Let $t-$ denotes a time before the time $t$. The update component consists of three sequential units: the interacting unit, the S- or G-update unit and the merging unit.
The interacting unit takes the source and target representations before the time $t$ as input, and outputs the joint representation of the interaction, i.e.
$x_e^{(t)}=\rho\left(W_sx_s^{(t-)}+W_gx_g^{(t-)}+b_e\right)$. The S- and G-update units employ the LSTM \cite{160} to respectively update the cell states and hidden states of
the source and target, i.e.
\begin{displaymath}
\begin{array}{l}
\left(C_{v_s}^{(t)},h_{v_s}^{(t)}\right)=\text{LSTM}_s\left(C_{v_s}^{(t-)},h_{v_s}^{(t-)},\Delta t_s\right),\quad \left(C_{v_g}^{(t)},h_{v_g}^{(t)}\right)=\text{LSTM}_g\left(C_{v_g}^{(t-)},h_{v_g}^{(t-)},\Delta t_g\right).
\end{array}
\end{displaymath}
The merging unit adopts the similar functions to the interacting unit to respectively merge $h_{v_s}^{(t)}$ and $h_{v_s}^{t-}$, and $h_{v_g}^{(t)}$ and $h_{v_g}^{t-}$.
The propagation component can propagate information from two interacting nodes ($v_s$ and $v_g$) to influenced nodes (i.e. their neighbors). It also consists of three units: the interacting unit,
the propagation unit and the merge unit, which are defined similarly to the update component except that they have different learnable parameters.
The literature \cite{105} addresses the vertex- and graph-focused prediction tasks on dynamic graphs with a fixed node set by combining GCNs, LSTMs and fully connected layers.
The Variational Graph Recurrent Neural Networks (VGRNNs) \cite{113} is essentially a variational graph auto-encoder whose encoder integrates the GCN and RNN into a graph RNN (GRNN) framework
and decoder is a joint probability distribution of a multi-variate Gaussian distribution and Bernoulli Distribution. The semi-implicit variational inference is employed to approximate the posterior
so as to generate the node embedding.
\subsubsection{GRNNs based on Vanilla RNNs}
The GRNNs based on vanilla RNNs firstly employ random walk techniques or traversal methods, e.g. Breadth-First Search (BFS) and Depth-First Search (DFS),
to obtain a collection of node sequences, and then leverage a RNN, e.g. LSTM and GRU, to capture long short-term dependencies.
The literature \cite{116} performs joint random walks on attributed networks, and utilizes them to boost the deep node representation learning.
The proposed GraphRNA in \cite{116} consists of two key components, namely a collaborative walking mechanism AttriWalk and a tailored deep embedding architecture for joint random walks GRN.
Suppose $\mathcal{A}_G$ denotes the node-attribute matrix of size $N\times M$. The AtriWalk admits the transition matrix of size $\mathbb{R}_+^{(N+M)\times(N+M)}$ which is written as
\begin{displaymath}
\mathcal{T}=
\left[
\begin{array}{cc}
\alpha A_G & (1-\alpha)\mathcal{A}_G \\
(1-\alpha)\mathcal{A}_G^T & 0
\end{array}
\right].
\end{displaymath}
After obtaining the sequences via the collaborative random walk, the bi-directional GRU \cite{115} and pooling operator are employed to learn the global representations of sequences.
The literature \cite{111} leverages the BFS node ordering and truncation strategy to obtain a collection of node representation sequences,
and then uses the GRU model and variational auto-regression regularization to perform the graph classification.
\section{Extensions and Applications}
The aforementioned architectures essentially provide ingredients of constructing the GNNs for us. Below, we investigate the extensions of the GNNs from the next 8 aspects:
GCNNs on spectral graphs, capability and interpretability, deep graph representation learning, deep graph generative models, combinations of the PI and GNNs,
adversarial attacks for the GNNs, graph neural architecture search and graph reinforcement learning, and briefly summarize the applications of the GNNs at last.
\subsection{GCNNs on Special Graphs}
The vanilla GCNNs aims at learning the representations of input graphs (directed or undirected, weighted or unweighted).
The real-world graphs may have more additional characteristics, e.g. spatial-temporal graphs, heterogeneous graphs, hyper-graphs, signed graphs and so all.
The GCNN for signed graphs \cite{28} leverage the balance theory to aggregate and propagate information through positive and negative links.
\subsubsection{Heterogeneous Graphs.}
Heterogeneous Graphs are composed of nodes and edges of different types, and each type of edges is called a relation between two types of nodes.
For example, a bibliographic information network contains at least 4 types of nodes, namely Author, Paper, Venue and Term, and at least 3 types of edges, namely Author-Paper, Term-Paper and Venue-Paper \cite{395}.
The heterogeneity and rich semantic information brings great challenges for designing heterogeneous graph convolutional neural networks. In general, a heterogeneous graph can be denoted as $H=(V,E,\nu,\zeta)$,
where $\nu(v)$ denotes the type of node $v\in V$ and $\zeta(e)$ denotes the type of edge $e\in E$. Let $\mathcal{T}^v$ and $\mathcal{T}^e$ respectively denote the set of node types and edge types.
Below, we summarize the vanilla heterogeneous GCNNs, namely HetGNNs \cite{61}.
\textbf{Vanilla Heterogeneous GCNNs.}
The Heterogeneous Graph Neural Networks (HetGNNs) \cite{61} aims to resolve the issue of jointly considering heterogeneous structural information
as well as heterogeneous content information of nodes. It firstly samples a fixed size of strongly correlated heterogeneous neighbors for each node via a Random Walk with Restart (RWR)
and groups them into different node types. Then, it aggregates feature information of those sampled neighboring nodes via a bi-directional Long Short Term Memory (LSTM) and attention mechanism.
Running RWR with a restart probability $p$ from node $v$ will yield a collection of a fixed number of nodes, denoted as $\text{RWR}(v)$. For each node type $t$,
the $t\text{-type}$ neighbors $N_G^t(v)$ of node $v$ denotes the set of $\text{top-}k_t$ nodes from $\text{RWR}(v)$ with regard to frequency.
Let $\mathcal{C}_v$ denote the heterogeneous contents of node $v$, which can be encoded as a fixed size embedding via a function $f_1(v)$, i.e.
\begin{displaymath}
f_1(v)=\dfrac{\sum_{j\in\mathcal{C}_v}\left[\overrightarrow{\text{LSTM}}\left(\mathcal{FC}_{\theta_x}(x_j)\right)\bowtie\overleftarrow{\text{LSTM}}\left(\mathcal{FC}_{\theta_x}(x_j)\right)\right]}{|\mathcal{C}_v|},
\end{displaymath}
where $\mathcal{FC}_{\theta_x}(\cdot)$ denotes feature transformer, e.g. identity or fully connected neural networks with parameter $\theta_x$, and
$\overrightarrow{\text{LSTM}}\left(\cdot\right)$ and $\overleftarrow{\text{LSTM}}\left(.\right)$ is defined by the Eq. (\ref{LSTM Formulation}).
The content embedding of the $t\text{-type}$ neighbors of node $v$ can be aggregated as follows,
\begin{displaymath}
\begin{array}{lcl}
f_2^t(v) & = & \text{\textsc{Aggregate}}^T\left(\left\{f_1(v'):v'\in N_G^t(v)\right\}\right) \\[2mm]
& = & \dfrac{\sum_{v'\in N_G^t(v)}\left[\overrightarrow{\text{LSTM}}(f_1(v'))\bowtie\overleftarrow{\text{LSTM}}(f_1(v'))\right]}{|N_G^t(v)|}.
\end{array}
\end{displaymath}
Let $\mathcal{F}(v)=\{f_1(v)\}\cup\{f_2^t(v):t\in\mathcal{T}^v\}$.
As a result, the output embedding of node $v$ can be obtained via the attention mechanism, i.e. $\mathcal{E}_v=\sum_{f_j(v)\in\mathcal{F}(v)}\omega_{v,j}f_j(v)$,
where the attention weights is computed by
$\omega_{v,j}=\dfrac{\exp\left(\rho\left(u^t\left[f_j(v)\bowtie f_1(v)\right]\right)\right)}{\sum_{f_j(v)\in\mathcal{F}(v)}\exp\left(\rho\left(u^t\left[f_j(v)\bowtie f_1(v)\right]\right)\right)}$.
In addition, the GraphInception \cite{397} can be employed to learn the hierarchical relational features on heterogeneous graphs
by converting the input graph into a multi-channel graph (each meta path as a channel) \cite{62}.
\textbf{Heterogeneous Graph Attention Mechanism.} The literature \cite{75} firstly proposes a hierarchical attention based
heterogeneous GCNNs consisting of node-level and semantic-level attentions. The node-level attention aims to learn the attention weights of a node and its meta-path-based neighbors,
and the semantic-level attention aims to learn the importance of different meta-paths. More specifically, given a meta path $\Phi$, the node-level attention weight of a node $v_i$
and its meta-path-based neighbors $v_j\in N_G^{\Phi}(v_i)$ is defined to be
\begin{displaymath}
\omega_{j,k}^{\Phi}=\dfrac{\exp\left(\rho\left(a_{\Phi}^T(M_{\nu(v_j)}x_j\bowtie M_{\nu(v_k)}x_k)\right)\right)}{\sum_{v_k\in N_G^{\Phi}(v_j)}\exp\left(\rho\left(a_{\Phi}^T(M_{\nu(v_j)}x_j\bowtie M_{\nu(v_k)}x_k)\right)\right)},
\end{displaymath}
where $M_{\nu(v_j)}$ transforms the feature vectors of nodes of type $\nu(v_j)$ in different vector spaces into a unified vector space.
The embedding of node $v_i$ under the meta path $\Phi$ can be computed by
\begin{displaymath}
x_j^{\Phi,l+1}=\bowtie_{k=1}^K\rho\left(\sum_{k\in N_G^{\Phi}(v_j)}\omega_{j,k}^{\Phi}M_{\nu(v_k)}x_k^{\Phi,l}\right).
\end{displaymath}
Given a meta-path set $\left\{\Phi_0,\cdots,\Phi_P\right\}$, performing the node-level attention layers under each meta path will yield a set of semantic-specific node representations,
namely $\left\{X^{\Phi_0},\cdots,X^{\Phi_P}\right\}$. The semantic-level attention weight of the meta path $\Phi_j$ is defined as
\begin{displaymath}
\beta^{\Phi_j}=\dfrac{\exp\left(\omega^{\Phi_j}\right)}{\sum_{p=1}^P\exp\left(\omega^{\Phi_p}\right)},
\end{displaymath}
where $\omega^{\Phi_p}=\frac{1}{|V|}\sum_{v_k\in V}q^T\tanh\left(Wx_k^{\Phi_p}+b\right)$. As a result, the embedding matrix $X=\sum_{p=1}^P\beta^{\Phi_p}X^{\Phi_p}$.
In addition, there are some available studies on the GCNNs for multi-relational graphs \cite{36,68,361} and the transformer for dynamic heterogeneous graphs \cite{60,114}.
\subsubsection{Spatio-Temporal Graphs.}
Spatio-temporal graphs can be used to model traffic networks \cite{59,71} and skeleton networks \cite{55,57}.
In general, a spatio-temporal graph is denoted as $G_{ST}=(V_{ST},E_{ST})$ where $V_{ST}=\{v_{t,j}:t=1,\cdots,T,j=1,\cdots,N_{ST}\}$.
The edge set $E_{ST}$ is composed of two types of edges, namely spatial edges and temporal edges. All spatial edges $(v_{t,j},v_{t,k})\in E_{ST}$ are collected in the intra-frame edge set $E_S$,
and all temporal edges $(v_{t,j},v_{t+1,j})\in E_{ST}$ are collected in the inter-frame edge set $E_T$.
The literature \cite{59} proposes a novel deep learning framework, namely Spatio-Temporal Graph Convolutional Networks (STGCN), to tackle the traffic forecasting problem.
Specifically, STGCN consists of several layers of spatio-temporal convolutional blocks, each of which has a "sandwich" structure with two temporal gated convolution layers (abbreviated as Temporal Gated-Conv)
and a spatial graph convolution layer in between (abbreviated as Spatial Graph-Conv). The Spatial Graph-Conv exploits the conventional GCNNs to extract the spatial features,
whereas the Temporal Gated-Conv the temporal gated convolution operator to extract temporal features. Suppose that the input of the temporal gated convolution for each node is a $\text{length-}M$ sequence
with $C_{\text{in}}$ channels, i.e. $X\in\mathbb{R}^{M\times C_{\text{in}}}$. The temporal gated convolution kernel $\Gamma\in\mathbb{R}^{K\times C_{\text{in}}\times 2C_{\text{out}}}$ is used to
filter the input $Y$, i.e.
\begin{displaymath}
\Gamma\ast_T Y=P\circledast\sigma(Q)\in\mathbb{R}^{(M-K+1)\times C_{\text{out}}},
\end{displaymath}
to yield an output $P\bowtie Q\in\mathbb{R}^{(M-K+1)\times(2C_{\text{out}})}$.
The Spatial Graph-Conv takes a tensor $\underline{X}^{(l)}\in\mathbb{R}^{M\times N_{ST}\times C^{(l)}}$ as input, and outputs a tensor $\underline{X}^{(l+1)}\in\mathbb{R}^{(M-2(K-1))\times N_{ST}\times C^{(l+1)}}$, i.e.
\begin{displaymath}
\underline{X}^{(l+1)}=\Gamma_1^{(l)}\ast_T\rho\left(\Theta^{(l)}\ast_G\left(\Gamma_0^{(l)}\ast_T X^{(l)}\right)\right),
\end{displaymath}
where $\Gamma_{0}^{(l)},\Gamma_1^{(l)}$ are the upper and lower temporal kernel and $\Theta^{(l)}$ is the spectral kernel of the graph convolution.
In addition, some other studies pay attention to the GCNNs on the spatio-temporal graphs from other perspectives, e.g. Structural-RNN \cite{56} via a factor graph representation of the spatio-temporal graph
and GCRNN \cite{108,109} combining the vanilla GCNN and RNN.
\subsubsection{Hypergraphs.}
The aforementioned GCNN architectures are concerned with the conventional graphs consisting of pairwise connectivity between two nodes.
However, there could be even more complicated connections between nodes beyond the pairwise connectivity, e.g. co-authorship networks. Under such circumstances, a hypergraph, as a generalization to the convectional graph,
provides a flexible and elegant modeling tools to represent these complicated connections between nodes. A hypergraph is usually denoted as $\mathcal{G}=(\mathcal{V},\mathcal{E},\omega)$,
where $\mathcal{V}=\{v_1,\cdots,v_N\}$ like the conventional graph, $\mathcal{E}=\{e_1,\cdots,e_M\}$ is a set of $M$ hyperedges.
$\omega(e_k)$ denote weights of hyperedges $e_k\in\mathcal{E}$. A non-trivial hyperedge is a subset of $\mathcal{V}$ with at least 2 nodes.
In particular, a trivial hyperedge, called a self-loop, is composed of a single node. The hypergraph $\mathcal{G}$ can also be denoted by an incidence matrix $\mathcal{H}_{\mathcal{G}}\in\mathbb{R}^{N\times M}$,
i.e.
\begin{displaymath}
\mathcal{H}_{\mathcal{G}}[j,k]=
\left\{
\begin{array}{ll}
0, & v_j\notin e_k \\
1, & v_j\in e_k.
\end{array}
\right.
\end{displaymath}
For a node $v_j\in\mathcal{V}$, its degree $\text{deg}_{\mathcal{V}}(v_j)=\sum_{e_k\in\mathcal{E}}\omega(e_k)\mathcal{H}_{\mathcal{G}}[j,k]$.
For a hyperedge $e_k\in\mathcal{E}$, its degree $\text{deg}_{\mathcal{E}}(e_k)=\sum_{v_j\in e_k}\mathcal{H}_{\mathcal{G}}[j,k]$.
Let $\mathcal{D}_{\mathcal{V}}=\text{diag}\left(\text{deg}_{\mathcal{V}}(v_1),\cdots,\text{deg}_{\mathcal{V}}(v_N)\right)$,
$\mathcal{D}_{\mathcal{E}}=\text{diag}\left(\text{deg}_{\mathcal{E}}(e_1),\cdots,\text{deg}_{\mathcal{E}}(e_M)\right)$, and
$\mathcal{W}_{\mathcal{G}}=\text{diag}\left(\omega(e_1),\cdots,\omega(e_M)\right)$.
The hypergraph Laplacian \cite{65} $\mathcal{L}_{\mathcal{G}}$ of $\mathcal{G}$ is defined to be
\begin{displaymath}
\mathcal{L}_{\mathcal{G}}=\mathcal{I}_N-\mathcal{D}_{\mathcal{V}}^{-\frac{1}{2}}\mathcal{H}_{\mathcal{G}}\mathcal{W}_{\mathcal{G}}\mathcal{D}_{\mathcal{E}}^{-1}\mathcal{H}_{\mathcal{G}}^T\mathcal{D}_{\mathcal{V}}^{-\frac{1}{2}}.
\end{displaymath}
It can also be factorized by the eigendecomposition, i.e. $\mathcal{L}_{\mathcal{G}}=\mathcal{U}\Lambda\mathcal{U}^T$.
The spectral hypergraph convolution operator, the Chebyshev hypergraph convolutional neural network and the hypergraph convolutional network
can be defined in analogy to the Eqs (\ref{Spectral Graph Convolution},\ref{Chebyshev Spectral Filter},\ref{GCN-Layer}).
The HyperGraph Neural Network (HGNN) architecture proposed in the literature \cite{65} is composed of multiple layers of the hyperedge convolution, and each layer is defined as
\begin{displaymath}
X^{(l+1)}=\rho\left(\mathcal{D}_{\mathcal{V}}^{-\frac{1}{2}}\mathcal{H}_{\mathcal{G}}\mathcal{W}_{\mathcal{G}}\mathcal{D}_{\mathcal{E}}^{-1}\mathcal{H}_{\mathcal{G}}^T\mathcal{D}_{\mathcal{V}}^{-\frac{1}{2}}X^{(l)}\Theta^{(l)}\right).
\end{displaymath}
In essence, the HGNN essentially views each hyperedge as a complete graph so that the hypergraph is converted into a conventional graph.
Treating each hyperedge as a complete graph obviously incurs expensive computational cost. Hence, some studies \cite{63,64} propose various approaches to approximate the hyperedges.
The HNHN \cite{174} interleaves updating the node representations with the hyperedge representations by the following formulas,
\begin{displaymath}
\mathcal{X}_{\mathcal{V}}^{(l+1)}=\rho\left(\mathcal{D}_{\mathcal{V}}^{-1}\mathcal{H}_{\mathcal{G}}\mathcal{X}_{\mathcal{E}}^{(l)}\Theta_{\mathcal{V}}^{(l)}\right),\quad
\mathcal{X}_{\mathcal{E}}^{(l+1)}=\rho\left(\mathcal{D}_{\mathcal{E}}^{-1}\mathcal{H}_{\mathcal{G}}^{T}\mathcal{X}_{\mathcal{V}}^{(l)}\Theta_{\mathcal{E}}^{(l)}\right).
\end{displaymath}
\subsection{Capability and Interpretability}
The GCNNs have achieved tremendous empirical successes over the supervised, semi-supervised and unsupervised learning on graphs. Recently,
many studies start to put their eyes on the capability and interpretability of the GCNNs.
\subsubsection{Capability}
The capability of the GCNNs refers to their expressive power. If two graphs are isomorphic, they will obviously output the same representations of nodes/edges/graph.
Otherwise, they should output different representations. However, two non-isomorphic graphs maybe output the same representations in practice.
This is the theoretical limitations of the GCNNs.
As described in the literatures \cite{83,88,360}, The $1\text{-hop}$ spatial GCNNs (1-GCNNs) have the same expressive power as the $1\text{-dimensional}$ Weisfeiler-Leman (1-WL)
graph isomorphism test in terms of distinguishing non-isomorphic graphs.
The 1-WL iteratively update the colors of nodes according to the following formula
\begin{displaymath}
C_{l}^{(l+1)}(v)=\text{\textsc{Hash}}\left(C_{l}^{(l)}(v),\left\{\hspace{-1.5mm}\left\{C_{l}^{(l)}(u):u\in N_G(v)\right\}\hspace{-1.5mm}\right\}\right).
\end{displaymath}
According to the literatures \cite{83,88}, we have that the 1-GCNN architectures do not have more power in terms of distinguishing two non-isomorphic graphs than the 1-WL heuristic. Nevertheless,
they have equivalent power if the aggregation and update functions are injective. In order to overcome
the theoretical limitations of the GCNNs, the literature \cite{83} proposes a Graph Isomorphism Network (GIN) architecture, i.e.
\begin{displaymath}
\begin{array}{rl}
A_v^{(l+1)}& =\text{\textsc{Aggregate}}^{(l+1)}\left(\left\{\left\{X^{(l)}[u,:]:u\in N_G(v)\right\}\right\}\right) \\[2mm]
& \triangleq\sum_{u\in N_G(v)}X^{(l)}[u,:] \\[2mm]
X^{(l+1)}[v,:] & =\text{\textsc{Update}}^{(l+1)}\left(X^{(l)}[v,:],A_v^{(l+1)}\right) \\ [2mm]
& \triangleq\text{MLP}\left((1+\epsilon^{(l+1)})X^{(l)}[v,:]+A_v^{(l+1)}\right),
\end{array}
\end{displaymath}
where $\epsilon^{(l)}$ is a scalar parameter. The literature \cite{163} studies the expressive power of the spatial GCNNs, and presents two results:
(1) The spatial GCNNs are shown to be a universal approximator under sufficient conditions on their depth, width, initial node features and layer expressiveness;
(2) The power of the spatial GCNNs is limited when their depth and width is restricted.
In addition, there are some other studies on the capability of the GCNNs from different perspectives, e.g. the first order logic \cite{164},
$p\text{-order}$ graph moments \cite{167}, algorithmic alignment with the dynamic programming \cite{401}, generalization and representational limits of the GNNs \cite{407}.
\subsubsection{Interpretability}
Interpretability plays a vital role in constructing a reliable and intelligent learning systems.
Although some studies have started to explore the interpretability of the conventional deep learning models, few of studies put their eyes on the interpretability of the GNs \cite{85}.
The literature \cite{165} bridges the gap between the empirical success of the GNs and lack of theoretical interpretations. More specifically, it considers two classes of techniques:
(1) gradient based explanations, e.g. sensitivity analysis and guided back-propagation; (2) decomposition based explanations, e.g. layer-wise relevance propagation and Taylor decomposition.
The GNNExplainer \cite{166} is a general and model-agnostic approach for providing interpretable explanations for any spatial GCNN based model in terms of graph machine learning tasks.
Given a trained spatial GCNN model $\Phi$ and a set of predictions, the GNNExplainer will generate a single-instance explanation by identifying a subgraph
of the computation graph and a subset of initial node features, which are the most vital for the prediction of the model $\Phi$. In general, the GNNExplainer can be formulated as an optimization problem
\begin{equation}\label{GNNExplainer optimization framework}
\max_{G_S,X_S^F} I\left(Y,(G_S,X_S^F)\right)=H\left(Y\right)-H\left(Y|G=G_S,X=X_S^F\right),
\end{equation}
where $I(\cdot,\cdot)$ denotes the mutual information of two random variables,
$G_S$ is a small subgraph of the computation graph and $X_S^F$ is a small subset of node features $\left\{X^F[j,:]:v_j\in G_S\right\}$.
The entropy term $H(Y)$ is constant because the spatial GCNN model $\Phi$ is fixed.
In order to improve the tractability and computational efficiency of the GNNExplainer, the final optimization framework is reformulated as
\begin{displaymath}
\min_{M,F} -\sum_{c=1}^C\mathbb{I}[y=c]\log P_{\Phi}\left(Y=y|G=A_G\circledast\sigma(M),X=X_S^F\right).
\end{displaymath}
In addition, the GNNExplainer also provides multi-instances explanations based on graph alignments and prototypes so as to answer questions like "How did a GCNN predict that a given set of nodes all have label $c$?".
\subsection{Deep Graph Representation Learning}
Graph representation learning (or called network embedding) is a paradigm of unsupervised learning on graphs. It gains a large amount of popularity since the DeepWalk \cite{154}.
Subsequently, many studies exploit deep learning techniques to learn low-dimensional representations of nodes \cite{139}. In general, the network embedding via the vanilla deep learning
techniques learn low-dimensional feature vectors of nodes by utilizing either stacked auto-encoders to reconstruct the adjacent or positive point-wise mutual information features \cite{124,132,133,136,141,142}
or RNNs to capture long and short-term dependencies of node sequences yielded by random walks \cite{134,135}. In the following, we introduce the network embedding approaches based on GNNs.
\subsubsection{Network Embedding based on GNNs}
In essence, the GNNs provides an elegant and powerful framework for learning node/edge/graph representations. The majority of the GCNNs and GRNNs are concerned with
semi-supervised learning (i.e. node-focused tasks) or supervised learning (i.e. graph-focused) tasks. Here, we review the GNN based unsupervised learning on graphs.
In general, the network embedding based on GNNs firstly utilize the GCNNs and variational auto-encoder to generate gaussian-distributed hidden states of nodes, and then reconstruct the adjacency matrix
and/or the feature matrix of the input graph \cite{87,127,129,131,138,148}. A representative approach among these ones is the Variational Graph Auto-Encoder (VGAE) \cite{87}
consisting of a GCN based encoder and an inner product decoder.
The GCN based encoder is defined to be
\begin{displaymath}
q(Z|X,A_G)=\prod_{j=1}^Nq(Z[j,:]|X,A_G)=\prod_{j=1}^N\mathcal{N}\left(Z[j,:]|\mu_j,\text{diag}(\sigma_j^2)\right),
\end{displaymath}
where $\mu_j=\text{GCN}_{\mu}\left(X,A_G\right)$ and $\log\sigma_j=\text{GCN}_{\sigma}(X,A_G)$.
The inner product decoder is defined to be
\begin{displaymath}
p(A_G|Z)=\prod_{j=1}^N\prod_{k=1}^Np\left(A_G[j,k]|Z[j,:],Z[k,:]\right)=\prod_{j=1}^N\prod_{k=1}^N\sigma\left(Z[j,:]Z[k,:]^T\right).
\end{displaymath}
They adopt the evidence lower bound \cite{151} as their objective function.
The adversarially regularized (variational) graph autoencoder \cite{129} extends the VGAE by adding an adversarial regularization term to the evidence lower bound.
The literature \cite{128} proposes a symmetric graph convolutional autoencoder which produces a low-dimensional latent nodes representations. Its encoder employs
the Laplacian smoothing \cite{16} to jointly encode the structural and attributed information,
and its decoder is designed based on Laplacian sharpening as the counterpart of the Laplacian smoothing of the encoder. The Laplacian sharpening is defined to be
$X^{(l+1)}=(1+\gamma)X^{(l)}-\gamma D^{-1}AX^{(l)}=X^{(l)}+\gamma(\mathbb{I}_N-D^{-1}A)X^{(l)}$,
which allows utilizing the graph structure in the whole processes of the proposed autoencoder architecture.
In addition, there are some other methods to perform the unsupervised learning on graphs, which do not rely on the reconstruction of the adjacency and/or feature matrix,
e.g. the graph auto-encoder on directed acyclic graphs \cite{79}, pre-training GNNs via context prediction and attribute masking strategies \cite{130},
and deep graph Infomax using a noise-contrastive type objective with a standard binary cross-entropy loss between positive examples and negative examples \cite{140}.
\subsection{Deep Graph Generative Models}
The aforementioned work concentrates on embedding an input graph into a low-dimensional vector space so as to perform semi-supervised/supervised/unsupervised learning tasks on graphs.
This subsection introduces deep graph generative models aiming to mimic real-world complex graphs.
Generating complex graphs from latent representations is confronted with great challenges due to high nonlinearity and arbitrary connectivity of graphs.
Note that graph translation \cite{145} is akin to graph generation. However, their difference lies in that the former takes two graphs, i.e. input graph and target graph, as input,
and the latter only takes a single graph as input. The NetGAN \cite{154} utilizes the generative adversarial network \cite{149} to mimic the input real-world graphs.
More specifically, it is composed of two components, i.e. a generator $G$ and a discriminator $D$, as well. The discriminator $D$ is modeled as a LSTM
in order to distinguish real node sequences, which are yielded by the second-order random walks scheme node2vec \cite{155}, from faked ones.
The generator $G$ aims to generate faked node sequences via another LSTM, whose generating process is as follows.
\begin{displaymath}
\begin{array}{l}
v_0=0, \quad z\sim N_m(0,\mathbb{I}_m), \quad (C_0,h_0)=g_{\theta'}(z), \\ [2mm]
(C_j,h_j,o_j)=\text{LSTM}_G(C_{j-1},h_{j-1},v_{j-1}),\quad v_j\sim\text{Cat}\left(\text{Softmax}(o_j)\right).
\end{array}
\end{displaymath}
The motif-targeted GAN \cite{147} generalizes random walk based architecture of the NetGAN to characterize mesoscopic context of nodes.
Different from \cite{147,154}, the GraphVAE \cite{146} adopts a probabilistic graph decoder to generate a probabilistic fully-connected graph, and then employs approximate graph matching to reconstruct the input graph.
Its reconstruction loss is the cross entropy between the input and reconstructed graphs.
The literature \cite{150} defines a sequential decision-making process to add a node/edge via the graph network \cite{85}, readout operator and softmax function.
The GraphRNN \cite{152} is a deep autoregressive model, which generates graphs by training on a representative set of graphs and decomposes the graph generation process into
a sequence of node and edge formations conditioned on the current generated graph.
\subsection{Combinations of the PI and GNNs}
The GNNs and PI are two different learning paradigms for complicated real-world data.
The former specializes in learning hierarchical representations based on local and global structural information, and the latter learning the dependencies between random variables.
This subsection provides a summarization of studies of combining these two paradigms.
\subsubsection{Conditional Random Field Layer Preserving Similarities between Nodes}
The literature \cite{162} proposes a CRF layer for the GCNNs to enforce hidden layers to preserve similarities between nodes.
Specifically, the input $X^{(l)}$ to $(l+1)\text{-th}$ layer is a random vector around the output $B^{(l)}=\text{GCNN}(X^{(l-1)},A_G, X)$ of the $(l-1)\text{-th}$ layer.
The objective function for the GCNN with a CRF layer can be reformulated as
\begin{displaymath}
\begin{array}{l}
J\left(W;X,A_G,Y\right) =\mathcal{L}\left(Y;B^{(L)}\right)+\displaystyle\sum_{l=1}^{L-1}\left(\frac{\gamma}{2}\|X^{(l)}-B^{(l)}\|_F^2+\mathcal{R}(X^{(l)})\right),
\end{array}
\end{displaymath}
where the first term after "=" is the conventional loss function for semi-supervised node classification problem, and
the last term is a regularization one implementing similarity constraint. The similarity constraint $\mathcal{R}(X^{(l)})$ is modeled as a CRF, i.e.
$p\left(X^{(l)}|B^{(l)}\right)=\frac{1}{Z(B^{(l)})}\exp\left(-E\left(X^{(l)}|B^{(l)}\right)\right)$
where the energy function $E\left(X^{(l)}|B^{(l)}\right)$ is formulated as
\begin{displaymath}
\begin{array}{ll}
& E\left(X^{(l)}|B^{(l)}\right) \\
= & \displaystyle\sum_{v\in V}\varphi_v(X^{(l)}[v,:],B^{(l)}[v,:])+\displaystyle\sum_{(u,v)\in E}\varphi_{u,v}\left(X^{(l)}[u,:],X^{(l)}[v,:],B^{(l)}[u,:],B^{(l)}[v,:]\right).
\end{array}
\end{displaymath}
Let $s_{u,v}$ denote the similarity between $u$ and $v$. The unary energy component $\varphi_v\left(\cdot,\cdot\right)$ and pairwise energy component
$\varphi_{u,v}\left(\cdot,\cdot,\cdot,\cdot\right)$ for implementing the similarity constraint are respectively formulated as
\begin{displaymath}
\begin{array}{l}
\varphi_v\left(X^{(l)}[v,:],B^{(l)}[v,:]\right)=\|X^{(l)}[v,:]-B^{(l)}[v,:]\|_2^2, \\[1.5mm]
\varphi_{u,v}\left(X^{(l)}[u,:],X^{(l)}[v,:],B^{(l)}[u,:],B^{(l)}[v,:]\right)=s_{u,v}\|X^{(l)}[u,:]-X^{(l)}[v,:]\|_2^2.
\end{array}
\end{displaymath}
The mean-field variational Bayesian inference is employed to approximate the posterior $p(B^{(l)}|X^{(l)})$. Consequently, the CRF layer is defined as
\begin{displaymath}
\left(X^{(l)}[v,:]\right)^{(k+1)}=\frac{\alpha B^{(l)}[v,:]+\beta\sum_{u\in N_G(v)}s_{u,v}\left(X^{(l)}[u,:]\right)^{(k)}}{\alpha+\beta\sum_{u\in N_G(v)}s_{u,v}}.
\end{displaymath}
\subsubsection{Conditional GCNNs for Semi-supervised Node Classification}
The conditional GCNNs incorporate the Conditional Random Field (CRF) into the conventional GCNNs so that the semi-supervised node classification can be enhanced by
both the powerful node representations and the dependencies of node labels.
The GMNN \cite{80} performs the semi-supervised node classification by incorporating the GCNN into the Statistical Relational Learning (SRL).
Specifically, the SRL usually models the conditional probability $p(Y_V|X_V)$ with the CRF, i.e.
$p(Y_V|X_V)=\frac{1}{Z(X_V)}\prod_{(i,j)\in E}\varphi_{i,j}\left(y_i,y_j,X_V\right)$,
where $y_*=Y_V[*,:], *=i,j$. Note that $Y_V$ is composed of the observed node labels $Y_L$ and hidden node labels $Y_U$.
The variational Bayesian inference is employed to estimate the posterior $p(Y_U|Y_L,X_V)$.
The objective function is defined as
\begin{displaymath}
\text{ELBO}\left(q_{\theta_v}(Y_U|X_V)\right)=\mathbb{E}_{q_{\theta_v}(Y_U|X_V)}\left[\log\left(p_{\theta_l}(Y_L,Y_U|X_V)\right)-\log\left(q_{\theta_v}(Y_U|X_V)\right)\right].
\end{displaymath}
This objective can be optimized by the variational Bayesian EM algorithm \cite{161}, which iteratively updates the variational distribution $q_{\theta_v}(Y_U|X_V)$ and the likelihood $p_{\theta_l}(Y_U|Y_L,X_V)$.
In the VBE stage, $q_{\theta_v}(Y_U|X_V)=\prod_{v\in U}q_{\theta_v}(y_v|X_V)$, and $q_{\theta_v}(y_v|X_V)$ is approximated by a GCNN.
In the VBM stage, the pseudo-likelihood is employed to approximate
\begin{displaymath}
\mathbb{E}_{q_{\theta_v}(Y_U|X_V)}\left[\sum_{v\in V}\log p_{\theta_l}\left(y_n|Y_{V\backslash v},X_V\right)\right]
=\mathbb{E}_{q_{\theta_v}(Y_U|X_V)}\left[\sum_{v\in V}\log p_{\theta_l}\left(y_n|Y_{N_G(v)},X_V\right)\right],
\end{displaymath}
and $p_{\theta_l}\left(y_n|Y_{N_G(v)},X_V\right)$ is approximated by another GCNN. The literature \cite{159} adopts the similar idea to the GMNN.
Its posterior is modeled as a CRF with unary energy components and pairwise energy components whose condition is the outputs of the prescribed GCNN.
The maximum likelihood estimation employed to estimate the model parameters.
\subsubsection{GCNN-based Gaussian Process Inference}
A Gaussian Process (GP) defines a distribution over a function space and assumes any finite collection of marginal distributions follows a multivariate Gaussian distribution.
A function $f:\mathbb{R}^d\rightarrow\mathbb{R}$ follows a Gaussian Process $\text{GP}(m(\cdot),\kappa(\cdot,\cdot))$ iff $\left(f(X_1),\cdots,f(X_N)\right)^T$ for any $N$ $d\text{-dimensional}$ random vectors.
follows a $N\text{-dimensional}$ Gaussian distribution $\mathcal{N}_N\left(\mu,\Sigma\right)$, where $\mu=\left(m(X_1),\cdots,m(X_N)\right)^T$ and $\Sigma=\left[\kappa(X_j,X_k)\right]_{N\times N}$,
For two $d\text{-dimensional}$ random vectors $X$ and $X'$, we have $\mathbb{E}\left[f(X)\right]=m(X)$ and $\text{Cov}(f(X),f(X'))=\kappa(X,X')$.
Given a collection of $N$ samples $\mathcal{D}=\left\{(X_j,y_j):j=1,\cdots N\right\}$, the GP inference aims to calculate the probability $p(y|X)$ for predictions, i.e.
\begin{displaymath}
f\sim\text{GP}(0(\cdot),\kappa(\cdot,\cdot)), y_j\sim\text{DIST}(\lambda(f(X_j))),
\end{displaymath}
where $\lambda(\cdot)$ is a link function and $\text{DIST}(\cdot)$ denotes an arbitrary feasible noise distribution. To this end, the posterior $p(f|\mathcal{D})$ needs to be calculated out firstly.
The literature \cite{156} employs amortized variational Bayesian inference to approximate $p(f|\mathcal{D})$, i.e. $f=\mu+L\epsilon$, and the GCNNs to estimate $\mu$ and $L$.
\subsubsection{Other GCNN-based Probabilistic Inference}
The literature \cite{157} combines the GCNNs and variational Bayesian inference to infer the input graph structure.
The literature \cite{158} infers marginal probabilities in probabilistic graphical models by incorporating the GCNNs to the conventional message-passing inference algorithm.
The literature \cite{168} approximates the posterior in Markov logic networks with the GCNN-enhanced variational Bayesian inference.
\subsection{Adversarial Attacks for the GNNs}
In many circumstances where classifiers are deployed, adversaries deliberately contaminate data in order to fake the classifiers \cite{149,364}.
This is the so-called adversarial attacks for the classification problems. As stated previously, the GNNs can solve semi-supervised node classification problems and supervised graph classification tasks.
Therefore, it is inevitable to study the adversarial attacks for GNNs and defense \cite{365}.
\subsubsection{Adversarial Attacks on Graphs}
The literature \cite{182} firstly proposes a reinforcement learning based attack method, which can learn a generalizable attack policy, on graphs.
This paper provides a definition for a graph adversarial attacker.
Given a sample $(G,c,y)\in\{(G_j,c_j,y_j):j=1,\cdots,M\}$ and a classifier $f\in\mathcal{F}$,
the graph adversarial attacker $g:\mathcal{F}\times\mathcal{G}\rightarrow\mathcal{G}$ attempts to modify a graph $G=(V,E)$ into $\widetilde{G}=(\widetilde{V},\widetilde{E})$, such that
\begin{displaymath}
\begin{array}{ll}
\max_{g} & \mathbb{I}(f(\widetilde{G},c)\neq y) \\ [1.5mm]
\text{s.t.} & \widetilde{G}=g(f,(G,c,y)), \\ [1.5mm]
& \mathcal{I}(G,\widetilde{G},c)=1,
\end{array}
\end{displaymath}
where $\mathcal{I}: \mathcal{G}\times\mathcal{G}\times V\rightarrow\{0,1\}$, named an equivalence indicator,
checks whether two graph $G$ and $\widetilde{G}$ are equivalent under the classification semantics.
The equivalence indicator are usually defined in two fashions, namely explicit semantics and small modifications.
The explicit semantics are defined as $\mathcal{I}\left(G,\widetilde{G},c\right)=\mathbb{I}\left(f^*(G,c)=f^*(\widetilde{G},c)\right)$ where $f^*$ is a gold standard classifier,
and the small modifications
are defined as
$\mathcal{I}\left(G,\widetilde{G},c\right)=\mathbb{I}\left(|(E-\widetilde{E})\cup(\widetilde{E}-E)|<m\right)\cdot\mathbb{I}\left(\widetilde{E}\subseteq N(G,b)\right)$
where $N(G,b)=\left\{(u,v):u,v\in V,d_G(u,v)<=b\right\}$. In order to learn an attack policy,
the attack procedure is modeled as a finite horizon Markov Decision Process (MDP) $\mathcal{M}_m(f,G,c,y)$ and is therefore optimized by Q-learning with a hierarchical Q-function.
For the MDP $\mathcal{M}_m(f,G,c,y)$, its action $a_t\in\mathcal{A}\subseteq V\times V$ at time step $t$ is defined to add or delete edges in the graph,
its state $(\widehat{G}_t,c)$ at time step $t$ is a partially modified graph with some of the edges added/deleted from $G$,
and the reward function is defined as
\begin{displaymath}
R\left(\widetilde{G},c\right)=\left\{
\begin{array}{ll}
1 & f(\widetilde{G},c)\neq y \\ [2mm]
-1 & f(\widetilde{G},c)=y.
\end{array}
\right.
\end{displaymath}
Note that the GCNNs are employed to parameterize the Q-function.
The \textsc{Nettack} \cite{181} considers attacker nodes in $\mathcal{A}$ to satisfy a feature attack constraint $X'_{u,j}\neq X_{u,j}^{(0)}\Longrightarrow u\in\mathcal{A}$,
a structure attack constraint $A'_{u,v}\neq A_{u,v}^{(0)}\Longrightarrow u\in\mathcal{A}\vee v\in\mathcal{A}$ and an equivalence indicator constraint
$\sum_{u}\sum_{j}|X_{u,j}^{(0)}-X'_{u,j}|+\sum_{u<v}|A_{u,v}^{(0)}-A'_{u,v}|\leq\Delta$, where $G'$ is derived by perturbing $G^{(0)}$.
Let $\mathcal{P}_{\Delta,\mathcal{A}}^{G0}$ denote the set of all perturbed graphs $G'$ satisfying these three constraints.
The goal is to find a perturbed graph $G'=(A',X')$ that classifies a target node $v_0$ as $c_{\text{new}}$
and maximizes the log-probability/logit to $c_{\text{old}}$, i.e.
$\max_{(A',X')\in\mathcal{P}_{\delta,\mathcal{A}}^{G0}}\max_{c_{\text{new}}\neq c_{\text{old}}}\log Z_{v_0,c_{\text{new}}}^*-\log Z_{v_0,c_{\text{old}}}^*$
where $Z^*=f_{\theta^*}(A',X')$ with $\theta^*=\arg\min_{\theta}\mathcal{L}(\theta;A',X')$. The \textsc{Nettack} employs the GCNNs to model the classifier.
The literature \cite{183} adopts the similar equivalence indicator, and poisoning attacks are mathematically formulated as a bilevel optimization problem, i.e.
\begin{equation}\label{bilevel optimization}
\begin{array}{ll}
\min_{\widetilde{G}} & \mathcal{L}_{\text{attack}}\left(f_{\theta^*}\left(\widetilde{G}\right)\right) \\ [2mm]
s.t. & \widetilde{G}\in\mathcal{P}_{\Delta,\mathcal{A}}^{G0} \\ [2mm]
& \theta^*=\arg\min_{\theta}\mathcal{L}_{\text{train}}\left(f_{\theta}\left(\widetilde{G}\right)\right).
\end{array}
\end{equation}
This bilevel optimization problem in formula (\ref{bilevel optimization}) is then tackled using meta-gradients,
whose core idea is to treat the adjacency matrix of the input graph as a hyperparameter.
\subsubsection{Defense against the Adversarial Attacks} A robust GCNN requires that it is invulnerable to perturbations of the input graph.
The robust GCN (RGCN) \cite{180} can fortify the GCNs against adversarial attacks. More specifically, it adopts Gaussian
distributions as the hidden representations of nodes, i.e.
\begin{displaymath}
X^{(l+1)}[j,:]\sim N\left(\mu_j^{(l+1)},\text{diag}(\sigma_j^{(l+1)})\right),
\end{displaymath}
in each graph convolution layer so that the effects of adversarial attacks can be absorbed into the variances of the Gaussian distributions.
The Gaussian based graph convolution is defined as
\begin{displaymath}
\mu_j^{(l+1)}=\rho\left(\displaystyle\sum_{v_k\in N_G(v_j)}\frac{\mu_k^{(l)}\circledast\alpha_k^{(l)}}{\sqrt{\widetilde{D}_{j,j}\widetilde{D}_{k,k}}}W_{\mu}^{(l)}\right),
\sigma_j^{(l+1)}=\rho\left(\displaystyle\sum_{v_k\in N_G(v_j)}\frac{\sigma_k^{(l)}\circledast\alpha_k^{(l)}\circledast\alpha_k^{(l)}}{\widetilde{D}_{j,j}\widetilde{D}_{k,k}}W_{\sigma}^{(l)}\right),
\end{displaymath}
where $\alpha_j^{(k)}$ are attention weights. Finally, the overall loss function is defined as regularized cross-entropy.
The literature \cite{176} presents a batch virtual adversarial training method
which appends a novel regularization term to the conventional objective function of the GCNNs, i.e.
\begin{displaymath}
\mathcal{L}=\mathcal{L}_0+\alpha\cdot\frac{1}{N}\sum_{u\in V}E(p(y|X_u,W))+\beta\cdot\mathcal{R}_{vadv}(V,W),
\end{displaymath}
where $\mathcal{L}_0$ is an average cross-entropy loss of all labelled nodes, $E(\cdot)$ is the conditional entropy of a distribution,
and $\mathcal{R}_{\text{vadv}}(V,W)$ is the average Local Distributional Smoothness (LDS) loss for all nodes. Specifically,
$\mathcal{R}_{\text{vadv}}(V,W)=\frac{1}{N}\sum_{u\in V}\text{LDS}(X[u,:],W,r_{\text{vadv},u})$ where $\text{LDS}(x,w,r_{\text{vadv}})=D_{KL}\left(p(y|x,\widehat{W})||p(y|x+r_{\text{vadv}},W)\right)$
and $r_{\text{vadv}}$ is the virtual adversarial perturbation. Additionally, there are some other studies aiming at verifying certifiable (non-)robustness to structure and feature perturbations for the GCNNs
and developing robust training algorithm \cite{177,178}. The literature \cite{180} proposes to improve GCN generalization by minimizing the expected loss
under small perturbations of the input graph.
Its basic assumption is that the adjacency matrix $A_G$ is perturbed by some random noises. Under this assumption, the objective function is defined as
$\min_{W}\int q(\epsilon|\alpha)\mathcal{L}(X,Y,A_G(\epsilon),W)d\epsilon$, where $A_G(\epsilon)$ denotes the perturbed adjacency matrix of $G$ and $q(\epsilon|\alpha)$ is a zero-centered density of the noise $\epsilon$
so that the learned GCN is robust to these noises and generalizes well.
\subsection{Graph Neural Architecture Search}
Neural Architecture Search (NAS) \cite{171} has achieved tremendous success in discovering
the optimal neural network architecture for image and language learning tasks. However, existing NAS algorithms cannot be
directly generalized to find the optimal GNN architecture. Fortunately, there have been some studies to bridge this gap.
The graph neural architecture search \cite{172,173} aims to search for an optimal GNN architecture within a designed search space.
It usually exploits a reinforcement learning based controller, which is a RNN, to greedily validate the generated architecture, and then the validation results are fed back to the controller.
The literature \cite{175} proposes a Graph HyperNetwork (GHN) to amortize the search cost of training thousands of different networks,
which is trained to minimize the training loss of the sampled network with the weights generated by a GCNN.
\subsection{Graph Reinforcement Learning}
The GNNs can also be combined with the reinforcement learning so as to solve sequential decision-making problems on graphs.
The literature \cite{188} learns to walk over a graph from a source node towards a target node for a given query via reinforcement learning.
The proposed agent M-Walk is composed of a deep RNN and Monte Carlo Tree Search (MCTS). The former maps a hidden vector representation $h_t$,
yielded by a special RNN encoding the state $s_t$ at time $t$, to a policy and Q-values, and the latter is employed to generate trajectories yielding more positive rewards.
The NerveNet \cite{187} propagates information over the underlying graph of an agent via a GCNN, and then predicts actions for different parts of the agent.
The literature \cite{186} combines the GNNs and Deep Reinforcement Learning (DRL), named DRL+GNN, to learn, operate and generalize over arbitrary network topologies.
The DRL+GNN agent employs a GCNN to model the Q-value function.
\subsection{Applications}
In this subsection, we introduce the applications of the GNNs.
Due to the space limitation, we only list the application fields, including complex network analysis \cite{233,234,238}, combinatorial optimization \cite{215,216,223,224},
knowledge graph \cite{272,273,289}, bioinformatics \cite{201,210,211}, chemistry \cite{197,202,203}, brain network analysis \cite{205,206},
physical system \cite{102,338,339}, source code analysis \cite{340,341,342}, intelligent traffic \cite{353,355,403}, recommender systems \cite{343,344,349},
computer vision \cite{112,343,344,345,346} and natural language processing \cite{303,308,312,318,326}.
\section{Benchmarks and Evaluation Pitfalls}
In this section, we briefly introduce benchmarks and evaluation Pitfalls. The benchmarks provide ground truth for various GNN architectures so that different GNNs can be compared fairly,
the evaluation pitfalls empirically show that the existing evaluation criterion have potential pitfalls.
\subsubsection{Benchmarks}
Graph neural networks have become a powerful toolkit for mining complex graphs.
It becomes more and more critical to evaluate the effectiveness of new GNN architectures and
compare different GNN models under a standardized benchmark with consistent experimental settings and large datasets.
A feasible benchmark for the GNNs should include appropriate graph datasets, robust coding interfaces and experimental settings
so that different GNN architectures can be compared in the same settings. The literature \cite{169} makes a pioneering effort to
construct a reproducible GNN benchmarking framework in order to facilitate researchers to gauge the effectiveness of different GNN architectures.
Specifically, it releases an open-source benchmark infrastructures for GNNs, hosted on GitHub based on PyTorch and DGL libraries \cite{402},
introduces medium-scale graph datasets with 12k-70k graphs of variable sizes 9-500 nodes, and identifies important building blocks of GNNs
(graph convolutions, anistropic diffusion, residual connections and normalization layers) with the proposed benchmark infrastructures.
The literature \cite{372} presents an Open Graph Benchmark (OGB) including challenging, real-world and large-scale benchmark graph datasets, encompassing
multiple important graph deep learning tasks ranging from social and information networks to biological networks, molecular graphs and knowledge graphs,
to facilitate scalable, robust and reproducible deep learning research on graphs. The OGB datasets, which provide a unified evaluation protocol
using application-specific train/validation/test dataset splits and evaluation metrics,
releases an end-to-end graph processing pipeline including graph data loading, experimental settings and model evaluations.
\subsubsection{Evaluation Pitfalls}
The literature \cite{170} compares four typical GCNN architectures: GCN \cite{6}, MoNet \cite{18},
GAT \cite{66} and GraphSAGE using three aggregation strategies \cite{104}
against 4 baseline models: logistic regression, multi-layer perceptron, label propagation and normalized laplacian label propagation,
and uses a standardized training and hyper-parameter tuning procedure for all these models so as to perform a more fair comparison.
The experimental results show that different train/validation/test splits of datasets lead to dramatically different rankings of models.
In addition, their findings also demonstrate that simpler GCNN architectures can outperform more sophisticated ones
only if the hyper-parameters and training procedures are tuned fairly for all models.
\section{Future Research Directions}
Although the GNNs have achieved tremendous success in many fields, there still exists some open problems.
This section summarizes the future research directions of the GNNs.
\subsubsection{Highly Scalable GNNs}
The real-world graphs usually contain hundreds of millions of nodes and edges, and have dynamically evolving characteristics.
It turns out that it is difficult for the existing GNN architectures to scale up to the huge real-world graphs.
This motivates us to design highly scalable GNN architectures which can efficiently and effectively learn node/edge/graph representations
for the huge dynamically-evolving graphs.
\subsubsection{Robust GNNs}
The existing GNN architectures are vulnerable to adversarial attacks.
That is, the performance of the GNN models will sharply drop once the structure and/or initial features of the input graph are attacked by adversaries.
Therefore, we should incorporate the attack-and-defense mechanism into the GNN architectures, i.e. constructing robust GNN architecture,
so as to reinforce them against adversarial attacks.
\subsubsection{GNNs Going Beyond WL Test}
The capabilities of the spatial GCNNs are limited by the 1-WL test,
and the higher-order WL test is computationally expensive. Consequently, two non-isomorphic graphs
will produce the same node/edge/graph representations under appropriate conditions.
This motivates us to develop a novel GNN framework going beyond WL test, or design an elegant higher-order GNN architectures
corresponding to the higher-order WL test.
\subsubsection{Interpretable GNNs}
The existing GNNs work in a black box. We do not understand why they achieve state-of-the-art performance in terms of
the node classification task, graph classification task and graph embedding task etc. Interpretability has become a major obstacle to
apply the GNNs to real-world issues. Although there have been some studies to interpret some specific GNN models,
they cannot interpret general GNN models. This motivates us to construct a unified interpretable framework for the GNNs.
\section{Conclusions}
This paper aims to provide a taxonomy, advances and trends for the GNNs. We expand the content of this paper from 4 dimensions: architectures, extensions and applications, benchmarks and evaluation pitfalls,
and future research directions.
The GNN architectures are expanded from 4 perspectives: graph convolutional neural networks, graph pooling operators, graph attention mechanisms and graph recurrent neural networks.
The extensions are expanded from 8 perspectives: GCNNS on special graphs, capability and interpretability, deep graph representation learning, deep graph generative models, combinations of the PI and GNNs
adversarial attachks for the GNNs, graph neural architecture search and graph reinforcement. In the future directions, we propose 4 prospective topics on the GNNs: highly scalable GNNs, robust GNNs,
GNNs going beyond WL test and interpretable GNNs. We expect that the relevant scholars can understand the computational principles of the GNNs,
consolidate the foundations of the GNNs and apply them to more and more real-world issues, through reading this review.
\begin{acks}
This work was supported by National Natural Science Foundation of China (Grant No. 62002255, 62076177, 61972273).
\end{acks}
\bibliographystyle{ACM-Reference-Format}
| eb5ac9b9632e6013cb318e682c6afed2fa7a855e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
With its high-resolution image and low cost, CT scan is critical in clinical decision and holds the key for making precise medical-care accessible to everyone around the world.
Recently, deep learning methods have been introduced to detect lesions in CT slices~\cite{3DCE,MSB, Retina, MVP, MULAN}.
Since it is difficult to distinguish lesions within a single axial slice, exploiting sufficient 3D context for accurate detection in volumetric CT data has emerged as a significant research focus.
Various architectures have been proposed for proper modeling of 3D context from neighboring CT slices.
Yan \textit{et al.}~\cite{3DCE} adopts a late fusion strategy which stacked 2D features of neighboring slices to build 3D context enhanced features.
Although the pseudo-3D contextual information has provided prominent performance gain~\cite{3DCE,MSB, Retina, MVP, MULAN}, its late fusion strategy leads to notable losses of context information from early stages of the network.
A direct way to address these issues is to employ 3D convolutions which introduce inter-slice connections hierarchically to learn 3D representations end to end. 3D convolutional filters can well preserve the 3D structure and texture information,
but intensive memory and computation demands hinder its wide application in the universal lesion detection problem.
What's worse, although 3D network pre-training has raised significant research attention~\cite{MG, Med3D,P3D,ACS}, the lack of good pre-trained 3D models makes it even harder to achieve good performance with 3D based detectors.
In this paper, we focus on the problem of universal lesion detection in CT slices, where multiple adjacent CT slices are taken into consideration to localize 2D lesions for the target slice.
We aim to develop a generic and efficient 3D backbone for 2D lesion detection with enhanced context modeling ability from multiple CT slices and devise a supervised pre-training method to boost its performance.
Specifically, pseudo-3D convolutional filters~\cite{P3D} which use depth-wise separable convolution are adopted to reduce the memory and computation overhead.
The backbone in our method is a Modified Pseudo-3D ResNet (MP3D ResNet), which extracts context enhanced 3D features from multiple neighboring CT slices (9 in our case) and then converted the 3D features into 2D ones with a group transform module (GTM) for further 2D lesion detection in the target slice.
Then, we feed backbone features extracted from MP3D ResNet into the neck of Feature Pyramid Network (FPN) to form the MP3D FPN for effective multi-scale detection.
Finally, to facilitate efficient training of the MP3D FPN, we designed a novel supervised pre-training method, which exploits supervised signals from large-scale 2D natural image object detection dataset to pre-train the proposed MP3D detector. In summary, the main contributions of our paper are three folds:
\paragraph{1.} We have proposed a generic framework to employ 3D network for 2D lesion detection in CT slices. The proposed MP3D FPN is computational and memory efficient, and it achieves state-of-the-art performance on the DeepLesion dataset.
\paragraph{2.} We have derived a novel and effective way to adopt 2D natural images to pre-train 3D network with supervised labels, whose pre-trained weights can potentially benefit other 3D medical image analysis tasks (e.g. segmentation).
\paragraph{3.} We have conducted comprehensive experiments to explore the effects of pre-trained weights for deep medical image analysis. The results suggest that pre-trained weights can not only lead to faster convergence in all sized datasets, but also help to achieve better results in smaller-scale ones.
\section{Methodology}
Fig.\ref{flowchart} gives an overview of the proposed lesion detection framework.
The proposed MP3D FPN comprises an MP3D ResNet as the backbone, a 2D FPN~\cite{FPN} as the neck and a 2D RPN/RCNN head.
The MP3D ResNet takes multiple consecutive CT slices (e.g. 9) as input and generates 3D feature maps which bear the ability of 3D context modeling.
Then a conversion block (GTM) further transforms the 3D feature maps into 2D ones for further 2D detection.
Detailed architecture designs of the proposed MP3D backbone and the novel supervised pre-training scheme will be elaborated in the following sections.
\begin{figure}[!tbp]
\centering
\includegraphics[width=\textwidth]{framework2.png}
\caption{Overview of the proposed MP3D FPN. MP3D ResNet extracts context enhanced 3D features and converts them to 2D ones with a group transform module (GTM). These context enhanced 2D features are then fed into the FPN neck and the RPN/RCNN head for further 2D lesion detection. The MP3D FPN is pre-trained on Microsoft COCO object detection dataset~\cite{COCO}. } \label{flowchart}
\end{figure}
\subsection{3D Context Modeling with an MP3D ResNet Backbone}
In this work, we explore to employ 3D convolutions for effective 3D context modeling in the problem of lesion detection from consecutive CT slices (e.g. 9 slices). To advance the time and memory efficiency of normal 3D ResNet, we adopt the Pseudo-3D Residual Network (P3D ResNet)~\cite{P3D} as the prototype of our backbone network.
The pseudo-3D convolution simulates $3\times 3\times 3$ convolution with $1\times 3\times 3$ filter on axial-view slices plus $3\times 1\times 1$ filter to build inter-slice connections on adjacent CT slices.
Lesion detection in CT slices aims to predict 2D bounding boxes in a certain slice, thus it requires 2D feature maps corresponding to the target slice for further prediction.
Therefore, we need to convert the 3D feature maps to 2D ones for further prediction, meanwhile preserving the precise information of the target CT slice for accurate localization and classification.
The designed Modified Pseudo-3D Residual Network (MP3D ResNet) highlights two aspects of modifications to fulfill such demands:
1) Instead of conducting isotropic pooling as in the original P3D ResNet, we neglect pooling operation in the inter-slice dimension.
2) A group transform module is introduced to generate the desired 2D feature maps from the context enhanced 3D features.
Neglecting pooling operation in the inter-slice dimension can help to preserve precise information of the target slice. In the meantime, since the number of input slices (e.g. 9) is rather small, we can get enough receptive field in the inter-slice dimension without downsampling.
Regarding 2D feature map conversions, Fang \textit{et al.}~\cite{PFN} proposed to extract $C$ 2D feature maps ($1\times 1\times H\times W$) corresponding to the center slice and concatenate them to form the converted 2D feature map of size ($C\times H\times W$).
However, this method can not fully exploit the 3D context information resided in other adjacent slices.
We, on the other hand, propose a group transform module (GTM) instead to includes all slice's features to compensate for the information loss.
Specifically, we view 3D features ($C\times D\times H\times W$) into 2D ($CD\times H\times W$) and apply a group convolutional layer with the group size of $C$ (every $D$ channel is a group) to fuse all neighboring features to yield the final 2D feature maps ($C\times H\times W$).
\begin{figure}[!tbp]
\includegraphics[width=\textwidth]{pretrain1.png}
\caption{Comparison between 1) using 2D Image-Net pre-trained weights for multi-slice medical image analysis and 2) decomposing 2D natural image to simulate multi-slice medical image for 3D network pre-training. } \label{pretrain}
\end{figure}
\subsection{Supervised 3D Pre-training with COCO Dataset}
\begin{table}[!htbp]
\caption{Sensitivities (\%) at various FPs per image on the test set of DeepLesion. $\mathbf{^*}$ indicates re-implementation of 3DCE using ResNet-50 FPN with the same configuration as our MP3D FPN.}
\label{tab1}
\centering
\footnotesize
\setlength{\tabcolsep}{8pt
\renewcommand{\arraystretch}{1.2
\centering
\label{tab:sample_1}
\begin{tabular}{l|ccccccccc}
\hline
\textbf{Methods} & $\mathbf{0.5}$ & $\mathbf{1}$ & $\mathbf{2}$ & $\mathbf{4}$ & $\mathbf{[email protected]}$ \\
\hline
\hline
3DCE, 27 slices\cite{3DCE} & 52.86 & 64.80 & 74.84 & 84.38 &-\\
MSB, 3 slices\cite{MSB} & 67.00 & 76.80 & 83.70 & 89.00 &-\\
RetinaNet, 3 slices\cite{Retina} & 72.15 & 80.07 & 86.40 & 90.77 &-\\
MVP-Net, 9 slices\cite{MVP} & 73.83 & 81.82 & 87.60 & 91.30 &-\\
MULAN, 9 slices\cite{MULAN} & \textbf{76.12} & \textbf{83.69} & \textbf{88.76} & \textbf{92.30} &-\\
\hline
FPN+3DCE, 3 slices$\mathbf{^*}$ &68.52 & 77.59 & 83.91 & 88.33 & 64.41\\
FPN+3DCE, 9 slices$\mathbf{^*}$ &74.06 & 82.00 & 87.58 & 91.56 & 70.28\\
FPN+3DCE, 27 slices$\mathbf{^*}$ &\textbf{74.67} & \textbf{82.89} & \textbf{88.17} & \textbf{91.62} & \textbf{70.82}\\
\hline
\textbf{MR3D FPN, 9 slices} & 79.09 & 84.84 & 89.18 & 92.06 & 76.57\\
\textbf{MP3D FPN, 9 slices} & \textbf{79.60} & \textbf{85.29} & \textbf{89.61} & \textbf{92.45} & \textbf{76.87} \\
\hline
\textbf{Imp over} MULAN, 9 slices & $\uparrow$\textbf{3.48} & $\uparrow$1.60 & $\uparrow$0.85 & $\uparrow$0.15&- \\
\end{tabular}
\end{table}
Supervised pre-training from natural images has proven to be an effective way for 2D medical image transfer learning\cite{3DCE,MSB,Retina,MVP,MULAN, Chest}. This indicates that using supervised pre-training models from another domain can actually benefit the medical image analysis application.
What's more, compared to self-supervised signals, we believe that supervised labels which carry the semantic information could enable the model to learn semantically invariant and discriminative features more effectively.
Therefore, in this section, we aim to develop a method to exploit supervised labels from large-scale 2D natural image object detection dataset (e.g. coco~\cite{COCO}) to pre-train our MP3D FPN.
Previous works~\cite{3DCE} have shown that by grouping 3 consecutive CT slices (which is natively 3D data) as a 3-channel RGB image, we can boost the detection performance with Image-Net pre-trained weights, indicating the feasibility of simulating RGB natural image with natively 3D CT slices.
This inspires us to reversely decompose the 3 channels of natural RGB images into 3 consecutive CT slices, and train an MP3D FPN with such simulated 3D data.
Fig~\ref{pretrain} illustrates a comparison of the two correlative strategies.
For implementation details, we train the MP3D FPN on COCO dataset for 72 epochs and the final weights are used to initialize MP3D ResNet.
To drive the network to learn useful 3D contextual features from inter-slice connections, it is essential to keep the resolution in the inter-slice dimension unchanged for all stages of the backbone
The MP3D detector trained with a slice number of 3 can be used to initialize lesion detectors which takes variable number of slices as network input.
\section{Experiments}
\subsection{Experimental Setup}
\noindent\textbf{Dataset and Metric:} The NIH DeepLesion is a large-scale dataset for lesion detection, which contains 32,735 lesions on 32,120 axial CT slices captured from 4,427 patients. DeepLesion is splitted into training (70\%), validation (15\%), and test (15\%) sets. We evaluate our MP3D FPN and all the compared methods on the test set by reporting the mean average precision ([email protected]) and average sensitivities at different false positives (FPs) per image.
\noindent\textbf{Implementation Details:} As in~\cite{Retina}, the Hounsfield units (HU) are clipped into the range of $[-1024,1050]$.
We interpolate in the z-axis to normalize the intervals of all CT slices to 2.5mm.
Anchor scales are set to $\{16, 32, 64, 128, 256\}$ in FPN.
Apart from horizontal and vertical flip, we resize the image to different scales of $\{448, 512, 576\}$ for data augmentation.
MP3D-63 with group normalization\cite{GN} is used as the backbone in all our experiments, which has similar depth with the ResNet3D-50 model. The MP3D-63 model is derived from the conventional P3D-63\cite{P3D} model with the proposed modifications.
Unless otherwise specified, the MP3D FPN takes 9 consecutive slices as input.
We train all the models for 24 epochs at the base learning rate of 0.02, and reduce it by a factor of 10 after the 16-th and 22-th epoch (corresponding to the 2x learning schedule\cite{Rethinking} on COCO dataset).
We conduct experiments on the NVIDIA TITAN V GPU with 12GB of memory, and mixed-precision training strategy is used in all our experiments to save memory.
\subsection{Comparison with State-of-the-arts}
Table \ref{tab1} presents the comparisons with the previous state-of-the-art (SOTA) methods.
Our model surpasses all the SOTA methods on sensitivities at different FPs and [email protected], which includes 3DCE\cite{3DCE}, MSB\cite{MSB}, RetinaNet\cite{Retina}, MVP-Net\cite{MVP} and MULAN\cite{MULAN}.
Without using any auxiliary supervision, MP3D FPN outperforms MULAN, the previous SOTA which additionally employs multi task learning and a deeper backbone (DenseNet-121) to improve the detection accuracy, by up to \textbf{3.48\%} on the sensitivity of [email protected].
We re-implement 3DCE with ResNet-50 FPN using the same configuration as our MP3D FPN for fair comparison. Our proposed MP3D achieve a performance gain of \textbf{6.05\%} on [email protected] compared with this 2D convolution based context encoding method, demonstrating the superior 3D context modeling ability of our MP3D backbones.
As shown in Table \ref{tab1}, MP3D FPN (248.93 GFLOPS, 45.16 $M$ Params) and MR3D FPN (Modified ResNet 3D, 415.81 GFLOPS, 64.03 $M$ Params) based detector achieve comparable results, but the MP3D based detector consumes much less time and memory.
This strongly proves the efficacy and the thrift of our MP3D model.
\subsection{Ablation Study}
We perform a number of ablations to probe into our MP3D FPN. The results are shown as follows:
\begin{table}[!t]
\caption{Detection performance and computational cost with variable numbers of input slice. GFLOPS is used to characterize the computational cost.}
\label{tab:slice_num}
\centering
\footnotesize
\setlength{\tabcolsep}{8pt
\renewcommand{\arraystretch}{1.2
\centering
\begin{tabular}{lccccccccc}
\textbf{Methods} & $\mathbf{0.5}$ & $\mathbf{1}$ & $\mathbf{2}$ & $\mathbf{4}$ & $\mathbf{[email protected]}$ & $\mathbf{GFLOPS}$\\
\hline
\hline
MP3D, 5 slices &76.86 &83.44&88.13 &91.54&75.01 & 156.84\\
MP3D, 7 slices &78.22 &84.45&88.90 &91.50&76.69 & 202.88\\
MP3D, 9 slices &79.60 &85.29 & \textbf{89.61} &92.45 &76.87 & 248.93\\
MP3D, 11 slices&\textbf{80.05}&\textbf{85.77}&89.55 &\textbf{92.45}&\textbf{77.64} & 294.97\\
\end{tabular}
\end{table}
\noindent\textbf{Input Slices:} Table \ref{tab:slice_num} shows the performance of the MP3D detector when applying 5,7,9 and 11 slices as input.
The detector achieves higher detection accuracy as more slices are used, meanwhile consuming more time and memory. MP3D with 7 slices as input get the best trade-off between effectiveness and efficiency.
\begin{table}[!t]
\caption{Comparison of different conversion modules and different pooling strategies for pre-training.}
\label{tab:gtm}
\centering
\footnotesize
\setlength{\tabcolsep}{8pt
\renewcommand{\arraystretch}{1.2
\centering
\begin{tabular}{lccccccccc}
\textbf{Methods} & $\mathbf{0.5}$ & $\mathbf{1}$ & $\mathbf{2}$ & $\mathbf{4}$ & $\mathbf{[email protected]}$\\
\hline
\hline
MP3D w/ CTM &79.18 &84.90 &88.96 &91.90 &76.30 \\
MP3D w/ GTM &\textbf{79.60} &\textbf{85.29} & \textbf{89.61} & \textbf{92.45} &\textbf{76.87} \\
\hline
MP3D w/ isotropic pooling &78.24 &84.41 &88.82 &91.98 &75.06\\
MP3D w/ proposed pooling &\textbf{79.60} &\textbf{85.29} & \textbf{89.61} & \textbf{92.45} &\textbf{76.87} \\
\end{tabular}
\end{table}
\noindent\textbf{Conversion Type:} Table \ref{tab:gtm} demonstrates the comparisons of proposed GTM with the center-cropping transform module (CTM), which is proposed by Fang \textit{et al.}~\cite{PFN}. The proposed GTM brings better results as it can efficiently aggregate information from all adjacent slices for further detection.
\subsection{Effectiveness of the 3D Pre-trained Model}
We conducted three groups of experiments to explore effectiveness of the pre-training method.
\noindent\textbf{Comparison to Isotropic Pooling:} In this work, to achieve 3D context modeling ability in the z-axis, we neglect pooling operation in the inter-slice dimension when pre-training the MP3D model on the Microsoft COCO dataset. We compared our proposed method to isotropic pooling for validation.
The pre-trained model takes three slices as input. When training with isotropic pooling, the z-axis degenerates to a single slice after the first two pooling layers, preventing further 3D convolution layers from learning useful 3D contextual information. As shown in Table\ref{tab:gtm}, pre-trained weights learned from isotropic pooling gives worse results than the proposed method. This also proves that using decomposed natural image as input can actually helps the 3D model to gain context-encoding ability. Thus the learned weights can potentially be used to boost the performance of other 3D medical image analysis tasks.
\begin{table}[!t]
\caption{Comparison of model performance with and without pre-training with different learning schedules. 1x, 2x and 6x indicates max training epochs of 12, 24 and 72 separately.}
\label{tab:scratch}
\centering
\footnotesize
\setlength{\tabcolsep}{8pt
\renewcommand{\arraystretch}{1.2
\centering
\label{tab:scratch}
\begin{tabular}{lccccccccc}
\textbf{Methods} & $\mathbf{0.5}$ & $\mathbf{1}$ & $\mathbf{2}$ & $\mathbf{4}$ & $\mathbf{[email protected]}$ \\
\hline
\hline
MP3D 1x w/o pretrain & 70.12& 78.00& 83.95& 88.23&67.60 \\
MP3D 2x w/o pretrain & 76.11& 82.65& 87.70& 91.17&74.00 \\
MP3D 6x w/o pretrain &\textbf{79.60} & 85.29& 89.26& 92.19& 76.75\\
\hline
MP3D 1x w/ pretrain &78.02& 84.33& 88.84& 91.74&75.78\\
MP3D 2x w/ pretrain&79.58& \textbf{85.29}& \textbf{89.61}& \textbf{92.45}&\textbf{76.87}\\
\end{tabular}
\end{table}
\begin{table}[!t]
\caption{Training with variable dataset sizes ($100\%$ to $20\%$). For simplicity, we present the results of [email protected].}
\label{tab4}
\centering
\footnotesize
\setlength{\tabcolsep}{8pt
\renewcommand{\arraystretch}{1.2
\centering
\label{tab:datasize}
\begin{tabular}{lcccccccc}
\textbf{Methods} & $\mathbf{100\%}$ & $\mathbf{80\%}$ & $\mathbf{60\%}$ & $\mathbf{40\%}$ & $\mathbf{20\%}$ \\
\hline
\hline
3DCE 9 slices 2x &70.28 &69.22 &67.08 &63.61 &57.02 \\
3DCE 27 slices 2x &70.82 &69.96 &68.08 &65.36 &58.82 \\
\hline
MP3D w/o pre-train 2x &74.00 &71.58 &68.79 &63.40 &50.67 \\
MP3D w/o pre-train 6x &76.75 &75.43 &72.87 &68.14 &58.98 \\
MP3D w/ pre-train 2x & \textbf{76.87} &\textbf{75.66} &\textbf{73.33} &\textbf{71.07} &\textbf{65.55} \\
\end{tabular}
\end{table}
\noindent\textbf{Comparison to Training From Scratch:}
He \textit{et al.}\cite{Rethinking} demonstrated that with sufficient training data (around 35k from its experiment) and longer training schedule (6x), models trained from scratch could achieve comparable results to models training with pre-trained weights. Therefore, we examined the effectiveness of our proposed pre-training method by comparing MP3D with pre-training to model trained from scratch with longer schedule.
As shown in Table\ref{tab:scratch}, when both trained for 1x learning schedule (12 epochs), MP3D with pre-trained weights significantly outperforms the one without pre-training, demonstrating faster convergence speed. And it turns out that with 2x learning schedule (24 epochs), model trained with the proposed pre-training weights can achieve comparable results with MP3D model trained from scratch with 6x learning schedule (72 epochs). These results validate the effectiveness of our proposed pre-training scheme.
\noindent\textbf{Performance on Variable Dataset Sizes:} In medical image analysis tasks, annotated data is often scarce. Therefore, it is appealing to gain a better understanding of the effects of pre-trained weights when dataset size is small. In this subsection, we compare the model performance of 2x, 6x training from scratch and 2x with pre-training on variable dataset sizes by randomly choosing 20\%, 40\%, 60\% and 80\% of the whole training data. Pre-training based models achieve better performance with less training time on all the cases, and the smaller the size of the dataset, the larger the gap. A dramatic drop of performance starts when training with only 40\% of the whole data. And when training with only 20\% of the dataset, which is around 4,500 images, the model trained with our proposed pre-trained weights achieves an absolute performance gain of 6.57\% on [email protected], accounting for an 11\% relative gain.
\section{Conclusions}
In this paper, we propose a generic model architecture to exploit 3D network for 2D lesion detection in CT slices.
The proposed MP3D FPN can reduce computation and memory cost while providing enhanced 3D context modeling ability.
A simple yet effective way for 3D network pre-training is also derived to facilitate efficient training.
Without sophisticated structures and multi-supervision signals, it significantly improves the detection performance on the DeepLesion dataset, surpassing all the SOTAs.
We have proved the benefits of pre-trained weights for variable dataset size, and we expect that the MP3D ResNet along with its pre-trained weights can serve as a benchmark backbone for 3D medical image analysis, making contributions towards accessible precise medication.
\subsubsection{Acknowledgements.} This work is funded by National Key Research and Development Program of China (No. 2019YFC0118101), MOST-2018AAA0102004 and NSFC-61625201. We would like to thank Yemin Shi for valuable discussions.
\input{paper62.bbl}
\end{document} | 3e12ffca40a22788e64cc795251a92a5f42bcd2c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\label{sec:intro}Introduction}
Let $X$ be a finite set, let $\mathcal{F}\subseteq2^{X}$ be a family
of subsets of $X$, and let $b$ be a positive integer. In the $\left(1:b\right)$
\textsc{Maker}\textendash \textsc{Breaker} game $\left(X,\mathcal{F}\right)$,
two players, called \textsc{Maker} and \textsc{Breaker}, take turns
in claiming previously unclaimed elements of $X$. On \textsc{Maker}'s
move, she claims one element of $X$, and on \textsc{Breaker}'s move,
he claims~$b$ elements (if less than $b$ elements remain before
\textsc{Breaker's} last move, he claims all of them). The game ends
when all of the elements have been claimed by either of the players.
\textsc{Maker} wins the game $\left(X,\mathcal{F}\right)$ if by the end of
the game she has claimed all the elements of some $F\in\mathcal{F}$;
otherwise \textsc{Breaker} wins. The description of the game is completed by stating which of the players
is the first to move, though usually it makes no real difference. For convenience, we typically assume
that $\mathcal{F}$ is closed upwards, and specify only the inclusion-minimal
elements of $\mathcal{F}$. Since these are finite, perfect information
games with no possibility of draw, for each setup of $\mathcal{F},b$
and the choice of the identity of the first player, one of the players has a strategy
to win regardless of the other player's strategy. Therefore, for a
given game we may say that the game is \textsc{Maker}\textquoteright s
win, or alternatively that it is \textsc{Breaker}\textquoteright s
win. The set $X$ is referred to as the \emph{board} of the game,
and the elements of $\mathcal{F}$ are referred to as the \emph{winning
sets}.
When $b=1$, we say that the game is \emph{unbiased}; otherwise it
is \emph{biased}, and $b$ is called the \emph{bias} of \textsc{Breaker}.
It is easy to see that \textsc{Maker}\textendash \textsc{Breaker}
games are \emph{bias monotone}. That is, if \textsc{Maker} wins some
game with bias $(1:b)$, she also wins this game with bias $(1:b')$
for every $b'\leq b$. Similarly, if \textsc{Breaker} wins a game
with bias $(1:b)$, he also wins this game with bias $(1:b')$ for
every $b'\geq b$. This bias monotonicity allows us to define the
\emph{threshold bias}: for a given game $\mathcal{F}$, the threshold
bias $b^{*}$ is the value for which \textsc{Breaker} wins the game
$\mathcal{F}$ with bias $(1:b)$ if and only if $b>b^{*}$. It is
quite easy to observe that it is never a disadvantage in a \textsc{Maker}\textendash \textsc{Breaker}
game to be the first player, and that if a player has a winning strategy
as the second player, essentially the same strategy can be used to
also win the game as the first player. Hence, when we describe a strategy
for \textsc{Maker} we assume that she is the second player, implying
that under the conditions described she can win as either a first
or a second player. The same goes for \textsc{Breaker}'s strategy.
\medskip{}
In this paper, our attention is dedicated to the $\left(1:b\right)$
\textsc{Maker}\textendash \textsc{Breaker} $s$-component game on
the binomial random graph $\mathcal{G}\left(n,p\right)$,
in which each of the ${n \choose 2}$ possible edges appears independently
with probability $p = p(n)$; that is, the board is the edge set of $\ggnp$
and the (inclusion-minimal) winning sets are the trees of $G$ with $s$
vertices. Since the board is random, our results hold with high probability
(WHP), i.e., with probability tending to $1$ as $n$ tends to infinity.
\medskip{}
For more on \textsc{Maker}\textendash \textsc{Breaker} games as well as other positional games, please see the books by Beck~\cite{B-08} and by Hefetz et al.~\cite{HKSS-14}.
\subsection{Previous results}
A natural case to consider is $s=n$; that is, the winning sets are
the spanning trees of the graph the game is played on. This $\left(1:b\right)$
$n$-component game is known as the \emph{connectivity} game.
The unbiased game was completely solved by Lehman~\cite{Lehman-64},
who showed that \textsc{Maker}, as a second player, wins the $\left(1:1\right)$
connectivity game on a graph $G$ if and only if $G$ contains two
edge-disjoint spanning trees. It follows easily from~\cite{Nash-61,Tutte-61}
that if $G$ is $2k$-edge-connected then it contains $k$ pairwise
independent spanning trees; thus, \textsc{Maker} wins the $\left(1:1\right)$
connectivity game on $4$-regular $4$-edge-connected graphs, whereas
\textsc{Breaker} trivially wins the $\left(1:1\right)$ connectivity
game on graphs with less than $2n-2$ edges, i.e., average degree under 4.
For denser graphs, since \textsc{Maker} wins the unbiased game by
such a large margin, it only seems fair to even out the odds by strengthening
\textsc{Breaker}, giving him a bias $b\ge2$. The first and most natural
board to consider is the edge set of the complete graph $K_{n}$.
Chv\'{a}tal and Erd\H{o}s~\cite{CH-78} showed that $\left(\frac{1}{4}-o\left(1\right)\right)n/\log n\le b^{*}\left(K_{n}\right)\le\left(1+o\left(1\right)\right)n/\log n$;
the upper bound was proved to be tight by Gebauer and Szab\'{o}~\cite{GS-09};
that is, \textsc{$b^{*}\left(K_{n}\right)=\left(1+o\left(1\right)\right)n/\log n$.}
Returning to $\mathcal{G}\left(n,p\right)$, Stojakovi\'{c} and Szab\'{o}~\cite{SS-05}
showed that WHP $b^{*}\left(\mathcal{G}\left(n,p\right)\right)=\Theta\left(np/\log n\right)$,
where \textsc{Breaker}'s win holds for any $0\le p\le1$, while \textsc{Maker}'s
win requires large enough $p$ (\textsc{Maker} cannot win for small
$p$ since $\ggnp$ is WHP disconnected). This was improved by Ferber
et al.~in~\cite{FGKN-15}, who showed for $p=\omega\left(\log n/n\right)$
that WHP $b^{*}\left(\mathcal{G}\left(n,p\right)\right)=\left(1+o\left(1\right)\right)np/\log n$.
A different random graph model, the random $d$-regular graph $\mathcal{G}\left(n,d\right)$
on $n$ vertices, was considered by Hefetz et al.~in~\cite{HKSS-11}.
They showed that WHP $b^{*}\left(\mathcal{G}\left(n,d\right)\right)\ge\left(1-\epsilon\right)d/\log_{2}n$
for $d=o\left(\sqrt{n}\right)$. Note that when $d=\Omega\left(\sqrt{n}\right)$,
the model $\mathcal{G}\left(n,d\right)$ is quite close to $\mathcal{G}\left(n,p\right)$
for $p=d/n$, since for this value of $p$ all degrees in $\ggnp$ are WHP $(1 + o(1))d$. Moreover,
they showed that $b^{*}\left(G\right)\le\max\left\{ 2,\bar{d}/\log n\right\} $
for any graph $G$ of average degree $\bar{d}$, so the result is
asymptotically tight.
\medskip{}
\textsc{Breaker}'s strategy in practically all results mentioned above
is to deny connectivity by isolating a single vertex. Much less is
known, however, for the case $s<n$. It seems that even if \textsc{Breaker}
is able to isolate a vertex in a constant number of moves, it does
little to prevent \textsc{Maker} from winning the $s$-component game
for $s=\Omega\left(n\right)$.
Instead of considering the threshold bias $b^{*}$, we shift the focus
to the maximal component size $s$ achievable by \textsc{Maker} in
the $\left(1:b\right)$ component game, for a given bias $b$ (assuming optimal play of both players). Let
us denote this quantity by $s_{b}^{*}\left(G\right)$. Bednarska and
\L uczak considered in~\cite{BL-01} the $\left(1:b\right)$ component
game on the complete graph. They showed that $s_{b}^{*}\left(K_{n}\right)$
undergoes a certain type of phase transition around $b=n$; specifically, that $s_{n+t}^{*}\left(K_{n}\right)=\left(1-o\left(1\right)\right)n/t$
for $\sqrt{n}\ll t\ll n$ but $s_{n-t}^{*}\left(K_{n}\right)=t+O\left(\sqrt{n}\right)$
for $0\le t\le n/100$.
The component game on $d$-regular graphs for fixed $d\ge3$ was considered
by Hod and Naor in~\cite{HN-14}. They showed
a similar phase transition of $s_{b}^{*}$ around $b=d-2$: for any
$d$-regular graph $G$ on $n$ vertices,
\[
s_{b}^{*}\left(G\right)=\begin{cases}
O\left(1\right), & b\ge d-1;\\
O\left(\log n\right) & b=d-2
\end{cases}
\]
whereas $s_{d-3}^{*}\left(\mathcal{G}\left(n,d\right)\right)=\Omega\left(n\right)$
WHP.
\subsection{\label{sec:our-results}Our results}
Given previous results, it is not surprising that the component game
on the binomial random graph $\mathcal{G}\left(n,p\right)$
undergoes a phase transition too. Writing $p=c/n$ \textemdash{} so
the expected average degree in $\mathcal{G}\left(n,p\right)$ is $c$
\textemdash{} we could perhaps guess the phase transition occurs around
$c=b+2$ in accord with the results for $d$-regular graphs. Another
plausible approach would be to consider the so-called \emph{random
graph intuition}, as first observed by Chv\'{a}tal and Erd\H{o}s in~\cite{CH-78}:
it turns out that in many cases, the winner of a game in which both
sides play to their best is the same as if both sides were to play
randomly. We may thus guess that the transition occurs around $c=b+1$,
since in the random players scenario \textsc{Maker} would end up with
the edge set of $\mathcal{G}\left(n,p/\left(b+1\right)\right)$, which
contains a linear-size component if and only if $c>b+1$.
It turns out, however, that neither of these heuristics gives the correct answer. The key graph parameter is degeneracy, which is related to the minimum degree of subgraphs,
rather than average degree; consequently, the critical $c$ is (somewhat) larger
than $b+2$.
\begin{defn*}
For an integer $k\ge1$, the $k$\emph{-core} of a graph $G=\left(V,E\right)$
is its largest subgraph $K$ of minimum degree $\delta\left(K\right)\ge k$.
If no such subgraph exists, we say $G$ has an empty $k$-core, or
that $G$ has no $k$-core, or that $G$ is $\left(k-1\right)$-\emph{degenerate}.
\end{defn*}
Pittel, Spencer and Wormald~\cite{PSW-96} (see also~\cite{CW-06,JL-07,Rio-08})
proved that for every $k\ge3$ there exists a threshold constant $c_{k}$
for the appearance of the $k$-core in the binomial
random graph. That is, $\ggnc$ has WHP an empty $k$-core for $c<c_{k}$
and a linear $k$-core for $c>c_{k}$. It was previously shown by
\L uczak~\cite{Luczak-91,Luczak-92} that $\ggnp$ WHP has either
an empty or a linear-size $k$-core, for every fixed $k\ge3$. The constant $c_k$ is implicitly defined and satisfies $c_{k}=k+\sqrt{k\log k}+O \big(\sqrt{k/\log k} \big)$ (see~\cite[Lemma 1]{PVW-11}). For small values of $k$ we have
$c_{3}\approx3.351,$ $c_{4}\approx5.149$, $c_{5}\approx6.799$, $c_{6}\approx8.365$.
\medskip{}
The following theorems show that for any constant $b$, the phase transition for the $(1:b)$ component
game on $\ggnc$ occurs at $c=c_{b+2}$.
\begin{thm}
\label{thm:breaker-win}For any two constants $b$ and $c<c_{b+2}$ we have WHP $s_{b}^{*}\left(\gnc\right)=o\left(\log^{3}n\right)$.
\end{thm}
\begin{thm}
\label{thm:maker-win}For any two constants $b$ and $c>c_{b+2}$ we have WHP $s_{b}^{*}\left(\gnc\right)=\Omega\left(n\right)$.
\end{thm}
Fix an integer $k\ge3$. The standard algorithm for finding the
$k$-core of a graph $G$ on $n$ vertices is the following, called
the $k$-peeling process: starting from $G_{0}=G$, let $\left(G_{t}\right)_{t\ge0}$
be the sequence of subgraphs of $G$, where $G_{t+1}$ is obtained
from $G_{t}$ by deleting all edges incident with vertices of degree
at most $k-1$. Since $G_{t+1}\subseteq G_{t}$, this (deterministic)
process stabilises after some finite time $T^*\left(G\right)\le n$.
If $G$ is $\left(k-1\right)$-degenerate, the graph $G_{ T^*}$
is empty; otherwise, $G_{ T^*}$ is the $k$-core of $G$, plus,
possibly, some isolated vertices (we only delete edges, so $G_{t}$
has $n$ vertices for all $t\ge0$, although some of them may become
isolated along the process).
\begin{rem*}
Two variations of this process are: $\left(i\right)$ delete vertices
instead of edges; $\left(ii\right)$ process in every step only a
single vertex of degree less than $k$.
\end{rem*}
Jiang, Mitzenmacher, and Thaler used a branching process argument
in~\cite{JMT-16} (see also~\cite{Gao-14}) to bound the typical value of the stabilization
time $T^*\left(n,c\right)$ of the peeling process on $\gnc$
in the subcritical regime $c<c_{k}$:
\begin{thm}[{\cite[Theorems~1 and~2]{JMT-16}}]
\label{thm:stabilization-time}Fix $k\ge3$ and $c<c_{k}$. Then WHP
\[T^*\left(n,c\right)=\log_{k-1}\log n+\Theta\left(1\right).\]
\end{thm}
\begin{rem*}
For $c>c_{k}$, the typical stabilization time is $T^*\left(n,c\right)=\Theta\left(\log n\right)$;
see~\cite[Theorem~4]{AM-15} for the upper bound and~\cite[Theorem~3]{JMT-16}
for the lower bound.
\end{rem*}
In Section~\ref{sec:technical-background} we analyse the
peeling process further, and prove a related result, which may be of independent
interest. While it takes $\log_{k-1}\log n+\Theta\left(1\right)$ time by
Theorem~\ref{thm:stabilization-time} to peel the \emph{entire} graph,
the graph is WHP already shattered into tiny fragments after a \emph{constant
}number of iterations:
\begin{thm}
\label{thm:shatter-time}Fix $k\ge3$ and $c<c_{k}$. There exists
a constant $t^{\dagger}=t^{\dagger}\left(c\right)$ such that, in
the $k$-peeling process on $\ggnc$, WHP all connected components
of $G_{t^{\dagger}}$ have size $o\left(\log^{3}n\right)$.
\end{thm}
\begin{rem*}
Our proof actually yields the somewhat better bound, $O\left(\log^3n/\log^2\log n\right)$, in Theorem~\ref{thm:shatter-time}, and consequently in Theorem~\ref{thm:breaker-win}. We use the term $o\left(\log^3n\right)$ in both theorems for brevity.
\end{rem*}
In Section~\ref{sec:prelim}
we provide some notations and a few bounds we use throughout the paper.
In Section~\ref{sec:breaker}, we prove Theorem~\ref{thm:breaker-win},
by showing how \textsc{Breaker} can WHP limit the radius of \textsc{Maker}'s
components (via Theorem~\ref{thm:shatter-time}) and hence limit
their size to poly-logarithmic. In Section~\ref{sec:maker} we prove
Theorem~\ref{thm:maker-win}, by showing how \textsc{Maker} can WHP
build a tree of linear size within the $\left(b+2\right)$-core of
$\gnc$. This is done in a two-step strategy:
\textsc{Maker} carefully nurtures a small sapling, which she afterwards
grows arbitrarily into a full-scale tree. The proofs of Theorem~\ref{thm:shatter-time}
and of two ingredients of the proof of Theorem~\ref{thm:maker-win}
are rather technical; we provide required background in Section~\ref{sec:technical-background}
and the proofs themselves in Section~\ref{sec:technical-proofs}.
\section{\label{sec:prelim}Preliminaries}
\subsection{Notation}
We use standard graph theory terminology, and in particular the following.
For a given graph $G$ we denote by $V(G)$ and $E(G)$ the set of
its vertices and the set of its edges, respectively. The \emph{excess}
of $G$ is defined as $\xs\left(G\right)=\left|E\left(G\right)\right|-\left|V\left(G\right)\right|$.
For two disjoint sets of vertices $A,B\subseteq V$ we denote by $E(A,B)$
the set of all edges $ab\in E$ with $a\in A$ and $b\in B$. For
a subgraph $F\subset G$ we denote its \emph{edge boundary} (with
respect to $G$) by $\partial F=E\left(V\left(F\right),V\left(G\right)\setminus V\left(F\right)\right)$.
For a positive integer $t$, the \emph{$t$-neighbourhood} of a vertex
$v\in V\left(G\right)$, also referred to as the \emph{ball} of radius
$t$ around $v$, is the subset of vertices of $G$ whose distance
to $v$ is at most $t$.
\medskip{}
Given a rooted tree $T$ and an integer $k\ge3$, a non-leaf vertex
$v\in V\left(T\right)$ is called \emph{$k$-light} if $v$ has less
than $k$ children in $T$ and \emph{$k$-heavy} otherwise. A \emph{tree
path} of $T$ is a sequence $\left(v_{0},v_{1},\ldots,v_{j}\right)$
of vertices of $T$ such that $v_{i-1}$ is the parent of $v_{i}$
for all $i=1,\ldots,j$. A tree path is called \emph{$k$-light} if
all its vertices are $k$-light.
The \emph{level} of any vertex $v\in V(T)$ is the length of the unique path in $T$ between $v$ and the root, and the \emph{height} of $T$ (assuming $T$ is finite) is the maximum level of a vertex $v \in V(T)$.
\medskip{}
For two integers $n>0$ and $m\ge0$, let $\D_{n,2m}$ denote the set of all nonnegative integer vectors $\dd = (d_1,\dots,d_n)$ such that $\sum d_i = 2m$. Each $\dd \in \D_{n,2m}$ is called a \emph{degree sequence}, and denote its maximum degree by~$\Delta(\dd) := \max_i d_i$.
\medskip{}
At any point during the game, an unclaimed edge is called \emph{free}. The
act of claiming one free edge by one of the players is called a \emph{step}.
\textsc{Breaker}'s $b$ successive steps (and \textsc{Maker}'s single
step) are called a \emph{move}. A \emph{round} in the game consists
of one move of the first player, followed by one move of the second
player. Since being the first player is never a disadvantage, we will
assume \textsc{Maker} starts when proving Theorem~\ref{thm:breaker-win}, and assume \textsc{Breaker} starts when proving Theorem~\ref{thm:maker-win}.
\medskip{}
Throughout the paper we use the well-known bound $\binom{n}{k}\le n^{k}/k!\le\left(en/k\right)^{k}$
for nonnegative integers~$n$ and~$k$. Let $\left[n\right]_{k}=n\left(n-1\right)\cdots\left(n-k+1\right)$,
and note that $\left[n\right]_{k}=n!/\left(n-k\right)!$ for $n\ge k$.
For a positive integer $m$, let $\left(2m-1\right)!!=\left(2m-1\right)\left(2m-3\right)\cdots3\cdot1$
and $\left(2m\right)!!=\left(2m\right)\left(2m-2\right)\cdots4\cdot2=2^{m}\cdot m!$.
\medskip{}
In this paper we make extensive use of the Poisson distribution. For
an integer $j\ge0$ and a real number $\lambda\ge0$, let $\Psi_{j}\left(\lambda\right)=\prob\left[\Poisson\left(\lambda\right)=j\right]=e^{-\lambda}\lambda^{j}/j!$, let $\Psi_{\ge j}\left(\lambda\right)=\prob\left[\Poisson\left(\lambda\right)\ge j\right]=\sum_{i\ge j}\Psi_{i}\left(\lambda\right)$, and let $\Psi_{<j}\left(\lambda\right)=\prob\left[\Poisson\left(\lambda\right) < j\right]= 1-\Psi_{\ge j}$.
Note that for any real number $\lambda \ge 0$ and for any two integers $j\ge\ell\ge 0$ we have
\begin{equation}\label{eq:poisson-obs}
\left[j\right]_{\ell}\Psi_{j}\left(\lambda\right)
= e^{-\lambda}\lambda^{j}/(j-\ell)!
= \lambda^{\ell}\Psi_{j-\ell}\left(\lambda\right).
\end{equation}
Given a positive integer $\ell>0$ and a real number $\lambda>0$,
let $Z_{\ell}\left(\lambda\right)$ denote an $\ell$-truncated Poisson
random variable, which is a $\Poisson\left(\lambda\right)$ random
variable conditioned on being at least $\ell$. In other words, $\Pr\left[Z_{\ell}\left(\lambda\right)<\ell\right]=0$
and $\Pr\left[Z_{\ell}\left(\lambda\right)=j\right]=\Psi_{j}\left(\lambda\right)/\Psi_{\ge\ell}\left(\lambda\right)$
for $j\ge\ell$.
\medskip{}
Our proofs are asymptotic in nature and whenever necessary, we assume $n$ is large enough. We omit floor and ceiling signs when these are not crucial. All logarithms in this paper, unless specified otherwise, are natural.
\subsection{Local structure of $\protect\gnc$ }
We also provide several results about the typical local structure
of $\gnc$, which will be useful later. We begin with the following
bound on the volume of balls in $\gnc$.
\begin{lem}[{\cite[Lemma 1]{CL-01}}]
\label{lem:volume-of-balls-in-Gnp}Let $\ggnc$ for $c>1$. Then
WHP, for every vertex $v\in V\left(G\right)$ and for every $1\le t\le n$,
there are at most $2t^{3}c^{t}\log n$ vertices in $G$ within distance
$t$ of $v$.
\end{lem}
Next we show that even though $G$ is not acyclic WHP for $c>1$,
we do not expect short cycles; in other words, local neighbourhoods
in $G$ are trees.
\begin{claim}
\label{clm:no-cycles}The probability of having a cycle in the $t$-neighbourhood
of a given vertex $v$ in $\ggnc$ is $o\left(1\right)$ for $t = \lceil \alpha\log n\rceil$, where $\alpha = \alpha(c)$ is some positive constant.
\end{claim}
\begin{proof}
It suffices to prove this for $c>1$. By Lemma~\ref{lem:volume-of-balls-in-Gnp},
a breadth-first search from $v$ discovers WHP at most $s=\left\lfloor 2t^{3}c^{t}\log n\right\rfloor $
vertices in the $t$-neighbourhood of $v$; the probability of having
either a back-edge or a side edge closing a cycle, in addition to the tree edges, among the first $s$ vertices discovered, is bounded
by
\[
\binom{s}{2}c/n<s^{2}c/n\le4t^{6}c^{1+2t}\log^{2}n/n,
\]
which is $o\left(1\right)$ for our choice of $t$ and for sufficiently small $\alpha$ (which only depends on $c$).
\end{proof}
Last, we show that sufficiently large connected subgraphs of $G$
expand by only a constant factor WHP.
\begin{claim}
\label{clm:vertex-expansion}Let $\ggnc$ for $c>1$. Then WHP, every
connected subset $U\subseteq V\left(G\right)$ of size $\left|U\right|\ge\log n$
has at most $3c\left|U\right|$ neighbours in $G$.
\end{claim}
\begin{proof}
Fix a set $U\subseteq V\left(G\right)$ of size $u=\left|U\right|$.
It is connected (and thus contains at least one of the $u^{u-2}$ possible
spanning trees of $U$) with probability at most $u^{u-2}\left(c/n\right)^{u-1}$.
The size of the external neighbourhood of $U$ is distributed $\Bin\left(n-u,1-\left(1-c/n\right)^{u}\right)$, stochastically dominated by
$\Bin\left(n,cu/n\right)$,
and thus by the Chernoff bound, it has more than $\left(1+\delta\right)cu$
neighbours with probability at most $\exp\left(\delta^{3}-\delta^{2}cu/2\right)$.
Therefore, the probability that, for some $u\ge\log n$, there exists
a connected set $U$ of $u$ vertices with more than $3cu$ neighbours
is bounded by
\begin{align*}
\sum_{u=\left\lceil \log n\right\rceil }^{n}\binom{n}{u}u^{u-2}\left(c/n\right)^{u-1}\exp\left(8-2cu\right) & \le\frac{e^{8}n}{c}\sum_{u=\left\lceil \log n\right\rceil }^{n}u^{-2}\left(ce^{1-2c}\right)^{u}\\
& <\frac{e^{8}n}{c\log^{2}n}\sum_{u=\left\lceil \log n\right\rceil }^{\infty}e^{-u}\le\frac{e^{9}}{c\left(e-1\right)\log^{2}n}=o\left(1\right).\qedhere
\end{align*}
\end{proof}
\section{\label{sec:breaker}His Side}
We now present a strategy for \textsc{Breaker} to use in the $\left(1:b\right)$
component game on $\gnc$ for $c<c_{b+2}$,
and show that it WHP restricts \textsc{Maker}'s connected components
to poly-logarithmic size.
To facilitate the description of \textsc{Breaker}'s strategy, we introduce
some terminology. Denote the \emph{rank} of a vertex~$v$ with respect
to the $\left(b+2\right)$-peeling process $\left(G_{t}\right)_{t\ge0}$
of a graph $G=G_{0}$ by $\rho\left(v\right)=\inf\left\{ t\ge0\mid\deg_{G_{t}}\left(v\right)<b+2\right\} $.
When $G$ is $\left(b+1\right)$-degenerate, $\rho\left(v\right)\le T^*\left(G\right)$
is finite for all $v$. In particular, this holds WHP for $\ggnc$
with $c<c_{b+2}$.
An edge $uv\in E\left(G\right)$ is called \emph{horizontal} if $\rho\left(u\right)=\rho\left(v\right)$
and \emph{vertical} otherwise. A \emph{horizontal component} (H-comp,
for short)~$C$ is a connected component in the graph consisting
of horizontal edges claimed by \textsc{Maker}, and its \emph{rank}
$\rho\left(C\right)$ is the rank shared by all its vertices. A vertical
edge $e$ is \emph{above} $C$ if $V\left(C\right)$ contains $e$'s
endpoint of lesser rank. Finally, denote by $F\left(C\right)$ the
set of free edges incident with $V\left(C\right)$ in $G_{t}$, where
$t=\rho\left(C\right)$. We partition $F\left(C\right)$ to vertical
and horizontal edges, writing $F\left(C\right)=F_{V}\left(C\right)\cup F_{H}\left(C\right)$.
\paragraph{\textsc{Breaker}'s Strategy $\mathcal{S}_{B}$:}
We assume \textsc{Maker} is the first player, so each round consists
of a move by \textsc{Maker} and a response by \textsc{Breaker}. For
a particular round, let $e=uv$ be the edge claimed by \textsc{Maker}
on her move. Assume without loss of generality that $\rho\left(u\right)\le\rho\left(v\right)$,
and let $C$ be the H-comp containing $u$ in \textsc{Maker}'s graph
after this move. \textsc{Breaker}'s strategy $\mathcal{S}_{B}$ is
to claim in each of his $b$ steps during this round:
\begin{itemize}
\item an arbitrary edge from $F_{V}\left(C\right)$, if $F_{V}\left(C\right)$
is nonempty; otherwise
\item an arbitrary edge from $F_{H}\left(C\right)$, if $F_{H}\left(C\right)$
is nonempty; otherwise
\item an arbitrary free edge.
\end{itemize}
It is clear that \textsc{Breaker} can follow $\mathcal{S}_{B}$ throughout
the game. The following claim characterises the horizontal components
allowed by $\mathcal{S}_{B}$:
\begin{claim}
\label{clm:H-comp-invariants}At the beginning of every round, every
H-comp $C$ satisfies exactly one of the following:
\begin{enumerate}
\item[$(i)$] \textsc{ Maker} claimed exactly one edge above $C$ and $F\left(C\right)$
is empty.
\item[$\left(ii\right)$] \textsc{ Maker} claimed no edges above $C$ and $\left|F\left(C\right)\right|\le\max\left\{ 0,b+2-\left|V\left(C\right)\right|\right\} $.
\end{enumerate}
\end{claim}
\begin{proof}
We prove this by induction on the number of rounds. Before the game
starts, $\left(ii\right)$ holds for every H-comp $C$ since no edges
have been claimed, $\left|V\left(C\right)\right|=1$ and $\left|F\left(C\right)\right|\le b+1$
by definition of $\left(b+2\right)$-rank. Assume the claim holds
at the beginning of round $r\ge0$; we show it holds at the end
of that round. Let $e=uv$ be the edge claimed by \textsc{Maker} on
her $r$th move, and assume WLOG that $\rho\left(u\right)\le\rho\left(v\right)$.
Denote by $C_{u}$ and $C_{v}$, respectively, the H-comps containing
$u$ and $v$ at the beginning of round $r$, and by $C$ the H-comp
containing $u$ after her move. It remains to show that after \textsc{Breaker}'s
move $C$ satisfies either $\left(i\right)$ or $\left(ii\right)$
since no other H-comp is affected by \textsc{Maker}'s move, and \textsc{Breaker}'s
move cannot disrupt an H-comp which was already satisfying $\left(i\right)$
or $\left(ii\right)$ from doing so. We distinguish between two cases:
\begin{casenv}
\item If $e$ is vertical, observe that $C=C_{u}$. Since we had $e\in F\left(C\right)$
before \textsc{Maker}'s move, $e$ must be the first edge \textsc{Maker}
claimed above $C$ by assumption. Before \textsc{Breaker}'s move we
have $\left|F\left(C\right)\right|\le b$ (as $e$ is already claimed),
so he claims all of $F\left(C\right)$ according to $\mathcal{S}_{B}$.
\item If $e$ is horizontal, $C=C_{u}\cup C_{v}$. Since both $F\left(C_{u}\right)$
and $F\left(C_{v}\right)$ contained $e$ at the beginning of the
round, both satisfied $\left(ii\right)$ and thus\textsc{ Maker} claimed
no edges above $C$. Moreover, before \textsc{Maker}'s move we had
$0<\left|F\left(C_{u}\right)\right|\le b+2-\left|V\left(C_{u}\right)\right|$,
and similarly for $v$, thus before \textsc{Breaker}'s move we have
\[
\left|F\left(C\right)\right|=\left|\left(F\left(C_{u}\right)\setminus\left\{ e\right\} \right)\cup\left(F\left(C_{v}\right)\setminus\left\{ e\right\} \right)\right|\le2\left(b+1\right)-\left|V\left(C\right)\right|
\]
and after his move, according to $\mathcal{S}_{B}$, either $F\left(C\right)$
is empty or $\left|F\left(C\right)\right|\le b+2-\left|V\left(C\right)\right|$.\qedhere
\end{casenv}
\end{proof}
Claim~\ref{clm:H-comp-invariants} yields the following two corollaries:
\begin{cor}
\label{cor:H-comp-size}Throughout the game $\left|V\left(C\right)\right|\le2\left(b+1\right)$
for every H-comp $C$.
\end{cor}
\begin{proof}
Indeed, no free edges are incident with H-comps of size $b+2$ or more
by Claim \ref{clm:H-comp-invariants}, and thus the largest H-comp
possibly achievable by \textsc{Maker} is obtained by claiming a free
horizontal edge between two H-comps of size $b+1$ each.
\end{proof}
\begin{cor}
\label{cor:T-gamma}Given a connected component $\Gamma$ of \textsc{Maker}'s
graph, if we contract every H-comp in $\Gamma$ to a single vertex,
we get a tree $T_{\Gamma}$ of height at most $\rho\left(C_{0}\right)$,
where $C_{0}$ is the H-comp in $\Gamma$ of maximal rank (that is,
the H-comp corresponding to the root of $T_{\Gamma}$).
\end{cor}
At this point we can already establish a weaker version of Theorem~\ref{thm:breaker-win}.
Indeed, using the above corollaries, Lemma~\ref{lem:volume-of-balls-in-Gnp}
and Theorem~\ref{thm:stabilization-time}, we obtain the following
poly-logarithmic bound on the size of \textsc{Maker}'s connected components.
\begin{prop}
\label{prop:breaker-strategy-is-ok}For $c<c_{b+2}$, $\ggnc$ is
WHP such that by playing according to $\mathcal{S}_{B}$, \textsc{Breaker}
ensures that all connected components in \textsc{Maker}'s graph are
of size $O\left(\log^{4b+5}n\right)$.
\end{prop}
\begin{proof}
Fix a connected component $\Gamma$ in \textsc{Maker}'s graph. The
height of $T_{\Gamma}$ from Corollary~\ref{cor:T-gamma} is bounded
by the stabilization time, and we have WHP $T^*= T^*\left(n,c\right)\le\log_{b+1}\log n+O\left(1\right)$
by Theorem~\ref{thm:stabilization-time}. Moreover, the distance
between two vertices in the same H-comp is at most $2b+1$ by Corollary~\ref{cor:H-comp-size},
so $\Gamma$ is contained in a ball of radius $t=2\left(b+1\right) T^*$
around any vertex $v\in V\left(\Gamma\right)$ of maximal rank. Applying
Lemma~\ref{lem:volume-of-balls-in-Gnp} yields the bound
\[
2t^{3}c^{t}\log n=O\left(\log^{3}\log n\cdot c^{2\left(b+1\right)\log_{b+1}\log n}\cdot\log n\right)=O\left(\log^{1+2\left(b+1\right)\log_{b+1}c}n\cdot\log^{3}\log n\right).
\]
Note that $c<c_{b+2}<2\left(b+1\right)$ and thus
\[
1+2\left(b+1\right)\log_{b+1}c<1+2\left(b+1\right)\left(1+\log_{b+1}2\right)\le4b+5,
\]
establishing the bound on the size of $C$.
\end{proof}
Proposition~\ref{prop:breaker-strategy-is-ok} gives a nonuniform
poly-logarithmic bound (i.e., the exponent depends on $b$, even when
computed precisely). Using Theorem~\ref{thm:shatter-time} and Claim~\ref{clm:vertex-expansion},
we improve the bound to a uniform $o\left(\log^{3}n\right)$ for any
value of $b$, and establish Theorem~\ref{thm:breaker-win}, via
the following proposition:
\begin{prop}
\label{prop:breaker-strategy-is-awesome}For $c<c_{b+2}$, $\ggnc$
is WHP such that if \textsc{Breaker} plays according to $\mathcal{S}_{B}$,
all connected components in \textsc{Maker}'s graph are of size $o\left(\log^{3}n\right)$.
\end{prop}
\begin{proof}
Fix a connected component $\Gamma$ in \textsc{Maker}'s graph. If
$\left|V\left(\Gamma\right)\right|<\log n$, we are done; otherwise,
let $U'\subseteq V\left(\Gamma\right)$ be the set of vertices of
rank at least $t^{\dagger}$, where $t^{\dagger}=t^{\dagger}\left(c\right)$
is the constant from Theorem~\ref{thm:shatter-time}, and let $U$
be the vertex set of a connected $\Gamma'\subseteq\Gamma$ of minimal
size such that $U'\subseteq U$ and $\left|U\right|\ge\log n$. By
Theorem~\ref{thm:shatter-time}, $\left|U\right|=o\left(\log^{3}n\right)$.
Furthermore, $\Gamma$ is contained in a ball of radius $t=2\left(b+1\right)t^{\dagger}$
around $U$, thus $t$ applications of Claim~\ref{clm:vertex-expansion}
yield the desired bound
\[
\left|V\left(\Gamma\right)\right|\le\left(3c\right)^{t}\left|U\right|=O\left(\left|U\right|\right)=o\left(\log^{3}n\right).\qedhere
\]
\end{proof}
\begin{rem*}
A slightly refined strategy by \textsc{Breaker} would be to claim
edges from $F_{H}\left(C\right)$ before edges from $F_{V}\left(C\right)$
if $\left|F_{V}\left(C\right)\right|>b$. It is not very hard to see
that by this he makes sure that throughout the game all H-comps except
the ``root'' ones are of size at most two. This improves the exponent
in Proposition~\ref{prop:breaker-strategy-is-ok} to $3+\log_{b+1}4$,
which approaches~$3$ from above as~$b$ increases. However, for
all $b\ge1$ this is still inferior to the bound given in Proposition~\ref{prop:breaker-strategy-is-awesome}
(which does not benefit from the modified strategy). We thus chose
to present the simpler strategy.
\end{rem*}
\section{\label{sec:maker}Her Side}
It is quite easy to prove a weaker version of Theorem~\ref{thm:maker-win}:
that WHP $s_{b}^{*}\left(\gnc\right)=\Omega\left(n\right)$
when $c>c_{b+3}$. Indeed, in this case $\mathcal{K}'$, the $\left(b+3\right)$-core
of $\gnc$, is WHP of linear size and a very
good edge expander. It follows that the naive tree-building strategy
within $\mathcal{K}'$ (i.e., the strategy that builds a single tree
$T\subset\mathcal{K}'$ by repeatedly claiming an arbitrary free edge
of $\partial T$, starting from an arbitrary vertex of $\mathcal{K}'$)
is successful for \textsc{Maker}, since as long as $T$ is sublinear in size, i.e., has $o(n)$ vertices, we have
$\left|\partial T\right|>b\left|V\left(T\right)\right|$ and \textsc{Breaker}
cannot claim all the boundary of $T$. This is the same strategy \textsc{Maker}
uses in~\cite{HN-14} when playing on a random $\left(b+3\right)$-regular
graph.
For $c_{b+2}<c<c_{b+3}$, however, the above strategy applied to~$\mathcal{K}$, the $\left(b+2\right)$-core of $\ggnc$, is prone to
being shut down by \textsc{Breaker} in the early stages of the game
despite the minimum degree in $\mathcal{K}$ being $b+2$. Thus in the proof of Theorem~\ref{thm:maker-win},
\textsc{Maker} uses a tree-building strategy that begins in a more
refined manner. In the tree $T\subset\mathcal{K}$ \textsc{Maker}
builds, she makes sure to have many heavy vertices (i.e., of degree
at least $b+3$ in $\K$). Only after securing sufficiently many such
vertices does she continue by growing $T$ naively.
\medskip{}
Recall that for a rooted tree, a non-leaf vertex is $k$-light if it has less
than $k$ children, and it is $k$-heavy otherwise, and that a tree path (a descending path towards the leaves) is $k$-light if it consists solely of $k$-light vertices.
Note that since leaves are not considered light, any tree path terminating in a leaf is not $k$-light by definition (although permitting leaves to be light would require only a trivial modification of our arguments).
\medskip
For simplicity, in the remainder of the paper $k$ stands for $b+2$, and $\K$ is the $k$-core of $G$.
The proof of Theorem~\ref{thm:maker-win} contains three main ingredients, stated in three lemmas. The first two lemmas require the following definition.
\begin{defn*}
An $\left(N,L\right)$-tree is a balanced tree (i.e., all of its leaves are at the same level) of height $N$ with no $\left(k-1\right)$-light vertices and no $k$-light tree paths of length $L$.
\end{defn*}
\begin{lem}
\label{lem:accumulate-heavy}In the $\left(1:k-2\right)$ \textsc{Maker\textendash Breaker}
game played on the edge set of an $\left(N,L\right)$-tree $T^*$,
\textsc{Maker} has a strategy to build a subtree $T\subset T^*$
with at least $\alpha^{N}$ vertices that are $k$-heavy
in $T^*$, for some constant $\alpha=\alpha\left(k,L\right)>1$.
Moreover, using this strategy \textsc{Maker}'s graph is a tree throughout
the game, until $T$ is achieved.
\end{lem}
We assume the hypothesis of Theorem~\ref{thm:maker-win}, i.e., $c>c_k$, in the following two lemmas.
\begin{lem}
\label{lem:good-tree} $\mathcal{K}$ contains WHP an $\left(N,L\right)$-tree
for $N=\log^{2}\log n$ and some constant $L=L\left(k,c\right)$.
\end{lem}
\begin{lem}
\label{lem:no-bad-two-cores}There exists some $\eps=\eps\left(k,c\right)>0$
such that WHP any subgraph $C\subset\mathcal{K}$ with minimum degree
at least~$2$, $\xs\left(C\right)\ge\log^{4}n$ and $\left|\partial C\right|\le (k-2)\left|V\left(C\right)\right|$
satisfies $\left|V\left(C\right)\right|\ge\eps n$.
\end{lem}
Lemma~\ref{lem:accumulate-heavy} is proved in Section~\ref{sec:accumulate-heavy}, while the proofs of Lemmas~\ref{lem:good-tree} and~\ref{lem:no-bad-two-cores}
are postponed to Sections~\ref{sec:good-tree} and~\ref{sec:no-bad-two-cores}, respectively.
Next we show how these lemmas imply Theorem~\ref{thm:maker-win}, but first we need the following easy claim.
\begin{claim}
\label{clm:2-core}Let $F$ be an induced subgraph of $\mathcal{K}$
with non-negative excess such that $\left|\partial F\right|\le b\left|V\left(F\right)\right|$,
and let $C$ be the $2$-core of $F$ (note that $C$ is not empty since $\xs\left(F\right)\ge0$ and thus
$F$ contains a cycle). Then $\xs\left(C\right)\ge\xs\left(F\right)$ and $\left|\partial C\right|\le b\left|V\left(C\right)\right|$.
\end{claim}
\begin{proof}
In order to obtain $C$ from $F$ we use a sequential vertex-deleting
variant of the $2$-peeling process. Starting from $F_{0}=F$, as
long as the minimum degree of $F_{t}$ is less than two, we obtain
$F_{t+1}$ from $F_{t}$ by deleting one arbitrary vertex $v\in V\left(F_{t}\right)$
of degree $\deg_{F_{t}}\left(v\right)\le1$. Regardless of the order
of deletions, this process terminates with the 2-core $C=F_{ T^*}$.
We prove that $\xs\left(F_{t}\right)\ge\xs\left(F\right)$ and $\left|\partial F_{t}\right|\le b\left|V\left(F_{t}\right)\right|$
for all $t\ge0$ by induction.
For $t=0$ there is nothing to prove; assuming it holds for $t$,
let $v\in V\left(F_{t}\right)$ be the vertex selected for deletion,
whose degree in $F_{t}$ is $0$ or $1$. In $F_{t+1}$ there is one
less vertex (i.e., $v$) and at most one less edge than $F_{t}$,
hence $\xs\left(F_{t+1}\right)\ge\xs\left(F_{t}\right)\ge\xs\left(F\right)$;
furthermore
\[
\partial F_{t+1}=\partial F_{t}\cup E\left(v,V\left(F_{t+1}\right)\right)\setminus E\left(v,V\left(\mathcal{K}\setminus F_{t}\right)\right)
\]
so
\[
\left|\partial F_{t+1}\right|=\left|\partial F_{t}\right|+\deg_{F_{t}}\left(v\right)-\left(\deg_{\mathcal{K}}\left(v\right)-\deg_{F_{t}}\left(v\right)\right)\le b\left|V\left(F_{t}\right)\right|+1-\left(b+1\right)=b\left|V\left(F_{t+1}\right)\right|.\qedhere
\]
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:maker-win}]
By Lemma~\ref{lem:good-tree} \textsc{Maker} can WHP locate a $\left(\log^{2}\log n,L\right)$-tree
$T^*$ in $\mathcal{K}$. She first builds a subtree $T\subset T^*$
with at least $\alpha^{\log^{2}\log n}\gg2\log^{4}n$ vertices of
degree at least $b+3$ in $\K$, which she can do by Lemma~\ref{lem:accumulate-heavy}. She then proceeds naively by claiming in every move an arbitrary
free edge in $\partial T$ as long as possible. \textsc{Maker} can
no longer proceed with her strategy only when \textsc{Breaker} has
claimed the entire boundary of $T$, and in particular $\left|\partial T\right|\le b\left|V\left(T\right)\right|$.
When this happens, consider the subgraph $F\subset\mathcal{K}$ induced by $V\left(T\right)$.
Clearly $V\left(F\right)=V\left(T\right)$ and thus $\partial F=\partial T$.
Now $F$ satisfies
\begin{align*}
2\left|E\left(F\right)\right| =\sum_{v\in V\left(F\right)}\deg_{F}\left(v\right) & =\sum_{v\in V\left(F\right)}\deg_{\mathcal{K}}\left(v\right)-\left|\partial F\right| \\
& \ge \left(b+2\right)\left|V\left(T\right)\right|+2\log^{4}n-b\left|V\left(T\right)\right| = 2\left(\left|V\left(F\right)\right|+\log^{4}n\right),
\end{align*}
meaning $\xs\left(F\right)\ge\log^{4}n$. By Claim~\ref{clm:2-core}
the 2-core $C$ of $F$ has large excess and small boundary, namely
$\xs\left(C\right)\ge\xs\left(F\right)\ge\log^{4}n$ and $\left|\partial C\right|\le b\left|V\left(C\right)\right|$.
Finally, by Lemma~\ref{lem:no-bad-two-cores}, WHP $C$ has at least
$\eps n$ vertices and thus
\[
\left|V\left(T\right)\right|=\left|V\left(F\right)\right|\ge\left|V\left(C\right)\right|\ge\eps n,
\]
establishing the theorem.
\end{proof}
\subsection{\label{sec:accumulate-heavy}Proof of Lemma~\ref{lem:accumulate-heavy}:
accumulating heavy vertices}
An $\left(N,L\right)$-tree is \emph{simple} if its root is $k$-light, and all $k$-heavy vertices in the tree have exactly $k$ children. Clearly, any $\left(N,L\right)$-tree contains a simple $\left(N,L\right)$-tree as a subtree, so we may assume without loss of generality that $T^*$ is simple. Note that $w$, the root of $T^*$, has degree $k-1$, every other $k$-light vertex but the leaves has degree $k$, and all $k$-heavy vertices of $T^*$, to which we refer from now on simply as heavy, have degree $k+1$. In this subsection $l(v)$ denotes the level of any vertex $v\in V( T^*)$; in particular, $l(w)=0$ and $l(v)=N$ for every leaf $v$. Whenever we refer to an edge $uv\in E( T^*)$ we assume $l(u)<l(v)$, and define the level of~$uv$ to be~$l(v)$.
We now describe \textsc{Maker}'s strategy $\mathcal{S}_{\textsc{M}}$. Throughout the game, \textsc{Maker}'s graph is a single tree $T=T^{(r)}\subset T^*$, where $T^{(r)}$ denotes her tree after the $r$th round. Initially, $T^{\left(0\right)}$ consists of $w$ solely. In
each move, as long as there exist free edges of level smaller than $N$ in $\partial T$, \textsc{Maker} enlarges her tree by claiming one of these edges, selecting an arbitrary edge of minimum level. The game stops when \textsc{Maker} cannot proceed with this strategy, i.e., when all edges in $\partial T$ have already been claimed by \textsc{Breaker} (note that by definition none of them could have been previously claimed by \textsc{Maker}), except perhaps for some edges in level $N$.
Given $\mathcal{S}_{\textsc{M}}$, we may assume that \textsc{Breaker} only claims edges from $\partial T$ as well. Indeed, suppose that \textsc{Breaker}, according to his strategy, wishes to claim an edge $uv \not\in \partial T$ in one of his steps. He can claim instead the (unique) edge $u'v' \in \partial T$ such that $v'$ is an ancestor of $v$. By $\mathcal{S}_{\textsc{M}}$, this prevents \textsc{Maker} from claiming any edge in the subtree of $T^*$ rooted at $v'$, and particularly the edge $uv$. If the edge $u'v'$ was already claimed by \textsc{Breaker}, then he claims an arbitrary free edge from $\partial T$ (if none of those exists the game ends anyway). It is evident that \textsc{Breaker} cannot be harmed from this modification of his strategy, and our assumption is justified.
We now assume that \textsc{Maker} plays second and follows $\mathcal{S}_{\textsc{M}}$ (it is trivial to see that she can do so), and show that it is a winning strategy for her, that is, we show that when the game is stopped, her tree $T$ contains sufficiently many heavy vertices. Before doing so we need some additional terminology. First, since both players only claim edges from $\partial T$ by assumption, we naturally redefine free edges to be unclaimed edges from $\partial T$ (instead of all unclaimed edges in $T^*$) for the remainder of this subsection. For $j=1,2,\ldots,N-1$, level $j$ is \emph{complete}
when no free edges remain in level $j$ or less. An edge $uv\in E(T)$ \emph{survives} in level $j$ for $j\ge l\left(v\right)$ if $T$ contains a level $j$ descendant of $v$. Finally, we define $C=C(k,L)=1+k^{L+1}$.
\begin{claim}
\label{clm:FreeEdgesNum} \emph{For $r\ge0$ let $h_{r}$ denote the
number of heavy vertices in $T^{(r)}$, and for $r\ge1$
let $f_{r}$ denote the number of free edges before }\textsc{Maker}\emph{'s
$r$th move. Then $f_{r+1} = h_{r}+1$ for every $r\ge0$.}
\end{claim}
\begin{proof}
Since $T^*$ is simple, and since \textsc{Maker} never claims an edge in level $N$, it follows
that
\[
\left|\partial T^{(r)}\right|=\sum_{v\in V\left(T^{(r)}\right)}\deg_{ T^*}(v)-2\left|E\left(T^{(r)}\right)\right|=(k-1)+rk+h_{r}-2r.
\]
In each of his first $r+1$ moves \textsc{Breaker} claims $b=k-2$
edges, each of them incident with some $v \in V(T)$, so $f_{r+1} = \left|\partial T^{(r)}\right|-(r+1)(k-2)=h_{r}+1$.
\end{proof}
An immediate corollary of Claim~\ref{clm:FreeEdgesNum} is that at
the end of the game $T$ is of height $N-1$, and level $N-1$ is complete. Indeed, since there exist free edges before each of \textsc{Maker}'s moves, the game stops only when all free edges are in level $N$. Having a free edge in level $N$ means that $T$ must have reached level $N-1$.
\begin{claim}
\label{cl:constant} Let $s$ and $j<N-L$
be two integers, and assume that there are~$s$ free edges in level~$j$ right before \textsc{Maker} claims her first edge
in level~$j$. Then at least~$s/C$ of them will be claimed by \textsc{Maker} and survive in level~$j+L$.
\end{claim}
\begin{proof}
Consider first the situation right before \textsc{Maker} claims her first edge in level $j$, and denote the $s$ free edges in this level by $u_{1}v_{1},\dots,u_{s}v_{s}$ (the
parents $u_{i}$ are not necessarily distinct, but the children $v_{i}$
are). For each $1\le i \le s$, let $T_{i}\subset T^*$ be the subtree rooted at~$u_{i}$
consisting of the edge $u_{i}v_{i}$ and the subtree of $T^*$
rooted at $v_{i}$ of height $L$. Note that all free edges are in level $j$ at this point by $\mathcal{S}_{\textsc{M}}$. By the strategies of \textsc{Maker} and \textsc{Breaker}, it follows that all edges in each $T_i$ are still unclaimed, and that exactly one of them, namely $u_iv_i$, is considered free in our new terminology.
Now let us examine the game when level $j+L$ is complete. Denote by $M\subseteq\left\{ 1,2,\ldots,s\right\} $
the set of indices of the edges that survived in level $j+L$ of $T$,
and by $B$ its complement, so $|M|+|B|=s$. We need to show that
$|M|>s/C$.
Let $m_{i}$ and $b_{i}$ denote the number of steps that were played
in $T_{i}$ by \textsc{Maker} and \textsc{Breaker}, respectively.
Recalling that $T^*$ is simple, for every $i=1,2,\ldots,s$, we have the trivial bound
\[
m_{i}\le|E(T_{i})|\le\sum_{t=0}^{L}k^{t}=\frac{k^{L+1}-1}{k-1} < \frac{k^{L+1}}{k-2}.
\]
Now let $i\in B$. Since \textsc{Maker} did not reach level $L$ in
$T_{i}$, i.e., did not reach any of its leaves, it follows by the assumption on \textsc{Breaker}'s strategy that every
step \textsc{Maker} made in $T_{i}$ increased the number of free edges in this
tree by at least $k-2$. Since no free edges remain in $T_{i}$, and
there was one free edge there at the beginning of the analysis, it
follows that $b_{i}\geq m_{i}(k-2)+1$.
During this analysis, which begins with \textsc{Maker}'s move, she
only plays in $T_{1},\ldots,T_{s}$. Thus $\sum_{i=1}^{s}b_{i}\leq(k-2)\sum_{i=1}^{s}m_{i}$.
The left hand side can be bounded from below by
\[
\sum_{i=1}^{s}b_{i}\geq\sum_{i\in B}b_{i}\geq\sum_{i\in B}\left(m_{i}(k-2)+1\right)=(k-2)\sum_{i\in B}m_{i}+|B|,
\]
while the right hand side can be bounded from above by
\[
(k-2)\sum_{i=1}^{s}m_{i}=(k-2)\left(\sum_{i\in B}m_{i}+\sum_{i\in M}m_{i}\right)<(k-2)\sum_{i\in B}m_{i}+|M|k{}^{L+1}.
\]
Putting it all together, we get $|B|<|M|k^{L+1}$, which implies $s=|M|+|B|<C|M|$,
thus the proof is complete.
\end{proof}
\begin{claim}
\label{cl:heavy} For every $0\le i\le\lfloor N/(C(L+1))\rfloor$,
when level $iC(L+1)$ is complete, $T$ contains at least $(2^{i}-1)C$
heavy vertices.
\end{claim}
\begin{proof}
We prove by induction on $i$. The claim holds trivially for $i=0$.
Assuming it holds for $i$, we show that it holds for $i+1$ as well.
Let $0\le j<C$ and write $J_{j}^{i}=(iC+j)(L+1)$. By the induction
hypothesis and by Claim~\ref{clm:FreeEdgesNum}, right before \textsc{Maker}
claims her first edge in level $J_{j}^{i}+1>iC\left(L+1\right)$ there
are at least $(2^{i}-1)C+1$ free edges, all of them in level $J_{j}^{i}+1$
by $\mathcal{S}_{M}$. By Claim~\ref{cl:constant}, at least $2^{i}$
of these free edges will be claimed by \textsc{Maker} and survive
in level $J_{j+1}^{i}$, resulting in at least $2^{i}$ vertex disjoint
tree paths of length $L$ in $T$, each of them containing at least
one heavy vertex by the property of $T^*$. It follows that $T$
contains at least $2^{i}C$ heavy vertices between levels $J_{0}^{i}+1$
and $J_{0}^{i+1}$. By the induction hypothesis $T$ also contains
at least $(2^{i}-1)C$ heavy vertices until level $J_{0}^{i}$, and
the claim holds.
\end{proof}
The game ends when level $N-1$ is complete, and~$T$ then contains by Claim~\ref{cl:heavy}
at least $(2^{\lfloor (N-1)/(C(L+1))\rfloor}-1)C>\alpha^{N}$
heavy vertices for an appropriate $\alpha=\alpha\left(k,L\right)>1$,
establishing Lemma~\ref{lem:accumulate-heavy}.
\section{\label{sec:technical-background}Technical Background}
In this section we describe the technical background required for
Section~\ref{sec:technical-proofs}.
\subsection{The configuration model }
We begin this section with a description of the so-called \emph{configuration model}, introduced by Bollob{\'a}s~\cite{B-80}, which is extremely useful for generating random graphs with a given degree sequence.
Given two positive integers $n$ and $m$, fix an arbitrary degree sequence $\dd \in \D_{n,2m}$. Let $V = \{v_1,\dots,v_n\}$ be a set of vertices, where each $v_i$ is incident with $d_i$ labelled half-edges. Let $W$ denote the set of all half-edges, and let $F$ be a partition of $W$ into $m$ pairs; such a partition, which may also be viewed as a perfect matching of the half-edges, is called a \emph{configuration}. Note that there are exactly $(2m-1)!!$ configurations for $\dd$. By forming an edge from any two half-edges which belong to same pair in $F$, we obtain a multigraph $H = H(F)$ on the vertex set $V$, such that $d_H(v_i) = d_i$.
Let $\Omega^{*}(n,\dd) = \{H(F) \mid \textrm{$F$ is a configuration for $\dd$}\}$ be the set of all multigraphs on $n$ labelled vertices with degree sequence $\dd$, and let $\mathcal{G}^{*}(n,\dd)$ be the probability space of $\Omega^{*}(n,\dd)$ when $F$ is chosen uniformly at random from all possible configurations for $\dd$. Let $\Omega(n,\dd)$ be the set of all simple graphs in $\Omega^{*}(n,\dd)$, and let $\mathcal{G}(n,\dd)$ be the uniform distribution over $\Omega(n,\dd)$. We will make use of the following theorem, due to Frieze and Karo{\'n}ski~\cite{FK-15}.
\begin{thm}[{\cite[Theorem~10.3]{FK-15}}]\label{thm:prob-simple}
Let $\dd = (d_1,\dots,d_n)$ and assume that $\Delta(\dd) \le n^{1/6}$ and $\sum_{i=1}^{n} [d_i]_2 = \Omega(n)$. Then for any multigraph property $\P$
$$\Pr_{G \sim \mathcal{G}(n,\dd)}\left[G \in \P\right] \le (1+o(1))e^{\lambda(\lambda+1)}\cdot\Pr_{G^* \sim \mathcal{G}^*(n,\dd)}\left[G^* \in \P\right],$$
where
$\lambda = \lambda(\dd) = \frac12 \sum_{i=1}^{n}[d_i]_2/ \sum_{i=1}^{n}d_i$.
\end{thm}
\subsection{\label{sec:branching-gnp}Exploring $\gnc$
via a Poisson branching process}
The main ingredient in the proof of Theorem~\ref{thm:shatter-time} is coupling local behavior
in $\gnc$ with an appropriate branching
process. We describe the coupling quickly, borrowing much of the notation
from~\cite{Rio-08}.
We also refer the reader to~\cite{CW-06,JL-07}. Recall that $\Psi_{j}\left(\lambda\right)$, $\Psi_{\ge j}\left(\lambda\right)$ and $\Psi_{<j}\left(\lambda\right)$ denote the probabilities of a $\Poisson(\lambda)$ random variable being equal to $j$, at least $j$, or less than $j$, respectively.
Let $X_{c}$ be a Galton\textendash Watson branching process that
starts with a single particle $x_{0}$ in generation zero, where the
number of children of each particle is an independent $\Poisson$$\left(c\right)$
random variable. Define a sequence $B_{0}\supseteq B_{1}\supseteq\cdots$
of events: for an integer $t\ge0$, let $B_{t}=B_{t}\left(c\right)$
be the event that $X_{c}$ contains a complete $\left(k-1\right)$-ary
tree of height $t$ rooted at $x_{0}$, and let $B=\bigcap_{t\ge0}B_{t}=\lim_{t\to\infty}B_{t}$
be the event that $X_{c}$ contains an infinite $\left(k-1\right)$-ary
tree rooted at $x_{0}$. Denote the probability of $B_{t}$ by $\beta_{t}$
and the probability of $B$ by $\beta=\lim_{t\to\infty}\beta_{t}$.
Then $\beta_{0}=1$. Also, each particle in the first generation of
$X_{c}$ has probability $\beta_{t}$ of having property $B_{t}$,
independently, so the number of such particles is distributed $\Poisson\left(c\beta_{t}\right)$.
Thus, $\beta_{t+1}=\Psi_{\ge k-1}\left(c\beta_{t}\right)$.
Since $x\mapsto\Psi_{\ge k-1}\left(cx\right)$ is a continuous increasing
function, $\beta=\beta\left(c\right)$ is the maximum solution to
the equation $x=\Psi_{\ge k-1}\left(cx\right)$. Recall $c_{k}$,
defined in Section~\ref{sec:intro} as the threshold for the appearance
of a nonempty $k$-core in $\gnc$; in terms
of the process $X_{c}$, we have
\begin{equation}
c_{k}=\inf\left\{ c:\beta\left(c\right)>0\right\} .\label{eq:ck-definition}
\end{equation}
\medskip{}
Given a graph $G$ with degree sequence $(d_1,\dots,d_n)$, the \emph{degree histogram} of $G$ is the sequence $D=\left(D_{0},D_{1},\ldots\right)$ such that $D_j=|i:d_i=j|$ for all $j$.
Let us describe the likely degree histogram $D^{t}$ of the
graph $G_{t}$ obtained after $t$ steps of the peeling process. For
$t=0$ the binomial degree distribution of $G_{0}\sim\gnc$
is asymptotically $\Poisson\left(c\right)$. In particular, it is easy to see (e.g., via the second moment method) that WHP
\begin{equation}
D_{j}^{0}\left(n\right)=\begin{cases}
\left(1+o\left(1\right)\right)\Psi_{j}\left(c\right)n, & j= o\left(\log n/\log\log n\right);\\
\Theta\left(\Psi_{j}\left(c\right)n\right), & j= \Theta\left(\log n/\log\log n\right);\\
0, & j= \omega\left(\log n/\log\log n\right).
\end{cases}\label{eq:degree-distribution-gnc}
\end{equation}
Now, for integers $j\ge0$ and $t>0$ let
\begin{equation}
\delta_{j}^{t}=\Psi_{j}\left(c\beta_{t}\right)\Psi_{\ge k-j}\left(c\beta_{t-1}-c\beta_{t}\right).\label{eq:branching-process-degrees-after-t-steps}
\end{equation}
\begin{claim}
\label{clm:degree-distribution-after-t-steps}For fixed $t\ge 0$ and $j\ge 0$
we have $\E\left[D_{j}^{t}\left(n\right)\right]=\left(1+o\left(1\right)\right)\delta_{j}^{t}n$. Moreover, WHP
$D_{j}^{t}\left(n\right)= \left(1+o\left(1\right)\right)\delta_{j}^{t}n$.
\end{claim}
\begin{proof}
We show that a vertex $v$ has degree $j$ in $G_{t}$ with probability
$\left(1+o\left(1\right)\right)\delta_{j}^{t}$. By Claim~\ref{clm:no-cycles}
we only need to consider the case where the local neighbourhood of
$v$ is a tree, i.e., behaves like the branching process $X_{c}$.
Consider first the case $j\ge k$, and note that here $\Psi_{\ge k-j}\left(c\beta_{t-1}-c\beta_{t}\right)=1$.
Each neighbour of $v$ survives $t$ steps of the $k$-peeling process
with probability $\beta_{t}$, independently, and thus the number
of such neighbours is distributed $\Poisson\left(c\beta_{t}\right)$.
Thus, the number of neighbours of $v$ in $G_{t}$ is exactly $j$
with probability $\left(1+o\left(1\right)\right)\Psi_{j}\left(c\beta_{t}\right)$.
For $0<j<k$, having $j$ neighbours is not sufficient for $v$ to
actually survive $t$ steps; we need $v$ to have at least $k$ neighbours
in $G_{t-1}$, of which exactly $j$ have survived $t$ steps. Thus,
the event $B_{t-1}\setminus B_{t}$, whose probability is $\beta_{t-1}-\beta_{t}$,
holds for at least $k-j$ neighbours of $v$. Again, the number of
such neighbours is distributed $\Poisson\left(c\beta_{t-1}-c\beta_{t}\right)$
and thus~$v$ has~$j$ neighbours in~$G_{t}$ and at least~$k-j$ neighbours
in $G_{t-1}\setminus G_{t}$ with probability
$$\left(1+o\left(1\right)\right)\Psi_{j}\left(c\beta_{t}\right)\Psi_{\ge k-j}\left(c\beta_{t-1}-c\beta_{t}\right)=\left(1+o\left(1\right)\right)\delta_{j}^{t}.$$
By linearity of expectation, we have $\E\left[D_{j}^{t}\left(n\right)\right]=\left(1+o\left(1\right)\right)\delta_{j}^{t}n$,
proving the first statement. Calculations similar to the above for
a pair of vertices $u,v$ show that $\var\left[D_{j}^{t}\left(n\right)\right]=o(\E\left[D_{j}^{t}\left(n\right)\right]^{2})$.
Since $\E\left[D_{j}^{t}\left(n\right)\right]\to\infty$, this establishes the sharp concentration of $D_{j}^{t}\left(n\right)$.
\end{proof}
\begin{rem*}
For every vertex of degree $0$ in $G_t$ we have two options. If the vertex had degree at least $k$ in $G_{t-1}$, then the analysis is identical to that of the case $0<j<k$ in Claim~\ref{clm:degree-distribution-after-t-steps}. Otherwise, less than $k$ of its neighbours survived $t-1$ peeling steps. Thus, WHP
\begin{equation}\label{eq:degree-0-after-t-steps}
D_{0}^{t}\left(n\right)=\left(1+o\left(1\right)\right)\left(\delta_{0}^{t}+\Psi_{<k}\left(c\beta_{t-1}\right)\right)n.
\end{equation}
\end{rem*}
\subsection{\label{sec:new}Degree histogram of the $k$-core}
Considering the peeling process of Section~\ref{sec:branching-gnp}, the limit of the formula in Claim~\ref{clm:degree-distribution-after-t-steps} for large $t$ is suggestive of the asymptotics of the degree histogram of the $k$-core $\K$ of $\ggnc$, but it is difficult to make this approach rigorous.
We quote~\cite{CW-06} for a convenient source of the following two lemmas.
With $\beta$ defined as in Section~\ref{sec:branching-gnp}, we define $\mu_c=c\beta$. We let $\nn$ denote the number of vertices in $\K$, and $\hat m$ the number of edges.
\begin{lem}\label{CWnm}
The number of vertices in $\K$ WHP satisfies
\begin{equation}\label{eq:number-of-vertices-in-k-core}
\hat{n} = \left(1+o\left(1\right)\right)\Psi_{\ge k}\left(\mu_{c}\right)n,
\end{equation}
and the number of edges in $\K$ WHP satisfies
\begin{equation}\label{eq:number-of-edges-in-k-core}
\hat{m} = \left(1+o\left(1\right)\right)\frac 12\mu_{c}\Psi_{\ge k-1}\left(\mu_{c}\right)n.
\end{equation}
\end{lem}
\begin{lem}\label{CWdeg}
WHP the degree distribution in $\K$ is asymptotically that of $Z_{k}\left(\mu_{c}\right)$, i.e., the $k$-truncated
Poisson with parameter $\mu_{c}$. Equivalently, the number of vertices of degree $j$ is 0 for $j<k$, whilst for fixed $j\ge k$ it is
\begin{equation}\label{degree-asy}
(1+o(1))\Pr[Z_{k}\left(\mu_{c}\right)=j]\hat n.
\end{equation}
\end{lem}
We will also find an alternative expression for the average degree useful. For every $\mu>0$ and $j\ge k$ we have
\begin{align*}
\frac{\Pr\left[Z_{k}\left(\mu\right)=j\right]}{\Pr\left[Z_{k-1}\left(\mu\right)=j-1\right]} & =\frac{\Psi_{j}\left(\mu\right)/\Psi_{\ge k}\left(\mu\right)}{\Psi_{j-1}\left(\mu\right)/\Psi_{\ge k-1}\left(\mu\right)}\\
& =\frac{e^{-\mu}\mu^{j}/j!}{e^{-\mu}\mu^{j-1}/\left(j-1\right)!}\cdot\frac{\Psi_{\ge k-1}\left(\mu\right)}{\Psi_{\ge k}\left(\mu\right)}=\frac{\mu}{j}\cdot\frac{\Psi_{\ge k-1}\left(\mu\right)}{\Psi_{\ge k}\left(\mu\right)},
\end{align*}
so, denoting the average degree in $\K$ by $\hat{d}:=2\hat{m}/\hat{n}$, for every $j \ge k$ we have
\begin{equation}\label{eq:average-degree-in-k-core}
\hat{d}=\frac{2\hat{m}}{\hat{n}}=\left(1+o\left(1\right)\right)\frac{\mu_{c}\Psi_{\ge k-1}\left(\mu_{c}\right)n}{\Psi_{\ge k}\left(\mu_{c}\right)n}=\left(1+o\left(1\right)\right)\frac{j\Pr\left[Z_{k}\left(\mu_{c}\right)=j\right]}{\Pr\left[Z_{k-1}\left(\mu_{c}\right)=j-1\right]}.
\end{equation}
There is one more fact that we will need about the joint degree distribution of vertices in the $k$-core.
Let $\gnm$ denote the Erd\H{o}s--R\'enyi random graph model, which is the uniform distribution over all graphs with $n$ vertices and $m$ edges.
For a fixed $k$, let $\K(n,m,k)$ be the probability space of graphs with minimum degree at least $k$, distributed as the $k$-core of $\gnm$, and similarly define $\K(n,c,k)$ for $\gnc$. When defining these graphs, we only consider the vertices in the $k$-core, so the vertex numbering is compressed into the range $[1, \nn]$, where $\nn \le n$ is the (random) number of vertices in the $k$-core, while maintaining their order from the original graph.
For the $\K(n,m,k)$ case, a restatement of~\cite[Corollary~1]{CW-06} gives the following approximation to the distribution of its degree sequence. Let ${\cal M}(\nn,\hat m, k)$ denote the probability space of sequences $(M_1,\ldots, M_{\nn})$ with the multinomial distribution, with parameters $\nn$ and $2\hat m$, but conditioned upon $M_i\ge k$ for all $i$. Let ${\cal M}(\nn,\hat m)$ denote the probability space of sequences $(X_1,\ldots, X_{\nn})$ summing to $2\mm$ with the multinomial distribution, that is, for every $(d_1, \dots, d_{\nn}) \in \D_{\nn,2\mm}$, the probability that $X_i = d_i$ for all $i$ is $(2\mm)!/(\nn^{2\mm}\prod d_i!)$. Let ${\cal M}(\nn,\hat m, k)$ denote the same probability space, but conditioned upon $X_i\ge k$ for all $i$.
\begin{prop}\label{thm:maincorfirst}
Let $k\ge 3$ and $c>c_k$ be fixed, and $m = (1+o(1))cn/2$. Let $H_n$ be any event in
the probability space defined by the random vector distributed as the degree
sequence of $\K(n,m,k)$.
Suppose that whenever $\hat n$ and $\hat m$ have the asymptotic behaviour given in~\eqref{eq:number-of-vertices-in-k-core} and~\eqref{eq:number-of-edges-in-k-core} respectively, it follows that
$$
\prob_{{\cal M}(\nn,\hat m, k)} (H_n) <P_n.
$$
Then $\prob(H_n)=O( P_n)$+o(1).
\end{prop}
Next, let $\P(\nn,\mm,k)$ denote the probability space of sequences consisting of $\hat n$ independent copies of $Z_k(\lambda)$, where $\lambda$ is chosen so that
\begin{equation}\label{lambdadef}
\E \left[Z_k(\lambda)\right] = \frac{2\hat{m}}{\hat{n}}.
\end{equation}
As observed in the proof of~\cite[Lemma~1]{CW-06}, the probability that the sum of $\nn$ copies of $Z_k(\lambda)$ is $2\hat m$ is $\Omega(1/\sqrt{\nn})$. It follows that we may replace $\prob_{{\cal M}(\nn,\hat m, k)}$ by $\prob_{\P(\hat n,\mm,k)}$ in the above proposition, as long as we replace $O( P_n)$ by $O(\sqrt n \cdot P_n)$. We may also replace $\K(n,m,k)$ by $\K(n,c,k)$ where $c=2m/n$, using the well-known strong connection between $\gnm$ and $\gnp$ in this case. That is, we have the following.
\begin{thm}\label{thm:maincor}
Let $k\ge 3$ and $c>c_k$ be fixed, and let $H_n$ be any event in
the probability space defined by the random vector distributed as the degree
sequence of $\K(n,c,k)$.
Suppose that whenever $\hat n$ and $\hat m$ have the asymptotic behaviour given in~\eqref{eq:number-of-vertices-in-k-core} and~\eqref{eq:number-of-edges-in-k-core} respectively, it follows that
$$
\prob_{\P(\hat n,\mm,k)} (H_n) <P_n.
$$
Then $\prob(H_n)=O(\sqrt n \cdot P_n)$+o(1).
\end{thm}
\section{\label{sec:technical-proofs}Technical Proofs}
In this section we prove Theorem~\ref{thm:shatter-time} (in Section~\ref{sec:shatter-proof}),
Lemma~\ref{lem:good-tree} (in Section~\ref{sec:good-tree}) and Lemma~\ref{lem:no-bad-two-cores}
(in Section~\ref{sec:no-bad-two-cores}). We first provide some more results that will be used in the proofs of these lemmas.
Let $\nn$ and $\mm$ denote the number of vertices and edges in $\K$, respectively. Since in the proof of Lemma~\ref{lem:no-bad-two-cores} we work with fixed degree sequences, we wish to characterise a set of sequences which contains the degree sequences typical for~$\K$, and is in particular compliant with the typical asymptotic values of~$\nn$ and~$\mm$, as well as the typical degree histogram of~$\K$, stated in Section~\ref{sec:new}. Note that the following definition technically and tacitly applies to a fixed sequence of degree sequences $\dd$, one for each $n$, since it describes an asymptotic property of the degree sequence.
\begin{defn*}
A degree sequence $\dd \in \D_{\nn,2\mm}$ is \emph{proper} (with respect to the underlying parameters $n$,$k$ and $c$) if~$\nn$ and~$\mm$ satisfy~\eqref{eq:number-of-vertices-in-k-core} and~\eqref{eq:number-of-edges-in-k-core} respectively, $\Delta(\dd) \le \log n$,
$\sum [d_i]_2 = \Theta(n)$, and the degree distribution follows the asymptotics in~\eqref{degree-asy}. In this case, we have
for each fixed integer $j\ge 0$ that
\begin{equation*}
\left|\{i \mid d_i = j\}\right| = \begin{cases}
0, & j < k;\\
\left(1+o\left(1\right)\right)\Psi_{j}\left(\mu_c\right)n & j\ge k.
\end{cases}
\end{equation*}
\end{defn*}
Given the degree sequence of $\K$, we perform some of the analysis in the proof of Lemma~\ref{lem:no-bad-two-cores} by using the configuration model. In order to make the results we obtain via this model applicable, we need the following immediate corollary of Theorem~\ref{thm:prob-simple}.
\begin{cor}\label{cor:prob-proper}
Let $f(n)$ be any function satisfying $f(n) \to \infty$. Then
\[
\Pr_{G \sim \mathcal{G}(n,\dd) }\left[G\in \P\right] \le f(n)\cdot\Pr_{G^* \sim \mathcal{G}^*(n,\dd)}\left[G^* \in \P\right]
\]
for any proper degree sequence $\dd \in \D_{\nn,2\mm}$ and any multigraph property $\P$.
\end{cor}
To show that we may restrict to proper degree sequences, and other kinds to be defined below, we first show that the upper tail of the sum of squared degrees is negligible in $\mathcal{G}(n,p)$, as follows. \begin{lem}\label{l:second_moment_degrees}
Let $(D_0, D_1,\ldots )$ be the degree histogram of $G\in\gnc$. Then WHP
\begin{description}
\item{(a)}
$D_j=0$ for all $j\ge \log n$;
\item{(b)} for every $\epsilon>0$ there exists an integer $j_0$ such that $ \sum_{j\ge j_0} [j]_2D_j/n <\epsilon$.
\end{description}
\end{lem}
\begin{proof} Part (a) is well-known and follows from the third case in~(\ref{eq:degree-distribution-gnc}). Part (b) also follows by standard methods, for instance as follows. First, note that we may assume that (a) holds and hence restrict the summation to $j<\log n$. Standard computations show for such $j$ that $\E [D_j]<c^jn/j!$ and the variance of $D_j$ is $O(n \log^2 n)$. Chebyshev's inequality, together with the union bound, then implies that WHP $D_j<\E [D_j]+n^{3/4}$ for all $j<\log n$. The result now follows, given the above bound on $\E [D_j]$.
\end{proof}
Note that if $G$ satisfies Part~$(b)$ of the lemma, then $\sum [d_i]_2 = O(n)$.
Indeed, let $j_0$ such that $ \sum_{j\ge j_0} [j]_2D_j/n < 1$. Then all vertices of $G$ of degree at least $j_0$ contribute at most $n$ to $\sum [d_i]_2$, while all other vertices contribute at most $[j_0]_2n$.
Since the bounds on degree counts of $G\in\gnc$ are also bounds for its core $\K$, an immediate consequence of Lemmas~\ref{CWnm},~\ref{CWdeg} and~\ref{l:second_moment_degrees} is the following.
\begin{cor}\label{l:proper_core}
The degree sequence of the core $\K$ is WHP proper.
\end{cor}
\medskip
The following claim, related to moments of the Poisson distribution,
is used in the proof of Theorem~\ref{thm:shatter-time}.
\begin{claim}
\label{clm:poisson-convolution}For real numbers $\mu\ge\lambda\ge0$
and integers $k\ge\ell\ge0$ we have
\[
\sum_{j=0}^{\infty}\left[j\right]_{\ell}\Psi_{j}\left(\lambda\right)\Psi_{\ge k-j}\left(\mu-\lambda\right)=\lambda^{\ell}\Psi_{\ge k-\ell}\left(\mu\right).
\]
\end{claim}
\begin{proof}
First we prove the claim for $\ell = 0$, that is
\begin{equation}\label{eq:poisson-convolution-0}
\sum_{j=0}^{\infty}\Psi_{j}\left(\lambda\right)\Psi_{\ge k-j}\left(\mu-\lambda\right)=\Psi_{\ge k}\left(\mu\right).
\end{equation}
Let $X\sim \Poisson(\lambda)$ and $Y\sim \Poisson(\mu-\lambda)$ be independent. Then $X+Y\sim\Poisson(\mu)$ and thus
\begin{align*}
\sum_{j=0}^{\infty}\Psi_{j}\left(\lambda\right)\Psi_{\ge k-j}\left(\mu-\lambda\right)
&= \sum_{j=0}^{\infty} \Pr[X=j]\Pr[Y\ge k-j]
= \sum_{j=0}^{\infty} \Pr[X=j \wedge Y\ge k-j] \\
&= \Pr[X+Y\ge k]
= \Psi_{\ge k}\left(\mu\right).
\end{align*}
Now, using the fact that $\left[j\right]_{\ell}=0$ for every $0\le j<\ell$ we get that the claim holds for all $\ell$:
\begin{align*}
\sum_{j=0}^{\infty}\left[j\right]_{\ell}\Psi_{j}\left(\lambda\right)\Psi_{\ge k-j}\left(\mu-\lambda\right)
& = \sum_{j=\ell}^{\infty}\left[j\right]_{\ell}\Psi_{j}\left(\lambda\right)\Psi_{\ge k-j}\left(\mu-\lambda\right)\\
& \stackrel{\eqref{eq:poisson-obs}}{=} \lambda^{\ell}\sum_{j=\ell}^{\infty}\Psi_{j-\ell}\left(\lambda\right)\Psi_{\ge k-j}\left(\mu-\lambda\right)\\
& = \lambda^{\ell}\sum_{j=0}^{\infty}\Psi_{j}\left(\lambda\right)\Psi_{\ge (k-\ell)-j}\left(\mu-\lambda\right)\\
& \stackrel{\eqref{eq:poisson-convolution-0}}{=} \lambda^{\ell}\Psi_{\ge k-\ell}\left(\mu\right).\qedhere
\end{align*}
\end{proof}
\medskip
The following lemma, bounding from above the probability of a truncated Poisson random variable achieving its minimum, is used in the proofs of Lemmas~\ref{lem:good-tree} and~\ref{lem:no-bad-two-cores}.
\begin{lem}
\label{lem:technical} For $k\ge3$ and $c>c_{k}$ let
\[
\delta=\delta\left(k,c\right)=\frac{1}{2}\left(1-\frac{c_{k}}{c}\right)\frac{k-2}{k-1}.
\]
Then
\[
\Pr\left[Z_{k-1}\left(\mu_{c}\right)=k-1\right]<\frac{1-2\delta}{k-1}.
\]
\end{lem}
\begin{proof}
Let
\[
F(\mu):=\frac{\Psi_{\ge k-1}(\mu)}{\Psi_{k-1}(\mu)}=\frac{1}{\Pr\left[Z_{k-1}(\mu)=k-1\right]},
\]
so we need to show that $k-1\le(1-2\delta)F(\mu_{c})$.
Let $h(\mu)=\mu/\Psi_{\ge k-1}(\mu)$, defined for all $\mu>0$. Recall
that $\beta(c)$ was determined to be the maximum solution of $x=\Psi_{\ge k-1}(cx)$,
which enables us to express $c_{k}$ as in~(\ref{eq:ck-definition}).
In terms of~$h$, we can define~$\mu_{c}$ as the maximum solution
of $h(\mu)=c$, which exists if and only if $\beta(c)>0$. We can
therefore express~$c_{k}$ again in the following way:
\[
c_{k}=\inf\{c\mid\exists\mu>0~h(\mu)=c\}=\inf\{h(\mu)\mid\mu>0\}.
\]
Clearly~$h$ is differentiable and its derivative is
\begin{align}
h'(\mu) & =\frac{1}{\Psi_{\ge k-1}(\mu)}\left(1-\mu\frac{(\Psi_{\ge k-1})'(\mu)}{\Psi_{\ge k-1}(\mu)}\right)\nonumber \\
& =\frac{1}{\Psi_{\ge k-1}(\mu)}\left(1-\mu\frac{\Psi_{k-2}(\mu)}{\Psi_{\ge k-1}(\mu)}\right)\nonumber \\
& =\frac{1}{\Psi_{\ge k-1}(\mu)}\left(1-(k-1)\frac{\Psi_{k-1}(\mu)}{\Psi_{\ge k-1}(\mu)}\right)\nonumber \\
& =\frac{1}{\Psi_{\ge k-1}(\mu)}\left(1-\frac{k-1}{F(\mu)}\right)\nonumber \\
& <\frac{1}{\Psi_{\ge k-1}(\mu)}.\label{eq:deriv}
\end{align}
Note that $h'(\mu)$ is positive if and only if $F(\mu)>k-1$, and
since~$F$ is an increasing function approaching~$1^{+}$ and~$\infty$
as~$\mu$ approaches~$0^{+}$ and~$\infty$, respectively, the
infimum~$c_{k}$ of~$h$ is actually its minimum, attained at a
unique point~$\mu_{c_{k}}$. In particular,~$h$ is increasing for
$\mu>\mu_{c_{k}}$ and since $h'(\mu_{c_{k}})=0$ we have
\begin{equation}
F\left(\mu_{c_{k}}\right)=k-1.\label{eq:zeroDeriv}
\end{equation}
By the mean value theorem (applied to~$h$) there exists some $\tilde{\mu}\in(\mu_{c_{k}},\mu_{c})$
such that
\begin{align*}
\frac{c-c_{k}}{\mu_{c}-\mu_{c_{k}}} & =\frac{h(\mu_{c})-h(\mu_{c_{k}})}{\mu_{c}-\mu_{c_{k}}}=h'(\tilde{\mu})\\
& \stackrel{\eqref{eq:deriv}}{<}\frac{1}{\Psi_{\ge k-1}(\tilde{\mu})}<\frac{1}{\Psi_{\ge k-1}(\mu_{c_{k}})}=\frac{h(\mu_{c_{k}})}{\mu_{c_{k}}}=\frac{c_{k}}{\mu_{c_{k}}},
\end{align*}
where the second inequality is due to the monotonicity of $\Psi_{\ge k-1}$
in~$\mu$. Rearranging, we get
\begin{equation}
\mu_{c_{k}}<\left(c_{k}/c\right)\mu_{c}.\label{eq:fact1}
\end{equation}
Recall that~$F$ is an increasing function of~$\mu$, so
\begin{align}
1-2\delta & =1-\left(1-\frac{c_{k}}{c}\right)\left(1-\frac{1}{k-1}\right)\nonumber \\
& \stackrel{\eqref{eq:zeroDeriv}}{=}1-\left(1-\frac{c_{k}}{c}\right)\left(1-\frac{1}{F\left(\mu_{c_{k}}\right)}\right)\nonumber \\
& >1-\left(1-\frac{c_{k}}{c}\right)\left(1-\frac{1}{F\left(\mu_{c}\right)}\right)\nonumber \\
& =\frac{1+\left(c_{k}/c\right)\left(F\left(\mu_{c}\right)-1\right)}{F\left(\mu_{c}\right)}.\label{eq:fact2}
\end{align}
Finally, observe that for every $\mu>0$
\begin{align*}
F(\mu)-1 & =\frac{\Psi_{\ge k-1}(\mu)-\Psi_{k-1}(\mu)}{\Psi_{k-1}(\mu)}=\frac{\Psi_{\ge k}(\mu)}{\Psi_{k-1}(\mu)}\\
& =\sum_{j=k}^{\infty}e^{-\mu}\frac{\mu^{j}}{j!}\Big/\left[e^{-\mu}\frac{\mu^{k-1}}{(k-1)!}\right]\\
& =(k-1)!\mu^{1-k}\sum_{j=1}^{\infty}\frac{\mu^{j+k-1}}{(j+k-1)!}\\
& =\sum_{j=1}^{\infty}\frac{\mu^{j}}{\left[j+k-1\right]_{j}},
\end{align*}
and thus $F(\alpha\mu)-1<\alpha(F(\mu)-1)$ for every $0<\alpha<1$,
implying
\[
k-1\stackrel{\eqref{eq:zeroDeriv}}{=}F\left(\mu_{c_{k}}\right)\stackrel{\eqref{eq:fact1}}{<}F\left(\left(c_{k}/c\right)\mu_{c}\right)<1+\left(c_{k}/c\right)\left(F\left(\mu_{c}\right)-1\right)\stackrel{\eqref{eq:fact2}}{<}\left(1-2\delta\right)F\left(\mu_{c}\right),
\]
which establishes the lemma.
\end{proof}
\subsection{\label{sec:shatter-proof}Proof of Theorem~\ref{thm:shatter-time}}
Recall that degree histograms were defined in Section~\ref{sec:branching-gnp}. An \emph{asymptotic degree histogram} is a sequence $D=\left(D_{0},D_{1},\ldots\right)$
of functions $D_{j}:\mathbb{N}\to\mathbb{N}$ such that $\sum_{j=0}^{\infty}D_{j}\left(n\right)=n$
and $\sum_{j=0}^{\infty}jD_{j}\left(n\right)$ is even for all $n$.
For a given asymptotic degree histogram $D$, denote by $\Omega\left(n,D\right)$ the set of all simple graphs on $n$ vertices with degree histogram $\left(D_0(n),D_1(n),\ldots\right)$.
If $\Omega\left(n,D\right)\neq\varnothing$ for all $n\ge1$, $D$ is \emph{feasible}; in this case, let $\mathcal{G}\left(n,D\right)$ be the uniform distribution over $\Omega\left(n,D\right)$.
A feasible asymptotic degree histogram $D$ is \emph{sparse}
if $\sum_{j=0}^{\infty}jD_{j}\left(n\right)/n=\kappa_{D}+o\left(1\right)$
for some constant~$\kappa_{D}$, called the asymptotic edge density
of $\mathcal{G}\left(n,D\right)$; $D$ is \emph{well-behaved} if:
\begin{enumerate}
\item There exist constants $\left\{ \delta_{j}\right\} _{j=0}^{\infty}$
such that $\lim_{n\to\infty}D_{j}\left(n\right)/n=\delta_{j}$ for
all fixed $j\ge0$.
\item $\left\{ j\left(j-2\right)D_{j}\left(n\right)/n\right\} _{j=0}^{\infty}$
tends uniformly to $\left\{ j\left(j-2\right)\delta_{j}\right\} _{j=0}^{\infty}$.
\item $\lim_{n\to\infty}\sum_{j=0}^{\infty}j\left(j-2\right)D_{j}\left(n\right)/n$
exists, and the sum uniformly approaches the limit $$Q_{D}:=\sum_{j=0}^{\infty}j\left(j-2\right)\delta_{j}.$$
\end{enumerate}
Molloy and Reed~\cite{MR-95} showed that the sign of $Q_{D}$ WHP determines the
existence of a giant component in $\mathcal{G}\left(n,D\right)$:
\begin{lem}[{\cite[Theorem~1]{MR-95}}]
\label{lem:molloy-reed}Let $D$ be a feasible well-behaved sparse
asymptotic degree histogram, and let $\Delta\left(n\right)=\max\left\{ j\in\mathbb{N}\mid D_{j}\left(n\right)>0\right\}$.
\begin{enumerate}
\item If $\Delta\left(n\right)=o\left(n^{1/4}\right)$ and $Q_{D}>0$ then
WHP $\mathcal{G}\left(n,D\right)$ has a linear-size connected component;
\item If $\Delta\left(n\right)=o\left(n^{1/8}\right)$ and $Q_{D}<0$ then
the size of the largest connected component in $\mathcal{G}\left(n,D\right)$
is WHP $O\left(\Delta^{2}\left(n\right)\log n\right)$.
\end{enumerate}
\end{lem}
Having computed the likely degree histogram $D^{t}$ of $G_{t}$
in Claim~\ref{clm:degree-distribution-after-t-steps}, we are ready
to prove Theorem~\ref{thm:shatter-time}.
\begin{proof}[Proof of Theorem~\ref{thm:shatter-time}]
We can view $G_{t}$ as drawn from $\mathcal{G}\left(n,D^{t}\right)$,
since an iteration of the peeling process can be carried out as follows:
first expose the set of vertices $V_{t-1}=\left\{ v\mid\rho\left(v\right)=t-1\right\} $,
then expose their degrees in $G_{t-1}$; finally expose and delete
all edges incident with $V_{t-1}$. Given the degree histogram $D^{t}$,
all edges inside $G_{t}$ remain unexposed.
Implicitly, $G_{t}$ can be regarded as a sequence of random graphs, one for each $n$, and their degree histograms $D^t$ determine an asymptotic degree histogram which we denote by $(D^t)$. Although results like Claim~\ref{clm:degree-distribution-after-t-steps} only describe events that hold WHP, they can easily be converted to statements about an asymptotic degree histogram such that the events hold for all $n$, and the asymptotic degree histogram WHP coincides with the random graph. (This can be done by altering the histogram on those values of $n$ which violate the required properties of being well behaved.) When we make statements about $(D^t)$ in the following, we assume these slight adjustments are made automatically.
Fix $t\ge 0$. By definition $D^{t}$ is feasible and it is easy to verify that $D^{t}$
is WHP well-behaved using Claim~\ref{clm:degree-distribution-after-t-steps} for bounded degrees, together with Lemma~\ref{l:second_moment_degrees} for the unbounded degrees. (Later we will choose a particular value of the constant $t$.)
Applying Claim~\ref{clm:poisson-convolution} with $\lambda=c\beta_{t}$, $\mu=c\beta_{t-1}$ and $\ell=1$, we get that $D^{t}$ is also sparse, with asymptotic edge density
\[
\kappa_{t}:=\kappa_{D^{t}}=\sum_{j=0}^{\infty}j\delta_{j}^{t}=\sum_{j=0}^{\infty}j\Psi_{j}\left(c\beta_{t}\right)\Psi_{\ge k-j}\left(c\beta_{t-1}-c\beta_{t}\right)=c\beta_{t}\Psi_{\ge k-1}\left(c\beta_{t-1}\right)=c\beta_{t}^{2}.
\]
We now bound the parameter $Q_t := Q_{D^t}$, by another application of Claim~\ref{clm:poisson-convolution} with $\lambda=c\beta_{t}$ and $\mu=c\beta_{t-1}$, but this time with $\ell=2$. We get
\[
\sum_{j=0}^{\infty}\left[j\right]_2\delta_j^t
=\sum_{j=0}^{\infty}\left[j\right]_2\Psi_j\left(c\beta_t\right)\Psi_{\ge k-j}\left(c\beta_{t-1}-c\beta_t\right)
=\left(c\beta_t\right)^2\Psi_{\ge k-2}\left(c\beta_{t-1}\right)
=c\kappa_t\Psi_{\ge k-2}\left(c\beta_{t-1}\right),
\]
hence
\begin{align*}
Q_{t} & =\sum_{j=0}^{\infty}j\left(j-2\right)\delta_{j}^{t}=\sum_{j=0}^{\infty}\left(\left[j\right]_{2}-j\right)\delta_{j}^{t}\\
& =c\kappa_{t}\Psi_{\ge k-2}\left(c\beta_{t-1}\right)-\kappa_{t}\\
& =\left(c\Psi_{\ge k-2}\left(c\beta_{t-1}\right)-1\right)\kappa_{t}.
\end{align*}
The decreasing sequence $\left(\beta_{t}\right)_{t\ge0}$ converges
to $\beta=0$ in the subcritical regime, and $x\mapsto c\Psi_{\ge k-2}\left(cx\right)$
is a continuous function, so
\[
\lim_{t\to\infty}c\Psi_{\ge k-2}\left(c\beta_{t}\right)=c\Psi_{\ge k-2}\left(c\beta\right)=0.
\]
In particular, there exists some constant $t^{\dagger}$ such that
$c\Psi_{\ge k}\left(c\beta_{t^{\dagger}}\right)<1$, implying $Q_{t^{\dagger}}<0$.
Moreover, $\Delta\left(G_{t^{\dagger}}\right)\le\Delta\left(G_{0}\right)=O\left(\log n/\log\log n\right)$
by~(\ref{eq:degree-distribution-gnc}), and we can finally apply
the second part of Lemma~\ref{lem:molloy-reed} and complete the
proof.
\end{proof}
\begin{rem*}
By~\cite[Lemma~6]{JMT-16} we have $t^{\dagger}=\Theta\left(1/\sqrt{c_{k}-c}\right)$.
\end{rem*}
\subsection{\label{sec:good-tree}Proof of Lemma~\ref{lem:good-tree}: the existence
of a $\left(\log^{2}\log n,L\right)$-tree in $\K$}
We prove Lemma~\ref{lem:good-tree} for $L:=\left\lceil \log_{1-\delta}\left(\delta^{2}/(2\mu_{c})\right)\right\rceil $, where $\delta < 1/2$ is the constant from Lemma~\ref{lem:technical}. Throughout this subsection we let $N = \left\lceil \log^{2}\log n\right\rceil$ and $p_j = \Pr[Z_{k-1}(\mu_{c})=j-1]$.
In addition, we set $C = (1-\delta)^2/(1-2\delta)>1$, and let $d_0$ be the minimal integer satisfying $1-\sum_{i \le d_0} p_i/C \le \delta^2/(2\mu_{c})$. Note that $L,C$ and $d_0$ are all constants depending only on $k$ and $c$.
\medskip
We consider an exploration process in the $k$-core $\K$, attempting to reveal an $(N,L)$-tree in it, but instead of analysing the exploration process on $\K$ itself, we condition on it having a proper degree sequence $\dd$ and apply the exploration process to the configuration model for the sequence~$\dd$. In view of Lemma~\ref{CWnm} and Corollary~\ref{l:proper_core}, Theorem~\ref{thm:prob-simple} implies that it is enough to show that the multigraph of this configuration model WHP contains an $(N,L)$-tree, where the convergence implicit in WHP is uniform over all proper degree sequences $\dd$.
The exploration starts with an arbitrary vertex $v_0$ in this configuration model, and explores its $(2N)$-neighbourhood in DFS manner. In each exploration step, an unmatched half-edge, say $x$, belonging to an exposed vertex at distance at most $2N-1$ from $v_0$, is matched to some other half-edge, say $y$, chosen uniformly at random from the set of all unmatched half-edges. We refer to the vertex containing $y$ as the \emph{next encountered vertex}. The selection of $x$ in each step is arbitrary among those in the vertex currently being treated by the DFS algorithm. Initially, $v_0$ is the only exposed vertex, and whenever a new half-edge is being matched, its vertex (i.e.\ the next encountered vertex) becomes exposed. Unless that vertex was already exposed, the new edge and vertex are added to the growing DFS tree.
Let $T$ denote the tree resulting from the exploration described above. The root of T is $v_0$, and all other vertices of $T$ have distance at most $2N$ from $v_0$ in $T$.
We assign each vertex in $T$
a \emph{type} from $\left\{ 0,1,\ldots,L+1\right\} $ in the following
manner. The type assignment for a vertex $u$ is performed at the point when the DFS algorithm has finished fully exploring the subtree of $T$ below $u$ and looks to move back to the parent of $u$ (or terminate, if $u = v_0$). First, if any back-edge has been encountered up to this point in the exploration process, $u$ is assigned type $L+1$. If no back-edge has been encountered, the following rules are applied.
If $u$ is a leaf, i.e. in level $2N$, its type is set to $0$.
Otherwise, $u$ is in level $i<2N-1$, and all its children have been assigned types already; denote by $S(u)$ the set of its children of type less than $L$.
Note that $d_T(u) = d_\K(u)$ in this case, so we can omit the subscript, and set
\[
\type\left(u\right)=\begin{cases}
0, & \left|S\left(u\right)\right|\ge k\quad\ \ \mbox{ and } d(u)\le d_0;\\
1+\max\left\{ \type\left(v\right)\mid v\in S\left(u\right)\right\} , & \left|S\left(u\right)\right|=k-1\mbox{ and } d(u)\le d_0;\\
L, & \left|S\left(u\right)\right|<k-1\ \ \, \mbox{ or } d(u)>d_0.
\end{cases}
\]
For $v\in V(T)$, let $T(v)$ denote the subtree of $T$ consisting of $v$ and all its descendants. Given the types of vertices as defined above, let $T^*(v)$ denote the result of removing from $T(v) $ all subtrees
rooted at vertices of type $L$ or $L+1$. Then for every vertex $u$ in
$T^*(v)$, the number of vertices in a longest $k$-light tree path originating at $u$ is exactly $\type(u) <L$. If $\type(u) = 0$ it simply means that no such paths exist as $u$ itself is not $k$-light. In particular, we have the following.
\begin{obs}\label{obs:type}
Let $v\in V(T)$ at level $i$. If $\type(v)<L$ then $T^*(v)$ is a $(2N-i,L)$-tree.
\end{obs}
Let us now take a closer look at the process of matching half-edges. When the first random half-edge is chosen, the probability that its vertex $u$ has degree $j$ is weighted by a multiplicative factor of $j$ (so-called ``degree-biased'' selection). Hence, for any fixed $j \ge k$ and a proper degree sequence, we have by~\eqref{degree-asy} that the probability that $u$ has degree $j$ is
\begin{equation}\label{eq:deg-biased}
(1+o(1))j\Pr[Z_{k}\left(\mu_{c}\right)=j]\nn/2\mm \stackrel{\eqref{eq:average-degree-in-k-core}}{=} (1+o(1))p_j.
\end{equation}
As the exploration carries on, the degree sequence of the unmatched half-edges does not represent the degrees of the vertices any more, but their ``remaining'' degrees, that is, the number of unmatched half-edges incident with each vertex at that point. Of course, this distinction only applies to the exposed vertices. Additionally, this degree sequence contains values smaller than $k$. To handle these subtleties, we define a new class of sequences, closely related to proper sequences.
For a constant $\eta > 1$, a degree sequence $\dd \in \D_{\nn,2\mm}$ is \emph{$\eta$-normal} (with respect to $n,k,c$ and $d_0$), if $n \ge \nn \ge \Psi_{\ge k}\left(\mu_{c}\right)n/2$ and $\mm \ge \mu_{c}\Psi_{\ge k-1}\left(\mu_{c}\right)n/4$, its maximum degree is at most $\log n$, and for every $j \le d_0$ the degree-biased probability that a half-edge selected uniformly at random belongs to a vertex of (remaining) degree $j$ is between $p_j/\eta$ and $\eta p_j$. Note that in particular, for every $j \le d_0$, the number of `$j$' entries in any $\eta$-normal sequence is at least $2\mm p_j/(j\eta) = \Omega(n)$. It is immediate to see that for any fixed $\eta > 1$, every proper sequence is $\eta$-normal for $n$ sufficiently large.
We are finally ready for the main part of the proof.
Recall the definition of $C$ from the beginning of this subsection, and let $C' = (C+1)/2 > 1$.
\begin{claim}\label{clm:pruning}
Assume that $\dd$ is a $C'$-normal sequence and consider any moment during the exploration process when a random half-edge is about to be chosen. Conditional upon the exploration so far, the probability that the next encountered vertex will eventually have type $L$ is at most $\delta/\mu_{c}$.
\end{claim}
We first show why Claim~\ref{clm:pruning} implies Lemma~\ref{lem:good-tree}.
Assume $\dd$ is $C'$-normal and consider any moment at which an unmatched half-edge incident with a vertex at level $N-1$ in $T$ is about to be treated, i.e., the next encountered vertex will belong to level $N$ unless already exposed. At this point the types of all vertices so far encountered at level $N$ have been assigned. We may therefore apply
Claim~\ref{clm:pruning}
to deduce that, conditional on the labels of the previously encountered vertices in level $N$, the probability that the next one receives type $L$ is at most $ \delta/\mu_c $. By coupling this process with a sequence of independent Bernoulli trials each with parameter $\delta/\mu_c$, we conclude that, for any $t>0$, the probability that at least $t$ vertices are encountered at level $N$ and all are given type $L$ is at most $(\delta/\mu_c)^t$.
Since $\dd$ is $C'$-normal, the degrees are all at most $\log n$, and so the number of vertices reached in the exploration is at most $(\log n)^{2N}= \exp\big(O(\log^{3}\log n)\big)=o(n^{1/3})$. Consequently, since there are $\Theta(n)$ half-edges altogether, each step of the exploration process chooses an exposed vertex with probability $o(n^{-2/3}\log n)$. Thus, the probability that at least one of the $o(n^{1/3}\log n)$ steps encounters a back-edge is $o(1)$. Since all vertices of $\K$ have degree at least $k$, it follows that WHP, $T$ has
at least $ (k-1)^{N} = \omega(1)$ vertices at level $N$. From the previous paragraph, the probability that these are all assigned type $L$ is $o(1)$. As there are WHP no back-edges, this implies that WHP some vertex $v$ at level $N$ receives a type less than $L$. By Observation~\ref{obs:type}, this event implies that $T^*(v)$ is an $(N,L)$-tree.
One can easily check that the convergence in the above WHP statements is uniform over all $C'$-normal degree sequences $\dd$. Since WHP $\K$ has a $C'$-normal degree sequence, Lemma~\ref{lem:good-tree} follows, and it only remains to prove the claim.
\begin{proof}[Proof of Claim~\ref{clm:pruning}]
For the given degree sequence $\dd = (d_1,\dots,d_{\nn}) \in \D_{\nn,2\mm}$,
let $R(\dd)$ be the set of all sequences $\dd' = (d_1',\dots,d_{\nn}') \in \D_{\nn,2\mm'}$, such that $d_i \ge d_i'$ for all $i$, and $m-m' \le n^{1/3}\log n$. By the arguments above, at every step of the exploration, the degree sequence of the unmatched half-edges in $\K$ is some element of $R(\dd)$.
Now let $\dd' \in R(\dd)$, and for every $j \le d_0$, let $n_j$ and $n_j'$ denote the number of vertices of degree $j$ in $\dd$ and in $\dd'$, respectively. Since $n_j = \Theta(n)$, and since $d_i \neq d_i'$ for $O(n^{1/3}\log n)$ coordinates, we have $n_j' = (1+o(1))n_j$. Similarly, $m' = (1+o(1))m$, and thus $jn_j'/(2m') = (1+o(1))jn_j/(2m)$. In short, for every $\dd' \in R(\dd)$ and $j \le d_0$, the probability of choosing a half-edge belonging to a vertex of (remaining) degree $j$, asymptotically equals the probability of the same event for $\dd$. Since, in addition, the probability that a randomly selected half-edge belongs to an already exposed vertex is $o(1)$, and since $C > C'$, we can state the following.
\begin{obs}\label{obs:Cbound}
At every step of the exploration, conditioning upon the exploration steps taken previously, the probability that the next encountered vertex is unexposed and has degree~$j \le d_0$ is bounded between $p_j/C$ and $Cp_j$.
\end{obs}
To complement this observation, consider any moment during the exploration, let $S$ denote the exploration sequence up to that point, and let $p_{>d_0} (S)$ denote the probability that the next encountered vertex will have degree larger than $d_0$, conditional on $S$. Let $p_{>d_0}$ denote the maximum of $p_{>d_0} (S)$, taken over all possible (partial) exploration sequences $S$. Then by Observation~\ref{obs:Cbound} and the definition of $d_0$ (at the beginning of this subsection), we have ${p_{>d_0} \le \delta^2/(2\mu_{c})}$.
\medskip
Before proceeding with the proof of the claim we introduce more terminology. We say a vertex $v$ of $T$ has height $h$ if it is in level $2N-h$, i.e.\ $T(v)$ has height $h$.
The height of an unmatched half-edge belonging to an exposed vertex is the same as the height of that vertex.
Similarly to the definition of $p_{>d_0}$, for every $1 \le h \le 2N $ and $0\le i\le L$, let $P_{i,h}$ denote the maximum, over all possible exploration sequences $S$ up to any step in which a half-edge at height $h$ is being matched, of the probability that the next encountered vertex will be assigned type $i$, conditional on $S$. Note that this next encountered
vertex is at height $h-1$ unless it was already exposed.
We prove the claim by showing that the following hold for every $1 \le h \le 2N $:
\begin{enumerate}
\item[$(a)$] $P_{L,h}\le\delta/\mu_{c}$;
\item[$(b)$] $P_{i,h}\le(1-\delta)^{i}$ for every $0\le i<L$.
\end{enumerate}
Recall that if a back-edge occurs before assigning the type of a vertex $u$, it is given type $L+1$, and that otherwise, and if $u$ is also not a leaf, then $d_T(u) = d_\K(u)$ and we simply refer to the degree of $u$ with no specification.
Observe that $(b)$ holds trivially for $i=0$ (for every $h$), so from now on we assume for simplicity $i>0$. Since the only positive type a leaf can be assigned is $L+1$, there is nothing to prove for $h = 1$.
We now prove $(a)$ and $(b)$ for $h > 1$ by induction on $h$, beginning with $h = 2$.
\medskip
Assume that no back-edge has yet occurred when a vertex $u$ at height 1 is assigned its type. Then every child of $u$ is a leaf of type 0, and there are therefore exactly three options. If $d(u) > d_0$ then $\type(u) = L$; if $d(u) = k$, i.e., $u$ has $k-1$ children in $T$, then $\type(u) = 1$; otherwise, $\type(u) = 0$.
So when a half-edge at height 2 is being matched, the probability that the next encountered vertex will eventually have type 1 or $L$ is bounded from above by the probability that the vertex will have degree $k$ or larger than $d_0$, respectively.
We therefore immediately obtain $(a)$ since
\[
P_{L,2} \le p_{>d_0} \le \delta^2/(2\mu_{c}) \le \delta/\mu_{c}.
\]
As for $(b)$, we only have to show that $P_{1,2} \le 1-\delta $, and this is true since by the above argument, Observation~\ref{obs:Cbound} and Lemma~\ref{lem:technical} we have
\[
P_{1,2}
\le Cp_k
<C\frac{1-2\delta}{k-1}
=\frac{(1-\delta)^2}{k-1}
<1-\delta.
\]
Assume now that $(a)$ and $(b)$ hold for $1 < h - 1 < 2N$; we prove $(a)$ and $(b)$ for $h$.
\medskip
Observe that if a vertex $v$ at height~$h -1$ with degree~$j$ is assigned type
$0<i<L$ (implying in particular that $j \le d_0$ by definition of type), then no back-edge has yet been encountered, exactly $j-k$ of $v$'s $j-1$ children must be assigned type $L$,
and the maximum type among its other $k-1$ children must be exactly $i-1$.
Now consider any moment when a half-edge at height $h$ is matched, and let $w$ denote the next encountered vertex. By Observation~\ref{obs:Cbound}, the probability that $w$ will be an unexposed vertex with degree $j$ is at most $Cp_j$.
Since the exploration is DFS, the bounds $P_{i,h-1}$ (for any $0 \le i \le L$) can be applied to each of the children of $w$ consecutively and conditionally. Thus at any given step in which a half-edge at height~$h$ is being matched, the probability that the next encountered vertex
will have degree $j$ and type $0<i<L$, conditional on the previous history of the process, can be bounded from above by
\[
f_{i,h}(j) := Cp_{j}\binom{j-1}{j-k}(P_{L,h-1})^{j-k}(k-1)P_{i-1,h-1}.
\]
In the same manner, consider the event that a vertex $v$ at height $h-1$ with degree $j \le d_0$ is assigned
type $L$. (Recall that a vertex will also be assigned type $L$ if it has degree larger than $d_0$, which happens with probability at most $p_{> d_0}$.) Then no back-edge has yet occurred, and either exactly $j-k$ of $v$'s children are of
type $L$ and there is at least one child of type $L-1$, or there
are at least $j-k+1$ children of $v$ of type $L$. So at any moment in which a half-edge at height~$h$ is being matched, we can bound
from above the probability that the next encountered vertex will have degree $j \le d_0$ and be assigned type $L$, conditional on the previous history of the process, by
\[
f_{h}(j) := Cp_{j}\binom{j-1}{j-k}(P_{L,h-1})^{j-k}(k-1)(P_{L,h-1}+P_{L-1,h-1}).
\]
\medskip
Next, we observe that for every $0<i<L$ and for every $k \le j < d_0$ the following holds:
\begin{equation}\label{eq:quotient}
\frac{f_{i,h}(j+1)}{f_{i,h}(j)}
=\frac{f_{h}(j+1)}{f_{h}(j)}
=\frac{p_{j+1}}{p_{j}}\cdot \frac{\binom{j}{j-k+1}}{\binom{j-1}{j-k}}\cdot P_{L,h-1}
=\frac{\mu_{c}}{j}\cdot\frac{j}{j-k+1}\cdot P_{L,h-1}
\le\delta,
\end{equation}
where the last inequality follows from the induction hypothesis.
We now use Lemma~\ref{lem:technical} and the induction hypothesis to get
\[
f_{i,h}(k)
=Cp_{k}(k-1)P_{i-1,h-1}
< C(1-2\delta)\left(1-\delta\right)^{i-1}
=\left(1-\delta\right)^{i+1},
\]
which together with~(\ref{eq:quotient}) yields
\[
P_{i,h}
\le \sum_{j = k}^ {d_0} f_{i,h}(j)
\le \sum_{j = k}^ {d_0} f_{i,h}(k)\delta^{j-k}
\ <\ \frac{f_{i,h}(k)}{1-\delta}<\left(1-\delta\right)^i,
\]
establishing $(b)$.
To prove $\left(a\right)$, we similarly use
\[
f_{h}(k)=Cp_{k}(k-1)(P_{L,h-1}+P_{L-1,h-1})<\left(1-\delta\right)^2 \left(\frac{\delta}{\mu_{c}}+\left(1-\delta\right)^{L-1}\right)
\]
and~(\ref{eq:quotient}) to obtain
\begin{eqnarray*}
\sum_{j = k}^ {d_0}f_{h}(j)
& \le & \sum_{j = k}^ {d_0}f_{h}(k)\delta^{j-k}\ <\ \frac{f_{h}(k)}{1-\delta}\\
& < & \left(1-\delta\right)\frac{\delta}{\mu_{c}}+\left(1-\delta\right)^{L}
=\frac{\delta}{\mu_{c}}-\left(\frac{\delta^{2}}{\mu_{c}}-\left(1-\delta\right)^{L}\right)\le\frac{\delta}{\mu_{c}} - \frac{\delta^2}{2\mu_{c}},
\end{eqnarray*}
where the last inequality holds by the choice of $L$. We conclude that (a) holds by using the bound on $p_{> d_0}$ and the fact that
\[
P_{L,h} \le p_{> d_0} + \sum_{j = k}^ {d_0}f_{h}(j).\qedhere
\]
\end{proof}
\subsection{\label{sec:no-bad-two-cores}Proof of Lemma~\ref{lem:no-bad-two-cores}:
high excess, small boundary 2-cores are WHP linear}
Throughout this subsection we denote by $\nn$ and $\mm$ the number of vertices and edges, respectively, in~$\K$, the $k$-core of $G$, and denote by $\hd = 2\mm/\nn$ the average degree in~$\K$. For a subset $U \subset V(\K)$ we write $t=\left|V\left(U\right)\right|$, $s=\left|E\left(U\right)\right|$ and $r=\xs\left(U\right)=s-t$. We refer to any graph of minimum degree at least 2 as a\emph{ 2-core}. A 2-core $C\subset\mathcal{K}$ is \emph{bad} if $C$ has large excess
$\xs\left(C\right)\ge\log^{4}n$, small boundary $\left|\partial C\right|\le\left(k-2\right)V\left(C\right)$,
and small size $\left|V\left(C\right)\right|<\eps n$. Lemma~\ref{lem:no-bad-two-cores}
claims that WHP $\mathcal{K}$ has no bad 2-cores for some constant
$\eps=\eps\left(k,c\right)>0$.
We will consider bad 2-cores in $\K$ with exactly $t$ vertices and $s$ edges for pairs $(t,s) \in \mathcal I$, where
\[
\mathcal{I}=\left\{\left(t,s\right)\mid t<\eps n \text{~~and~~} t+\log^{4}n\le s\le\binom{t}{2}\right\}.
\]
Observe that in particular $\log^{4}n < \binom{t}{2}<t^{2}$, and so from now on we assume $t>\log^{2}n$.
\medskip
Recall the constant $\delta$ from Lemma~\ref{lem:technical}, and let $\delta_{1}<1/e$ be constant sufficiently small that $(1-\delta/2)\delta_{1}^{-4\delta_{1}}<1-\delta/4$. Such a $\delta_{1}$ exists since $x^{-x}\rightarrow1^{+}$ as $x\rightarrow0^{+}$.
In the proof we separate potential bad 2-cores into two classes: dense (i.e.,
with excess $r\ge\delta_{1}t$) and sparse (i.e., with excess $r\le\delta_{1}t$),
and partition $\mathcal{I}$ into $\mathcal{I}=\mathcal{I}_{\dns}\cup\mathcal{I}_{\sprs}$
accordingly.
We are now finally ready to define $\eps$. Since the discussion is
restricted to the $k$-core $\mathcal{K}$, which is WHP of linear
size $\hat{n} = (1+o(1))\Psi_{\ge k}(\mu_{c})n$, it will be convenient to define a constant
\[
\eps_{1}=\min\left\{(e^2c)^{-1-1/\delta_{1}},\frac{\delta}{2-\delta},\frac{1}{1+2e^{4}k^{6}}\right\}
\]
and set $\eps=\eps_{1}\Psi_{\ge k}\left(\mu_{c}\right)/2$.
\subsubsection{Dense 2-cores: $r\ge\delta_{1}t$}
We show that WHP, not only $\K$ does not contain dense bad 2-cores, but $G$ does not contain \emph{any} dense subgraphs of relevant size, without further restrictions. Clearly, it therefore suffices to only consider the case $r = \delta_1t$. So, for $\log^2n < t < \eps n$, let $N_t$ denote the expected number of subgraphs of $G$ with $t$ vertices and $(1+\delta_1)t$ edges. Then
\begin{align*}
N_t & \le \binom {n}{t} \binom {\binom t2}{(1+\delta_1)t} \left(\frac{c}{n}\right)^{(1+\delta_1)t} \\
& \le \left(\frac{en}{t}\right)^{t}\left(\frac{et^2/2}{(1+\delta_1)t}\right)^{(1+\delta_1)t} \left(\frac{c}{n}\right)^{(1+\delta_1)t} \\
& \le \left(\frac tn \right)^{\delta_1t} \left(\frac{e^2c}{2}\right)^{(1+\delta_1)t} \\
& \le \left(\eps\left(e^2c\right)^{1+1/\delta_1}\right)^{\delta_1t}2^{-t} \\
& \le 2^{-t},
\end{align*}
where the last inequality is due to the fact that $\eps<\eps_{1}$
and the definition of $\eps_{1}$. Summing over $t$, we get
$$\sum_{t=\log^2n}^{\eps n} N_t \le \sum_{t=\log^2n}^{\eps n} 2^{-t} \le n2^{-\log^2n} = o(1),$$
implying that WHP $G$ does not contain any subgraph with $t$ vertices and $s$ edges for any $(t,s)\in\cI_{\dns}$.
\subsubsection{Sparse 2-cores: $r\le\delta_{1}t$}
By using Corollary~\ref{l:proper_core}, we will be able to restrict to $k$-cores with proper degree sequences. Of course, ``the degree sequence is proper'' is not an event but an asymptotic statement. Strictly, what Corollary~\ref{l:proper_core} means is that there is a concrete specification of the asymptotic bounds in the definition of ``proper'' such that such bounds hold WHP for the degree sequence of the $k$-core. When we refer to the event that the sequence is proper below, we mean the event that a set of such bounds hold.
In view of Theorem~\ref{thm:maincor}, we consider a sequence $\dd$ of $\hat n$ independent copies of $Z_k(\lambda)$ where $\lambda$ is determined by~\eqref{lambdadef}. By Lemma~\ref{CWnm} we only need to consider $\nn = \left(1+o\left(1\right)\right)\Psi_{\ge k}\left(\mu_{c}\right)n$ and $\mm = \left(1+o\left(1\right)\right)\mu_{c}\Psi_{\ge k-1}\left(\mu_{c}\right)n/2$, and consequently, the estimation of $\hd$ given in~\eqref{eq:average-degree-in-k-core} holds. Since $\E \left[Z_k(\lambda)\right] = \lambda \Psi_{\ge k-1}(\lambda) / \Psi_{\ge k}(\lambda)$, it follows by the definition of $\lambda$ and Lemma~\ref{CWnm} that ${\lambda = (1+o(1))\mu_c}$.
Let $A_n$ denote the event that $(i)$ $\dd$ is a proper sequence, and $(ii)$ a random $k$-core $\K$ with degree sequence $\dd$ has probability at least $1/n$ of containing a 2-core with parameters $(t,s)\in \mathcal{I}_{\sprs}$. Implicitly, this event is contained in the event that the sum of components of $\dd$ is even. Note that the restriction $t < \eps n$ and the definition of $\eps$ imply that $t < \eps_1\nn$.
\medskip
In this subsection we make use of two types of degree sequences; proper degree sequences $\dd \in \D_{\nn,2\mm}$ for $\K$, and degree sequences $\dd \in \mathbb{N}^{t}$ for subsets of $V(\K)$ of size $t$. In order to distinguish between these two types we write either $\ddn$ or $\ddt$, respectively. When referring to a subset $U\subset V(\K)$ of size $t$, we use $u_{1},\dots,u_{t}$ to denote its vertices, even when this is not written explicitly.
Since estimating the expected number of sparse 2-cores involves some
tedious calculations, we make them in several steps. We begin with
a few bounds which are given without context at this moment and will
be useful later. First, for a given $\ddt\in \mathbb{N}^{t}$, by using simple combinatorial
identities and by letting $r=s-t$ and $h_{i}'=h_{i}-2$ for every $i$, we have
\begin{eqnarray}
\sum_{\substack{h_{1},\ldots,h_{t}\ge2\\
\sum h_{i}=2s
}
}\hspace{4pt}\prod_{i=1}^{t}\binom{d_{i}}{h_{i}} & = & \sum_{\substack{h_{1},\ldots,h_{t}\ge2\\
\sum h_{i}=2s
}
}\hspace{4pt}\prod_{i=1}^{t}\binom{d_{i}-2}{h_{i}-2}\frac{[d_{i}]_{2}}{[h_{i}]_{2}}\nonumber \\
& \le & \sum_{\substack{h_{1}',\ldots,h_{t}'\ge0\\
\sum h_{i}'=2r
}
}\hspace{4pt}\prod_{i=1}^{t}\binom{d_{i}-2}{h_{i}'} [d_{i}]_{2}/2\nonumber \\
& = & 2^{-t}\binom{\sum_{i=1}^{t}(d_{i}-2)}{2r}\prod_{i=1}^{t}[d_{i}]_{2}~.\label{eq:shortCalc}
\end{eqnarray}
Second, for $j\le2r$ let $\widetilde\D_{j}=\left\{ \ddt \in \D_{t,kt+j} \mid \forall i\,d_{i}\ge k\right\}$, and note that $\left|\widetilde\D_{j}\right|=\binom{j+t-1}{t-1}$ and that
\begin{equation}\label{eq:sum-Dj}
\sum_{j=0}^{2r}\binom{j+t-1}{t-1} = \binom{2r+t}{t} \le \binom{2t}{2r}\le \left(\frac{et}{r}\right)^{2r}.
\end{equation}
Furthermore, at least $t-j$ entries in every $\ddt\in \widetilde\D_{j}$ equal~$k$,
and $\prod_{i=1}^{t}[d_{i}]_{2}$ is maximised when the entries in
$\ddt\in \widetilde\D_{j}$ are as equal as possible; that is, when~$j$ of them
equal~$k+1$. Recall that the distribution of a single component of $\dd$ is $Z_{k}(\lambda)$, abbreviated to $Z$ for the
remainder of this subsection. Since $\lambda = (1+o(1))\mu_c$ we may estimate the probabilities in the distribution of $Z$ asymptotically by using $Z_{k}(\mu_c)$. By Lemma~\ref{lem:technical}, for $n$ sufficiently large we have
\begin{eqnarray}
\left[k\right]_{2}\Pr\left[Z=k\right] & \stackrel{\eqref{eq:average-degree-in-k-core}}{=} & \left(1+o\left(1\right)\right)\left[k\right]_{2}\frac{\hat{d}}{k}\Pr\left[Z_{k-1}\left(\mu_{c}\right)=k-1\right]\nonumber\\
& \le & \left(1+o\left(1\right)\right)\left(1-2\delta\right)\hat{d}<\left(1-\delta\right)\hat{d}.\label{eq:fact3}
\end{eqnarray}
Hence,
\begin{eqnarray}
\sum_{\substack{d_{1},\ldots,d_{t}\ge k\\
\sum d_{i}\le kt+2r
}
}\prod_{i=1}^{t}[d_{i}]_{2}\Pr\left[Z=d_{i}\right] & = & \sum_{j=0}^{2r}\hspace{4pt}\sum_{\ddt\in \widetilde\D_{j}}\hspace{4pt}\prod_{i=1}^{t}[d_{i}]_{2}\Pr\left[Z=d_{i}\right]\nonumber \\
& \le & \sum_{j=0}^{2r}\left|\widetilde\D_{j}\right|\left([k+1]_{2}\right)^{j}\left([k]_{2}\right)^{t-j}\Pr\left[Z=k\right]^{t-j}\nonumber \\
& \stackrel{\eqref{eq:fact3}}{\le} & \left([k+1]_{2}\right)^{2r}\sum_{j=0}^{2r}\left|\widetilde\D_{j}\right|\left((1-\delta)\hd\right)^{t-j}\nonumber \\
& \le & \left((1-\delta)\hd\right)^{t}\left(2k^{2}\right)^{2r}\sum_{j=0}^{2r}\binom{j+t-1}{t-1}\nonumber \\
& \stackrel{\eqref{eq:sum-Dj}}{\le} & \left((1-\delta)\hd\right)^{t}\left(\frac{2ek^{2}t}{r}\right)^{2r}.\label{eq:longCalc}
\end{eqnarray}
We now wish to bound the probability that a given set of vertices induces a sparse bad 2-core in a random $k$-core $\K$ with a given degree sequence, and this is where the configuration model becomes useful. For a given degree sequence $\ddn \in \D_{\nn,2\mm}$ and for positive integers $t < \nn$ and $s<\mm$, let us count how many configurations $F$ for $\ddn$ yield a multigraph $H = H(F)$ such that $H[U]$ is a 2-core with $s$ edges, where $U = \{u_{1},\dots,u_{t}\}$ is the set of the first $t$ vertices in the sequence. First we have to choose a degree sequence for $H[U]$, that is, choose $\hh = \{h_{1},\dots,h_{t}\} \in \D_{t,2s}$ such that $d_{H[U]}(u_i) = h_i \ge 2$ for every $i$. Given $\hh$, there are $\prod_{i=1}^{t}\binom{d_{i}}{h_{i}}$ possibilities to determine for each $u_{i}$ which $h_{i}$ of its $d_{i}$ half-edges go inside $U$ (while the rest go outside). Finally, there are $(2s-1)!!$ configurations for $H[U]$, and $(2\mm-2s-1)!!$ configurations for the rest of $H$. It follows that if $F$ is chosen uniform
ly at random from all $(2\mm-1)!!$ possible configurations for~$\ddn$ then the probability of $H[U]$ being a 2-core with $s$ edges is at most
\begin{equation}\label{eq:prob-config-dense}
\sum_{\substack{h_{1},\ldots,h_{t}\ge2\\\sum h_{i}=2s}}\hspace{4pt}\prod_{i=1}^{t}\binom{d_{i}}{h_{i}} \frac{(2s-1)!!(2\mm-2s-1)!!}{(2\mm-1)!!}
\stackrel{\eqref{eq:shortCalc}}{\le} \frac{2^{s}s!}{\left(2\mm-2s\right)^{s}} \cdot2^{-t}\binom{\sum_{i=1}^{t}(d_{i}-2)}{2r}\prod_{i=1}^{t}[d_{i}]_{2}~.
\end{equation}
Note that this bound does not depend on $d_{t+1},\dots,d_{\nn}$.
Now, for integers $t$ and $s$, an arbitrary subset $U\subset V(\K)$ of size $t$,
and a degree sequence $\ddt\in\mathbb{N}^{t}$, denote by $P(U,\ddt,s)$
the probability that $\K[U]$ is a bad 2-core with $s$ edges,
where $\K$ is as in the definition of the event $A_n$ but conditioned upon satisfying $d_\K(u_i) = d_i$ for every $1 \le i \le t$. By the condition on the boundary
of bad 2-cores, we only need to consider sequences $\ddt$ such that $\sum_{i=1}^{t}d_{i}\le(k-2)t+2s=kt+2r$,
and in particular $\sum_{i=1}^{t}(d_{i}-2)\le kt$, since $r<t$.
Recall that $2\mm=\hd\nn$ and note that for sparse 2-cores we have
\begin{equation}
2s=2(t+r)\le3t\le kt\le\hd\eps_{1}\nn.\label{eq:two-ess}
\end{equation}
Using~\eqref{eq:prob-config-dense}, and Corollary~\ref{cor:prob-proper} with $2^r$ standing for $f(n)$, we can bound $P(U,\ddt,s)$ from above by
\begin{eqnarray}
P(U,\ddt,s) & \le & 2^r\frac{2^{s}s!}{\left(2\mm-2s\right)^{s}}\cdot2^{-t}\binom{\sum_{i=1}^{t}(d_{i}-2)}{2r}\prod_{i=1}^{t}[d_{i}]_{2}\nonumber \\
& \stackrel{\eqref{eq:two-ess}}{\le} & \frac{2^{2r}s!}{\left(\hd\nn-\hd\eps_{1}\nn\right)^{s}}\binom{kt}{2r}\prod_{i=1}^{t}[d_{i}]_{2}\nonumber \\
& \le & \frac{2^{2r}s!}{\left((1-\eps_{1})\hd\nn\right)^{s}}\left(\frac{ekt}{2r}\right)^{2r}\prod_{i=1}^{t}[d_{i}]_{2}\nonumber \\
& = & \frac{s!}{\left((1-\eps_{1})\hd\nn\right)^{s}}\left(\frac{ekt}{r}\right)^{2r}\prod_{i=1}^{t}[d_{i}]_{2}~.\label{eq:prob_admits}
\end{eqnarray}
For the random sequence $\dd$ of independent truncated Poissons under consideration, define $A_n(t,s)$ to be the expected number of bad 2-cores, that have $t$ vertices and $s$ edges, in the random $k$-core $\K$ (in the case the sequence is proper and has even sum --- otherwise treat the number as 0). It then follows from above that for any given pair $(t,s)\in\cI_{\sprs}$ we have
\begin{eqnarray*}
A_n(t,s) & \le & \sum_{\substack{U\subset V(\K) \\ |U| = t}}\hspace{4pt}\sum_{\substack{d_{1},\ldots,d_{t}\ge k\\
\sum d_{i}\le kt+2r
}
}\Pr\left[\deg_{\K}(u_{i})=d_{i}\text{ for }i=1,2,\ldots,t\right]P(U,\ddt,s)\\
& \stackrel{(\ref{eq:prob_admits})}\le & \binom{\nn}{t}\sum_{\substack{d_{1},\ldots,d_{t}\ge k\\
\sum d_{i}\le kt+2r
}
}\left(\prod_{i=1}^{t}\Pr\left[Z=d_{i}\right]\right)\cdot\frac{s!}{\left((1-\eps_{1})\hd\nn\right)^{s}}\left(\frac{ekt}{r}\right)^{2r}\prod_{i=1}^{t}[d_{i}]_{2}\\
& \le & \frac{\nn^{t}}{t!}\frac{s!}{\left((1-\eps_{1})\hd\nn\right)^{s}} \left(\frac{ekt}{r}\right)^{2r}\sum_{\substack{d_{1},\ldots,d_{t}\ge k\\
\sum d_{i}\le kt+2r
}
}\prod_{i=1}^{t}[d_{i}]_{2}\Pr\left[Z=d_{i}\right]\\
& \stackrel{\eqref{eq:longCalc}}{{\le}} & \frac{[s]_{r}}{\left((1-\eps_{1})\hd\right)^{s}\nn^{r}} \left(\frac{ekt}{r}\right)^{2r}\left((1-\delta)\hd\right)^{t}\left(\frac{2ek^{2}t}{r}\right)^{2r}\\
& \le & \left(\frac{1-\delta}{1-\eps_{1}}\right)^{t}\left(\frac{t}{r}\right)^{4r}\left(\frac{4se^{4}k^{6}}{(1-\eps_{1})\hd\nn}\right)^{r}\\
& \stackrel{\eqref{eq:two-ess}}{\le} & \left(\frac{1-\delta}{1-\eps_{1}}\left(\frac{t}{r}\right)^{4r/t}\right)^{t}\left(\frac{\eps_{1}2e^{4}k^{6}}{1-\eps_{1}}\right)^{r}.
\end{eqnarray*}
By definition of $\eps_{1}$ we have $\frac{1-\delta}{1-\eps_{1}}\le1-\delta/2$
and $\eps_{1}2e^{4}k^{6}\le1-\eps_{1}$. Using the fact that $x\mapsto x^{-x}$
is increasing for $0<x<1/e$, and by definition of $\delta_{1}$,
for every $(t,s)\in\cI_{\sprs}$ we obtain
\[
A_n(t,s)\le \left(\frac{1-\delta}{1-\eps_{1}}\left(\frac{r}{t}\right)^{-4r/t}\right)^{t}
\le\left(\left(1-\frac{\delta}{2}\right)\delta_{1}^{-4\delta_{1}}\right)^{t}\le\left(1-\frac{\delta}{4}\right)^{t}.
\]
Recalling that $t>\log^{2}n$, and observing that $\left|\mathcal{I}\right|<n^{3}$, we get
\[
\sum_{(t,s)\in\cI_{\sprs}}A_n(t,s)
\le \sum_{(t,s)\in\cI_{\sprs}}\left(1-\delta/4\right)^{t}
\le \left|\cI\right|\left(1-\delta/4\right)^{\log^{2}n}=O(1/n^2).
\]
The probability of the event $A_n$ is hence $O(1/n)$. Thus, by Theorem~\ref{thm:maincor}, if $\dd$ is distributed as the degree
sequence of $\K(n,c,k)$, the probability that a random core with degree sequence $\dd$ has a sparse 2-core and is proper is $o(1)$. That is, WHP $\K(n,c,k)$ does not have both a proper degree sequence and a sparse 2-core. Recalling that it has proper degree sequence WHP, Lemma~\ref{lem:no-bad-two-cores} follows.
| 4f38d6dfa8eeb517cab678c00a32626454c0b176 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The objective of this paper is to discuss research needs and requirements towards the predictive simulation of turbulent reacting flows. While significant progress has been made on the fundamental understanding of turbulent combustion, the modeling of stationary combustion processes and the qualitative evaluation of combustion dynamics, deficiencies remain in regard to quantitative predictions and reliable simulations of combustion processes, involving complex geometries, unsteady operating conditions, pollutants, and the consideration of multiphase flows. Faced with these challenges, questions arise on how to select a particular simulation approach to predict turbulent reacting flows under specific constraints about solution accuracy and computational resource allocation, and what are techniques that allow us to assess the accuracy of a turbulent-combustion simulation.
In the following, we consider simulation methods for modeling gaseous turbulent combustion, consisting of the following components: a discretized representation of the flow geometry; initial and boundary conditions; a mathematical model for the numerical solution of the governing equations, such as Reynolds-averaged Navier-Stokes (RANS), large-eddy simulation (LES), and direct numerical simulation (DNS); a combustion model for the representation of energy release, species conversion and the coupling to density, pressure, and constitutive properties; closure models for the description of the feedback of turbulence with the gas-phase combustion; a chemical-kinetic mechanism; and constitutive relations for the thermo-chemical quantities. Furthermore, the selection of a particular simulation approach is constrained by user-specific requirements on the accuracy in predicting certain responses functions (RFs), comprising quantities of interest (QoIs) or processes of interest (PoIs). These requirements have to be considered in the presence of computational resources that impose limitations on the choice of certain simulation approaches and modeling strategies.
In the following, we seek to address several questions with the goal of examining the current state-of-the-art in combustion modeling, identifying research gaps, and stipulating ideas to overcome these challenges. Specifically,~\cref{SEC_SIM_PURP} is concerned with the selection of a simulation approach that aligns with the purpose of a simulation. We then proceed by discussing the selection of combustion models and discuss different modeling strategies to represent turbulent combustion. Aspects on the importance of numerical errors and the interaction of these errors with physical models are discussed in~\cref{SEC_NUM_DISCR}.~\Cref{SEC_MODEL_ACCURACY} is concerned with the qualitative assessment of the simulation accuracy, and recent developments on the probabilistic analysis and dynamic simulation content are discussed. Emerging concepts for augmenting simulations with experimental data is addressed in~\cref{SEC_AUG_SIM}, and the paper finishes with a summary in~\cref{SEC_CONCL}.
\section{\label{SEC_SIM_PURP}What is the purpose of a combustion simulation?}
\begin{figure}[!htb!]
\centering
\includegraphics[width=0.6\textwidth, clip=, keepaspectratio]{modelPerformance2.pdf}
\caption{\label{FIG_MODEL_PERFORMANCE}Schematic illustrating the selection of simulation approach and dependence on requirements about accuracy and computational cost.}
\end{figure}
The first and perhaps most important question to ask is in regard to the objective of a particular combustion simulation. The answer to this questions will depend on the field of application (see~\cref{FIG_MODEL_PERFORMANCE}): while industrial applications might utilize simulation tools to support the design optimization with the goal of minimizing emissions or exploring new combustion strategies, academia and research laboratories might employ simulations for gaining principal understanding about flame-structure, energy transfer or other fundamental combustion-physical processes. Other applications may rely on simulations to support the diagnosis of failure modes or for risk assessment. As such, it becomes apparent that the specific purpose determines the choice of a particular simulation approach, the model selection, as well as requirements on the accuracy of the simulation and computational resource allocation.
Most common is the prediction of certain scalar response quantities, such as total heat release, fuel conversion, temperature, or pollutants. These response quantities are either directly solved for in a simulation or can be determined by post-processing the simulation results. Results are commonly reported in the form of statistical quantities, ensemble-averaged quantities, or conditional data in order to facilitate analysis or comparison with measurements. Temperature, species of \ce{CO2} and \ce{H2O}, and heat release, which is controlled by few exothermic reaction pathways, can all be predicted to good accuracy independent of a particular combustion regime~\cite{WU_SEE_WANG_IHME_CF2015,SEE_IHME_PCI2015}. This, however, is not the case for minor species and pollutants such as CO, \ce{NO_x}, or soot. Since these quantities are tightly regulated in aviation gas turbines, stationary power generation, and automotive applications, their reliable prediction introduces significant modeling challenges.
In contrast to predicting QoIs, the simulation of transient processes becomes relevant for determining combustion dynamics, such as ignition, blow-out limits, stability boundaries, onset of engine-knock conditions, extinction, or thermoacoustic instabilities. Since these processes of interest (PoIs) are deterministic chaotic phenomena that are characterized by aperiodic long-term behavior with strong sensitivity to initial conditions, probabilistic techniques are most likely more suitable to describe these events.
The purpose of a simulation is linked to specific requirements on solution accuracy, model fidelity, and resource allocation, as illustrated in~\cref{FIG_MODEL_PERFORMANCE}. For instance, the computational exploration of a high-dimensional parameter space to optimize the burner performance emphasizes the need for low-order models at the expense of solution accuracy, and targeted simulations of higher fidelity are then conducted to validate the most promising design candidates. Time-to-solution and computational resource allocation are constraining factors on computational simulations.
While the computational cost can be estimated reasonably well from the knowledge about the algorithmic complexity, spatio-temporal resolution, and code-scalability, quantifying improvements in solution accuracy of a specific model prior to performing the simulation is not trivial.
\begin{tcolorbox}[colback=myCol,breakable]
Recommendations:
\begin{itemizePacked}
\item Establish formal methods for down-selecting specific simulation approaches in order to meet user-specific requirements of combustion simulations.
\item Articulate specific requirements on solution accuracy, response functions, and resource allocations to obtain improvements in solution accuracy ${\cal{E}}$ for certain computational cost ${\cal{C}}$.
\end{itemizePacked}
\end{tcolorbox}
\nomenclature{${\cal{E}}$}{error in solution accuracy}
\nomenclature{${\cal{C}}$}{computational cost}
\section{\label{SEC_COMB_MODEL}How should a combustion model be selected?}
The computational modeling of turbulent reacting flows is concerned with overcoming challenges of resolving all relevant spatio-temporal scales and the consideration of a large number of chemical species that participate in the combustion process. The evolution of the chemical species can be obtained as solution to a advection-diffusion-reaction equation, which is here written in general form:
\begin{equation}
\label{EQ_SPECIES_TE}
\partial_t (\rho {\mbox{\boldmath$\phi$}}) + \nabla\cdot(\rho{\mbox{\boldmath$u$}}{\mbox{\boldmath$\phi$}}) = -\nabla\cdot{\bf j}_{{\phi}}+\rho\dot{{\mbox{\boldmath$\omega$}}}_{{\phi}}\;,
\end{equation}
where $\rho$ is the density, ${\mbox{\boldmath$u$}}\in\mathbb{R}^{N_d}$ is the velocity vector, ${\mbox{\boldmath$\phi$}}\in\mathbb{R}^{N_s}$ is the vector of species mass fractions, ${\bf{j}}_{{\phi}}\in\mathbb{R}^{N_s\times N_d}$ is the mass-diffusion flux matrix, and $\dot{{\mbox{\boldmath$\omega$}}}_{{\phi}}\in\mathbb{R}^{N_s}$ is the vector of reaction rates.
\nomenclature{$\rho$}{density}
\nomenclature{${\mbox{\boldmath$u$}}$}{velocity vector}
\nomenclature{${\mbox{\boldmath$\phi$}}$}{vector of species mass fractions}
\nomenclature{${\bf{j}}_{\zeta}$}{diffusion flux matrix of quantity $\zeta$}
\nomenclature{$\dot{{\mbox{\boldmath$\omega$}}}$}{vector of reaction rates}
\nomenclature{$\wt{(\cdot)}$}{Favre-filtered quantity}
\nomenclature{$\overline{(\cdot)}$}{Reynolds-filtered quantity}
To overcome the challenge of describing effects that occur at computationally unresolved scales, involving turbulent stresses, turbulent scalar transport, and turbulence/chemistry interaction,~\cref{EQ_SPECIES_TE} is filtered or averaged using a Favre-filtering procedure:
\begin{equation}
\label{EQ_SPECIES_TE_FILTER}
\partial_t (\overline{\rho} \wt{{\mbox{\boldmath$\phi$}}}) + \nabla\cdot(\overline{\rho}\wt{{\mbox{\boldmath$u$}}}\wt{{\mbox{\boldmath$\phi$}}}) = -\nabla\cdot\overline{\bf j}_{{\phi}}-\nabla\cdot{\bf j}_{{\phi}}^{\text{sgs}}+\overline{\rho}\wt{\dot{{\mbox{\boldmath$\omega$}}}}_{{\phi}}\;,
\end{equation}
which introduces unclosed terms for the subgrid scalar flux ${\bf j}_{\phi}^{\text{sgs}}$ and the filtered chemical source term $\wt{\dot{{\mbox{\boldmath$\omega$}}}}_{{\phi}}$. These terms require modeling. Here, the low-pass filtering follows the common definition of a Favre-filter with $\wt{\phi}=\overline{\rho}^{-1}\int \rho({\mbox{\boldmath$\xi$}},t)\phi({\mbox{\boldmath$\xi$}},t)G({\mbox{\boldmath$x$}}-{\mbox{\boldmath$\xi$}},t) \textup d{\mbox{\boldmath$\xi$}}$ with $G$ being the compact filter kernel. Closure models for ${\bf j}_{\phi}^{\text{sgs}}$ are commonly adopted from non-reacting and isothermal flow models and corrections are introduced to account for variable transport properties and density variations.
To reduce the chemical complexity,~\cref{EQ_SPECIES_TE_FILTER} is replaced by transport equations associated with appropriate low-dimensional manifold methods, in which the chemical state vector is represented in terms of a low-dimensional manifold:
\begin{equation}
{\mbox{\boldmath$\phi$}}\simeq{{\mbox{\boldmath$\varphi$}}} = {\cal{M}}({\mbox{\boldmath$\psi$}})\;,
\end{equation}
\nomenclature{$t$}{time}
\nomenclature{${\mbox{\boldmath$\psi$}}$}{state vector describing combustion manifold}
\nomenclature{${\cal{M}}$}{low-dimensional manifold}
\nomenclature{sgs}{subgrid-scale quantity}
\nomenclature{$N$}{number}
\nomenclature{$T$}{temperature}
\nomenclature{$Z$}{mixture fraction}
where ${\mbox{\boldmath$\psi$}}\in\mathbb{R}^{N_{\cal{M}}}$ is the state vector that is used to parameterize the manifold. The vector ${\mbox{\boldmath$\psi$}}$ may include a subset of species mass fractions, mixture fraction, reaction progress, as well as other flow-field describing quantities, such as strain rate or scalar dissipation. The topology of the manifold and with this the manifold-spanning quantities depend on a particular combustion model. The evolution of the manifold-spanning state-vector ${\mbox{\boldmath$\psi$}}$ is then obtained as solution to a transport equation:
\begin{equation}
\label{EQ_SPECIES_TE_FILTER_MFOLD}
\partial_t (\overline{\rho} \wt{{\mbox{\boldmath$\psi$}}}) + \nabla\cdot(\overline{\rho}\wt{{\mbox{\boldmath$u$}}}\wt{{\mbox{\boldmath$\psi$}}}) = -\nabla\cdot\overline{\bf j}_{{\psi}}-\nabla\cdot{\bf j}_{{\psi}}^{\text{sgs}}+\overline{\rho}\wt{\dot{{\mbox{\boldmath$\omega$}}}}_{{\psi}}\;.
\end{equation}
Therefore, by using the manifold formulation, the number of equations that is solved reduces from $N_s$ to $N_{\cal{M}}$. In the case that $N_{\cal{M}}=N_s,$ the manifold model recovers the chemical complexity of the complete chemical system.
Over recent years, different modeling approaches have been developed that address the reduction of the chemical complexity and the modeling of the turbulence/chemistry interaction. At a fundamental level, these approaches can be distinguished into \emph{topology-free} and \emph{topology-based} combustion models~\cite{WU_SEE_WANG_IHME_CF2015}. Examples of topology-free combustion models are finite-rate chemistry models, the eddy-dissipation concept~\cite{MAGNUSSEN_HJERTBERG_SYMP16}, probability-density function (PDF) methods~\cite{POPE_PECS85,HAWORTH_PECS2010}, and deconvolution methods~\cite{PANTANO_SARKAR_POF2001,DOMINGO_VERVISCH_PCI2014,WANG_IHME_CF2017,WANG_IHME_CF2019}. These combustion models introduce limited assumptions about the flame structure, and are therefore considered to be applicable to a wide range of combustion problems. Topology-free models enable the incorporation of different combustion-physical processes, but are accompanied by higher computational complexity. In contrast, topology-based combustion models exploit the topological structure of the flame by constructing the manifold from the solution of canonical flame configurations, such as laminar counterflow diffusion flames, freely propagating premixed flames, or one-dimensional embedded flame elements. Examples of topology-based combustion models are the class of flamelet models~\cite{PETERS_PECS84,GICQUEL_DARABIHA_THEVENIN_PCI28,OIJEN_DEGOEY_CST2000,CHUCK_JFM04,IHME_PCI04}, thickened flame models~\cite{COLIN_VEYNANTE_POINSOT_POF2000}, and conditional moment closures~\cite{KLIMENKO_BILGER_PECS99,KLIMENKO_CTM01}.
Common to all models are assumptions that are invoked to make them tractable. These assumptions include the consideration of particular combustion modes, transport properties, chemical complexity, contraction of chemical time-scales, or the omission of multidimensional, non-equilibrium and radiation effects. These models have been examined through \emph{a priori} analyses or \emph{a posteriori} comparisons with experimental observations and high-resolution simulation data. Model extensions have been proposed to address shortcomings that are encountered in applications to new flame configurations, combustion regimes, or operating conditions. However, in these extensions it is critical to ensure that the parent model constitutes a proper subset of the extended model in order to recover the performance of the underlying model (see~\cref{DG_MODEL_EXTENSION}). Examples of model extensions are the consideration of heat-loss effects~\cite{MA_WU_IHME_HICKEY_AIAAJ2018,FIORINA_CTM03}, radiation~\cite{IHME_PITSCH_POF2008}, multistream flows~\cite{HASSE_PETERS_PCI30,IHME_SEE_PCI33,IHME_ZHANG_HE_DALLY_FTC2012,CHEN_IHME_CF2013}, and transient processes~\cite{IHME_SEE_CF2010}. While these model extensions fulfill the requirement on the subset-completeness, increasing storage requirements of these higher-dimensional manifold representations can impact the numerical accuracy and may challenge the resolution requirements in accurately representing the combustion manifold. Furthermore, the solution of additional transport equations that often involve high-order moments introduce additional challenges in obtaining stable solutions and the modeling of unclosed terms. As such, it becomes clear that further extensions of manifold-based combustion models will face limitations. In light of rapidly increasing computational resources, efficient time-integration schemes with linear complexity~\cite{WU_MA_IHME_CPC2019}, efficient chemical reduction techniques~\cite{LU_LAW_PCI30,PEPIOT_PITSCH_CF2008,NIEMEYER_SUNG_RAJU_CF2010,JARAVEL_WU_IHME_CF2019}, and the development of realizable reconstruction techniques for subgrid contributions~\cite{WANG_IHME_CF2017,WANG_IHME_CF2019,NIKOLAOU_CANT_VERVISCH_PRF018}, it is expected that the utilize of topology-based combustion models will decline in the foreseeable future.
\begin{figure}[!htb!]
\centering
\includegraphics[width=0.5\textwidth, clip=, keepaspectratio]{modelExtension1.pdf}
\caption{\label{DG_MODEL_EXTENSION}Radar chart, schematically illustrating the model performance and impact of partial and complete model extension. Directions along arrow indicates improved model performance, multiple thin arrows emphasize that several processes or quantities of interest require consideration.}
\end{figure}
Combustion models are commonly employed in a \emph{monolithic} form, meaning that only a single model is utilized to describe the entire combustion process encountered in a combustor configuration. While this is certainly appropriate for flames that operate in the vicinity of asymptotic regimes of premixed, non-premixed, or near-equilibrium conditions, these monolithic modeling strategies introduce challenges for representing complex combustion processes involving multimode combustion, multiphase flows, or the consideration of localized combustion events such as flame/wall interaction, ignition, or flame/turbulence coupling. This is illustrated in~\cref{FIG_COMB_MODE}, showing instantaneous simulation results of a single-element coaxial rocket injector~\cite{MA_WU_IHME_HICKEY_AIAAJ2018}. The instantaneous temperature field is shown in the left panel and combustion relevant combustion regimes are presented on the right, emphasizing the complexity of turbulent combustion.
\begin{figure}[!htb!]
\centering
\includegraphics[width=0.71\textwidth, clip=, keepaspectratio]{combustionModes.pdf}
\caption{\label{FIG_COMB_MODE}Simulation of uni-element rocket injector~\cite{MA_WU_IHME_HICKEY_AIAAJ2018}, showing instantaneous temperature field (left) and combustion modes (right).}
\end{figure}
To reduce the modeling complexity, opportunities arise for combining different combustion models in such a way as to optimally represent specific thermochemical processes that are encountered in a combustion simulation. This principle is encapsulated in the Pareto-Efficient Combustion (PEC) framework~\cite{WU_SEE_WANG_IHME_CF2015,WU_MA_JARAVEL_IHME_PCI2019}, in which an optimal combustion-submodel assignment is employed to meet user-specific requirements about solution accuracy on specific QoIs under consideration of computational-cost constraints. This submodel assignment takes advantage of existing combustion models that are readily available in existing CFD-solvers. PEC is a trust-region formulation and utilizes a \emph{Pareto efficiency} to facilitate an optimal combustion submodel assignment. The notion of Pareto efficiency describes an optimality between two competing conditions -- in the PEC model these two conditions are (i) the accuracy in predicting user-specific quantities of interest and (ii) computational cost. Pareto-optimality is then achieved when it is impossible to make any one condition better without making the other condition worse. In the present case, this results in a Pareto front.
By combining different combustion models, PEC enables the general model-adaptation to the underlying flow-field representation so that regions of different combustion-physical complexities are represented by the most appropriate model without violating intrinsic model assumptions. Key attributes of this PEC-formulation are (i) the local adaptation of the model fidelity and computational complexity to the underlying flow-field-specific combustion processes, (ii) the control of the model accuracy and computational cost through the selection of error threshold, model compliance, and quantities of interest, and (iii) the representation of realistic fuel chemistry through detailed and reduced kinetic models. To enable the application to turbulent combustion requires the development of algorithms that take into consideration a robust model selection, error-evaluation, subzone coupling, conservation properties, and considerations in regard to load-balancing and dynamic submodel adaptation.
In this context, we note that other adaptation techniques have been developed, which are concerned with locally adapting the chemical complexity to thermodynamic activity in flow simulation~\cite{LIANG2009527,SHI_LIANG_GE_REITZ_CTM2010,XU_AMEEN_SOM_CHEN_REN_LU_CF2018,
GOU2013225,
HE_ETAL_EF2011,
REN201419,
BANERJEE2006619,
LIANG20153236,
GREEN_BARTON_ETAL_IECR2001,
SCHWER2003451}. As such, PEC generalizes these dynamic adaptive chemistry techniques by adapting the combustion-model assignment to meet user-specific requirements about solution-accuracy with respect to user-selected quantities of interest.
\begin{tcolorbox}[colback=myCol,breakable]
Recommendations:
\begin{itemizePacked}
\item Develop reliable indicators for assessing the compliance of combustion models with underlying CFD-solution.
\item Consider criteria that indicate when models are employed outside its intended validity and violate intrinsic model assumptions.
\item Develop hierarchical modeling strategies to meet user-specific requirements on accuracy and model fidelity in predicting specific response data of interest.
\item Formulate closure models that employ consistent principles for the representation of turbulence/chemistry interaction, turbulent scalar fluxes, and turbulent stresses.
\item Ensure and demonstrate that a model extension is a proper superset of the underlying parent model.
\item Develop subgrid-closure models for turbulent transport, turbulence/chemistry coupling, and turbulence/radiation interaction and other unclosed terms that are based on fundamental compressible reacting flow-theory instead of relying on non-reacting and isothermal flow models.
\end{itemizePacked}
\end{tcolorbox}
\section{\label{SEC_NUM_DISCR}What is the impact of the numerical discretization on the combustion simulation?}
In simulations of turbulent reacting flows, it is difficult to separate numerical errors from modeling errors~\cite{POPE_NJP2004}. This is attributed to the overlap of mesh resolution and physical scales of the flow as well as the non-linear interaction of physical models and filters with the numerical discretization scheme~\cite{Klein2008,KAUL_RAMAN_BALARAC_PITSCH_POF2009}. Commonly employed in simulations of turbulent reacting flows are low-order schemes that utilize
finite-difference (FD) or finite-volume (FV) discretizations. However, by representing complex geometries that require unstructured and skewed meshes, these schemes can introduce appreciable amounts of numerical dissipation and dispersion errors, which can deteriorate the representation of the flame structure, turbulent stresses, transport, and turbulence/chemistry coupling. Mitigating the role of discretization errors is therefore critical to ensure the general behavior of a combustion model that is based on physical principles. Explicitly filtered LES-methods enable the formal separation of numerical discretization and modeling errors. Success of these methods has been demonstrated in \emph{a priori} and \emph{a posteriori} studies~\cite{RADHAKRISHNAN_BELLAN_JFM2012,COCKS20153394,GALLAGHER_SANKARAN_AIAAJ2019}. More recently, coupling effects between numerical discretization and filtering on numerical errors were investigated~\cite{EDOH_GALLAGHER_CTR2019}. Different spatial discretization schemes were examined in the context of explicitly filtered LES, showing that increasing the filter-to-grid ratio reduces the impact of numerical errors on the simulation results.
A common technique for mitigating discretization errors is mesh adaptation. Perhaps the simplest approach is to employ local and static mesh stretching. In this method, the mesh is locally stretched or contracted to either reduce or refine the resolution in certain regions of interest. Mesh stretching can be accomplished through directional refinement for structured discretizations, and the order of accuracy can be preserved through mapping onto a regular reference grid. Unstructured solution methods are not constrained by directional stretching, and provide greater flexibility in locally adapting to complex flow features and complex flow geometries.
Other approaches for mesh adaptation are nested grids or dynamically adaptive mesh refinement (AMR). In nested-grid techniques, which has been employed in weather forecasting~\cite{GIORGI_MEARNS_JGRA1999}, one or multiple finer meshes are statically embedded in a background mesh. To increase the physical fidelity, different submodels and additional physical representations can be utilized in the nested region. This approach enables communication of coarse-scale simulations with more refined mesoscale formulations. Coupling between the outer and nested domains is implemented either through one-way coupling, in which only the outer domain is providing boundary conditions to the nested region, or two-way coupling, in which both domains are tightly coupled through the exchange of boundary conditions. With relevance to embedding, it is important to recognize that the accuracy and performance of a nested-grid method is inherently dependent on constraints and matching conditions at the static boundaries. Another issue is to enforce conservation properties for mass, momentum, species, and energy across domain interfaces. Although the lack of conservation can affect the simulation accuracy, these methods are commonly employed for short-term predictions that do not rely on exact conservation properties. Nested-grid methods with two-way coupling can be considered as a special case of static non-uniform meshing. In contrast to nested-grid techniques with heterogeneous model description, AMR is based on a dynamic mesh adaptation in which different patch- or block-based refinement regions are represented using the same discretization and physical submodels~\cite{BERGER_COLELLA_JCP1989,BELL_BERGER_SALTZMAN_WECLOME_SIAMJ1994,AMR_BOOK2005}. For steady-state flows, AMR reduces to the nested grid approach. AMR has been employed for simulating shock waves, flow-discontinuities, and large-scale flames~\cite{GAO_GROTH_JCP2010,HILL_PULLIN_JCP2004,BELL_ETAL_PNAS2005,AMROC,AMREX,LAPOINTE_ETAL_FSJ2020}.
Apart from mesh-refinement, the adaptation of the solution-representation offers another approach to resolve flow-field features of importance. High-order methods, such as the class of spectral and discontinuous Galerkin (DG) schemes, are techniques that -- in addition to mesh-adaptation ($h$-adaptation) -- enable refinement in polynomial order ($p$-refinement). These variational discretization techniques provide opportunities for mitigating numerical approximation errors by enabling the separation between numerical discretization and physical modeling. Specifically, compared to conventional FV/FD-schemes, advantages of DG-methods are that they (i) provide high-order accuracy on unstructured grids and complex geometries, (ii) are well suited for advanced refinement strategies, using local mesh adaptation and refinement in polynomial order, (iii) enable a compact discretization with subcell resolution, (iv) have low numerical dissipation and dispersion, and (v) the large degree of structured computations and data locality introduce a high level of parallelism, making these methods particularly suitable for high-performance computing. Since variational methods employ an element-local discretization, they are suitable for unstructured meshes, and can therefore be combined with local mesh refinement. Rigorous convergence proofs for DG-methods have been established, showing optimal convergence $\epsilon \sim h^{p+1}$ for general meshes, and the solution accuracy does not deteriorate with element type.
\nomenclature{$\epsilon$}{error}
\nomenclature{$h$}{mesh resolution}
\nomenclature{$p$}{polynomial order}
DG-methods support local and flexible $hp$-strategies, which offers greater flexibility compared to conventional $h$-adaptation via nesting, AMR, or grid stretching alone. Theory and recent computational results have shown that for turbulent flows at low and moderate Mach numbers, an increase in polynomial order ($p$-refinement) is more effective in capturing turbulence characteristics, while for highly compressible turbulence regimes, $h$-refinement becomes beneficial for representing shows and flow-discontinuities. Significant progress has been made on extending DG-methods for reacting turbulent flows~\cite{LV_IHME_JCP2014,LV_IHME_PCI35,LV_IHME_AMS2017,WU_MA_LV_IHME_AIAA2017}. However, in order to fully utilize the potential of these high-order variational methods for simulating turbulent flows~\cite{WANG_ETAL_IJNMF2013,GASSNER_BECK_TCFD2013,DEWIART_HILLEWAERT_ETAL_IJNMF2015,LV_MA_IHME_CF2018}, open research issues remain regarding the formulation of stabilization techniques, the development of combustion-physical models, and the construction of subgrid-closures that are consistent with the high-order discretization for large-eddy simulations.
\begin{tcolorbox}[colback=myCol,breakable]
Recommendations:
\begin{itemizePacked}
\item Develop methods to characterize and separate contributions of numerical discretization errors, physical models, and filtering operators.
\item Develop novel analytical methods for examining the impact of numerical discretization for non-linear PDEs.
\item Establish robust measures to quantify mesh dependence on numerical simulations.
\item Explore and quantify the merit of high-order discretization techniques for application to turbulent reacting flows.
\end{itemizePacked}
\end{tcolorbox}
\section{\label{SEC_MODEL_ACCURACY}How to assess the simulation accuracy?}
Directly connected to the performance evaluation of a combustion model is the quantification of the agreement with measurements and between models. So far, such comparisons have been largely performed by considering individual scalar quantities. These evaluations follow conventional statistical analysis in which moments (typically mean and root-mean-square) and conditional data are compared along axial and radial locations in the flame. Qualitative comparisons of scatter data are commonly performed to examine as to whether a particular combustion model is able to capture certain combustion-physical events such as reignition, extinction, or the departure from thermochemical equilibrium.
Another issue towards the quantitative comparison of simulation results with measurements is the multitude of experimental techniques for data acquisition, which include single-point data, line measurements, line-of-sight absorption, or multidimensional imaging at acquisition rates ranging from single-shot to high-repetition rate measurements to resolve turbulent dynamics~\cite{ALDEN_BOOD_LO_RICHTER_PCI2011}. This data is then processed in the form of statistical results from Favre and Reynolds averaging, conditional data, probability density functions, and scatter data.
Different methods have been proposed to measure model errors, and most common is the linear scalarization and comparison of single-point statistics. Since thermochemical species in chemically reacting flows are strongly coupled, the comparison of individual quantities neglects these interdependencies, often showing that major species and temperature exhibit good qualitative agreement while sensitive quantities and minor species exhibit significant deficiencies. As such, the comparison of individual scalar quantities limits a holistic examination of interscalar dependencies and the identification of correlations between reacting and hydrodynamic flow-field quantities. This has the consequence that these comparisons often only provide an inconclusive assessment of the model performance, and limit a quantitative comparison among different modeling approaches.
Recently, the Wasserstein metric was introduced as a generalized measure for the quantitative evaluation of combustion models~\cite{JOHNSON_WU_IHME_CF2017}. Compared to commonly employed techniques that consider low-order statistical moments, this probability metric is formulated in distribution space, thereby enabling the direct consideration of instantaneous data that are obtained from transient simulations and high-speed measurements without the need for data reduction to low-order statistical moments. The Wasserstein metric and related probabilistic measures~\cite{GIBBS_SU_ISR2002}, such as the earth mover's distance, Kullback-Leibler divergence, or the Kolmogorov metric, represent frameworks for combining different data that originate from scatter plots, instantaneous simulation results, or the reconstruction of statistical moments in the form of empirical distributions. Compared to other metrics, an attractive feature of the Wasserstein metric is that it is equipped with essential properties of metric spaces.
\nomenclature{$W_2$}{Wasserstein metric}
The Wasserstein metric has been employed to assess different modeling approaches for simulating turbulent flames~\cite{JOHNSON_WU_IHME_CF2017,WU_MA_LV_IHME_AIAA2017,BARLOWTNF2018}. Representative results are illustrated in~\cref{FIG_WSM}, showing comparisons of the multiscalar Wasserstein metric for QoIs of mixture fraction, temperature, and species mass fractions of \ce{CO2} and \ce{CO} in simulations of a turbulent \ce{CH4}/air jet flame with inhomogeneous inlets. In this LES study, a flamelet model in conjunction with a regularized deconvolution model was employed to examine the impact of successive mesh-refinement on convergence~\cite{WANG_IHME_CF2019}. Convergence analysis of the Wasserstein metric with respect to mesh refinement provides a quantitative evaluation of the performance of a particular combustion model and the impact of subgrid closures. Different granularities of this metric in the form of planar data, axial comparisons, and radial profiles allow for systematically assessing simulation results in predicting user-specific QoIs. As such, the Wasserstein metric allows for the quantitative evaluation of the influence of boundary conditions and mesh resolution, for isolating regions of physical complexity that demand further experimental probing, for identifying model deficiencies, and for guiding the selection of modeling strategies to meet user-specific requirements on simulation accuracy and computational cost.
\begin{figure}[!htb!]
\centering
\includegraphics[width=0.55\textwidth, clip=, keepaspectratio]{wassersteinMetricSydneyFlame.pdf}
\caption{\label{FIG_WSM}Quantitative evaluation of multiscalar Wasserstein metric~\cite{JOHNSON_WU_IHME_CF2017}, $W_2(Z,T,Y_{\ce{CO2}}, Y_{\ce{CO}})$ LES-calculations of the piloted turbulent \ce{CH4}/air jet flame with inhomogeneous inlets~\cite{BARLOW_MEARES_MAGNOTTI_CUTCHER_MASRI_CF2015} as a function of successive mesh refinement~\cite{WANG_IHME_CF2019} (Mesh I: $26\times10^5$, Mesh II: $207\times10^5$, Mesh III: $1.66\times10^6$, Mesh IV: $13.2\times10^6$, and Mesh V: $53\times10^6$ control volumes). The four bar-graphs from each contribution correspond to axial locations of $x/D=\{1,5,10,15\}$. This quantitative validation analysis enables models to be compared objectively.}
\end{figure}
While the Wasserstein metric targets assessing the accuracy in predicting QoIs, metrics for evaluating the accuracy in predicting transient combustion processes requires the consideration of the dynamics of a simulation. To this end, the Lyapunov exponent has been suggested as a possible metric for assessing the chaotic dynamics of LES calculations~\cite{NASTAC_LABAHN_MAGRI_IHME_PRF2017}. The Lyapunov exponent, $\lambda$, is amenable to a simple physical interpretation: If a system is chaotic, given an infinitesimal initial perturbation to the solution, two trajectories of the system separate in time exponentially until nonlinear saturation. The average exponential separation is the Lyapunov exponent. As such, the Lyapunov exponent provides a convenient measure of the dynamic nature in turbulent flows, and its reciprocal is related to the predictability horizon of a chaotic solution. The Lyapunov exponent is (i) a robust indicator of chaos, (ii) a global quantity describing the strange attractors, and (iii) relatively simple to calculate~\cite{NoninfinitesimalPerturbations,Boffetta2002}. The Lyapunov exponent can be employed in transient simulations where a statistically stationary flow is not present. The calculation of the Lyapunov exponent can be performed on arbitrary meshes or geometries and incorporates information about the numerical discretization, combustion models, and contains non-local information about the turbulence and boundary conditions. Unlike the Wasserstein metric, a limitation of the Lyapunov exponent is that it does not exhibit metric properties, so that theoretical information about asymptotic limits and convergence rates is not available and specific to particular applications of interest.
\nomenclature{$\lambda$}{Lyapunov exponent}
\begin{tcolorbox}[colback=myCol,breakable]
Recommendations:
\begin{itemizePacked}
\item Quantitative methods for assessing the accuracy of simulation results are needed. The Wasserstein metric and related probabilistic measures offer compact representations of model performance at different levels of granularity for objectively analysis of simulation results. Other methods that improve the fidelity, robustness, and convergence are desired that broaden the usability of these analysis methods.
\item The Lyapunov exponent provides a convenient method for evaluating the dynamic content and predictability horizon of a simulation; further developments on quantifying the accuracy in assessing the prediction of dynamical processes are needed that overcome deficiencies of global exponents.
\item The development of methods are necessary that address the lack of measures for the quantitative assessment of time-dependent processes; methods such as proper orthogonal decomposition, dynamic mode decomposition, or Koopman operator provide viable directions for combustion applications~\cite{HOLMES_LUMLEY_BERKOOZ_ROWLEY_BOOK2012,SCHMID_JFM2010,TAIRA_ETAL_AIAAJ2017,ROWLEY_DAWSON_ARFM2017}.
\end{itemizePacked}
\end{tcolorbox}
\section{\label{SEC_AUG_SIM}How can simulations be augmented with experimental data?}
The quantitative prediction of user-specific quantities and processes of interest within prescribed accuracy requirements introduces significant challenges. These challenges arise from the chaotic nature of turbulent reacting flow systems, uncertainties in thermochemical properties, reaction rates, constitutive relations, incomplete specifications of boundary and initial conditions, numerical errors that are introduced through the discretization, and closure models to represent the turbulence/chemistry interaction, turbulent scalar fluxes, and subgrid contributions that evolve at the unresolved scales. While progress in any of these aspects will have a direct impact on our ability to predict turbulent flows, enormous advances and innovative approaches are required to resolve all of these aspects.
The enrichment of simulations with data either from experiments or detailed simulations can hereby help to constrain deficiencies of current simulation approaches. In particular, remarkable advances on the development and utilization of multi-dimensional high-speed measurement techniques has provided quantitative information about instantaneous velocity fields, temperature, and species at acquisition rates and spatial resolutions that are adequate for resolving relevant turbulent scales. Although these measurements have been used in evaluating combustion models, this has been done in the context of \emph{a posteriori} comparisons of statistical results.
Data assimilation provides interesting opportunities for integrating measurements into numerical simulations~\cite{EVENSEN_BOOK2009}, and has been used extensively to produce initial conditions for weather prediction models and to provide representations of the spatio-temporally evolving atmospheric state~\cite{Kalnay2003}. Assimilation techniques enable the estimation of the state of a complex, dynamical system by combining incomplete and sparse experimental data with erroneous models. Data assimilation can spread information from observations in space and time to unobserved state quantities, filter the effect of random observation noise from state estimates, provide estimates of observation and model errors, and determine unknown or uncertain model parameters.
Different assimilation techniques have been developed~\cite{ASCH_BOCQUET_NODET_BOOK2016,LABAHN_WU_HARRIS_CORITONE_FRANK_IHME_FTC2020}, they can be categorized into variational techniques such as nudging, adjoint methods, and 3D/4D-Var, and statistical techniques such as optimal interpolation and the class of Kalman-filtering methods. While variational techniques rely on optimal control theory to minimize a deterministic cost function, statistical techniques incorporate stochastic information in the form of probability distributions of model uncertainties and observation errors. In particular, the ensemble Kalman filter offers advantages for incorporating quantities not contained in the solution vector, the applicability to large-scale problems, and the robust evaluation of the error covariance.
\begin{figure}[!htb!]
\centering
\includegraphics[width=0.8\textwidth, clip=, keepaspectratio]{enkfJetFlame2.pdf}
\caption{\label{FIG_ENKF}Assimilation of simultaneous measurements into simulation of a turbulent jet flame, showing sequence of hydroxyl-LIF signal (left panel) and axial velocity (right panel) at four distinct simulation times. Top row: baseline simulation without assimilation, middle: experimental data, and bottom: results after assimilation step. Black box indicates the location of the measurement window. Figure adapted from~\cite{LABAHN_WU_CORITON_FRANK_IHME_PCI2019}.}
\end{figure}
Because of its generality, data assimilation provides interesting opportunities for turbulent-combustion applications in several ways. Specifically, assimilation can be employed for incorporating measurements into simulations under consideration of experimental uncertainties, sparsity in the data since measurements are only collected at specific locations and sample rates, and measurements are only obtained for a subset of quantities. This is illustrated in~\cref{FIG_ENKF}, where an ensemble Kalman filter (EnKF) was employed to assimilate simultaneous tomographic PIV/OH-PLIF measurements into LES with the goal of capturing local extinction and reignition events~\cite{LABAHN_WU_CORITON_FRANK_IHME_PCI2019}. While the application of data assimilation for state estimation is a common approach for atmospheric-flow predictions, forecasting has only limited value for combustion simulations. The main reasons for this are offline measurements and the current lack of real-time simulations and analysis tools. More specifically, weather-forecasting is concerned with the prediction of changes in the weather over several hours and significant efforts have been made to achieve real-time simulations. In contrast, reacting flows evolve on significantly shorter time-scales on the order of few milliseconds, and it remains illusive to capture this time horizon in real-time using existing LES-modeling capabilities. Nevertheless, state estimations can be employed for complementing experimental observations with simulations to uncover combustion-physical processes or for spinning up numerical simulations. Another attractive opportunity is to utilize data assimilation for parameter estimations and model evaluation. The merit of this concept was demonstrated in examining the performance of a flamelet-based combustion model in capturing extinction events~\cite{LABAHN_WU_CORITON_FRANK_IHME_PCI2019}, showing that the model was required to significantly reduce the mixing rate and attenuate the reactivity in order to reproduce the experimentally observed behavior of the local extinction and reignition sequence (see~\cref{FIG_ENKF}).
\begin{tcolorbox}[colback=myCol,breakable]
Recommendations:
\begin{itemizePacked}
\item Explore data assimilate techniques to integrate experimental measurements into simulations for state estimation, parameter estimations, and model evaluations.
\item Extend data assimilation methods to turbulent combustion applications by addressing aspects on scalar boundedness, ensemble density, and non-Gaussianity.
\item Utilize experimental data and simulations for physics-informed learning, to enhance physical models, and to support model classification.
\end{itemizePacked}
\end{tcolorbox}
\section{\label{SEC_CONCL}Summary and Conclusions}
This paper examines requirements towards improving the quantitative capability of current simulation and analysis tools for predicting turbulent reacting flows. Throughout this paper, attempts are made to stipulate further development by offering recommendation for further developments. To this end, aspects are identified that are concerned with the selection of specific simulation approaches and combustion models to achieve reliable predictions of response function of interest to the user under consideration of requirements on accuracy and computational resource allocation. Aspects on the interaction between numerical discretization and physical models are addressed; high-order methods and explicitly filtered LES strategies are discussed as promising methods for mitigating and controlling numerical discretization errors and for separating the impact of numerical discretizations from physical modeling. Related to the evaluation of numerical models, we introduce probabilistic and dynamic analysis tools in the form of the Wasserstein metric and the Lyapunov exponent, respectively, in order to support the quantitative analysis of solution accuracy and the estimation of the predictability horizon and dynamic behavior in unsteady combustion simulations; we highlight limitations of dynamic metrics in quantifying transient combustion processes. With the increasing availability of high-resolution, time-resolved, and simultaneous measurements, we examine emerging opportunities for utilizing this data to improve computational models, to obtain reliable state estimates, and to identify model deficiencies.
\section*{Acknowledgments}
Acknowledgment is made to NASA (award NX15AV04A), AFOSR (Award FA9300-19-P-1502), and the donors of The American Chemical Society Petroleum Research Fund for support of this research. Helpful discussions with Hao Wu, Peter Ma, Qing Wang, Jeff Labahn, and Alan Kerstein are gratefully acknowledged.
| 4ec55ce17832268e4b158fd813583680bb3f1237 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Grasping is one of the important actions in people's daily life, but the limbs of some disabled people cannot implement the grasping function because of paralysis, amputation, etc. Wearable robots including exoskeletons~\cite{dollar2008lower}, prosthetics~\cite{belter2013mechanical}, and supernumerary robotic fingers~\cite{wu2015hold, hussain2016soft, ciullo2018analytical} can serve as alternative actuators.
Recognition of grasping intention is the premise for controlling wearable robots. The current main methods include EMG-based methods~\cite{salvietti2016compensating, marco2017surface}, EEG-based methods~\cite{penaloza2018bmi, downey2018intracortical}, and action-based methods~\cite{huang2015control, de2015recognizing}. Among these methods, the EMG-based method is not suitable for patients with hemiplegic paralysis, because the EMG signals of affected limbs have been distorted, from which it is difficult to identify the intention of patients. The accuracy and discriminability of EEG signals obtained by the noninvasive brain-machine interface are relatively low, and it is easy to be interfered by external factors. Although the invasive interface can obtain accurate EEG signals, most patients cannot accept craniotomy. The action-based methods use sensors such as airbags and IMUs to obtain the action states of the user's limb to infer the user's intention, but these methods are impossible for those patients who have lost their locomotivity.
Eye-tracking technology can be the feasible means of grasping intention recognition for such patients. The use of eye movement as a control interface relies on the fact that our gaze is proactive and directly correlated to action intentions and cognitions~\cite{hayhoe2005eye}. Furthermore, the oculomotor system control function is usually retained even in the most severe cases of paralysis related to muscular dystrophies or brain stroke. And in neurodegenerative diseases that affect the motor system (such as stroke, Parkinson's disease, or Multiple Sclerosis), it is easier for patients to control their gaze point rather than conducting stable skeletal movements~\cite{cipresso2011combined}.
Most of the studies using eye movement technology only focus on reaching tasks, e.g., Roni et al. used gaze controlling robotic arms to assist human hands in reaching tasks~\cite{maimon2017towards} , and Shaft et al. used eye movement controlling robotic arms to perform tasks\cite{shafti2019gaze}. Other studies have adopted unnatural methods to recognize the grasping intention of patients, such as asking users to deliberately blink or prolong the gaze, which will increase the cognitive load on users. It is not easy for patients with nerve damage to complete these unnatural actions.
In these case, natural human-robot interaction can be used to recognize the patients' intention. Natural human-robot interaction uses natural action to obtain the user's intention~\cite{ferland2013natural, valli2008design}, without imposing extra cognitive load on the users. This is similar to the user's control of their own limbs for activities without long-term training.
Previous studies have shown that there are differences in natural gaze between grasping and viewing behaviours~\cite{brouwer2009differences}. However, these studies focus on the effects of different objects on the first two fixations. To the best of our knowledge, there are only qualitative summaries of differences from experimental observations but no quantitative features that can be used for grasping intention recognition reported in existing literatures.
On this basis, we study the relationships of the fixations in the grasping tasks and viewing tasks when the thumb and index finger are visible, which is a common situation while humans are conducting grasping tasks. We analyze these relationships and extract some quantitative features to identify gasping intention.
Using the obtained features, we recognize the user's grasping intention and verify the effectiveness of the method. Our main contributions are as follows:
\begin{enumerate}[]
\item We discover the relationships between the fixation selves, the fixations and objects, the fixations and grasping points in the grasping tasks and the viewing tasks. Compared with the viewing tasks, the fixations of the user are more centralized and tend to be close to the grasp point of the index finger in the grasping task.
\item We propose four features from the relationships for grasping intention recognition. Then we compare the accuracy of different features in recognition of grasping intention. On this basis, we implement a natural method to recognize the grasping intention by eye movement.
\item We carry out grasping experiments with humanoid robotic arms in actual daily scenes, which verifies the effectiveness of the proposed method.
\end{enumerate}
\section{Methods}
\subsection{System overview and architecture}
We aim to create a system to recognize grasping intention based on human natural gaze data. The system consists of a binocular eye-tracker, a scene camera, a head stand and a black background (see Fig.~\ref{setup}). The block diagram in Fig.~\ref{block} depicts an overview of our system. Individual modalities are described in detail in the following.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{1-1.png}
\caption{The block diagram of our system. The user's fixation positions are obtained by the eye-tracker, the object information is obtained by the image processing algorithms, and the grasping position is obtained by pre-experiment. The features proposed in this paper are extracted from these data and used to train the intent recognition model. In the task of intent recognition, the humanoid robot arm is controlled to complete the grasping action according to the model recognition result.}
\label{block}
\end{figure*}
\subsubsection{Binocular eye-tracker}
For 2D eye tracking, we use the Pupil Core~\cite{kassner2014pupil}.
The eye-tracker is built from two infrared (IR) cameras used to photograph the eyes, yielding a resolution of 192 $\times$ 192 pixels at a maximum frame-rate of 120 Hz. It has two extendable mounts for the infrared cameras, which can be adjusted to focus the IR cameras on each eye respectively for individuals. There are two IR LEDs attached to the cameras to illuminate the pupil of the user. Using the obtained infrared images, we can select the appropriate eye model and apply ellipse fitting parameters during system operation to detect the pupil of the user, and then use this information to estimate the user's fixation point.
An Intel Realsense D435 RGB-D camera (Intel Corporation, Santa Clara, California, USA) is mounted on the top of the eye-tracker as a scene camera, with a 3D printed frame. This camera is used to capture the user's ego-centric scene.
\subsubsection{Object centroid detection}
In the experiment environment, we use the OpenCV image processing library~\cite{kaehler2016learning} to process live scene video streams and detect the object's centroid in each frame with a contour-based algorithm.
To avoid the interference from ambient light, we use a solid color background. Besides, we filter the centroid of some disturbing objects based on the area of the object contour. Because it is impossible to detect the real centroid of 3D objects in RGB images, we use the 2D centroid of objects in images instead.
In real environment applications, we use Mask R-CNN~\cite{he2017mask} to quickly and accurately obtain object information. This machine learning-based method can effectively improve the robot system's ability to perceive the environment and identify objects that patients need to grasp in their daily lives.
\subsubsection{Grasp position detection}
In grasping tasks the thumb and the other fingers play different roles. The thumb guides the hand straight to the object's location, and then the other fingers represented by the index finger grip around the object to ensure a safe grasp~\cite{haggard1997hand}. We record the positions of the participants' digits to obtain the relationship between the fixation position and the grasping position. The grasping positions are obtained through pre-experiment.
In the preliminary experiment, we ask participants to grasp each object in repeated fashion and use the average positions as the grasping positions of the thumb and index finger.
\subsection{Experiment}
There are two categories of gaze behaviour: fixations and saccades. Fixations are those times when our eyes essentially stop scanning about the scene, holding the central foveal vision in place so that the visual system can take in detailed information about what is being looked at~\cite{rayner200935th}.
The main goal of the experiment is to investigate the difference of fixation modes of the participants under different task conditions of grasping and viewing.
Eight different participants carry out both viewing and grasping tasks. Their ages range between 20 and 30. Prior to the experiments, a short introduction is provided to these subjects, including technologies involved, the system setup and the research purpose. During the experiment, a head stand is used to fix the subjects' head. In practical applications, a head tracking module can be added to track the movement of the use's head, which is not covered in this paper.
\subsubsection{Procedure}
The shapes that the participants are asked to grasp or view are magnetically attached to a black background, which is placed in front of an iron plate (Fig.~\ref{setup}). We use three different white plastic objects for training, depicted in Fig.~\ref{setup}(b): a square, a cross, a T-shape (presented in four different orientations).
They are 10 mm thick.
There are two types of grasping tasks, horizontal grasping and vertical grasping. The vertical T-shape is used for vertical grasping tasks, while the horizontal T-shape is used for horizontal grasping tasks. The square and cross shapes are used in both grasping tasks.
For each type of the grasping task, participants perform 4 shapes * 5 repetitions = 20 practice trials, presented in random order.
For the viewing task, participants perform the same number of trials.
We use two other shapes: a triangle and a bar (presented in 2 different orientations) for testing.
\begin{figure*}[thbp]
\centering
\subfigure[]{
\begin{minipage}[t]{0.45\linewidth}
\centering
\includegraphics[width=1\linewidth,height=3.8cm]{3-1.png}
\end{minipage}%
}%
\subfigure[]{
\begin{minipage}[t]{0.45\linewidth}
\centering
\includegraphics[width=1\linewidth]{3-2.png}
\end{minipage}%
}%
\centering
\caption{The experimental scene and target shapes. (a) Schematic depiction of the setup and the way of grasping in the experiment. (b) Different shapes in the experiment.}
\label{setup}
\end{figure*}
The eye-tracker needs to be calibrated before the experiment starts. Participants rest their head on a head stand, 50 cm in front of the monitor, which is the same distance as in the experiment. Both eyes are calibrated using a nine point calibration/validation procedure on the computer monitor. After the participants have performed half trials, the calibration is repeated. At the start of each trial, the participants fix at the start point and closes their eyes after fixation.
For grasping tasks, participants start a trial with the eyes closed and the hands resting on the lap. When hearing a ``grasp'' command, they open their eyes and immediately fix their eyes on the start point. Participants then stretch out their thumb and index finger to grasp the shape (Fig.~\ref{setup}(a)). Different grasping tasks have different grasping modes. The user's fingers of the horizontal grasping tasks contact with the left and right edges of the object, while the fingers of the vertical grasping tasks contact with the top and bottom edges.
The trial is ended when the ``finish" command is heard. After completing the trial, the participants return to the start point with their hands on the lap and their eyes closed, until hearing the verbal command to start next trial.
The procedure of the viewing task is similar to the grasping task, with the exception that instead of being asked to grasp, participants are asked to ``view the shape'' until hearing a "finish" command. The viewing process lasts for three seconds. The duration is measured by the participants through the pre-experiment. In the experiment, participants perform 10 grasping tests repeatedly and we select the maximum time of task execution as the benchmark time in the experiment.
\section{Data analysis}
Among 640 grasping and viewing trails, the data of 16 grasping trails (5\%) and 24 viewing trails (7.5\%) are rejected due to insufficient quality of eye data. We use the fixation detection algorithm to distinguish fixations from fast saccades. The parameters of the algorithm recommended by the manufacturer are set as follows:
\begin{equation}
\begin{aligned}
&dp{s_{\max }}{\rm{ = 3}}{\rm{.01}},~du{r_{\min }}{\rm{ = 80}}ms,~du{r_{\max }} = 400ms
\end{aligned}
\end{equation}
$dps_{max}$ indicates the threshold of the distance between all gaze locations during a fixation. $dur_{min}$ indicates the minimum duration in which the distance between every two gaze locations must not exceed $dps_{max}$ and $dur_{max}$ indicates the maximum duration.
After processing the gaze information, the two-dimensional coordinates of the fixation points of participants in the trial are obtained. The coordinates are the relative positions of the fixations to the centroid of the shape.
We compare the result of the grasping and viewing tasks through repeated measures analysis of variance (ANOVAs), taking the objects as within-subjects variable and the tasks as between-subjects variable. For all statistical tests, we use 0.05 as the level of significance.
\begin{table*}[thbp]
\centering
\caption{Significant results for two kinds of grasping tasks and viewing tasks repeated measures ANOVAs on the ADF2C, the ADF2I, the ADF2T, and the VAR}
\begin{tabular}{lllll}
\toprule
\textbf{} & \textbf{ADF2C} & \textbf{ADF2I} & \textbf{ADF2T} & \textbf{VAR} \\ \hline
\textbf{Vertical grasping and viewing} & \textit{F}=0.046, \textit{p} = 0.829 & \textit{F}=142.699, \textit{p} $<$0.001 & \textit{F}=5.699, \textit{p}=0.0176 & \textit{F}=284.699, \textit{p} $<$0.001 \\ \midrule
\textbf{Horizontal grasping and viewing} & \textit{F}=3.781, \textit{p} = 0.053 & \textit{F}=192.511, \textit{p} $<$0.001 & \textit{F}=164.964, \textit{p} $<$0.001 & \textit{F}=171.128, \textit{p} $<$0.001 \\ \midrule
\textbf{Grasp and viewing} & \textit{F}=3.209, \textit{p} = 0.074 & \textit{F}=270.596, \textit{p} $<$0.001 & \textit{F}=80.568, \textit{p} $<$0.001 & \textit{F}=437.500, \textit{p} $<$0.001 \\ \midrule
\end{tabular}
\label{ANOVAS}
\end{table*}
\subsection{Number of fixation}
Table~\ref{fixation} shows the average fixation times in one trial. On average, there are 8.18 (\textit{STD}=0.99) fixation points for each grasping task and 8.62 (\textit{STD}=0.76) fixations for each viewing task. The results of repeated measures ANOVA show that the task type affects the number of fixations (\textit{F}=270.60, \textit{p}$<$0.001). But in one trial, the number of fixations will only be an integer, which makes it difficult to distinguish between grasping and viewing by the number of fixations.
\begin{table*}[thbp]
\centering
\caption{Fixations in different trials}
\begin{tabular}{cccccccc}
\toprule
\textbf{Shape} & \textbf{square} & \textbf{cross} & \textbf{cross\_down} & \textbf{cross\_up} & \textbf{cross\_right} & \textbf{cross\_left} & \textbf{mean} \\ \midrule
\textbf{Viewing} & 8.71$\pm$0.83 & 8.35$\pm$0.70 & 8.60$\pm$0.89 & 8.57$\pm$0.70 & 8.08$\pm$0.87 & 8.62$\pm$0.66 &8.62$\pm$0.80 \\ \midrule
\textbf{Grasping} & 8.28$\pm$1.09 & 8.11$\pm$0.69 & 8.64$\pm$1.11 & 8.23$\pm$0.70 & 8.08$\pm$0.36 & 8.08$\pm$0.36 &8.18$\pm$0.86 \\ \midrule
\end{tabular}
\label{fixation}
\end{table*}
\subsection{Fixations locations}
Fig.~\ref{mean_location} illustrates the distribution of the fixation location by showing the fixations in certain trials in the grasping and viewing tasks respectively. The centroid of the shapes is represented by the blue point. We can find that fixations have different characteristics under different tasks. To effectively distinguish between grasping and viewing, we propose some features and quantitatively analyze their relationship in different tasks.
In one trial, the user's fixations are expressed as:
\begin{equation}
F({t_i}) = ({f_x}({t_i}),{f_y}({t_i}))
\label{fix_coordinate}
\end{equation}
$i$ represents the number of fixation, and $t_i$ represents the time corresponding to fixation $i$. The two-dimensional centroid coordinates of the object are expressed as follows:
\begin{equation}
C(t) = ({c_x}(t),{c_y}(t))
\label{cog_coordinate}
\end{equation}
where $t$ represents the time of coordinates. The coordinates of the user's thumb and index finger grasp points in the egocentric video are expressed as:
\begin{equation}
\begin{aligned}
{G_1}(t) = (g{1_x}(t),g{1_y}(t))\\
{G_2}(t) = (g{2_x}(t),g{2_y}(t))
\label{grasp_coordinate}
\end{aligned}
\end{equation}
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{mean_location.png}
\caption{The fixation position of participants on four shapes in certain trials. The centroid of the shapes is represented by the blue point.}
\label{mean_location}
\end{figure}
\subsubsection{The average distance from the fixations to the centroid}
The first feature we propose is the average distance from the user's fixation points to the centroid of the shape (ADF2C). This feature is used to describe the relationship between the participants' fixations and the object:
\begin{equation}
\begin{aligned}
ADF2C = \frac{1}{n}\sum\limits_{i = 1}^n {\sqrt {{{({f_x}({t_i}) - {c_x}({t_i}))}^2} + ({f_y}({t_i}) - {c_y}{{({t_i})}^2}} }
\label{ADF2C_equ}
\end{aligned}
\end{equation}
If the detected centroid does not belong to the user's target object, this feature will appear abnormal (too large), which means that the user's target object can be determined by this feature.
Fig.~\ref{dis_cog} shows the distance between the fixations and the centroid of the shape in grasping and viewing, including the mean value of the distances. Obviously, the curves of ADF2Cs intertwine in the grasping and viewing tasks, and their average values are so close that there is no significant difference.
A repeated measures ANOVA on the ADF2C with the tasks as variable indicates that the ADF2C does not differ between grasping and viewing (\textit{F}=3.20, \textit{p} = 0.07).
\begin{figure}[thp]
\centering
\includegraphics[width=0.5\textwidth]{distance_to_COG.png}
\caption{The ADF2Cs in all trials. The solid line represents the result of each trial, and the dotted line represents the average of all results. }
\label{dis_cog}
\end{figure}
\subsubsection{The average distance from the fixations to the grasping point}
We find that there are also different relationships between the fixations and the grasp position in different tasks. The grasping positions of the participants are obtained from the pre-experiment. We consider using the average distance between the fixations and the grasping positions in each trial to reflect these relationships.
These two features are the average distance from the fixations to the index finger (ADF2I) and the average distance from the fixations to the thumb (ADF2T):
\begin{equation}
\begin{aligned}
ADF2T = \frac{1}{n}\sum\limits_{i = 1}^n {\sqrt {{{({f_x}({t_i}) - g{1_x}({t_i}))}^2} + ({f_y}({t_i}) - g{1_y}{{({t_i})}^2}} } \\
ADF2I = \frac{1}{n}\sum\limits_{i = 1}^n {\sqrt {{{({f_x}({t_i}) - g{2_x}({t_i}))}^2} + ({f_y}({t_i}) - g{2_y}{{({t_i})}^2}} }
\label{ADF2TI_equ}
\end{aligned}
\end{equation}
Fig.~\ref{dis_grasp} shows the relationship between ADF2I and ADF2T in all trials, respectively. We can find that ADF2I and ADF2T are obviously different in grasping and viewing tasks.
Furthermore, the mean value of ADF2I is significantly lower than that of ADF2T in grasping tasks.
The reason is that in the grasping tasks, the participants' eyes focus on the index finger more than the thumb, which may help stabilize the grasps. A repeated measures ANOVA on the ADF2I shows a significant effect of the tasks (\textit{F}=270.60, \textit{p}$<$0.001). Similar to ADF2I, ADF2T is also affected by the task(\textit{F}=80.57, \textit{p}$<$0.001). Table~\ref{ANOVAS} reveals that this significant effect is more obvious in horizontal grasping tasks than in vertical grasping tasks (the value of \textit{F} indicates the extent of the effect).
\begin{figure}[thp]
\centering
\includegraphics[width=0.5\textwidth]{distance_to_indexandthumb.png}
\caption{The average distance from the fixation to the grasping point. The solid line represents the result of each trial, and the dotted line represents the average of all results.}
\label{dis_grasp}
\end{figure}
\subsubsection{The relationship between fixation points}
We also find that in one trial, there is a relationship between the fixations. We first propose the variance (VAR) of the distances from all the fixation points to the center as a quantitative description of concentration:
\begin{equation}
\begin{aligned}
&O = ({o_x},{o_y}) = (\frac{1}{n}\sum\limits_{i = 1}^n {{f_x}({t_i})} ,\frac{1}{n}\sum\limits_{i = 1}^n {{f_y}({t_i})} )\\
&{d_i} = \sqrt {{{({f_x}({t_i}) - {o_x})}^2} + {{({f_y}({t_i}) - {o_y})}^2}}\\
&M = \frac{1}{n}\sum\limits_{i = 1}^n {{d_i}}\\
&VAR = \frac{{\sum\limits_{i = 1}^n {{{({d_i} - M)}^2}} }}{n}
\end{aligned}
\end{equation}
It can be seen from Fig.~\ref{var} that the VARs in the grasping tasks are significantly smaller than those in the view tasks. In the grasping tasks, the mean value of the VARs is 29.85. \textit{(STD}=44.08), which is much smaller than the viewing tasks (\textit{MEAN}=256.67, \textit{STD}=181.05).
This can be explained by the fact that in the grasping tasks, the participant's attention and sight need to be more concentrated to complete the grasp, and the participant's sight changes randomly across the target during the viewing task.
A repeated measures ANOVA on the VARs show a significant effect of the tasks (\textit{F}=437.50, \textit{p}$<$0.001). This effect is similar in horizontal grasping and vertical grasping.
\begin{figure}[thp]
\centering
\includegraphics[width=0.5\textwidth]{var_of_fixation.png}
\caption{The VARs in all trials. The solid line represents the result of each trial, and the dotted line represents the average of all results.}
\label{var}
\end{figure}
Table~\ref{ANOVAS} shows all significant results of the repeated measures ANOVA on the ADF2C, the ADF2I, the ADF2T, and the VAR.
\section{Grasping intention recognition}
From the above data analysis, we can know that in the viewing and grasping tasks, participants' fixations have different features, which show the potential of accurately estimating the participants' grasping intention. Thanks to the use of human natural eye behavior, people's cognitive load is reduced and can be quickly accepted by users. This method is especially suitable for patients with upper limb injuries. They can use gaze data to control and express their grasping intention, instead of relying on residual EMG or movement.
To evaluate the effect of different features on grasping intention recognition, we selected five feature combinations.
1) Combination1:ADF2T+VAR
2) Combination2:ADF2I+VAR
3) Combination3:ADF2C+ADF2T+VAR
4) Combination4:ADF2C+ADF2I+VAR
5) Combination5:ADF2C+ADF2I+ADF2T+VAR
Several different machine learning methods are used, including support vector machine (SVM), K-Nearest Neighbor (KNN), stochastic gradient descent (SGD), and decision tree (DT).
The total samples are randomly partitioned into 5 equal-sized subsets. Among the 5 subsets, four subsets are used as the training set while the remaining one is used as the testing set (Test1). For each methods,we use the 5-fold cross-validation to evaluate the accuracy of recognition. The other test set (Test2) is collected from three other shapes that does not exist in the training set, each of which contains 20 samples.
From Fig.~\ref{acc} we can learn that
for different feature combinations, Combination3 has the highest average accuracy on Test1, which is 89.7\%$\pm$1.9\%. The average recognition accuracy of the other four feature combinations on Test1 also exceeds 88.8\%.
On Test2, Combination2 has the highest average recognition accuracy of 90.6\%$\pm$4.5\%. The average recognition accuracy of other feature combinations is also acceptable, the lowest of which is higher than 87\%.
For different classifiers, the KNN has the highest average accuracy on Test1, which is 91.6\%$\pm$0.3\%. The SGD has the lowest grasping intention recognition accuracy on Test1, with an average of 85.1\%$\pm$2.5\%. But this is also an acceptable result, which means that the influence of classifiers is small.
On Test2, the SVM (linear) achieves the best accuracy of 93.6\%$\pm$0.3\%. The accuracy of other classifiers also exceeds 82\%. The reason why the SVM classifier has higher accuracy on Test2 than Test1 may be that Test1 has a longer experimental process. This makes the participants become fatigued and lead to abnormal fixation mode.
From the analysis, we can know that the features we discovered can effectively identify the user's grasping intention and can achieve excellent results with different classifiers. Furthermore, we have successfully applied the proposed feature combinations to the newly appeared objects in the test set to identify the user's grasping intention. These results show that the features we extracted from the gaze data are robust and can reflect human intention, which can be effectively applied to the recognition of users' grasping intention.
\begin{figure*}[htbp]
\centering
\subfigure[]{
\begin{minipage}[bp]{0.5\linewidth}
\centering
\includegraphics[width=1\linewidth]{acc_test1.png}
\end{minipage}
}%
\subfigure[]{
\begin{minipage}[bp]{0.5\linewidth}
\centering
\includegraphics[width=1\linewidth]{acc_test2.png}
\end{minipage}
}%
\centering
\caption{Classification accuracy of each classifier and combinations for grasping intention recognition. We perform 100 repeated experiments and used the average accuracy and standard deviation. (a) The accuracy of Test1. (b) The accuracy of Test2}
\label{acc}
\end{figure*}
\section{Grasping experiment in real environment}
We have carried out the application experiment in the actual daily scene with the humanoid arm. The experimental scene is shown in the Fig.~\ref{realexp}. We chose a cup that is commonly used in daily life as the intended target. We used the Mask R-CNN~\cite{he2017mask} object detection algorithm to extract the target information we need from the real and complex environment. Through the selected algorithm, we can detect the mask of the cup from the user's egocentric scene and use it to calculate the two-dimensional centroid of the object.
We set up a sliding window with a duration of three seconds to sample the data. Then, the features are extracted by combining the centroid of the object, the grasping points with the user's fixation information. At the same time, we use the Combination4 and the KNN as features and the classifier for grasping intention recognition. The experimental results show that, combined with the current advanced image processing algorithms, our proposed method can effectively identify the user's grasping intention and successfully implement the grasp. The complete experimental process is shown in the video in the material.
Four subjects participated in our experiment. To complete the objective evaluation, the subjects were asked to complete the Quebec User Evaluation of Satisfaction with assistive Technology (QUEST) questionnaire~\cite{demers2002quebec}. The results are shown in Table~\ref{question}. From the QUEST questionnaire we find that although there is general agreement with effectiveness, easiness to use, and safety, there is displeasure on comfort. This is probably because the scene camera is relatively heavy and the eye-tracker is not a mature wearable product.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{real_experiment.png}
\caption{Application experiment in the real environment. The system recognizes the user's grasping intention and controls the humanoid robotic arm to complete the grasp.}
\label{realexp}
\end{figure}
\begin{table}[htbp]
\centering
\caption{QUEST evaluation results for the humanoid robotic arm assist system}
\begin{tabular}{ccccc}
\hline
\textbf{} & \textbf{Safety} & \textbf{\begin{tabular}[c]{@{}c@{}}Easy \\ to Use\end{tabular}} & \textbf{Comfort} & \textbf{\begin{tabular}[c]{@{}c@{}}Effec\\ -tive\end{tabular}} \\ \hline
\textbf{Not satified at all} & 0 & 0 & 0 & 0 \\ \hline
\textbf{Not very satisfied} & 0 & 0 & 1 & 0 \\ \hline
\textbf{More or less satisfied} & 1 & 0 & 1 & 0 \\ \hline
\textbf{Quite satisfied} & 3 & 3 & 2 & 2 \\ \hline
\textbf{Very satisfied} & 0 & 1 & 0 & 2 \\ \hline
\end{tabular}
\label{question}
\end{table}
\section{Discussion and Conclusion}
Recognition of grasping intentions is the premise for wearable robots to assist users in grasping, but in some cases, the eye movements are the only intention related activities that can be observed. So we study the eye movements related to grasping.
We obtain the fixation messages of participants in grasping and viewing tasks through experiments. These messages are combined with the target centroid and the grasping positions to study the fixation laws under different tasks. Through the experimental results and statistical analysis, we find that fixations are significantly different in the two tasks.
In the grasping task, the participants' fixations are closer to the grasp position of the index finger and concentrate on a small area. While in viewing tasks, the distribution of participants' fixations are arbitrary, which may be distributed in multiple parts of the target.
This phenomenon has inspired us to use these natural gaze fixations to identify the grasping intention of participants.
We first propose four features- the ADF2C, the ADF2I, the ADF2T, and the VAR to describe these differences quantitatively. Utilizing the quantitative features, we can use machine learning methods to recognize the user's grasp intention, and further control the robot to help the user complete the grasp.
It is crucial that only the natural gaze fixations of the human are used in our method, and no additional actions or commands are required. Thus, this method is very suitable for patients with hemiplegia.
However, not all users' fixations follow the rules we have found. Some users' fixations are close the centroid or the grasp point of the thumb rather than the grasp point of the index finger when grasping. When the user knows the pose and shape of the object in advance, he/she may be able to grasp it without staring at it.
Additionally, the work of this paper explores the relationships of the users' fixations in different tasks when the thumb and index finger are visible.
For some grasping tasks, the contact position of the index finger is not visible. In the next work, we will explore the laws of fixations in grasping and viewing tasks when not all the grasping points are visible. Furthermore, we intend to apply the method to patients with hemiplegia so that they can complete daily grasping tasks.
In conclusion, we are exploiting the clues behind eye movements, which are highly correlated to our motor actions. They are key factors in human motor planning and imply our movement intention. Consequently, gaze tracking becomes a rich source of information about user intent that can be utilized. Fixations reflect the focus of human gaze. We have proved that only a few fixations are needed to quickly and accurately recognize the user's grasping intention. This technology bears a tremendous potential as an interface for human-robot interaction. Moreover, eye-movements are retained in the majority of motor disabilities, which enhances the applicability of eye-tracking technology in the field of human-robot interfaces for the paralysed and in prosthetics. Unlike EMG based and EEG based methods that require long-term training, our method only needs natural eye movement and does not need training. The intention recognition performance of our system are subject to the accuracy of gaze estimation. We hope to improve the accuracy of gaze estimation in the future to achieve more accurate recognition results. Our system allows for straightforward control of the robot for grasping. The application in the real environment has proved its effectiveness. Furthermore, the user's fixations can also provide the location information of the target. If we combine the user's intention with the position information of the object, the user can control the humanoid robotic arm to grasp the object with his eyes only. The ability to recognize the user's grasp intention from fixations enables novel ways of achieving the embodiment in robotic solutions that restore or augment human functions.
\addtolength{\textheight}{-12cm}
\bibliographystyle{IEEEtran}
| 93b6ef7a078c44a8075f31fcfd7039e9e245104d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
Communities around the world are fighting the COVID-19 pandemic. In the United States alone, more than 50 million cases and 800,000 deaths have been recorded \cite{CDCCOVID}. The most reliable way to end transmission of this and other rapidly spreading diseases is to stop any and all contact between individuals in the population; no interaction means no transmission. While this strategy is extreme, many communities (even nations) implemented some version of this plan in the early weeks and months of the COVID-19 pandemic. These guidelines and mandates for social separation (distancing) were largely successful in slowing the spread of the disease \cite{socialDistancing1, socialDistancing2, socialDistancing3}, but they also prevented healthy people from interacting with each other, hindering global economic activity \cite{econ1, econ2} and leaving many feeling isolated and lonely \cite{lonely1, lonely2}.
Reopening schools, businesses and other organizations and returning to a more normal level of interaction is therefore a priority for many communities. However, this does not mean that we must expect and accept community-wide outbreaks; other interventions can be implemented to limit the risk of transmission. Such interventions include (but are not limited to) mask use, vaccinations, viral testing, and contact tracing. This paper quantifies the effect of each of these mitigation strategies on the spread of disease and presents practical guidance for creating policies to maintain low risk of disease outbreak. The analysis that leads to this guidance is performed on a stochastic compartmental model for disease transmission, and the result is a single algebraic criterion (an effective disease reproduction number) based on the implemented interventions. In simulations of infections on random networks (agent-based simulations), this single criterion is able to predict the presence or absence of outbreaks across a vast array of population and disease parameters, whose true values will be unknown to those implementing intervention strategies. The practical guidance produced here was also the foundation of the policies regarding interventions implemented at the Massachusetts Institute of Technology (MIT), which has gradually resumed its normal functioning over the last two years without widespread outbreaks. Policies at other colleges and universities can also be considered in the framework of this analysis, and the outbreak trends at these other institutions are consistent with the analysis presented below given their chosen disease mitigation policies.
\section{METHODS}
\subsection{Modeling foundations}
We begin by applying the principles established by Kermack and McKendrick \cite{SIR_OG} for modeling the spread of infection. This model relies on partitioning the population into different categories (e.g.~Susceptible, Infectious, Recovered or SIR) based on disease status and applying a conservation law on the total population size. The rate of change of the number of people in a particular partition is determined only by the net flow of individuals into or out of that group. Stable dynamics (i.e.~a shrinking infectious population) can be equivalently defined as a positive net flow of people out of the infectious compartment.
Many extensions of the SIR model include additional classifications of individuals \cite{SIQRmodel, SEIRModel2, berger2020seir, ageModel}. In our cases, we partition the population into four groups: susceptible, infectious, isolated, and recovered.\footnote{An exposed/latent group, which separates presymptomatic from symptomatic individuals, is not considered in this study due to the high number of COVID-19 cases that never become symptomatic \cite{COVIDAsymptomatic, asympSpread}. Additionally, by ignoring the development of symptoms (and thus detection), we will end up \textit{overestimating} the number of total undetected infections, leading to more conservative policies to prevent outbreaks.} The choice of partitions and the flows between the compartments are depicted in Fig.~\ref{fig:CVdiagram}. The rate of the flow into the infectious class is denoted as $\Phi_{\mathrm{in}}$. The flow out is driven by two mechanisms: recovery ($\Phi_{\mathrm{rec}}$) and detection ($\Phi_{\mathrm{det}}$).
\begin{figure*
\centering
\includegraphics[width=0.9\linewidth]{Figure1.pdf}
\caption{Compartmental model for COVID-19 and similar diseases. Individuals join the infectious class (red) by contracting the disease (at a rate $\Phi_{\mathrm{in}}$) and leave by recovering ($\Phi_{\mathrm{rec}}$) or by being detected and subsequently isolated ($\Phi_{\mathrm{det}}$).}
\label{fig:CVdiagram}
\end{figure*}
We can now write a conservation equation for the infectious group:
\begin{equation}
I_{t+1} - I_t = \Phi_{\mathrm{in}} - \Phi_{\mathrm{rec}} - \Phi_{\mathrm{det}}
\label{eq:CVeqn}
\end{equation}
where $I_t$ is the number of infectious individuals at time $t$. This conservation equation highlights the three levers that we can operate to reduce the spread of the disease and achieve controlled, stable dynamics. The first is reducing the rate of infection ($\Phi_{\mathrm{in}}$), which can be achieved by increasing social distancing, wearing masks, or vaccinating the population. The second is increasing the rate of recovery. This lever is not easy to operate and is dictated by biological constraints and available medical treatments. The final lever, increasing detection, is achieved by ramping up infection testing and contact tracing measures.
The objective of this analysis is to provide practical guidance to organizations looking to reopen. For this reason, the model must capture the fact that, in practice, compliance to these intervention strategies may not be perfect; organizations may not be able to mandate particular mitigation strategies and even if they do, individuals may choose not to abide. In order to capture the effect (and danger) of this non-compliance, the model presented here allows for some members of the population to employ the intervention strategies while others abstain. The fraction of the population who are vaccinated $f_v$, who opt into random surveillance testing $f_T$, and who wear masks $f_m$ are therefore all considered. While a person's decision to get vaccinated or to opt into a surveillance testing program is fixed, their decision to wear a mask may vary day-to-day. Such considerations are made in formalizing the partitions and flows between partitions of the model.
Stability in the disease dynamics is defined as the case where the size of the infectious population (shown in red in Fig.~\ref{fig:CVdiagram}) is shrinking, meaning this is the partition we are most concerned with modeling. For this study, we will assume that vaccinated individuals and unvaccinated individuals will behave the same way when they are infectious, i.e.~they will have the same ability to transmit the disease and will be equally likely to recover or be tested/detected on a given day. Additionally, because we do not assume masking behavior is constant day-to-day, we need not separate the infectious population into masked and unmasked partitions. Opting into or out of a surveillance testing program, however, is a different story. An individual's testing status (their opt-in or opt-out decision) remains the same over time and infected individuals who are tested will become isolated and leave the infectious population much faster than those who opt-out of testing. Therefore, we need to consider the behavior of the opt-into testing population and the opt-out-of testing population separately, meaning we will examine two conservation principles, one for those opting into testing
\begin{equation}
I_{t+1,\mathrm{oi}} - I_{t, \mathrm{oi}} = \Phi_{\mathrm{in}, \mathrm{oi}} - \Phi_{\mathrm{rec},\mathrm{oi}} - \Phi_{\mathrm{det},\mathrm{oi}}
\label{eq:CVeqnOI}
\end{equation}
and one for those opting out
\begin{equation}
I_{t+1,\mathrm{oo}} - I_{t, \mathrm{oo}} = \Phi_{\mathrm{in}, \mathrm{oo}} - \Phi_{\mathrm{rec},\mathrm{oo}} - \Phi_{\mathrm{det},\mathrm{oo}}.
\label{eq:CVeqnOO}
\end{equation}
The next step is defining these flows into and out of the infectious populations. First, we consider the infection process. Infection occurs when an infectious individual successfully transmits the disease to one of their susceptible contacts. To estimate the likelihood of transmission, we must first introduce the basic reproduction number of a disease, denoted $\mathcal{R}_0$. This classic epidemiological parameter is defined as the number of secondary cases resulting from a single infected individual in a fully susceptible population \cite{IntroMathEpid, SIR_OG}. If, for instance, that single infected individual was contagious for $d$ days, had $\hat{x}$ daily contacts (all susceptible individuals), and there was a likelihood of transmitting the disease to each contact (each day) of $\rho$, the basic reproduction number would be
\begin{equation}
\mathcal{R}_0 = \rho \hat{x} d.
\label{eq:R0}
\end{equation}
This relationship defines the transmission likelihood in the absence of mitigation strategies ($\rho$) in terms of the parameters most often reported for a particular infectious disease, i.e.~$\mathcal{R}_0$ and $d$. The daily number of susceptible contacts $\hat{x}$ will depend on the particular community and will be discussed further below.
The transmission likelihood in a given interaction, however, will be reduced when interventions such as mask-wearing and vaccination are implemented. If one person is wearing a mask, the likelihood of transmission is reduced by a factor of $\left(1-\epsilon_m\right)$ where $\epsilon_m$ is the efficacy of mask use at preventing the spread of the disease. If both people in the interaction are wearing masks, the reduction factor is $\left(1-\epsilon_m\right)^2$. Similarly, if the susceptible individual is vaccinated, the transmission likelihood is reduced by a factor of $\left(1-\epsilon_v\right)$ with vaccine efficacy to breakthrough infectious $\epsilon_v$. If the susceptible individual is a testing opt-in individual, they will join the testing opt-in infectious population when they contract the disease. If instead the susceptible individual is a testing opt-out individual, they will join the opt-out infectious population when they contract the disease.
All together, these parameters can be used to define the flow of susceptible individuals into the infectious compartments:
\begin{equation}
\begin{split}
\Phi_{\mathrm{in},oi} &= \sum_{j=1}^{I_{t,\mathrm{oi}}}\sum_{k=1}^{X_j} T_k Y_{jk} + \sum_{j=1}^{I_{t,\mathrm{oo}}}\sum_{k=1}^{X_j} T_k Y_{jk} \\
\Phi_{\mathrm{in},oo} &= \sum_{j=1}^{I_{t, \mathrm{oi}}}\sum_{k=1}^{X_j} \left(1-T_k\right) Y_{jk} + \sum_{j=1}^{I_{t, \mathrm{oo}}}\sum_{k=1}^{X_j} \left(1-T_k\right) Y_{jk}
\end{split}
\end{equation}
where $X_j$ is a random variable representing the number of susceptible contacts of infected individual $j$. For now, we will assume we know this distribution and that it has a mean $\mu_x$ and a variance $\sigma^2_x$, though we will find that in determining stability to outbreaks, the basic reproduction number $\mathcal{R}_0$ is sufficient for capturing the critical aspects of this distribution. The testing status (opt-in or opt-out) of the susceptible contact $k$ is captured by the Bernoulli random variable $T_k$ which has parameter $f_T$ and indicates whether or not that individual will join the opt-in or opt-out population upon infection.
The random variable $Y_{jk}$ is a Bernoulli random variable with parameter equal to the effective transmission likelihood
\begin{equation}
\rho_{\mathrm{eff}} = \rho \left(1-\epsilon_m M_j\right)\left(1- \epsilon_m M_k \right)\left( 1- \epsilon_v V_k \right)
\label{eq:rhoEff}
\end{equation}
which itself depends on the normal (no intervention) transmission probability $\rho$, the intervention efficacies $\epsilon_m$ and $\epsilon_v$, and the intervention statuses of both individuals determined by the Bernoulli random variables $M_j$ (mask status of $j$, parameter $f_m$), $M_k$ (mask status of $k$, parameter $f_m$), and $V_k$ (vaccine status of $k$, parameter $f_v$).
Next we consider the recovery process. A simple approximation is to assume that each infected individual is equally likely to recover on any given day and that the expected duration for which they are infectious (in the absence of interventions) is $d$. The resulting recovery flows are
\begin{equation}
\begin{split}
\Phi_{\mathrm{det}, \mathrm{oi}} &= \sum_{j=1}^{I_{t, \mathrm{oi}}} Z_{j}\\
\Phi_{\mathrm{det}, \mathrm{oo}} &= \sum_{j=1}^{I_{t, \mathrm{oo}}} Z_{j}
\end{split}
\end{equation}
where $Z_j$ is a Bernoulli random variable with parameter $1/d$.
Finally, we consider detection through infection testing and follow-up contact tracing. While only opt-in individuals will be detected through random surveillance testing, both their opt-in and opt-out contacts may be identified and isolated through follow-up contact tracing. In the surveillance (random) testing group, any member is equally likely to be selected for testing on a given day (with probability $\nu$, the daily testing frequency). If an individual tests positive, their contacts are traced and tested. Isolating contacts only has a substantial effect on disease dynamics when those contacts themselves are infectious at the time of tracing. At the previous time step, each contact had a likelihood $\rho_{\mathrm{eff}}$ of having been infected by person $j$. Using this as the likelihood that contact $k$ is infected at the current time step will \textit{underestimate} the number of infectious contacts (because person $j$ may have been infecting their contacts for more than one day), which will lead to more conservative stability bounds. We also consider that not all contacts will be successfully traced, an event which happens with likelihood $\epsilon_c$. Each contact therefore has a probability $\epsilon_c \rho_{\mathrm{eff}}$ of being detected as infectious through contact tracing. Together, the number of infectious individuals removed from the infectious groups by detection via both random testing and contact tracing are
\begin{equation}
\begin{split}
\Phi_{\mathrm{det}, \mathrm{oi}} &= \sum_{j=1}^{I_{t,\mathrm{oi}}} V_j \left(1 + \sum_{k=1}^{X_j} T_k W_{jk} \right) \\
\Phi_{\mathrm{det}, \mathrm{oo}} &= \sum_{j=1}^{I_{t,\mathrm{oi}}} V_j \left(0+ \sum_{k=1}^{X_j} \left(1-T_k\right)W_{jk} \right)
\end{split}
\end{equation}
where $V_j$ (random testing) and $W_{jk}$ (tracing) are Bernoulli random variables with parameters $\nu$ and $\epsilon_c \rho_{\mathrm{eff}}$ respectively. Again, the testing status is captured in the random variable $T_k$ as defined above and the transmission probability $\rho_{\mathrm{eff}}$ is affected by the mask and vaccine statuses of the individuals involved (see Eq.~\ref{eq:rhoEff}).
Using the the constitutive relationships defined above and the conservation laws of Eq.~\ref{eq:CVeqnOI}-\ref{eq:CVeqnOO}, we can rewrite the equations for the evolution of the number of infectious individuals in each testing group:
\begin{equation}
\begin{split}
I_{t+1, \mathrm{oi}} &= I_{t, \mathrm{oi}} + \sum_{j=1}^{I_{t, \mathrm{oi}}} \left[\left(\sum_{k=1}^{X_j} T_k \left(Y_{jk} - V_j W_{jk}\right)\right) - Z_j \left(1-V_j\right) - V_j \right] + \sum_{j=1}^{I_{t,\mathrm{oo}}}\left[ \sum_{k=1}^{X_j} T_k Y_{jk} \right] \\
I_{t+1, \mathrm{oo}} &= I_{t, \mathrm{oo}} + \sum_{j=1}^{I_{t, \mathrm{oi}}} \left[\sum_{k=1}^{X_j} \left(1-T_k\right) \left(Y_{jk} - V_j W_{jk}\right)\right] + \sum_{j=1}^{I_{t,\mathrm{oo}}}\left[\left( \sum_{k=1}^{X_j} \left(1-T_k\right) Y_{jk} \right) - Z_j\right].
\end{split}
\label{eq:stochasticDiff}
\end{equation}
Here, we have made one additional correction. The recovery removal for each member of the opt-in population is $Z_j\left(1-V_j\right)$ to avoid double counting people who recover on the same time step that they test positive. This ensures we are not overestimating the flow out of the infectious population, hence our results will remain conservative.
\subsection{Stability analysis}
The disease dynamics in a community are considered stable if a randomly infected individual will not be able to spread the disease to others in the community before being detected or recovering. In a real (stochastic) environment it will always be possible for a single infection to lead to a widespread outbreak, as in the case of ``superspreader" events where a single infectious individual has many more contacts than usual on a single day and will likely be able to spread the disease before being identified. To capture the dynamics, therefore, we consider the mean and variance of the evolution of the number of infectious individuals at a given time.
The mean dynamics are found by taking the expectation of Eq.~\ref{eq:stochasticDiff} for the full stochastic process. Doing so leads us to find a 2-by-2 transition matrix:
\begin{equation}
\textbf{A} =
\left[\begin{array}{cc}
a_{11} & a_{12} \\
a_{21} & a_{22}
\end{array} \right]
=
\left[\begin{array}{cc}
1+\hat{\rho}_{\mathrm{eff}}\mu_x f_T \left(1-\nu \epsilon_c\right) - \left(\frac{1}{d} + \nu - \frac{\nu}{d}\right)& \hat{\rho}_{\mathrm{eff}} \mu_x f_T \\
\hat{\rho}_{\mathrm{eff}}\mu_x \left(1-f_T\right)\left(1-\nu \epsilon_c\right) & 1+ \hat{\rho}_{\mathrm{eff}} \mu_x \left(1-f_T\right) - \frac{1}{d}
\end{array} \right]
\end{equation}
such that
\begin{equation}
\left(\begin{array}{cc}
\mathbb{E}\left[ I_{t+1, \mathrm{oi}}\right] \\
\mathbb{E} \left[I_{t+1, \mathrm{oo}}\right]
\end{array}\right) =
\textbf{A}
\left(\begin{array}{cc}
\mathbb{E} \left[I_{t, \mathrm{oi}} \right]\\
\mathbb{E} \left[I_{t, \mathrm{oo}}\right]
\end{array} \right).
\label{eq:2by2}
\end{equation}
Here, we have used
\begin{equation}
\hat{\rho}_{\mathrm{eff}} = \mathbb{E}\left[\rho_{\mathrm{eff}}\right] = \rho \left(1- \epsilon_m f_m\right)^2 \left(1-\epsilon_v f_v\right).
\end{equation}
Stability in the mean dynamics therefore is achieved when the eigenvalues of $\textbf{A}$, the transition matrix, are both less than 1. Finding the eigenvalues of the matrix in Eq.~\ref{eq:2by2} leads to a single stability criterion; this system is stable if
\begin{equation}
1 > \rho \mu_x d \left(1-\epsilon_m f_m\right)^2 \left(1- \epsilon_v f_v\right)\left(f_T \frac{1-\nu \epsilon_c}{1+\nu\left(d-1\right)} + \left(1-f_T\right)\right)
\label{eq:stability}
\end{equation}
Notice, the mean value of the distribution $X$ (the number of susceptible contacts of an infectious individual) appears in this expression, implying that we need to know something about the underlying distribution, information that is typically unavailable. In analyzing stability to create practical guidance for intervention measures, this distribution (particularly its mean) need not be known exactly, but we do need an upper bound on this value to ensure that we overestimate disease transmission and determine conservative values for mitigation requirements. To estimate this upper bound, we return to the discussion of the basic reproduction number $\mathcal{R}_0$ (see Eq.~\ref{eq:R0}). Certainly, assuming that all of an individual's contacts are susceptible (which is implied in the $\hat{x}$ of the basic reproduction number) is an overestimate of the true number of susceptible contacts that an individual has.
However, we are not considering the number of susceptible contacts that any random individual in the population has; we are considering the number of susceptible contacts that an \textit{infected} individual has, which, akin to the friendship paradox, is expected to be higher than that of non-infected individuals \cite{KojakuSadamori2021_friendship}. In this case, $\hat{x}$ is still a reasonable (over)estimate for $\mathbb{E}\left[X\right]$ because it is inferred from the measured $\mathcal{R}_0$ (based on infection case data), which naturally incorporates the tendency of a disease to spread to highly-connected individuals \cite{HeffernanJM2005_measureR0, KEELINGMATTJ2000_measureR0, DelamaterPaulL2019ComplexR0, LiJing2011T_failureR0}. In conclusion, the term $\rho\mu_x$, which contains parameters that are unknown and difficult to measure, can be rewritten as $\mathcal{R}_0/d$ without jeopardizing confidence in the resulting stability bound.
Now we return to Eq.~\ref{eq:stability} to redefine it in more familiar terms, i.e.~in terms of an effective reproduction number $\mathcal{R}_{\mathrm{eff}}$:
\begin{equation}
\mathcal{R}_{\mathrm{eff}} = \mathcal{R}_0 \left(1-\epsilon_m f_m\right)^2 \left(1- \epsilon_v f_v\right)\left(f_T \frac{1-\nu \epsilon_c}{1+\nu\left(d-1\right)} + \left(1-f_T\right)\right) < 1
\label{eq:ReffStability}
\end{equation}
which ensures stability of the system. The implications of this expression, i.e.~how it can be used to quantify trade-offs between mitigation strategies, are discussed in the following sections. One important note, however, is that this expression depends on the underlying network of the community only through the basic reproduction number $\mathcal{R}_0$, a commonly measured and reported value for a disease. No further information about the network, such as the variance in the distribution of contacts $X$, is required to determine a sufficient balance of interventions to ensure stability.
Having explored the evolution of the mean number of infectious individuals, we now turn to the evolution of the variance. All of these dynamic equations together can be written as
\begin{equation}
\left(\begin{array}{cc}
\mathbb{E}\left[ I_{t+1, \mathrm{oi}}\right] \\
\mathbb{E} \left[I_{t+1, \mathrm{oo}}\right] \\
\mathrm{var} \left(I_{t+1, \mathrm{oi}} \right)\\
\mathrm{var} \left(I_{t+1, \mathrm{oo}} \right)
\end{array}\right) =
\left[\begin{array}{cccc}
a_{11} & a_{12} & 0 & 0 \\
a_{21} & a_{22} & 0 & 0 \\
b_{11} & b_{12} & a_{11}^2 & a_{12}^2 \\
b_{21} & b_{22} & a_{21}^2 & a_{22}^2
\end{array} \right]
\left(\begin{array}{cc}
\mathbb{E} \left[I_{t, \mathrm{oi}} \right]\\
\mathbb{E} \left[I_{t, \mathrm{oo}}\right] \\
\mathrm{var} \left(I_{t, \mathrm{oi}} \right)\\
\mathrm{var} \left(I_{t, \mathrm{oo}} \right)
\end{array} \right).
\label{eq:4by4}
\end{equation}
where the elements $b_{11}, b_{12}, b_{21},$ and $b_{22}$ are functions of the disease and mitigation parameters defined as:
\begin{equation}
\begin{split}
b_{11} &= \mu_x f_T \hat{\rho}_{\mathrm{eff}} \left(1- f_T \hat{\rho}_{\mathrm{eff}}\right)+ \sigma_x^2 \left(f_T \hat{\rho}_{\mathrm{eff}}\right)^2 \\
& \qquad + \nu \left( \mu_x f_T \hat{\rho}_{\mathrm{eff}} \epsilon_c \left(1- f_T \hat{\rho}_{\mathrm{eff}} \epsilon_c\right) + \sigma_x^2 \left(f_T \hat{\rho}_{\mathrm{eff}}\epsilon_c\right)^2 \right) \\
& \qquad + \nu \left(1-\nu\right) \left(1+ \mu_x f_T \hat{\rho}_{\mathrm{eff}}\epsilon_c\right)^2 + \frac{1}{d} \left(1- \frac{1}{d}\right) \\
b_{12} &= \mu_x f_T \hat{\rho}_{\mathrm{eff}} \left(1- f_T \hat{\rho}_{\mathrm{eff}} \right) + \sigma_x^2 \left (f_T \hat{\rho}_{\mathrm{eff}} \epsilon_c \right)^2 \\
b_{21} &= \mu_x \left(1-f_T\right) \hat{\rho}_{\mathrm{eff}} \left(1- \left(1-f_T\right)\hat{\rho}_{\mathrm{eff}}\right) + \sigma_x^2 \left(\left(1-f_T\right)\hat{\rho}_{\mathrm{eff}}\right)^2 \\
& \qquad + \nu \left(\mu_x \left(1-f_T\right)\hat{\rho}_{\mathrm{eff}} \epsilon_c \left(1- \left(1-f_T\right)\hat{\rho}_{\mathrm{eff}}\epsilon_c\right) + \sigma_x^2 \left(\left(1-f_T\right)\hat{\rho}_{\mathrm{eff}}\epsilon_c\right)^2 \right) \\
& \qquad +\nu \left(1-\nu\right) \left(\mu_x \left(1-f_T\right) \hat{\rho}_{\mathrm{eff}}\epsilon_c \right)^2\\
b_{22} &= \mu_x \left(1-f_T\right) \hat{\rho}_{\mathrm{eff}} \left(1-\left(1-f_T\right) \hat{\rho}_{\mathrm{eff}}\right) + \sigma_x^2 \left(\left(1-f_T\right)\hat{\rho}_{\mathrm{eff}}\right)^2 + \frac{1}{d}\left(1-\frac{1}{d}\right)
\end{split}.
\end{equation}
While these expressions are cumbersome, there are two important results to discuss. First, while only the mean number of contacts of an infectious individual ($\mu_x$) appeared in the expression for the mean dynamics (which was then related to $\mathcal{R}_0$), the variance in this distribution $\sigma_x^2$ appears in this set of equations. Second, even though the variance $\sigma_x^2$ appears, it does not affect the stability requirements. Given the structure of this transition matrix, its eigenvalues are less than 1 when the eigenvalues of the matrix $\textbf{A}$ are less than 1. The effective reproduction number $\mathcal{R}_{\mathrm{eff}}$ from Eq.~\ref{eq:ReffStability} remains a sufficient stability criterion for both the mean and variance of the number of infectious individuals.
\subsection{Simulating infection}
To test the analytic result derived above in Eq.~\ref{eq:ReffStability}, we implemented an infection simulation on a fixed population. In these simulations, a population network is established at the outset, where every member of the population is a node in the network and every edge represents a contact between two people. Such networks are difficult to estimate for a real population, but generating random networks with wide ranges of structures and properties is straightforward. Fig.~\ref{fig:NetworkComparison} shows random network realizations for four different structure types---Erd\H{o}s-R\'{e}nyi, Uniform, Scale-free, and Small World---and their resulting degree distributions.
\begin{figure}[htp]
\hspace{-1cm}
\includegraphics[clip,width=0.6\columnwidth]{Ploterdos.png}%
\hspace{-1cm}
\includegraphics[clip,width=0.6\columnwidth]{Plotuniform.png}%
\hspace{-1cm}
\includegraphics[clip,width=0.6\columnwidth]{Plotbarabasi.png}%
\hspace{-1cm}
\includegraphics[clip,width=0.6\columnwidth]{Plotstrogatz.png}%
\caption{Sample realizations of four different network structures and their resulting degree distributions. Intervention-free infection was simulated on each network to make a measurement of $\mathcal{R}_0$. The required testing to stabilize the system was determined using Eq.~\ref{eq:testingReqFull} and another infection simulation was implemented with testing just above this stabilizing value. Both infection progressions are show: intervention-free (red) and with stabilizing testing (blue).}
\label{fig:NetworkComparison}
\end{figure}
Once the underlying network of connections is established, each individual is assigned to one of the partitions of the model as well as to vaccine and testing statuses. At each time step, the infection status of each individual is updated. If a person is infectious, they infect their susceptible neighbors with probability $\rho_{\mathrm{eff}}$ at each time, where $\rho_{\mathrm{eff}}$ is determined by the vaccine and masking status of the individuals involved. If an individual is infectious at time $t$, they recover at $t+1$ with probability $1/d$. The opt-in individuals will become isolated if they are detected through random testing with probability $\nu$. If they are detected, each of their infectious neighbors has a probability $\epsilon_c$ of also joining the isolated population. Isolated individuals will eventually recover and do so with probability $1/d$ at each time step. The simulation continues until there are no longer any infectious individuals remaining or for a fixed number of days.
\section{RESULTS}
\subsection{Balancing mitigation strategies: analysis}
Having performed the stability analysis, we turn our attention to the practical implications of the effective reproduction number of Eq.~\ref{eq:ReffStability}. The stability criterion ($\mathcal{R}_{\mathrm{eff}} < 1$) can be rearranged to provide guidance on how much of each mitigation strategy an organization should employ to ensure their reopening does not lead to widespread outbreaks.
\subsubsection{Without testing}
Let us first consider the case where no testing is done $\nu = 0$, which is also equivalent to the case where testing is available (nonzero $\nu$) but no one opts into the program ($f_T= 1$). This will be the case for the many organizations that do not have access to testing resources or the funding to secure them. It would also be the case for organizations where the vast majority of individuals are opposed to being testing and a testing requirement cannot be implemented. In this case, the stability criterion simplifies to
\begin{equation}
1 > \mathcal{R}_0 \left(1-\epsilon_m f_m\right)^2 \left(1- \epsilon_v f_v\right) = \mathcal{R}_{\nu = 0} \label{eq:noTesting}
\end{equation}
In the absence of masking and vaccinations, this simplifies to the well-known result for stable dynamics in the mitigation-free case: $1 > \mathcal{R}_0$. Implementing masking and/or vaccination requirements, however, will allow for controlled dynamics even in the case of basic reproduction numbers larger than 1.
First let us consider the effect of masking in the absence of vaccines, as would be the case at the beginning of a disease epidemic when vaccines may not exist or may not be widely available. Given a disease ($\mathcal{R}_0$) and the efficacy of masks at preventing transmission of that disease $\epsilon_m$, we can determine the fraction of the population that must be masked in order to ensure stability to disease outbreaks:
\begin{equation}
f_m > \frac{1- \frac{1}{\sqrt{\mathcal{R}_0}}}{\epsilon_m}.
\end{equation}
We are of course limited by the fact that no more than 100\% of the population can be masked at a given time. Meaning if $\mathcal{R}_0 > 1/\left(1-\epsilon_m\right)^2$, the disease cannot be controlled by masking alone.
Similar analysis can be performed to determine the effect of vaccination on the spread of a disease. Because compliance to masking can be difficult to ensure and because masking can be unpleasant and a hindrance to some activities, we now aim to determine what fraction of the population needs to be vaccinated to ensure that, in the absence of testing, no masking is required. The required vaccinated fraction is
\begin{equation}
f_v > \frac{1- \frac{1}{\mathcal{R}_0}}{\epsilon_v}.
\label{eq:vaxMandate}
\end{equation}
This recovers the classical definition of \textit{herd immunity}, i.e.~the fraction of the population that needs to be vaccinated to control the spread of a disease. Similar to the analysis of masking, we are limited by the fact that no more than 100\% of the population can be vaccinated, the result being that only diseases with $R_0 < 1/\left(1-\epsilon_v\right)$ can be controlled by vaccine mandates alone. In reality, less than 100\% of the population will be vaccinated, and if Eq.~\ref{eq:vaxMandate} does not hold for that value of $f_v$, additional mitigation strategies will be required to ensure stable disease dynamics. Such additional strategies may include masking (and the combined effects of these two interventions together is captured in Eq.~\ref{eq:noTesting}) or testing and contact tracing, as discussed below.
\subsubsection{With testing}
Now we will consider the case where testing resources are available to a given organization. We will first assume that participation in the testing program is mandatory for all individuals, i.e.~$f_T = 1$. In this case, the stability criterion of Eq.~\ref{eq:ReffStability} simplifies to
\begin{equation}
1 > \mathcal{R}_{\nu=0} \left(\frac{1-\nu \epsilon_c}{1 + \nu \left(d-1\right)}\right).
\end{equation}
where $\mathcal{R}_{\nu=0}$ is the effective reproduction number in the absence of testing but includes the effects of masking and vaccinations (see Eq.~\ref{eq:noTesting}). This expression can be rearranged to determine the minimum required daily testing fraction
\begin{equation}
\nu_{f_T = 1} > \frac{\mathcal{R}_{\nu=0} -1}{\epsilon_c \mathcal{R}_{\nu=0} + d - 1}.
\label{eq:testingMin}
\end{equation}
Again, this expression holds for the case where everyone opts into the testing program. We also recover the fact that if the masking and vaccine interventions can drive the effective reproduction number $\mathcal{R}_{ \nu=0}$ below 1, then no testing is required ($\nu_{f_T=1} = 0$). We further recover the result that if $R_0 < 1$, even without masking and vaccines, no testing is required to ensure stability to disease outbreaks.
Unlike the masking and vaccinated fractions ($f_m$ and $f_v$), the daily testing fraction can exceed $1$, meaning individuals are tested (on average) more than once per day. In reality, this is unlikely due to available testing resources. Additionally, rather than random testing, organizations may prefer to perform fixed-cadence testing for simpler logistical implementation. For instance, organizations may choose to test once a day (with an effective testing fraction $\nu=1$), once a week ($\nu = 1/7$), or once every two weeks ($\nu = 1/14$). Even if the testing capability is larger than the given effective testing fraction, organizations may still opt to test on a regular cadence to keep the implementation as easy as possible.
Finally, before discussing the case where individuals only participate in the testing program voluntarily ($f_T < 1$), the effect of contact tracing must be noted. By examining Eq.~\ref{eq:testingMin}, we see that less contact tracing (lower efficacy $\epsilon_c$) corresponds to increased required daily testing fractions and thus increased required daily testing resources. Organizations with the capability to do contact tracing (and to do so at little to no cost) should therefore employ this strategy to reduce the burden and cost of testing.
Next we consider the case where not all individuals are required to participate in surveillance testing (or if they are required to but are not compliant). Eq.~\ref{eq:ReffStability} can again be rearranged to provide guidance on how much testing must be done in this case:
\begin{equation}
\nu > \frac{\mathcal{R}_{\nu=0} -1}{\left[f_T\epsilon_c - \left(1-f_T\right) \left(d-1\right)\right]\mathcal{R}_{\nu=0} + d - 1 }
\label{eq:testingReqFull}
\end{equation}
where again, $\mathcal{R}_{\nu=0}$ is the effective reproduction number in the presence of masking and vaccinations but the absence of testing (see Eq.~\ref{eq:noTesting}). This expression captures the effect of all of the mitigation strategies we have explored. Given the masking ($f_m$), vaccination ($f_v$), and testing ($f_T$) requirements of an organization, this relationship determines the amount of daily testing that must be done ($\nu$) to ensure reopening without risk of causing widespread outbreaks of a specific disease ($\mathcal{R}_0, d$). If an organization does not have sufficient resources to perform this level of daily testing, they can alter their requirements for the other mitigation strategies to lower the needed testing.
\subsection{Balancing mitigation strategies: example}
While the trade-offs between the discussed mitigation strategies are fully expressed in Eq.~\ref{eq:testingReqFull}, they are nonlinear and perhaps difficult to imagine. In this section, an illustration of these trade-offs is provided for a specific set of disease parameters. These parameters loosely correspond to the reported values for the Delta variant of SARS-CoV-2. The selected basic reproduction number is $\mathcal{R}_0 = 5$ \cite{deltaR0_5}; the length of the contagious period is $d=14$ days \cite{d_14}; and the efficacies of masking and vaccinations are $\epsilon_m = 0.25$ \cite{hosoi2021estimating} and $\epsilon_v = 0.65$ \cite{vaccineEfficacy}.
For this particular disease, we consider four scenarios: no masking and no vaccines ($f_m = f_v = 0$), only masking with full compliance ($f_m = 1, f_v = 0$), only vaccinating with full compliance ($f_m = 0, f_v =1$), and full compliance on both masking and vaccination mandates ($f_m=f_v=1$). In each scenario, we consider the full spectrum of compliance to participation in a testing program ($f_T = [0,1]$), and testing scenarios from none at all to daily testing for everyone ($\nu = [0,1]$). In all scenarios, contacts of infectious individuals are assumed to be successfully identified and tracked down with efficacy $\epsilon_c = 0.8$.
Using these values and Eq.~\ref{eq:testingReqFull}, a stability boundary was determined in the space spanning the testing frequency ($\nu$) and testing compliance ($f_T$) for each of these scenarios. This stability boundary is shown for each of the four scenarios as white lines in Fig.~\ref{fig:tradeoffs}. If the combination of daily testing fraction and fraction of the population opting-in to the testing program falls above the stability boundary (in a blue region in Fig.~\ref{fig:tradeoffs}), the mitigation strategies in place will lead to successful prevention of widespread outbreaks of this disease. If instead a community's testing program parameters fall below the stability boundary (in a red region in Fig.~\ref{fig:tradeoffs}), that community would be susceptible to outbreaks.
\begin{figure
\centering
\includegraphics[trim = 100 50 100 0, clip, width=1\columnwidth]{tradeoffsFinal.png}
\caption{Trade-offs between mitigation strategies. Infection of a disease similar to the Delta variant of SARS-CoV-2 with $\mathcal{R}_0 = 5$ and $d = 14$ was simulated on random networks of various structures under random intervention combinations. The analytical stability boundary is shown as a white line separating the controlled region (blue background) from the uncontrolled region (red background). Points represent the average over 100 infection simulations for the same set of parameters on a random network whose structure is represented by point shape.}
\label{fig:tradeoffs}
\end{figure}
To further illustrate the separation of stable and unstable dynamics, the propagation of a disease with these parameters was simulated on 200 random networks (50 for each scenario) comprised of 5000 nodes with a randomly selected testing program ($f_T, \nu$ combination). The structure and properties of these networks were selected randomly from the ranges specified in Table \ref{table:MC}. In Fig.~\ref{fig:tradeoffs}, each simulated network is represented as a point whose shape indicates its network structure and whose color represents the percent of the population that became infected with the disease over 180 days (averaged over 100 random infection progressions for initial seeded infections of $I_0 = 5$). Notice, the color scale is nonlinear; blue-to-white points represent those where each initially infected individual ($5$ people) spread the disease to, on average, fewer than one other individual ($<10$ people infected total), while white-to-red points represent those where the disease spread more than that, up to propagating through the entire population (5000 people). The simulation results are consistent with the stability analysis performed above.
The results in Fig.~\ref{fig:tradeoffs} are an illustration of the analysis presented in this paper only for one specific disease (based on the Delta variant of SARS-CoV-2), but the analysis will hold for diseases of all parameters. For instance, in the case of both mask and vaccine mandates (the lower right plot of Fig.~\ref{fig:tradeoffs}), some amount of regular testing and compliance to testing is required to ensure a community will be stable to widespread outbreaks of this disease. However, if a different set of disease parameters were selected, e.g.~ones based on the Alpha variant of SARS-CoV-2 which will have a lower basic reproduction number and against which vaccines will be more effective \cite{vaccineEfficacy, d_14}, mask and vaccine mandates would be sufficient for controlling the spread of the disease (see Eq.~\ref{eq:noTesting}), and testing would no longer be needed. Finally, while Fig.~\ref{fig:tradeoffs} only shows four scenarios based on the absence/presence of fully compliant masking and/or vaccine mandate, compliance to either mandate in practice would likely fall somewhere in between. The stability boundary in these scenarios can still be calculated using Eq.~\ref{eq:testingReqFull} and used to determine the requirements of a successful testing program.
\subsection{Network (in)dependence}
The simulations whose results are shown in Fig.~\ref{fig:NetworkComparison} and Fig.~\ref{fig:tradeoffs} both begin to suggest the minimal role of the network parameters on the stability criterion. All relevant network information is captured in the value of $\mathcal{R}_{0}$; no additional information about the network structure or degree distribution is required. This result is clear from the expression for the effective reproduction number (Eq.~\ref{eq:ReffStability}) but is further illustrated by performing a Monte Carlo simulation of the parameter space.
\begin{figure}
\includegraphics[trim = 0 125 0 0, clip, width=0.8\columnwidth]{plot1Dfinal.png}
\caption{Monte Carlo infection simulation. Intervention strategies, intervention efficacies, disease parameters, and network parameters were all generated randomly. The mean total fraction of the population which became infected over 100 trials at each combination of parameters is shown (via point color) against the effective reproduction number $\mathcal{R}_{\mathrm{eff}}$. This single parameter successfully separates the controlled and uncontrolled infection outbreaks denoted by blue and red backgrounds respectively.}
\label{fig:oneDplot}
\end{figure}
The results of this Monte Carlo simulation are shown in Fig.~\ref{fig:oneDplot} and the ranges from which all the parameters were sampled are listed in Table \ref{table:MC}. Networks of 5000 nodes were generated and seeded randomly with 5 initial infections. Fig.~\ref{fig:oneDplot} presents the fraction of the population that became infected over the whole simulated epidemic (on average over 100 random infection progressions) against the $\mathcal{R}_{\mathrm{eff}}$ (computed from the randomly selected combination of parameters using Eq.~\ref{eq:ReffStability}). For all combinations and networks, only this effective reproduction number is required to predict whether outbreaks are limited to $<1\%$ of the population or spread to nearly everyone. While the exact number of individuals who become infected may vary, the presence or absence of an outbreak is well predicted by the effective reproduction number, regardless of the network structure. Again, this is consistent with the analysis in which $\mathcal{R}_0$ captures all of the relevant network information; no further information about the distribution of edges (contacts) in the network, e.g.~the mean or variance of this distribution, is required to determine what combination of mitigation strategies will successfully lead to controlled disease dynamics.
\begin{table}[htbp]
\caption{The ranges for each parameter that was sampled randomly and used in an infection simulation.}
\begin{center}
\begin{tabular}{l c|c|c}
\multicolumn{2}{c|}{\textbf{Parameter}}&\textbf{Description}&\textbf{Range}\\
\hline{}
Disease & $\mathcal{R}_0$ & basic reproduction number & 3--13 \\
& $d$ & contagious period (days) & 5--15 \\
\hline
Intervention & $\epsilon_m$ & mask efficacy & 0--1 \\
& $\epsilon_v$ & vaccine efficacy & 0--1 \\
& $\epsilon_c$ & contact tracing efficacy & 0--1 \\
& $f_m$ & fraction masked & 0--1 \\
& $f_v$ & fraction vaccinated & 0--1 \\
& $f_T$ & fraction opting-in to testing & 0--1 \\
& $\nu$ & daily testing fraction & 0--1 \\
\hline
Network & & & \\
\textit{Erd\H{o}s-R\'{e}nyi} & $\mu_{\mathrm{ER}}$ & mean degree & 5--25 \\
\textit{Uniform} & $x_{\mathrm{min,U}}$ & minimum degree & 1--10 \\
& $x_{\mathrm{max,U}}$ & maximum degree & 10--20 \\
\textit{Scale-free} & $\alpha_{\mathrm{SF}}$ & distribution power law & 3--5 \\
& $x_{\mathrm{min,SF}}$ & minimum degree & 1--20 \\
\textit{Small World} & $\mu_{\mathrm{SW}}$ & mean degree & 5--25 \\
& $\beta_{\mathrm{SW}}$ & rewiring probability & 0--1
\end{tabular}
\label{table:MC}
\end{center}
\end{table}
\section{DISCUSSION}
\subsection{Comparison to university case data}
This analysis was performed to inform policy making for semi-contained organizations with a limited number of individuals, universities in particular. Since the large-scale shutdowns of the early COVID-19 pandemic, most universities have returned to nearly full operation using some combination of the mitigation strategies discussed above. Few universities have reported widespread outbreaks since their reopenings. This is consistent with our analysis and may be because schools are implementing a sufficient combination of mitigation strategies to prevent such outbreaks. However, a lack of reported outbreaks does not necessarily imply stability; universities may not be reporting outbreaks because they are simply not detecting them. Detection requires regular viral testing of some kind. If schools are testing regularly, they would detect outbreaks. However, testing weekly would also be a sufficient way to ensure that there are no outbreaks.\footnote{The Delta variant of SARS-CoV-2 was discovered after the vaccine was developed, and as seen in Fig.~\ref{fig:tradeoffs}, weekly testing in the case of a highly vaccinated population is sufficient for controlling the spread of this disease regardless of masking behavior. In reality, compliance to vaccine requirements will be less that $100\%$, but weekly testing is likely still sufficient provided enough individuals opt-into the program. Therefore, schools that are testing their members once a week are implementing mitigation strategies that our analysis suggests are sufficient for preventing outbreaks.} Universities that are not testing may not be in the stable disease dynamics regime but may also not detect instability.
Case data can also be misleading as cases can go up at a university even if appropriate mitigation strategies are in place. This happens because each day there can be interactions between those inside the community (e.g.~students at the school) and those outside the community (e.g.~individuals in the town or city in which the school is located). These interactions effectively lead to new ``seed" infections in the community population. While the policies in place at the university can control whether or not these seeds grow, they may be limited in the control of how often the seeds are planted; the seeding rate will depend more on the overall prevalence of the disease in the area. Separating case data at a university by the origin of the infection (from intra-community or extra-community interactions) is a complex data task beyond the scope of this analysis.
\subsection{Generality of results}
While this work was originally inspired by the necessity to use testing and other interventions to control the spread of COVID-19, the results depend neither on the details of the disease nor on the details of the population structure. Given some estimate of the parameters of a disease's proliferation, the stability analysis performed in this paper gives general guidelines on what successful testing programs require based on the adherence to masking and/or vaccination policies by individuals in the population.
The resulting stability criterion interestingly \emph{does not} depend on the variance of the underlying network of the community. Highly connected individuals only impact the dynamics through their contribution to the measured basic reproduction number $\mathcal{R}_0$. This simplifies the analysis of the spread of an epidemic as all the relevant network details are captured in the measured $\mathcal{R}_0$, allowing one to evaluate mitigation strategies for all communities, not just those whose interaction behaviors are roughly homogeneous. Hence in examining the impact of interventions on \textit{stability}, higher complexity network models (those with additional information about the underlying community structure) are unnecessary. Analyses that go beyond stability, e.g.~predicting the exact size of an outbreak rather than whether or not an outbreak will exist, require models that contain more detailed network information \cite{SakaguchiInhomogeneous,RadicchiOutbreakSize}. However, for the prevention of widespread outbreaks, the stability criterion presented in Eq.~\ref{eq:ReffStability} is sufficient for capturing the effect of different mitigation strategies in a way that allows policy makers to balance trade-offs, given their available resources, how to reopen their schools, businesses, and other organization safely amidst a pandemic.
\bibliographystyle{plain}
| 5cc51239ff6d15de9d0d0840bd0ab47d86ba1a0c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Spin is a fundamental quantum mechanical property of electrons and
other particles. The spin states can be used as the basis of quantum
bits in quantum information science (QIS)\citep{awschalom2018quantum},
in addition to being used in spintronics analogous to electrical charge
in conventional electronics\citep{vzutic2004spintronics}. The key
property for spintronics and spin-based QIS is the lifetime of spin
states. Stable manipulations of spin states in practical applications
require lifetimes on the order of hundreds of nanoseconds or even
milliseconds. Determining the underlying mechanisms and controlling
spin relaxation are vital to reach long spin lifetimes at room temperature.
Experimentally spin relaxation can be studied through ultrafast magneto-optical
pump-probe\citep{meier2012optical,dean2016ultrafast} and spin transport
measurements\citep{avsar2020colloquium}, allowing the direct observations
of dynamical processes and quantitative determination of spin relaxation
time, $\tau_{s}$.
Despite significant experimental progresses and several proposed systems
in the past decades\citep{pla2012single,vzutic2004spintronics}, materials
with properties required for practical QIS and spintronics applications
such as long $\tau_{s}$ at room temperature remain to be found\citep{awschalom2018quantum,avsar2020colloquium,RevModPhys.89.011001}.
Theoretical predictions of materials properties have been mostly focused
on electronic excitations\citep{reviewDreyer,Wu2017,Smart2018} and
electron-hole recombinations\citep{Wu2019,Wu20192,smart2020} of potential
spin defects for QIS applications. Reliable prediction of spin lifetime
and dominant relaxation mechanism will allow rational design of materials
in order to accelerate the identification of ideal materials for quantum
technologies, while forgoing the need of experimental search over
a large number of materials.
Until recently, most state-of-the-art theoretical methods to study
spin dynamics of solid-state materials are limited to simplified and
system-specific models that require prior input parameters \citep{vzutic2004spintronics,dyakonov1972spin,cummings2017giant,afanasiev2019ultrafast}.
These methods laid important theoretical foundation for spin dynamics,
such as the spin-Bloch kinetic equations developed from Non-equilibrium
Green's Function theory (NEGFT)\citep{wu2010spin}. Ref.\citenum{rosati2014derivation} derived a closed equation of motion for the electronic single-particle density matrix, including different scattering matrices, which may be applicable to spin dynamics. However, because
of the simplified electronic structure and electron-phonon coupling
matrices, quantitative prediction of spin relaxation remains out of
reach. Occasionally, even trends in $\tau_{s}$ predicted by such
models may be incorrect as shown recently for graphene.\citep{habib2020electric}
Furthermore, these models are unable to provide predictive values
for new materials where prior inputs are not available.
Prior to our work, the existing first-principles methodology for spin
lifetime has been mostly based on spin-flip matrix elements in a specialized
Fermi's Golden rule\citep{restrepo2012full,park2020spin,fedorov2013impact},
which is only applicable to systems with Kramers' degeneracy or spatial
inversion symmetry, not suitable to lots of materials promising for
quantum computing and spintronics applications~\citep{vzutic2004spintronics,avsar2020colloquium}.
Other first-principles techniques like real-time Time-Dependent Density
Functional Theory (TDDFT)\citep{marques2012fundamentals} are challenging
for crystalline systems due to high computational cost for describing
phonon relaxations that require large supercells. More importantly,
long simulation time over nanoseconds often required by spin relaxation
is a major difficulty for TDDFT, which is only practical for tens
to a few hundred femtoseconds. While spin dynamics based on TDDFT
has been recently performed for ultrafast demagnetization of magnetic
systems within tens of fs\citep{chen2019role,acharya2020ultrafast,krieger2015laser},
the intrinsic time scale and supercell limitations mentioned above
remain.
We recently derived a generalized rate equation~\citep{xu2020spin}
based on first-principles density matrix (DM) Lindblad dynamics framework,
which provides accurate spin relaxation time due to spin-orbit and
electron-phonon couplings for a broad range of materials, with arbitrary
symmetry. However, our previous work requires the system is already
at a quasi-equilibrium state when the dynamics can be described by
a single-exponential decay, which can not describe coupled spin and
carrier dynamics at an out-of-equilibrium state in ultrafast pump-probe
experiments. In this work, we develop a real-time \textit{ab initio} DM dynamics method
based on this theoretical framework, with complete
descriptions of scattering processes including electron-phonon (e-ph),
electron-impurity (e-i), and electron-electron (e-e), being adequate
for over ns to ms simulation time. Specifically, compared to the generalized
rate equation in our previous work, DM dynamics with explicit real-time
evolutions allow coupled carrier and spin relaxation away from quasi-equilibrium
with all decoherence pathways simultaneously. This will facilitate
direct prediction of experimental signatures in ultrafast magneto-optical
spectroscopy to unambiguously interpret experimental probes of spin
and electron dynamics.
We will demonstrate the generality of our approach by considering two prototypical and disparate systems - GaAs and few-layer WSe$_2$, which have very different spin relaxation mechanisms.
We will first apply our DM dynamics methodology
to investigate ultrafast spin dynamics of GaAs, which has broad interest
in spintronics over past decades\citep{vzutic2004spintronics,kikkawa1998resonant,hilton2002optical,jiang2009electron,kamra2011role}
and more recently\citep{PhysRevX.7.031010,PhysRevLett.121.033902,belykh2018quantum},
partly due to its long spin lifetime especially in the n-doped material
at relatively low temperature\citep{kikkawa1998resonant}. Despite
various experimental~\citep{kikkawa1998resonant,hilton2002optical,ohno1999spin,kimel2001room,hohage2006coherent}
and theoretical~\citep{vzutic2004spintronics,yu2005spin,jiang2009electron,mower2011dyakonov,kamra2011role,marchetti2014spin}
(mostly using parameterized model Hamiltonian) studies previously,
the dominant spin relaxation mechanism of bulk GaAs under various
temperatures and doping levels remains unclear. For example, Refs.\citenum{jiang2009electron}
and \citenum{mower2011dyakonov} claimed e-i and e-ph scatterings
dominate spin relaxation at low and room temperatures, respectively;
however, Refs.~\citenum{kamra2011role} and \citenum{marchetti2014spin}
conclude that e-e may be more important at room temperature and even
more at lower temperatures. Moreover, electron-phonon scattering matrices
which can be accurately obtained from first-principles, are very difficult
to be precisely described in parameterized models used previously.
Most importantly, the applicability of empirical D'yakonov-Perel'
(DP) relation, which is widely used for describing inversion-asymmetric
systems including GaAs, needs to be carefully examined. Throughout
this work, we provide complete and unambiguous insights on the underlying
mechanism of spin relaxation and applicability of the DP relation
for GaAs from first-principles DM dynamics.
Due to broken inversion symmetry and strong SOC,
monolayer transition metal dichalcogenides (TMDs) exhibit exciting
physical properties including valley-specific optical excitation and spin-valley locking effects. In Ref. \citenum{dey2017gate,li2021valley},
it has been shown that by introducing doping in monolayer TMDs, ultraslow
decays of Kerr rotations, which correspond to ultralong spin/valley
lifetimes of resident carriers especially resident holes
can be
observed at low temperatures. Those features make monolayer TMDs advantageous
for spin-valleytronics and (quantum) information processing.
Besides monolayers, bilayer TMDs recovering inversion symmetry
have also attracted significant interests because of the
new ``layer'' degree of freedom or layer pseudospin in addition
to spin and valley pseudospin\cite{gong2013magnetoelectric,xu2014spin,khani2020gate}.
Previous studies already concluded that electronic states at $K$/$K'$
valleys of a bilayer TMD are approximately a superposition of those of two monolayers. This allows us to tune which layer carriers/spins
are localized by a perpendicular electric field $E_{z}$, and make
use of the spin-valley-layer locking effects for spin-valleytronic
applications.
Although spin/valley relaxation of resident carriers
in monolayer TMDs, which is most relevant to spin-valleytronic
applications, have been extensively examined\citep{xu2020spin,dey2017gate,li2021valley,song2016long}, the underlying
dynamics especially the effects of different types of impurities
have not been investigated through predictive \emph{ab initio} simulations.
Furthermore, for bilayer TMDs, the study on spin/valley dynamics is still in infancy with few ultrafast measurements which however do not
exhibit long spin relaxation time and are lack of spin-valley-layer locking property \cite{guimaraes2018spin,ye2019ultrafast,bertoni2016generation}.
There is a lack of knowledge of the role of scattering processes and
the scattering pathways for spin/valley dynamics of free carriers
in bilayers,
which prevents researchers to realize and manipulate spin-valley-layer locking effects for designing spin-valleytronic devices.
In this work, we will answer the above questions
by performing \emph{ab initio} real-time dynamics simulations with
a circularly polarized pump pulse and relevant scattering mechanisms.
We focus on WSe$_{2}$ due to its larger valence band SOC splitting
and focus on dynamics of holes since $\tau_{s}$ of holes seem longer than electrons.
In the following, we first introduce our theoretical formalism of
real-time density-matrix approach with various scattering processes
and pump-probe spectroscopy. In particular, we focus on spin-orbit
mediated spin relaxation and decoherence processes under the existences
of electron scatterings, which are rather common in semiconductors
and metals\citep{wu2010spin,vzutic2004spintronics}. We use this method
to simulate pump-probe Kerr rotation and real-time spin dynamics,
by using GaAs as a prototypical example and comparing with experiments.
Next, we study spin lifetime dependence on the temperature and doping
level, where the dominant mechanisms can vary significantly. We then
discuss the roles of different scattering mechanisms and phonon modes
in carrier and spin relaxations, respectively, in order to resolve
related long-standing controversies. We further simulate
ultrafast dynamics in monolayer and bilayer WSe$_{2}$ and extract
useful dynamical properties. Our work provides predictive theory
and computational platform for open quantum dynamics, and offers new
and critical insights for spin relaxation and decoherence in general
solid-state systems.
\section{Theory}
\subsection{Real-time density-matrix dynamics and spin relaxation time\label{subsec:theory_dynamics}}
To provide a general formulation of quantum dynamics in solid-state
materials, we start from the Liouville-von Neumann equation in the
interaction picture,
\begin{align}
\frac{d\rho\left(t\right)}{dt} & =-i[H'\left(t\right),\rho\left(t\right)],\label{eq:Liouville}\\
H'\left(t\right) & =H\left(t\right)-H_{0}\left(t\right),
\end{align}
where $H$, $H_{0}$ and $H'$ are total, unperturbed and perturbed
Hamiltonian, respectively. In this work, the total Hamiltonian is
\begin{align}
H= & H_{0}+H\sub{pump}+H\sub{e-i}+H\sub{e-ph}+H\sub{e-e},\label{eq:H}\\
H_{0}= & H_{e,0}+H\sub{efield}+H_{z}+H\sub{ph},
\end{align}
where $H_{e,0}$ is electronic Hamiltonian under zero external field.
In this work, $H\sub{efield}$ is Hamiltonian induced by a perpendicular electric field $E_z$ along the vacuum direction.
$H_{z}$ is the Zeeman Hamiltonian corresponding
to an external magnetic field $\vec{B}$, $H_{z}=g_{s}\mu_{B}\vec{B}\cdot\vec{s}$,
where $\vec{s}=\left(s_{x},s_{y},s_{z}\right)$ and $s_{i}$ is spin
matrix in Bloch basis under zero field. $g_{s}$ is $g$ factor and
$\mu_{B}$ is the Bohr magneton. $H\sub{pump}$ is the Hamiltonian
of the pump pulse and will be described below. $H\sub{ph}$ is the
phonon Hamiltonian, while $H\sub{e-i}$, $H\sub{e-ph}$ and $H\sub{e-e}$
describe the electron-impurity, electron-phonon and electron-electron
interactions respectively. The detailed forms of the interaction Hamiltonians
are given in Appendix A.
In practice, the many-body density matrix master equation in Eq.~\ref{eq:Liouville}
is reduced to a single-particle one and the environmental degrees
of freedom are traced out\citep{rossi2002theory}. The total rate
of change of the density matrix is separated into terms related to
different parts of Hamiltonian,
\begin{align}
\frac{d\rho}{dt}= & \frac{d\rho}{dt}|\sub{coh}+\frac{d\rho}{dt}|\sub{scatt},\label{eq:dynamics}
\end{align}
where $\rho$ is the density matrix of electrons. Above, $\frac{d\rho}{dt}|\sub{coh}$
describes the coherent dynamics of electrons under potentials or fields,
e.g. the applied pump pulse, while $\frac{d\rho}{dt}|\sub{scatt}$
captures the scattering between electrons and other particles.
To obtain Eq. \ref{eq:dynamics} which involves only
the dynamics of electrons or the electronic subsystem, we have assumed
the environmental subsystem is not perturbed by the change of the electronic subsystem,
which in this work means there is no dynamics of phonons.
This assumption is valid when the system is not far from equilibrium, e.g.,
when excitation is weak.
In most spin dynamics experiments, it is desirable to work in the low excitation density limit to avoid additional complexities and focus on the physics of spin dynamics.
Indeed in many experiments, e.g., in Refs. \citenum{kikkawa1998resonant} and \citenum{yang2015long}, pump fluence and excitation density are controlled to be low, e.g., excitation density 2$\times$10$^{14}$ cm$^{-2}$ for GaAs.
Therefore, phonon dynamics can be safely excluded in the current stage.
The inclusion of phonon degrees of freedom in the density-matrix dynamics has been discussed
in detail in Refs. \citenum{iotti2017phonon,rossi2002theory} with model Hamiltonian, which can be our future work to implement from first-principles.
To define spin lifetime, we follow the time evolution of the observable
\begin{align}
S_{i} & =\mathrm{Tr}\left(s_{i}\rho\right),\label{eq:spin_observable}
\end{align}
where $s_{i}$ is the spin operator ($i=x,y,z$). This time evolution
must start at an initial state (at $t=t_{0}$) with a net spin i.e.
$\delta\rho(t_{0})=\rho(t_{0})-\rho\super{eq}\neq0$ such that $\delta S_{i}(t_{0})=S_{i}\left(t_{0}\right)-S_{i}\super{eq}\neq0$,
where ``eq'' corresponds to the final equilibrium state. We evolve
the density matrix through Eq.~\ref{eq:dynamics} using an adaptive
Runge-Kutta fourth-order method for a long enough simulation time,
typically from tens of ps to several ns, until the evolution of $S_{i}\left(t\right)$
can be reliably fitted by
\begin{align}
S_{i}\left(t\right)-S_{i}\super{eq}= & \left[S_{i}\left(t_{0}\right)-S_{i}\super{eq}\right]exp\left[-\frac{t-t_{0}}{\tau_{s,i}}\right]\nonumber \\
& \times\mathrm{cos}\left[\omega_{B}\left(t-t_{0}\right)+\phi\right].
\end{align}
to extract the relaxation time, $\tau_{s,i}$. Above, $\omega_{B}$
is oscillation frequency due to energy splitting in general, which
under an applied magnetic field $\vec{B}$ would include a contribution
$\approx0.5g_{s}\mu_{B}\left(\vec{B}\times\hat{\vec{S_{i}}}\right)$.
In order to examine whether the spin relaxation time depends on how
the spin imbalance is generated, we implement two general ways to
initialize $\delta\rho(t_{0})$. First, for simulating pump-probe
experiments, we choose $\delta\rho(t_{0})$ corresponding to interaction
with a pump pulse. Second, we use the technique proposed previously
in Ref. \citenum{xu2020spin} by applying a test magnetic field
at $t=-\infty$, allowing the system to equilibrate with a net spin
and then turning it off suddenly at $t_{0}$.
\subsection{Scattering terms}
The scattering part of the master equation can be separated into contributions
from several scattering channels,
\begin{align}
\frac{d\rho}{dt}|\sub{scatt}= & \sum_{c}\frac{d\rho}{dt}|_{c},
\end{align}
where $c$ labels a scattering channel. Under Born-Markov approximation,
in general we have\citep{rosati2014derivation}
\begin{align}
\frac{d\rho_{12}}{dt}|_{c}= & \frac{1}{2}\sum_{345}\left[\begin{array}{c}
\left(I-\rho\right)_{13}P_{32,45}^{c}\rho_{45}\\
-\left(I-\rho\right)_{45}P_{45,13}^{c,*}\rho_{32}
\end{array}\right]+H.C.,\label{eq:scattering}
\end{align}
where $P^{c}$ is the generalized scattering-rate matrix and H.C.
is Hermitian conjugate. The subindex, e.g., ``1'', is the combined
index of k-point and band. The weights of k points must be considered
when doing sum over k points. Note that $P^{c}$ in the interaction
picture is related to its value $P^{S,c}$ in the Schrodinger picture
as
\begin{align}
P_{1234}^{c}\left(t\right)= & P_{1234}^{S,c}\mathrm{exp}\left[i\left(\epsilon_{1}-\epsilon_{2}-\epsilon_{3}+\epsilon_{4}\right)t\right],
\end{align}
where $\epsilon_{i}$ are single-particle eigenvalues of $H_{0}$.
Below, we consider three separate scattering mechanisms - electron-impurity
(e-i), electron-phonon (e-ph) and electron-electron (e-e), and describe
the matrix elements for each.
For electron-phonon scattering, the scattering matrix is given by\citep{rosati2014derivation}
\begin{align}
P_{1234}\super{S,e\text{-}ph}= & \sum_{q\lambda\pm}A_{13}^{q\lambda\pm}A_{24}^{q\lambda\pm,*},\\
A_{13}^{q\lambda\pm}= & \sqrt{\frac{2\pi}{\hbar}}g_{12}^{q\lambda\pm}\sqrt{\delta_{\sigma}^{G}\left(\epsilon_{1}-\epsilon_{2}\pm\omega_{q\lambda}\right)}\sqrt{n_{q\lambda}^{\pm}},
\end{align}
where $q$ and $\lambda$ are phonon wavevector and mode, $g^{q\lambda\pm}$
is the electron-phonon matrix element, resulting from the absorption
($-$) or emission ($+$) of a phonon, computed with self-consistent
spin-orbit coupling from first-principles,\citep{giustino2017electron}
$n_{q\lambda}^{\pm}=n_{q\lambda}+0.5\pm0.5$ in terms of phonon Bose
factors $n_{q\lambda}$, and $\delta_{\sigma}^{G}$ represents an
energy conserving $\delta$-function broadened to a Gaussian of width
$\sigma$.
Next, for electron-impurity scattering, the scattering matrix is given
by
\begin{align}
P_{1234}\super{S,e\text{-}i}= & A_{13}^{i}A_{24}^{i,*},\\
A_{13}^{i}= & \sqrt{\frac{2\pi}{\hbar}}g_{13}^{i}\sqrt{\delta_{\sigma}^{G}\left(\epsilon_{1}-\epsilon_{3}\right)}\sqrt{n_{i}V\sub{cell}},\\
g_{13}^{i}= & \left\langle 1\right|V^{i}\left|3\right\rangle ,
\end{align}
where $n_{i}$ and $V\sub{cell}$ are impurity density and unit cell
volume, respectively, and $V^{i}$ is the impurity potential. In this
work, we deal with ionized and neutral impurities differently. For ionized impurities, $V^{i}$ is proportional
to screened Coulomb potential\citep{jacoboni2010theory}; for neutral impurities, we compute impurity potentials with supercell methods from DFT. (See Appendix
A for further details).
Finally, for electron-electron scattering, the scattering matrix is
given by\citep{rosati2014derivation}
\begin{align}
P_{12,34}\super{S,e\text{-}e}= & 2\sum_{56,78}\left(I-\rho\right)_{65}\mathscr{A}_{15,37}\mathscr{A}_{26,48}^{*}\rho_{78},\\
\mathscr{A}_{1234}= & \frac{1}{2}\left(A_{1234}-A_{1243}\right),\\
A_{1234}= & \frac{1}{2}\sqrt{\frac{2\pi}{\hbar}}\left[g_{1234}^{e-e}(\delta_{\sigma,1234}^{G})^{1/2}+g_{2143}^{e-e}(\delta_{\sigma,2143}^{G})^{1/2}\right],\\
g_{1234}\super{e\text{-}e}= & \left\langle 1\left(r\right)\right|\left\langle 2\left(r'\right)\right|V\left(r-r'\right)\left|3\left(r\right)\right\rangle \left|4\left(r'\right)\right\rangle ,
\end{align}
where $V\left(r-r'\right)$ is the screened Coulomb potential and
$\delta_{\sigma,1234}^{G}=\delta_{\sigma}^{G}\left(\epsilon_{1}+\epsilon_{2}-\epsilon_{3}-\epsilon_{4}\right)$
is a Gaussian-broadened energy conservation function. The screening
is described by Random-Phase-Approximation (RPA) dielectric function
(details in Appendix A). Although the above equations describe all
possible scattering processes between electrons and holes, we only
consider those between conduction electrons here, which are appropriate
for n-type Group III-V semiconductors\citep{jiang2009electron,mower2011dyakonov}.
The electron-hole scattering can be important for intrinsic and p-type
material.\citep{jiang2009electron,mower2011dyakonov} We note that
unlike the e-ph and e-i channels, $P\super{S,e\text{-}e}$ (as well as the dielectric screening in $V$) is a function
of $\rho$ and needs to be updated during time evolution of $\rho$.
This is a clear consequence of the two-particle nature of e-e scattering.
$P\super{S,e\text{-}e}$ can be written as the difference between
a direct term and an exchange term,
\begin{align}
P\super{S,e\text{-}e}= & P\super{S,e\text{-}e,d}-P\super{S,e\text{-}e,x},\\
P\super{S,e\text{-}e,d}= & \sum_{56,78}\left(I-\rho\right)_{65}A_{15,37}A_{26,48}^{*}\rho_{78},\\
P\super{S,e\text{-}e,x}= & \sum_{56,78}\left(I-\rho\right)_{65}A_{15,37}A_{26,84}^{*}\rho_{78}.
\end{align}
According to Ref. \citenum{rossi2002theory}, the direct term is
expected to dominate the dynamical scattering processes between conduction
or valence electrons, allowing us to neglect the exchange term here.
\subsection{Pump-probe simulation}
In nonrelativistic limit, the light-matter interaction
Hamiltonian operator ($\widehat{H}_{e-p}$) reads\cite{joly2009interaction}
\begin{align*}
\widehat{H}_{e-p}= & \frac{e}{m_{e}}\vec{A}\left(t\right)\cdot\widehat{\vec{p}}\\
& +\frac{e}{2m_{e}}\vec{A}\left(t\right)\cdot\vec{A}\left(t\right)+g_{e}\mu_{B}\widehat{\vec{s}}\cdot\left(\bigtriangledown\times\vec{A}\left(t\right)\right),
\end{align*}
where $\vec{A}\left(t\right)$ is the vector potential
and $\vec{A}\left(t\right)=\vec{A}_{0}\left(t\right)e^{-i\omega t}+\vec{A}_{0}^{*}\left(t\right)e^{i\omega t}$
with $\vec{A}_{0}\left(t\right)$ being the complex amplitude and
$\omega$ being photon frequency. $\widehat{\vec{p}}$ is momentum
operator. $g_{e}\approx2.0023192$ is anomalous gyromagnetic ratio.
The second quadratic term plays a role
only when pump fluence is higher by several orders of magnitude
than that in usual spin dynamics experiments and can be safely neglected.
Since $\bigtriangledown\times\vec{A}\left(t\right)=-i\vec{q}_{\mathrm{photon}}\times\vec{A}\left(t\right)$\cite{joly2009interaction}
and the photon wavevector $\vec{q}_{\mathrm{photon}}$ is quite small (the photon wavelength is much longer than the scale of unit cells),
the third term is also negligible. Therefore, we will only keep the
first $\vec{A}\left(t\right)\cdot\widehat{\vec{p}}$ term.
The interaction with a pump pulse of frequency $\omega\sub{pump}$
in the interaction picture is given by
\begin{align}
H_{\mathrm{pump},k,mn}\left(\omega\sub{pump},t\right)= & \frac{e}{m_{e}}\vec{A}_{0}\left(t\right)\cdot\vec{p}_{k,mn}e^{it\left(\epsilon_{m}-\epsilon_{n}-\omega\sub{pump}\right)}\nonumber \\
& +H.C.,
\end{align}
where $m,n$ represent the band indices and $k$ represents the k point
sampling in the first Brillouin zone. For a Gaussian pulse centered
at time $t\sub{center}$ with width $\tau\sub{pump}$,
\begin{align}
\vec{A}_{0}\left(t\right) & =\vec{A}_{0}\frac{\mathrm{1}}{\sqrt{\sqrt{\pi}\tau\sub{pump}}}\mathrm{exp}\left[-\left(t-t\sub{center}\right)^{2}/\left(2\tau\sub{pump}^{2}\right)\right].
\end{align}
Note that the corresponding pump fluence is $I\sub{pump}=\omega\sub{pump}^{2}|A_{0}|^{2}/\left(8\pi\alpha\right)$,
where $\alpha$ is fine structure constant. As a part of the coherent
portion of the time evolution, the dynamics due to this term are captured
directly in the Liouville form,\citep{d2020real,hannewald2000quantum}
\begin{align}
\frac{d\rho}{dt}|\sub{pump}= & -i[H\sub{pump},\rho].
\end{align}
The probe pulse interacts with the material similarly to the pump
pulse, and could be described in exactly the same way in principle.
However, this would require repeating the simulation for several values
of the pump-probe delay. Instead, since the probe is typically chosen
to be of sufficiently low intensity, we use second-order time-dependent
perturbation theory to capture its interaction with the system,
\begin{align}
\Delta\rho\super{probe}= & \frac{1}{2}\sum_{345}\left\{ \begin{array}{c}
\left[I-\rho\left(t\right)\right]_{13}P_{32,45}\super{probe}\rho\left(t\right)_{45}\\
-\left[I-\rho\left(t\right)\right]_{45}P_{45,13}\super{probe,*}\rho\left(t\right)_{32}
\end{array}\right\} +H.C.,\label{eq:drho_probe}
\end{align}
where $P\super{probe}$ is the generalized scattering-rate matrix
for the probe in the interaction picture. Its corresponding Schrodinger-picture
quantity is
\begin{align}
P_{1234}\super{S,probe}= & \sum_{\pm}A_{13}\super{probe,\pm}A_{24}\super{probe,\pm,*},\\
A_{13}\super{probe,\pm}= & \sqrt{\frac{2\pi}{\hbar}}\frac{e}{m_{e}}\left(\vec{A}_{0}\super{probe}\cdot\vec{p}\right)\sqrt{\delta_{\sigma}^{G}\left(\epsilon_{1}-\epsilon_{3}\pm\omega\sub{probe}\right)}.
\end{align}
The dielectric function change $\Delta\epsilon$ between the excited
state and ground state absorption detected by the probe is then
\begin{align}
\mathrm{Im}\Delta\epsilon= & \frac{2\pi}{\left(\omega\sub{probe}\right)^{3}|A_{0}\super{probe}|^{2}}\mathrm{Tr}\left(H_{0}\Delta\rho\super{probe}\right).
\end{align}
Note that $\Delta\rho\super{probe}$ contains $|A_{0}\super{probe}|^{2}$
so that $\mathrm{Im}\Delta\epsilon$ is independent of $A_{0}\super{probe}$.
The above $\mathrm{Im}\Delta\epsilon$ is a functional of the density
matrix according to Eq.~\ref{eq:drho_probe} and is an extension
of the usual independent-particle $\mathrm{Im}\epsilon$ depending
on just occupation numbers.\citep{molina2017ab} After computing $\mathrm{Im}\Delta\epsilon$
above, the real part $\mathrm{Re}\Delta\epsilon$ can be obtained
from the Krames-Kronig relation.
By summing up the dielectric function change $\Delta\epsilon$ computed
above with the dielectric function for ground state absorption, we
can obtain the excited-state $\epsilon$ as inputs for Kerr and Faraday
rotation calculations.\citep{mainkar1996first} These correspond to
the rotations of the polarization plane of a linearly polarized light,
reflected by (Kerr) and transmitted through (Faraday) the material,
after a pump excitation with a circularly-polarized light. Time-Resolved
Kerr/Faraday Rotation (TRKR/TRFR) has been widely used to study spin
dynamics of materials\citep{kikkawa1998resonant,kimel2001room}. In
a TRKR experiment, a circularly-polarized pump pulse is used to excite
valence electrons of the sample to conduction bands. The transitions
approximately satisfy the selection rule of $\Delta m_{j}=\pm1$ for
left and right circularly-polarized pulses, respectively, where $m_{j}$
is secondary total angular momentum. TRKR works by measuring the changes
of polarization of reflected light, which qualitatively is proportional
to the small population imbalance of electronic states with different
$m_{j}$.
Specifically, the Kerr rotation angle $\theta_{K}$ is computed with
dielectric functions by
\begin{align}
\theta_{K}= & \mathrm{Im\frac{\sqrt{\epsilon_{+}}-\sqrt{\epsilon_{-}}}{1-\sqrt{\epsilon_{+}}\sqrt{\epsilon_{-}}}},\label{eq:kerr}
\end{align}
where $\pm$ denotes the left and right circular polarization, respectively.
\section{Computational details}
The ground-state electronic structure, phonon, and e-ph matrix element
calculations of GaAs and few-layer WSe$_{2}$ are first calculated
using Density Functional Theory (DFT) with relatively coarse $k$
and $q$ meshes in the JDFTx plane-wave DFT code.\citep{sundararaman2017jdftx}
For GaAs, we use the experimental lattice constant of 5.653~$\text{Å}$,\citep{GaAs_lattice}
and select the SCAN exchange-correlation functional\citep{sun2015strongly}
for an accurate description of the electron effective mass (see section
II in Supplemental Materials\citep{supplementalmaterial}). We also
apply a scissor operator to the DFT values to reach experimental band
gap 1.43 eV\citep{madelung1987semiconductors}.
For WSe$_{2}$, we used PBE exchange correlation functional along with
the DFT-D2 pair potential dispersion corrections\citep{grimme2006semiempirical}.
The resulting
lattice constant is 3.32 $\text{Å}$ and distance between two W-atom planes is 6.419~$\text{Å}$
close to experimental values of bulk WSe$_2$, 3.297 and 6.491~$\text{Å}$\citep{agarwal1979growth}.
The phonon calculations
of GaAs and WSe$_{2}$ employ a $4\times4\times4$ and $6\times6$
supercell, respectively. We use Optimized Norm-Conserving Vanderbilt
(ONCV) pseudopotentials\citep{hamann2013optimized} with self-consistent
spin-orbit coupling throughout, which we find converged at a plane-wave
kinetic energy cutoff of 34 and 62 Ry for GaAs and WSe$_{2}$, respectively.
With these computational parameters, we find the effective mass of
conduction electrons of GaAs to be 0.054$m_{e}$, close to the experimental
value of 0.067$m_{e}$\citep{madelung1987semiconductors}. (More convergence
tests can be found in Supporting Information (SI)\citep{supplementalmaterial}).
We then transform all quantities from plane wave basis to maximally
localized Wannier function basis\citep{marzari1997maximally}, and
interpolate them\citep{PhononAssisted,giustino2017electron,GraphiteHotCarriers,brown2017experimental,NitrideCarriers,TAparameters}
to substantially finer k and q meshes. The Wannier interpolation approach
fully accounts for polar terms in the e-ph matrix elements and phonon
dispersion relations, using the approach developed by Verdi and Giustino\citep{verdi2015frohlich} for 3D and using the methods in Ref.~\citenum{Sohier2016} and Ref.~\citenum{Sohier2017} for 2D systems.
The Born effective charges and dielectric constants are calculated
from open-source code QuantumESPRESSO\citep{giannozzi2009quantum}.
For GaAs, the fine $k$ and $q$ meshes are $288\times288\times288$
for simulations at 300 K and are finer at lower temperature, e.g.,
$792\times792\times792$ for simulations at 30 K. This is necessary
to sample enough electronic states around band edges and for spin
lifetime convergence within 20$\%$. The $k$ and $q$ convergence
are easier for WSe$_{2}$ due to much larger effective masses and
we used $168\times168$ and $600\times600$ meshes at 50 and 10 K,
respectively. The computation of e-i and e-e matrix elements and the
real-time dynamics simulations are done with a new custom code interfaced
to JDFTx. The energy-conservation smearing parameter $\sigma$ is
chosen to be comparable or smaller than $k_{B}T$ for each calculation.
Detailed convergence tests of number of k points and energy window
for electronic states at various smearing parameters can be found
in Supplemental Materials\citep{supplementalmaterial}.
\section{Results and discussions\label{sec:results}}
\subsection{Applications to n-doped GaAs}
\subsubsection{Spin dynamics and its relation to TRKR\label{subsec:GaAs_ultrafast}}
\begin{figure}
\includegraphics[scale=0.4]{n_and_Sz} \caption{The energy-resolved dynamics of carriers (a) $\Delta n\left(\epsilon,t\right)\,=\,n\left(\epsilon,t\right)-n\left(\epsilon,0\right)$
and (b) spins $S_{z}\left(\epsilon,t\right)$ of conduction electrons
with a circularly polarized pump pulse centered
at 0.5 ps. The insets on the top right of both panels show $\Delta n\left(\epsilon,t\right)$
and $S_{z}\left(\epsilon,t\right)$ at $\epsilon$=1.45 eV. The pump
energy $\omega\sub{pump}$=1.47 eV is chosen to be higher than band
gap 1.43 eV\citep{madelung1987semiconductors}. The width of the pump
pulse $\tau\sub{pump}$ is 100 fs. The pump fluence $I\sub{pump}$ is
low at 0.01 $\mathrm{\mu J}$ cm$^{-2}$. The dynamics can be approximately
divided into three regions - Region I, II and III labeled in this
figure. In Region I, the system is excited by a pump pulse. In Region
II, pump processes are already finished, then both carriers and spins
relax simultaneously. In Region III, carrier distribution stays unchanged
while spins keep decaying.\label{fig:energy-resolved-dynamics}}
\end{figure}
In general, time evolution of Kerr rotation angle $\theta_{K}$ (see
Eq. \ref{eq:kerr}) is not equivalent to that of spin along the direction
of reflected light, and in fact, they can be quite different in some
cases\citep{dey2017gate}. There are few first-principles studies
of TRKR considering scattering processes in a form of semiclassical
Boltzmann equation\citep{molina2017ab}. A full quantum description
of scatterings with non-diagonal density matrix in TRKR has not been
presented in previous first-principles studies, to the best of our
knowledge. And the relation between dynamics of $\theta_{K}$ and
spin observable for general systems including GaAs has not yet been
well examined.
Using our density-matrix approach, we are able to directly simulate
the nonequilibrium ultrafast dynamics of optically excited systems
during which the dynamics of different electronic quantities such
as spin and carriers can be strongly coupled. We include all scattering
terms in a full quantum description as shown in the theory section
II.B and Appendix A. We perform the real-time dynamics simulations
of n-type GaAs for tens of ps at room temperature and several ns at
low temperature until the fitted spin lifetime does not change any
more. Having temporal density matrix, we can further analyze the dynamics
of various observables, including occupation, spin and Kerr rotation
angle easily. We then examine the relation between $\theta_{K}$ and
spin in the dynamics.
\begin{figure}
\includegraphics[scale=0.25]{kerr} \caption{(a) Compare the dynamics of Kerr rotation angle $\theta_{K}$ at different
probe energies $\omega\sub{probe}$ excited by a circularly
polarized pump pulse in first few ps. Fast oscillations in first
2 ps are due to the pump pulse and coupled spin and carrier relaxation.
(b) Compare relaxation of different observables - $\theta_{K}$ with
different $\omega\sub{probe}$ (denoted by black and red lines) and
$S_{z}$ with initial spin imbalance generated by a pump pule (Pump)
or a test magnetic field along z direction $\lambda_{s,z}\sim0.001-0.1$
Tesla (blue and green lines). $\omega\sub{pump}$=1.47 eV. The pump
pulse is centered at 0.5 ps. The longer time-scale dynamics (over
10 ps) for $\theta_{K}$ and $S_{z}$ has similar relaxation time
independent on the generation method of spin imbalance and specific
probe energies.
\label{fig:spin_generators_and_observables}}
\end{figure}
Figure \ref{fig:energy-resolved-dynamics} shows the energy-resolved
dynamics of carriers $\Delta n\left(\epsilon,t\right)\,=\,n\left(\epsilon,t\right)-n\left(\epsilon,0\right)$
and spins $S_{z}\left(\epsilon,t\right)$. The energy-resolved observable
$O\left(\epsilon\right)$ is defined as $\mathrm{Re}\left[\sum_{k,mn}o_{k,mn}\rho_{k,nm}\delta\left(\epsilon-\epsilon_{km}\right)\right]$,
where $o$ is operator matrix. We can see that during the first ps
(region I in Fig.~\ref{fig:energy-resolved-dynamics}), both observables
vary quickly due to the existence of the pump processes and both have
their maximum at an energy slightly lower than the pump energy, 1.47
eV (slightly larger than the band gap 1.43 eV), at a time shortly
after the time center of the pump pulse - 0.5 ps. Interestingly, after
pump being not active or after 0.8-1 ps, carriers and spins simultaneously
relax until 2-3 ps (region II in Fig.~\ref{fig:energy-resolved-dynamics}a
and \ref{fig:energy-resolved-dynamics}b). Afterward (region III in
Fig.~\ref{fig:energy-resolved-dynamics}a and \ref{fig:energy-resolved-dynamics}b),
carriers stay unchanged but spins $S_{z}\left(\epsilon,t\right)$
decay exponentially as shown in the insets of Fig. \ref{fig:energy-resolved-dynamics}a
and \ref{fig:energy-resolved-dynamics}b.
We have further analyzed the dynamics of Kerr rotation angle $\theta_{K}$
and compared it with spin dynamics. From Fig.~\ref{fig:spin_generators_and_observables}a,
we can see that during pump processes and shortly after them (from
0 to 2 ps), $\theta_{K}\left(t\right)$ has strong oscillations and
sensitive to the probe energy $\omega\sub{probe}$. The $\omega\sub{probe}$-sensitivity
may be partly attributed to the energy dependence of carrier and spin
dynamics. From Fig.~\ref{fig:spin_generators_and_observables}a and
\ref{fig:spin_generators_and_observables}b, it can be seen that after
3 ps (or in time region III defined in Fig.~\ref{fig:energy-resolved-dynamics}),
$\theta_{K}$ with different $\omega\sub{probe}$ decay exactly the
same. We can also find that with a pump pulse, relaxation time of
the Kerr rotation is the same as that of $S_{z}$, i.e. $\tau_{s,z}$.
Moreover, it turns out that $\tau_{s,z}$ does not depend on how spin
imbalance is generated - by a circularly polarized
pump pulse or by turning off a test magnetic field along $z$ direction
(see Sec. \ref{subsec:theory_dynamics}). This may indicate that if
the system is not extremely far from equilibrium, spin relaxation
along direction $i$ is not sensitive to the way of generating spin
imbalance, as long as the degrees of freedom other than $S_{i}$ are
not relevant or disappear in a short time. According to these observations,
hereinafter, we will do real-time dynamics starting from a $\delta\rho$
generated by turning off a test magnetic field and fit $\tau_{s,z}$
from time evolution of $S_{z}$.
\begin{figure}
\includegraphics[scale=0.35]{Compare-Exp} \caption{Theoretical spin lifetime with (black solid square) and without (black
empty square) the electron-electron scattering compared with experimental
data. Exp. A, B, C and D are experimental data from Refs. \citenum{kimel2001room,bungay1997direct,hohage2006coherent}
and \citenum{kikkawa1998resonant}, respectively.\label{fig:compare_exp}}
\end{figure}
We have also studied the effects of $\omega\sub{pump}$ and pump fluence
$I\sub{pump}$ on spin relaxation of $n$-GaAs at 300 K. We find that
$\omega\sub{pump}$ has very weak effects on spin relaxation but $\tau_{s,z}$
decreases with pump fluence. See more details in Appendix C.
\subsubsection{Temperature-dependence of spin lifetime and its dominant relaxation
mechanism\label{subsec:temperature}}
As discussed earlier, long-standing controversies remain for the dominant
spin relaxation mechanism of GaAs at different temperature and doping
level~\citep{jiang2009electron, mower2011dyakonov, kamra2011role, marchetti2014spin},
which will be resolved in the following sections. We start from study
$\tau_{s,z}$ of $n$-GaAs as a function of temperature at a moderate
doping level ($2\times10^{16}$ cm$^{-3}$). For simplicity, we assume
all impurities are fully ionized, so that the impurity density $n_{i}$
is equal to the free carrier density $n\sub{free}$. We first compared
our calculated spin lifetime with experimental results in Fig. \ref{fig:compare_exp}.
Our results of $\tau_{s,z}$ of $n$-GaAs give good agreement with
experiments at various temperatures\citep{kimel2001room,bungay1997direct,hohage2006coherent,kikkawa1998resonant}.
Different experiments have slight variations between each other due
to sample preparation conditions and specific measurement techniques.
The spin lifetime increases from tens of ps at room temperature to
tens of ns at low temperature. Note that e-e scattering plays an essential
role at low temperatures, i.e. by comparing with (black solid square)
and without (black empty square) in Fig.~\ref{fig:compare_exp}.
The correct temperature dependence of $\tau_{s,z}$ can be reproduced
only if e-e scattering is included.
\begin{figure}
\includegraphics[scale=0.35]{scattering_channels} \caption{Spin and carrier lifetimes of $n$-GaAs with $n_{i}=2\times10^{16}$
cm$^{-3}$ with different scattering mechanisms and different phonon
modes. In (a) and (b), ``All'' represents all the e-ph, e-i and
e-e scattering mechanisms being considered. TA, LA and LO represent
transverse acoustic, longitudinal acoustic and longitudinal optical
modes, respectively. The carrier lifetimes $\overline{\tau}_{p}$
present are the inverse of averaged carrier scattering rates $\left\langle \tau_{p}^{-1}\right\rangle $.
The method of carrier lifetime calculations is given in Appendix B.
$\langle\rangle$ means taking average around chemical potential $\mu$.
For a state-resolved quantity $A_{kn}$, its average is defined as
$\left\langle A\right\rangle =\sum_{kn}A_{kn}\left[f^{\mathrm{eq}}\right]'\left(\epsilon_{kn}\right)/\sum_{kn}\left[f^{\mathrm{eq}}\right]'\left(\epsilon_{kn}\right)$,
where $\left[f^{\mathrm{eq}}\right]'$ is the derivative of Fermi-Dirac
function.\label{fig:compare_spin_carrier_relaxation}}
\end{figure}
We further examine the contributions of different scattering mechanisms
to carrier and spin lifetime respectively, as a function of temperature.
Different from spin lifetime obtained from real-time DM dynamics including
all scattering processes simultaneously, the carrier lifetime ($\overline{\tau}_{p}$)
is defined through the inverse of the averaged carrier scattering
rate($\tau_{p}^{-1}$): $\overline{\tau}_{p}=1/\left\langle \tau_{p}^{-1}\right\rangle $.
Various scattering processes (e-e, e-i and e-ph) contribute to the
total carrier scattering rates through $\tau_{p}^{-1}=(\tau_{p}\super{e-e})^{-1}+(\tau_{p}\super{e-i})^{-1}+(\tau_{p}\super{e-ph})^{-1}$.
$\langle\rangle$ means taking average around chemical potential $\mu$.
For a state-resolved quantity $A_{kn}$, its average is defined as
$\left\langle A\right\rangle =\sum_{kn}A_{kn}\left[f^{\mathrm{eq}}\right]'\left(\epsilon_{kn}\right)/\sum_{kn}\left[f^{\mathrm{eq}}\right]'\left(\epsilon_{kn}\right)$,
where $\left[f^{\mathrm{eq}}\right]'$ is the derivative of Fermi-Dirac
function. For both carrier and spin lifetime, the lifetime due to
the most dominant scattering channel is the closest to the one including
all processes (black squares in both Fig.~\ref{fig:compare_spin_carrier_relaxation}a
and b). For spin relaxation in Fig.~\ref{fig:compare_spin_carrier_relaxation}a,
at low temperature below 50 K, e-e scattering is the most dominant
process as discussed above. However, the e-ph process becomes more
dominant above 100K. On the other hand, for carrier relaxation in
Fig.~\ref{fig:compare_spin_carrier_relaxation}b, the e-i process
is dominant over a wide temperature range from low to right below
room temperature. At room temperature, for both spin and carrier lifetimes,
the e-ph scattering is the most important process (closest to the
total lifetime with all scattering processes).
Our observations differ from those in Refs. \citenum{kamra2011role}
and \citenum{marchetti2014spin}, where the authors also found that
e-e scattering dominates spin relaxation at lower temperatures, e.g.
77 K, but their results showed that at room temperature e-e scattering
can be more important than other scatterings and enhances $\tau_{s,z}$
of $n$-GaAs by about 100$\%$ with moderate doping concentrations.
The overestimate of the effects of e-e scattering at room temperature
is most likely a limitation of the semiclassical method employed therein.
Similarly, we also find that different phonon modes can play different
roles in carrier and spin relaxations as shown in Fig. \ref{fig:compare_spin_carrier_relaxation}c
and \ref{fig:compare_spin_carrier_relaxation}d. For example, at room
temperature, LO (longitudinal optical) mode is most important for
carrier relaxation but seems less important than TA (transverse acoustic)
modes for spin relaxation. The situation is the opposite at 100 K
where TA/LO is most important for carrier/spin relaxation. Our finding
that TA modes are slightly more important than LO mode in spin relaxation
at room temperature is different from what have been believed in previous
model studies\citep{marchetti2014spin,jiang2009electron}, where they
declared that the electron-LO-phonon scattering dominates spin relaxation
at high temperatures especially at room temperature. This disparity
is most likely due to differences in the e-ph matrix elements and
electronic quantities, where we used fully first-principles approaches
instead of parameterized models in previous work.
In addition, we find the total spin lifetime is the longest when considering
all scattering processes in Fig.~\ref{fig:compare_spin_carrier_relaxation}a;
in contrast, the carrier lifetime is the shortest including all scattering
mechanism in Fig.~\ref{fig:compare_spin_carrier_relaxation}b. This
follows the inverse relation between spin and carrier lifetime in
the empirical D'yakonov--Perel' (DP) mechanism\citep{dyakonov1972spin,vzutic2004spintronics}
for systems without inversion symmetry, as will be discussed in more
details in next section.
\subsubsection{Doping-level-dependence of spin lifetime and its dominant relaxation
mechanism\label{subsec:doping}}
\begin{figure*}
\includegraphics[scale=0.24]{Doping_spin_lifetime} \includegraphics[scale=0.24]{Doping_carrier_lifetime}
\includegraphics[scale=0.24]{BinterSq} \caption{(a) Spin and (b) carrier lifetimes of $n$-GaAs with different doping
concentrations at 30 K with different scattering mechanisms. ``All''
represents all the e-ph, e-i and e-e scattering mechanisms being considered.
(c) $\left\langle \vec{\Omega}^{2}\right\rangle -\left\langle \Omega_{i}^{2}\right\rangle $
as a function of carrier density, where $\vec{\Omega}$ is the Larmor
frequency due to the ``internal'' magnetic field computed from first-principles,
which describes the SOC term induced by inversion asymmetry.\label{fig:doping}}
\end{figure*}
Figure~\ref{fig:doping} shows the carrier and spin lifetimes with
different doping density $n_{i}$ at 30 K with individual and total
scattering pathways, respectively. Similar to temperature dependence
and phonon contributions, it is also found that the roles of different
scattering mechanism differ considerably between spin and carrier
relaxation processes. Specifically, for the carrier relaxation in
Fig.~\ref{fig:doping}b, except when $n_{i}$ is very low (e.g. at
$10^{14}$ cm$^{-3}$), the electron-impurity scattering (e-i) dominates,
similar to the case of carrier lifetime over a large range of temperature
at a moderate doping in Fig.~\ref{fig:compare_spin_carrier_relaxation}b.
On the other hand, for the spin relaxation in Fig.~\ref{fig:doping}a,
the e-e scattering dominates except at very high concentration (above
$10^{17}$ cm$^{-3}$), while e-i scattering is only important in
the very high doping region (close to or above $10^{17}$ cm$^{-3}$).
Figure~\ref{fig:doping} shows the calculated $\tau_{s}$ has a maximum
at $n_{i}\,=\,1\text{-}2\times10^{16}$ cm$^{-3}$, and $\tau_{s}$
decreases fast with $n_{i}$ going away from its peak position. This
is in good agreement with the experimental finding in Ref. \citenum{kikkawa1998resonant},
which also reported $\tau_{s}$ at $n_{i}=10^{16}$ cm$^{-3}$ is
longer than $\tau_{s}$ at other lower and higher $n_{i}$ at a low
temperature (a few Kelvin). The $n_{i}$ dependence of $\tau_{s}$
may be qualitatively interpreted from the commonly used empirical
DP relation\citep{vzutic2004spintronics} for inversion-asymmetric
systems, $\tau_{s,i}\sim\tau_{s,i}\super{DP}=1/\left[\overline{\tau}_{p}\cdot\left(\left\langle \vec{\Omega}^{2}\right\rangle -\left\langle \Omega_{i}^{2}\right\rangle \right)\right]$,
where $\overline{\tau}_{p}$ is the carrier lifetime, and $\vec{\Omega}$
is the Larmor frequency due to the ``internal'' magnetic field,
which describes the SOC term induced by inversion asymmetry. For spin
1/2 systems, the internal magnetic field at ${\bf k}$ ($\vec{\Omega}_{{\bf k}}$)
will induce an energy splitting $\Delta_{{\bf k}}$ and polarize the
spin along the direction of $\vec{\Omega}_{{\bf k}}$. Previously,
$\vec{\Omega}_{{\bf k}}$ was mostly obtained with model Hamiltonian
with Dresselhaus SOC field\citep{dresselhaus1955spin}, which is rather
qualitative. Instead, we obtained k-dependent internal magnetic field
$\vec{\Omega}_{{\bf k}}$ from first-principles calculations, by using
$\Omega_{{\bf k},i}=2\Delta_{{\bf k}}\cdot s_{{\bf k},i}\super{exp}/\hbar$,
where $s_{{\bf k},i}\super{exp}$ is the spin expectation value.
From Fig.~\ref{fig:doping}, we find that with $n_{i}$ from $10^{14}$
cm$^{-3}$ to $5\times10^{15}$ cm$^{-3}$, carrier lifetime $\overline{\tau}_{p}$
decreases rapidly (black curve in Fig.~\ref{fig:doping}b) and $\left\langle \vec{\Omega}^{2}\right\rangle -\left\langle \Omega_{i}^{2}\right\rangle $
remains flat in Fig.~\ref{fig:doping}c, which may explain why spin
lifetime ($\tau_{s}$) increases in Fig.~\ref{fig:doping}a based
on the DP relation; however, when $n_{i}>10^{16}$ cm$^{-3}$, $\overline{\tau}_{p}$
decreases with a similar speed but $\left\langle \vec{\Omega}^{2}\right\rangle -\left\langle \Omega_{i}^{2}\right\rangle $
experiences a sharp increase, which may explain why spin lifetime
decreases in Fig.~\ref{fig:doping}b and owns a maximum at $10^{16}$
cm$^{-3}$.
Note that although the above empirical DP relation is intuitive to
understand the cause of doping-level dependence of spin lifetime,
it may break down when we evaluate individual scattering processes.
For example, when $n_{i}$ increases from $10^{14}$ cm$^{-3}$ to
$10^{15}$ cm$^{-3}$, both carrier lifetime $\overline{\tau}_{p}$
and spin lifetime $\tau_{s,z}$ due to e-i scattering decrease while
the internal magnetic field remains unchanged. Moreover, the simple
empirical relation cannot possibly explain our first-principles results
that the e-e and e-i scatterings have largely different contributions
in carrier and spin relaxation. First-principles calculations are
critical to provide unbiased mechanistic insights to spin and carrier
relaxation of general systems.
\subsection{Applications to few-layer WSe$_{2}$}
\subsubsection{Spin/valley relaxation of resident holes of monolayer WSe$_{2}$}
For holes of monolayer WSe$_2$, spin/valley relaxation is mostly
determined by intervalley spin-flip scattering processes between $K$ and $K'$
valleys because of the spin-valley locking. Previously, we reported spin/valley
lifetimes of resident holes of monolayer TMDs at T$\geq$50 K with
e-ph scattering\cite{xu2020spin}. At very low temperatures, e.g.,
10 K, intervalley e-ph scattering is however not activated as the
corresponding phonon occupation is negligible; therefore, other scattering
mechanisms are necessary to be included. Note that e-e scattering should not play
an important role in spin relaxation of holes of TMDs. The reason
is: The e-e scattering is a two-particle process where a transition
is accompanied by another transition with energy and momentum being
conserved. Considering the fact that only the highest occupied band
is involved (see band structure in Fig. S5) in dynamics of TMD holes, for an e-e process, a $K$$\rightarrow$$K'$
($K'$$\rightarrow$$K$) spin-flip transition must be accompanied
by an opposite $K'$$\rightarrow$$K$ ($K$$\rightarrow$$K'$) spin-flip
transition. Overall, e-e scattering processes have negligible contributions to spin relaxation
of TMD semiconductors. As a result, we will include only e-ph and e-i scatterings for WSe$_2$.
We use the supercell method to compute e-i scattering matrix elements
for neutral defects with self-consistent SOC and more details can be found in Appendix A.
\begin{figure}
\includegraphics[scale=0.15]{fig6}
\caption{
(a) The schematics of four types of impurities in WSe$_{2}$. (b)
Spin lifetimes of holes of monolayer WSe$_{2}$ with a relatively
low hole density $10^{11}$ cm$^{-2}$ with impurities compared with
experimental data. Exp. A, B, C and D are experimental data from Refs. \citenum{li2021valley}, \citenum{goryca2019detection}, \citenum{song2016long} and \citenum{yan2017long}, respectively.
The choices of impurity concentration $n_{i}$ of different impurities
are given in the main text.\label{fig:monolayer}}
\end{figure}
Experimentally several types of impurities/defects exist in TMD
samples. Here we pick four types of impurities with different symmetries
and chemical bonds (see Fig. \ref{fig:monolayer}(a)) - Se vacancy
($\mathrm{V_{Se}}$), two neighboring Se vacancies ($\mathrm{V_{2Se-N}}$), W vacancy ($\mathrm{V_{W}}$) and two Se vacancies with the same in-plane position ($\mathrm{V_{2Se-S}}$). As most point defects are relatively deep with large ionization energies in semiconducting TMDs\citep{wang2019excitation}, we mostly consider neutral defects here. According to Refs. \citenum{edelberg2019approaching,rhodes2019disorder,yankowitz2015local}, the impurity concentration
$n_{i}$ ranges from 8$\times$10$^{10}$ to 10$^{14}$ cm$^{-2}$
depending on samples. Considering that $\mathrm{V_{Se}}$ is often regarded as
the most abundant impurity, we choose a reasonable impurity density $n_{i}$ of $\mathrm{V_{Se}}$ - 7$\times$10$^{11}$ cm$^{-2}$, within the experimental range and for better comparison with experimental $\tau_{s}$ at T$\geq$20 K shown in Fig. \ref{fig:monolayer}(b).
$n_{i}$ of $\mathrm{V_{2Se-N}}$ is chosen
as 8$\times$10$^{9}$ cm$^{-2}$, two order of magnitude lower than $\mathrm{V_{Se}}$
because of its larger formation energy\cite{yankowitz2015local} and
better comparison with experimental $\tau_{s}$. $n_{i}$ of $\mathrm{V_{W}}$
and $\mathrm{V_{2Se-S}}$ are chosen arbitrarily as we find they have rather
weak effects on spin relaxation and are 7$\times$10$^{11}$ and 3.5$\times$10$^{11}$
cm$^{-2}$, respectively.
\begin{figure}
\includegraphics[scale=0.146]{fig7}
\caption{
(a) Schematic diagram of scattering pathways of excited holes in 1$\%$
compressed bilayer WSe$_{2}$ with a low equilibrium hole density
and 1.4$\times$10$^{12}$ cm$^{-2}$ $\mathrm{V_{Se}}$ (double of the monolayer value in Fig.~\ref{fig:monolayer}) under finite $E_{z}$ during and after a circularly polarized pump,
which initially excite holes (labelled as half circles)
at both $K$ and $K'$ valleys and two top valence bands (``A$\Downarrow$'' at $K'$ and ``B$\Downarrow$'' at $K$).
The lower-energy excited holes will decay to the band edge
(most to ``A$\Downarrow$'' at $K'$ and fewer to ``A$\Uparrow$'' at $K$)
through different scattering pathways.
A state being labeled by ``A''(``B'') means the wavefunction
of this state is mostly localized in layer A(B) of the bilayer.
$\Uparrow$ and $\Downarrow$ represent spin-up and spin-down, respectively.
The color of electronic state represents spin polarization - red and blue
mean spin-up and spin-down, respectively.
(b) Time evolutions of valley-
and band-resolved (excited or excess) hole densities at 50 K under
two $E_{z}$ with a circularly polarized pump. The pump energy is
selected to excite electrons at two top valence bands. The pump center
is at 0.5 ps. $n_{V}$ represents excess hole density at valley $V$.
The insets of panel (b) are the schematics of energies of two top
valence bands at $K$, $K'$ and $\Gamma$ valleys under two $E_{z}$ (see calculated band structures in SI Fig.~S7).
\label{fig:bilayer_dynamics}}
\end{figure}
From Fig. \ref{fig:monolayer}, we first find that
at T\textgreater 20 K, spin relaxation is almost driven
by e-ph scattering and impurities can only affect spin relaxation
at T$\leq$20 K. For the effects of different impurities on spin relaxation,
we have $\mathrm{V_{2Se-N}}$ $\gg$ $\mathrm{V_{Se}}$ $\gg$ $\mathrm{V_{W}}$ $\sim$ $\mathrm{V_{2Se-S}}$.
Such differences are directly related to the large differences among various impurities in
electron-impurity matrix elements for
the intervalley processes (scattering between $K$ and $K'$ valley), i.e. much larger matrix elements $|g_i|$ for intervalley scattering at $\mathrm{V_{2Se-N}}$ and $\mathrm{V_{Se}}$ compared to the ones at $\mathrm{V_{W}}$ and $\mathrm{V_{2Se-S}}$
as shown in SI Fig.~S9.
Moreover, the temperature dependence of $\tau_{s}$ with $\mathrm{V_{2Se-N}}$
is much weaker and in better agreement with experiments than that
with $\mathrm{V_{Se}}$. Therefore, the observed weak temperature dependence
in some experiments is probably related to the existence of larger size impurities
with lower symmetries (e.g. $\mathrm{V_{2Se-N}}$). Our observations suggest that the local symmetry
and chemical bonds surrounding an impurity have large impact on spin relaxation.
Additionally, we have simulated spin lifetimes of
monolayer WSe$_{2}$ at 15 K with different hole densities and find
a strong hole density dependence. The related results are shown in SI Fig. S12.
\subsubsection{Ultrafast dynamics of holes of bilayer WSe$_{2}$}
Understanding detailed dynamical processes and related scattering mechanism can help develop strategy of controlling and manipulating spin/valley relaxation through tuning external fields, materials composition, and strain. In the following, we will first identify the scattering pathways of excited holes and spins in bilayer WSe$_2$ (AB stacking as shown in Fig.~\ref{fig:bilayer_dynamics}a).
Next through real-time simulations, we determine dynamical quantities like spin lifetimes and valley polarization at different external fields and strain.
Finally, we show the carrier occupation on each layer is fully spin polarized and can be switched by an external electric field.
According to previous studies\cite{kim2021thickness,zhao2013origin} and our calculations, for valence bands
of unstrained bilayer WSe$_{2}$, $\Gamma$ valley is slightly higher
than $K$/$K'$ valley, which is usually undesirable.
The $K$/$K'$ valley can be pushed higher than $\Gamma$ valley by applying slight
in-plane compressive strain (corresponding band structures can be found in SI Fig.~S6).
Moreover, under zero $E_{z}$, the bilayer WSe$_2$ has inversion symmetry, leading to carriers being equally populated at $K$ and $K'$ valleys and at two layers, with spin up and down degeneracy.
A finite $E_z$ can break inversion symmetry (thus break Kramers degeneracy) and induce non-zero layer polarization or layer population difference. For example, two top valence bands from layer A with $\Uparrow$ (up spin) and layer B with $\Downarrow$ (down spin) are degenerate at $K$ without electric field but split under electric field.
Thus each band is associated with a particular spin channel, valley and layer, i.e. spin-valley-layer locking effect.
Also by tuning the sign and magnitude of $E_z$,
we are able to control various physical quantities like layer pseudospin, band splitting energy, etc.
Therefore, to ensure spin-valley-layer locking effects
being observed, we will study slightly compressed inversion-symmetric
bilayer WSe$_{2}$ under finite $E_{z}$.
Figure \ref{fig:bilayer_dynamics}(a) shows the scattering pathway schematics for hole bands (half circles) during the first few ps of slightly compressed (1$\%$) bilayer WSe$_{2}$ excited by a circularly polarized pump pulse under finite $E_{z}$ at
50 K, before exciton recombination processes happen typically at tens of ps timescale at this temperature. \cite{palummo2015exciton,goodman2017exciton}
Similar to the GaAs case (see Sec. \ref{subsec:GaAs_ultrafast}), the
hole spins will undergo the following processes: optical generation, decay to band edges and at the end slow relaxation.
Initially, holes with the same spin polarization are excited at
both $K$ and $K'$ valleys and two layers equally (e.g. down spin holes generated at $K$ valley and layer B and $K'$ valley and layer A as shown in Fig.~\ref{fig:bilayer_dynamics}(a)). During and after
the excitations, lower-energy holes at $K$ ($K'$) valley will decay
to the band edge through two possible
scattering pathways: (i) Direct pathway through interlayer spin-conserving
scattering (solid blue line); (ii) Indirect pathway through $\Gamma$-valley-related
scattering (dashed lines for both spin conserving and flip processes). After all holes decay to the band edge,
most of their carried spins are ``locked'' at a certain layer and valley (e.g., ``A$\Downarrow$'' at $K'$ in Fig.~\ref{fig:bilayer_dynamics}) due to weak intervalley spin-flip scattering (green arrow in Fig.~\ref{fig:bilayer_dynamics}(a)), which
is the so-called ``spin-valley-layer locking''.
\begin{figure}
\includegraphics[scale=0.13]{fig8}
\caption{
Spin lifetime and maximum valley polarization $P_{V}=|n_{K}-n_{K'}|/n_{tot}$
of bilayer WSe$_{2}$ at 50 K, where $n_{K(K')}$ is excess hole density
at $K\left(K'\right)$ valley and $n_{tot}$ is total excess hole density.
Negative strain means compressive strain.
$P_{V}$ will be 90$\%$ of its maximum shortly after the pump and
reach the maximum at a time from 1 to 20 ps depending on $E_z$ and the strain.
After reaching its maximum, since most carriers/spins have already decayed to the band edge, $P_{V}$ can relax only very slowly through intervalley spin-flip scattering.
This implies that having a high maximum of $P_{V}$ is important
to ensure a high $P_{V}$ during a long time.
}\label{fig:spin_lifetime_and_valley_pol}
\end{figure}
In Fig. \ref{fig:bilayer_dynamics}(b), we show time evolutions of
valley- and band-resolved (excited or excess) hole densities under
two $E_{z}$. It can be seen that under a low $E_{z}$ (0.5 V/nm, Fig. \ref{fig:bilayer_dynamics}(b) left panel),
the main scattering pathway is the direct one mentioned above, i.e. the spin down holes scattered from B $\Downarrow$ at $K$ in dashed blue line with decreasing population to A $\Downarrow$ at $K'$ in solid blue line with increasing population. Under
a higher $E_{z}$ (2 V/nm, Fig. \ref{fig:bilayer_dynamics}(b) right panel), although the direct scattering pathway
still exists, the indirect one through the $\Gamma$ valley also becomes important because the band energy at $\Gamma$ is pushed higher than the second valence band at $K$ and $K'$ under this electric field (see inset of Fig. \ref{fig:bilayer_dynamics}(b) right panel). Here occupation at B $\Downarrow$ at $K$ in dashed blue line rapidly decreased while occupation at $\Gamma$ with both up and down spins in solid black temporarily increased through indirect scattering,
and most importantly the A $\Uparrow$ at $K$ in solid red also increased due to the scattering through $\Gamma$ valley. Increased population at A $\Uparrow$ at $K$ represents weakening the spin-valley-locking effect. Therefore, the indirect scattering pathway will lead to the reduction of spin density and valley polarization.
\begin{figure}
\includegraphics[scale=0.35]{fig9}
\caption{
Time evolution of top layer hole occupation $f_{h}^{\mathrm{top}}$
and spin $s_{z}^{\mathrm{top}}$ (see their definitions in Appendix D)
normalized by corresponding total quantities of compressed bilayer WSe$_{2}$ at 50 K after the sign
of $E_{z}$ is switched at $t=0$. At $t=-10$ ps, a pump pulse centered
at $t=-9.5$ ps with $\tau_{\mathrm{pump}}=100$ fs is started to
be applied and real-time density matrix dynamics is run under $E_{z}=0.5$
V/nm until $t=0$ to allow holes/spins to decay to the band edge where
states are localized in bottom layer. When the sign of $E_{z}$ is
suddenly switched, holes/spins are still localized in the same bottom
layer but the new eigenstates around the band edge are localized in
top layer, thus holes/spins will transfer from bottom to top layer.
\label{fig:switching}
}
\end{figure}
We then show two key dynamical quantities - spin lifetime $\tau_{s}$ in Fig.~\ref{fig:spin_lifetime_and_valley_pol}(a) and maximum
valley polarization of bilayer WSe$_{2}$ in Fig.~\ref{fig:spin_lifetime_and_valley_pol}(b) as a function of $E_{z}$
with different percentages of strain. Valley polarization is defined as $P_{V}=|n_{K}-n_{K'}|/n_{tot}$,
where $n_{K(K')}$ is excess hole density at $K\left(K'\right)$ valley
and $n_{tot}$ is total excess hole density.
A high maximum of $P_{V}$ is necessary to ensure a high $P_{V}$ for extended time.
This is because: when $P_{V}$ reaches its maximum, since most carriers/spins have already decayed to the band edge at the same time,
$P_{V}$ will decay very slowly through intervalley spin-flip scattering.
Obviously, unstrained system is not suitable to utilize spin-valley-layer locking considering
its short spin lifetimes and low maximum $P_{V}$ under a relatively low electric field. As we discussed
above, a slight strain is helpful to ensure $K/K'$ valleys are sufficiently
higher than $\Gamma$ valley. Moreover, from Fig. \ref{fig:spin_lifetime_and_valley_pol}, an optimized range of $E_{z}$ is from $\sim$0.2 to $\sim$2
V/nm, where the system shows long $\tau_{s}$ and high maximum $P_{V}$, consistent
with the fact mentioned above: indirect scattering pathway, which becomes
important under high $E_{z}$, will cause loss of spin and
valley polarization.
Finally, in a bilayer, the percentage of carriers/spins localized in
top/bottom layer at equilibrium can be controlled by a static $E_{z}$.
For device applications, it may be desirable to dynamically and spatially
tune the locations of holes/spins by controlling $E_{z}$. For this
purpose, a knowledge of the speed of carriers/spins transferring between
two layers can be useful.
To extract the time scale of such transfer, we first generate
holes/spins by applying a circularly polarized pump pulse centered
at $-9.5$ ps and let the system evolves until $t=0$ under $E_{z}=0.5$
V/nm to ensure almost all holes/spins being localized in bottom layer,
then by switching $E_{z}$ suddenly to $-0.5$ V/nm at $t=0$, holes/spins
will start to transfer to top layer.
In Fig. \ref{fig:switching}, we show time evolution of
top layer hole occupation $f_{h}^{\mathrm{top}}$ and spin $s_{z}^{\mathrm{top}}$
(see their definitions in Appendix D)
normalized by corresponding total quantities of 2\% compressed bilayer
after the sign of $E_{z}$ is switched.
From Fig. \ref{fig:switching}, at 50 K, 90\% switching of $f_{h}^{\mathrm{top}}$
and $s_{z}^{\mathrm{top}}$ takes $\sim$6 ps. This time constant
is much shorter than $\tau_{s}$, which means such tuning is fast enough to use an electric field as a "switch" in spintronic devices.
\section{Conclusions}
In this article, we present a first-principles real-time density-matrix
approach to simulate ultrafast spin-orbit-mediated spin dynamics in
solids with arbitrary crystal symmetry. The complete \emph{ab initio}
descriptions of pump, probe and three scattering processes - the electron-phonon,
electron-impurity and electron-electron scattering in the density-matrix
master equation, allows us to directly simulate the nonequilibrium
ultrafast pump-probe measurements and makes our method applicable
to any temperatures and doping levels. This method has been applied
to simulate spin relaxation of $n$-GaAs. We confirm that relaxation
time of Kerr rotation and that of spin observables are almost identical
and find that relaxation time of spin polarization is relatively robust,
i.e. insensitive to how spin imbalance is initialized. Furthermore,
we have studied the temperature and doping-level dependencies of spin
lifetime and examined the roles of various scattering mechanisms.
Overall our theoretical results are in good agreement with experiments.
Importantly, our first-principles simulations provide rich mechanistic
insights of spin relaxation of $n$-GaAs: we point out that although
at low temperatures and moderate doping concentrations e-i scattering
dominates carrier relaxation, e-e scattering is the most dominant
process in spin relaxation. The relative contributions of phonon modes
also vary considerably between spin and carrier relaxation. We have
further examined ultrafast dynamics in few-layer WSe$_{2}$ with
realistic impurities. We find that spin relaxation can highly depend on local symmetry and chemical bonds
surrounding impurities. For the bilayer, we identify the scattering
pathways of holes in ultrafast dynamics and determine relevant dynamical
properties, including $\tau_{s}$, maximum valley polarization and
layer population/spin switch time, which are essential to utilize
its unique spin-valley-layer locking effects. Our method opens up
the pathway to predict spin relaxation and decoherence for general
materials and provide unbiased insights and guidelines to experimental
materials design, which have the potential to revolutionize the field
of spintronics and quantum information technologies.
\section*{Acknowledgements}
We thank Hiroyuki Takenaka for helpful discussions. This work is supported
by National Science Foundation under grant No. DMR-1956015 and the Air Force Office of Scientific Research under AFOSR Award No. FA9550-YR-1-XYZQ. A. H.
acknowledges support from the American Association of University Women(AAUW)
fellowship program. This research used resources of the Center for
Functional Nanomaterials, which is a US DOE Office of Science Facility,
and the Scientific Data and Computing center, a component of the Computational
Science Initiative, at Brookhaven National Laboratory under Contract
No. DE-SC0012704, the lux supercomputer at UC Santa Cruz, funded by
NSF MRI grant AST 1828315, the National Energy Research Scientific
Computing Center (NERSC) a U.S. Department of Energy Office of Science
User Facility operated under Contract No. DE-AC02-05CH11231, the Extreme
Science and Engineering Discovery Environment (XSEDE) which is supported
by National Science Foundation Grant No. ACI-1548562 \citep{xsede},
and resources at the Center for Computational Innovations at Rensselaer
Polytechnic Institute.
\section*{Appendix A: Interaction Hamiltonian terms and matrix elements}
Three interaction Hamiltonian terms in Eq. \ref{eq:H} read
\begin{align}
H\sub{e\text{-}ph}= & \sum_{12q\lambda}c_{1}^{\dagger}c_{2}\left(g_{12}^{q\lambda-}b_{q\lambda}+g_{12}^{q\lambda+}b_{q\lambda}^{\dagger}\right),\\
H\sub{e\text{-}i}= & n_{i}V\sub{cell}\sum_{12}c_{1}^{\dagger}c_{2}g_{12}^{i},\\
H\sub{e\text{-}e}= & \sum_{1234}c_{1}^{\dagger}c_{2}^{\dagger}c_{3}c_{4}g_{1234}\super{e\text{-}e}.
\end{align}
The e-ph matrix $g^{q\lambda\pm}$ are computed with self-consistent
spin-orbit coupling and Wannier interpolation by the supercell method.
Here we assume impurity density is sufficiently low
and the average distance between neighboring impurities is sufficiently long
so that the interactions between impurities are negligible.
The e-i matrix $g^{i}$ is
\begin{align}
g_{13}^{i}= & \left\langle 1\right|\Delta V^{i}\left|3\right\rangle ,\\
\Delta V^{i}= & V^{i}-V^{0},
\end{align}
where $V^{i}$ is the potential of the impurity system
and $V^{0}$ is the potential of the pristine system. In this work, $g^{i}$ for neutral and ionized impurities
are computed differently as follows.
For neutral impurities, $V^{i}$ is computed using a large supercell including an impurity with self-consistent SOC at DFT. This is important for including the detailed potential profile and chemical bonding environment for different impurities.
To speed up the
supercell convergence, we used the potential alignment method developed
in Ref. \citenum{sundararaman2017first}.
We checked the supercell
size convergence of $g^{i}$ for neutral $\mathrm{V_{Se}}$ in monolayer WSe$_{2}$
and found that the corresponding spin lifetime with 6$\times$6
and 8$\times$8 supercells differs by only a few percent.
Thus
6$\times$6 supercell is enough for $g^{i}$ of
neutral impurities in monolayer WSe$_{2}$.
For ionized impurities, we use approximate impurity potentials as detailed below.
In general, $\Delta V^{i}$ may be separated into
two terms $\Delta V^{i}=\Delta V_{\mathrm{ns}}^{i}+\Delta V_{\mathrm{soc}}^{i},$
where $\Delta V_{\mathrm{ns}}^{i}$ is the spin-independent part and
$\Delta V_{\mathrm{soc}}^{i}=\left[\hbar/\left(4m^{2}c^{2}\right)\right]\nabla\left(\Delta V_{\mathrm{ns}}^{i}\right)\times\vec{p}\cdot\vec{\sigma}$
is the SOC correction. For ionized impurities, we approximate $\Delta V_{\mathrm{ns}}^{i}$
as the potential of point charge and is simply the product of the
impurity charge $Z$ and the screened Coulomb potential\citep{jacoboni2010theory}, i.e., $\Delta V_{\mathrm{ns}}^{i}=ZV^{\mathrm{scr}}$.
Such approximate describes the long-range part of spin-independent differential impurity potential $\Delta V_{\mathrm{ns}}^{i}$
accurately, which is often the most important contribution from ionized impurities. Considering that $\Delta V_{\mathrm{soc}}^{i}$ is commonly
neglected in previous theoretical studies on spin relaxation\cite{tamborenea2003spin,jiang2009electron}
whenever screened Coulomb potential $V^{\mathrm{scr}}$ is used, we will not include $\Delta V_{\mathrm{soc}}^{i}$
for ionized impurities either. Note that such potential for ionized impurities still relaxes spin through spin-mixing and spin-precession.
\begin{comment}
A more rigorous treatment of $\Delta V^{i}$ of ionized
impurities is: we separate $\Delta V^{i}$ into the short-range $V^{\mathrm{SR}}$
and long-range part $V^{\mathrm{LR}}$ similarly to the electron-phonon
interaction\cite{verdi2015frohlich}. $V^{\mathrm{LR}}$ can be computed
based on the screened Coulomb potential as above and $V^{\mathrm{SR}}$ can
be computed using the supercell method as in the neutral impurity
case. Such treatment will be our future work.
\end{comment}
The e-e matrix $g\super{e\text{-}e}$ is
\begin{align}
g_{1234}\super{e\text{-}e}= & \left\langle 1\left(r\right)\right|\left\langle 2\left(r'\right)\right|V\left(r-r'\right)\left|3\left(r\right)\right\rangle \left|4\left(r'\right)\right\rangle ,
\end{align}
where $V\left(r-r'\right)$ is the screened electron-electron interaction.
The SOC corrections on $V\left(r-r'\right)$ \citep{aryasetiawan2008generalized,grimaldi1997theory} will not be included
similar to the ionized impurity case.
Thus, $V\left(r-r'\right)$ is simply the screened Coulomb potential $V^{scr}$.
Therefore, in both calculations of $g^{i}$ and $g\super{e\text{-}e}$, the computation of the screened Coulomb potential $V^{scr}$ is of key importance.
Currently, we use the static RPA (Random Phase Approximation) dielectric
function for the screening and neglect local-field effects. We then
show the e-e self-energy ($\mathrm{Im\Sigma}$) obtained with such
dielectric function well reproduces the one obtained with dynamically
screened Coulomb interaction with full RPA dielectric matrix in the
relevant energy range as shown in Fig.~\ref{fig:ImSigma_benchmark}.
The dielectric function has the form
\begin{align}
\epsilon\left(\vec{q}\right)= & \epsilon_{s}\epsilon\super{intra}\left(\vec{q}\right),
\end{align}
where $\epsilon_{s}$ is the static background dielectric constant
and can be calculated by Density Functional Perturbation Theory (DFPT)\citep{wu2005systematic}.
$\epsilon\super{intra}\left(\vec{q}\right)$ is the intraband contribution
which involves only states with free carriers and is critical for
doped semiconductors. It is computed using Random Phase Approximation
(RPA),
\begin{align}
\epsilon\super{intra}\left(\vec{q}\right)= & 1-V\super{bare}\left(\vec{q}\right)\sum_{\vec{k}mn}\left(\begin{array}{c}
\frac{f_{\vec{k}-\vec{q},m}-f_{\vec{k}n}}{\epsilon_{\vec{k}-\vec{q},m}-\epsilon_{\vec{k},n}}\times\\
|\left\langle u_{\vec{k}-\vec{q},m}|u_{\vec{k}n}\right\rangle |^{2}
\end{array}\right),
\end{align}
where the sum runs over only states having free carriers, e.g., for
a n-doped semiconductor, $m$ and $n$ are conduction band indices.
In the above formula, $f$ is time-dependent non-equilibrium
occupation instead of the equilibrium one $f^{\mathrm{eq}}$. Therefore,
if pump is activated or optical field $\vec{A}_{0}\left(t\right)$ of
the pump pulse is not negligible, $\epsilon\super{intra}\left(\vec{q}\right)$
will be updated in every time step, as $f$ will differ from $f^{\mathrm{eq}}$
and the magnitude of difference depends on the excitation density. $V\super{bare}\left(\vec{q}\right)=e^{2}/\left(V\sub{cell}\varepsilon_{0}|q|^{2}\right)$
is the bare Coulomb potential with $V\sub{cell}$ the unit cell volume
and $\varepsilon_{0}$ vacuum permittivity. $u_{\vec{k}n}$ is the
periodic part of the Bloch wave function.
We then have the matrix elements in reciprocal space,
\begin{align}
g_{13}^{i}= & ZV\super{scr}\left(\vec{q}_{13}\right)\left\langle u_{1}|u_{3}\right\rangle ,\\
g_{1234}\super{e\text{-}e}= & V\super{scr}\left(\vec{q}_{13}\right)\delta_{\vec{k}_{1}+\vec{k}_{2},\vec{k}_{3}+\vec{k}_{4}}\left\langle u_{1}|u_{3}\right\rangle \left\langle u_{2}|u_{4}\right\rangle ,\\
V\super{scr}\left(\vec{q}_{13}\right)= & V\super{bare}\left(\vec{q}_{13}\right)/\epsilon\left(\vec{q}_{13}\right),\label{eq:scr}
\end{align}
where $V\super{scr}\left(\vec{q}\right)$ is the screened Coulomb
potential and $\vec{q}_{13}=\vec{k}_{1}-\vec{k}_{3}$. $\delta_{\vec{k}_{1}+\vec{k}_{2},\vec{k}_{3}+\vec{k}_{4}}$
is Kronecker delta function and means $\vec{k}_{1}+\vec{k}_{2}=\vec{k}_{3}+\vec{k}_{4}$.
$\left\langle u_{1}|u_{3}\right\rangle $ is the overlap matrix element
between two periodic parts of the Bloch wave functions.
\section*{Appendix B: Carrier scattering rate and Im$\Sigma$ from the density-matrix
approach}
At the semiclassical limit, density matrix $\rho$ is replaced by
(non-equilibrium) occupation $f$, then the scattering
term originally with a full quantum description in Eq.~\ref{eq:scattering}
required by DM dynamics becomes:
\begin{align}
\frac{df_{1}}{dt}|_{c}= & \mathop{\sum_{2\neq1}}\left[\left(1-f_{1}\right)P_{11,22}^{c}f_{2}-\left(1-f_{2}\right)P_{22,11}^{c}f_{1}\right],\label{Eq.f1}
\end{align}
using the facts that $P_{11,22}$ is real and ``2=1'' term is zero.
``c'' represent a scattering channel. Note that the weights of k
points must be considered when doing sum over k points.
Suppose $f$ is perturbed from its equilibrium value by $\delta f$,
i.e., $f=f\super{eq}+\delta f$, then insert $f$ after perturbation
into Eq.~\ref{Eq.f1} and linearize it,
\begin{align}
\frac{df_{1}}{dt}|_{c}= & -\mathop{\sum_{2\neq1}}\left[P_{11,22}^{c}f_{2}^{\mathrm{eq}}+\left(1-f_{2}^{\mathrm{eq}}\right)P_{22,11}^{c}\right]\delta f_{1},
\end{align}
using the fact that $\delta P_{11,22}$ is always zero, even for the
e-e scattering.
Define carrier relaxation time of state ``1'' $\tau_{p,1}^{c}$
by $\frac{df_{1}}{dt}|_{c}=-\frac{\delta f_{1}}{\tau_{p,1}^{c}}$,
we have
\begin{align}
\frac{1}{\tau_{p,1}^{c}}= & \mathop{\sum_{2\neq1}}\left[P_{11,22}^{c}f_{2}^{\mathrm{eq}}+\left(1-f_{2}^{\mathrm{eq}}\right)P_{22,11}^{c}\right].\label{eq:carrier_lifetime}
\end{align}
The linewidth or the imaginary part of the self-energy for the scattering
channel $c$ is related to the carrier relaxation time by $\mathrm{Im}\Sigma_{1}^{c}=\hbar/\left(2\tau_{p,1}^{c}\right)$.
Using Eq.~\ref{eq:carrier_lifetime}, we have calculated the e-ph
scattering rates and they are in good agreement with previous theoretical
results \citep{zhou2016ab}. For e-ph scattering, Eq.~\ref{eq:carrier_lifetime}
will reproduce the imaginary part of the well-known Fan-Migdal self-energy\citep{giustino2017electron}.
For e-i scattering, we have
\begin{align}
\frac{1}{\tau_{p,1}\super{e\text{-}i}}= & \frac{2\pi}{\hbar}n_{i}V\sub{cell}\sum_{2}|g_{12}^{i}|^{2}\delta_{\sigma}^{G}\left(\epsilon_{1}-\epsilon_{2}\right).\label{tau-e-i}
\end{align}
The above equation (Eq.~\ref{tau-e-i}) is consistent with Ref. \citenum{jacoboni2010theory}.
\begin{figure}
\includegraphics[scale=0.38]{ImSigma_benchmark} \caption{$\mathrm{Im\Sigma}$ due to e-e scattering of valence electrons of
p-type silicon computed by Eq.~\ref{eq:ee_lifetime} (Density-Matrix)
compared with those calculated by the finite-temperature GW method
(FT-GW) \citep{benedict2002quasiparticle}. $\mu$ is set to 0.05
eV lower than Valence Band Maximum (VBM). For simplicity, SOC is not
considered in this test.\label{fig:ImSigma_benchmark}}
\end{figure}
For e-e scattering, neglecting the exchange contribution, which is
a commonly-used approximation\cite{rossi2002theory,benedict2002quasiparticle},
\begin{align}
\frac{1}{\tau_{p,1}\super{e\text{-}e}}= & \frac{2\pi}{\hbar}\sum_{2\neq1,34}|A_{1324}|^{2}\left[\begin{array}{c}
f_{2}^{\mathrm{eq}}f_{4}^{\mathrm{eq}}\left(1-f_{3}^{\mathrm{eq}}\right)+\\
\left(1-f_{2}^{\mathrm{eq}}\right)f_{3}^{\mathrm{eq}}\left(1-f_{4}^{\mathrm{eq}}\right)
\end{array}\right].\label{eq:ee_lifetime}
\end{align}
To verify our implementation of e-e scattering term, we have calculated
$\mathrm{Im\Sigma}$ due to e-e scattering of valence electrons of
p-type silicon based on the above equation and compare it with those
calculated by the finite-temperature GW method from first-principles,
\citep{benedict2002quasiparticle} as implemented in JDFTx.\citep{sundararaman2017jdftx}
The JDFTx implementation, in turn, has been benchmarked to reproduce
the expected dependence with temperature and carrier energy, Im$\Sigma\super{e\text{-}e}\propto(\varepsilon-\varepsilon_{F})^{2}+(\pi k_{B}T)^{2}$,
as expected for metals.\citep{TAparameters}
From Fig.~\ref{fig:ImSigma_benchmark}, we can see the results by
two methods agree well for the energy range close to the Fermi level
which is relevant to e-e scatterings due to energy conservation.
This verifies our implementation of e-e scattering part.
\section*{Appendix C: The effects of $\omega\sub{pump}$ and pump fluence on
spin relaxation of GaAs at 300 K}
\begin{figure}
\includegraphics[scale=0.38]{pumpE}
\caption{$S_{z}\left(t\right)$ of $n$-GaAs with $n_{i}=2\times10^{16}$ cm$^{-3}$
at 300 K with different pump pulse energies ($\omega\sub{pump}$)
varying with several $k_{B}T$. \label{fig:pumpE}}
\end{figure}
In Fig.~\ref{fig:pumpE}, we study the $S_{z}$ relaxation dependence
on pump-pulse energy changes with several $k_{B}T$. We can see that
variation of $\omega\sub{pump}$ has very weak effects on spin dynamics
of $n$-GaAs at 300 K.
\begin{figure}
\includegraphics[scale=0.24]{fig12a.png} \includegraphics[scale=0.24]{fig12b.png}
\caption{(a) The excitation density as a function of the pump fluence (left panel)
and (b) the spin lifetime as a function of the excitation density
generated by a circularly polarized pump pulse for
$n$-GaAs with $n_{i}=10^{14}$ cm$^{-3}$ at 300 K. $\omega\sub{pump}$=1.47
eV.\label{fig:pumpfluence}}
\end{figure}
In Fig.~\ref{fig:pumpfluence}a, we study the effects of the pump fluence
$I\sub{pump}$ on spin relaxation. Firstly, we can see that in the
low pump fluence region or when $I\sub{pump}<1$ $\mu$J/cm$^{-2}$,
the excitation density increases linearly with $I\sub{pump}$ but
when $I\sub{pump}>1$ $\mu$J/cm$^{-2}$, the excitation density increases
slower. This is because in high fluence cases, during the excitation
by a pump pulse, a significant amount of conduction states have been
already filled, which reduce the probability of the transitions from
valence bands to conduction bands. From Fig.~\ref{fig:pumpfluence}b,
we find that spin lifetime of $n$-GaAs decreases with the excitation
density. This dependence may be explained based on the empirical DP
relation\citep{vzutic2004spintronics} $\tau_{s,i}\sim\tau_{s,i}^{DP}=1/\left[\overline{\tau}_{p}\cdot\left(\left\langle \vec{\Omega}^{2}\right\rangle -\left\langle \Omega_{i}^{2}\right\rangle \right)\right]$
as we discussed in Sec.~\ref{subsec:temperature}. At 300 K, generally
$\left\langle \vec{\Omega}^{2}\right\rangle -\left\langle \Omega_{i}^{2}\right\rangle $
will increase with increasing free carrier density (through an increase
of excitation density here), similar to what we find at 30 K shown
in Fig. \ref{fig:doping}c. On the other hand, $\overline{\tau}_{p}$
due to the electron-phonon scattering, which dominates carrier relaxation
at 300 K, is less sensitive to the variation of excitation density.
Therefore, it is the increase of $\left\langle \vec{\Omega}^{2}\right\rangle -\left\langle \Omega_{i}^{2}\right\rangle $
causing the decrease of spin lifetime when increasing excitation density.
\section*{Appendix D: Layer-resolved quantities of bilayer WSe$_{2}$}
For a bilayer with one layer above $z=0$ and another
below, the operator $\hat{l}^{\mathrm{top}}$ projecting any local
quantity $A(\vec{r})$ to the top layer can be defined through the
relation
\begin{align}
\hat{l}^{\mathrm{top}}A(\vec{r})= & H\left(z\right)A(\vec{r}),
\end{align}
where $z$ is the third component of $\vec{r}$ and
$H\left(z\right)$ is Heaviside step function
\begin{align}
H\left(z\right)= & \{\begin{array}{c}
1,z>0\\
0,z\leq0
\end{array}.
\end{align}
Therefore, top layer hole occupation $f_{h}^{\mathrm{top}}$ is
\begin{align}
f_{h}^{\mathrm{top}}= & \mathrm{Tr}\left(\hat{l}^{\mathrm{top}}\left(1-f\right)\right)\\
= & \sum_{kn}l_{k,nn}^{\mathrm{top}}\left(1-f_{kn}\right),
\end{align}
where $f$ is occupation and $l_{k,mn}^{\mathrm{top}}=\left\langle km\right|\hat{l}^{\mathrm{top}}\left|kn\right\rangle =\left\langle km\right|H\left(z\right)\left|kn\right\rangle $.
$\mathrm{Tr}$ means taking trace. $k$ is k-point index. $n$ and
$m$ are band indices.
Moreover, top layer spin $s_{z}^{\mathrm{top}}$
can be defined using the operator $\hat{l}^{\mathrm{top}}\hat{s}_{z}$,
\begin{align}
s_{z}^{\mathrm{top}}= & \mathrm{Tr}\left(\hat{l}^{\mathrm{top}}\hat{s}_{z}\hat{\rho}\right)\\
= & \sum_{k,lmn}l_{k,lm}^{\mathrm{top}}s_{z,k,mn}\rho_{k,nl}.
\end{align}
Note that the commutator $\left[l^{\mathrm{top}},s_{z}\right]$
is found numerically close to zero for bilayer WSe$_2$.
This indicates that we can safely define $s_{z}^{\mathrm{top}}$
as an observable using the above equations.
| 791ed3f23a2a9d3fda82e280f22152a6ef5186ad | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
For a number $h\in \mathbb{N}$, a red-blue coloring of a finite set $A$ is said to be \emph{$h$-balanced} if the numbers of red and blue elements in $A$ differ by at most $h$. If $h=1$, then we call it \emph{perfectly balanced}. For positive integers $s,k\in \mathbb N$, we denote by $A_{s,k}=\{s,2s,\dots,ks\}$ the homogeneous arithmetic progression of length $k$ and step $s$.
In 1932 Erd\H {o}s \cite{Erdos} asked if there exists a constant $h$ and a red-blue coloring of $\mathbb{N}$ such that \emph{every} homogeneous arithmetic progression is $h$-balanced. The property in question seems unbelievable, and in fact, he expressed a guess that there is no such constant. This was confirmed only recently by Tao \cite{Tao}, with a support of collective efforts in a Polymath Project \cite{Polymath}.
We prove in this note that for every fixed $k\in \mathbb{N}$ there is a red-blue coloring of $\mathbb N$ which is perfectly balanced on all sets $A_{s,k}$.
\begin{theorem}\label{Theorem Main}
For every fixed $k\in\mathbb N$ there exists a red-blue coloring of $\mathbb N$ such that every homogeneous arithmetic progression $A_{s,k}$ is perfectly balanced.
\end{theorem}
The proof uses completely multiplicative functions with values in the set $\{+1,-1\}$ and some estimates for primes in arithmetic progressions. We give it in Section 2. In Section 3 we present another approach that did not appear to be successful, but leads to an intriguing open problem. In Section 4 we propose a new variant of the Erd\H {o}s Discrepancy Problem, motivated by the \emph{majority} coloring of graphs. This leads in turn to a question concerning completely multipliactive functions, resembling the famous conjecture of P\'{o}lya \cite{Polya}, concerning partial sums of the \emph{Liouville function} $\lambda(n)$. Finally, in Section 5, we briefly describe our initial motivation that led us to Theorem \ref{Theorem Main}, and discuss the related problem concerning \emph{rainbow} homogeneous arithmetic progressions of fixed length $k$.
\section{The proof}
Recall that an arithmetic function $f$ is \emph{completely multiplicative} if it satisfies $f(ab)=f(a)f(b)$ for every pair of positive integers $a,b\in \mathbb{N}$. Notice that this implies that $f(1)=1$. Since we will consider only functions with two values $\{+1,-1\}$ (corresponding to colors red and blue), we will call them shortly \emph{multiplicative colorings}.
We start with the following simple lemma.
\begin{lemma}\label{Lemma Multiplicative Balanced}
Let $k$ be a fixed positive integer. Suppose that $c$ is a multiplicative coloring of the set $\{1,2,\dots,k\}$ which is perfectly balanced. Then $c$ can be extended to a multiplicative coloring of $\mathbb N$ which is perfectly balanced on every set $A_{s,k}$.
\end{lemma}
\begin{proof}
Let $C=(c(1),c(2),\dots,c(k))$ be the initial color sequence satisfying the assumptions of the lemma. In particular, it satisfies
\begin{equation}
c(1)+c(2)+\cdots +c(k) \in \{+1,-1,0\}.
\end{equation}
We extend the coloring $c$ to the whole of $\mathbb N$ in a natural way. First, if $p\geqslant k+1$ is a prime number, then we may chose for $c(p)$ any value from $\{+1,-1\}$. If $n=p_1p_2\cdots p_r$ is a product of primes $p_i$, then we compute $c(n)=c(p_1)c(p_2)\cdots c(p_r)$. So, the coloring $c$ is multiplicative. Consequently, the color sequence of every set $A_{s,k}$ satisfies
\begin{equation}
(c(s),c(2s),\dots,c(ks)) =c(s)(c(1),c(2),\dots,c(k))=\pm C,
\end{equation}
and is therefore perfectly balanced.
\end{proof}
The next lemma comes from the paper by Borwein, Choi, and Coons \cite{BorweinCC}.
\begin{lemma}[Borwein, Choi, Coons, \cite{BorweinCC}]\label{Lemma Borwein}
Let $b$ be a multiplicative coloring of $\mathbb{N}$ defined by $b(3)=+1$ and $b(p)\equiv p(\modulo 3)$ for all other primes $p$. Then for every $k\geqslant 1$, the sum $\Sigma_{i=1}^{k}b(i)$ equals the number of $1$'s in base $3$ expansion of $k$. In consequence, for every $k\geqslant 1$ we have
\begin{equation}
0\leqslant\Sigma_{i=1}^{k}b(i)\leqslant \lceil \log_3 k\rceil +1.
\end{equation}
\end{lemma}
We will also need some estimates on the number of primes of the form $3m+1$ in the interval $(N,2N)$. Recall that the Chebyshev function $\theta(x;3,1)$ is defined by
\begin{equation}
\theta(x;3,1)=\sum_{p\equiv1(\modulo3),p\leqslant x}\log p.
\end{equation}
We will use the following result of McCurley \cite{McCurley}.
\begin{lemma}[McCurley \cite{McCurley}]\label{Lemma McCurley}
For $x\geqslant 17377$ we have
\begin{equation}
0.49x\leqslant \theta(x;3,1)\leqslant 0.51x.
\end{equation}
\end{lemma}
Using this lemma we get a useful lower bound for the number of primes of the form $3m+1$ between $x$ and $2x$.
\begin{lemma}\label{Lemma f(N)}
Let $f(x)$ denote the number of primes of the form $3m+1$ in the interval $(x,2x)$. Then for every $x\geqslant17377$ we have
\begin{equation}
f(x)\geqslant 0.47\frac{x}{\log(2x)}.
\end{equation}
\end{lemma}
\begin{proof}
Let $f(x)=r$, and let $q_1,q_2,\dots, q_r$ be all the primes of the form $3m+1$ in the interval $(x,2x)$. Then we may write
\begin{equation}
\theta(2x;3,1)-\theta(x;3,1)=\sum_{i=1}^{r} \log q_i \leqslant \log ((2x)^r)=r\log(2x).
\end{equation}
Using Lemma \ref{Lemma McCurley} we get
\begin{equation}
\theta(2x;3,1)-\theta(x;3,1)\geqslant 0.98x-0.51x=0.47x.
\end{equation}
Hence, we obtain $r\geqslant \frac{0.47x}{\log(2x)}$ for all $x\geqslant 17377$.
\end{proof}
Now we are ready to prove Theorem \ref{Theorem Main}
\begin{proof}[Proof of Theorem \ref{Theorem Main}]
Let $k\in \mathbb{N}$ be fixed. By Lemma \ref{Lemma Multiplicative Balanced} it is sufficient to construct a multiplicative perfectly balanced coloring of $\{1,2,\dots,k\}$. For small values of $k$, say for $k\leqslant 10^6$, this can be done by computer. So, assume that $k\geqslant 17377$, and let $b$ be a multiplicative coloring form Lemma \ref{Lemma Borwein}.
We will switch the colors of prime numbers of the form $p=3m+1$ lying in the range $(\frac{k}{2},k)$, from plus to minus, so that the resulting coloring will be balanced. This will not affect multiplicativity. Moreover, there are sufficiently many such primes since by Lemma \ref{Lemma Borwein} and Lemma \ref{Lemma f(N)}, their number satisfies
\begin{equation}
f\left(\frac{k}{2}\right)\geqslant 0.47\frac{k}{2\log k}\geqslant \lceil \log_3{k} \rceil+1 \geqslant \sum_{i=1}^{k}b(i).
\end{equation}
This completes the proof.
\end{proof}
\section{Greedy muliplicative coloring}
To prove Theorem \ref{Theorem Main} we firstly considered a different approach proposed by Rejmer (personal communication). It is a simple algorithm producing a perfectly balanced multiplicative coloring of the set $\{1,2,\dots,k\}$ in a \emph{greedy} way.
Let $k$ be a fixed positive integer. We start with putting $c(1)=+1$. In each consecutive step we color the next integer so that the new coloring is perfectly balanced and multiplicative. So, in the second step we put $c(2)=-1$. For a more precise description, suppose that after $j-1$ steps, $j\geqslant 2$, we obtained a perfectly balanced multiplicative coloring $(c(1),c(2),\dots, c(j-1))$. In the next step we distinguish two cases.
\begin{case}[$j-1=2m$]\emph{In this case we must have $\sum_{i=1}^{2m}c(i)=0$. Thus, any choice for $c(j)$ will not destroy the perfect balance. If $j$ is composite, then $c(j)$ is determined by multiplicativity. If $j$ is prime, then we may put $c(j)=-1$.}
\end{case}
\begin{case}[$j-1=2m-1$]\emph{Then $j$ is even, so $c(j)$ is determined by multiplicativity. Since $\sum_{i=1}^{2m-1}c(i)=\pm 1$, we may have either $\sum_{i=1}^{2m}c(i)=0$ or $\sum_{i=1}^{2m}c(i)=\pm 2$. In the former case we do nothing. In the later case, we find the largest prime $p>\frac{j}{2}$ such that $c(p)$ has \textquotedblleft wrong\textquotedblright sign, and switch it. This makes the new coloring $(c(1),c(2),\dots ,c(j))$ perfectly balanced.}
\end{case}
Notice that by Bertrand's Postulate, there is always a prime between $\frac{j}{2}$ and $j$. However, it is not clear that there will always be a prime whose sign-switching would improve balance. For instance, in the $16$th step of the algorithm we get the following coloring:
$$
\begin{array}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
1&2&3&4&5&6&7&8&9&10&11&12&13&14&15&16 \\
\hline
+&-&-&+&-&+&-&-&+&+&+&-&-&+&+&+\\
\hline
\end{array}
$$
with $9$ pluses and $7$ minuses. To fix this imbalance, we go back to the first prime to the left, which is $p=13$. However, $c(13)=-1$, so switching this color into $c(13)=+1$ would only increase imbalance. Fortunately, for the next prime $p=11$ we have $c(11)=+1$, so we may switch it to $c(11)=-1$ and get in this way a perfectly balanced multiplicative coloring.
We do not known if this algorithm ever stops.
\begin{conjecture}\label{Conjecture Rejmer}
Rejmer's algorithm runs ad infinitum.
\end{conjecture}
Rejmer made some computational experiments with his algorithm. In particular, he run it up to $10^9$ steps producing in this way perfectly balanced multiplicative colorings of $A_{1,k}$ for all $k\leqslant 10^9$. Notice that the first half terms of the last coloring will not be changed in the future. Thus, assuming validity of Conjecture \ref{Conjecture Rejmer}, the algorithm defines an intriguing recursive binary sequence $R(n)$ over $\{-1,+1\}$. Up to $n=40$ Rejmer's sequence coincides with the Liouville function $\lambda(n)$ (defined by $\lambda(p)=-1$ for all primes $p$), but $R(41)=+1$. The same happens for many other primes, in particular $R(97)=R(101)=+1$. One may suspect that there will be infinitely many primes $p$ with $R(p)=+1$.
\begin{conjecture}\label{Conjecture Rejmer 2}
There exist infinitely many primes $p$ for which $R(p)=+1$.
\end{conjecture}
\section{Majority version of the Erd\H {o}s Discrepancy Problem}
Let $h$ be a positive integer and let $c$ be a red-blue coloring of the set $A_{s,k}$. We say that $c$ is a \emph{majority} coloring if more than a half of the elements of $A_{s,k}-\{s\}$ have color different than the element $s$. Notice that a perfectly balanced coloring of $A_{s,k}$ satisfies the majority condition, while a majority coloring of $A_{s,k}$ can have all elements, except one, in the same color.
Is it possible that there is a red-blue coloring of $\mathbb{N}$ which satisfies the majority condition on \emph{every} homogeneous arithmetic progression $A_{s,k}$? Interpreting colors as numbers $\{+1,-1\}$, we may express the majority coloring of $A_{s,k}$ via the inequality:
\begin{equation}
c(a)\sum_{j=1}^{k}c(ja)\leqslant 0.
\end{equation}
So, the answer to the above question would be positive if we could find a completely multiplicative function $c$ satisfying the inequality
\begin{equation}
\sum_{j=1}^{k}c(j)\leqslant 0,
\end{equation}
for every $k\geqslant2$. A natural candidate for such negativity property is the Liouville function $\lambda(n)$. Actually in 1919 P\'{o}lya \cite{Polya} conjectured that $\sum_{i=1}^{n}\lambda(i)\leqslant 0$ for all $n\geqslant 2$, and proved that this would imply the Riemann Hypothesis (see \cite{BorweinCRW}). Unfortunately, this supposition occurred to be far from the truth, but the smallest counter-example is $n=906150257$.
As in the original Erd\H {o}s Discrepancy Problem, one may consider a relaxed version of majority coloring with some parameter $h$. Let us call a coloring $c$ of $A_{s,k}$ an \emph{$h$-majority} coloring if it satisfies:
\begin{equation}
c(s)\sum_{j=1}^{k}c(js)\leqslant h.
\end{equation}
This leads to the following conjecture.
\begin{conjecture}\label{Conjecture Merdos}
There exists a constant $h$ and a red-blue coloring of $\mathbb N$ which is $h$-majority on every homogeneous arithmetic progression $A_{s,k}$.
\end{conjecture}
A natural first attempt is to look for an appropriate multiplicative coloring, which leads to the following problem.
\begin{conjecture}\label{Conjecture Merdos Multiplicative}
There exists a constant $h$ and a completely multiplicative function $c$ satisfying
\begin{equation}
\sum_{j=1}^{k}c(j)\leqslant h,
\end{equation}
for all $k\geqslant1$.
\end{conjecture}
We made some computer experiments with functions that are close to the Liouville function $\lambda(n)$ in the sense that only a few small primes have sign $+1$. For instance, switching only the sign of one small prime gives usually a function with much smaller partial sums. However, a standard argument using the Riemann Zeta function shows that in order to get a function satisfying Conjecture \ref{Conjecture Merdos Multiplicative} one has to switch signs of infinitely many primes.
\section{Final remarks}
Curiously, our initial impulse for Theorem \ref{Theorem Main} came from a different direction and was related to the following problem posed independently by Pach and P\'{a}lv\"{o}lgyi (see \cite{CaicedoCP}).
\begin{conjecture}\label{Conjecture Rainbow}
For every $k\in \mathbb{N}$ there is a $k$-coloring of $\mathbb N$ such that every set $A_{s,k}$ is \emph{rainbow}.
\end{conjecture}
Notice that the above statement easily implies the assertion of Theorem \ref{Theorem Main}. Indeed, a desired red-blue coloring can be obtained by splitting the set of $k$ colors into two subsets of (almost) the same cardinality.
Another consequence of Conjecture \ref{Conjecture Rainbow} is a positive answer to the following question of Graham \cite{Graham}: Is it true that among any $n$ distinct positive integers $a_1,a_2,\dots,a_n$ there is always a pair $a_i,a_j$ satisfying $\frac{a_i}{\gcd (a_i,a_j)}\geqslant n$? The problem was solved in the affirmative for sufficiently large $n$ by Szegedy \cite{Szegedy} and independently by Zaharescu \cite{Zaharescu}. Then Balasubramanian and Soundararajan \cite{BalasubramanianS} gave a complete solution by using methods of Analytic Number Theory.
To see a connection between these two problems, consider a graph $G_k$ on positive integers in which two numbers $r,s$ are joined by an edge if and only if $\frac{r}{\gcd(r,s)}\leqslant k$ and $\frac{s}{\gcd(r,s)}\leqslant k$. Let $\omega(G_k)$ and $\chi(G_k)$ denote the clique number and the chromatic number of the graph $G_k$, respectively. Then Graham's problem is equivalent to $\omega(G_k)=k$, while Conjecture \ref{Conjecture Rainbow} is equivalent to a much stronger statement that $\chi(G_k)=k$ (see \cite{BosekDGSSZ DM}, \cite{CaicedoCP}).
Going back to the Erd\H {o}s Discrepancy Problem, it is natural to wonder to what extent the original question can have a positive answer. Let us call a pair of sets $(S,K)$ \emph{cute} if there is a constant $h$ and a red-blue coloring of $\mathbb N$ such that every set $A_{s,k}$ is $h$-balanced, for all $s\in S$ and $k\in K$. So, the result of Tao \cite{Tao} says that the pair $(\mathbb{N},\mathbb{N})$ is not cute, while Theorem \ref{Theorem Main} asserts that $(\mathbb{N},K)$ is cute for every singleton $K=\{k\}$. It is not hard to prove that there are infinite sets $K$ for which $(\mathbb{N},K)$ is still cute. For instance, one may use the multiplicative coloring $b$ from Lemma \ref{Lemma Borwein} to infer that $(\mathbb{N},K)$ is cute if $K$ is the set of positive integers avoiding $1$'s in their base $3$ expansion. Notice however, that this set $K$ has density zero. Is there a cute pair $(\mathbb{N},K)$ with $K$ of positive density?
On the other hand, there exist dense sets $S$ for which the pair $(S,\mathbb N)$ is cute. For instance, the alternating red-blue coloring of $\mathbb{N}$ is perfectly balanced on all sets $A_{s,k}$ with $s$ odd (see \cite{LeongShallit}). Is there a cute pair $(S,K)$ with both sets of density one?
| 9162f59f29ecb934eebfc67684300fe7b559645c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
\def\arabic{section}.\arabic{equation}{\arabic{section}.\arabic{equation}}
\setcounter{equation}{0}
The effective chiral Lagrangian formalism has become a well-established methodology to study the interactions
of the Goldstone bosons with or without other particle species, like e.g. pions and nucleons, respectively
\cite{ccsc}.
The most significant example is Chiral Perturbation Theory (ChPT) \cite{weinberg.200221.1,leutwyler.200220.1},
which is the low-energy effective field theory (EFT) of Quantum Chromodynamics (QCD).
For introduction and reviews, see e.g. \cite{u1,u2,u3}.
The use of perturbative calculations within ChPT as input for
non-perturbative $S$-matrix based methods is a general procedure several decades old.
Due to the fact that ChPT results are perturbative, given in terms of an expansion
organized in increasing powers of the external four-momenta and light pseudoscalar masses,
unitarity is only satisfied in the perturbative sense, similarly as in a standard Born series
(perturbative unitarity is discussed in Sec.~\ref{sec.200525.1}).
A well-known example in this regard is the calculations in Quantum Electrodynamics with Feynman diagrams,
where the expansion is done in powers of $\alpha$ (the fine structure constant),
so that if the leading-order calculation is ${\cal O}(\alpha)$ then unitarity contributions
start at ${\cal O}(\alpha^2)$
from one-loop diagrams.
However, the fulfillment of unitarity implies to square the calculated amplitudes,
and not to expand the latter only up to the order in which the scattering amplitude is calculated
(for an explicit example the reader can consult e.g. Sec.~7.3 of Ref.~\cite{peskin}.)
It is somewhat astonishing that already in 1970 one can read about motivations
for unitarizing phenomenological chiral Lagrangians, introduced
to construct realizations of the current algebra approach \cite{weinberg.200221.1}.
Rephrasing the original remarks by Schnitzer \cite{schnitzer.200221.1,schnitzer.200302.1}, the ideas he put forward are
still the main reasons to advocate the unitarization of ChPT amplitudes:
\begin{enumerate}
\item The tree approximation to the scattering amplitudes violate badly unitarity. This could also be
said for perturbative unitarity, at least in some partial waves.
\item The Lagrangians are nonlinear and nonrenormalizable, which makes difficult to compute higher-order corrections.
Nowadays, we would better say that there is a rapid proliferation of
counterterms as the order of the calculation increases in ChPT, with the state of the art at the
two-loop level in ChPT.
It is typically simpler and much more predictive to implement lower-order
calculations of ChPT within non-perturbative methods.\footnote
Several examples are given along this review
related to meson-meson scattering and spectroscopy, like e.g. the $\pi\pi$ phase shifts, scalar and vector
pion form factors, impact of the resonances $f_0(500)$ and $\rho(770)$ in the low-energy phenomenology,
$\eta\to 3\pi$ decays, etc.
}
\item Even if such corrections could be computed, the resultant renormalized perturbation series would
probably diverge, since the perturbation parameter has the strength characteristic of strong interactions.
This is clear from phenomenology because hadronic interactions are characterized by
plenty of resonances and a rapid saturation of unitarity in many partial-wave amplitudes (PWAs).
\end{enumerate}
Although the interest in the present writing is on ChPT and the associated chiral expansion, among those
early papers of Schnitzer we also quote Ref.~\cite{schnitzer.200302.1}.
This paper builds a particular realization of the current-algebra, which
satisfies the associated Ward-Takahashi identities
and two-body unitarity is implemented by means of an effective-range-type parameterization (a
unitarization method discussed in Sec.~\ref{sec.200523.1}).
One possibility to improve the agreement with data of the perturbative calculations within ChPT is
to apply the chiral series expansion to an interaction element of the amplitude, which
is afterwards implemented within non-perturbative techniques.
This is one of the basic ideas behind unitarization methods for the chiral series of
scattering amplitudes.
The first works along these lines considered the application of an
effective-range-type parameterization
to unitarize $\pi\pi$ scattering \cite{brown.200302.1,schnitzer.200302.1},
once the $\pi\pi$ scattering amplitude was calculated at leading order in ChPT
by the application of the current-algebra techniques and
the partial conservation of the axial-vector current (PCAC) \cite{weinberg.200302.1,brown.200302.1}.
A similar unitarization method was later applied to the first calculation at next-to-leading order (NLO)
in the chiral counting of the $\pi\pi$ PWAs in the chiral limit ($m_\pi=0$).
The calculation of the latter ones, as well as their
unitarization by applying a generalized effective-range expansion (ERE) \cite{chew},
were undertaken in Ref.~\cite{lehmann.200302.1},
as discussed in more detail in Sec.~\ref{sec.200511.1}.
This calculation explicitly shows that the $\pi\pi$ scattering in the chiral limit is finite.
The pioneering works by Truong and collaborators \cite{truong.200211.1,truong.200211.2,truong.200211.3}
deserve special mention,
in which the role of the isoscalar $S$-wave $\pi\pi$ final-state interactions (FSI) are stressed,
having significant effects on several physical processes.
In a first instance \cite{truong.200211.1}, the authors correct the current-algebra result for $\eta\to 3\pi$
by the $S$-wave $\pi\pi$ rescattering, a reaction which is also discussed by the application of the
Khuri-Treiman (KT) \cite{khuri}
formalism in Sec.~\ref{sec.200522.1}.
For that Ref.~\cite{truong.200211.1} multiplies the current-algebra transition
amplitude by an Omn\`es function \cite{oller.book,musk.book},
in which the isoscalar scalar $\pi\pi$ phase shifts were, however, taken from experiment.
As a result, the Watson final-state theorem is fulfilled \cite{watson.200211}.
In another work of the saga \cite{truong.200211.2}, the input phase shifts were generated consistently by the
theoretical scheme followed after taking the one-loop ChPT result for the scalar and vector pion form factors and
imposing the fulfillment of unitarity, as discussed in Sec.~\ref{sec.200515.1}.
It is also stressed that in this form a resummation of the ChPT series is achieved
that may also give rise to resonant effects.
It is also remarkable the confirmation by unitarization methods
of the existence of the $\sigma$ resonance in pion-pion interactions
at low energies. This resonance is nowadays called $f_0(500)$ in the PDG \cite{pdg} and its pole position is
given there at $(400-550)-i(200-350)$~MeV. The standard view of ChPT,
based on the spontaneous symmetry breaking of $SU(2)\times SU(2)$ chiral symmetry \cite{leutwyler.200220.1},
considered as highly unlikely that such a low-mass resonance could happen in $\pi\pi$ scattering,
where the small expansion parameter is claimed to be $m_\pi^2/m_\rho^2\simeq 0.03$, with $m_\rho$ the mass of the
$\rho(770)$ meson. However, for the isoscalar scalar $\pi\pi$ scattering the unitarity corrections are affected by a
large numerical factor that could actually make the expansion parameter in the momentum-squared dependence of
these PWAs to be much larger.
This was explicitly shown in Ref.~\cite{nd} by performing the exercise of determining the value of the
renormalization scale $\mu$ needed in order to generate the $\rho(770)$ pole by unitarizing the leading-order (LO)
ChPT amplitude. It was obtained that a huge unnatural value for the realm of QCD was needed, with $\mu \simeq 1$~TeV,
while the same value for generating the $\sigma$ resonance had the natural value in QCD of $\mu\simeq 1$~GeV.
It is instructive to also show the main equations for the completion of this exercise. For the isovector vector
$\pi\pi$ interactions, where the $\rho(770)$ resonance appears, one has the unitarized expression of the
LO ChPT PWA $T_{11}(s)$, which reads \cite{nd} (as also discussed in Sec.~\ref{sec.200517.1})
\begin{align}
\label{200508.1}
T_{11}(s)&=\left[\frac{6f_\pi^2}{s-4m_\pi^2}+g(s)\right]^{-1}~.
\end{align}
Here, $T_{IJ}$ is the PWA of the two-pion system with isospin $I$, $J$ is the angular momentum and
the LO ChPT amplitude is $(s-4m_\pi^2)/6f_\pi^2$, with $f_\pi=92.4$~MeV the pion weak decay constant.
The function $g(s)$ corresponds to the two-pion unitarity loop function, given by
\begin{align}
\label{200508.2}
g(s)&=\frac{1}{(4\pi)^2}\left[\log\frac{m_\pi^2}{\mu^2}+\sigma(s)\log\frac{\sigma(s)+1}{\sigma(s)-1}\right]~,\\
\sigma(s)&=\sqrt{1-4m_\pi^2/s}~.\nonumber
\end{align}
In turn, the unitarized expression for the $I=J=0$ PWA, which contains the $f_0(500)$ pole, is \cite{npa,nd}
\begin{align}
\label{200509.3}
T_{00}(s)&=\left[\frac{f_\pi^2}{s-m_\pi^2/2}+g(s)\right]^{-1}~,
\end{align}
where the LO ChPT PWA is $(s-m_\pi^2/2)/f_\pi^2$.
The main difference between Eqs.~\eqref{200508.1} and \eqref{200509.3} is the factor 6 dividing the LO ChPT $T_{11}(s)$ compared to
$T_{00}(s)$, because $s\gg m_\pi^2$ in the region where the $\sigma$ or $\rho$ poles lie.
Indeed, in order to get a resonance of mass $m_\rho$ in the $I=J=0$ PWA one needs a $\mu$ of around 1.8~GeV,
in comparison with $\mu$ around 1~TeV that is needed in the $I=J=1$ case.
The reason for this dramatic change in the needed values of $\mu$ is because $g(s)$ only depends logarithmically on
this parameter.
This fact reflects that the unitarity corrections for the scalar isoscalar sector are numerically enhanced.
This
enhancement is enough to generate resonant effects that strongly impact the phenomenology
and make fallacious to think in the possibility to reach accuracy by a straightforward application
of ChPT for many reactions. As a result, the infinite set of unitarity bubble diagrams should be resummed
in order to account for this numerical enhancement.
This phenomenon is also seen in the strong corrections affecting the $I=J=0$ $\pi\pi$
scattering length originally calculated by Weinberg at LO \cite{weinberg.200302.1} with current algebra methods.
The expressions for the $a_{IJ}$ scattering lengths up to NLO or ${\cal O}(p^4)$ in ChPT
from Ref.~\cite{leutwyler.200220.1} contain chiral loops which are the dominant NLO contributions in the limit
$M_\pi\to 0$. They read:
\begin{align}
\label{200509.1}
a_{00}&=\frac{7m_\pi^2}{32\pi f_\pi^2}\left\{1-\frac{9m_\pi^2}{32\pi^2 f_\pi^2}\log\frac{m_\pi^2}{\mu^2}+\ldots\right\}~,\\
a_{20}&=-\frac{m_\pi^2}{16\pi f_\pi^2}\left\{1-\frac{5m_\pi^2}{32\pi^2 f_\pi^2}\log\frac{m_\pi^2}{\mu^2}\ldots\right\}~,\nonumber\\
a_{11}&=\frac{1}{24\pi f_\pi^2}\left\{1-\frac{m_\pi^2}{8\pi^2f_\pi^2}\log\frac{m_\pi^2}{\mu^2}+\ldots\right\}~.\nonumber
\end{align}
It follows then that $a_{00}$ has the largest NLO contribution in the limit $m_\pi\to 0$. In order to
appreciate better the relatively large size of this correction, it is worth comparing it with
the pion mass calculated in ChPT up to NLO \cite{leutwyler.200220.1},
\begin{align}
\label{200509.2}
m_\pi^2&=\bar{m}_\pi^2\left\{1+\frac{m_\pi^2}{32\pi^2 f_\pi^2}\log\frac{m_\pi^2}{\mu^2}+\ldots\right\}~,
\end{align}
with $\bar{m}_\pi^2$ the bare mass squared.
The NLO term here is a factor 9 smaller in absolute value than that for $a_{00}$.
Indeed, this was one of the reasons for developing a non-perturbative dispersive approach
that could provide an improvement in the prediction of the $\pi\pi$ scattering lengths.
The idea is to make use of the Roy equations \cite{roy.200509.1} and to match with ChPT
in the subthreshold region, where the ChPT expansion is better behaved,
since it is away from the threshold cusps \cite{anant,leutwyler.200509.1}. In this way, the two subtraction constants
needed for solving the Roy equations can be predicted by ChPT, applied at different orders. The resulting
convergence properties of the prediction for the $\pi\pi$ scattering lengths is much improved and a
reliable estimate of the uncertainties can be also provided.
Another more recent advance was the development of a new set of Roy-like equations
in Refs.~\cite{217.pela,111.pela}, the so-called GKPY equations.
The difference is that these equations have only one subtraction instead of two and, e.g.,
they have given rise to an accurate determination of the $f_0(500)$ pole
from experimental data, without relaying on the ChPT expansion.
These unitarity techniques have also other interesting fields of application beyond meson physics.
Indeed,
the 90's of the past century experienced a boost in the interest of applying chiral
EFTs for the study of nuclear interactions.
To large extent this was triggered by the seminal articles of Weinberg \cite{weinberg.200301.1},
in which the systematic application of ChPT order by order to calculate the nuclear potentials $V$ is established.
As the chiral order increases, however, extra derivatives with respect to $r$ act on the potential, so that
it becomes more singular for $r\to 0$.
Because of this complication the application of ChPT for the calculation of
the low-energy $NN$ PWAs by implementing the chiral potentials in quantum-scattering integral equations
is not yet fully satisfactory.
In atomic and molecular physics the scattering by a singular potential is of great importance too, a well-known
example being the Van der Waals force among atoms or molecules.
This is a problem in which recent advances giving rise to the exact $N/D$ method in non-relativistic scattering
\cite{oller.lhc.aop,oller.lhc.plb}
are showing themselves as very powerful and promising.
This new method is briefly reviewed in Sec.~\ref{sec.200517.4}.
The application of ChPT with barons to $NN$ scattering also triggered the use of this EFT to
the study of the non-perturbative $\bar{K}N$ scattering in coupled channels, particularly in connection with the
$\Lambda(1405)$ \cite{kaiserweise,or,plb,ooj}.
The non-perturbative character of the $ NN $ interactions, which requires the full iteration of the potential,
is due to two basic aspects.
(i) One of them is a quantum effect of kinematical origin within the typical scales of the problem.
The typical distance of propagation of two nucleons as virtual particles is
$l_{NN} \sim 1/ E_{\text{kinetic}} \sim m/p^2 \sim (m/m_\pi) b_\pi $, where $m$ and $m_\pi$ are the
nucleon and pion masses, respectively.
The range of the $ NN $ interactions is given by the Compton wavelength of the pion
$b_\pi=m_\pi^{-1}$ (in our units $\hbar=c=1$).
As $ m/m_\pi\gg 1 $ this travel distance for virtual particles
is large enough for having several repetitive collisions
between the propagating two nucleons.
The same conclusion is reached if one focuses on the propagation of real nucleons.
For a typical three-momentum $m_\pi$ they have a velocity of order $m_\pi/m$.
Thus, the time for crossing a distance $b_\pi$ is $\sim m/ m_\pi^2\gg b_\pi$.
(ii) Nonetheless, if the coupling between two nucleons
were small enough the scattering would be perturbative despite (i).
This does not happen since the coupling due to one-pion exchange between two nucleons
is of the order $g_A^2 m_\pi^3/f_\pi^2$, where
$g_A\simeq 1.26$ is the axial coupling of the nucleon.
This factor times $l_{NN}$ implies the dimensionless number
\begin{align}
\label{180507.2}
\frac{l_{NN}}{16\pi}\frac{g_A^2 m_\pi^3}{f_\pi^2}=\frac{m g_A^2}{16\pi f_\pi^2} m_\pi~,
\end{align}
which is about 0.5. Therefore, the $NN$ interactions should be treated non-perturbatively as a general rule.
In this equation the phase-space factor $1/16\pi$ is included,
which accounts for the two-nucleon propagation in all directions.
We also distinguish in Eq.~\eqref{180507.2} the scale \cite{oller.180722.1,birse.180506.1}
\begin{align}
\label{180613.1}
\Lambda_{NN}= \frac{16\pi f_\pi^2}{m g_A^2}\simeq 2 m_\pi~,
\end{align}
which has a striking small size despite it is not proportional to $m_\pi$.
This is another consequence of the non-perturbative character of the $NN$ interactions.
The unnaturally large size of the $NN$ $S$-wave scattering lengths ($a_s$),
so that they are much bigger in absolute value than the Compton wavelength of the pion, $|a_s|\gg b_\pi $,
introduces a new scale at low energies.
For instance, the scattering length for the isovector $^1S_0$ $NN$ PWA is $a_s \simeq -25$~fm.
As a result, the dimensionless number in Eq.~\eqref{180507.2} becomes even larger
by a factor $|a_s|m_\pi \gg 1 $.
Therefore, when the center of mass (CM) three-momentum is smaller than $m_\pi$,
in which case the ERE applies, the $NN$ interactions are manifestly non-perturbative
and the $NN$ potential has to be iterated. Precisely, in this energy region one finds the
bound state of the Deuteron in the coupled ${^3S}_1-{^3D}_1$ PWAs and an antibound state for the $^1S_0$.
One close field is infinite nuclear matter,
where resummation techniques based on the $N/D$ method, discussed in Sec.~\ref{sec.200517.1},
were applied in Refs.~\cite{lacour,llanes} to work out the $NN$ scattering amplitude
in the nuclear medium. From this result,
equations of state for neutron and symmetric nuclear matter were derived each containing only a
free parameter, and showing themselves as very successful from the phenomenological point of view.
See Ref.~\cite{oller.nm.review} for a recent review on these and other connected works.
Related resummations were achieved in Refs.~\cite{kaiser,boulet} to address the unitary limit in normal
nuclear matter for a Fermi Gas.
This issue concerns both nuclear physics, condensed matter and atomic, molecular, and optical physics.
At higher energies, one also finds examples of the application of unitarization techniques, some of them, like the
$N/D$ method or the Inverse-Amplitude Method (IAM), discussed here.
Regarding this point, there have been recently a series of works applying these two methods to study the
scattering and spectrum of the longitudinal components of the electroweak gauge vector bosons $W$ and $Z$
by taking advantage of the equivalence theorem, which is applicable to energies much larger than the masses
of the $W$ and $Z$ bosons \cite{f1,f2,f3,f4}. These studies are very timely due to the experimental program at the
LHC, which reinforces their interest.
Quantum gravity is another field in which unitarization techniques have been applied in the last years to
study the $2\to 2$ scattering, due to one-graviton exchange in the $s$ channel, of $N_S$ scalars, $N_V$ vector and
$N_f$ fermions.
Notice that this set of fields comprises all the particles in the standard model as a particular case.
The two particles making up an initial or final two-body state are selected so as to avoid the graviton $t$- and/or $u$-channel
exchanges. The reason is because these exchanges drive to infrared divergences (gravity is a force of infinite range)
that invalidate a standard
partial-wave amplitude expansion. The quantum corrections are implementing within the low-energy EFT of
Quantum Gravity \cite{weinphys,burgess,donoghue.200205.1}.
The interested reader can consult Refs.~\cite{han.200204.1,donoghue.200204.1,calmet.200204.1}.
These works employ the one-loop vacuum polarization due to matter fields (gravitons are excluded),
and resum its iteration plus the tree-level contribution.
Of course, a similar situation also
arises in the electromagnetic case by the exchange of a photon in the $t$- and $u$-channels.
A prominent example of it being the Coulomb scattering.
An interesting future prospect is to develop unitarization methods appropriate for infinite-range interactions.
It could then handle crossed one-graviton exchanges and allow to study those $2\to 2$ scattering processes
disregarded in Refs.~\cite{han.200204.1,donoghue.200204.1,calmet.200204.1}.
In this work we review a set of unitarization methods and we always
follow the order of first discussing scattering, mostly in PWAs, and then FSI.
We also develop links between the different methods discussed.
The unitarization techniques selected are popular ones within the hadron physics community.
One of the reasons for their popularity is because they have proven to be very powerful
in phenomenological applications, so that they are certainly of interest.
It was not the aim of this work to be exhaustive and give a comprehensive review
discussing every unitarization method used in the literature.
Historical reasons are behind the inclusion of the (generalized) ERE unitarization, widely used in the
earlier papers of the 60's and 70's, since later on this method was replaced by the IAM,
$K$-matrix parameterizations, $N/D$ method, etc, in relativistic hadron-hadron scattering (not so for non-relativistic applications).
The contents of this work are organized as follows.
After a brief review on the $S$-matrix and unitarity in Sec.~\ref{sec.200525.1} we then move to discuss several
unitarization methods in Sec.~\ref{sec.200511.1}.
The generalized ERE, the $K$-matrix approach, the IAM
and the Pad\'e resummation are then considered. The Sec.~\ref{sec.200523.4} is dedicated to
the implementation of re-scattering effects in probes and several methods are presented, with some of them clearly
related to the already presented ones in Sec.~\ref{sec.200511.1}
dedicated to scattering. Subsequently, other methods are introduced that could be applied to any given set of PWAs.
We discuss in Sec.~\ref{sec.200517.1} the $N/D$ method for PWAs and FSI. This section ends with a brief
account of the exact $N/D$ method recently developed for non-relativistic scattering.
The last section contains our conclusions with extra discussions included.
\section{Unitarity}
\setcounter{equation}{0}
\label{sec.200525.1}
The $S$-matrix operator $S$ gathers the transition probability amplitudes between in and out states
in a scattering process.
Let us denote by $|\alpha\rangle_{\rm in}$ and
$|\beta\rangle_{\rm out}$ an `in' and an `out' state in the Heisenberg picture,
respectively.
Then, the matrix elements of the $S$ matrix, $S_{\beta\alpha}$, correspond to the scalar products
\begin{align}
\label{200608.1}
S_{\beta\alpha}&={_{\rm out}\langle}\beta|\alpha{\rangle_{\rm in}}~.
\end{align}
The $S$ matrix plays a central role in Quantum Field Theory (QFT) \cite{weinberg.vol1}.
One is typically concerned with the matrix elements of the $S$-matrix so as
to extract scattering observables out of a QFT.
A crucial property in this regard is that the (on-shell) matrix elements of the $S$-matrix are
invariant under reparameterization of the quantum fields in QFT \cite{haag.58,ccsc}.
The analytical continuation of the $S$ matrix in the complex-energy plane allows to determine the
spectrum of the theory. Its continuum part corresponds to branch cuts and the
bound states, virtual states and resonances are poles of the $S$ matrix.
Furthermore it is very suitable to implement a relativistic formalism since
the $S$-matrix elements are covariant under the Poincar\'e group.
In the Dirac or interacting picture of QFT the $S$ matrix is given by
\begin{align}
\label{200608.3}
S_{\beta\alpha}&
=\frac{\langle \beta|e^{i\int d^4 x {\cal L}_{\rm int}} |\alpha\rangle}{\langle 0|e^{i\int d^4 x {\cal L}_{\rm int}}|0\rangle} ~,
\end{align}
where ${\cal L}_{\rm int}$ is the interacting Lagrangian,
$|\alpha\rangle$ and $|\beta\rangle$ are free particles states and
$ |0\rangle$ is the $0_{\rm th}$-order perturbative vacuum.
In Eq.~\eqref{200608.3} $U(+\infty,-\infty)=\exp i\int d^4 x {\cal L}_{\rm int}(x)$ is the evolution operator in the
Dirac picture from/to asymptotic times.
The denominator is a normalization factor that cancels the disconnected contributions
without involving any external particle in the matrix elements of $U(+\infty,-\infty)$.
A crucial point is that the $S$ matrix is unitary because of the completeness relation of either the
`in' or `out' states. However, it is important to emphasize that in the case of the $S$ matrix its unitarity
refers to the subset of states that are open for a given energy.
This is different to the typical sum over intermediate states covering a resolution of the
identity for the whole Fock space.
E.g. within ordinary Quantum Mechanics (conserving the number of particles)
one can insert a resolution of the identity by plane waves
within the product $A B$ of two one-particle operators as
\begin{align}
\langle \beta|AB|\alpha\rangle &=\int\frac{d^3p}{(2\pi)^3}\langle \beta| A|{\mathbf{p}}\rangle\langle {\mathbf{p}}|B|\alpha\rangle~.
\end{align}
Here ${\mathbf{p}}$ takes any value, so that its kinetic energy is arbitrary large and not constrained by the
available energy fixed by the external states $|\alpha\rangle$ and $|\beta\rangle$.
After this qualification, we can show that $S^\dagger S$ is unitary by employing the completeness relation
associated with the `out' states, so that
\begin{align}
\label{200608.2}
\int d\beta S_{\beta\gamma}^* S_{\beta\alpha}=
\int d\beta \, {{_{\rm out}\langle}\beta|\gamma{\rangle_{\rm in}}\!\!\!\!\!^*} \,\,\, {_{\rm out}\langle}\beta|\alpha{\rangle_{\rm in}}
={_{\rm in}\langle}\gamma|\alpha{\rangle_{\rm in}}=\delta_{\gamma\alpha}~.
\end{align}
Analogously, we can also derive $S S^\dagger=I$ by attending to the completeness relation of the
`in' states. Therefore,
\begin{align}
\label{200220.1}
SS^\dagger=S^\dagger S=I~.
\end{align}
The scattering operator $T$, also called the $T$ matrix, is introduced such that in terms of it the $S$ matrix reads
\begin{align}
\label{200220.2}
S&=I+iT~.
\end{align}
The unitarity of the $S$ matrix implies in turn that $T$ fulfills that
\begin{align}
\label{200304.1}
T-T^\dagger&=iT^\dagger T=iTT^\dagger~,
\end{align}
which is the unitarity relation for the $T$ matrix.
The last expression on the right-hand side (rhs) of the previous equation allows to derive the Boltzmann $H$-theorem
in Statistical Mechanics, which is one of the most fundamental theorem in physics.
It drives to the increase of entropy with time until the equilibrium is reached.
It is also well-known that unitarity implies the optical theorem and the existence of the
diffraction peak at high energies. For derivations of these points the reader can consult Sec.~3.6 of
Ref.~\cite{weinberg.vol1}.
The unitarity relation satisfied by the $T$ matrix is central in the $S$-matrix theory
in which the scattering amplitudes are analytically continued in their kinematical arguments \cite{martin.290916.1}.
In the development of this program one also employs the property of hermitian analyticity,
so that the matrix elements of $T^\dagger$ can be also expressed in terms of those of $T$ by
an analytical continuation in the complex $s$ plane of the (sub)process in question.
This fact allows an extension of the standard unitarity relation of Eq.~\eqref{200304.1}, such that
its left-hand side (lhs) provides the discontinuity of the analytical scattering amplitudes across
the normal cuts due to intermediate states. This discontinuity implies the existence of the so-called
right-hand cut (RHC), or unitarity cut, in the scattering amplitudes.
Due to the hermitian analyticity the unitarity relation could also involve on-shell intermediate states,
because the total energy is above their thresholds, but with some other kinematical variables
taking non-physical values (e.g. the Mandelstam variable $t$ could be away from the physical process).
For more details the reader can consult Sec.~4.6 of Ref.~\cite{olive.book}.
An explicit example is developed in Sec.~\ref{sec.200522.1}, where an analytical extrapolation
in the mass of the $\eta$ squared is used for the KT formalism.
Multiplying
both sides of Eq.~\eqref{200304.1}
to the left by $T^{-1}$ and to the right by ${T^\dagger}^{-1}$ we have
the interesting equation
\begin{align}
\label{200304.2b}
{T^\dagger}^{-1}-T^{-1}&=i I~.
\end{align}
The unitarity constraints are more easily expressed in terms of partial-wave amplitudes (PWAs),
in which the matrix elements of the $T$-matrix are taking between asymptotic states having well-defined
angular momentum. For instance, for two particles without spin, like in $\pi\pi$ scattering, the PWAs $T_\ell(s)$ are
giving by the standard expression
\begin{align}
T_\ell(s)&=\frac{1}{2}\int_{-1}^{+1}d\!\cos\theta\, P_\ell(\cos\theta)T({\mathbf{p}}',{\mathbf{p}})~,
\end{align}
where $\ell$ is the angular momentum,
${\mathbf{p}}'$ and ${\mathbf{p}}$ are the final and initial three-momenta, $\theta$ is their relative angle
and $P_\ell(\cos\theta)$ is a Legendre polynomial.
The general formulas relating the PWAs and the scattering amplitudes can be
found in Refs.~\cite{oller.book,oller.review}, to which we refer for further details.
Ref.~\cite{martin.290916.1} offers a rather thorough treatment on PWAs within the helicity formalism.
Because of time-reversal symmetry the $T$ matrix is symmetric in partial waves.
If we write this matrix in brief as $T_L$, and denote its matrix elements by $T_{L,ij}$, by its symmetric
character we mean that $T_{L,ij}(s)=T_{L,ji}(s)$.
Eq.~\eqref{200304.2b} then implies that the imaginary part of the inverse of the PWA matrix is fixed by unitarity.
In the region of energy in which the resolution of the identity is saturated by two-body intermediate states,
the rhs of Eq.~\eqref{200304.2b} can be written as
\begin{align}
\label{200304.3b}
\Im T_L^{-1}&=-\frac{q}{8\pi\sqrt{s}}\theta(s)~.
\end{align}
In this equation, $q$ is the diagonal matrix of the CM three-momentum for every two-body intermediate state and
$\theta(s)$ is also another diagonal matrix whose matrix elements are 1 for $\sqrt{s}$ larger than the threshold and 0 otherwise.
Eq.~\eqref{200304.3b} is equivalent to the probably more familiar unitarity equation for PWAs
\begin{align}
\label{200304.3c}
\Im T_L&=T_L^* \frac{q}{8\pi\sqrt{s}}\theta(s) T_L~.
\end{align}
The phase-space diagonal matrix $q(s) \theta(s)/(8\pi\sqrt{s})$ is sometimes denoted for short by $\rho(s)$.
The previous relation is not linear because its rhs is quadratic.
This fact drives to the concept of perturbative unitarity, which applies when perturbation theory is employed to calculate
the PWAs up to some order in a dimensionless parameter, let us call it $\varepsilon$.
Therefore, if the PWA is calculated up to ${\cal O}(\varepsilon^n)$,
Eq.~\eqref{200304.3c} indeed implies that
\begin{align}
\label{200611.1}
\Im T_L^{(n)}&\neq T_L^{(n)*} \frac{q}{8\pi\sqrt{s}}\theta(s) T_L^{(n)}~,
\end{align}
because the rhs contains contributions of ${\cal O}(\varepsilon^{2n})$, while the lhs only does so
up to ${\cal O}(\varepsilon^n)$. The consistent procedure is to expand the rhs in powers of
$\varepsilon$ and keep only terms up to ${\cal O}(\varepsilon^n)$. For instance, up to second
order in $\varepsilon$ one has that
\begin{align}
\label{200611.2}
\Im T_L^{(2)}&=T_L^{(1)*} \frac{q}{8\pi\sqrt{s}}\theta(s) T_L^{(1)}~.
\end{align}
The discontinuity across intermediate states in the crossed channels gives rise to the crossed-channel cuts in the PWAs after
the angular projection required to calculate them. We denote this kind of cuts generically as left-hand cuts (LHCs).
The interested reader could consult the Sec.~2 of Ref.~\cite{oller.review} for a handy pedagogical introduction
to the notions of RHC, LHC and crossing.
Now, if we consider simultaneously stronger and weaker interactions,\footnote{Where the weaker interactions,
which are supposed to be proportional to some small dimensionless parameter, could
correspond e.g. to actually electromagnetic or weak probes, while the stronger ones typically refer to the
strong interactions among hadrons.} the unitarity relation, at leading order in the weaker interaction, reads
\begin{align}
\label{200220.3}
F-F^\dagger=iT^\dagger F~.
\end{align}
In this equation $F$ represents the matrix elements of the $T$ matrix involving the weaker
interactions, so that they vanish if these interactions are neglected altogether, while still the stronger ones
would be acting.
In the latter equation we have taken that the weaker interactions act in the initial state, otherwise
write $iF^\dagger T$ on the rhs of Eq.~\eqref{200220.3}.
In PWAs the unitarity relation of Eq.~\eqref{200220.3} gets its simplest form.
In the physical region for the reactions to occur it reads
\begin{align}
\label{200220.4}
\Im F_i(s)&=\sum_jF_j(s)\rho_j(s)T_{L,ij}(s)^*=\sum_jF_j(s)^*\rho_j(s)T_{L,ij}(s)~,
\end{align}
where $\rho_j(s)$ corresponds to the phase space of the
intermediate hadronic states (integrations could also be involved for multiparticle states) and $s$ is the standard Mandelstam variable
corresponding to the total CM energy squared. The opening of the threshold for the channel $j$,
$s_{{\rm th},j}$, is accounted for by a Heaviside function $ \theta(s-s_{{\rm th},j})$ included as part of $\rho_j(s)$.
For the one-channel case the sum on the rhs of the Eq.~\eqref{200220.4} collapses to just one term,
\begin{align}
\label{200220.5}
\Im F(s)&=F(s)\rho(s)T_\ell(s)^*=F(s)^*\rho T_\ell(s)~,
\end{align}
where $T_\ell(s)$ is the corresponding uncoupled PWA.
Since the lhs of the equation is real then
it follows that the phase of the form factor $F(s)$ and the phase shift of $T_\ell(s)$ are the same modulo $\pi$.
This is the well-known Watson final-state theorem.
\section{ERE, $K$-matrix, IAM and Pad\'e approximants}
\label{sec.200511.1}
\setcounter{equation}{0}
Along this section we follow a multifaceted discussion relating different unitarization approaches, like the
(generalized) ERE, $K$-matrix parameterizations, the IAM and
the Pad\'e approximants.
\subsection{ERE and $K$-matrix approaches}
\label{sec.200523.1}
In the early days of PCAC, soft pions theorems and realizations based on chiral Lagrangians, it was
customary to refer as (generalized) ERE to a unitarization method based on the
identification of a remnant in the inverse of a PWA free of RHC which was expanded in powers of $p^2$.
The standard ERE was originally derived in Ref.~\cite{bethe.ere} for $NN$ interactions which, for an uncoupled PWA, has the form
\begin{align}
\label{200512.1}
T_\ell&=\frac{p^{2\ell}}{p^{2\ell+1} \cot\delta_\ell-ip^{2\ell+1} }~.
\end{align}
The remaining part is identified with $p^{2\ell+1}\cot\delta_\ell$ as it is well known, because of the
relation between the $T$ and $S$ matrices in the normalization used typically for the ERE, which is the
one in Eq.~\eqref{200512.1}. Namely, the steps are
\begin{align}
\label{200512.2}
S&=e^{2i\delta_\ell}=1+i2 p T_\ell \rightarrow T_\ell=\frac{e^{2i\delta_\ell}-1}{2ip}~,\\
T_\ell^{-1}&
ip\frac{e^{2i\delta_\ell}+1}{e^{2i\delta_\ell}-1}-ip=p\cot\delta_\ell-ip~.\nonumber
\end{align}
The $NN$ scattering is non-relativistic (NR), with $m^2\gg p^2$ at low energies, so that the expansion of
$p^{2\ell+1}\cot \delta_{\ell}$ is a Taylor series in $p^2$.
However, for pion-pion interactions, where $p^2\sim m_\pi^2$ in the
region of interest both theoretical and experimentally speaking, the series expansion in $p^2$ is a
Laurent series for the $S$ waves. The reason is because the Adler zeroes required by chiral symmetry
in the $S$-wave PWAs \cite{adler.181115.1}, despite there is no centrifugal barrier.
The latter is present for the higher partial waves, $\ell\geq 1$,
which implies the standard zero at threshold so that $T_\ell$ vanishes as $p^{2\ell}$ for $p\to 0$.
The phase space factor for relativistic systems changes in comparison with the NR expression of
Eq.~\eqref{200512.1}. The steps are the same as in Eqs.~\eqref{200512.1} and \eqref{200512.2}, but now
instead of $T_\ell$ one should use $T_\ell/\sqrt{s}$ so that $S_\ell=1+2ipT_\ell/\sqrt{s}$. Then,
\begin{align}
\label{200512.2b}
T_\ell&=\left[\frac{p}{\sqrt{s}}\cot\delta_\ell-i\frac{p}{\sqrt{s}}\right]^{-1}~.
\end{align}
In more recent times, the remaining part of $T_\ell^{-1}$ after discounting the factor $-ip/\sqrt{s}$, required by unitarity,
cf. Eq.~\eqref{200304.3b},
is called the inverse of the $K$-matrix, $K_\ell$, instead of $p\cot\delta_\ell$.
In this notation, $T_\ell$ is written as
\begin{align}
\label{200513.1}
T_\ell&=\left[K_\ell^{-1}-i\frac{p}{\sqrt{s}}\right]^{-1}~.
\end{align}
Of course, Eqs.~\eqref{200512.2}, \eqref{200512.2b} and \eqref{200513.1} can be generalized straightforwardly to a matrix notation for coupled-channel scattering,
with $T_\ell$ and $K_\ell$ replaced by the matrices $T_L$ and $K_L$, respectively. The inverse of the later is usually referred
as the $M_L$ matrix, $M_L=K_L^{-1}$ \cite{au}.
We are surprised that in these first works, e.g.
\cite{schnitzer.200221.1,schnitzer.200302.1,brown.200302.1,sakurai,chew,lehmann.200302.1,brehm},
it was common to refer to the (generalized) ERE without any mention at all to the $K$-matrix approach, a notion much more common
in later times and, in particular, for more recent papers based on the unitarization of ChPT.
Probably this is related to the fact that the $K$-matrix parameterizations
have been used in many instances in the literature over large energy intervals in order to fit experimental data.
As a result, it does not really make sense to keep
any memory of a particular threshold, as it is the case for the ERE.
Indeed, in those earlier papers referred the basic object of study was $\pi\pi$ scattering or the $\pi$ vector form factor.
Another fact worth stressing is that in those earlier references the expressions
finally used for $T_L^{-1}$ had better analytical properties than the ones typically
found later in papers using the $K$-matrix approach, as in Refs.~\cite{au,sarasa,moir} among many others
phenomenological studies.
The reason is because the later ones keep only the term $-ip/\sqrt{s}$ in $T_L^{-1}$ while, in the
first papers referred \cite{schnitzer.200221.1,schnitzer.200302.1,brown.200302.1,sakurai,chew,lehmann.200302.1,brehm},
the non-trivial analytical function $h(s)$, which is $8\pi g(s)$ modulo a constant, cf. Eq.~\eqref{200508.2},
was used by performing a dispersion relation (DR) along the RHC.
Namely,
\begin{align}
\label{200514.1}
h(s)&=8\pi g(s)-\frac{1}{\pi}\log\frac{m_\pi}{\mu}=\sigma(s)\log\frac{\sigma(s)+1}{\sigma(s)-1}~.
\end{align}
The function $g(s)$ is an analytical function of $s$ in the cut complex $s$ plane, having the RHC along the real $s$
axis for $s>4m_\pi^2$.
As a trivial byproduct, the zero at $s=0$ that occurs in the phase space factor $-ip/\sqrt{s}$
in the simplest $K$-matrix parameterizations is absent when using the function $g(s)$, which is the correct analytical
extrapolation of the two-body unitarity requirement above threshold.
Indeed, the removal of this spurious singularity at $s=0$ was the argument used in Ref.~\cite{brown.200302.1}
to construct the function $h(s)$ without using any DR. This reference also notices the presence of the Adler zeros
in the $I=0,$~2 S-wave $\pi\pi$ PWAs and similar expressions to Eq.~\eqref{200509.3} are proposed for these PWAs.
The main difference, an important one indeed, between Eq.~\eqref{200509.3} and Ref.~\cite{brown.200302.1}
is that the function $g(s)$, contrary to $h(s)$, contains a subtraction constant
\begin{align}
\label{200513.2}
\frac{1}{16\pi^2}\log\frac{m_\pi^2}{\mu^2}~,
\end{align}
which is absent in the function $h(s)$ of Brown and Gobble \cite{brown.200302.1}.
This is a crucial fact for the right reproduction of important features in low-energy $\pi\pi$ scattering,
like the generation of the $f_0(500)$ resonance pole in good agreement with the latest
and more sophisticated determinations \cite{pdg}.
As a matter of fact, the predicted $I=J=0$ $\pi\pi$ phase shifts in Ref.~\cite{brown.200302.1} are around a factor 2
smaller than data for the energies in between $500-700$~MeV, while the $I=2$ S-wave $\pi\pi$
phase shifts are too large in modulus by the same factor. These deficiencies in the approach of Ref.~\cite{brown.200302.1}
are cured once the subtraction constant
of Eq.~\eqref{200513.2}, with a natural value for $\mu\simeq 1$~GeV, is taken into account \cite{npa}.
For the $I=J=1$ $\pi\pi$ PWA Ref.~\cite{brown.200302.1} performs a generalized ERE up to and including the effective range,
\begin{align}
\label{200513.3}
T_{11}^{-1}-h(s)&=\frac{1}{a_1p^2}+\frac{r_1}{2}~.
\end{align}
The parameter $a_1$ is fixed from the current algebra prediction \cite{weinberg.200302.1},
$a_1=1/12\pi f_\pi^2$, while $r_1$ is determined by the vanishing of the real part of $T_{11}(s)^{-1}$ at $s=m_\rho^2$.
The resulting equation is therefore,
\begin{align}
\frac{1/a_1}{m_\rho^2/4-m_\pi^2}+\frac{r_1}{2}+\Re h(m_\rho^2)=0~.
\end{align}
Let us notice that $r_1/2$ in Eq.~\eqref{200513.3} can be also considered as a subtraction constant
of $g(s)$. Attending to Eq.~\eqref{200514.1} the relation is
\begin{align}
\label{140514.2}
\log\frac{m_\pi^2}{\mu^2}=\pi r\simeq -\frac{96\pi^2f_\pi^2}{m_\rho^2-4m_\pi^2}+\delta r~,
\end{align}
with $\delta r$ a correction of around a 20\% of the term explicitly shown.
This simple calculation illustrates the discussion at the Introduction regarding the
huge unnatural value $\mu\simeq 1.7~$TeV that results by the matching in Eq.~\eqref{140514.2},
while the expected value is around 1~GeV.
As a result of this analysis the authors of Ref.\cite{brown.200302.1}
predicted the width of the $\rho(770)$ to be 130~MeV and the
$I=J=1$ phase shifts up to 1000~MeV, in good shape compared with later experimental determinations.
They also gave an expression for the coupling of the $\rho\to\pi\pi$ ($g_{\rho\pi\pi}$)
in terms of $f_\pi$ and $m_\rho$,
which drives to the KSFR relation \cite{ksfr}, $f_\rho^2=m_\rho^2/2f_\pi^2$,
if one assumes vector-meson dominance (VMD) \cite{vmd1,vmd2}.
Here $f_\rho$ is the coupling of the $\rho$-photon transition which is equal to $g_{\rho\pi\pi}$ within VMD \cite{vmd2}.
The authors summarize their research by stating that
the fulfillment of the low-energy current-algebra constraints together with
the inclusion of extra energy
dependence as required by general principles, such as it follows by implementing
two-body unitarity and the correct analytical properties of PWAs, are able to provide good results in
a large energy range, much larger than the one naively expected for current-algebra results.
This is a conclusion that has been strengthened along the years, at the same time that the chiral calculations
have been improved going to higher orders and the unitarization methods have become more
sophisticated.
\subsection{ERE and IAM}
\label{sec.200523.2}
Already at 1972 the calculation of the NLO ChPT amplitude was worked by Lehmann \cite{lehmann.200302.1} in the
chiral limit ($m_\pi\to 0$), much earlier than the seminal paper by Gasser and Leutwyler
\cite{leutwyler.200220.1}, which established the general framework for ChPT at ${\cal O}(p^4)$.
The author did not need to work out the chiral Lagrangians at NLO order because he only used unitarity, crossing
symmetry and analyticity to work out the chiral loops. The point is that because of unitarity a PWA satisfies
Eq.~\eqref{200304.3c}. However, unitarity is only satisfied perturbatively in the chiral expansion,
so that if we denote by $T_4(s)$ a one-loop ChPT PWA and $T_2(s)$ its LO, then perturbative unitarity requires that
\begin{align}
\label{200514.6a}
\Im T_4(s)&=T_2(s)^2 \frac{p}{8\pi\sqrt{s}}\theta(s-4m_\pi^2) ~,
\end{align}
a particular example of Eq.~\eqref{200611.2}.
The PWA $T_4(s)$ has LHC and RHC.
The discontinuity along the RHC is twice $i\Im T_4(s)$, because of the
Schwarz reflection principle. A DR that results by considering a closed circuit engulfing the RHC,
implies the following contribution to $T_4(s)$,
\begin{align}
\label{200514.6}
a+bs+cs^2
+\frac{s^3}{8\pi^2}\int_{4m_\pi^2}^\infty ds'\frac{T_2(s')^2\sqrt{s'/4-m_\pi^2}/\sqrt{s'}}{(s')^3(s'-s)}~.
\end{align}
Three subtractions have been taken because $T_2(s)$ at most diverges like $s$ in the limit $s\to \infty$.
By invoking crossing one can build up the one-loop contributions from the $t$- and $u$-channels for a given
process. As usual the Mandelstam variables are indicated by $s$, $t$ and $u$ ($s+t+u=0$ for massless pions).
In Cartesian coordinates for the pions and treating all of them on equal footing,
so that they are all e.g. incoming, one can write for the scattering amplitude
$\pi_1(k_1)\pi(k_2)\pi_{i_3}(k_3)\pi_{i_4}(k_4)\to 0$,
where the $k_i$ are the on-shell four-momenta ($k_i^2=0$, $\sum_i k_i=0$), the expression
\begin{align}
\label{200515.1}
T(s,t,u)&=\delta_{i_1i_2}\delta_{i_3i_4}A(s,t,u)+\delta_{i_1i_3}\delta_{i_2i_4}A(t,s,u)+\delta_{i_1i_4}\delta_{i_2i_3}A(u,t,s)~.
\end{align}
Here crossing has also been used to properly exchange the arguments of the $A(s,t,u)$ function.
The previous expression is manifestly symmetric in the indices $i_3$ and $i_4$ which also implies that, because
the pions are bosons, $A(s,t,u)$ is symmetric under the exchange $t\leftrightarrow u$.
Since the isospin coordinates run only from 1 to 3, two out of the four pions have the same coordinates necessarily.
In the calculation of Ref.~\cite{lehmann.200302.1} the resulting expression for $A(s,t,u)$ has two parts.
One of them corresponds to DR integrals of the type in Eq.~\eqref{200514.6},
in all the $s$-, $t$- and $u$-channels, which can be evaluated in an algebraic close form.
The other contribution is a second-order polynomial in the Mandelstam variables,
whose general expression can be written as $a+bs+cs^2+c'(t^2+u^2)$, which can also be extra constrained.
In this respect, $a=0$ because Goldstone particles do not interact
in the limit in which masses and four-momenta vanish.
The term $b s$ is order $p^2$ and it is already accounted for in $T_2(s)$.
As a result, the one-loop calculation of Lehmann only involves two unknown parameters,
nowadays typically called counterterms because they are associated to bare parameters appearing
at the NLO ChPT Lagrangian.
In terms of the $A(s,t,u)$ amplitude one can calculate the different $\pi\pi$ isospin PWAs \cite{iam.oller.long}, $T_{IJ}$.
An interesting point of Ref.~\cite{lehmann.200302.1} is the perturbative matching in the chiral expansion
of the calculated PWAs at ${\cal O}(p^4)$ with the ERE expression for a PWA, cf. Eq.~\eqref{200512.2}.
The subtle point is that the former only satisfies unitarity in a perturbative way, as discussed above.
Therefore, writing in the massless case that
\begin{align}
\label{200515.2}
\frac{\Re T_{IJ}}{\Im T_{IJ}}&=\cot\delta_{IJ}~,
\end{align}
is not right. The correct procedure is to write a chiral expansion of $1/T_{IJ}$ up to NLO and from there
to identify $\cot\delta_{IJ}$,
\begin{align}
\label{200515.3}
\frac{1}{T_{IJ}}&=\frac{1}{T_2+T_4}+{\cal O}(p^6)=\frac{1}{T_2}-\frac{T_4}{T_2^2}+{\cal O}(p^6)~.
\end{align}
Taking into account the perturbative unitarity satisfied by $T_4$, one can extract from here the NLO expression
for $\cot\delta_{IJ}$ (with a numerical normalization factor properly chosen) as, cf. Eq.~\eqref{200512.2b},
\begin{align}
\label{200515.4}
\frac{p}{\sqrt{s}}\cot\delta_{IJ}&=\frac{1}{T_2}-\frac{\Re T_4}{T_2^2}+{\cal O}(p^6)~.
\end{align}
This is indeed the first example that we know of a paper in the literature deriving the
expression of a PWA as
\begin{align}
\label{200515.5}
T_{IJ}&=\frac{T_2^2}{T_2-T_4}~.
\end{align}
This formula, generalized to any other two-body PWA and also to coupled channels,
is the basic one for the so-called IAM \cite{dobado.90,iam.oller.long}.
It also illustrates the connection between these earlier treatments based on the ERE and
this more modern method, which was named IAM after the general framework for the one-loop calculations
in ChPT was established in Ref.~\cite{leutwyler.200220.1}. The approach of Ref.~\cite{lehmann.200302.1} has
the advantage over the previous ERE of
Refs.~\cite{schnitzer.200221.1,schnitzer.200302.1,brown.200302.1,sakurai,brehm} that chiral one-loop contributions in the crossed channels
are also kept, so that the LHC is reproduced up to NLO in the inverse of the PWA.
The extension of Eq.~\eqref{200515.5} up to two-loop ChPT can be done straightforwardly by expanding
the inverse of $(T_2+T_4+T_6)^{-1}$ up to next-to-next-to-leading order (NNLO), or ${\cal O}(p^2)$.
The result is,
\begin{align}
\label{200515.5b}
T_{IJ}&=\frac{T_2^3}{T_2^2-T_4 T_2+T_4^2-T_2 T_6}~.
\end{align}
Taking into account that perturbative unitarity requires that $\Im t_6=2 T_2\rho \Re T_4$, it follows that
$T_{IJ}$ given by Eq.~\eqref{200515.5b} fulfills exact unitarity, $\Im T_{IJ}^{-1}=-\rho$.
The Eq.~\eqref{200515.5b} is the IAM at the two-loop order \cite{arriola.2iam}.
\subsection{IAM and Pad\'e approximants}
\label{sec.200523.3}
Another non-perturbative method used with the aim of improving the convergence of the QFT
calculations in perturbation theory is the
Pad\'e resummation technique \cite{basdevant.zinn}.
It is also a unitarization method that
was applied since the early days of current algebra calculations by Refs.~\cite{basdevant.lee,basdevant.review}, in which the linear
$\sigma$ model was considered too. An interesting {\it qualitative} agreement with data
for the $\pi\pi$ $S$-, $P$- and $D$-waves was reported, despite the limitations of the theoretical input.
Given a function $f(z)$ that is analytic at $z=0$, its Taylor series expansion around this point
converges within the circle of radius $R$, which is the distance to the nearest singularity.
However, it is also known that the value of $f(z)$ at a point $z_1$ within its domain of analyticity,
but beyond the radius of convergence of the Taylor series around $z=0$, is fixed
by the coefficients in the later expansion.
The idea of the Pad\'e method is to provide a resummation of the Taylor series
and build an approximation of $f(z)$ beyond the radius of convergence of its Taylor series around $z=0$.
The Pad\'e approximant $[n,m]$ is given by the ratio of two polynomial functions $P_n(z)$ and $Q_m(z)$
of degrees $n$ and $m$, respectively, which has the same $n+m$ first derivatives as $f(z)$ at $z=0$. Namely,
\begin{align}
\label{200516.1}
f^{[n,m]}(z)&=\frac{P_n(z)}{Q_m(z)}=f(z)+{\cal O}(z^{n+m+1})~,~|z|<R~.
\end{align}
Notice that in particular the approximant $[n,0]$ is identical up
to ${\cal O}(z^n)$ with the Taylor series of $f(z)$ at $z=0$.
It is also typically the case that the Pad\'e approximants usually provide an acceleration in the rate of convergence
of the Taylor series itself. For instance, one can write that
\begin{align}
\label{200516.2}
\sqrt{1+z}&=1+\frac{z}{1+\sqrt{1+z}}~.
\end{align}
By iteration it can be expressed as a continued fraction, which are particular cases of Pad\'e approximants,
\begin{align}
\label{200516.3}
f^{[1,0]}&=\frac{2+z}{2}~,\\
f^{[1,1]}&=\frac{4+3z}{4+z}~,\nonumber\\
f^{[2,1]}&=\frac{8+8z+z^2}{8+4z}~,\nonumber\\
f^{[2,2]}&=\frac{16+20z+5z^2}{16+12z+z^2}~,\nonumber
\end{align}
etc. Let us compare the first four Pad\'e approximants with the first four terms in the Taylor series,
$\sqrt{1+z}=1+\frac{z}{2}-\frac{z^2}{8}+\frac{z^3}{16}+\ldots$ by
calculating $\sqrt{2}=1.4142$.
We then obtain the sequence of approximate results from the Pad\'e approach $\{1.5,1.4,1.4167,1.4138\}$, and the Taylor series $\{1,1.5,1.375,1.4375\}$. It is clear
the improvement in the convergence properties achieved by the Pad\'e method in this case.
The formulas for the IAM at one- and two-loop ChPT, Eqs.~\eqref{200515.5} and \eqref{200515.5b},
respectively, can also be obtained as Pad\'e approximants, where a generic small parameter $\varepsilon$ accounts
for the chiral order. Formally, we then write $T_2\to \varepsilon^2 t_2$, $T_4\to \varepsilon^4 t_4$ and $T_6\to \varepsilon^6 t_6$.
The one-loop IAM is a $[1,1]$ Pad\'e approximant:
\begin{align}
\label{200517.1}
t^{[1,1]}(s)&=\frac{a_0+\varepsilon^2 a_2}{1+\varepsilon^2 b_2}=\varepsilon^2 t_2+\varepsilon^4 t_4+{\cal O}(\varepsilon^6)~.
\end{align}
To solve this type of equation, typically found in Pad\'e approximants,
it is convenient to rewrite Eq.~\eqref{200517.1} as
\begin{align}
\label{200517.2a}
a_0+\varepsilon^2 a_2&=(1+\varepsilon^2 b_2)(\varepsilon^2 t_2+\varepsilon^4 t_4)+{\cal O}(\varepsilon^6)~.
\end{align}
By matching the different powers of $\varepsilon^2$ one has that
\begin{align}
\label{200517.2}
a_0&=0~,\\
a_2&=t_2~,\nonumber\\
b_2&=-t_4/t_2~.\nonumber
\end{align}
From which it follows that
\begin{align}
\label{200517.3}
t^{[1,1]}&=\frac{T_2}{1-T_4/T_2}=\frac{T_2^2}{T_2-T_4}~.
\end{align}
For the approximant $[1,2]$
\begin{align}
\label{200517.4}
t^{[1,2]}(s)&=\frac{a_0+\varepsilon^2 a_2}{1+\varepsilon^2 b_2+\varepsilon^4 b_4}
=\varepsilon^2 t_2+\varepsilon^4 t_4+\varepsilon^6 t_6+{\cal O}(\varepsilon^8)~.
\end{align}
The result of the matching is the same as in Eq.~\eqref{200517.2} for
$a_0$, $a_2$ and $b_2$, and the extra new parameter $b_4$ is
\begin{align}
\label{200517.5}
b_4&=(t_4^2-t_2 t_6)/t_2^2 ~.
\end{align}
Therefore,
\begin{align}
\label{200517.6}
t^{[1,2]}&=\frac{T_2}{1-\frac{T_4}{T_2}+\frac{T_4^2-T_2 T_6}{T_2^2}}
=\frac{T_2^3}{T_2^2-T_2 T_4+T_4^2-T_2 T_6}~,
\end{align}
as Eq.~\eqref{200515.5b}.
\section{Final-State Interactions}
\setcounter{equation}{0}
\label{sec.200523.4}
As a canonical example of taking into account the FSI that correct the production processes
due to weaker probes because of the rescattering by the stronger interactions, we start with the
unitarization of the vector pion form factor, $F_V(s)$, within the ERE approach of Ref.~\cite{sakurai}.
We next move to the Omn\`es solution for a form factor and also consider the scalar pion form factor, $F_S(s)$,
paying attention to a caveat in the use of an Omn\`es function that one should properly consider.
Along the discussion we introduce the way FSI are treated in Ref.~\cite{truong.200211.2}, as it is
probably the first paper in which NLO ChPT is unitarized to account for FSI following the basic notions of
unitarity, Watson final-state theorem and use of an Omn\`es function, which are the basic elements usually employed in
the different modern approaches to resum FSI \cite{oller.book,oller.review}.
We end this section with a basic account of the Khuri-Treiman approach for $\eta\to 3\pi$ decays.
\subsection{ERE, the Omn\`es solution and coupled channels}
\label{sec.200514.1}
The application of the ERE for implementing the FSI of the pion vector form factor was pioneered in Ref.~\cite{sakurai}.
The main aim of this paper concerns the corrections because of the finite width of the $\rho$
to the VMD dominance relation between $\Gamma(\rho\to e^+e^-)$ and $\Gamma(\rho\to\pi^+\pi^-)$,
as well as to characterize the energy shape of $\Gamma(e^+e^-\to \pi^+\pi^-)$.
Ref.~\cite{sakurai} implemented the relationship between the $I=J=1$ $\pi\pi$ PWA and
the pion form factor $F_V(s)$ by writing $F_V(s)=T_1(s)/t_2(s)$, with $t_2(s)$ the LO ChPT amplitude.
This relation is a consequence of the Omn\`es representation in the approximation in which:
i) One assumes that the only zero in $T_1(s)$ in the region of interest
is the one at threshold, $s=4m_\pi^2$, because of the $\ell=1$ centrifugal barrier;
ii) one also assumes the dominance of the $\rho(770)$ exchange so that it is a good approximation to
consider that $T_1(s)$ is dominated by $s$-channel dynamics
\footnote{Under these assumptions $T_1(s)$ is given by the Omn\`es function on the rhs of Eq.~\eqref{200216.4}
times $(s-4m_\pi^2)/48\pi f_\pi^2$.}
Thus,
\begin{align}
\label{200514.3}
F_V(s)&=\frac{48\pi f_\pi^2 T_1(s)}{s-4m_\pi^2}=\frac{T_1(s) a_1}{p^2}~,
\end{align}
guaranteeing that $F_V(0)=1$ because of conservation of total charge.
Next, Ref.~\cite{sakurai} performs the same ERE of Ref.~\cite{brown.200302.1},
which we have already discussed, cf. Eq.~\eqref{200513.3},
which allows to finally write the form factor in a successful manner as
\begin{align}
\label{200514.4}
F_V(s)=\frac{1}{1+\frac{r_1p^2}{2a_1}+\frac{p^2}{a_1}h(s)}~.
\end{align}
The authors of Ref.~\cite{sakurai} simplify further this expression by removing those terms involving the expansion of
the real part of $h(s)$ around $s=m_\rho^2$ that are at least of ${\cal O}(s-m_\rho^2)$.
They finally write
\begin{align}
\label{200514.5}
F_V(s)&=\frac{m_\rho^2[1+d_1 m_\rho/\Gamma_\rho]}{m_\rho^2-s-im_\rho \Gamma_\rho (p/p_\rho)^3(m_\rho/\sqrt{s})}~,\\
d_1&=\frac{3}{\pi}\frac{m_\pi^2}{p_\rho^2}\log\left(\frac{m_\rho+2p_\rho}{2m_\pi}\right)
+\frac{m_\rho}{2\pi p_\rho}-\frac{m_\pi^2m_\rho}{\pi p_\rho^3}~,\nonumber\\
p_\rho&=\sqrt{m_\rho^2/4-m_\pi^2}~.\nonumber
\end{align}
Again, one concludes that the extrapolation of the current-algebra results plus the extra energy dependence that
arises by implementing the basic principles
of two-body unitarity and analyticity allows one to reach much higher energies than
expected, even above the 1~GeV frontier.
Writing a form factor proportional to a given PWA is usually employed
in many cases in the literature. The basic reason is to provide an expression for the
coupled form factors $F_i(s)$ that automatically satisfies the constraint imposed by the two-body
unitarity, cf. Eq.~\eqref{200220.4}. Following Ref.~\cite{au} one then writes
\begin{align}
\label{200515.6}
F_i&=\sum_j \widetilde{\alpha}_j T_{ji}~,
\end{align}
where the sum is over the strongly-coupled channels.
The functions $\widetilde{\alpha}_i$ are real and they are also expected to be smooth because
all the RHC features in $F_i(s)$ are included in the PWAs $T_{ij}(s)$. As a result, the
$\widetilde{\alpha}_i$ should not have nearby singularities, if any. They could involve crossed-channel
cuts which could be mimicked typically by parameterizing these functions by low-degree polynomials.
Nonetheless, in the case of the low-energy interactions of the lightest pseudoscalars, like pions,
an extra feature is the presence of the Adler zeroes in the $S$ waves. In particular, for
$I=J=0$ we have already discussed that this Adler zero is around $s_A=m_\pi^2/2$, cf. Eq.~\eqref{200509.3}.
The existence of Adler zeros is a characteristic feature of the interactions of the Goldstone bosons, as said, but not
necessarily for their production through external currents. To handle with such cases Ref.~\cite{au}
proposes to explicitly remove the Adler zeroes in the $T_{ij}(s)$, when they are present, and any
necessary zero in the production process is then explicitly included in the prefactors. Denoting by
${\cal T}_{ij}(s)=T_{ij}(s)/(s-s_{A_{ij}})$, with $s_{A_{ij}}$ the Adler zero in $T_{ij}(s)$, the final expression proposed is
\begin{align}
\label{200515.7}
F_i&=\sum_j \alpha_j {\cal T}_{ji}~.
\end{align}
For the case of only one coupled channel, the form factor can be expressed in terms of an Omn\`es function
$\Omega(s)$.
Due to the Watson final-state theorem the continuous phase of the form factor $\varphi(s)$
is the same as the phase shift $\delta(s)$ for the PWA $T(s)$.
The Omn\`es function results by performing a DR for the logarithm of the function
$f(s)=F(s)Q(s)/P(s)$, where $P(s)$ and $Q(s)$ are the polynomials whose only roots are the possible
zeros and poles of $F(s)$, respectively, which are assumed to be finite in number.
The discontinuity of $\log f(s)$ along the RHC is the discontinuity of its imaginary
part, and it is given by $2i\varphi(s)$.
We can then write the following expression for the DR of $\omega(s)\equiv \log f(s)$,\footnote{For a more extensive discussion
on the Muskhelishvili-Omn\`es problem the reader can consult Refs.~\cite{oller.book,oller.review}.}
\begin{align}
\label{181120.8}
\omega(s)&=\sum_{i=0}^{n-1}a_i s^i+\frac{s^n}{\pi}\int_{s_{\rm th}}^\infty \frac{\varphi(s')ds'}{(s')^n(s'-s)}~,
\end{align}
where we have taken $n$ subtractions assuming that $\varphi(s)$ does not diverge stronger than $s^{n-1}$ when $s\to\infty$.
The Omn\`es function $\Omega(s)$ is defined in terms of $\omega(s)$ as
\begin{align}
\label{181120.9}
\Omega(s)=\exp{\omega(s)}~.
\end{align}
One can always normalize the Omn\`es function such that $\Omega(0)=1$, which fixes $a_0=1$.
In this manner we always take at least one subtraction.
It is also clear that the ratio
\begin{align}
\label{181120.10}
R(s)&=\frac{F(s)}{\Omega(s)}~,
\end{align}
is a meromorphic function of $s$ in the first RS of the cut complex $s$ plane,
being analytic in this whole
plane if $F(s)$ has no bound states.
As it is well known, any analytical function in the whole complex plane is either a constant or it is unbounded,
which is then the case for $R(s)$ too under the stated assumptions.
Therefore,
\begin{align}
\label{181120.11}
F(s)&=R(s)\Omega(s)~,
\end{align}
diverges as much as or stronger than $\Omega(s)$ for $s\to \infty$.
The function $\omega(s)$ would have severe divergences for $s\to\infty$ if its DR required for convergence more than one subtraction.
The reason is that if $\varphi(s)/s^{n-1}$ ($n \geq 2$) has no zero limit for $s\to \infty$,
the DR for $\omega(s)$ would be affected by logarithmic divergences like $s^{n-1} \log s$ which could not be
cancelled by the subtractive polynomial.
In such circumstances it would be required that $R(s)$ is a non-trivial analytical function in order to
cancel such divergences and guarantee that $F(s)$ can be represented as a DR.
If the conditions are met for a DR of $\log F(s)Q(s)/P(s)$, cf. Eq.~\eqref{181120.8},
then $R(s)=Q(s)/P(s)$ is a rational function. Thus, from the previous analysis, we conclude that
the DR of $\omega(s)$ in Eq.~\eqref{181120.8} involves only one subtraction and
it is then necessary that $|\varphi(s)/s|<s^{-\gamma}$ for some $\gamma>0$ in the limit $s\to \infty$.
We can then write the following representation for $F(s)$,
\begin{align}
\label{181121.1}
F(s)&=\frac{P(s)}{Q(s)}\Omega(s)~,\\
\label{181121.2}
\Omega(s)&=\exp\omega(s)~,\\
\label{181121.3}
\omega(s)&=\frac{s}{\pi}\int_{s_{\rm th}}^\infty \frac{\varphi(s')ds'}{s'(s'-s)}~.
\end{align}
The presence of $P(s)$ makes clear that one can fix de normalization of the Omn\`es function, $\Omega(0)=1$,
without any loss of generality.
The asymptotic behavior of $\Omega(s)$ in the limit $s\to \infty$ can be calculated as follows.
Let us rewrite $\omega(s)$ in Eq.~\eqref{181121.3} as
\begin{align}
\label{181122.1}
\omega(s)&=\varphi(\infty)\frac{s}{\pi}\int_{s_{\rm th}}^\infty \frac{ds'}{s'(s'-s)}
+\frac{s}{\pi}\int_{s_{\rm th}}^\infty \frac{\varphi(s')-\varphi(\infty)}{s'(s'-s)}ds'~,
\end{align}
with $\varphi(\infty)=\lim_{s\to\infty}\varphi(s+i\epsilon)$.
Then,
\begin{align}
\label{181122.2}
\omega(s+i\varepsilon)&\xrightarrow[s\to\infty]{} -\frac{\varphi(\infty)}{\pi}\log \frac{s}{s_{\rm th}} + i\varphi(\infty)
-\frac{1}{\pi}\int_{s_{\rm th}}^\infty\frac{\varphi(s')-\varphi(\infty)}{s'}ds'~,
\end{align}
being the limit $s\to\infty$ dominated by the logarithmic divergence,
as the other two terms in this equation are constants.
It follows from here the limit behavior
\begin{align}
\label{181122.3}
\Omega(s)&\xrightarrow[s\to \infty]{} {\cal C}_{\Omega}\, e^{i\varphi(\infty)}\times
\left(\frac{s_{\rm th}}{s}\right)^\frac{\varphi(\infty)}{\pi}~.
\end{align}
This result, together with Eq.~\eqref{181121.1}, implies that the asymptotic behavior for $F(s)$ is
\begin{align}
\label{181122.4}
F(s)&\xrightarrow[s\to\infty]{}{\cal C}_{F}\, e^{i\varphi(\infty)}\times s^{p-q-\frac{\varphi(\infty)}{\pi}}~,
\end{align}
where $C_{\Omega}$ and $C_{F}$ are constants, and $p$ and $q$ are the number of zeros and poles of $F(s)$, respectively (or equivalently,
the degrees of $P(s)$ and $Q(s)$, in this order).
Two interesting consequences follow from Eq.~\eqref{181122.4}:
\noindent
i) If the asymptotic high-energy behavior of $F(s)$ is known to be proportional to $s^\nu$, then
\begin{align}
\label{181122.5}
p-q-\frac{\varphi(\infty)}{\pi}=\nu~.
\end{align}
\noindent
ii) Under changes of the parameters when modeling strong interactions one should keep Eq.~\eqref{181122.5}
unchanged. As $\nu$ is a known constant, then
\begin{align}
\label{181122.6}
p-q-\frac{\varphi(\infty)}{\pi}={\rm fixed}~.
\end{align}
For instance, if $\varphi(\infty)/\pi$ increases by one and there are no bound states
then an extra zero should be introduced in the form factor to satisfy Eq.~\eqref{181122.6}.
A similar procedure would be applied for other scenarios.
It is worth stressing that by using Eq.~\eqref{181121.1} one can guarantee that Eq.~\eqref{181122.6} is fulfilled,
while this is not the case for $\Omega(s)$.
The use of this function without taking proper care of the rational function $P(s)/Q(s)$, included in the expression for
$F(s)$ in Eq.~\eqref{181121.1}, could drive to an unstable behavior under changes of the parameters, e.g. in a fit to data.
This problem was originally discussed in Ref.~\cite{ollerz.200220.1} in connection with the scalar form factor of the pion $F_S(s)$,
to which we refer for further details in the discussion that follows.
This form factor is associated with the light-quark scalar source, $\bar{u}u+\bar{d}d$, and is defined as
\begin{align}
\label{181122.7}
F(s)&=\hat{m}\int d^4 x e^{i(p+p')x}\langle0|\bar{u}(x)u(x)+\bar{d}(x)d(x)|0\rangle~,
\end{align}
where $u$ and $d$ are the up and down quarks, $\hat{m}$ is their masses, and $s=(p+p')^2$.
Because of the quantum numbers of the non-strange scalar source, the FSI
occur in the isoscalar scalar meson-meson scattering, introduced in Sec.~\ref{sec.200511.1}.
There, we discuss the Adler zero required by chiral symmetry and the pole of the
$\sigma$ or $f_0(500)$ resonance, being both of them related by unitarity, analyticity and chiral symmetry.
At around the two-kaon threshold, $\sqrt{s}=991.4$~MeV, the $K\bar{K}$ channel makes a big impact.
This energy almost coincides with the sharp emergence of the $f_0(980)$ resonance,
which gives rise to a rapid increase of the $\pi\pi$ isoscalar scalar phase shifts $\delta_{00}(s)$, since
it is a relatively narrow resonance \cite{pdg}, cf. Fig.~\ref{fig.200524.1}.
The elasticity parameter $\eta_{00}$ also experiences a sharp reduction as soon as the $K\bar{K}$ channel open,
since the $f_0(980)$ couples much more to $K\bar{K}$ than to $\pi\pi$ \cite{guo.181123.1}.
This phenomenon causes an active conversion of the pionic flux into the kaonic one.
The rapid rise of the isoscalar scalar $\pi\pi$ phase shifts,
also implies the corresponding rise of the phase of the isoscalar scalar PWA $T(s)$, $\varphi(s)$, because they coincide
below the $K\bar{K}$ threshold, i.e. for $\sqrt{s}<2m_K$.
However, above this energy the rise of $\varphi(s)$ is interrupted abruptly
if $\delta_{00}(s_K)<\pi$, with $s_K=4m_K^2$, while in the opposite case $\varphi(s)$ keeps increasing.
Quite interestingly, the two situations can be connected by tiny variations in the values of the parameters in the
hadronic model, while keeping compatibility with the experimental phase shifts at around the $f_0(980)$ mass.
As a result, there is a jump in the limiting value of $\Omega(s)$ because $\varphi(\infty)$ changes by $\pi$.
Thus, in order to keep constant Eq.~\eqref{181122.6} under an increase by $\pi$ in $\varphi(\infty)$ for
$\delta_{00}(s_K)>\pi$, it is necessary to increase $p$ by one unit, so that a zero is necessary in $F_S(s)$ that is not
present when $\delta_{00}(s_K)<\pi$.
For completeness, we also mention that had we required the continuity from $\delta_{00}(s_K)>\pi$
to $\delta_{00}(s_K)<\pi$ then an extra pole (in the first RS) should be added.
This latter scenario can be disregarded in $\pi\pi$ scattering because of the absence of bound states.\footnote{With
respect to the difference between $\varphi(s)$ and $\delta_{00}(s)$, as indicated above,
the $f_0(980)$ dominates the behavior of the isoscalar scalar meson-meson
scattering around 1~GeV, and couples much more strongly to kaons than to pions.
For instance, Ref.~\cite{guo.181123.1} calculates that its coupling to kaons is a factor 3 larger than that to pions.
This makes that the mixing between the pion and kaon scalar form factors is suppressed, following each of them
its own eigenchannel of the $I=J=0$ PWAs.}
Let $s_1$ be the value of $s$ at which the pion scalar form factor has a zero for $\delta(s_K)>\pi$. Then, we can write
an Omn\`es representation of the pion scalar form factor in terms of a modified Omn\`es function
\begin{align}
\label{181123.5}
\Omega(s)&=\left\{
\begin{array}{ll}
\exp\omega(s)& ~,~\delta(s_K)<\pi~,\\
\frac{s_1-s}{s_1}\exp\omega(s) & ~,~\delta(s_K)>\pi~,
\end{array}
\right.
\end{align}
such that $F_S(s)=F_S(0)\Omega(s)$. From here it is clear that $s_1$ can be fixed by the
requirement that $\Im F(s_1)=0$.
Because of the Watson final-state theorem in the elastic region
we can write that $\Im F(s)=|F(s)|\sin \delta_{00}(s)|/\rho(s)$ and it vanishes when
$\delta_{00}(s_1)=\pi$, which allows to determine $s_1$ from the knowledge of $\delta_{00}(s)$.
The context clarifies whether the same symbol $\Omega(s)$ actually refers to Eq.~\eqref{181120.9} or Eq.~\eqref{181123.5}.
A clear lesson from the discussion here is that possible troubles could occur when using an Omn\`es function
in fitting the free parameters because an unstable behavior could arise due to a jump in $\varphi(\infty)$.
These regions of dramatic differences in $\exp\omega(s)$ are separated by a discontinuity of $\varphi(s)$ in the parametric space.
As a consequence, it is important in the fitting process to satisfy the condition Eq.~\eqref{181122.6}.
In particular, for the $I=J=0$ $\pi\pi$ PWA the more elaborated function in Eq.~\eqref{181123.5} should be used, instead of just the
standard Omn\`es $\exp\omega(s)$ given in Eq.~\eqref{181121.2}.
This fact also affects studies of two-photon fusion into two
pions, like Ref.~ \cite{pen.prl}, as discussed in Ref.~\cite{schat.gama}.
\subsection{The IAM for FSI}
\label{sec.200515.1}
The first step of Ref.~\cite{truong.200211.2} is to write down twice subtracted DR expressions for the scalar and vector pion form factors,
$F_S(s)$ and $F_V(s)$, respectively, as
\begin{align}
\label{200216.1}
F_S(s)&=1+\frac{\langle r_S^2\rangle s}{6}+\frac{s^2}{\pi} \int_{4m_\pi^2}^\infty
\frac{F_S(s')e^{-i\delta_{00}}\sin\delta_{00}(s')ds'}{s^{'2}(s'-s-i\epsilon)}~,\\
\label{200216.2}
F_V(s)&=1+\frac{\langle r_V^2\rangle s}{6}+\frac{s^2}{\pi} \int_{4m_\pi^2}^\infty
\frac{F_V(s')e^{-i\delta_{11}}\sin\delta_{11}(s')ds'}{s^{'2}(s'-s-i\epsilon)}~.
\end{align}
Here, $\delta_{00}(s)$ and $\delta_{11}(s)$ are the $J=0$ and 1 isoscalar and isovector
$\pi\pi$ phase shifts, in this order.
These DRs can be interpreted as singular integral equations (IEs) for the form factors $F_S(s)$ and
$F_V(s)$ \cite{jamin.ff}.
Let us remark, as in Ref.~\cite{truong.200211.2},
that the solutions of the IEs of Eqs.~\eqref{200216.1} and \eqref{200216.2}
for $F_S(s)$ and $F_V(s)$, respectively, can be expressed in terms of the associated Omn\`es
functions \cite{meissgasser}.
In the approximation of identifying the phases of the form factors with
the phase shifts, strictly valid only for the elastic region, one has the approximate expressions
\begin{align}
\label{200216.3}
F_S(s)&=P_S(s)\exp\left[\frac{s}{\pi}\int_{4m_\pi^2}^\infty \frac{\delta_{00}(s')ds'}{s'(s'-s-i\epsilon)}\right]~,\\
\label{200216.4}
F_V(s)&=P_V(s)\exp\left[\frac{s}{\pi}\int_{4m_\pi^2}^\infty \frac{\delta_{11}(s')ds'}{s'(s'-s-i\epsilon)}\right]~,
\end{align}
where $P_S(s)$ and $P_V(s)$ are polynomials that take into account the zeros (if any) of the form factors in the
first or physical RS.
At the one-loop order in ChPT or, equivalently, at next-to-leading order NLO or ${\cal O}(p^4)$,
we can replace inside the dispersive integrals of Eq.~\eqref{200216.1} the $\pi\pi$ scattering PWAs
at leading order,
\begin{align}
\label{200219.1}
f_0(s)&=\sin\delta_{00}e^{i\delta_{00}}=\delta_{00}(s)+{\cal O}(p^4)=\frac{\sigma(s)}{16\pi}\frac{s-m_\pi^2/2}{f^2}+{\cal O}(p^4)~,\\
\label{200304.2}
f_1(s)&=\sin\delta_{11} e^{i\delta_{11}}=\delta_{11}(s)+{\cal O}(p^4)=\frac{\sigma(s)}{16\pi}\frac{s-4m_\pi^2}{6f^2}+{\cal O}(p^4)~.
\end{align}
The phase space function $\sigma(s)$ is defined in Eq.~\eqref{200508.2}.
Evaluating the dispersive integral in Eq.~\eqref{200216.1} with the approximation for $f_0(s)$
of Eq.~\eqref{200219.1},
Ref.~\cite{truong.200211.2} of course ends with the same expression for $F_S(s)$ as the NLO ChPT \cite{leutwyler.200220.1}
result,
\begin{align}
\label{200219.3}
F_S(s)&=1+\frac{s}{6}\langle r_S^2\rangle - \frac{1}{16\pi^2 f^2}\left[(s-m_\pi^2/2)[h(s)-h(0)]+\frac{m_\pi^2}{2}h'(0)s\right]+{\cal O}(p^6)~.
\end{align}
The function $h(s)$ is defined in Eq.~\eqref{200514.1}.
By proceeding in an analogous way,
a similar expression holds for the vector form factor at this level of accuracy, ${\cal O}(p^4)$,
\begin{align}
\label{200219.3b}
F_V(s)&=1+\frac{s}{6}\langle r_V^2\rangle - \frac{1}{96\pi^2 f^2}\left[(s-4m_\pi^2)[h(s)-h(0)]+ 4 m_\pi^2 h'(0)s\right]+{\cal O}(p^6)~.
\end{align}
There is an important difference between the scalar and vector form factors.
The unitarity corrections are enhanced by around a factor 6 for the former compared to the latter, because the
leading order ChPT amplitude is around a factor 6 larger, compared Eqs.~\eqref{200219.1} and \eqref{200304.2}, as
first noticed in Ref.~\cite{nd} and already discussed above in detail.
By invoking the Watson final-state theorem, one can calculate from the perturbative expressions
of $F_S(s)$ and $F_V(s)$ in Eqs.~\eqref{200219.3} and \eqref{200219.3b} the $\pi\pi$ phase shifts for $J=0$ and 1,
respectively.
Nonetheless, since the form factors are calculated perturbatively one should
proceed consistently in order to extract from this perturbative information the corresponding phase shifts.
In this way,
denoting by $F_2(s)$ the LO form factors and by $F_4(s)=F_4^r(s)+iF_4^i(s)$ their NLO contributions,
with the superscripts indicating the real ($r$) and imaginary ($i$) parts, we then have for the
Watson final-state theorem:
\begin{align}
\label{200304.3}
F(s)&=
|F(s)|e^{i\phi}=\sqrt{(F_2+F_4^r)^2+(F_4^i)^2}e^{i\phi}=
F_2\sqrt{ \left( 1 + \frac{F_4^r}{F_2} \right) ^2 + \left( \frac{F_4^i}{F_2} \right) ^2} e^{i\phi}\\
&=F_2+F_4^r+i\phi F_2+{\cal O}(p^6)~,\nonumber
\end{align}
from where the phase $\phi$ can be extracted.
Let us notice that
Ref.~\cite{truong.200211.2} compared directly the phase of the perturbative form factors
in Eqs.~\eqref{200219.3} and \eqref{200219.3b} with the
phase shifts of the $\pi\pi$ PWAs in its Fig.~1 and 2.
In this respect, it did no take account that this is not meaningful because
the Watson final-state theorem only holds perturbatively in ChPT, as explained.
\begin{figure}[H]
\begin{center}
\includegraphics[width=0.85\textwidth]{truprlfs.pdf}
\caption{{\small The top row concerns the $\pi\pi$ scalar form factor $F_S(s)$ and the
bottom one the $\pi\pi$ vector form factor $F_V(s)$.
In each row the left panel refers to the phase and the right panel to the modulus squared of the
corresponding form factor.
The perturbative calculations are indicated by the (magenta) dot-dashed lines in all cases.
The non-perturbative result for $F_S(s)$ are shown by the (black) solid lines.
For $F_V(s)$ we show two lines for the IAM resummation, Eq.~\eqref{200220.9},
the (black) solid lines and the (red) dashed ones.
The former employs $\langle r_V^2\rangle=0.42$~fm$^2$ (used in Ref.~\cite{truong.200211.2}) and the latter $\langle r_V^2\rangle=0.41$~fm$^2$.
The $\rho-\omega$ mixing, clearly visible at the top of $|F_V(s)|^2$, is not discussed here.
The experimental points for the $I=J=1$ $\pi\pi$ phase shifts are from Refs.~\cite{pipi11}, and
those for $|F_V(s)|^2$ were obtained in Ref.~\cite{1985.barkov}.
For the $I=J=0$ phase shifts we use the subset of points employed in Fig.~\ref{fig.200524.1}
and that appear on the top in the $f_0(500)$ region.}
\label{fig.200220.1}}
\end{center}
\end{figure}
We show in Fig.~\ref{fig.200220.1} the resulting form factors, so that the top line is dedicated to $F_S(s)$
and the bottom one to $F_V(s)$. The panels on the left correspond to the phases of these form factors and
the panels on the right to their module squared.
It is clear that there is a strong departure between the calculated phase shifts from the NLO ChPT form factors (magenta dashed lines) and the experimental values even at low values of $s$.
This is also clearly true for the modulus squared of $F_V(s)$, for which the perturbative calculation
again departures strongly from the experimental points.
It is particularly visible there the emergence of the resonance $\rho(770)$,
which dominates the phase shifts and $|F_V(s)|^2$,
with tails extending up to threshold and affecting the low-energy results.
This phenomenon can only be captured approximately in $SU(2)$ ChPT by the large size of the counterterm $\bar{\ell}_6$,
\begin{align}
\bar{\ell}_6&=(4\pi f_\pi)^2\langle r_V^2\rangle= 16.5\pm 1.1~,
\end{align}
as estimated in Ref.~\cite{leutwyler.200220.1}.
For the vector case, the cause of the large higher-order contributions is clearly associated with the
prominent role played by the $\rho(770)$ resonance.
In turn, for the scalar sector the enhanced RHC is the one blamed for such effects.
Indeed, these strong contributions from unitarity and analyticity even drive
to the emergence of a pole in the complex $s$ plane, the $\sigma$ or $f_0(500)$ resonance, as already
discussed, cf. Eq.~\eqref{200509.3}.
Ref.~\cite{truong.200211.2} discusses that the application of the chiral series expansion should be
performed on the inverse of the form factor rather than on the form factor itself.
The main reason lies on sound and general grounds, as provided by unitarity and analyticity.
Let us consider a DR representation of $F^{-1}(s)$, analogous to Eq.~\eqref{200216.1}.
The point to be stressed is that the imaginary part of $F^{-1}(s)$ is expected to be much smoother
than the imaginary part of $F(s)$ itself in the elastic region.
The reason is that the imaginary part of the inverse of the form factor satisfies,
because of unitarity in PWAs, that
\begin{align}
\label{200220.6}
\Im F^{-1}(s)=-\frac{\Im F(s)}{|F(s)|^2}=-\rho(s)\left(\frac{T(s)}{F(s)}\right)
=-\rho(s)\left(\frac{T(s)}{F(s)}\right)^*~.
\end{align}
As $F(s)$ and $T(s)$ share the same resonances, their propagators cancel in
the ratio $T(s)/F(s)$ that gives $\Im F^{-1}(s)$.
Then, this ratio is expected to be smoother than $\Im F(s)=\rho(s)F(s)^*T(s)$, where this
cancellation does not occur but rather the resonance effects in $F(s)$ and $T(s)$ mutually enhance each other
because of the product involved.
Then, let us write down a twice-subtracted DR for the inverses of the form factors $F_S(s)$ and $F_V(s)$.
First, we neglect by now the possible presence of zeroes in the form factors in the 1st Riemann sheet (RS),
which give rise to poles in the inverse of the form factors.
The issue of a zero in $F_S(s)$ for certain types of $T$ matrices was already discussed in
Sec.~\ref{sec.200514.1}, as first shown to happen in Ref.~\cite{ollerz.200220.1}.
This is not an issue here because we are considering the one-channel elastic scattering in the
isoscalar scalar $\pi\pi$ PWAs.
As a result we write,
\begin{align}
\label{200220.7}
F_S^{-1}(s)&=1-s\frac{\langle r_S^2\rangle}{6}-\frac{s^2}{\pi}\int_{4m_\pi^2}^\infty \frac{\rho(s')e^{i\delta_{00}}\sin\delta_{00}(s)\,F_S^{-1}(s')ds'}{(s')^2(s'-s-i\epsilon)}~,\\
\label{200220.8}
F_V^{-1}(s)&=1-s\frac{\langle r_V^2\rangle}{6}-\frac{s^2}{\pi}\int_{4m_\pi^2}^\infty \frac{\rho(s')e^{i\delta_{11}}\sin\delta_{11}(s)\,F_V^{-1}(s')ds'}{(s')^2(s'-s-i\epsilon)}~.
\end{align}
Then, up to ${\cal O}(p^4)$, in the integrand of these integrals one takes the leading order expressions in the
chiral expansion of $f_\ell(s)$, cf. Eqs.~\eqref{200219.1} and \eqref{200304.2}, and $F_{S,V}(s)=1$.
In this way, except for a global sign the same result as above is obtained for the dispersive integral as in the DR for $F_{S,V}(s)$.
Namely, the only difference is a flip of sign in the
NLO contributions
in Eqs.~\eqref{200219.3} and \eqref{200219.3b}.
Then, the
results for the form factors can be written as
\begin{align}
\label{200220.9}
F(s)&=\frac{1}{1-F_4(s)}~,
\end{align}
with $F(s)$ representing either $F_S(s)$ or $F_V(s)$ and $F_{4}(s)$ is the ${\cal O}(p^4)$ ChPT result. Similarly
$F_2(s)=1$ is the LO ChPT calculation.
Being more specific, Eq.~\eqref{200220.9} results after performing the
DR integrals, compare with Eqs.~\eqref{200216.1} and \eqref{200216.2},
\begin{align}
\label{200220.10}
F_S(s)&=\frac{1}{1-\frac{\langle r_S^2\rangle s}{6}-\frac{s^2}{(4\pi f_\pi)^2} \int_{4m_\pi^2}^\infty
\frac{\sigma(s')(s'-m_\pi^2)ds'}{s^{'2}(s'-s-i\epsilon)}}~,\\
\label{200220.11}
F_V(s)&=\frac{1}{1-\frac{\langle r_V^2\rangle s}{6}-\frac{s^2}{6(4\pi f_\pi)^2} \int_{4m_\pi^2}^\infty
\frac{\sigma(s')(s'-4m_\pi^2)ds'}{s^{'2}(s'-s-i\epsilon)}}~.
\end{align}
The resulting phase and modulus squared of $F_S(s)$
from Eq.~\eqref{200220.10} is shown by the (black) solid lines in the top panels of
Fig.~\ref{fig.200220.1}.
The resummed expression of $F_V(s)$ in Eq.~\eqref{200220.11}
gives rise to the results shown by the (black) solid and the (red) dashed lines
in the bottom panels of Fig.~\ref{fig.200220.1}.
They differ in the value of $\langle r_V^2\rangle$ employed, so that the former uses
0.42~fm$^2$ (as Ref.~\cite{truong.200211.2}),
and the latter takes the slightly lower value 0.41~fm$^2$, so as to agree better with the data on the
isovector vector $\pi\pi$ phase shifts.
We also use the updated value $f=92.4$~MeV, instead of 94~MeV used in Ref.~\cite{truong.200211.2}.\footnote{This reference indeed employs the normalization $f=133$~MeV=$94 \sqrt{2}$.}
It is clear that now, the resulting phase shifts calculated from the phases of the form factors in
Eqs.~\eqref{200220.10} and \eqref{200220.11}
are much closer to the experimental points than the perturbative ones form Eqs.~\eqref{200219.3} and \eqref{200219.3b}.
The same dramatic improvement also happens for the modulus squared of $F_V(s)$
calculated from Eq.~\eqref{200220.11}, as compared with the data points given by the empty circles.
In the peak of $|F_V(s)|^2$ it is clear the effect due to the $\rho-\omega$ mixing, which is not
treated here, see e.g. Ref.~\cite{palomar} for its implementation.
Notice that this improvement is achieved by employing the same perturbative input, namely the NLO ChPT results.
It is a matter of properly reshuffling the chiral expansion in a way clearly motivated by unitarity and analyticity.
We also show $|F_S(s)|^2$ in the right top panel of Fig.~\ref{fig.200220.1}, in which the resonance shape
due to the $f_0(500)$ is clearly visible. These resonance effects are not so evident in the case of the
isoscalar scalar phase shifts because of the Adler zero in this $\pi\pi$ PWA, which interferes strongly with the
pole contribution from the resonance itself.
\subsection{KT formalism}
\label{sec.200522.1}
The KT formalism was originally developed by Ref.~\cite{khuri} to study the $K\to 3\pi$ decays and,
up to including two-body intermediate states, it allows to implement unitarity and crossing symmetry.
Later on, this approach has been applied to study extensively the $\eta\to 3\pi$ decays, among others.
These decays violate isospin because the $G$ parity of the $\eta$ is $+1$ and that of the pion in $-1$,
so that it is proportional to $m_u-m_d$ in pure QCD.
The application of ChPT to the $\eta\to 3\pi$ decays has been controversial, until accepting
that FSI are so strong that a non-perturbative unitarization method is needed to be
implemented in order to be able to confront well with experimental data \cite{truong.200211.1}.
The earliest calculations using current-algebra techniques obtained a value for the $\eta\to\pi^+\pi^-\pi^0$ of
around 65~eV \cite{weinberg.190329.1}, too small compared with the experimental result
$\Gamma(\eta\to \pi^+\pi^+\pi^0)=(300 \pm 12)$~eV \cite{pdg}.
Roiesnel and Truong \cite{truong.200211.1}
stressed that a non-perturbative calculation taking care of the isoscalar-scalar
$\pi\pi$ FSI, by employing an Omn\`es function on top of the current-algebra result,
increases the decay width up to 200~eV.
A few years later, the NLO ChPT calculation \cite{CT.181127.3} gives $(160\pm 50)~$eV,
which implies a large correction by a
factor 2.4 over the LO calculation in the right direction, but still too small by around a factor of 2.
In addition, the parameter $ \alpha$, typically employed in the parameterization of the Dalitz plot for the decay
$\eta\to 3\pi^0$, is positive at NLO ChPT \cite{CT.181127.3}
while experimentally it is negative, $\alpha=-0.0318\pm 0.0015$ \cite{pdg}.
The calculation at NNLO in ChPT of the $\eta\to 3\pi$ decays was performed in Ref.~\cite{bijnens.190329.1} but
the proliferation of new counterterms prevented a sharp result.
If resonance saturation is assumed to estimate the NNLO ChPT counterterms then the Dalitz plot parameters
are not well reproduced.
One then concludes that the $\eta\to 3\pi$ decays are sensitive to the detailed values
of the ${\cal O}(p^6)$ counterterms, so that an accurate calculation requires a precise knowledge of their values.
This controversial situation stimulated the interest in developing sophisticated calculations
combining ChPT and non-perturbative methods, within
unitarized ChPT \cite{truong.200211.1,borasoy.190329.1,borasoy.190329.2} and the KT formalism
\cite{kambor.190329.1,anisovich.190329.1,jpac.190329.1,colangelo.190329.1,alba.190329.1}.
We now describe the basic points of the one-channel KT formalism for $\eta\to 3\pi$ decays and refer the reader
to Refs.~\cite{descotes.190329.1,alba.190329.1} and the recent review \cite{oller.review} for further details.
In particular, the generalization to coupled channels was worked out in Ref.~\cite{alba.190329.1}, given in a
more compact matrix notation in Ref.~\cite{oller.review}.
Let us consider the decay $\eta(p_0)\to\pi^+(p_1)\pi^-(p_2)\pi^0(p_3)$, which is
related by crossing symmetry to the scattering reactions $\eta(p_0)\pi^0(-p_3)\to\pi^+(p_1)\pi^-(p_ 2)$ in the $s$-channel,
$\eta(p_0)\pi^-(-p_1)\to\pi^0(p_3)\pi^-(p_ 2)$ in the $t$-channel,
and $\eta(p_0)\pi^+(-p_2)\to\pi^+(p_2)\pi^0(p_ 3)$ in the $u$-channel.
The Mandelstam variables $s$, $t$ and $u$ are given by
\begin{align}
\label{200522.1}
s&=(p_0-p_3)^2=(p_1+p_2)^2~,\\
t&=(p_0-p_1)^2=(p_2+p_3)^2~,\nonumber\\
u&=(p_0-p_2)^2=(p_1+p-3)^2~.\nonumber
\end{align}
The crossing-symmetry relations are
\begin{align}
\label{200522.2}
T(\eta\pi^0\to \pi^+\pi^-)&=A(s,t,u)~,\\
T(\eta \pi^- \to \pi^0\pi^-)&=A(t,s,u)~,\nonumber\\
T(\eta \pi^+\to \pi^+\pi^0)&=A(u,t,s)~.\nonumber
\end{align}
These amplitudes in turn can be decomposed in scattering amplitudes with well defined isospin, $M^I(s,t,u)$, as
\begin{align}
\label{200522.3}
A(\eta\pi^0\to \pi^+\pi^-)=A(s,t,u)&=-\frac{1}{3}M^2(s,t,u)+\frac{1}{3}M^0(s,t,u)~,\\
A(\eta\pi^+\to \pi^+\pi^0)=A(u,t,s)&=+\frac{1}{2}M^2(s,t,u)+\frac{1}{2}M^1(s,t,u)~,\nonumber\\
A(\eta\pi^-\to \pi^0\pi^-)=A(t,s,u)&=+\frac{1}{2}M^2(s,t,u)-\frac{1}{2}M^1(s,t,u)~.\nonumber
\end{align}
The inversion of these relations gives us the $M^I(s,t,u)$,
\begin{align}
\label{200522.4}
M^0(s,t,u)&=3A(s,t,u)+A(u,t,s)+A(t,s,u)~,\\
M^1(s,t,u)&=A(u,t,s)-A(t,s,u)~,\nonumber\\
M^2(s,t,u)&=A(u,t,s)+A(t,s,u)~.\nonumber
\end{align}
The PWA amplitudes are denoted by $M^{(IJ)}(s)$, and one has the standard relations
\begin{align}
\label{200522.5c}
M^{(IJ)}(s)&=\frac{1}{2}\int_{-1}^{+1}d\cos\theta\,P_J(\cos\theta)M^I(s,t,u)~,\\
M^I(s,t,u)&=\sum_{J=0}^\infty (2J+1)P_J(\cos\theta)M^{(IJ)}(\cos\theta)~.\nonumber
\end{align}
In the KT formalism the $S$ and $P$ waves are the ones that are subject to a non-perturbative
treatment.
The PWAs have a RHC above the two-pion threshold $s>4m_\pi^2$.
Instead of writing the unitarity constraint as in Eq.~\eqref{200304.3c}, one should consider it as giving
the discontinuity along the RHC because of the two on-shell intermediate pions. We then write\footnote{Due
to the fact that in the decay channel all the three pions are on-shell in the region $(m_\eta-m_\pi)^2\geq s \geq 4m_\pi^2$
this is another source of imaginary part from the crossed-channel cuts that are also on-shell.
For $s=(m_\eta^2-m_\pi^2)/2$ the branch point singularity at $t,u=4m_\pi^2$
happens for $\cos\theta=\mp 1$. These crossed-channel cuts can be separated from the RHC one by
giving a vanishing positive imaginary part to $m_\eta^2$ and then proceed by analytical continuation in
$m_\eta^2$ \cite{mandelstam.190403.1}.}
\begin{align}
\Im M^{(IJ)}&\to \frac{M^{(IJ)}(s+i\epsilon)-M^{(IJ)}(s-i\epsilon)}{2i}=\rho M^{(IJ)}(s+i\epsilon)M^{(IJ)}(s-i\epsilon)~,\nonumber \\
\label{200522.5a}
M^{(IJ)}(s+i\epsilon)&=\underbrace{\left(1+2i\rho M(s+i\epsilon)\right)}_{{\text{$S$-matrix~in~PWAs}}}M(s-i\epsilon)=e^{2i\delta^{(IJ)}(s)}M(s-i\epsilon)~.
\end{align}
From this last line we can write more conveniently the discontinuity of $M^{(IJ)}(s)$ along the RHC, $\Delta M^{(IJ)}(s)$, as
\begin{align}
\label{200522.5}
\Delta M^{(IJ)}(s)&=M^{(IJ)}(s+i\epsilon)-M^{(IJ)}(s-i\epsilon)=2i\sin\delta^{(IJ)}e^{-i\delta^{(IJ)}}M^{(IJ)}(s+i\epsilon)~,
\end{align}
which is the relation finally used.
A crucial feature of the KT formalism is to write down $A(s,t,u)$ as the sum of
three functions of only one Mandelstam variable, $M_0(s)$, $M_1(s)$ and $M_2(t)$ \cite{alba.190329.1,colangelo.190329.1}
\begin{align}
\label{200522.6}
A(s,t,u)&=M_0(s)-\frac{2}{3}M_2(s)+(s-u)M_1(t)+(s-t)M_1(u)+M_2(t)+M_2(u)~,
\end{align}
which is invariant under the exchange $t\leftrightarrow u$, a feature that can be seen as a consequence of
charge-conjugate invariance. This representation is valid up to ${\cal O}(p^8)$ in ChPT
\cite{bijnens.190329.1,colangelo.190329.1} because then the $\pi\pi$ $D$ waves
also contribute and higher polynomials in $(s-t)$ and $(s-u)$ would be required.
The derivation of Eq.~\eqref{200522.6} can be understood by considering only $J\leq 1$ PWAs in the $s$-channel
and taking into account the isospin decomposition
for the process $\eta\pi^0\to \pi^+\pi^-$ and the crossed-channel ones, cf. Eq.~\eqref{200522.3}.
In this way, for the $s$-channel process there is no $I=1$ contribution, which only happens in the
crossed ones, cf. Eq.~\eqref{200522.3}. As this is a $P$-wave we then write it as $M_1(t)(s-u)+M_1(u)(s-t)$, that
also keeps explicitly the symmetry under the exchange $t\leftrightarrow u$.
The $I=0$ contribution can only happen in the $s$-channel, because for the other channels the third component of isospin
is not zero.
This is the $M_0(s)$ contribution in Eq.~\eqref{200522.6}. Finally, regarding the $I=2$
it is clear from Eq.~\eqref{200522.3} that it appears in the combination
$-2 M_2(s)/3+M_2(t)+M_2(u)$.
Taking the expression for $A(s,t,u)$ in the ones of $M^{I}(s,t,u)$, as given in Eq.~\eqref{200522.4}, it follows that
\begin{align}
\label{200522.7}
M^0(s,t,u)&=3M_0(s)+M_0(t)+M_0(u)+\frac{10}{3}\big[M_2(t)+M_2(u)\big]+2(s-u)M_1(t)+2(s-t)M_1(u)~,\nonumber\\
M^1(s,t,u)&=2(u-t)M_1(s)+(u-s)M_1(t)-(t-s)M_1(u)+M_0(u)-M_0(t)+\frac{5}{3}\big[M_2(t)-M_2(u)\big]~,\nonumber\\
M^2(s,t,u)&=2M_2(s)+\frac{1}{3}\big[M_2(t)+M_2(u)\big]+M_0(t)+M_0(u)-(s-u)M_1(t)-(s-t)M_1(u)~.
\end{align}
Writing down the PWAs for $IJ=00,20$ and 11 we have
\begin{align}
\label{200522.8}
M^{00}(s)&\equiv 3\big[M_0(s)+\hat{M}_0(s)\big]\\
M^{11}(s)&\equiv -\frac{2}{3}\kappa\big[ M_1(s)+\hat{M}_1(s)\big]~, \nonumber\\
M^{20}(s)&\equiv 2\big[M_2(s)+\hat{M}_2(s)\big]~,\nonumber
\end{align}
where
\begin{align}
\label{200522.9}
\kappa(s)=\sqrt{\sigma(s)\lambda(s)}
\end{align}
with
\begin{align}
\label{200522.10}
\lambda(s)&=\lambda(s,m_1^2,m_2^2)=s^2+m_\eta^4+m_\pi^4-2s(m_\pi^2+m_\eta^2)-2m_\pi^2m_\eta^2~,\\
\sigma(s)&=1-\frac{4m_\pi^2}{s}~.\nonumber
\end{align}
We have also introduced in Eq.~\eqref{200522.8} the angular averages
\begin{align}
\label{200522.11}
\langle M_I\rangle_n&=\frac{1}{2}\int_{-1}^{+1}d\cos\theta\, \cos\theta^n
M^I(s,t(s,\cos\theta),u(s,\cos\theta))~,\\
\hat{M}_0(s)&=
\frac{2}{3}\langle M_0\rangle_0+\frac{20}{9}\langle M_2\rangle_0+2(s-s_0)\langle M_1\rangle_0
+\frac{2}{3}\kappa \langle M_1\rangle_0~,\nonumber\\
\kappa(s)\hat{M}_1(s)&=
\frac{9}{2}(s-s_0)\langle M_1\rangle_1 +\frac{3\kappa}{2}\langle M_1\rangle_2
+3\langle M_0\rangle_1-5\langle M_2\rangle_1 ~,\nonumber\\
\hat{M}_2(s)&=
\frac{1}{3}\langle M_2\rangle_0+ \langle M_0\rangle_0-\frac{3}{2}(s-s_0)\langle M_1\rangle_0
-\frac{\kappa}{2}\langle M_1\rangle_1~,\nonumber
\end{align}
and
\begin{align}
\label{200522.12}
3s_0&=m_\eta^2+3m_\pi^2~.
\end{align}
The function $\hat{M}_I(s)$ has no discontinuity across the RHC so that the
discontinuities of the PWAs $M^{(IJ)}(s)$ can be expressed as,
\begin{align}
\label{200522.13}
\Delta M_0(s)&=2i e^{-i\delta^{(00)}(s)}\sin\delta^{(00)}(s) \big[M_0(s)+\hat{M}_0(s)\big]~,\\
\Delta M_1(s)&=2i e^{-i\delta^{(11)}(s)}\sin\delta^{(11)}(s) \big[M_1(s)+\hat{M}_1(s)\big]~,\nonumber\\
\Delta M_2(s)&=2i e^{-i\delta^{(20)}(s)}\sin\delta^{(20)}(s) \big[M_2(s)+\hat{M}_2(s)\big]~.\nonumber
\end{align}
Following the same steps as above in Eq.~\eqref{200522.5} we can then also write that
\begin{align}
\label{200522.14}
M_I(s+i\epsilon)&=M_I(s-i\epsilon)e^{2i \delta_{IJ}}+2i\hat{M}_I(s)e^{i\delta^{(IJ)}}sin\delta^{(IJ)}~,
\end{align}
with $J=I$ except for $I=2$ for which $J=0$ (as it should be clear from the context in this section).
Dividing this expression by the corresponding Omn\`es function $\Omega^{(IJ)}(s)$, which fulfills that along
the RHC $\Omega^{(IJ)}(s+i\epsilon)=e^{i2\delta^{(IJ)}}\Omega^{(IJ)}(s-i\epsilon)$, $\Omega^{(IJ)}(s+i\epsilon)=|\Omega^{(IJ)}(s)|e^{i\delta^{(IJ)}}$,
we then obtain from Eq.~\eqref{200522.14} the
discontinuity of $M_I/\Omega^{(JI)}$ as
\begin{align}
\label{200522.15b}
\frac{M_I(s+i\epsilon)}{\Omega^{(IJ)}(s+i\epsilon)}-\frac{M_I(s-i\epsilon)}{\Omega^{(IJ)}(s-i\epsilon)}&
=2i\frac{\hat{M}_I(s)\sin\delta^{(IJ)}}{|\Omega^{(IJ)}(s)|}~.
\end{align}
The final step is to obtain IEs for $M_I(s)$ by writing down DRs for $M_I(s)/\Omega^{(IJ)}$ as
\begin{align}
\label{200522.15}
M_I(s)&=\Omega^{(IJ)}(s)\Big[
P_I^{(m)}(s)+\frac{s^n}{\pi}\int_{4m_\pi^2}^{\infty}ds'\frac{\hat{M}_I(s') \sin\delta^{(IJ)}(s')}{|\Omega^{(IJ)}(s')|(s')^n(s'-s)}\Big]~,
\end{align}
where $P_I^{(m)}(s)$ is a subtractive polynomial with $m\geq n-1$. Requiring that $A(s,t,u)$ diverges linearly at most at infinity in the
Mandelstam variables \cite{anisovich.190329.1},
then $M_1(s)$ should be bounded by a constant and $M_0(s)$, $M_2(s)$ would diverge linearly at most
in the limit $s\to\infty$. Furthermore, we also know the asymptotic behavior in the same limit for the Omn\`es functions,
cf. Eq.~\eqref{181122.3}, with $|\Omega^{(IJ)}(s)|\to s^{-\delta^{(IJ)}(\infty)/\pi}$. Depending on $\delta^{(IJ)}(\infty)$
the value of $m$ should be adjusted to the required asymptotic behavior of $M_I(s)$. For instance,
Ref.~\cite{alba.190329.1} assumes that $\delta^{(00)}(\infty)=\pi$, $\delta^{(11)}(\infty)=\pi$ and $\delta^{(20)}=0$, so that
$m=2$ for $I=0$ and $m=1$ for $I=1,$ and 2.
The DRs in Eq.~\eqref{200522.15} constitute a set of coupled linear IEs because the angular averages $\langle \hat{M}_I\rangle_n$
are also expressed in terms of the $M_I(s)$ functions. A standard way for solving these equations is by iteration.
The subtraction constants can be determined by matching with the NLO ChPT calculation of $A(s,t,u)$ and/or fitted to data,
as done in Refs.~\cite{anisovich.190329.1,alba.190329.1}. A clear improvement is obtained in the calculated decay width for the
$\eta\to\pi^+\pi^-\pi^0$ in Ref.~\cite{anisovich.190329.1}, where the value $\Gamma_{\eta\to\pi^+\pi}=283\pm 28$~eV was obtained.
Other improvements concern the parameter $\alpha$ for characterizing the amplitude for $\eta\to 3\pi^0$ in its Dalitz plot.
NLO ChPT gives a value $\alpha=0.0142$ while the KT treatment of Ref.~\cite{alba.190329.1}
gives $\alpha=-0.0337(12)$, to be compared with
the PDG average value of $\alpha=-0.0318(15)$.
\section{The $N/D$ method}
\setcounter{equation}{0}
\label{sec.200517.1}
In this section we elaborate on different aspects of the $N/D$ method, first introduced in Ref.~\cite{chew} to study uncoupled $\pi\pi$ PWAs. We first review on this method, discuss in more detail the limit
in which the crossed-channel dynamics is neglected \cite{nd}, and afterwards elaborate on how the latter can be treated
perturbatively within the $N/D$ method \cite{plb,ww}.
These results can also be used to take into account FSI in production
processes \cite{palomar,ozi}.
For the case of NR scattering, thanks to recent developments
\cite{oller.lhc.aop}, it is possible to know the
exact discontinuity of a PWA along the LHC for a given potential. In this way, one can
generate the same solutions as in the Lippmann-Schwinger (LS) equation, together with other ones
that cannot be obtained in a LS equation when mimicking the short-distance interactions
by contact terms in the potential \cite{oller.lhc.plb}.
\subsection{Scattering}
\label{sec.200517.2}
For the scattering of particles with equal masses there is only a
LHC for $s<s_{\rm Left}$ because of crossing.
However, when the particles involved have different masses there are also
other types of cuts in the complex $s$ plane due to crossing. For instance,
for the scattering of particles $a+b\rightarrow a+b$, in addition to a LHC there is also
a circular cut for $|s|=m_b^2-m_a^2$ \cite{martin.290916.1} where, for definiteness,
we have considered that $m_b>m_a$.
Nonetheless, when we refer in the following to the LHC we actually mean all the crossed-channel cuts.
Indeed, had we taken instead the complex $p^2$ plane all the cuts would be linear and only
a LHC would be present \cite{martin.290916.1}.
We introduce the $N/D$ method following Ref.~\cite{nd}.
The uncoupled case is discussed first and afterwards we move to coupled-channel scattering.
The discussion is restricted to two-body intermediate states.
The discontinuity of the inverse of a PWA $T_\ell(s)$ along the RHC is $2i$ times its imaginary part,
being the latter fixed by phase space because of unitarity, cf. Eq.~\eqref{200304.3b}.
In the $N/D$ method $T_\ell(s)$ is expressed as the quotient of two functions,
\begin{align}
\label{n/d}
T_\ell(s)=\frac{N_\ell(s)}{D_\ell(s)}~,
\end{align}
where $N_\ell(s)$ stands for the numerator function and $D_\ell(s)$ for the denominator one.
The former has only LHC and the later RHC.
To enforce the right kinematical threshold behavior of a PWA, vanishing as $p^{2\ell}$,
Ref.~\cite{nd} divides $T_\ell(s)$ by $p^{2\ell}$,
\begin{align}
\label{T'}
T'_\ell(s)=\frac{T_\ell(s)}{p^{2\ell}}~.
\end{align}
The $N/D $ method is then applied to this function,
\begin{align}
\label{n/d'}
T'_\ell(s)=\frac{N'_\ell(s)}{{D}'_\ell(s)}~.
\end{align}
It follows then that the discontinuities of ${N}'_\ell(s)$ and
${D}'_\ell(s)$ along the LHC and RHC, respectively, are
\begin{align}
\label{eqs1}
{\Im D}'_\ell &=\Im T'^{-1}_\ell \; {N}'_\ell=-\rho(s) {N}'_\ell p^{2\ell}~, &s>s_{\rm th}~, \\
{\Im D}'_\ell&=0~, &s<s_{\rm th}~,\nonumber\\
\label{eqs2}
{\Im N}'_\ell&=\Im T'_\ell \; {D}'_\ell=\Delta_\ell D'_\ell {p^{2\ell}}~, &s<s_{\rm Left} ~, \\
{\Im N}'_\ell&=0~, &s>s_{\rm Left}~,\nonumber
\end{align}
with $\Delta_\ell(s)=\Im T_\ell(s)$ along the LHC.
Let us discuss the DRs for ${D}'_\ell(s)$ and ${N}'_\ell(s)$ that result by taking into account
these discontinuities.
For $D'_\ell(s)$ one has,
\begin{align}
\label{d'}
{D}'_\ell(s)=\sum_{m=0}^{n-1}\overline a _m s^m-\frac{(s-s_0)^n}{\pi}\int^\infty_{s_{\rm th}} ds'
\frac{p(s')^{2\ell} \rho(s') {N}'_\ell(s')}{(s'-s)(s'-s_0)^n}~.
\end{align}
Here $n$ is, at least, the minimum number of subtractions required to guarantee the convergence of the
integral in the DR,
\begin{align}
\label{n}
\displaystyle \lim_{s \to \infty} \frac{{N}'_\ell(s)}{s^{n-\ell}}= 0~.
\end{align}
Consistently with Eq.~(\ref{n}), the DR for $N'_\ell(s)$ can be written as
\begin{align}
\label{n2}
{N}'_\ell(s)=\sum_{m=0}^{n-\ell-1} \overline b_m s^m+\frac{(s-s_0)^{n-\ell}}{\pi}\int_{-\infty}^{s_{\rm Left}} ds'
\frac{\Delta_\ell(s') {D}'_\ell(s')}{p(s')^{2\ell} (s'-s_0)^{n-\ell} (s'-s)}~.
\end{align}
The Eqs.~(\ref{d'}) and (\ref{n2}) are a system of coupled linear IEs whose input is $\Delta_\ell(s)$.
It is customary to substitute the expression for $N'_\ell(s)$ in $D'_\ell(s)$ and end with a linear IE for
$D'_\ell(s)$ along the LHC. Namely,
\begin{align}
\label{200519.5}
D'_\ell(s)&=\sum_{m=0}^{n-1}\overline{a}_m s^m
-\sum_{m=0}^{n-\ell-1}\overline b_m\frac{(s-s_0)^n}{\pi}\int_{s_{\rm th}}^\infty ds'\frac{p(s')^{2\ell}\rho(s'){s'}^m}{(s'-s)(s'-s_0)^n}\\
&+\frac{(s-s_0)^n}{\pi^2}\int_{-\infty}^{s_{\rm Left}}ds''\frac{\Delta_\ell(s'')D_\ell'(s'')}{(s''-s_0)^{n-\ell}p(s'')^{2\ell}}
\int_{s_{\rm th}}^\infty ds'\frac{p(s')^{2\ell}\rho(s')}{(s'-s)(s'-s'')(s'-s_0)^\ell }~,\nonumber
\end{align}
and the last integral can indeed be performed algebraically.
This is a linear IE for $D'_\ell(s)$ with $s$ along the LHC. Once this solved one can calculate
$D_\ell(s)$ for $s\in \mathbb{C}$ and, in particular, along the physical region, $s+i\epsilon$.
Other types of IEs could be deduced by taking more subtractions independently in $D_\ell(s)$ and $N_\ell(s)$.
Fore more details on this respect the reader can consult \cite{guo.rios.NN}.
The expression in Eq.~\eqref{200519.5} can be shortened and simplified for equal mass scattering
with mass $m$ by taking $s_0=4m^2$,
because then $p(s')^2=(s-4m^2)/4$. It follows that,
\begin{align}
\label{200519.6}
D'_\ell(s)&=\sum_{m=0}^{n-1}\overline{a}_m s^m
-\sum_{m=0}^{n-\ell-1}\overline b_m\frac{(s-s_0)^n}{4^\ell \pi}\int_{s_{\rm th}}^\infty ds'\frac{\rho(s'){s'}^m}{(s'-s)(s'-s_0)^{n-\ell}}\\
&+\frac{(s-s_0)^n}{\pi^2}\int_{-\infty}^{s_{\rm Left}}ds''\frac{\Delta_\ell(s'')D_\ell'(s'')}{(s''-s_0)^{n}}
\int_{s_{\rm th}}^\infty ds'\frac{\rho(s')}{(s'-s) (s'-s'')}~.\nonumber
\end{align}
The last integral in the previous expression can be written in terms of $g(s)$, Eq.~\eqref{200508.2}.
One of the subtraction constants can be fixed because we can freely choose the normalization of $D'_\ell(s)$, since
their ratio and analytical properties are invariant under a change in normalization. The standard choice is to take $D'_\ell(0)=1$.
However, given $\Delta'_\ell(s)$ along the LHC, the solution is not unique because of the addition of extra subtraction constants
in $D'_\ell(s)$ and $N'_\ell(s)$.
Historically, the possible addition of Castillejo-Dalitz-Dyson (CDD) poles \cite{Castillejo} was the clear
indication that extra solutions could be obtained even if $\Delta_\ell(s)$ is assumed to be known along the LHC.
They give rise to zeros of $T_\ell(s)$ along the RHC and each zero comprises two real parameters, its residue and
position. Phenomenologically the CDD poles
correspond to the short-distance dynamics underneath the scattering process and
might also be related to the addition of bare states \cite{dyson}.
Let us notice that $T_\ell(s)^{-1}$ does not exist at a zero of $T_\ell(s)$ and, therefore, Eq.~\eqref{200304.3b}
is not defined there. As in Ref.~\cite{Castillejo} let us introduce the auxiliary function $\lambda(s)$ such that
\begin{align}
\label{dl1}
{\Im D}'_\ell(s)=\frac{d \lambda (s)}{ds}~,
\end{align}
and rewrite Eq.~(\ref{eqs1}) as,
\begin{eqnarray}
\label{dl2}
{\displaystyle \frac{d\lambda}{ds}}&=-\rho(s) p^{2\ell}{N}'_\ell~, & s>s_{\rm th}~, \\
{\displaystyle \frac{d\lambda}{ds}}&=0~.~~~~~~~~~~~~~~ & s<s_{\rm th}~. \nonumber
\end{eqnarray}
Denoting by $s_i$ the zeros of $T_\ell(s)$ along the real axis above threshold,
we can write $\lambda(s)$ from Eq.~\eqref{dl2} as
\begin{align}
\label{l2}
\lambda(s)=-\int_{s_{\rm th}}^s p(s')^{2\ell} \rho(s') {N}'_\ell(s') ds'+
\sum_i \lambda(s_i)\theta(s-s_i)~,
\end{align}
where the $\lambda(s_i)$ are a priori unknown.
Thus, Eqs.~(\ref{dl1}) and (\ref{l2}) allow us to write
\begin{align}
\label{tower1}
{D}'_\ell(s) &= \sum_{m=0}^{n-1} \overline a_m s^m +
\frac{(s-s_0)^n}{\pi}\int_{s_{\rm th}}^\infty \frac{
{\Im D}'_\ell(s') ds'}{(s'-s)(s'-s_0)^n} \\
& =\sum_{m=0}^{n-1} \overline a_m
s^m -\frac{(s-s_0)^n}{\pi}\int_{s_{\rm th}}^\infty \frac{p(s')^{2\ell} \rho(s')
{N}'_\ell(s')}{(s'-s)(s'-s_0)^n}ds' +
\frac{(s-s_0)^n}{\pi} \int_{s_{\rm th}}^\infty \frac{\sum_i \lambda(s_i)
\delta(s'-s_i)}{(s'-s)(s'-s_0)^n}ds'\nonumber\\
&=\sum_{m=0}^{n-1} \overline a_m s^m
-\frac{(s-s_0)^n}{\pi}\int_{s_{\rm th}}^\infty \frac{p(s')^{2\ell} \rho(s')
{N}'_\ell(s')}{(s'-s)(s'-s_0)^n}ds'+\sum_i \frac{\lambda(s_i)}{\pi(s_i-s) }\frac{(s-s_0)^n}{(s_i-s_0)^n}~.\nonumber
\end{align}
The last term in the previous equation can be rewritten as
\begin{align}
\label{cano}
\frac{(s-s_0)^n}{s-s_i}&=\sum_{i=0}^{n-1}(s-s_0)^{n-1-i}(s_i-s_0)^i+\frac{(s_i-s_0)^n}{s-s_i}~.
\end{align}
The contribution $\displaystyle{\sum_{i=0}^{n-1}(s-s_0)^{n-1-i}(s_i-s_0)^i}$ can be reabsorbed in
$\displaystyle{\sum_{m=0}^{n-1} \overline a _m s^m}$ and Eq.~\eqref{tower1} can be rewritten as
\begin{align}
\label{d'2}
{D}'_\ell(s)&=\sum_{m=0}^{n-1} \widetilde{a}_m s^m +
\sum_i \frac{\widetilde{\gamma}_i}{s-s_i}
-\frac{(s-s_0)^n}{\pi}\int_{s_{\rm th}}^\infty
\frac{p(s')^{2\ell} \rho(s') {N}'_\ell(s')}{(s'-s)(s'-s_0)^n}ds' ~,
\end{align}
where $\widetilde{a}_m$, $\widetilde{\gamma}_i$ and $s_i$ are constants not fixed by the knowledge of $\Delta_\ell(s)$,
and the CDD poles give rise to the last term.
Interesting results can be deduced under the approximation of neglecting the LHC,
$\Delta_\ell(s)\to 0$. Eq.~(\ref{n2}) then becomes
\begin{align}
\label{naprox}
{N}'_\ell(s)=\sum_{m=0}^{n-\ell-1} \overline{b}_m s^m=\overline b_{n-\ell-1} \prod_{j=1}^{n-\ell-1}(s-s_j)~,
\end{align}
and $N'_\ell(s)$ is just a polynomial, which can be reabsorbed in $D'_\ell(s)$ by dividing simultaneously both functions by
$N'_\ell(s)$ itself. The expression for $T'_\ell(s)$ then becomes
\begin{eqnarray}
\label{fin/d}
{T}'_\ell(s)&=&\frac{1}{{D}'_\ell(s)}~,\\
{N}'_\ell(s)&=&1~,\nonumber\\
{D}'_\ell(s)&=&-\frac{(s-s_0)^{L+1}}{\pi}\int_{s_{\rm th}}^\infty
\frac{p(s')^{2\ell} \rho (s')}{(s'-s)(s'-s_0)^{L+1}}ds'+\sum_{m=0}^L a_m s^m+
\sum_i^{M_\ell} \frac{R_i}{s-s_i}~.\nonumber
\end{eqnarray}
The number of real free parameters in the previous equation is $\ell+1+2M_\ell$, with $M_\ell$ the number of
CDD poles. A priori there is nothing to prevent the generalization of Eq.~\eqref{fin/d} such that
some $s_i$ could also lie below threshold.
We could adjust the position and residue of a CDD pole such that the real part of
$D'_\ell(s)$ vanishes at the desired position.
This would give rise to typical resonance behavior above threshold, or to a bound-state pole if
this happens below threshold.
This is why the parameters of the CDD poles are typically associated with the coupling constants and masses
of the poles in the $S$ matrix.
In other instances, the CDD poles are needed because the presence of a zero cannot be related
to $\Delta_\ell(s)$, but they respond to fundamental constraints in the theory.
This is the case of the Adler zeroes in QCD \cite{adler.181115.1}, which already
occur at LO in the chiral expansion,
while $\Delta_\ell(s)\neq 0$ only at NLO and higher orders.
It is therefore necessary to account for them by including CDD poles,
such that the derivative of the PWA at the zero corresponds to
the inverse of the residue of the CDD pole, $R_i$.
For the $\pi\pi$ Adler zeroes the latter could be fixed in good approximation by the LO ChPT result.
The other $\ell+1$ parameters emerge by having enforced the correct behavior of a PWA near threshold,
which should vanish as $p^{2\ell}$.
Let us stress that Eq.~(\ref{fin/d}) gives the general form of an elastic PWA
when the LHC contributions are neglected.
Phenomenologically this assumption could be suited if the LHC is far away and/or if
it is suppressed for some reason \cite{X3872}.
The free parameters in Eq.~\eqref{fin/d} can be fixed by fitting experimental data
and/or by reproducing the Lattice QCD (LQCD) results at finite
volume or when varying some of the QCD parameters, like $N_c$ or the quark masses
\cite{guo.190126.1,guo.190126.3,albaladejo.190126.1,roca.old}.
Ref.~\cite{nd} focuses on meson-meson scattering, whose basic theory is QCD.
It studied the $S$- and $P$-wave two-body scattering between the lightest pseudoscalars ($\pi$, $K$ and $\eta$),
as well as the related spectroscopy.
It was found that the full nonet of scalar resonances \cite{oller.mix} $f_0(500)$, $f_0(980)$, $a_0(980)$ and $K^*_0(800)$ arose
from the self-interactions among the lightest pseudoscalars, while the more massive
resonances $f_0(1370)$, $f_0(1500)$, $a_0(1450)$ and $K^*_0(1430)$ stem from a nonet of bare resonances
with a mass around 1.4~GeV.
In addition, Ref.~\cite{nd} included a bare scalar singlet with a mass around 1~GeV which
gives also a contribution to the $f_0(980)$ \cite{guo.nc.traj}.
Later on, Ref.~\cite{alba.oller.glue} extended this model by including more channels
and could determine a glueball state
affecting mainly the $f_0(1700)$ with a reflection (because of the $\eta\eta'$ threshold)
on the $f_0(1500)$ as well.
Of course, the same Eq.~\eqref{fin/d} can be applied to other interactions,
e.g. Ref.~\cite{ww} studied $W_LW_L$ scattering in the electroweak symmetry breaking sector.
The generalization of Eq.~\eqref{fin/d} to coupled channels is rather straightforward by employing a
matrix notation, where
the $T$ matrix in coupled channels is a matrix denoted by $T_L(s)$.
As in Eq.~\eqref{fin/d} we take from the onset that crossed-channel dynamics can be
neglected in a first approximation.
Thus, the matrix element ${T}_{L,ij}(s)$ is proportional to $p_i^{\ell_i} p_j^{\ell_j}$,
which gives rise for odd orbital angular momentum (unless $i=j$)
to another cut between $s_{{\rm th}; i}$ and $s_{{\rm th}; j}$
due to the square roots in the expressions of $p_i$ and $p_j$ as a
function of $s$.
To avoid this cut we define the matrix $T'_L$, analogously to Eq.~\eqref{T'}, as
\begin{align}
\label{A.1}
{T}'_L(s)=p^{-L} {T}_L(s) p^{-L}~.
\end{align}
In this equation, the symbol $p^L$ corresponds to a diagonal matrix with matrix elements
\begin{align}
p^L_{ij}&=p_i^{\ell_i} \delta_{ij}~,\\
p_i&=\frac{\lambda^{1/2}(s,m_{1i}^2,m_{2i}^2)}{2\sqrt{s}}~,\nonumber
\end{align}
and $m_{1i}$ and $m_{2i}$ are the masses of the two particles in the same channel $i$.
The matrix unitarity relation along the RHC then reads
\begin{align}
\label{A.2}
{\Im T}'^{ -1}_L(s)=-p^L \rho(s) p^L=-\rho(s) p^{2L}~,
\end{align}
where $\rho(s)$ is another diagonal matrix whose elements are $\rho_i(s)$.
The next step proceeds with the generalization to coupled channel of Eq.~\eqref{n/d} by writing ${T}'_L$ as
\begin{align}
\label{A.4}
{T}'_L={D}'^{ -1}_L {N}'_L~,
\end{align}
with ${N}'_L$ and ${D}'_L$ two matrices, the former only involves LHC and the later RHC, respectively.
In our present case without LHC, the matrix elements of $N'_L$ are polynomials functions. Multiplying
$N'_L$ and $D'_L$ in Eq.~\eqref{A.4} to the left by ${N'_L}^{-1}$ we can always make that ${N}'_L=I$ and write,
\begin{eqnarray}
\label{A.8}
{T}'_L&=&\widetilde{{D}}'^{-1}_L~, \\
\widetilde{{N}}'_L&=&I~, \nonumber\\
\widetilde{{D}}_L'&=&-\frac{(s-s_0)^{L+1}}{\pi}\int_0^\infty ds'
\frac{\rho(s')p^{2L}(s')}{(s'-s)(s'-s_0)^{L+1}}+{R}(s)~,\nonumber
\end{eqnarray}
with ${R}(s)$ a matrix of rational functions which poles produce the CDD poles
in $\widetilde{{D}}_L'$.
Let us notice that all the zeros in ${\rm det}{T}'_L$ correspond to CDD poles in the ${\rm det}\widetilde{{D}}_L'(s)$.
This is the generalization of the CDD poles for the coupled-channel case.
The resulting expression for $T_L(s)$ in Eq.~\eqref{A.8} can be also recast as
\begin{align}
\label{200518.1}
T_L(s)&=\left[V_L^{-1}+g(s)\right]^{-1}~,
\end{align}
with $g(s)$ the diagonal matrix with matrix elements $g_i(s)$ defined as
\begin{align}
\label{200518.2}
g_i(s)&=a_i(s_0)-\frac{s-s_0}{\pi}\int_{s_{{\rm th};i}}^\infty \frac{\rho_i(s')ds'}{(s'-s_0)(s'-s)}~,
\end{align}
where $a_i(s_0)$ is a subtraction constant and $s_0$ the subtraction point.
The result of this integration can also be written as
\begin{align}
\label{200518.3}
g_i(s
&=\frac{1}{16\pi^2}\left[
a_i(\mu)+\log\frac{m_{1i}^2}{\mu^2}-x_+\log\frac{x_+-1}{x_+}-x_-\log\frac{x_--1}{x_-}\right]~,\nonumber\\
x_\pm&=\frac{s+m_{2i}^2-m_{1i}^2}{2s}\pm\frac{1}{2s}\sqrt{(s+m_{2i}^2-m_{1i}^2)^2-4s(m_{2i}^2-i0^+)}~.
\end{align}
The parameter $\mu$ is a renormalization scale, such that a change in the value of $\mu$ can always be
reabsorbed in a corresponding variation of $a_i(\mu)$, while the combination $a_i(\mu)-2\log\mu$ is independent of $\mu$.
The unitarity loop function $g_i(s)$ corresponds to the one-loop two-point function
\begin{align}
\label{181106.5}
g_i(s)&=i\int\frac{d^4p}{(2\pi)^4}\frac{1}{[(P/2-p)^2-m_{1i}^2+i\varepsilon][(P/2+p)^2-m_{2i}^2+i\varepsilon]}\\
&=\int_0^\infty \frac{p^2 dp}{(2\pi)^2}\frac{\omega_1+\omega_2}{\omega_1\omega_2[s-(\omega_1+\omega_2)^2+i\varepsilon]}~,\nonumber
\end{align}
where $\omega_j=\sqrt{m_{ji}^2+{\mathbf{p}}^2}$ and the total four-momentum $p_1+p_2$ is indicated by $P$.
The integral in Eq.~\eqref{181106.5} diverges logarithmically, which is the reason why a subtraction
has been taken in Eq.~\eqref{200518.2}.
The Eq.~\eqref{200518.3} also results by employing dimensional regularization
and reabsorbing the diverging term in $a_i(\mu)$.
Let us elaborate on the so-called natural value for the subtraction constants.
The function $g_i(s)$ given by Eq.~\eqref{200518.3} has the value at threshold,
\begin{align}
\label{200518.4}
g_i(s_{\rm th})&=\frac{a_i(\mu)}{16\pi^2}+\frac{1}{8\pi^2(m_{1i}+m_{2i})}(m_{1i}\log\frac{m_{1i}}{\mu}+m_{2i}\log\frac{m_{2i}}{\mu})~.
\end{align}
This expression is compared with the one that results by evaluating $g_i(s)$ in terms of a three-momentum cutoff $\Lambda$.
The resulting expression for the function $g_i(s)$, and denoted by $g_{\Lambda i}(s)$,
can be found in Ref.~\cite{iam.oller.long}.
The natural size of a three-momentum cutoff in hadron physics is the inverse of the typical size of a
compact hadron, which is generated by the strong dynamics binding quarks and gluons.
Thus, according to this estimate we take $\Lambda \simeq 1$~GeV.
For NR scattering $g_i(s)$ and $g_{\Lambda i}(s)$ ($m_{1i},\,m_{2i} \gg |{\mathbf{p}}|$)
are given by the value at threshold of every function plus
$-ip/(8\pi(m_1+m_2))+{\cal O}({\mathbf{p}}^2)$ \cite{guo.rios.NN}.
The value at threshold of $g_{\Lambda i}(s_{\rm th})$ can be worked out
explicitly with the result \cite{guo.190126.1}
\begin{align}
\label{200518.5}
g_{\Lambda i}(s_{\rm th})&=-\frac{1}{8\pi^2(m_{1i}+m_{2i})}\left[
m_{1i}\log\left(1+\sqrt{1+m_{1i}^2/\Lambda^2}\right)\right.\\
&\left. +m_{2i}\log\left(1+\sqrt{1+m_{2i}^2/\Lambda^2}\right)
-m_{1i}\log\frac{m_{1i}}{\Lambda}-m_{2i}\log\frac{m_{2i}}{\Lambda}
\right]~.\nonumber
\end{align}
By equating Eqs.~\eqref{200518.4} and \eqref{200518.5} the following matching value for $a_i(\mu)$ results,
\begin{align}
\label{181106.8}
a_i(\mu)&=-\frac{2}{m_{1i}+m_{2i}}\left[m_{1i}\log\left(1+\sqrt{1+m_{1i}^2/\Lambda^2}\right)
+m_{2i}\log\left(1+\sqrt{1+m_{2i}^2/\Lambda^2}\right) \right]+\log\frac{\mu^2}{\Lambda^2}~.
\end{align}
One should employ $\mu\simeq \Lambda\simeq 1$~GeV in Eq.~\eqref{181106.8} to estimate the natural value for
the subtraction constants, a procedure originally established in Ref.~\cite{plb}.
In this way, both the renormalization scale $\mu$ and the cut off $\Lambda$ are used with values
suitable to the transition from the low-energy EFT
to the shorter-range QCD degrees of freedom.
As an example, let us take $\pi\pi$ scattering and $\Lambda=1$~GeV. Then, from Eq.~\eqref{181106.8}
\begin{align}
\label{181106.9}
a(\mu)&=-1.40+\log\frac{\mu^2}{\Lambda^2}~,~\Lambda=1~{\rm GeV}~.
\end{align}
The Eq.~\eqref{200518.1} is adequate for including perturbatively the LHC contributions in the $T$-matrix $T_L$.
This can be achieved by matching order by order with a calculation within an EFT. For instance, this has been used
many times taking as input one-loop calculations in ChPT
\cite{ww,plb,jamin,guo.181123.1,alba.oller.sigma,guo.prc,guo.190126.1,guo.190126.3,talks,kangP}.
The procedure is as follows.
Let us take a meson-meson scattering amplitude calculated in ChPT up to one-loop or ${\cal O}(p^4)$,
$T_L= T_2+T_4+{\cal O}(p^6)$. Then the chiral expansion of Eq.~\eqref{200518.1}, with $V=V_2+V_4+{\cal O}(p^6)$, $g={\cal O}(p^0)$ \cite{ww},
reads at LO,
\begin{align}
\label{200519.1}
T_2&=V_2+{\cal O}(p^4)~,
\end{align}
and at NLO,
\begin{align}
\label{200519.2}
T_4=V_4-V_2 g V_2+{\cal O}(p^6)~,
\end{align}
and similarly for higher orders.
Thus, up to NLO the matching equations fix $V_2$ and $V_4$ to
\begin{align}
\label{200519.3}
V_2&=T_2~,\\
V_4&=T_4+V_2 gV_2~.\nonumber
\end{align}
The LHC contributions arise because crossed-channel loops are calculated order by order in the ChPT
results for $T_L$.
At NLO in the calculation of $V_L$ we then have the expression
\begin{align}
\label{200613.1}
T_L(s)&=\left[(T_2+T_4+T_2 g T_2)^{-1}+g\right]^{-1}~.
\end{align}
If $(T_2+T_4+T_2 g T_2)^{-1}$ is further expanded we then recover the IAM result of Eq.~\eqref{200515.5} because
\begin{align}
\label{200613.2}
(T_2+T_4+T_2 g T_2)^{-1}+g=T_2^{-1}-T_2^{-1}T_4T_2^{-1}+{\cal O}(p^2)~,
\end{align}
so that
\begin{align}
\label{200613.3}
T_L(s)\to \left[T_2^{-1}-T_2^{-1}T_4T_2^{-1}\right]^{-1}=T_2^{-1}\left[T_2-T_4\right]^{-1} T_2~.
\end{align}
This is the formula for the IAM in coupled channels at NLO \cite{iam.oller.long,oop.prl}.
\begin{figure}[H]
\begin{center}
\includegraphics[width=0.85\textwidth]{erratumNPA.pdf}
\caption{{\small Results from Ref.~\cite{npa} with only one free parameter
for the $S$-wave meson-meson scattering with $I=0$ and 1.
From top to bottom and left to right, the isoscalar scalar $\pi\pi\to \pi\pi$ and $K\bar{K}\to \pi\pi$
phase shifts, the $\pi\pi$ inelastic cross-section with the same quantum numbers
and a $\pi^0\eta$ event distribution around the isovector scalar
$a_0(980)$ resonance are plotted.
For more details and references of the experimental papers we refer to Ref.~\cite{npa}.}
\label{fig.200524.1}}
\end{center}
\end{figure}
In order to appreciate the power of the method for some reactions we consider the LO matching, that is,
with $V=V_2$, applied in Ref.~\cite{npa} to study the meson-meson $S$-waves with $I=0$ and 1.
This is a coupled-channel study with $\pi\pi$ and $K\bar{K}$ for $I=0$ and $\pi\eta$ and $K\bar{K}$ for $I=1$.
It is certainly remarkable that only one free parameter entered in the successful calculation of the PWAs
from the $\pi\pi$ threshold up to around 1.2~GeV. This is shown in Fig.~\ref{fig.200524.1}
by the $\pi\pi\to \pi\pi$, $K\bar{K}\to\pi\pi$ phase shifts,
the inelastic $\pi\pi$ reaction and a $\pi^0\eta$ event distribution around the $a_0(980)$, from
top to bottom and left to right, respectively.
The resonances $f_0(500)$, $f_0(980)$ and $a_0(980)$,
clearly visible in Fig.~\ref{fig.200524.1}, are generated dynamically from the
interactions between the pseudoscalars. The free parameter is the three-momentum cut-off
with natural size $\Lambda\simeq 1~$GeV used in the evaluation of the unitarity-loop functions
$g_{\Lambda i}(s)$ employed in this study.
It is also the case in some instances \cite{nd,jamin,guo.181123.1,guo.190126.1,guo.190126.3,alba.oller.glue}
that the ChPT expansion is complemented with the exchange of bare resonance fields, so that the tree-level amplitude
is crossing symmetric.
One typically improves the convergence properties of the chiral expansion by including bare resonance fields
because of the (partial) saturation of the chiral counterterms
by the resonance exchanges \cite{EP}. Then,
the matching process is undertaken up to ${\cal O}(\hbar p^4,\hbar^2)$, which
means to neglect any two-loop contribution and any one-loop contribution beyond ${\cal O}(p^4)$.
In this way, one could consider one-loop contributions involving higher orders because of the explicit inclusion of the
resonance fields.
The matching proceeds as in Eqs.~\eqref{200519.1} and \eqref{200519.2}, with the difference that now
the LO amplitudes
include also the tree-level exchange of resonances
and $T_4$ involves the one-loop contributions up to ${\cal O}(p^4)$.
Then,
the Eq.~\eqref{200519.3} still holds and one has again Eq.~\eqref{200518.1} for $T_L$.
This equation is in appearance analogous to the $N/D$-method form of Eq.~\eqref{A.4}.
Indeed, if we identify $N_L$ with $V_L$ and $D_L$ with $I+V_L g(s)$, it
can be shown \cite{oller.review.old} that up to ${\cal O}(\hbar p^4,\hbar^2)$ the resulting functions satisfy
the $N/D$-method equations, cf.~Eq.~\eqref{eqs1}.
The perturbative solution of the $N/D$ equations with respect to the LHC contributions can also be
organized as an iterative solution in increasing number of insertions of $\Delta_\ell$. The first-iterated $N/D$ method
consists on taking only one power of $\Delta_\ell$ in the integrand of the DRs for $D_\ell(s)$ and $N_\ell(s)$.
The approximation is obtained by settling $D'_\ell(s)=1$ into the integrand for the DR of $N'_\ell(s)$, Eq.~\eqref{n2},
which is then denoted as $N'_{\ell;1st}$. Then,
\begin{align}
\label{n2b}
{N}'_{\ell;1st}(s)=\sum_{m=0}^{n-\ell-1} \overline b_m s^m+\frac{(s-s_0)^{n-\ell}}{\pi}\int_{-\infty}^{s_{\rm Left}} ds'
\frac{\Delta_\ell(s') }{p(s')^{2\ell} (s'-s_0)^{n-\ell} (s'-s)}~.
\end{align}
Since $\Delta_\ell(s)$ is known the DR integral could in principle be calculated. This is usually a tree-level amplitude
that can also be calculated in QFT, from which indeed $\Delta_\ell(s)$ is actually derived.
Therefore, we assume that in the first iterated $N/D$ method $N'_{\ell;1st}$ is also given.
As a result, the calculation of $D'_\ell(s)$ in this approximation, denoted by $D'_{\ell:1st}(s)$, just reduces
to perform the integration
\begin{align}
\label{d'b}
{D}'_\ell(s)=\sum_{m=0}^{n-1}\overline a _m s^m-\frac{(s-s_0)^n}{\pi}\int^\infty_{s_{\rm th}} ds'
\frac{p(s')^{2L} \rho(s') {N}'_{\ell;1st}(s')}{(s'-s)(s'-s_0)^n}~.
\end{align}
The first-iterated $N/D$ method was used in Ref.~\cite{igi} to discuss $\pi\pi$ scattering within linear realizations
of chiral symmetry, taking into account the exchanges of a $\sigma$ and $\rho$ resonances.
More recently, it has been employed to study $\rho\rho$ scattering in Ref.~\cite{gulmez.rr}
by taking the pure gauge-boson part
of the non-linear chiral Lagrangian with hidden-local symmetry \cite{38gulmez,39gulmez}. Its generalization to
the $SU(3)$-related vector-vector scattering was undertaken in Ref.~\cite{gulmez.guo}.
These studies were motivated by the earlier ones in Refs.~\cite{rr.review}, with still an on-going
productive discussion in interpreting the results.
\subsection{FSI}
\label{sec.200517.3}
Let us consider the unitarity relation for a form factor, Eq.~\eqref{200220.5},
with the expression of the $T$ matrix in PWAs $T_L$ as given in Eq.~\eqref{200518.1}.
It then results that
\begin{align}
\label{181119.2}
F(s)&=\left(V_L^{-1}+g\right)^{-1}\left(V_L^{-1}+g+2i\rho(s)\theta \right)F^*~.
\end{align}
Since $\Im g(s)=-\rho(s)$ it is clear that $g(s)+2i\rho(s)\theta(s)=g(s)^*$,
so that from Eq.~\eqref{181119.2} we have that along the RHC it is fulfilled that
\begin{align}
\label{181119.3}
\left(V_L^{-1}+g\right)F&=\left(V_L^{-1}+g^* \right)F^*~.
\end{align}
The cancellation of $V_L$ from both sides leads to
\begin{align}
\label{181119.4}
\left[I+V_L(s) g(s)\right]F(s)&=\left[I+V_L(s) g(s)^* \right]F(s)^*~.
\end{align}
From this equation it is clear that the combination
\begin{align}
\label{181119.5}
\left[I+V_L g(s)\right]F(s)
\end{align}
has no RHC \cite{basdevant.181119.1}. Then $F(s)$ can be expressed as
\begin{align}
\label{181119.6}
F(s)&= \left[I+V_L(s) g(s)\right]^{-1}L(s)~,
\end{align}
with $L(s)$ a column vector of $n$ functions without RHC, being $n$ the number of PWAs.
An analogous relation can be obtained if we write $T_L(s)$ as in the $N/D$ method in coupled channels,
$T_L(s)=D_L^{-1}(s)N_L(s)$. Following the same steps as in Eqs.~\eqref{181119.2}--\eqref{181119.6},
taking into account that $\Im D(s)=- N(s)\rho$, one ends with the relations
\begin{align}
\label{181120.1b}
D(s)F(s)&=D(s)^*F(s)^*~, \\
F(s)&=D(s)^{-1}L(s)~,\nonumber
\end{align}
and $L(s)$ is free of RHC. We can then write $F(s)$ as the product of
two matrices, the inverse of $D_L(s)$, which only has RHC, and $L(s)$, which could have LHC.
As a result, Eq.~\eqref{181120.1b} is the generalization of the $N/D$ method to production processes.
Coming back to Eq.~\eqref{181119.6}, let us remark that $I+V_L(s) g(s)$
could have the two types of cuts (since $V_L(s)$ in general has LHC).
For instance, for the case of the pion form factor if this is expressed as in Eq.~\eqref{181120.1b}
then $L(s)$ has no LHC, while if expressed as in Eq.\eqref{181119.6}
it would typically have one, if $V_L(s)$ has it.
However, for the relevant case for phenomenological applications in which $V_L(s)$
is driven by the $s$-channel dynamics and it
does not comprise explicit LHC, then $L(s)$ has no either LHC.
In this case, the matrix $D_L(s)$ and $I+V_l(s) g$ can be identified.
This formalism has been employed by Ref.~\cite{oller.gama}
to study the $\gamma\gamma\to$meson-meson fusion reactions.
Refs.~\cite{ozi,oller.D} used it to study the scalar form factor of the pion
(and of other pseudoscalar mesons) in connection with $J/\psi$ and $D$ decays,
and Ref.~\cite{palomar} analyzed the vector form factor of the pion.
This formalism was also very important to unveil the two-pole structure of
the $\Lambda(1405)$ in Ref.~\cite{plb}, because
in previous studies the $\pi\Sigma$ event distribution
for this resonance was always taken to be proportional to the
modulus squared of the $\pi\Sigma\to\pi\Sigma$ $I=0$ $S$-wave.
\subsection{The exact $N/D$ method in NR scattering}
\label{sec.200517.4}
For non-relativistic scattering one can calculate for a given potential the exact discontinuity of a PWA along
the LHC. This has been a recent advance in $S$-matrix theory achieved by Ref.~\cite{oller.lhc.aop}, to which
we refer the reader for further details.
The key point was to extrapolate analytically the LS equation to complex three-momenta for off-shell
scattering. The solution of the LS equation for half-off-shell scattering is an analytical function in the
off-shell three-momentum complex $q$ plane with vertical cuts which extend along the lines
$(\pm) p\pm i\lambda$, with $|\lambda|\geq \mu_0$. Here the $\pm$ symbols are unrelated,
$\mu_0$ is the lightest particle exchanged, and
$p$ is the on-shell three-momentum (fixed by the energy $E$ of the process,
$E=p^2/2\mu$, with $\mu$ the reduced mass).
E.g. for $NN$ scattering the lightest particle exchange is the pion and $\mu_0=m_\pi$.
We denote in the following a PWA for half-off-shell scattering as $T_\ell(q,p)$, where $q$ is the off-shell
three-momentum and $p$ the on-shell one.
The discontinuity we are interested in, e.g. for its later application to the $N/D$ method, is
\begin{align}
\label{200520.1}
\Delta_\ell(p^2)&=\frac{1}{2i}\left[T_\ell(p+i\epsilon,p+i\epsilon)-T(p-i\epsilon,p-i\epsilon)\right]=\Im T_\ell(p+i\epsilon,p+i\epsilon)~.
\end{align}
After some mathematical derivations that can be consulted in Ref.~\cite{oller.lhc.aop}, this discontinuity can be
obtained by solving an ordinary linear IE.
This IE is written in terms of the discontinuity of the potential in momentum space $v_\ell(q,p)$.
Its writing gets simplified by using $\hat{v}_\ell$ defined by
\begin{align}
\label{200520.2}
\hat{v}_\ell(q',q)&={q'}^{\ell+1}v_\ell(q,q)q^{\ell+1}~.
\end{align}
The discontinuity of the potential entering into the IE is
\begin{align}
\label{200520.3}
\Delta \hat{v}_\ell(\nu,\nu_1)&=
\Im \hat{v}_\ell(i\nu+\epsilon^-,i\nu_1+\epsilon)
-\Im \hat{v}_\ell(i\nu+\epsilon^+,i\nu_1+\epsilon)~,
\end{align}
with $\epsilon^-<\epsilon<\epsilon^+$ and $\epsilon^+\to 0$ at the end of the calculation.
After this preamble, the sought IE is ($p=ik$, $k\geq \mu_0$ and $n=2\ell+2$)
\begin{align}
\label{200520.4}
f(\nu)&=\Delta \hat{v}_\ell(\nu,k)+\frac{\theta(p-2\mu_0-\nu)\mu}{2\pi^2}\int_{\mu_0+\nu}^{k-\mu_0}\frac{d\nu_1\nu_1^2}{k^2-\nu_1^2}
\left\{\frac{1}{(i\nu_1+0^+)^n}+\frac{1}{(i\nu_1-0^+)^n}\right\}
\Delta\hat{v}_\ell(\nu,\nu_1)f(\nu_1)~.
\end{align}
In terms of $f(\nu)$ the discontinuity $\Delta_\ell(p^2)$ is given by
\begin{align}
\label{200520.5}
\Delta_\ell(p^2)&=(-1)^\ell \frac{f(-k)}{2k^{2\ell+2}}~.
\end{align}
Thus, we need to solve the IE for $\nu\in[-k+\mu_0,k-\mu_0]$, and the range of the integration in
the IE for $f(\nu)$ is finite for a given $p$, contrary to the LS equation.
This IE can be solved without ambiguity because $\Delta\hat{v}(\nu,\nu_1)$ can be
determined for a given potential and with it $f(\nu)$ by solving Eq.~\eqref{200520.4}.
For a general potential it is convenient to employ its spectral decomposition,
\begin{align}
v({\mathbf{q}},{\mathbf{p}})&=\int_{\mu_0}^\infty d\bar{\mu}^2\frac{\eta(\bar \mu^2)}{({\mathbf{q}}-{\mathbf{p}})^2+\bar{\mu}^2}+\ldots
\end{align}
where $\eta(\bar{\mu}^2)$ is the spectral function, and the ellipsis indicates possible subtractions
that due to its polynomial nature do not give contribution to the discontinuity of the potential.
In terms of the spectral decomposition we can write that
\begin{align}
\label{200520.7}
\Delta \hat{v}_\ell(\nu,\nu_1)&=-\frac{2}{\pi}\int_{\mu_0}^\infty d\bar{\mu}^2 \eta(\bar \mu^2)\rho(\nu^2,\nu_1^2;\bar{\mu}^2)
\theta(\nu_1-\nu-\bar{\mu})~.
\end{align}
The function $\rho(\nu^2,\nu^2;\bar{\mu}^2)$ is a polynomial in its argument and its fixed by the
partial-wave projection involved in the case of interest.
For brevity in the presentation offered here we have just referred to the uncoupled case, but the formalism
can also be generalized easily to evaluate the LHC discontinuity for coupled PWAs \cite{oller.lhc.aop}.
A potential is said to be singular if for $r\to 0$ it diverges stronger than $1/r^2$ or as
$\alpha/r^2$ for $\alpha+\ell(\ell+1)<-1/4$. In the opposite case the potential is said to be
regular \cite{oller.lhc.aop}. In the ChPT calculation of nuclear potentials the increase in the order of the calculation
implies typically an increase in the degree of divergence of the potential for $r\to 0$, because
off-shell momentum factors give rise to spatial derivatives. This fact is the main reason why
the original Weinberg's program for solving nuclear properties once the chiral potentials are
calculated order by order has not been taken to full completion.
The resulting $\Delta_\ell(p^2)$ obtained by solving the master Eq.~\eqref{200520.4} has a different qualitative
behavior depending on whether the potential is regular, attractive singular or repulsive singular.
General arguments, based on the scaling properties of the function $\rho(\nu^2,\nu_1^2;\bar{\mu}^2)$,
were given in Ref.~\cite{oller.lhc.aop} to explain such differences in the behavior of $\Delta_\ell(p^2)$.
Explicit examples were also worked out in Ref.~\cite{oller.lhc.aop} corresponding to actual PWAs in
$NN$ scattering, with the chiral potential calculated at different chiral orders, from LO up to NNLO.
The function $\rho(\nu^2,\nu_1^2;\bar{\mu}^2)$ is a polynomial in $\nu$ and $\nu_1$ of degree $m$.
Then, the argument of Ref.~\cite{oller.lhc.aop} follows by considering a re-scaling
by a parameter $\tau$ of the variables $k$, $\nu$ and $\nu_1$ in the limit $k\gg\mu_0$.
It follows from Eq.~\eqref{200520.4} that the $n_{th}$ iterated solution for $f(\nu)$
is subject to a re-scaling by
\begin{align}
\label{200520.8}
\tau^{(n+1)m-(2\ell+1)n}=\tau^{(m-2\ell-1)n+m}~.
\end{align}
The point is whether $m-2\ell-1$ is smaller or larger than zero. In the former case we have the
behavior corresponding to a regular potential, so that each extra iteration implies at least
an extra factor of $1/k$ and for $k\to\infty$ the discontinuity $\Delta_\ell(-k^2)$ tends to its Born
approximation. However, when $m-2\ell-1>0$ each iteration increases the power of $k$ in the asymptotic behavior
of $\Delta_\ell(-k^2)$, becoming more and more divergent as $n$ increases.
This is the situation for a singular potential.
For the regular potentials $\Delta_\ell(p^2)$ tends to its Born term contribution which vanishes at least as
$1/p^2$ for $p^2=-k^2$ and $k^2\to \infty$.
For such type of $\Delta_\ell(p^2)$ it was shown in Ref.~\cite{guo.rios.NN}
that any $N/D$ IE, irrespectively of the number of subtractions taken, has solution.
However, for singular potentials the resulting $|\Delta_\ell(p^2)|$ grows faster than any polynomial in
the same limit. This is clearly shown in Ref.~\cite{oller.lhc.aop} by log-log plots in which the
slop of $|\Delta(-k^2)|$ continuously grows with increasing $k^2$. As a dramatic consequence of this result is that
it is not possible to write down a DR representation for a NR PWA if the potential is singular. However,
it is still possible to use the $N/D$ method because what matters for the $N/D$ IEs is the product
$\Delta_\ell(-k^2)D_\ell(-k^2)$. The denominator function is known to behave asymptotically
as $s^{-\delta(\infty)/\pi}$, cf. Eq.~\eqref{181122.3}, and $\delta(\infty)=N\pi$,
with $N$ the number of bound states, because of the Levinson theorem.
It turns out that the number of such stats is infinite for attractive singular potentials \cite{singular.rev}
and, in this case, $D_\ell(-k^2)$ vanishes also faster than any power law.
The exact $N/D$ method is defined in Ref.~\cite{oller.lhc.aop} as the $N/D$ method
but using $\Delta_\ell(p^2)$ stemming from
Eqs.~\eqref{200520.4} and \eqref{200520.5}, which is the exact LHC discontinuity of the full PWA
for a given potential.
In this way, we showed in Ref.~\cite{oller.lhc.aop} that one reproduces exactly the
LS-equation solutions for regular potentials. This is also true for the singular potentials
when the potential is used in the whole range of integration
in the LS equation, that is, for $q\in[0,\infty]$ (the cut-off is sent to infinity).
For the singular-potential case we refer to the standard kind
of solutions, so that for a repulsive singular potential the solution has no free
parameters and is determined,
while for the attractive singular case the solution involves one free parameter that could be fixed e.g.
by imposing a given value for the scattering length \cite{singular.rev,singular.case,singular.arr}.
Several potentials were studied in Ref.~\cite{oller.lhc.aop}, both for uncoupled and coupled PWAs.
Within the latter group the $^3S_1$--$^3D_1$ coupled PWAs were studied and the $^3S_1$ scattering length was
taken as input. Needless to say, in all cases the LS equation with infinite cutoff and the $N/D$ method agree
perfectly in our numerical study.
The fact of having none or only one free parameter is a very constrained situation in practical applications,
and it is the reason why it has not been possible to achieve yet a good
agreement with data in $NN$ scattering in terms of regulator-independent solutions (i.e. in
which the three-momentum cut-off is taken to infinity). Notice that the number of free parameters in the solution
of the LS equation for singular potentials is then not linked with the chiral order
in the calculation of the chiral potential.
However, in terms of the $N/D$ method one can in principle add an arbitrary number of subtractions,
which allows one to look for extra solutions.
We have already discussed this point in connection with the ambiguity associated with the CDD poles
in Sec.~\ref{sec.200517.2}.
This possibility was explored in detail in Ref.~\cite{oller.lhc.plb} for the $^1S_0$ $NN$ PWA.
The NLO and NNLO ChPT potentials for this PWA are actually attractive and singular. The standard solutions of
the LS equation were reproduced, and a detailed numerical analysis was performed in order to show the
agreement between the LS equation and the exact $N/D$ method.
But we also showed in this reference that one can generate new solutions that cannot be achieved
by the LS equation when the three-momentum cut-off is taken to infinity
with contact interactions included in the potential to aim renormalization (in the form of
polynomial counterterms in its momentum expression).
In this way, a new solution was discussed that can reproduce the $^1S_0$
scattering length, effective range and shape parameter $v_2$.
For this solution the DRs for $N_\ell(s)$ and $D_\ell(s)$ converge separately.
It is also interesting to indicate that a solution within the exact $N/D$ method for this PWA
fixing only two parameters, the scattering length and the effective range were taken, could not be found.
Last but not least, a very attractive feature of the exact $N/D$ method is that it allows to evaluate the PWAs in
the whole complex $p^2$ plane. Then, it is very convenient to look for resonance and (anti)bound states.
In the case of the $^1S_0$ PWA there is an antibound state which is found at $p=-i0.066$~MeV both at NLO and NNLO
when all the first three ERE parameters are reproduced.
\section{Conclusions}
\label{sec:conclu}
\def\arabic{section}.\arabic{equation}{\arabic{section}.\arabic{equation}}
\setcounter{equation}{0}
We have elaborated on several unitarization methods of perturbative calculations
in Chiral Perturbation Theory (ChPT) that can be employed to study scattering and
the re-scattering corrections to an external probe. Special attention has been
given to the $N/D$ method both for scattering and for implementing the final-state interactions (FSI).
The unitarization methods, since the earlier papers on current algebra techniques,
have been able to extend to much larger energies the expected region of utility of
ChPT calculations.
This has been accomplished thanks to the extra energy and momentum dependence
generated by using a non-perturbative
theoretical framework which satisfies key properties of $S$-matrix theory,
which stem from two-body unitarity and analyticity.
Some of the most striking and important applications of the unitarization methods of input perturbative
calculations have occurred in the field of spectroscopy.
In this way,
it has been possible to study resonances and bound states, and even predict
some of them, while unexpected properties have been unveiled too, as e.g. the two-pole nature
of some resonances \cite{meissner.two}, as first shown for the $\Lambda(1405)$ in Ref.~\cite{plb}.
Along this review we have paid attention to establish links between different unitarization methods.
Thereby, by starting with the (generalized) relativistic effective-range expansion (ERE)
we have connected it with the $K$-matrix approach and then obtained from the former
the Inverse Amplitude Method (IAM) unitarization formula.
The IAM has been also connected with the Padd\'e approximation.
In the last part of the manuscript we have introduced and discussed the $N/D$ method.
A link between the $N/D$ and the IAM can also be established by employing
the solution to the $N/D$ method based on treating perturbatively the left-hand cut discontinuity.
This allows one to derive the IAM as a particular case of this method too.
The associated methods to take care of the FSI
corresponding to the unitarization techniques of scattering have been introduced as well. In addition,
we have discussed the (Muskhelishvili-)Omn\'es solution and the Khuri-Treiman approach.
An advantage of the unitarization technique based on the $N/D$ method is that it
can be applied to deliver the
unitarized partial-wave amplitudes (PWAs)
even if only the leading-order scattering amplitudes are employed.
A subtraction constant is then required, but it could be estimated making use of naturalness arguments.
In this way, one can study important resonances in hadron physics in a very constrained manner,
essentially without any free parameter. Good examples are
the $f_0(500)$, $f_0(980)$, $a_0(980)$, and $\kappa(800)$ in the scalar light mesonic sector, the
$\Lambda(1405)$ in the strangeness $-1$ $S$-wave meson-baryon scattering, etc.
Of course, one could also use as input higher-order scattering amplitudes provided by the effective field theory
of interest and perform a higher-order analysis in the input taken.
Regarding the $N/D$ method, we would like to stress that thanks to recent advances (in which the author has been involved),
it can be considered for non-relativistic scattering as an alternative formulation of scattering theory.
The qualitative leap forward has been the derivation of the exact discontinuity of a PWA along the
left-hand cut, which can then be employed to solve the $N/D$ integral equations. In this way, one can solve
standard regular potentials and reproduce the solutions obtained with the Lippmann-Schwinger equation
for PWAs.
But it also allows to obtain extra solutions for singular potentials without dependence on cutoff, which can be sent to infinity.
This method has the advantage that, in terms of the solution found, it is
straightforward to evaluate the on-shell scattering amplitudes in the complex energy plane too.
In this way, e.g. one could look for poles and their residues
(which give the resonance or bound-state couplings).
This is a very promising and exciting field of current research,
and first applications are being explored for $NN$ scattering.
\acknowledgments{This work has been supported in part by the MEC (Spain) and FEDER (EU) Grants FPA2016-77313-P and
PID2019-106080GB-C22.}
\reftitle{References}
| 43f075516e34642044025f3966236b4f702c785e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
In the US, the Airline Deregulation Act was established in 1978, which removed the governmental controls over fares, routes, and market entries of new air carriers (carriers for short)\cite{brown1987politics}. After the establishment of the law, the formerly regional carriers and completely new carriers grew in the market and increased the level of competition among carriers\cite{goetz2009good}. The deregulation also caused the emergence of low cost carriers (LCCs). In contrast to full service carriers (FSCs), an LCC offers low fares without many traditional passenger services. The growth of LCCs, especially Southwest Airlines in the case of the US, which is the largest LCC in the world in terms of passenger traffic\cite{ren2020fare}, elicited a dramatic downward pressure on fares\cite{brueckner2013airline}. After the deregulation in the US, similar airline deregulations and subsequent elevated competition among carriers have been duplicated around the world including in Europe\cite{alderighi2012competition} and Asia\cite{zhang2008low}.
In addition to fierce competition among carriers, the airline industry has been exposed to a highly competitive environment partly owing to various socioeconomic events. For example, the 9/11 attacks in 2001 had significant impacts on the US airline industry. This event resulted in a negative demand shock of more than 30\%\cite{ito2005assessing}. The US airline industry lost nearly \$35 billion between 2001 and 2005 in total\cite{goetz2009good}. In fact, a number of carriers filed liquidation bankruptcies (so-called Chapter 7 bankruptcy) or reorganization bankruptcies (so-called Chapter 11 bankruptcy) including major carriers such as American Airlines, United Airlines, and Delta Air Lines\cite{ciliberto2012bankruptcy}. Moreover, the US airline industry has faced massive restructuring with a number of mergers such as the merger of American Airlines and US Airways, and of Delta Air Lines and Northwest Airlines\cite{steven2016mergers}. In Europe, the volcanic eruptions in Iceland in 2010 had a devastating effect on the European air transport network, preventing air travel throughout most of Europe for six days\cite{oroian2010eyjafjallajokull}.
Overall, business environments in the airline industry have been tough on the global scale. For example, the International Air Transport Association (IATA) announced that year 2019 was the smallest in freight demand in the world since 2009\cite{IATA2020airline}. As of 2020, the airline industry is facing a financial crisis because of the COVID-19 pandemic. The Europe's biggest regional carrier of the time, Flybe, went bankrupt in March 2020, and many carriers may follow\cite{Harper2020airline}.
The airline industry is highly susceptible to such socioeconomic and natural events, and these events may influence the structure of air transport networks. Regardless of the reason, air transport networks can be regarded as temporal networks in which nodes are airports and edges are direct flight connections\cite{rocha2017dynamics}; in various types of networks such as social and economic networks, nodes and/or edges change over time, constituting temporal networks \cite{holme2012temporal,holme2015modern,masuda2020guidance,holme2019temporal}.
In fact, air transport networks consist of multiple carriers. Each carrier has its own corporate strategies, which may change over time depending on evolution of the market and knowledge that the carrier gains by learning from databases (e.g., about customers)\cite{bieger2011airline}. These time-varying corporate strategies of carriers may affect the evolution of their network structure.
A number of studies have analyzed structures of air transport networks. Most of these studies have examined static properties of air transport networks such as the degree distribution, average path length, clustering coefficient, assortativity, and centrality\cite{li2003structural, guimera2004modeling, guimera2005worldwide, guida2007topology, bagler2008analysis, xu2008exploring, cheung2012complex}. In recent years, several studies have investigated evolution of air transport networks.
For example, in an air transport network in Brazil, the number of nodes and edges decreased from 1995 to 2006, whereas the number of passengers and the amount of cargo increased during the same period\cite{da2009structural}.
In contrast, the connectivity of an air transport network in China increased from 1930 to 2012\cite{wang2014evolution}.
These and other\cite{rocha2017dynamics,dai2018evolving,sun2015temporal,azzam2013accelerated} studies are concerned with the evolution of air transport networks at the country, continental, or global level and do not distinguish networks by carriers. In fact, each carrier may be evolving in different manners, reflecting its own corporate strategies. Because carriers compete for a limited traffic demand, some carriers expanding their networks may cause other carriers to shrink their networks. Even if the structure of an entire network combining different carriers is relatively stable\cite{lin2014evolving}, each carrier may be drastically changing its network structure over time. To date, time-dependent structure of air transport networks at the carrier level is not well understood\cite{lordan2014study}, while a few studies have investigated static properties of airline networks by carriers\cite{han2009network, reggiani2009network}.
In the present study, we propose a novel framework for analyzing evolution of air transport networks. In particular, we examine temporal air transport networks from the viewpoint of recurrence. Recurrence implies that the structure of the network returns to one close to that of the same network in the past. Recurrent plots and recurrence quantification analysis are a set of data analysis methods to quantify recurrence, which was originally proposed for time series \cite{marwan2007recurrence} and recently extended to the case of temporal networks \cite{masuda2019detecting, lopes2020recurrence}. While we do not employ recurrence plots or recurrence quantification analysis in the exact sense in the present study, we will use the idea of recurrence to examine when air transport networks approach their past structure relatively closely as compared to in other times. We apply the proposed framework to four major carriers in the US, i.e., American Airlines (AA), United Airlines (UA), Delta Air Lines (DL), and Southwest Airlines (WN), from 1987 to 2019. In particular, we ask how the network structure of different carriers responds to major events such as mergers, which airports are involved in major changes in the network structure, and how different carriers may be different in the extent of recurrence and periodicity of the network structure.
\section*{Methods}
\subsection*{Data set and construction of monthly air transport networks}
We use a data set available at the US Bureau of Transportation Statistics. The data are based on the reports made by certified carriers that account for at least one percent of the domestic scheduled passenger revenue. By definition, a certified carrier holds an operating certificate issued by the US Secretary of Transportation. The data set contains the origin and destination airports of each flight, departure and arrival times, operating carriers, and so on. We downloaded monthly data between October 1987 and August 2019\cite{BTS2020air}.
We focus on air transport networks for the four most major carriers in the US in terms of revenue, i.e., AA, UA, DL, and WN. According to the US Bureau of Transportation Statistics, domestic revenue passenger-miles of each of these four carriers in 2019 is over 100 billions. Alaska Airlines, ranked fifth, earned domestic revenue passenger-miles less than 50 billions in 2019. For each carrier, we construct monthly snapshot networks from October 1987 to August 2019, in which a node is an airport. We connect two nodes by an edge if there is at least one direct commercial flight between the two airports in the month. For simplicity, we assume that each snapshot network is unweighted and undirected.
\subsection*{Distance matrix}
To measure the dissimilarity between two snapshot networks, we define the normalized network distance (network distance for short) between two networks $G$ and $G^{\prime}$ as
\begin{equation}
d(G,G^{\prime})= 1-\frac{M(G\cap G^{\prime})}{\sqrt{M(G)M(G^{\prime})}}\,,
\label{eq:normalized_network_distance}
\end{equation}
where $M(G)$ and $M(G^{\prime})$ are the numbers of edges in $G$ and $G^{\prime}$ respectively, and $M(G\cap G^{\prime})$ is the number of edges that $G$ and $G^{\prime}$ have in common. Network distance $d$ ranges between $0$ and $1$. The distance matrix for a carrier is a $t_{\max} \times t_{\max}$ symmetric matrix of which the $(t, t')$th entry is given by $d(G_t, G_{t'})$,
where $G_t$ is the network at month $t$ and $t_{\max}$ is the number of months observed. If $d(G_t, G_{t'})$ is small, where $t < t'$, we regard that the network at month $t'$ approximately recurs to that at month $t$. As we will show, the distance matrices based on Eq.~\eqref{eq:normalized_network_distance} show clear horizontal and vertical lines that correspond to abrupt changes in the network structure, which are synchronous with the sudden increases in the number of nodes and edges upon major mergers. We decided to use Eq.~\eqref{eq:normalized_network_distance} because we did not observe such abrupt changes in the distance matrices calculated from three other network distance measures (see Supplementary Fig.~S1 online).
\subsection*{Statistical analysis}
To measure differences in the distance matrices, we conduct statistical tests on distributions of $d$. First, we use the Kruskal-Wallis test, which is an nonparametric statistical test to assess the differences among three or more groups\cite{kruskal1952use}. We used ``Kruskal'' function in R (ver. 3.6.3). Second, as a post-hoc test, we use the Bonferroni-corrected Mann-Whitney U test \cite{mann1947test} to compare all pairs of groups. We used ``wilcox.test'' function in R (ver. 3.6.3). In addition to the $p$ values, we measure its effect size $r$, which is the $Z$ value obtained from the ``wilcox.test'' function divided by the square root of the number of observations\cite{rosenthal1994parametric}. According to a standard, the effect size is said to be strong, medium, or small when $r>0.5$, $r>0.3$, or $r>0.1$, respectively\cite{cohen2013statistical}.
\subsection*{Autocorrelation function and power spectral density}
To measure the similarity between air transport networks across time, we define an autocorrelation function (ACF) for a given carrier by
\begin{equation}
\mbox{ACF}(\tau)=1-\frac{1}{t_{\max}-\tau}\sum_{t=1}^{t_{\max}-\tau} d(G_{t},G_{t+\tau})\,,
\label{eq:ACF}
\end{equation}
where $\tau$ is the time lag. Note that Eq. \eqref{eq:ACF} is not a proper autocorrelation function. However, in analogy to the Wiener–Khinchin theorem\cite{wiener2019cybernetics}, we define a power spectral density (PSD) by the Fourier transform of the ACF as follows:
\begin{equation}
\mbox{PSD}(f)=\int_{-\infty}^{\infty} \mbox{ACF}(\tau)e^{-2\pi if\tau}{\rm d}\tau\,,
\label{eq:PSD}
\end{equation}
where $f$ is the frequency. The PSD shows the power of periodic variations.
\subsection*{State detection in temporal air transport networks}
We identify states of the temporal air transport networks for each carrier based on its distance matrix. Consider a sequence of $t_{\max}$ static networks. To assign a state to each of the $t_{\max}$ snapshot networks, we apply a hierarchical clustering algorithm to the $t_{\max} \times t_{\max}$ distance matrix\cite{masuda2019detecting}. We used ``linkage'' and ``fcluster'' functions in ``scipy'' module in Python. To define the distance between clusters, we used ``ward'' in the ``linkage'' function. The hierarchical clustering provides partitions of the monthly networks into $C$ discrete states, where the number of states, $C$, ranges between $1$ and $t_{\max}$. We adopt the value of $C\ (2 \le C \le t_{\max})$ that maximizes the Dunn's index, $D$ \cite{dunn1974fuzzy}, defined by
\begin{equation}
D=\frac{\min_{1 \leq c \neq c' \leq C} \min_{G_i \in c{\rm th\ state}, G_j \in c'{\rm th\ state}} d(G_i,G_j)}{\max_{1 \leq c'' \leq C} \max_{G_{i'}, G_{j'} \in c''{\rm th\ state}} d(G_{i'},G_{j'})}\,.
\label{eq:dunn}
\end{equation}
The numerator on the right-hand side of Eq.~\eqref{eq:dunn} represents the smallest distance between two states among all the pairs of states. The denominator represents the largest diameter of the state among all the states.
\section*{Results}
\subsection*{Evolution of the numbers of nodes and edges}
We show the evolution of the numbers of nodes and edges for each carrier in Fig.~\ref{fig:nodes_edges}. The three FSCs (i.e., AA, UA, and DL) experienced the Chapter 11 bankruptcies. This is a form of bankruptcy that permits reorganization under the bankruptcy laws. A debtor proposes a plan of reorganization, and it can exit from the bankruptcy if it completes the proposed plan. The dashed lines in Fig.~\ref{fig:nodes_edges} represent the times of the Chapter 11 bankruptcies (shown in red) and exits from them (shown in yellow). Additionally, we show the times of the 9/11 attacks and mergers by the dotted-dashed lines and dotted lines, respectively.
The three FSCs show some common patterns. Between the middle of the 1990s and around 2010, the numbers of nodes and edges largely decreased. After the 9/11 attacks, the three carriers experienced the bankruptcies. Then, after they exited from the bankruptcy, they merged with other carriers. Upon the mergers, the number of edges almost doubled suddenly. There was no notable change in the numbers of nodes and edges in response to the bankruptcies and the exits from them. In contrast, WN, which is an LCC, has never experienced bankruptcy. The numbers of nodes and edges in the WN network have roughly monotonically increased over the two decades despite the 9/11 attacks and the merger. In the next sections, we further study the structural evolution of the four carriers.
\subsection*{Distance matrix for each carrier}
\subsubsection*{Visualization and quantification of the distance matrices}
In Fig.~\ref{fig:distance_matrices}, we show the distance matrices for the four carriers. First, we notice that there are clear horizontal and vertical lines in the distance matrices. These lines indicate abrupt changes in the network structure. For example, the figure suggests that the three FSCs considerably changed their network structure upon the mergers after they exited from the bankruptcies (i.e., the merger between AA and US Airways (US) in 2015 in Fig.~\ref{fig:distance_matrices}a, that between UA and Continental Airlines (CO) in 2012 in Fig.~\ref{fig:distance_matrices}b, and that between DL and Northwest Airlines (NW) in 2010 in Fig.~\ref{fig:distance_matrices}c). These results are consistent with the approximately two-fold sudden increase in the numbers of edges shown in Figs.~\ref{fig:nodes_edges}a-c. However, the distance matrices additionally indicate that the network structure after the merger is considerably different from that before the merger for each FSC. For example, the DL network after the merger with NW in 2010 is considerably different from that in the late 1990s, although the numbers of nodes and edges are similar at the two time points (Fig.~\ref{fig:nodes_edges}c).
\begin{figure}[H]
\centering
\includegraphics[width=0.88\linewidth]{Fig1.pdf}
\caption{Evolution of the numbers of nodes and edges over time. (a) American Airlines, (b) United Airlines, (c) Delta Air Lines, and (d) Southwest Airlines.}
\label{fig:nodes_edges}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=0.99\linewidth]{Fig2.pdf}
\caption[]
{Network-distance matrices. (a) American Airlines, (b) United Airlines, (c) Delta Air Lines, and (d) Southwest Airlines.}
\label{fig:distance_matrices}
\end{figure}
Second, there are checkerboard patterns in the distance matrices, for example, around year 2010 for AA (Fig.~\ref{fig:distance_matrices}a). The checkerboard patterns imply seasonal periodicity in the network structure. These patterns are also consistent with the annual periodic fluctuations in the number of nodes and edges visible in Fig.~\ref{fig:nodes_edges}a. Note that the classical recurrence plots\cite{marwan2007recurrence}, in which the distance values are binarized into 0 or 1, would not uncover patterns of the distance matrix either at relatively large distance values (e.g., the sudden changes in response to the post-bankruptcy mergers) or at relatively small distance values (e.g., the checkerboard patterns). In other words, recurrence plots with a small threshold value for binarization will not have the long horizontal and vertical lines marking major mergers because relatively large distance values that contribute to generation of these lines are all binarized to 1. By contrast, recurrence plots with a large threshold value will miss the checkerboard patterns because relatively small distance values that contribute to the generation of the checkerboard patterns are all binarized to 0. In our framework, we analyze the original values of the network distances without binarizing them, which enables us to reveal various patterns in the distance matrices without varying a threshold parameter.
Third, we notice that the distance matrix for WN (Fig.~\ref{fig:distance_matrices}d) has many bright horizontal and vertical stripes after 1995. They suggest that WN has frequently changed their network, but for a short duration each time such as a single month.
\begin{table}[thb]
\centering
\caption{\label{tab:Statistical_analysis_1} Results of the Bonferroni-corrected Mann-Whitney U test to compare each pair of carriers in terms of $d$. ***: $p<0.001$.}
\begin{threeparttable}
\begin{tabular}{|c|c|}
\hline
& Effect size, $r$ \\ \hline
AA vs UA & $0.0129^{***}$ \\ \hline
AA vs DL & $0.211^{***}$ \\ \hline
AA vs WN & $0.325^{***}$ \\ \hline
UA vs DL & $0.225^{***}$ \\ \hline
UA vs WN & $0.337^{***}$ \\ \hline
DL vs WN & $0.162^{***}$ \\ \hline
\end{tabular}
\footnotesize
\end{threeparttable}
\end{table}
In Fig.~\ref{fig:distance_all_period}, we show the box plots of the distribution of the network distance between pairs of months. Each box plot summarizes each distance matrix shown in Fig.~\ref{fig:distance_matrices}. To be quantitative, we carry out statistical analyses on these distributions. First, we carry out the Kruskal-Wallis test to find that there is a significant difference among the four distributions shown in Fig.~\ref{fig:distance_all_period} ($p<0.001$). Second, as a post-hoc test, we run the Bonferroni-corrected Mann-Whitney U test to compare all six pairs of carriers. Table~\ref{tab:Statistical_analysis_1} indicates that each pair of carriers is significantly different, partly owing to large sample sizes. According to a standard practice, two pairs, AA-WN and UA-WN, are different with medium effect sizes ($r>0.3$), and three pairs, DL-WN, AA-DL, and UA-DL, are different with small effect sizes ($r>0.1$). The effect size for pair AA-UA is only 0.0129. These results support our casual observation with Fig.~\ref{fig:distance_matrices}, i.e., WN shows unique patterns and has kept changing the network structure more considerably than the other three carriers. Additionally, DL changed its structure significantly more than AA and UA did over time.
\subsubsection*{Periodicity}
In Fig.~\ref{fig:ACF_PSD_four_air_carriers}a, we show the ACFs for the four carriers, calculated from the distance matrices shown in Fig.~\ref{fig:distance_matrices}. This figure indicates that networks closer in time are more similar to each other. We also observe that the ACF for WN is the smallest for all the lag $\tau >0$, confirming that the WN network tends to vary over time to a larger extent than the other carriers. In addition, Fig.~\ref{fig:ACF_PSD_four_air_carriers}a reconfirms that the DL network tends to vary over time more than the AA and UA networks do.
The ACFs for the FSCs have local maxima at multiples of 12 months, suggesting annual periodic patterns. To quantify the strength of annual cycles, we show the PSDs for the four carriers in Fig.~\ref{fig:ACF_PSD_four_air_carriers}a. The PSDs have local maxima, corresponding to an annual cycle, for AA and UA. Here we define the strength of the annual periodicity, denoted by $s$, at $f=1/12$ [1/month] by comparing $\mbox{PSD}(1/12)$ with the middle point of the two neighbors, $\mbox{PSD}(1/15)$ and $\mbox{PSD}(1/10)$, as follows:
\begin{equation}
s=\frac{\mbox{PSD}(1/12)}{(\mbox{PSD}(1/15)+\mbox{PSD}(1/10))/2}\,.
\label{eq:strength}
\end{equation}
A value of $s$ larger than 1 suggests some periodicity with frequency $f=1/12$. We show $s$ for each carrier in Fig.~\ref{fig:strength_periodicity_four_airlines}. The figure suggests that AA has the strongest annual periodicity, followed by UA and DL. The value of $s$ for WN is less than 1, suggesting the lack of periodicity.
\begin{figure}[H]
\centering
\includegraphics[width=0.39\linewidth]{Fig3.pdf}
\caption{The distribution of the network distance $d$. The box plots show five-number summaries of the distribution: first quartile ($Q_1$), median, third quartile ($Q_3$), minimum without outliers ($Q_1-1.5 \, \times \, IQR$), and maximum without outliers ($Q_3+1.5 \, \times \, IQR$), where IQR $=Q_3-Q_1$. Open circles represent outliers.}
\label{fig:distance_all_period}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.99\linewidth]{Fig4.pdf}
\caption{Analysis of autocorrelation and periodicity. (a) ACFs and (b) PSDs for the four carriers.}
\label{fig:ACF_PSD_four_air_carriers}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.38\linewidth]{Fig5.pdf}
\caption{Strength of annual periodicity in the air transport networks.}
\label{fig:strength_periodicity_four_airlines}
\end{figure}
\subsubsection*{State transitions in temporal air transport networks}
In this section, we examine states of temporal air transport networks for each carrier.
In Figs.~\ref{fig:nodes_edges} and \ref{fig:distance_matrices}, we observed that the network structure for the three FSCs drastically changed upon the mergers after they exited from the bankruptcy. If we apply our algorithm to detect states of temporal networks to the entire observation, we detect two states for each FSC, one state corresponding to the time window before the merger and another state after the merger. However, a visual inspection of Fig.~\ref{fig:distance_matrices} suggests that we may benefit from distinguishing multiple states for the time period before the merger. Therefore, for each FSC, here we apply the hierarchical clustering algorithm to detect state dynamics to the distance matrix of the temporal network restricted to before the post-bankruptcy merger. We regard the snapshot networks after the post-bankruptcy merger as one distinct state. For WN, we applied the algorithm to detect system states to the entire observation window because the WN network has not considerably changed upon the merger with AirTran Airways (FL) in 2015 up to our visual inspections of Figs.~\ref{fig:nodes_edges}d and \ref{fig:distance_matrices}d.
We show the detected state dynamics for the four carriers in Fig.~\ref{fig:detected_states}. For all the four carriers, we detected a change point at January 1995. We identified three states for AA and UA, and five states for DL. The detected changes for DL from state 2 to 3 and state 3 to 4 are located at January 2001 and December 2005, respectively. For WN, we detected two states. Differently from the other carriers, state 1 for the WN network is not consecutive in time, i.e., state 1 reappears in October 1995 for a month after the state transition on January 1995. We notice that the detected change point at December 2005 for DL is three months after the bankruptcy. However, except for this instance and the state after the post-bankruptcy mergers for the FSCs, the detected change points coincide with neither the 9/11 attacks (i.e., September 2001), the bankruptcies (i.e., AA: November 2011, UA: December 2002), nor the exits from the bankruptcies (i.e., AA: December 2013, UA: February 2006, DL: April 2007).
\begin{figure}[bth]
\centering
\includegraphics[width=0.83\linewidth]{Fig6.pdf}
\caption{System-state dynamics of temporal air transport networks. (a) American Airlines, (b) United Airlines, (c) Delta Air Lines, and (d) Southwest Airlines. To aid visual understanding, we show the network distance on a color scale if and only if the two snapshot networks belong to the same state. Otherwise, the corresponding cell is shown in yellow.}
\label{fig:detected_states}
\end{figure}
To understand differences in the network structure across the states, we visualize deleted and added edges for each pair of consecutive states. We define deleted edges as the edges that exist at least once in the older state and never exist in the newer state. The added edges are defined as edges that never exist in the older state and exist at least once in the newer state. We show the results in Fig.~\ref{fig:change_networks}. The deleted and added edges are shown in red and blue, respectively.
Figure~\ref{fig:change_networks}a indicates that, in transition to state 2, AA withdrew edges especially at BNA (Nashville International Airport) and RDU (Raleigh-Durham International Airport) and added edges at STL (St. Louis Lambert International Airport) and LAX (Los Angeles International Airport). These results are consistent with the fact that AA closed two hub airports, BNA and RDU, in the middle of the 1990s and merged with Trans World Airlines (TW) in 2002 whose major airports were STL and LAX. Figure~\ref{fig:change_networks}b indicates that, in transition to state 3, AA withdrew edges by the largest number at STL and added many edges at CLT (Charlotte Douglas International Airport), PHX (Phoenix Sky Harbor International Airport), and PHL (Philadelphia International Airport). These results are consistent with the fact that AA ceased to use STL as a hub around 2010 and merged with US whose major airports were CLT, PHX, and PHL.
Figure~\ref{fig:change_networks}c indicates that, in transition to state 2, UA changed many routes connecting to IAD (Washington Dulles International Airport), which has been a hub airport for UA. Figure~\ref{fig:change_networks}d indicates that, in transition to state 3, UA added edges by the largest numbers at IAH (George Bush Intercontinental Airport) and EWR (Newark Liberty International Airport). This event originated from the merger with CO whose major airports had been IAH and EWR.
Figures~\ref{fig:change_networks}e and \ref{fig:change_networks}f indicate that, in transition from state 1 to 2 and from state 2 to 3, DL changed many edges connecting to some airports such as MCO (Orlando International Airport), CVG (Cincinnati/Northern Kentucky International Airport), and ATL (Hartsfield–Jackson Atlanta International Airport). Figure~\ref{fig:change_networks}g shows that, in transition to state 4, DL withdrew edges by the largest number at DFW (Dallas/Fort Worth International Airport). Figure~\ref{fig:change_networks}h is consistent with the merger with NW whose major airports were MSP (Minneapolis–Saint Paul International Airport) and DTW (Detroit Metropolitan Airport).
Figure~\ref{fig:change_networks}i indicates that, in transition to state 2, WN expanded the network by adding many edges at some airports such as MDW (Midway International Airport), DEN (Denver International Airport), and MCO.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{Fig7.pdf}
\caption{Deleted edges (shown in red) and added edges (shown in blue) for each pair of the consecutive states. In each panel, we placed 10 nodes with the largest number of added and deleted edges in total in the peripheral part of the visualization.}
\label{fig:change_networks}
\end{figure}
\subsubsection*{State-dependent periodicity}
In the previous section, we documented differences in the network structure across the states through visualization. We expect that the different states are systematically different in some quantitative aspects. Therefore, in this section, we quantitatively examine the temporal variability of the network and possible annual periodicity of network changes, which themselves may change on the timescale of decades. We measure the ACF and PSD for each state and each carrier. Note that a lower ACF suggests a larger temporal variability of time series\cite{murray2014hierarchy,cavanagh2016autocorrelation,watanabe2019atypical}. Because state 1 for WN is not consecutive in time, we exclude the single isolate month in state 1 (i.e., October 1995) and nine isolated months in state 2 (i.e., from January to September 1995) in the following analysis in the following analysis.
\begin{figure}[H]
\centering
\includegraphics[width=0.97\linewidth]{Fig8.pdf}
\caption{ACFs and PSDs for each network state. (a) American Airlines, ACF. (b) American Airlines, PSD. (c) United Airlines, ACF. (d) United Airlines, PSD. (e) Delta Air Lines, ACF. (f) Delta Air Lines, PSD. (g) Southwest Airlines, ACF. (h) Southwest Airlines, PSD.}
\label{fig:ACFs_PSDs_each_state}
\end{figure}
We show the ACFs and PSDs for each state and each carrier in Fig.~\ref{fig:ACFs_PSDs_each_state}. For AA and UA, the ACF is the largest in state 2, followed by state 1, and then by state 3, for all the lag values (Fig.~\ref{fig:ACFs_PSDs_each_state}a,c). This result indicates that the temporal variability is smaller (i.e., the ACF is larger) in state 2, which roughly spans from 1995 to 2013 for both carriers. In contrast, the ACF for DL and WN tends to be smaller at later times, at least up to $\tau=2$ years of the lag (Fig.~\ref{fig:ACFs_PSDs_each_state}e,g). This result indicates that the temporal variability becomes larger over time for DL and WN.
The PSDs for some combinations of the state and carrier have a local maximum at $f=1/12$. We measure the strength of periodicity in the same manner as Eq. \eqref{eq:strength}, i.e., by comparing $\mbox{PSD}(1/12)$ with the middle point of the two neighbors, i.e., $\mbox{PSD}(1/18)$ and $\mbox{PSD}(1/9)$. It should be noted that the largest lag assumed in Fig.~\ref{fig:ACFs_PSDs_each_state} is 30 as opposed to 50 in Fig.~\ref{fig:ACF_PSD_four_air_carriers}, which results in different $f$ values for the two neighbors of $f=1/12$ (i.e., $f=1/18$ and $1/9$) as compared to those for Fig.~\ref{fig:ACF_PSD_four_air_carriers}. We show the mean degree, $\langle k \rangle$, and the strength of periodicity, $s$, for the different states and carriers in Fig.~\ref{fig:strength_periodicity_all_states}. The figure suggests that both $\langle k\rangle$ and $s$ depend on the states but in different manners. For example, the periodicity is much larger in the final state in the UA and DL networks, which the state dependence of $\langle k\rangle$ does not explain. The figure further indicates that all the four carriers did not have a clear periodicity before 1995 (corresponding to state 1 for each carrier). After 1995, all carriers strengthened the periodicity. In the latest states, UA and DL have stronger periodicity than AA and WN.
\begin{figure}[h]
\centering
\includegraphics[width=0.65\linewidth]{Fig9.pdf}
\caption{Mean degree and strength of periodicity in each state.}
\label{fig:strength_periodicity_all_states}
\end{figure}
\subsection*{Network distance between carriers}
In this section, we examine similarity in the network structure across different carriers. We calculate the distance matrix for each pair of carriers by measuring the normalized network distance between a network at month $t$ for a carrier and a network at month $t'$ for another carrier, where $1\le t,\ t' \le t_{\max}$. The distance matrices in Fig.~\ref{fig:distance_matrices_different_airlines} indicate that the four carriers are all dissimilar to each other; note that the distance values range between 0 and 1 by definition. Nevertheless, the AA and UA networks are apparently more similar to each other than the other carrier pairs are on average. Another observation is that there are some visible horizontal and vertical lines in the figure (e.g., a horizontal line in 2005 in Fig.~\ref{fig:distance_matrices_different_airlines}b, a vertical line in 2015 in Fig.~\ref{fig:distance_matrices_different_airlines}c, and a horizontal line in 2008 in Fig.~\ref{fig:distance_matrices_different_airlines}e).
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{Fig10.pdf}
\caption{Network distance between different carriers over time. (a) American Airlines - United Airlines, (b) American Airlines - Delta Air Lines, (c) American Airlines - Southwest Airlines, (d) United Airlines - Delta Air Lines, (e) United Airlines - Southwest Airlines, and (f) Delta Air Lines - Southwest Airlines.}
\label{fig:distance_matrices_different_airlines}
\end{figure}
\begin{table}[h]
\caption{\label{tab:Statistical_analysis_2} Results of the Bonferroni-corrected Mann-Whitney U test to compare all 15 pairs of carriers in terms of $d$. ***: $p<0.001$.}
\begin{minipage}[t]{.33\textwidth}
\begin{center}
\begin{tabular}{|c|c|}
\hline
& Effect size, $r$ \\ \hline
AA-UA vs AA-DL & $0.512^{***}$ \\ \hline
AA-UA vs AA-WN & $0.866^{***}$ \\ \hline
AA-UA vs UA-DL & $0.866^{***}$ \\ \hline
AA-UA vs UA-WN & $0.855^{***}$ \\ \hline
AA-UA vs DL-WN & $0.865^{***}$ \\ \hline
\end{tabular}
\end{center}
\end{minipage}
%
\hfill
%
\begin{minipage}[t]{.33\textwidth}
\begin{center}
\begin{tabular}{|c|c|}
\hline
& Effect size, $r$ \\ \hline
AA-DL vs AA-WN & $0.742^{***}$ \\ \hline
AA-DL vs UA-DL & $0.693^{***}$ \\ \hline
AA-DL vs UA-WN & $0.585^{***}$ \\ \hline
AA-DL vs DL-WN & $0.621^{***}$ \\ \hline
AA-WN vs UA-DL & $0.304^{***}$ \\ \hline
\end{tabular}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[t]{.33\textwidth}
\begin{center}
\begin{tabular}{|c|c|}
\hline
& Effect size, $r$ \\ \hline
AA-WN vs UA-WN & $0.134^{***}$ \\ \hline
AA-WN vs DL-WN & $0.243^{***}$ \\ \hline
UA-DL vs UA-WN & $0.151^{***}$ \\ \hline
UA-DL vs DL-WN & $0.0112^{***}$ \\ \hline
UA-WN vs DL-WN & $0.0551^{***}$ \\ \hline
\end{tabular}
\end{center}
\end{minipage}
%
\end{table}
The box plots for the entries of the network distance matrix for each pair of carriers are shown in Fig.~\ref{fig:box_2}. Consistent with Fig.~\ref{fig:distance_matrices_different_airlines}a, the AA and UA networks seem to be more similar to each other on average than the other pairs. We then carried out the Kruskal-Wallis test and found that there was a significant difference in these distributions ($p<0.001$). The Bonferroni-corrected Mann-Whitney U test as a post-hoc test revealed that there were significant differences in the $p$-values for all 15 pairs with various effect sizes (see Table~\ref{tab:Statistical_analysis_2} for the statistical results). In particular, the network distance values for AA-UA are smaller than those for any other carrier pairs with strong effect sizes ($r>0.5$). The network distance values for AA-DL are also smaller than the other carrier pairs except for the AA-UA pair with strong effect sizes ($r>0.5$). These results indicate that these two carrier pairs are more similar than the other carrier pairs.
To search for the origin of the observed similarities and dissimilarities between carrier pairs, we show the time series of the total number of edges shared by two carriers in each month and the number of such shared edges connecting to some major airports in Fig.~\ref{fig:common_edges}. Fig.~\ref{fig:common_edges}a indicates that the relatively large similarity in the entire observation period between AA and UA originates from the shared edges connecting to ORD (O'Hare International Airport), which has been a major hub for both AA and UA. Fig.~\ref{fig:common_edges}b indicates that AA and DL became more dissimilar around 2005 because DL closed a hub at DFW, which we also observed in Fig.~\ref{fig:change_networks}g. Figure~\ref{fig:common_edges}c indicates that AA and WN became more similar around 2015. This is because AA merged with US, one of whose major airports had been PHX, as we also observed in Fig.~\ref{fig:change_networks}b. Figure~\ref{fig:common_edges}d indicates that most of the edges that UA and DL share originate from the operations at LAX. Figure~\ref{fig:common_edges}e,f indicates that WN became more similar to UA and DL because WN increased the number of edges at DEN and ATL, which we also observed in Fig.~\ref{fig:change_networks}i. These results indicate that large part of the similarity in each carrier pair originates from operations at a single airport shared by the two carriers.
\begin{figure}[h]
\centering
\includegraphics[width=0.52\linewidth]{Fig11.pdf}
\caption{Distribution of the network distance for the pairs of carriers.}
\label{fig:box_2}
\end{figure}
\section*{Discussion}
We proposed a framework for analyzing evolution patterns in temporal air transport networks. We applied the proposed methods to the four largest carriers in the US, i.e., AA, UA, DL, and WN, between 1987 and 2019. We found that their overall evolution patterns were different. Specifically, our analysis suggested that WN changed the network structure more than the other carriers and that DL changed the network structure more than AA and UA. Their evolution patterns are also different in terms of the temporal variability and the strength of periodicity. Furthermore, we found that each carrier experienced abrupt changes in their network structure. We also found that the temporal variability and the strength of periodicity tended to increase in the latest years.
\begin{figure}[H]
\centering
\includegraphics[width=0.98\linewidth]{Fig12.pdf}
\caption{ Evolution of the number of edges shared by pairs of carriers. We show the total number of shared edges and the number of shared edges connecting to major airports for each pair of carriers. (a) American Airlines versus United Airlines, (b) American Airlines versus Delta Air Lines, (c) American Airlines versus Southwest Airlines, (d) United Airlines versus Delta Air Lines, (e) United Airlines versus Southwest Airlines, and (f) Delta Air Lines versus Southwest Airlines. ORD: O'Hare International Airport, DFW: Dallas/Fort Worth International Airport, PHX: Phoenix Sky Harbor International Airport, LAX: Los Angeles International Airport, DEN: Denver International Airport, and ATL: Hartsfield–Jackson Atlanta International Airport.}
\label{fig:common_edges}
\end{figure}
All the four carriers drastically changed their network structure in January 1995. In addition, they strengthened the periodicity after 1995. To the best of our knowledge, this change point is not related to any major socioeconomic and industrial events. The reason for this change point is unclear. We also showed that the three FSCs (i.e., AA, UA, and DL) substantially changed their network structure upon the post-bankruptcy mergers. On the other hand, we did not detect notable changes in the network structure upon the 9/11 attacks, the bankruptcies, and the exits from the bankruptcies, except for a change point for DL three months after the bankruptcy. There are different possible mechanisms that may have produced these results. First, these events may have triggered drastic changes in the network structure, but on a shorter time scale than a month. Analysis of weekly or daily snapshot networks may be fruitful; our methods are not selective of the time scale, and detailed flight data for the US carriers with a time resolution of minutes is available at the US Bureau of Transportation Statistics\cite{BTS2020air}. Second, only the number of flights between airports may have primarily changed in response to such events. Although we have assumed unweighted networks in the present article, if one modifies the network distance measure to accommodate weighted networks, the methods are directly applicable to the case of weighted temporal networks. These research directions warrant future work.
We showed that WN, which is an LCC, has shown particularly distinct patterns of evolution over years as compared to the FSCs, such as a consistent growth trend in terms of the number of nodes and edges, relatively fast changes in the network structure, and lack of periodicity when the data from all years are combined. To generalize these results to differences between LCCs and FSCs, one needs more thorough analyses of different carriers, such as LCCs and FSCs in Europe and Asia, where LCCs have been rapidly growing. We opted to study the US case in the present study because of the availability of the data.
Our data-analytic framework is also applicable to different contexts. For example, an application to airline alliances (e.g., Oneworld, SkyTeam, and Star Alliance) may reveal influences of entries and exits of carriers on the networks of airline alliances over years. Another direction is to study influences of natural disasters on the structure of air transport networks. For example, ash cloud caused by the volcanic eruption in Iceland in 2010 affected the structure of the European air transport networks\cite{wilkinson2012vulnerability, dunn2016increasing}. Another possible application is to study influences of epidemic outbreaks including the COVID-19 pandemic, which is ongoing as of mid 2020. Obviously, carriers have been slashing down flights (corresponding to the edge weight) and edges due to the drop in the demand. In addition to this, some carriers may be strategically maintaining their network structure to be as similar as possible, or perhaps the converse, to that before COVID-19. Some carriers may also be changing the similarity of their network to other carriers' networks. Finally, the proposed methods including the analysis of periodicity are not specific to air transportation networks such that they are applicable to temporal networks in general.
| 65ea244bc7db40e468932989b1cf942e5ccc118c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{Section:Introduction_IFB}
\input{Introduction_IFB.tex}
\section{Problem Formulation}
\label{Section:Problem_IFB}
\input{Problem_IFB.tex}
\section{Statement of the Main Results}
\label{Section:Main_IFB}
\input{Main_IFB_symmetric.tex}
\section{Proof of Theorem~\ref{THM:Capacity_IFB}: Outer Bounds}
\label{Section:Converse_IFB}
\input{Converse_IFB.tex}
\section{Proof of Theorem~\ref{THM:Ach_IFB}: Transmission Protocol}
\label{Section:Achievability_IFB}
\input{Achievability_IFB_March2020.tex}
\section{Proof of Theorem~\ref{THM:InnerBound_IFB}}
\label{Section:Achievability_InnerBound_IFB}
\input{Achievability_InnerBound_IFB.tex}
\section{Discussion on Other Regimes and Connection to Blind Index Coding}
\label{Section:Discussion_IFB}
The authors conjecture that the outer-bound region of Theorem~\ref{THM:Capacity_IFB} is also fact the capacity region of the two-user erasure broadcast channel with intermittent feedback considered in Theorem \ref{THM:InnerBound_IFB}. We believe the transmission strategy needs to be improved to provide a smooth transition between global delayed CSI scenario and the one-sided scenario. As we discussed briefly in Section~\ref{Section:Achievability_IFB}, the transmission strategy under the one-sided feedback assumption~\cite{ISIT19sclin} requires careful utilization of the available feedback in all phases of communication. However, in Case 2 of Theorem~\ref{THM:Ach_IFB} and in the proof of Theorem~\ref{THM:InnerBound_IFB}, we do not utilize feedback during certain phases of communications. The reason why the achievable rates do not degrade by ignoring feedback during Phase 3 of each iteration of Case~2 lies in the specific assumptions on channel parameters: we assume fully correlated feedback links and thus, transmitter wither has delayed feedback from both receivers or from no one.
The story is more complicated for Theorem~\ref{THM:InnerBound_IFB}. In the BIC phase, we have a BC in which a fraction of the bits for each receiver is available to the other one but the transmitter does not know which ones. However, different to the BIC problem in \cite{kao2016blind}, now we have additional (intermittent) feedback. To simplify the scheme, we provide signals intended for each receiver to both. However, using feedback one can indeed improve upon this scheme and improve the rates. Unfortunately, we face the following challenges: (1) the scheme becomes rather complex even if we assume after the BIC step feedback links are no longer intermittent; (2) even for erasure BIC without feedback, the capacity is not provided in \cite{kao2016blind}. Not to mention our more complicated setting. Thus, the BIC with feedback is an interesting yet challenging problem that needs to be solved before we can fully address the capacity region of the two-user erasure broadcast channel with intermittent feedback. Ideally, the transmission strategy should cover both extreme cases of DD and DN, and this is an ongoing part of our research.
\section{Conclusion}
\label{Section:Conclusion_IFB}
We developed new outer bounds on the capacity region of two-user erasure BCs with intermittent feedback. We also showed that these outer bounds are achievable under certain assumptions on channel parameters. The next step is to characterize the capacity region for all channel parameters. We conjecture the outer bounds are tight, and the region given in Theorem~\ref{THM:Capacity_IFB} is in fact the capacity region. However, we need to solve BIC with (intermittent) feedback as an intermediate step before settling this conjecture. An interesting future direction is consider the scenario in which receivers can encode the feedback messages, and find the minimum required feedback bandwidth to achieve the global feedback performance with only intermittent feedback.
Finally, in~\cite{vahid2016two}, two-user erasure interference channels~\cite{vahid2014capacity} with local delayed feedback were studied. There, it was shown that each transmitter must at least know the delayed CSI of the links connected to it in order to achieve the global delayed CSI performance. For distributed transmitters, understanding the capacity region of the two-user erasure interference channel with intermittent feedback would be the starting point.
\bibliographystyle{ieeetr}
| 3ba1b74024621c35d296a2321fbaac7fb0dd5c4b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Flamelet models are attractive for modelling some subgrid behavior with reasonable fidelity for large-eddy simulations \cite{Nguyen2018,Nguyen_WAS2018}. Due to prohibitive computational costs, the typical turbulent combustor analysis with its inherently multiscale behavior cannot use direct numerical simulation to describe the flame behavior. While some success has been achieved with flamelet modelling in analyzing experimental combustors, there is need for improvement in the basic model. Current practice in the sub-grid modelling uses flamelets with only a single diffusion flame. Yet, we know from experience that multi-branched flames can occur. We need therefore advancement of the theory to address multi-branched flames undergoing strain over a wide range of pressure. Merging of the branches under high strain rates is also of interest in practical applications.\hfill \break
\indent Flamelet studies provide insights about common behaviors that can be observed in more convoluted turbulent flames. Flamelets subjected to high strain rates have been studied with findings of multi-branched flames and also merging of branches of triple flames \cite{Nguyen2018,Nguyen_WAS2018}. Triple flames are described as tri-brachial structures composed by a fuel-lean premixed flame, a fuel-rich premixed flame, and a diffusion flame derived from the unburned reactants \cite{Phillips,Dold,Buckmaster_1988}. These flames are crucial in several combustion applications such as ignition and extinction processes of non-premixed systems \cite{Reveillon}.\hfill \break
Rajamanickam et al. numerically investigated the influence of stoichiometry on strained triple flames in counterflow configuration \cite{RAJAMANICKAM}. They used a one-step chemical kinetic model with constant density and constant transport properties. Their main conclusion was that the classical tri-branchial structure might not be observed in some cases, especially at high equivalence ratio. However, that study and the majority of studies involving multi-branched flames are performed using global chemical kinetic models \cite{Dold,Hartley_1991,KIONI, Ruetsch_1995,JIMENEZ_2007}, which might not be sufficient to predict an accurate behavior of these flames as shown by previous authors \cite{Law_2015_reviewer1}. Use of unity reaction orders for triple flame studies with global kinetics are also common and result in unrealistically symmetric triple flames \cite{KIONI,Ruetsch_1995,AJJ_WAS}. They are also unable to predict flame-branch merging accurately.\hfill \break
The current study, however, considers all the possible scenarios referred to the mixture ratio of the reactants that could lead to a multi-branched flame. In particular, three cases are considered: (i) lean premixed mixture coming from one side and non-reactive mixture with fuel coming from the other side; (ii) rich premixed mixture from one side and non-reactive mixture with oxygen from the other side; and (iii) lean premixed mixture coming from one side and rich premixed mixture coming from the other side.\hfill \break
Therefore, this work advances further with examination of the flames that co-exist at different mixture ratios and pressures when a variable normal strain rate is applied in a transverse direction to the flame. In contrast to previous studies, the numerical simulations herein employ a detailed chemical kinetic model. The counterflow-burner configuration is used since it has features relevant to the downstream behavior of triple flames.
\section{Model and analysis}
\subsection{Chemical kinetic model and software}
Results are obtained using the opposed-flow flame module (OPPDIF) from Chemkin-Pro\textsuperscript{\textregistered} software, for all studied cases of counterflow flames in axisymmetric configuration. Individual specific heats and enthalpies are calculated from temperature polynomial fits. Chemkin-Pro\textsuperscript{\textregistered} offers two detailed formulations to estimate the transport properties: mixture-averaged and multicomponent. The latter is deemed superior and chosen herein. The ideal gas law is used as the equation of state. For more details on the formulation for thermo-physical and transport properties, see the Chemkin-Pro\textsuperscript{\textregistered} theory manual \cite{Chemkin_manual}. \hfill \break
Two parameters control the adaptive grid refinement in Chemkin-Pro\textsuperscript{\textregistered} based on the solution curvature and gradient. They are both set to 0.5. Another two parameters control the solution convergence. They are the absolute and relative tolerances, which are set to 10$^{-9}$ and 10$^{-4}$, respectively.\hfill \break
The chemical kinetic model here is the most recent version of the San Diego Mechanism, which was last updated in 2018 to include two new reactions related to CHCHO production and consumption. The San Diego Mechanism includes a total of 270 reactions and 58 species \cite{San_Diego_mech}.
\subsection{Boundary conditions}
Boundary conditions are prescribed for the species mass fractions, temperature, pressure, and inlet velocities for each nozzle (Tables \ref{tab:Chemkin_Conditions} and \ref{tab:Chemkin_Cases}). The selected equivalence ratios ensure to deliver the outcomes for the interest regimes of this study, which are lean flame (Case 1), rich flame (Case 2) and both lean and rich flames (Case 3). The distance between the two nozzles is 2 cm. It is chosen based on the literature \cite{JORDA_2017}, to prevent flame proximity to the boundaries from affecting too heavily the solution -- i.e. zero gradients are desired at boundaries on species mass fraction and temperature curves -- while minimizing the computational cost.\hfill
\begin{table}[h]
\centering
\caption{Boundary conditions.}
\label{tab:Chemkin_Conditions}
\begin{tabular}{cccccl}
\cline{2-5}
& Inlet (left) & Inlet (right) & Strain {($S$, [}s$^{-1}${])} & \\ \cline{1-5}
& 0.1 & 0.1 & 10 & \\
Velocities {($V$, [}m/s{]}): 4 cases & 0.5 & 0.5 & 50 & \\
& 1 & 1 & 100 & \\
& 1.5 & 1.5 & 150 & \\
Temperature {($T$, [}K{]}) & 298 & 298 & -- & \\
Pressure ({$P$, [}atm{]}) & 1 up to 20 & 1 up to 20 & -- & \\ \cline{1-5}
\end{tabular}
\end{table}
The global strain rate $S$ is calculated following the expression $S$ = 2 $V_{left}/L$, where $V_{left}$ is the velocity at the left inlet nozzle, and $L$ is the distance between the two nozzles.
\begin{table}[ht!]
\centering
\caption{Species boundary conditions. Mass fractions and equivalence ratios ($\phi$).}
\label{tab:Chemkin_Cases}
\begin{tabular}{c|cc|cc|cc}
\cline{2-7}
& \multicolumn{2}{c|}{Case 1} & \multicolumn{2}{c|}{Case 2} & \multicolumn{2}{c}{Case 3} \\\cline{1-7}
Nozzle: & Left & Right & Left & Right & Left & Right \\
Mixture Type: & Fuel Lean & One Reactant & Fuel Rich & One Reactant & Fuel Lean & Fuel Rich \\
& $\phi$=0.34 & & $\phi$=5.70 & & $\phi$=0.34 & $\phi$=5.70
\\\hline
Methane (CH$_{4}$) & 0.02 & 0.25 & 0.25 & 0 & 0.02 & 0.25 \\
Nitrogen (N$_{2}$) & 0.7546 & 0.75 & 0.5775 & 0.77 & 0.7546 & 0.5775 \\
Oxygen (O$_{2}$) & 0.2254 & 0 & 0.1725 & 0.23 & 0.2254 & 0.1725 \\ \hline
\end{tabular}
\end{table}
\section{Results and Discussion}
This section is divided in three sub-sections that correspond to each of the three studied cases according to Table \ref{tab:Chemkin_Cases}.\hfill \break
The non-orthodox variable $\xi$ is employed to separate the heat-release-rate peaks while normalizing the x-axis (Eq. \ref{eq:x-axis}). In Equation \ref{eq:x-axis}, $s.p$ stands for \textit{stagnation plane} and $\chi$ is a dummy variable, used only as integration limit, and corresponding to distance measured from the left nozzle $x$. In the plots, the stagnation plane location is marked by a green dashed vertical line and placed at zero. This should help visualize the position of the flames with respect to the stagnation plane and the mixing-layer thickness (denoted as $\delta$ in Eq. \ref{eq:MLT}). The range of $\xi$ magnitudes in the horizontal axis of each plot is unity, making them comparable visually even though their stagnation planes are shifted.
\begin{equation}
\label{eq:x-axis}
\xi=\frac{\int_0^\chi Tdx}{\int_0^L Tdx}- \frac{\int_0^{s.p} Tdx}{\int_0^L Tdx}
\end{equation}
\begin{equation}
\label{eq:MLT}
\delta\simeq\sqrt{{D_T}/{S}}
\end{equation}
The mixing-layer thickness is estimated using the expression in Equation \ref{eq:MLT}, where the thermal diffusivity $D_T$ is taken from the reacting mixture (Cases 1 and 2) or from the average of thermal diffusivities of both reacting mixtures (Case 3). RefProp \cite{LEMMON-RP91} is employed to calculate the properties of each mixture at the different pressures of study. Doing so, the Prandtl number in all cases is determined to be between 0.7 and 0.72. The definition of $\delta$ is arbitrary and results will show that significant mixing for both premixed and non-premixed flame structures can occur outside the indicated bounds.
For the following discussion, it is important to consider Tables \ref{tab:Adiabatic_flame_T} and \ref{tab:Adiabatic_flame_T_flame_front_Case1}, which show the computed adiabatic flame temperatures at stoichiometric mixture ratio, as well as for the relevant equivalence ratios for this study.\hfill \break
\begin{table}[ht!]
\centering
\caption{Adiabatic flame temperatures [K] for relevant mixtures at various pressures with initial temperature of 298 K. Upstream equivalence ratios ($\phi$).}
\label{tab:Adiabatic_flame_T}
\begin{tabular}{ccccc}
\hline
\multicolumn{ 1}{c}{$P$ [atm]} & {\bf Stoichiometric mixture} &{\bf Case 1} & {\bf Case 2} \\
\multicolumn{ 1}{c}{} &$\phi$=1 & $\phi$=0.34 & $\phi$=5.70 \\
\hline
1 & 2224.54 & 1152.33 & 884.9 \\
\hline
10 & 2267.12 & 1152.33 & 991.45 \\
\hline
20 & 2276.6 & 1152.33 & 1026.86 \\
\hline
\end{tabular}
\end{table}
\begin{table}[ht!]
\centering
\caption{Case 1. Local equivalence ratios at the premixed flame front ($\phi$) and their corresponding adiabatic flame temperatures ($T_{\text{adiabatic}}$ [K]) calculated with initial temperature of 298 K. Temperatures at the premixed flame heat-release-rate peak ($T_{\text{premixed flame}}$ [K]).}
\label{tab:Adiabatic_flame_T_flame_front_Case1}
\begin{tabular}{c|cc|c|c|c|c}
\cline{2-7}
& \multicolumn{6}{c}{\textbf{Case 1}} \\ \hline
$P$ {[}atm{]} & \multicolumn{2}{c|}{1} & \multicolumn{2}{c|}{10} & \multicolumn{2}{c}{20} \\ \hline
$S$ {[}s$^{-1}${]} & \multicolumn{1}{c|}{10} & 50 & 10 & 50 & 10 & 50 \\ \hline
$\phi$ & \multicolumn{1}{c|}{0.17} & 0.22 & 0.21 & 0.15 & 0.20 & 0.17 \\ \hline
T$_{\text{adiabatic}}$ & \multicolumn{1}{c|}{758.3} & 882.7 & 851.6 & 712.1 & 820.1 & 760.8 \\ \hline
T$_{\text{premixed flame}}$ & \multicolumn{1}{c|}{1363.9} & 1351.3 & 1525.8 & 1601.6 & 1521.5 & 1653.1 \\ \hline
\end{tabular}
\end{table}
\begin{table}[ht!]
\centering
\caption{Case 2. Local equivalence ratios at the premixed flame front ($\phi$) and their corresponding adiabatic flame temperatures ($T_{\text{adiabatic}}$ [K]) calculated with initial temperature of 298 K. Temperatures at the premixed flame heat-release-rate peak ($T_{\text{premixed flame}}$ [K]).}
\label{tab:Adiabatic_flame_T_flame_front_Case2}
\begin{tabular}{c|cccccc}
\cline{2-7}
& \multicolumn{6}{c}{\textbf{Case 2}} \\ \hline
$P$ {[}atm{]} & \multicolumn{2}{c|}{1} & \multicolumn{2}{c|}{10} & \multicolumn{2}{c}{20} \\ \hline
$S$ {[}s$^{-1}${]} & \multicolumn{1}{c|}{10} & \multicolumn{1}{c|}{50} & \multicolumn{1}{c|}{10} & \multicolumn{1}{c|}{50} & \multicolumn{1}{c|}{10} & 50 \\ \hline
$\phi$ & \multicolumn{1}{c|}{3.23} & \multicolumn{1}{c|}{3.07} & \multicolumn{1}{c|}{3.58} & \multicolumn{1}{c|}{3.11} & \multicolumn{1}{c|}{4.14} & 3.34 \\ \hline
T$_{\text{adiabatic}}$ & \multicolumn{1}{c|}{980.5} & \multicolumn{1}{c|}{1009.6} & \multicolumn{1}{c|}{1065.3} & \multicolumn{1}{c|}{1102.1} & \multicolumn{1}{c|}{1077.5} & 1121.2 \\ \hline
T$_{\text{premixed flame}}$ & \multicolumn{1}{c|}{1577.0} & \multicolumn{1}{c|}{1726.2} & \multicolumn{1}{c|}{1528.3} & \multicolumn{1}{c|}{1654.7} & \multicolumn{1}{c|}{1168.0} & 1619.6 \\ \hline
\end{tabular}
\end{table}
\subsection{Case 1}
A fuel-lean mixture is injected from the left nozzle while diluted fuel enters the domain from the right side. Two flames are expected a priori: one premixed flame burning all the fuel from the left and one diffusion flame burning the leftover oxidizer from the left with the fuel from the right. Results are shown in Figure \ref{fig:SDMech_298K_case1_species}. Pressure and strain rate effects are described in the next two Subsections \ref{Subsection:case1_increaseP} and \ref{Subsection:case1_increaseS}.
\subsubsection{Pressure Effects}
\label{Subsection:case1_increaseP}
The heat release rate is taken as a marker of flame presence in the figures. The dominant flame, in terms of the peak value of heat release rate in Figure \ref{fig:SDMech_298K_case1_species}, is the premixed flame on the left rather than the diffusion flame at low pressures (upper plot). However, this relative importance switches at higher pressures for both strain rates studied.\hfill \break
From the same figure, comparing the different heat release rates at constant strain rate, it is seen that the heat-release-rate peaks are more distinct, even though the flames are getting closer. As pressure increases, the heat release peaks are higher and physically closer together. This behavior is expected since the combustion rate is enhanced when increasing pressure and it is in agreement with recent analysis using one-step Westbrook-Dryer kinetics for propane \cite{WAS_WSCI2019}.\hfill \break
Note that an endothermic region is found in most cases. A plausible hypothesis based on the review of that region (Figure \ref{fig:SDMech_case1_S25_P20_speciesednothermic}) is that, under certain conditions, methane converts to methyl (CH$_3$), which might recombine to form ethane (C$_2$H$_6$) and then, ethane dehydrogenates to ethylene (C$_2$H$_4$). This is a highly endothermic process \cite{pyrolysis_larkins} that is only considered in detailed chemical kinetic models and explains the endothermic region found in these simulations. Similar trends and a characteristic endothermic region were observed when simulating this case using the DRM19 chemical kinetic model (84 reactions, 19 species \cite{DRM19}). The results shown using that model (Figure \ref{fig:DRM19_298K_case1_species}) at lower pressures at these strain rates qualitatively support the features that were obtained using the San Diego Mechanism. Therefore, this endothermic region is not restricted to the San Diego mechanism and it is expected to appear when detailed chemistry is employed.\hfill \break
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{Fig2_SDMech_S5_P1_Case1.eps}\quad
\includegraphics[width=.45\textwidth]{Fig2_SDMech_S25_P1_Case1.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig2_SDMech_S5_P10_Case1.eps}\quad
\includegraphics[width=.45\textwidth]{Fig2_SDMech_S25_P10_Case1.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig2_SDMech_S5_P20_Case1.eps}\quad
\includegraphics[width=.45\textwidth]{Fig2_SDMech_S25_P20_Case1.eps}
\caption{Comparison between two different strain rates for Case 1 at $1\:\text{atm}$, $10\:\text{atm}$ and $20\:\text{atm}$. $S = 10\text{\:s}^{-1}$ (left) and $50\text{\:s}^{-1}$ (right). San Diego Mechanism. Mole fractions of CH$_4$, O$_2$ and N$_2$ (black) and heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case1_species}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{Fig3_SDMech_S25_P20_case1_speciescomparison_SI1.eps}\quad
\includegraphics[width=.45\textwidth]{Fig3_SDMech_S25_P20_case1_speciescomparison_SI2.eps}\quad
\includegraphics[width=.45\textwidth]{Fig3_SDMech_S25_P20_case1_speciescomparison_SI3.eps}
\caption{Result for Case 1 with $S = 50\text{\:s}^{-1}$ at 20 atm. San Diego Mechanism. Mole fractions of CH$_3$, C$_2$H$_6$ and C$_2$H$_4$ (black). Heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_case1_S25_P20_speciesednothermic}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics [width=.45\textwidth]{Fig4_DRM_S5_P10_Case1_SI.eps}\quad
\includegraphics [width=.45\textwidth]{Fig4_DRM_S25_P10_Case1_SI.eps}
\caption{Comparison between two different strain rates for Case 1 at $10\:\text{atm}$. $S = 10\text{\:s}^{-1}$ (left) and $50\text{\:s}^{-1}$ (right). DRM19 model. Mole fractions of CH$_4$, O$_2$ and N$_2$ (black) and heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:DRM19_298K_case1_species}
\end{figure}
\begin{table}[ht!]
\centering
\caption{Characteristics of the heat-release-rate ($Q$) peaks for Cases 1 and 2. $T$ stands for temperature [K], $P$ for pressure [atm], and $S$ to strain rate [$\text{\:s}^{-1}$].}
\label{tab:Characteristics_flames_Case1_Case2}
\begin{tabular}{c|c|c|c|c|c}
\hline
Case & $P$ & $S$ & $\xi$ at $Q$ peak & Character of $Q$ peak & $T$ at $Q$ peak \\
\hline
\multicolumn{ 1}{c|}{} & \multicolumn{1}{|c|}{1} & \multicolumn{ 1}{|c|}{10} & -0.586 & Lean premixed flame with diffusive character & 1363.90 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.277 & Diffusion flame & 1945.35 \\
\cline{3-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{50} & -0.409 & Lean premixed flame with diffusive character & 1351.28 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.209 & Diffusion flame & 1896.72 \\
\cline{2-6}
\multicolumn{ 1}{c|}{1} & \multicolumn{ 1}{|c|}{10} & \multicolumn{ 1}{|c|}{10} & -0.339 & Lean premixed flame with diffusive character & 1525.77 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.210 & Diffusion flame & 2100.02 \\
\cline{3-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{50} & -0.181 & Lean premixed flame with diffusive character & 1601.63 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.114 & Diffusion flame & 2042.00 \\
\cline{2-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{20} & \multicolumn{ 1}{|c|}{10} & -0.264 & Lean premixed flame with diffusive character & 1521.45 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.162 & Diffusion flame & 2124.70 \\
\cline{3-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{50} & -0.135 & Lean premixed flame with diffusive character & 1653.12 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.093 & Diffusion flame & 2073.30 \\
\hline
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{1} & \multicolumn{ 1}{|c|}{10} & -0.440 & Rich premixed flame with diffusive character & 1576.99 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.240 & Diffusion flame & 2010.19 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.155 & Exothermic reactions (\textit{no flame}) & 2009.53 \\
\cline{3-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{50} & 0.039 & Rich premixed flame with diffusive character & 1726.23 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & 0.150 & Diffusion flame & 2040.70 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & 0.232 & Exothermic reactions (\textit{no flame}) & 1953.67 \\
\cline{2-6}
\multicolumn{ 1}{c|}{2} & \multicolumn{ 1}{|c|}{10} & \multicolumn{ 1}{|c|}{10} & -0.310 & Rich premixed flame with diffusive character& 1528.33 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.106 & Diffusion flame & 2217.03 \\
\cline{3-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{50} & 0.005 & Rich premixed flame with diffusive character & 1654.69 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & 0.107 & Diffusion flame & 2259.32 \\
\cline{2-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{20} & \multicolumn{ 1}{|c|}{10} & -0.305 & Rich premixed flame with diffusive character & 1168.04 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & -0.078 & Diffusion flame & 2253.03 \\
\cline{3-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{50} & -0.002 & Rich premixed flame with diffusive character & 1619.59 \\
\cline{4-6}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & \multicolumn{ 1}{|c|}{} & 0.085 & Diffusion flame & 2308.72 \\
\hline
\end{tabular}
\end{table}
\subsubsection{Strain Rate Effects}
\label{Subsection:case1_increaseS}
In all pressure and strain rate conditions, two flames can be observed. Both flames are placed on the left of the stagnation plane and the flame on the left becoming dominant over the diffusion flame at low pressures. Our results indicate that increasing strain rate at constant pressure favors flame merging, especially at low pressures. Therefore, lower strain rates allow easier distinction between existing flame peaks. \hfill \break
Analysing the character of these flames in depth, it is concluded that the flame expected to be premixed (left) does not follow the classical premixed characteristics where flame peak temperature is lower than adiabatic flame temperature, and velocity remains constant at constant pressure when increasing strain rate. Rather, the temperature of this flame is over the adiabatic flame temperature expected for the upstream equivalence ratio ($\phi$ = 0.34, see Tables \ref{tab:Adiabatic_flame_T} and \ref{tab:Adiabatic_flame_T_flame_front_Case1}). This is an indication that the premixed flame is diffusion controlled. The possibility that this high flame temperature is due to a variation on the local equivalence ratio with respect to its upstream value has also been explored. Figure \ref{fig:SDMech_298K_case1_phi} shows equivalence ratio as a function of $\xi$ at $1\:\text{atm}$ and $S = 10\text{\:s}^{-1}$. The heat release rate has been normalized by its maximum for graphical purposes. The local equivalence ratio at the premixed flame front is $\phi$ = 0.17, with a corresponding adiabatic flame temperature of 758.3 K, which is still lower than the premixed flame temperature reported in Table \ref{tab:Adiabatic_flame_T_flame_front_Case1}. Similar results are found for all the pressures and strain rates conditions considered in Case 1 (see Table \ref{tab:Adiabatic_flame_T_flame_front_Case1}). Hence, the conclusion that the premixed flame behaves as a diffusion flame still holds.\hfill \break
\begin{figure}[h!]
\centering
\includegraphics[width=.75\textwidth]{Lean_SDMech_S5_P1_Case1_phi.eps}
\caption{Case 1 at $S = 10\text{\:s}^{-1}$ and $1\:\text{atm}$. San Diego Mechanism. Mole fractions of CH$_4$, N$_2$ and O$_2$ (black), normalized heat release rate (orange), and local equivalence ratio (blue). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case1_phi}
\end{figure}
Additionally, Figure \ref{fig:SDMech_298K_case1_densityvelocity} and Table \ref{tab:SDMech_densityvelocity} show that at constant pressure, the velocity at the premixed flame front does not remain constant when varying the strain rate, but rather increases with it. These observations indicate that as strain rate increases, the distance between flames decreases, leading to steeper gradients in the axial direction for the reactants mass-fraction curves and temperature profile. Consequently, diffusion rates are enhanced. In other words, increasing the strain rate enhances reactants diffusion mass transport towards the diffusion flame, producing more heat, which in turn is diffused in the negative x-direction at a faster rate towards the premixed flame. The latter also produces more heat - highlighted by its peak temperature being higher than its corresponding adiabatic flame temperature- and is able to sustain higher velocities than it would if it behaved like a typical premixed flame. \hfill \break
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{Fig11_Lean_SDMech_S5_v_P1_T_dx_MLT_densityvelocity.eps}\quad
\includegraphics[width=.45\textwidth]{Fig11_Lean_SDMech_S25_v_P1_T_dx_MLT_densityvelocity.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig11_Lean_SDMech_S5_v_P10_T_dx_MLT_densityvelocity.eps}\quad
\includegraphics[width=.45\textwidth]{Fig11_Lean_SDMech_S25_v_P10_T_dx_MLT_densityvelocity.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig11_Lean_SDMech_S5_v_P20_T_dx_MLT_densityvelocity.eps}\quad
\includegraphics[width=.45\textwidth]{Fig11_Lean_SDMech_S25_v_P20_T_dx_MLT_densityvelocity.eps}
\caption{Comparison between two different strain rates for Case 1 at $1\:\text{atm}$, $10\:\text{atm}$ and $20\:\text{atm}$. $S = 10\text{\:s}^{-1}$ (left) and $50\text{\:s}^{-1}$ (right). Velocity ($v$) and density ($\rho$) times velocity are plotted. San Diego Mechanism. Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case1_densityvelocity}
\end{figure}
\begin{table}[ht!]
\centering
\caption{Flame front velocities [cm/s] at different pressure ($P$) and strain rate ($S$) conditions for Cases 1 and 2.}
\label{tab:SDMech_densityvelocity}
\begin{tabular}{c|c|c|c}
\centering
&\backslashbox{$P$ [atm]}{$S$ [s$^{-1}$]}&{10}&{50}\\\cline{1-4}
& 1 &17.0&33.01\\\cline{2-4}
Case 1 & 10 &4.76&10.16\\\cline{2-4}
& 20 &3.18&7.147\\\hline
& 1 &3.98&9.02\\\cline{2-4}
Case 2 & 10 &1.90&3.89\\\cline{2-4}
& 20 &1.69&3.03\\\hline
\end{tabular}
\end{table}
\subsection{Case 2}
In this case, a fuel-rich mixture is injected from the left nozzle while air enters the domain from the right side. Here, two flames are expected: a premixed flame burning the fuel stoichiometrically from the left and a diffusion flame burning the leftover fuel from the left with the air from the right. However, due to the same reasons discussed in Case 1, the premixed fuel-lean flame is also diffusion controlled. Its peak temperature is also higher than the corresponding adiabatic flame temperature at the in-flowing mixture ratio, and the velocity at the flame front increases with strain rate at constant pressure. Solutions for this case are portrayed in Figure \ref{fig:SDMech_298K_case2_species} and more details follow in the next Subsections \ref{Subsection:case2_increaseP} and \ref{Subsection:case2_increaseS}.
\subsubsection{Pressure Effects}
\label{Subsection:case2_increaseP}
As in the previous case, Figure \ref{fig:SDMech_298K_case2_species} shows that increasing pressure enhances the combustion; therefore, higher heat release rates are obtained. The diffusion flame (right heat-release-rate peak) is dominant for all pressures and strain rates. \hfill \break
At low pressure ($1\:\text{atm}$) and especially at high strain rate, merged flames are observed which become distinct and branch into two co-existing flames when increasing pressure and maintaining constant strain rate. Figure \ref{fig:SDMech_298K_case2_species} also shows that depending on the pressure and strain, the diffusion flame (right) possesses two local peaks in heat release rate: one at $\xi$ = - 0.240 and the other at $\xi$ = - 0.155 ($1\:\text{atm}$ and $10\text{\:s}^{-1}$), while the peak at $\xi$ = - 0.440 corresponds to the premixed fuel-rich flame (see Table \ref{tab:Characteristics_flames_Case1_Case2}). There is only one peak in temperature (see Figure \ref{fig:SDMech_case2_S5_P1_temperature}). In the local absence of methane, the heat-release-rate peak at $\xi$ = - 0.155 is mostly produced by reactions \ref{reac:Production_CO2} and \ref{reac:Consumption_H2}, which are highly exothermic (i.e. negative reaction enthalpy $\Delta H_{\ce{R}}$) and displaced to the formation of products.
\begin{chequation}
\begin{align}
\ce{CO + OH <=>> CO2 + H $\quad \Delta H_{\ce{R}} = -320.443 \;$ kJ $\;$ mol^{-1}} \label{reac:Production_CO2}\\
\ce{H2 + OH <=>> H2O + H $\quad \Delta H_{\ce{R}} = -279.328 \;$ kJ $\;$ mol^{-1} } \label{reac:Consumption_H2}
\end{align}
\end{chequation}
Figures \ref{fig:SDMech_case2_S5_P1_temperature} and \ref{fig:SDMech_case2_S5_P1_speciesdetail} show that the maximum temperature and concentration of CO$_2$ and H$_2$O occur at the valley between $\xi$ = - 0.240 and $\xi$ = - 0.155. However, concentrations of CO and H$_2$ peak in the region between the premixed rich flame and the diffusion flame. These two species are not fully depleted across the left peak of the diffusion flame, hence enabling reactions \ref{reac:Production_CO2} and \ref{reac:Consumption_H2} to be active in the right peak of the same flame. The importance of certain reaction pathways at the locations of the three reaction peaks are highlighted in Figures \ref{fig:SDMech_case2_S5_P1_pathways_C} and \ref{fig:SDMech_case2_S5_P1_pathways_H}. Bolder lines imply a more important pathway. Thus, it is expected that the pathways that relate to the premixed fuel-rich flame (Subfigure \ref{fig:SDMech_case2_S5_P1_pathways_a_C}, $\xi$ = - 0.440) are thicker than the ones presented for the diffusion flame (Subfigure \ref{fig:SDMech_case2_S5_P1_pathways_b_C}, $\xi$ = - 0.240) or when there is no flame (Subfigures \ref{fig:SDMech_case2_S5_P1_pathways_c_C} and \ref{fig:SDMech_case2_S5_P1_pathways_c_H}, $\xi$ = - 0.155). Moreover, Subfigures \ref{fig:SDMech_case2_S5_P1_pathways_c_C} and \ref{fig:SDMech_case2_S5_P1_pathways_c_H} further support that reactions \ref{reac:Production_CO2} and \ref{reac:Consumption_H2} are the cause for the right peak of the diffusion flame, since pathways that lead to reactions \ref{reac:Production_CO2} and \ref{reac:Consumption_H2} are represented by darker and thicker arrows. Figure \ref{fig:SDMech_case2_S5_P1_pathways_sensitivity_analysis} shows the contribution of the most important reactions on a linear scale to the absolute rate of production of H$_2$ at $\xi$ = -0.155, clarifying that the consumption of H$_2$ is mostly given by reaction \ref{reac:Consumption_H2}. The negative sign in the values at the left side of the vertical line indicates the reverse reaction rate while the positive values at the right side indicates that it is a forward reaction rate. The remainder of the reactions that are not shown in Figure \ref{fig:SDMech_case2_S5_P1_pathways_sensitivity_analysis} do not have a significant contribution to the production or consumption of H$_2$.\hfill \break
\begin{figure}[h!]
\centering
\includegraphics [width=.45\textwidth]{Fig5_SDMech_S5_P1_Case2}\quad
\includegraphics [width=.45\textwidth]{Fig5_SDMech_S25_P1_Case2}
\medskip
\includegraphics [width=.45\textwidth]{Fig5_SDMech_S5_P10_Case2}\quad
\includegraphics [width=.45\textwidth]{Fig5_SDMech_S25_P10_Case2}
\medskip
\includegraphics [width=.45\textwidth]{Fig5_SDMech_S5_P20_Case2}\quad
\includegraphics [width=.45\textwidth]{Fig5_SDMech_S25_P20_Case2}
\caption{Comparison between two different strain rates for Case 2 at $1\:\text{atm}$, $10\:\text{atm}$ and $20\:\text{atm}$. $S = 10\text{\:s}^{-1}$ (left) and $50\text{\:s}^{-1}$ (right). San Diego Mechanism. Mole fractions of CH$_4$, O$_2$ and N$_2$ (black) and heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case2_species}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{Rich_SDMech_S5_v_P1_T_SI}
\caption{Case 2 at $S = 10\text{\:s}^{-1}$ and $1\:\text{atm}$. San Diego Mechanism. Temperature (black) and heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_case2_S5_P1_temperature}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{Rich_SDMech_S5_v_P1_T_dx_MLT_species_SI2}\quad
\includegraphics [width=.45\textwidth]{Rich_SDMech_S5_v_P1_T_dx_MLT_species_SI3}
\caption{Case 2 at $S = 10\text{\:s}^{-1}$ and $1\:\text{atm}$. San Diego Mechanism. Mole fractions of CO (green), CO$_2$ (blue), H$_2$ (black), and H$_{2}$O (magenta) and heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_case2_S5_P1_speciesdetail}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.55\textwidth}
\caption{}
\includegraphics[width=.50\textwidth]{case2_s5_1atm_x0dot857143_Cpathway}
\label{fig:SDMech_case2_S5_P1_pathways_a_C}
\end{subfigure}
\begin{subfigure}[b]{0.7\textwidth}
\caption{}
\includegraphics[width=.78\textwidth]{case2_s5_1atm_x1dot026786_Cpathway}
\label{fig:SDMech_case2_S5_P1_pathways_b_C}
\end{subfigure}
\begin{subfigure}[b]{0.8\textwidth}
\caption{}
\includegraphics[width=.52\textwidth]{case2_s5_1atm_x1dot089286_Cpathway}
\label{fig:SDMech_case2_S5_P1_pathways_c_C}
\end{subfigure}
\caption{Main reaction pathways for species containing carbon at the three locations where reaction rate has a local maximum, (a) $\xi$ = -0.440 (b) $\xi$ = -0.240 (c) $\xi$ = -0.155. Case 2 with $S = 10\text{\:s}^{-1}$ at $1\:\text{atm}$. San Diego Mechanism. Thickness of the arrows represent the importance of the reaction pathway. Generated with ANSYS Chemkin-Pro\textsuperscript{\textregistered} Reaction Path Analyzer (RPA) \cite{Chemkin_RPA_manual}.}
\label{fig:SDMech_case2_S5_P1_pathways_C}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.55\textwidth}
\caption{}
\includegraphics[width=1.2\textwidth]{case2_s5_1atm_x0dot857143_Hpathway}
\label{fig:SDMech_case2_S5_P1_pathways_a_H}
\end{subfigure}
\begin{subfigure}[b]{0.7\textwidth}
\caption{}
\includegraphics[width=.90\textwidth]{case2_s5_1atm_x1dot026786_Hpathway}
\label{fig:SDMech_case2_S5_P1_pathways_b_H}
\end{subfigure}
\begin{subfigure}[b]{0.6\textwidth}
\caption{}
\includegraphics[width=.90\textwidth]{case2_s5_1atm_x1dot089286_Hpathway}
\label{fig:SDMech_case2_S5_P1_pathways_c_H}
\end{subfigure}
\caption{Main reaction pathways for species containing hydrogen at the three locations where reaction rate has a local maximum, (a) $\xi$ = -0.440 (b) $\xi$ = -0.240 (c) $\xi$ = -0.155. Case 2 with $S = 10\text{\:s}^{-1}$ at $1\:\text{atm}$. San Diego Mechanism. Thickness of the arrows represent the importance of the reaction pathway. Generated with ANSYS Chemkin-Pro\textsuperscript{\textregistered} Reaction Path Analyzer (RPA) \cite{Chemkin_RPA_manual}.}
\label{fig:SDMech_case2_S5_P1_pathways_H}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.75\textwidth]{PNG_SENSITYVITY_H2_x1dot089}
\caption{Contribution of each reaction on a linear scale to the absolute rate of production of H$_2$ at $\xi$ = -0.155. Case 2 with $S = 10\text{\:s}^{-1}$ at $1\:\text{atm}$. San Diego Mechanism. Rate values at the right and left of the vertical line represent the production and consumption of H$_2$ [mol/(cm$^3$ s)], respectively.}
\label{fig:SDMech_case2_S5_P1_pathways_sensitivity_analysis}
\end{figure}
\subsubsection{Strain Rate Effects}
\label{Subsection:case2_increaseS}
Figure \ref{fig:SDMech_298K_case2_species} shows that lowering strain rate separates the different flame branches, while increasing it at constant pressure causes merger. The effect of strain rate is better observed at low pressure ($1\:\text{atm}$) and agrees with the observations from Sirignano \cite{WAS_WSCI2019}.\hfill \break
Variations in strain rate have a major impact on the location and the character of the flames. In this Case 2, as in the previous case and contrary to expectation, the character of the premixed flame does not follow the classical premixed flame behavior for any of the conditions studied. The temperature of this flame is, again, higher than the adiabatic flame temperature expected for the equivalence ratio ($\phi$)= 5.70 (see Tables \ref{tab:Adiabatic_flame_T} and \ref{tab:Characteristics_flames_Case1_Case2}). However, similarly to Case 1, Figure \ref{fig:SDMech_298K_case2_phi} shows that the equivalence ratio varies locally due to a local mixture fraction change. Also in Case 2, the adiabatic flame temperature corresponding to the local equivalence ratio at the premixed flame front ($\phi$ = 3.23) is 980.5K for 1 atm, which is inferior to the flame temperature. The same trend is observed for all the pressures and strain rates studied for Case 2 (see Table \ref{tab:Adiabatic_flame_T_flame_front_Case2}).\hfill \break
\begin{figure}[h!]
\centering
\includegraphics[width=.75\textwidth]{Rich_SDMech_S5_P1_Case2_phi.eps}
\caption{Case 2 at $S = 10\text{\:s}^{-1}$ and $1\:\text{atm}$. San Diego Mechanism. Mole fractions of CH$_4$, N$_2$ and O$_2$ (black), normalized heat release rate (orange), and local equivalence ratio (blue). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case2_phi}
\end{figure}
Figure \ref{fig:SDMech_298K_case2_densityvelocity} and Table \ref{tab:SDMech_densityvelocity} also show that, similarly to Case 1, the velocity does not remain constant when varying the strain rate. Analogous examination to that for Case 1 in Subsection \ref{Subsection:case1_increaseS} is followed. Consequently, similar to Case 1, the premixed flame is heat-diffusion controlled. Production of active radicals -- such as H, O, and OH -- goes along with the heat production. These radicals could be transported upstream by mass diffusion and also contribute to the temperature increment \cite{Gomez_reviewer1}. Nevertheless, literature has shown that heat diffusion alone is sufficient to cause that effect \cite{WAS_WSCI2019}. The observed diffusive character is strengthened for the higher strain rate ($50\text{\:s}^{-1}$) cases, where both flames are within the mixing-layer boundaries. For all pressures at high strain rate, the left flame falls on the stagnation plane while the right flame is placed on the right side. This implies once again that the premixed left flame has diffusion character even though it receives both reactants from a single side. \hfill \break
At low strain rate ($10\:\text{s}^{-1}$), the purely diffusion flame (right flame) is found just where the mixing-layer ``edge" is. The edge point is based only on an order-of-magnitude estimate and the flame actually lies in the mixing layer. All the remaining methane that passes through the premixed flame without burning is consumed in this flame. For all pressures at low strain rate, both flames are placed to the left side of the stagnation plane. \hfill \break
\begin{figure}[h!]
\centering
\includegraphics[width=.45\textwidth]{Fig12_Rich_SDMech_S5_v_P1_T_dx_MLT_densityvelocity.eps}\quad
\includegraphics[width=.45\textwidth]{Fig12_Rich_SDMech_S25_v_P1_T_dx_MLT_densityvelocity.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig12_Rich_SDMech_S5_v_P10_T_dx_MLT_densityvelocity.eps}\quad
\includegraphics[width=.45\textwidth]{Fig12_Rich_SDMech_S25_v_P10_T_dx_MLT_densityvelocity.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig12_Rich_SDMech_S5_v_P20_T_dx_MLT_densityvelocity.eps}\quad
\includegraphics[width=.45\textwidth]{Fig12_Rich_SDMech_S25_v_P20_T_dx_MLT_densityvelocity.eps}
\caption{Comparison between two different strain rates for Case 2 at $1\:\text{atm}$, $10\:\text{atm}$ and $20\:\text{atm}$. $S = 10\text{\:s}^{-1}$ (left) and $50\text{\:s}^{-1}$ (right). Velocity ($v$) and density ($\rho$) times velocity are plotted. San Diego Mechanism. Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case2_densityvelocity}
\end{figure}
\subsection{Case 3}
A fuel-lean mixture is injected from the left nozzle while a fuel-rich mixture enters the domain from the right side. Three flames are expected in this case: one premixed flame burning the fuel stoichiometrically from the right, one premixed flame burning all the fuel from the left, and one diffusion flame in the middle burning the leftover fuel and air coming from the right and left sides, respectively.\hfill \break
For this more complex case, two extra strain-rate values have been studied ($100\:\text{s}^{-1}$ and $150\:\text{s}^{-1}$) and the corresponding pressure effects can be seen in Figures \ref{fig:SDMech_298K_case3_species} and \ref{fig:SDMech_298K_case3_S50_S75}.
\subsubsection{Pressure Effects}
As described for Cases 1 and 2, flames become more distinct from each other as pressure is increased. This effect can be observed for all strain rates, but it manifests best at the highest strain rate. In the cases when $S = 100\text{\:s}^{-1}$ and $150\text{\:s}^{-1}$, only two flames are observed at low pressure, being a premixed fuel-lean diffusion character flame (left) and a diffusion flame (right). Notice that Figure \ref{fig:SDMech_298K_case3_S50_S75} clearly shows that the flame on the right has a diffusive character since reactants arrive from opposite sides of the domain and get consumed where the heat-release-rate peaks. Therefore, it is concluded that the expected premixed fuel-rich flame is merged with the diffusion flame. \hfill \break
The character of these flames can be classified, from left to right, as fuel-lean premixed and diffusion controlled, diffusion, and fuel-rich having both diffusion and premixed character (see Table \ref{tab:Characteristics_flames_Case3} for more details). For most configurations, the peak heat release rate is higher from left to right, with the fuel rich flame being the weakest of the three (see Figures \ref{fig:SDMech_298K_case3_species} and \ref{fig:SDMech_298K_case3_S50_S75}). However, at the highest pressure for the lower strain rates of $10\:\text{s}^{-1}$ and $50\:\text{s}^{-1}$ (Figure \ref{fig:SDMech_298K_case3_species}), the diffusion-flame reaction peak is the highest. See Figure \ref{fig:SDMech_case3_maxQ_differentSP} where this is highlighted; there is a crossover at a given strain rate where the lean-premixed flame switches from weaker to stronger compared to the diffusion flame. This crossover point occurs at higher strain rate with increasing pressure. Residence time for the diffusion flame is proportional to the reciprocal of strain rate. The diffusion flame weakens with decreasing residence time since a certain time is needed for complete reaction. That being said, the integral of the heat release rate versus the axial coordinate shown in Figure \ref{fig:SDMech_298K_case3_Q_int} clearly highlights how the jump corresponding to the diffusion flame is larger than those related to the other heat-release-rate peaks. Figure \ref{fig:SDMech_298K_case3_Q_int} also shows that the total heat release for the diffusion flame is dominant. Furthermore, profiles of mole fractions (Figure \ref{fig:SDMech_298K_case3_species}) indicate that the mass diffusion rate on major species at the premixed flame is determined by the demand created by the diffusion flame. Therefore, the diffusion flame produces more heat even for the cases where the fuel-lean premixed flame heat-release-rate peak is higher. That is explained as follows: for the case with a mass fraction of 0.02 for the fuel to be burned in the lean flame, only an O$_2$ fraction of a 0.04 will be consumed there, leaving a mass fraction of about 0.185 to be consumed in the diffusion flame. \hfill \break
Examining the case at the lowest pressure and strain rate, the lean and rich premixed flames are represented by the heat-release-rate peaks placed at $\xi$ = - 0.612 and $\xi$ = - 0.045, respectively. The premixed flame establishes its residence time based on pressure and mixture ratio without regard to strain rate. It simply relocates its position so that its propagation velocity matches the incoming stream velocity. The heat-release-rate peak observed at $\xi$ = - 0.340 refers to the zone where exothermic reactions \ref{reac:Production_CO2} and \ref{reac:Consumption_H2} are occurring. Therefore, the region from $\xi$ = - 0.340 to the end of the domain shows similar behavior to the one explained for Case 2 in Section \ref{Subsection:case2_increaseP}. In this case, this heat-release-rate peak is placed at the left side of the rich premixed flame peak and on the right of the diffusion flame peak, since the rich premixed mixture comes from the left side of the domain - i.e., opposite side than in Case 2.
\begin{center}
\LTcapwidth=\textwidth
\begin{longtable}{c|c|c|c|c}
\caption{Characteristics of the heat-release-rate ($Q$) peaks for Case 3. $T$ stands for temperature [K], $P$ for pressure [atm], and $S$ to strain rate [$\text{\:s}^{-1}$].}
\label{tab:Characteristics_flames_Case3}\\
\hline $P$ [atm] & $S$ [$\text{\:s}^{-1}$] & $\xi$ at $Q$ peak & Character of $Q$ peak & $T$ at $Q$ peak [K]\\ \hline
\endfirsthead
\multicolumn{4}{c}%
{{\bfseries \tablename\ \thetable{} -- continued from previous page}} \\
\hline $P$ & $S$ & $\xi$ at $Q$ peak & Character of $Q$ peak & $T$ at $Q$ peak [K] \\ \hline
\endhead
\hline \multicolumn{4}{|r|}{{Continued on next page}} \\ \hline
\endfoot
\hline \hline
\endlastfoot
\multicolumn{1}{c|}{1} & \multicolumn{ 1}{|c|}{10} & -0.612 & Lean premixed flame with diffusive character& 1364.42 \\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.340 & Exothermic reactions (\textit{no flame})&2067.00\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.245 & Diffusion flame & 2051.22\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.045 & Rich premixed flame with diffusive character&1606.51\\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{50} & -0.436 & Lean premixed flame with diffusive character&1380.28 \\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.187 & Diffusion flame&2007.65\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.059 & Rich premixed flame with diffusive character&1670.89\\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{100} & -0.389 & Lean premixed flame with diffusive character& 1350.18\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.186 & Diffusion flame & 1979.44\\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{150} & -0.361 & Lean premixed flame with diffusive character& 1369.71\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.192 &Diffusion flame& 1960.90\\
\cline{1-5}
\multicolumn{ 1}{c|}{10} & \multicolumn{ 1}{|c|}{10} & -0.376 & Lean premixed flame with diffusive character& 1526.31\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.230 & Diffusion flame &2255.97\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & 0.000 & Rich premixed flame with diffusive character&1520.22\\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{50} & -0.210 & Lean premixed flame with diffusive character &1593.89 \\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.135 &Diffusion flame & 2193.047\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.019 & Rich premixed flame with diffusive character&1626.38\\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{100} & -0.177 &Lean premixed flame with diffusive character & 1633.23\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.126 & Diffusion flame& 2155.22\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.047 & Rich premixed flame with diffusive character &1710.69 \\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{150} & -0.161 & Lean premixed flame with diffusive character& 1643.07\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.120 &Diffusion flame & 2132.86\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.058 & Rich premixed flame with diffusive character & 1745.86\\
\cline{1-5}
\multicolumn{ 1}{c|}{20} & \multicolumn{ 1}{|c|}{10} & -0.304 & Lean premixed flame with diffusive character&1519.78 \\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.188 &Diffusion flame & 2283.75\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & 0.063 & Rich premixed flame with diffusive character & 1134.40\\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{50} & -0.160 &Lean premixed flame with diffusive character & 1645.40\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.106 &Diffusion flame & 2232.40 \\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.014 & Rich premixed flame with diffusive character&1642.81 \\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{100} & -0.134 &Lean premixed flame with diffusive character & 1674.86\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.096 & Diffusion flame & 2206.48\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.032 & Rich premixed flame with diffusive character & 1664.28 \\
\cline{2-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{150} & -0.121 &Lean premixed flame with diffusive character &1698.14 \\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.092 & Diffusion flame& 2182.94\\
\cline{3-5}
\multicolumn{ 1}{c|}{} & \multicolumn{ 1}{|c|}{} & -0.042 & Rich premixed flame with diffusive character & 1713.78\\
\hline
\end{longtable}
\end{center}
\begin{figure}[ht!]
\centering
\includegraphics[width=.45\textwidth]{Fig7_SDMech_S5_P1_Case3.eps}\quad
\includegraphics[width=.45\textwidth]{Fig7_SDMech_S25_P1_Case3.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig7_SDMech_S5_P10_Case3.eps}\quad
\includegraphics[width=.45\textwidth]{Fig7_SDMech_S25_P10_Case3.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig7_SDMech_S5_P20_Case3.eps}\quad
\includegraphics[width=.45\textwidth]{Fig7_SDMech_S25_P20_Case3.eps}
\caption{Comparison between two different strain rates for Case 3 at $1\:\text{atm}$, $10\:\text{atm}$ and $20\:\text{atm}$. $S = 10\text{\:s}^{-1}$ (left) and $50\text{\:s}^{-1}$ (right). San Diego Mechanism. Mole fractions of CH$_4$, O$_2$ and N$_2$ (black) and heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case3_species}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=.45\textwidth]{Fig8_SDMech_S50_P1_Case3.eps}\quad
\includegraphics[width=.45\textwidth]{Fig8_SDMech_S75_P1_Case3.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig8_SDMech_S50_P10_Case3.eps}\quad
\includegraphics[width=.45\textwidth]{Fig8_SDMech_S75_P10_Case3.eps}
\medskip
\includegraphics[width=.45\textwidth]{Fig8_SDMech_S50_P20_Case3.eps}\quad
\includegraphics[width=.45\textwidth]{Fig8_SDMech_S75_P20_Case3.eps}
\caption{Comparison between two different strain rates for Case 3 at $1\:\text{atm}$, $10\:\text{atm}$ and $20\:\text{atm}$. $S = 100\text{\:s}^{-1}$ (left) and $150\text{\:s}^{-1}$ (right). San Diego Mechanism. Mole fractions of CH$_4$, O$_2$ and N$_2$ (black) and heat release rate (orange). Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case3_S50_S75}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{Fig10_SDMech_maxQ_Case3.eps}
\caption{Case 3. Heat-release-rate peaks for different strain rates at pressures of $1\:\text{atm}$ (red), $10\:\text{atm}$ (dark blue), $20\:\text{atm}$ (green), $30\:\text{atm}$ (light blue), 40 atm (pink). Only lean-premixed flame (triangles) and diffusion flame (circles) peaks are represented. San Diego Mechanism. See the online version for color references.}
\label{fig:SDMech_case3_maxQ_differentSP}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=.45\textwidth]{Both_SDMech_S5_P10_T_dx_MLT_dQ_x_SI}\quad
\includegraphics[width=.45\textwidth]{Both_SDMech_S25_P10_T_dx_MLT_dQ_x_SI}
\medskip
\includegraphics[width=.45\textwidth]{Both_SDMech_S5_P20_T_dx_MLT_dQ_x_SI}\quad
\includegraphics[width=.45\textwidth]{Both_SDMech_S25_P20_T_dx_MLT_dQ_x_SI}
\caption{Comparison between two different strain rates for Case 3 at $10\:\text{atm}$ and $20\:\text{atm}$. $S = 10\text{\:s}^{-1}$ (left) and $50\text{\:s}^{-1}$ (right). San Diego Mechanism. Stagnation plane location (green) and the estimated mixing-layer edge (red). See the online version for color references.}
\label{fig:SDMech_298K_case3_Q_int}
\end{figure}
\subsubsection{Strain Rate Effects}
At high strain rate ($100\:\text{s}^{-1}$ and $150\:\text{s}^{-1}$), the heat-release-rate peak of the flames increases from right to left, with the fuel rich premixed flame being the weakest of the three. This is also true at low strain rates. Note that the important factor is the integral of the reaction rate through the reaction zone rather than the peak value in terms of total heat release rate. Not many general flame behavioral differences were found between the high strain rates studied. However, the peaks of the flame are getting closer and further from the stagnation point, approaching the estimated mixing-layer edge. This behavior indicates a likely flame merging at still higher strain rates.\hfill \break
\section{Conclusions}
Flame merging has been observed numerically and experimentally for practical turbulent combustion situations. The novelty of the current numerical study is the in-depth analysis of the flame merging and separation processes entailing detailed chemistry and transport, and employing the counterflow canonical configuration commonly used in experimental studies. This is achieved by examining previously unreported mixture ratios in-flowing from each nozzle of the counterflow configuration.\hfill \break
This study describes the important roles that normal strain rate and pressure have with flames. An increase in the normal strain rate and/or a decrease in the pressure causes flames to merge while strain rate decreases and/or pressure increases lead to multiple flames.\hfill \break
Moreover, an endothermic region is observed for Case 1. This finding has to be further investigated. However, a preliminary explanation related to the production of ethylene (C$_2$H$_4$) is plausible.\hfill \break
At low strain rates and pressures, and only when a rich premixed mixture is injected, an extra heat-release-rate peak is obtained. After our investigation, this heat-release-rate peak is linked to high exothermic reactions producing CO$_2$ and H$_2$O, and consuming CO and H$_2$.\hfill \break
In the different multi-branched flames studied, unexpected character of the lean and rich premixed flames indicates that they are diffusion controlled rather than possessing a classical wave-like nature strongly dependent on chemical kinetic rates. \hfill \break
Further research should be conducted to characterize the extinction rate for the studied cases.
\section{Acknowledgements}
This work was supported by the Air Force Office of Scientific Research [FA9550-18-1-0392] with Dr. Mitat Birkan as the scientific officer.\hfill \break
The authors would also like to thank Prof. Hai Wang (Stanford University) and Prof. Antonio L. Sanchez (UC San Diego) for the discussions and suggestions.
| 7ec6db4b339853ed99421c8c3d9cd5bcb8368847 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{secintro}
Supervised machine-learning algorithms allow one to train sophisticated statistical models to associate the different instances of a quantum system to the corresponding physical properties~\cite{dunjko2018machine,carleo2019machine,carrasquilla2020machine}.
These algorithms have already been used to boost some of the most important computational tasks in quantum chemistry and in material science, including: molecular dynamics simulations~\cite{blank1995neural,behler2007generalized,behler2011neural,bartok2017machine,zhang2018end}, electronic-structure calculations based on density-functional theory~\cite{snyder2012finding,li2016understanding,brockherde2017bypassing,ryczko2019deep,moreno2019deep}, and molecular-property predictions from structural information~\cite{hansen2013assessment,schutt2014represent,hansen2015machine}.
They have also been used in drug-design research to predict binding affinities of protein-molecule complexes~\cite{ballester2010machine,khamis2015machine,jimenez2018k}.
Several statistical models have been adopted, including kernel-ridge regression, support vector machines, random forest, and artificial neural networks.
Like in many other fields of science and engineering, deep neural networks are emerging as the most promising candidates. This is mostly due to their ability to automatically extract the relevant features out of many system descriptors~\cite{mills2017deep} --- thus avoiding hand-crafted features --- and to systematically improve in accuracy as the amount of training data increases~\cite{pilati2019supervised}.
However, for small training sets the deep networks are plagued by the overfitting problem, meaning that they fail to accurately generalize to previously unseen instances. This is a critical problem, given that producing copious training sets for large quantum systems is computationally unfeasible, unless one accepts crude approximations.
In this article, we investigate the scalability of various neural networks in the supervised learning of quantum systems. For scalability, we mean their ability to address arbitrary system sizes -- without changing the network structure, namely, the number of neurons and of connecting weights -- while maintaining satisfactory performance as the system size increases.
One of our main goals is to develop statistical models that provide accurate predictions for large quantum systems, even when the training set is sparse.
In previous studies performed in the field of molecular dynamics, approximately scalable models have been implemented by computing potential energies as the sum of single-atom contributions~\cite{behler2016perspective}. These contributions take into account the atomic environment, but only up to a finite cut-off distance.
This approach could lead to uncontrolled approximations in the presence of long-range interactions or correlations~\cite{bartok2017machine}.
In quantum chemistry, a limited form of scalability has been achieved by representing instances of different sizes using the same number of descriptors~\cite{rupp2012fast,hansen2015machine,huo2017unified,stuke2019chemical,jungsize} (usually extending with zeros the descriptor vectors of the small instances).
These descriptors are fed to intrinsically non-scalable models (i.e., models accepting a fixed-number of input descriptors) as, e.g., kernel ridge regression.
This approach becomes impractical for large systems, and it does do not allow extrapolating beyond the sizes included in the training set.
Very recently, a scalable neural network designed to learn the extensive properties of solids has been introduced~\cite{mills2019extensive}. It is based on partitioning the systems into partially overlapping tiles and computing the (extensive) target value as the sum of single-tile contributions.
The overlap regions ought to take into account boundary effects.
The possible drawback of this model is that it appears not to be suitable for intensive properties.
The statistical models mentioned above have been trained and tested against approximate predictions based on density functional theory and, for small chemical systems, on the coupled-cluster method. Therefore, it has not been verified whether they can actually learn the exact properties of large quantum many-body systems, or merely the simplification implied by the adopted approximate theories.
In this article, we introduce a novel architecture based on a deep convolutional neural network complemented by a global pooling layer. This layer allows the network addressing arbitrary system sizes, without necessarily retraining on each size.
Furthermore, it allows us to adopt a transfer learning technique~\cite{caruana1997multitask}, whereby the learning of large systems is accelerated by a pre-training performed on smaller sizes~\cite{zen2020transfer}. As we demonstrate, scalability also allows us to perform extrapolations to sizes larger than the ones included in the training set~\cite{efthymiou2019super,mills2019extensive}.
This novel scalable model is compared to the extensive architecture of Ref.~\cite{mills2019extensive}, and also to (non-scalable) dense networks with all-to-all interlayer connectivity.
To rigorously quantify the performances of these networks, we benchmark their predictions against the exact ground-state energies of disordered quantum Hamiltonians.
In fact, synthetic disordered systems have emerged as suitable stringent testbeds for deep neural networks~\cite{mills2017deep,pilati2019supervised,ryczko2019deep} in the quantum-physics domain.
The first testbed system we consider is a one-dimensional continuous-space Hamiltonian for a single particle in a disordered potential.
This model describes early cold-atom experiments on Anderson localization~\cite{roati2008anderson,billy2008direct}.
In this case, the ground-state energy is not extensive, meaning that it does not increase when the size of the optical field increases.
The second testbed is an (extensive) quantum Ising chain with disordered couplings. We consider a setup with only nearest-neighbor interactions, and also one augmented with far-neighbor interactions and one with frustrated couplings. The ferromagnetic quantum critical point is also addressed.
The ground-state energies are exactly computed via the Jordan-Wigner transformation, via exact diagonalization, and via unbiased quantum Monte Carlo simulations, depending on the specific setup.
Our findings indicate that the scalable model with the global pooling layer retains high accuracy for increasing system size, both for the single-particle Hamiltonian and for the quantum Ising chain. The extensive network performs well only in the latter testbed.
Transfer learning drastically accelerates the learning of large systems, allowing reaching high accuracy with training sets two orders of magnitude smaller than the ones needed without pre-training.
Remarkably, the global pooling network is able to accurately extrapolate the ground-state energies of large Ising chains, including sizes that can be addressed only via computationally expensive quantum Monte Carlo simulations.
The rest of the article is organized as follows: the testbed systems, namely the continuous-space Hamiltonian and the different setups of the quantum Ising chain, are described in Section~\ref{secmodels}.
The networks and the training algorithm are described in Section~\ref{secnetworks}.
Our results on the supervised learning are reported in Section~\ref{secresults}.
Section~\ref{secconclusions} summarizes the main findings, with our conclusions and some future perspectives.
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig1.eps}
\caption{(color online).
Profile of an instance of an optical speckle field $V(x)$ (blue curve) as a function of the spatial coordinate $x/\gamma$. The system size is $L=20\gamma$ where $\gamma$ is the disorder correlations length. The (black) horizontal line represents the average intensity $V_0=\left<V(x)\right>$, while the (red) dot-dashed segment indicates the ground-state energy $E$. Energies are expressed in units of the disorder correlation energy $E_c$.
The dashed (dark green) curve represents the squared modulus of the ground-state wave function $\phi(x)$ with $\ell_2$ normalization. The corresponding (logarithmic) scale is indicated on the right vertical axis.
}
\label{fig1}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.6\columnwidth]{fig2.jpg}
\caption{(color online).
Representation of the connectivity structure of a dense neural network. The (green) leftmost layer represents the $N_d$ input neurons. The three intermediate (red) layers include $N_n$ hidden neurons each. The output neuron on the right is associated to the target value. All neurons are connected to all neurons in the adjacent layers.
}
\label{fig2}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig3.pdf}
\caption{(color online).
Schematic representation of the scalable convolutional network with the global pooling layer.
The input layer (on the left) is followed by $N_c$ convolutional layers with $N_f$ filters, which create filtered maps of the input.
The maps are organized in depth. Local pooling layers are inserted between the convolutional layers, reducing the number of neurons.
The convolutional layers are connected to the dense layers (on the right) via the global pooling layer.
}
\label{fig3}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.85\columnwidth]{fig4.jpg}
\caption{(color online).
Caricature of the extensive convolutional network, inspired by the architecture introduced in Ref.~\cite{mills2019extensive}.
The input system is divided into partially overlapping tiles, represented by alternating colors. The descriptors of each tile are fed into parallel identical networks represented by the (green) rectangles with stripes. The outputs of these parallel networks are summed to obtain the target value, which is the total ground-state energy.
}
\label{fig4}
\end{center}
\end{figure}
\section{Testbed models}
\label{secmodels}
The first testbed system we consider is a one-dimensional continuous-space Hamiltonian for a single quantum particle. This Hamiltonian is defined as:
\begin{equation}
\hat{H}=-\frac{\hbar^2}{2m} \frac{\mathrm{d}^2}{\mathrm{d}x^2} + V(x),
\label{H}
\end{equation}
%
where $\hbar$ is the reduced Planck constant and $m$ is the particle mass. Periodic boundary conditions are adopted.
$V(x)$ is a random external potential designed to describe the effect of optical speckle patterns on alkali atoms.
The Hamiltonian~\eqref{H} describes the early experiments on the Anderson localization phenomenon performed with ultracold atomic gases~\cite{roati2008anderson,billy2008direct}.
Experimentally, optical speckle patterns are generated by shining coherent light through rough semitransparent surfaces. The transmitted light is then focused onto the atomic cloud.
In computer simulations, the corresponding potential can be generated using the numerical algorithm described in Refs.~\cite{huntley1989speckle,PhysRevA.73.013606}.
The generated potential satisfies periodic boundary conditions.
The potential is defined on a discrete spatial grid: $x_g=g\delta x$, where $\delta x=L/N_g$, $L$ is the system size, and the integer $g=0,1,\dots, N_g-1$. The number of grid points $N_g$ has to be large enough, as discussed below.
For the blue-detuned optical fields considered in this article, the probability distribution of the local potential intensities $V_{\mathrm{loc}}=V(x)$, for any $x\in[0,L]$, is $P(V_{\mathrm{loc}})=\exp(-V_{\mathrm{loc}}/V_0)/V_0$ for $V_{\mathrm{loc}}\geqslant 0$, and $P(V_{\mathrm{loc}})=0$ otherwise.
The parameter $V_0 \geqslant 0$ fixes the average intensity, $\langle V(x) \rangle=V_0$.
Different instances of the optical speckle field can be generated using in the numerical algorithm different pseudo-random numbers.
The average intensity $V_0$ also coincides with the the standard deviation of the speckle field: $\sqrt{\langle V(x)^2 \rangle-V_0^2}=V_0$.
Therefore, $V_0$ is the only parameter that determines the amount of disorder in the system.
We normalize the optical speckle field so that its spatial average over the finite system size $L$ is equal to $V_0$, for each individual instance, eliminating small fluctuations due to finite size effects.
The local intensities at two positions ${x}^\prime+{x}$ and ${x}^\prime$ are statistical correlated. The decay of these correlations is characterized by the following autocorrelation function:
\begin{equation}
\Gamma(x)=\frac{\langle V({ x}^\prime)V({x}^\prime+{x})\rangle}{V_0^2}-1 = \frac{\sin^2\left(\pi x/\gamma\right)}{\left(\pi x/\gamma\right)^2}.
\label{speckle6}
\end{equation}
The parameter $\gamma$ characterizes length-scale of the spatial correlations.
The ground-state wave-function $\phi(x)$ of the Hamiltonian~\eqref{H} and the corresponding energy $E$ are determined via exact numerical diagonalization of the matrix obtained from a finite-difference approach. Specifically, the second derivative is represented via an $11$-point formula.
%
For all system sizes, we set the number of grid points $N_g$ so that $\delta x = 0.0195 \gamma$. This is sufficiently small to essentially eliminate any discretization error.
%
We consider different instances of the optical speckle field, fixing the average intensity at the moderately large value $V_0=5E_{\mathrm{c}}$, where $E_{\mathrm{c}}= \hbar^2/(2m \gamma^2)$ is the correlation energy.
This intensity is sufficiently strong to observe the Anderson localization phenomenon in the finite sizes considered in this article~\cite{PhysRevA.100.013603,falco2010density}.
As a consequence of Anderson localization, the wave function $\phi(x)$ has non-negligible values only in a small region of space.
Away from this core region, $\phi(x)$ displays an approximately exponential decay.
This effect is visualized in Fig.~\ref{fig1} for an instance of the speckle field of size $L=20\gamma$.
The profile of the speckle potential $V(x)$ is also shown, together with corresponding ground-state energy $E$.
This energy level randomly fluctuates for different instances of the speckle field. In Section~\ref{secresults}, various networks are trained to predict the ground-state energies of new speckle-field instances.
The second testbed system we consider is a quantum Ising chain. In general, this model is defined as:
\begin{equation}
\hat{H}=-\sum_{j=1}^{N_s} J_j{\sigma}^{z}_{j} {\sigma}^{z}_{j+1}- \sum_{j=1}^{N_s} J^{\prime}_j{\sigma}^{z}_{j} {\sigma}^{z}_{j+d} -\Gamma \sum_{j=1}^{N_s} {\sigma}^{x}_{j}.
\label{H2}
\end{equation}
$\sigma^x_j$ and $\sigma^z_j$ are conventional Pauli matrices at the lattice sites $j=1,\dots,N_s$. $N_s$ is the number of spins, and we consider again periodic boundary conditions.
The couplings $J_j$ fix the strength of the interactions between the nearest-neighbor spins $j$ and $j+1$.
The couplings $J^{\prime}_j$ fix the one between the spins $j$ and $j+d$. The integer $d>1$ fixes the range of this interaction term, as specified below.
$\Gamma$ is the intensity of the (uniform) transverse magnetic field.
We consider various setups of the Hamiltonian~\eqref{H2}. In the first setup, beyond nearest-neighbor couplings are set to $J^{\prime}_j =0$.
The nearest-neighbor couplings $J_j$ are sampled from the uniform probability distribution $\mathcal{P}(J)=\theta(J+1)\theta(1-J)$, where $\theta(x)$ is the unit step function: $\theta(x)=1$ for $x>0$ and $\theta(x)=0$ otherwise. %
The transverse field intensity is set at $\Gamma=0.5$.
We also consider a second setup with nonnegative couplings sampled from the distribution $\mathcal{P}_{>}(J)=\theta(J)\theta(1-J)$, with the transverse-field tuned at the critical point $\Gamma \cong 0.36792$ of the ferromagnetic quantum phase transition~\cite{shankar1987nearest}.
The target value of the supervised learning procedure is, again, the ground-state energy $E$. In some cases, we consider the energy per spin $E/N_s$, as specified in the Appendix~\ref{Appendix}.
The ground-state energy can be exactly computed at a modest computational cost by performing a Jordan-Wigner transformation to a free fermion model~\cite{pfeuty1970one}. For this computation, we follow the numerical algorithm of Ref.~\cite{young1996numerical}.
The third setup we consider for the Hamiltonian~\eqref{H2} includes both nearest-neighbor and next-nearest neighbor couplings. The latter corresponds to the range $d=2$. Both couplings are sampled from the (nonnegative) uniform distribution $\mathcal{P}_{>}(J)$. In this setup, we set $\Gamma=1$.
For the fourth setup, a far-neighbor term corresponding to the range $d=10$ is chosen, again with nonnegative uniform random couplings.
The fifth (and last) setup is analogous to the previous one, but with possibly negative (nearest neighbor and far-neighbor) couplings sampled from $\mathcal{P}(J)$.
Note that this choice leads to frustration effects, which are known to favor emergence of competing states. Therefore, we expect this setup to be particularly challenging for supervised learning algorithms.
In the presence of beyond nearest-neighbor interactions, the ground-state energy cannot be determined via the Jordan-Wigner transformation.
We resort to exact numerical diagonalization of the Hamiltonian matrix represented in the basis of eigenstates of the $\sigma^z_j$ Pauli matrices. The Hamiltonian is stored in computer memory via a sparse representation, and the diagonalization is performed with the MKL Intel library.
Due to the exponential increase of the size of the Hilbert space with the number of spins, the accessible system sizes are limited to $N_s=25$.
To generate a copious enough test set for larger Ising chains, we employ a recently introduced self-learning projective quantum Monte Carlo algorithm~\cite{pilati2019self}.
This algorithm provides unbiased stochastic estimates of the ground-state energy~\cite{pilati2020simulating}, affected only by statistical fluctuations of the order of $10^{-4}\%$. As shown in Ref.~\cite{pilati2019supervised}, such small random fluctuations do not affect the training process.
The largest chain size we address is $N_s=50$, which is far out of reach for exact diagonalization calculations.
These data represent a challenging benchmark for the accuracy of deep neural networks for quantum systems.
It is worth mentioning that artificial neural networks have been recently employed also to accelerate the search of the ground-state of classical disordered Ising models~\cite{mcnaughton2020boosting}.
\section{Artificial neural networks}
\label{secnetworks}
This Section describes the networks considered in this article.
The first network is a conventional dense feed-forward architecture. Its structure is visualized in Fig.~\ref{fig2}.
The input layer includes $N_d$ neurons, which assume the values of the descriptors associated to each system instance.
The neuron $h=1,\dots,N_n^l$ in the hidden layer $l=1,\dots,N_l+1$ assumes the activation value $a_h^l=g(\sum_j w_{h,j}^l a_{j}^{l-1} + b_h^l)$. The index $j$ labels the neurons in the previous layer. The weights $w_{h,j}^l$ couple the layers $l$ and $l-1$. With $b_h^l$ we indicate the bias terms.
For the activation function $g(x)$, we choose the rectified linear unit $ g_{\mathrm{relu}}(x) = \mathrm{max}(0,x)$, or the exponential linear activation $g_{\mathrm{elu}}(x) = x$ if $x > 0$, and $g_{\mathrm{elu}}(x) = \exp(x)-1$ if $ x \le 0$, as specified in the Appendix~\ref{Appendix}.
The output layer $l=N_l+1$ includes one neuron only. Its activation function is the identity.
It is worth emphasizing that in the dense architecture all neurons are coupled to all neurons in the adjacent layers with distinct weights and biases.
The training process consists in optimizing the coefficients $w_{h,j}^l$ and $b_h^l$ so that when the network is fed with the $N_d$ dimensional descriptor vector $\bold{f}_t$ of a system instance, the activation of the output neuron, indicated in the following as $F(\bold{f}_t)$, closely approximates the target value $y_t$.
The index $t=1,\dots,N_t$ labels the instances in the training set.
The optimization algorithm minimizes the loss function $\mathcal{L}(\bold{W})=\frac{1}{N_t}\sum_t \left(F(\bold{f}_t)-y_t\right)^2 + \alpha \| \bold{W}\|_{2}^2$, including the mean-squared error and a regularization term based on the $\ell_2$-norm $\| \bold{W}\|_{2}$ of the vector formed with the weights.
For the networks and the datasets considered in this article, the regularization parameter $\alpha$ is tuned to negligible values, unless otherwise specified.
The optimization method we adopt is the Adam algorithm~\cite{kingma2014adam}.
The dense networks are not scalable, since the number of weights and biases depends on the number of descriptors. This implies that this architecture can operate only on a unique input size.
Furthermore, using dense networks for large systems is impractical. In fact, to be flexible enough to accurately approximate general functions, the required number of hidden neurons $N_n$ has to be of the order of $N_d$.
As discussed in Section~\ref{secresults}, $N_d$ is proportional to the system size ($L$ or $N_s$, depending on the physical system under consideration).
This implies an approximately quadratic growth of the number of weights with the system size. As a consequence, the training process requires rapidly divergent training-set sizes and computational times.
This problem is often encountered when applying artificial neural networks to image analysis.
For this reason, images are usually analyzed via convolutional networks.
The structure of one of the convolutional networks employed in this article is visualized in Fig.~\ref{fig3}.
These networks include $N_c$ proper convolutional layers, followed by $N_l$ dense layers with all-to-all inter-layer connectivity. In standard practice, the last convolutional layer is connected to the dense part through the so-called flatten layer, whereby all activations are concatenated in a unique one-dimensional array.
Notice that in Fig.~\ref{fig3} the standard flatten layer is replaced by a global pooling layer, as explained in the next paragraph.
Each convolutional layer includes a certain number of filters $N_f$. These filters create filtered maps of the previous layer via a convolution operation. The filtered maps are associated to the activations of $N_f$ parallel hidden layers, organized in depth within each convolutional layer.
The neuron activations are again computed via an activation function evaluated on a linear combination of the activations of the previous layer, similarly as in dense networks. However, in the convolutional layers each neuron is connected only to a tile of size $f$ of the previous layer, covering the full depth (with different weights). Theses tiles (in general) partially overlap and they span the whole layer. We consider unit stride and zero padding, so that the size of the filtered maps coincides with the input size $N_d$. Notably, all neurons in a hidden layer are coupled to their corresponding tile with the same weights and biases. This drastically reduces the number of parameters to be optimized compared to a dense network, in particular for large systems.
Another frequently adopted strategy to boost the training process consists in inserting so-called local pooling layers between some or all convolutional layers. The pooling layers down-sample the neurons by performing
averages, or by selecting the maximum value, of the activations within small tiles of size $p$.
The remarkable efficiency of the convolutional networks for large systems originates from the ability of the convolutional layers to automatically extract the most relevant features out of the (possibly many) system descriptors. Notably, this avoids recurse to hand-crafted features, making the networks quite generally applicable.
The role of the dense layers is to operate high-level operations on the extracted features.
However, a standard convolutional network as the one described above is not scalable. In fact, while the convolutional layers are scalable by construction, the size of the flatten layer scales with the input system size. This implies that at least the final dense layers have to be retrained for each system size.
In this article, we implement a fully scalable convolutional network by introducing a global pooling layer in the place of the flatten layer.
We consider pooling with the maximum or the average operation, as specified in Section~\ref{secresults}.
The activations in the global pooling layer correspond to the average or to the maximum values of the whole filtered maps in the last convolutional layer.
They differ from more conventional local pooling layers, which perform pooling operations on small tiles.
Clearly, the number of neurons in the global pooling layer coincides with the number of filters in the last convolutional layer, and it is therefore independent on the system size. This makes the network fully scalable.
We demonstrate in Section~\ref{secresults} that convolutional networks with a global pooling layer, which we refer to as global (maximum or average) networks, display stable performances for increasing system sizes. Furthermore, the scalability property allows one to train them via transfer learning from small to large systems, and to perform extrapolations to sizes larger than the ones included in the training set.
Recently, another strategy to implement a scalable network has been introduced~\cite{mills2019extensive}. A caricature of the corresponding structure is shown in Fig.~\ref{fig4}. This strategy is designed for the extensive properties of solid-state systems. It consists in dividing the system into a set of partially overlapping tiles.
Each tile is formed by a focus region including $\mathcal{F}$ descriptors, and by a context region including $\mathcal{C}$ descriptors. The context regions overlap with the focus regions of the adjacent tiles.
In our implementation, the tile subdivision satisfies the periodic boundary conditions.
The descriptors corresponding to each tile are fed to a conventional convolutional network.
The output value of each network represents the contribution of the corresponding tile to the total ground-state energy, which is computed as the sum of these contributions. The role of the context regions is to take into account the boundary effects. An important feature is that the parallel convolutional networks corresponding to each tile share the same coefficients.
The networks described above, namely the dense, the global (maximum and average), and the extensive networks, are implemented using the Keras library for machine learning~\cite{chollet2015keras}, with TensorFlow backend.
The structural parameters of all models, including number and type of layers and of neurons, are detailed in the Appendix~\ref{Appendix}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig5.pdf}
\caption{(color online).
Coefficient of determination $R^2$ on the test set for the one-dimensional continuous-space Hamiltonian~\eqref{H}.
Two architectures are considered, namely the global maximum and the dense networks.
They are trained and tested on the system size $L$ (horizontal axis).
Here and in the following figures, the error-bars represent the estimated standard deviation of the average over several repetitions of the training process with different pseudo-random numbers. The connecting segments are guides to the eye.
}
\label{fig5}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig6.pdf}
\caption{(color online).
Transfer learning for the one-dimensional continuous-space Hamiltonian~\eqref{H}.
The coefficient of determination $R^2$ for the test size $L=80\gamma$ is shown as a function of the number of instances in the training (or re-training) set $N_t$.
Three global maximum networks are compared: the first one is trained from scratch on the size $L=80\gamma$ without pre-training (green circles); the second one is pre-trained on $L=20\gamma$ (red triangles); the third is pre-trained on $L=20\gamma$ and only its dense layers are retrained on $L=80\gamma$ (black squares).
}
\label{fig6}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig7.pdf}
\caption{(color online).
Extrapolation for the one-dimensional continuous-space Hamiltonian~\eqref{H}.
The coefficient of determination $R^2$ is shown as a function of the size $L$ of the test systems.
Three global maximum networks with different numbers of convolutional filters $N_f$ are compared.
The first three symbols in the key correspond to homogeneous training on the systems size $L=20\gamma$.
The last three symbols correspond to heterogeneous training on the systems sizes $L=20\gamma$ and $L=40\gamma$.
}
\label{fig7}
\end{center}
\end{figure}
\section{Results}
\label{secresults}
The networks described in Section~\ref{secnetworks} are trained to predict the ground-state energies of the disordered Hamiltonians defined in Section~\ref{secmodels}.
The first testbed we consider is the single-particle continuous-space model~\eqref{H}. The system instances are represented by spatial descriptors corresponding to the values of the disordered potential $V(x)$ on a discrete grid with fixed spacing $\delta x$ (see Section~\ref{secmodels}). Notice that for the smallest size we consider, namely $L=20\gamma$, the number of descriptors is as large as $N_d=1024$, and it increases proportionally to $L$.
To quantify the network's accuracy, we compute the coefficient of determination, defined as:
\begin{equation}
R^2=1- \frac{ \sum_{k=1}^{N_k} \left(F(\bold{f}_k)-y_k\right)^2} { \sum_{k=1}^{N_k} \left(y_k-\bar{y}\right)^2}.
\end{equation}
Here, $y_k$ corresponds to the ground-state energy of the test instance $k$, and $N_k$ is the number of instances in the test set. $\bar{y}$ is the average ground-state energy of the test set.
Clearly, none of the instances of the test set is included in the training set.
If the predictions are exact one obtains $R^2=1$. The constant function $F(\bold{f})=\bar{y}$ corresponds to the score $R^2=0$.
$R^2$ is a fair figure of merit since it takes into account the different intrinsic variances of the target values in the different benchmarks.
In particular, these variances change with the system size.
The performances of the different networks are compared in Fig.~\ref{fig5}, as a function of the system size $L$.
For all benchmarks we consider, the training set includes $N_t \approx 10^5$ instances, while the test set includes $N_k\approx 10^3$ or $N_k\approx 10^4$ instances, unless otherwise specified.
One notices that the accuracy of the dense network substantially decreases with $L$. We attribute this loss of performance to the rapid increase of the number of coefficients to be optimized (see Section~\ref{secnetworks}). Optimizing them would require even larger training sets, which are computationally prohibitive, and/or more powerful optimization algorithms.
The global network retains excellent performance $R^2\simeq 1$ with the available training sets. Here, the global pooling layer extracts the maximum value. This choice is motivated by the idea that in an Anderson localized system, as the one under investigation, the network has to identify the spatial region where localization occurs.
The extensive network displays unremarkable performance (data not visualized). For example, for $L=80\gamma$ the highest score we obtain is $R^2 \simeq 0.70$. This is not surprising, since in the single-particle model the ground-state energy is not extensive. In fact, since deep wells are more likely to occur in larger optical speckle fields, low energy values become more likely as $L$ increases.
Still, it is worth pointing out that the extensive network outperforms the dense model.
The scalability property allows one to adopt the transfer-learning protocol commonly employed by computer scientists working on image analysis. In that context, transfer learning is implemented by pre-training very deep networks on large databases of generic images from the world-wide-web, comprising $\sim 10^6$ images. Then, the networks are specialized on the desired classification task on the available, typically relatively small, training set.
Relevant examples of pre-trained deep neural networks for image analysis are the ResNet~\cite{he2016deep} and the VGG models~\cite{simonyan2014very}.
Transfer learning has recently proven very helpful also in quantum-physics research~\cite{zen2020transfer}. It has been employed to accelerate the optimization -- performed via a reinforcement learning algorithm -- of variational wave-functions built with a (non-scalable) generative neural network, namely, the restricted Boltzmann machine.
In this article, a transfer learning protocol is used to accelerate the supervised learning of large-system properties via a pre-training performed on smaller systems, for which copious training sets can be created at a limited computational cost.
Our protocol exploits the scalability property, meaning that a network with the same structure (number of neurons and connectivity) can address different system sizes.
As shown in Fig.~\ref{fig6}, the global maximum network, when pre-trained on the system size $L=20\gamma$, reaches the remarkable performance $R^2\simeq 0.98$ for the size $L=80\gamma$ with as few as $N_t \approx 5\times 10^2$ instances in the retraining set. For larger $N_t$, the score rapidly converges towards the ideal performance $R^2 = 1$.
Instead, extremely small sets with $N_t \ll 5\times 10^2$ instances for the retraining stage become problematic, due to the risk of overfitting the (few) available instances.
An equivalent network trained from scratch (without pre-training) on the $L=80\gamma$ systems requires at least as many as $N_t \sim 5\times 10^4$ instances to reach $R^2\simeq 1$.
It is also worth mentioning that, for small $N_t\lesssim 10^3$, the $R^2$ scores obtained without pre-training require appropriate tuning of the regularization parameter.
Fig.~\ref{fig6} displays also the results of a partial optimization in the retraining stage. In this protocol, the coefficients of the convolutional layers are fixed at the values obtained in the pre-training stage. This procedure is potentially useful since it significantly reduces the computational cost of retraining.
Again, the learning of the larger system is significantly accelerated compared to the optimization without pre-training. This indicates that once the convolutional layers have learned how to extract the relevant features, they can be transferred to different sizes.
Clearly, the fully retrained network reaches higher accuracy due to the superior flexibility.
The remarkable efficiency of the transfer learning protocol motivates us to attempt performing extrapolations to system sizes larger than those included in the training set. The results are reported in Fig.~\ref{fig7}. Three global maximum networks with different numbers of convolutional filters $N_f$ are considered.
These networks are trained on a heterogeneous set, including as many as $1.6\times 10^5$ instances of size $L=20\gamma$, and also a much smaller set of $4.5\times 10^3$ instances of size $L=40\gamma$. The training process includes small cycles of $5$ epochs for each system size in an alternated fashion.
Remarkably, these networks provide accurate predictions up to the largest size considered in our tests, namely $L=320\gamma$.
The heterogeneous training with two system sizes is essential. It ensures that the network does not specialize on a unique size. In fact, three analogous networks trained on a homogeneous set with only $L=20\gamma$ instances fail to accurately extrapolate to larger systems (see Fig.~\ref{fig7}).
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig8.pdf}
\caption{(color online).
Coefficient of determination $R^2$ on the test set for the quantum Ising chain~\eqref{H} with only nearest-neighbor couplings (first setup, see Section~\ref{secmodels}).
Four networks are considered: the dense, the global maximum, the extensive, and the global average networks.
They are trained and tested on the system sizes $N_s$ (horizontal axis).
}
\label{fig8}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig9.pdf}
\caption{(color online).
Transfer learning for the quantum Ising chain~\eqref{H} with only nearest-neighbor interactions (first setup).
The coefficient of determination $R^2$ for the test system size $N_s=100$ is shown as a function of the number of instances in the training (or retraining) set $N_t$.
Three networks are compared: the global average, the global maximum, and the extensive networks.
The first three symbols in the key correspond to networks pre-trained on the size $N_s=20$, and then retrained on $N_s=100$. The last two symbols correspond to networks obtained without pre-training.
}
\label{fig9}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig10.pdf}
\caption{(color online).
Extrapolation for the quantum Ising chain~\eqref{H} with only nearest-neighbor interactions, tuned at the quantum critical point (second setup).
The coefficient of determination $R^2$ is shown as a function of the test system size $N_s$.
The global average network and the extensive network are compared. The training is performed on the sizes indicated in the key.
}
\label{fig10}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig11.pdf}
\caption{(color online).
Extrapolation for the quantum Ising chain~\eqref{H} in two setups (third and fourth setups, see text): nearest neighbor plus next-nearest-neighbor interactions (i.e., range $d=2$); nearest neighbor plus long-range interaction (range $d=10$) .
The coefficient of determination $R^2$ is shown as a function of the test system size $N_s$.
The global average network is considered, and it is trained on the sizes indicated in the key.
}
\label{fig11}
\end{center}
\end{figure}
The second testbed we consider is the random quantum Ising chain~\eqref{H2}. Here, the ground-state energy $E$ is extensive, since on average it increases with the system size $N_s$.
For the first (and also the second) setup (see details in Section~\ref{secmodels}), which include only nearest-neighbor interactions, $E$ is efficiently obtained via the Jordan-Wigner transformation.
The $N_s$ nearest-neighbor couplings $J_j$ are used as system descriptors.
Fig.~\ref{fig8} analyses the performances of different networks, training and testing on the same system size.
The training set includes $N_t\simeq 3 \times 10^4$ instances.
While the accuracy of the dense network rapidly decreases as the system size increases, the extensive network and the two types of global networks maintain $R^2 \simeq 1$.
For the extensive network, the optimal focus and context sizes are $\mathcal{F}=1$ and $\mathcal{C}=1$, respectively, as previous found in a study on classical spin models~\cite{mills2019extensive}.
In the first global network, the global pooling layer extracts the maximum values of the filtered maps. In the second, it computes the average value. One might expect the latter choice to be more appropriate for the extensive property under consideration. This expectation is confirmed below.
Transfer learning is analyzed in Fig.~\ref{fig9}. The global average network and the extensive network, when pre-trained on the small size $N_s=20$, accurately predict the ground-state energy of $N_s=100$ spin chains even with as few as $N_t=30$ instances in the retraining stage. To reach a comparable accuracy without pre-training, they need at least $N_t \approx 10^3$ instances. As anticipated, the global maximum network is much less efficient, requiring $N_t \gg 10^3$ even with pre-training.
The accuracy of the extrapolations is analyzed in Fig.~\ref{fig10}. Here, we consider the second setup. It corresponds to the quantum critical point of the Hamiltonian~\eqref{H2} with only nearest-neighbor interactions. The critical point is particularly interesting since long-range (ferromagnetic) correlations develop. In fact, one might speculate these correlations to be difficult to describe by neural network models. In Fig.~\ref{fig10}, a comparison is made between networks trained on different system sizes. Both the global average network and the extensive network demonstrate remarkable accuracies, even for the largest system sizes considered in our tests. For the former network, the performance improves if pre-training is performed on larger sizes. For the latter, the performance is remarkable even when training is performed on the smallest system size $N_s=10$.
The next benchmarks we consider are the third and fourth setups of the random Ising chain (see Section~\ref{secmodels}). Beyond the nearest-neighbor interaction, they include a next-nearest-neighbor term (corresponding to the range $d=2$) and a far-neighbor term ($d=10$), respectively.
The uniform random couplings $J_i$ and $J^{\prime}_j$ are non-negative. This choice avoids frustration effects.
These Hamiltonians represent a more challenging benchmark, since one cannot speculate that the networks are simply learning to perform the Jordan-Wigner calculation. It is also worth mentioning that long-range interactions, e.g., electrostatic forces, are not addressed by common supervised-learning models for electronic systems~\cite{bartok2017machine}. We argue that the $d=10$ term considered here is suitable to test whether the networks can capture non-local effects.
In these two setups, the system instances are represented by a descriptor vector of size $N_d=2N_s$, including the nearest-neighbor and the beyond nearest-neighbor couplings
in the alternate ordering $\bold{f}=\left(J_1,J^\prime_1,J_2,J^\prime_2,\dots, J_{N_s},J^\prime_{N_s}\right)$.
In Fig.~\ref{fig11} we analyze the extrapolations of two global average networks, trained on the sizes $N_s=12$ and $N_s=15$ for the $d=2$ setup, and on $N_s=21$ and $N_s=22$ for the $d=10$ setup.
Remarkably, these networks accurately reproduce the ground-state energies of $N_s = 50$ spin chains, which are computed via quantum Monte Carlo simulations~\cite{pilati2020simulating} ($N_k \simeq 200$ test instances are considered). The performance improves when training is performed on the larger size.
The last benchmark we consider is a $N_s=21$ spin chain in the fifth setup, which includes nearest-neighbor and far-neighbor interactions ($d=10$) with both positive and negative (random) couplings, leading to frustration effects.
The representation we adopt in this case is based on a two-channel descriptor matrix of size $N_s \times 2$, where each row includes the pair ($J_j$, $J_j^\prime$), for $j=1,\dots,N_s$.
The global average network provides accurate predictions, namely $R^2 = 0.991(3)$, even for this challenging
benchmark~\footnote{With a single channel description, we obtain $R^2=0.980(5)$.
Furthermore, a preliminary analysis with the extensive network provides mediocre results, possibly indicating a difficulty in describing long-range interactions}.
These findings indicate that the scalable networks provide accurate predictions for large-scale complex quantum systems.
Notably, these predictions are obtained at a minuscule computational cost, giving access to system sizes which are out of reach for other accurate computational techniques.
\section{Conclusions}
\label{secconclusions}
We have analyzed the scalability of different artificial neural networks in the supervised learning of disordered quantum systems.
The accuracies of the networks have been rigorously quantified as a function of the system size.
We have introduced a scalable network based on a convolutional architecture complemented by a global pooling layer. This layer allows the network addressing arbitrary system sizes.
This scalable architecture has been compared to (non-scalable) dense networks and to the (scalable) extensive network introduced in Ref.~\cite{mills2019extensive}.
These networks have been tested for various benchmarks, namely a continuous-space single-particle model relevant for cold-atom experiments, and a random quantum Ising chains in different setups, including one with only short-range interaction, one with long-range and frustrated interactions, and one tuned at the ferromagnetic quantum critical point.
The novel scalable architecture retains high accuracy as the system size increases, both for the (intensive) continuous-space Hamiltonian and for the (extensive) quantum Ising chain. As expected, the extensive network performs well only in the latter testbed. Both networks outperform conventional dense networks with all-to-all connectivity.
As discussed in Section~\ref{secresults}, the scalability property allows one to adopt the transfer-learning protocol familiar from the field of image analysis. As we demonstrated, this protocol allows accurately learning the properties of large quantum systems with small training sets, orders of magnitude sparser than those required without pre-training.
Remarkably, the scalable architecture is also able to extrapolate to system sizes larger than those included in the training set, accurately reproducing the results of state-of-the-art quantum Monte Carlo simulations.
Our study highlights the crucial role of the network scalability in the supervised learning of complex quantum systems.
This property allows accurately learning large-system properties with computationally feasible training sets. It also gives access to sizes that are out of reach for other accurate computational techniques.
We demonstrated that scalable networks can learn exact ground-state properties, not only approximate results based on, e.g., density functional theory, as shown in previous studies. Notably, our testbeds included long-range interactions, which have been often omitted in previous investigations.
The scalable network we have introduced does not require manual development of ad-hoc features. In fact, it can be applied to rather generic physical systems with straightforward system representations, allowing the number of descriptors to scale with the system size.
For example, this network could be directly applied to solid-state or molecular systems, predicting properties such as atomization or ionization energies. As descriptor vectors, one could employ images representing the atomic density distribution, as in Ref.~\cite{mills2019extensive}.
The instances in the training set and in the test set would represent different molecular configurations or crystals with different porosities obtained by randomly removing a certain number of atoms.
We leave these endeavors to future investigations.
\section*{Acknowledgements}
\noindent
We acknowledge useful discussions with P. Mujal, A. Polls, B. Juli\'{a}-D\'{i}az, P. Pieri, and A. Perali.\\
S. P. and S. C. acknowledge financial support from the FAR2018 project titled ``Supervised machine learning for quantum
matter and computational docking'' of the University of Camerino and from the Italian MIUR under the project PRIN2017 CEnTraL 20172H2SC4.
S. P. also acknowledges the CINECA award under the ISCRA initiative, for the availability of high performance computing resources and support.
| d72f6f71bd544745fe70e5b1198a29cbd011f840 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\label{sec:intro}Introduction}
The flow periodicity due to the vortex shedding from a bluff body exposed to a fluid stream can cause structural vibration if the body is flexible or it is elastically mounted, which is referred to as `vortex-induced vibration'. In practical applications, compliant structures usually have degrees of freedom to move both along and across the incident flow. Much of the fundamental research on the problem has dealt with rigid circular cylinders as bluff bodies, constrained elastically so as to have a single degree of freedom to oscillate either in-line with a free stream (the streamwise direction) or transversely (the cross-stream direction). The circular cylinder spawns the characteristic that is not prone to galloping vibration and vortex-induced vibration occurs in its purest form. In an early review on vortex shedding and its applications, \cite{King1977} noted that maximum in-line amplitudes are approximately 0.2 diameters, peak-to-peak, or about one-tenth of the corresponding maximum cross-stream amplitudes. As a consequence, subsequent research has mostly concentrated on purely transverse vortex-induced vibration as attested in later reviews on the topic \citep[see, e.g.,][]{Bearman1984,Sarpkaya2004,Williamson2004,Gabbai2005,Bearman2011,Paidoussis}. Insight into the fundamentals of vortex-induced vibration can be gained by the complementary study of either in-line or transverse vortex-induced vibrations since both share the same excitation mechanism. The present study deals with the in-line case.
\subsection{Characteristics of in-line free response }
Figure \ref{fig:sketch} shows the flow-structure configuration considered in the present study.
The elastically-mounted cylinder is modelled using the conventional mass-spring-damper system. The cylinder is constrained so that it can oscillate only in-line with a uniform free stream. The motion of the cylinder is governed by Newton's second law, which can be expressed per unit span as
\begin{equation}\label{eq:motion1}
m\, \ddot{x}_c + c\, \dot{x}_c + k\, x_c = F_x(t),
\end{equation}
where $x_c$, $\dot{x}_c$, and $\ddot{x}_c$ respectively are the displacement, velocity and acceleration of the cylinder, $m$ is the mass of the cylinder, $c$ is the structural damping, $k$ is the spring stiffness, and $F_x(t)$ is the time-dependent sectional fluid force acting on the cylinder.
\begin{figure}\vspace{5mm}
\centerline{\includegraphics[width=0.63\textwidth]{Fig1.eps}}
\caption{\label{fig:sketch} Schematic of the flow-structure configuration. }
\end{figure}
The condition for the onset of vortex-induced in-line vibration can be broadly expressed as $U^*_a\approx 1/(2S)$, whereas the corresponding value for cross-stream vibration is $U^*_a\approx 1/S$, where $U_a^*= U_\infty/f_{n,a}D$ is the reduced velocity and $S=f_{v0}D/U_\infty$ is the Strouhal number; here $U_\infty$ is the velocity of the free stream, $D$ is the diameter of the cylinder, $f_{v0}$ is the frequency of vortex shedding for a stationary cylinder, and $f_{n,a}$ is the natural frequency of the structure in still fluid, i.e.\ including the `added mass'. Traditionally, response data from experimental studies obtained as the flow velocity is varied over the attainable range of experimental facilities are presented as a function of the reduced velocity based on the structural frequency in still fluid. Other non-dimensional parameters governing the structural response are: the ratio of the cylinder mass to the fluid mass displaced by the cylinder, denoted as the mass ratio $m^*$, the ratio of the structural damping to the critical damping at which the mechanical system can exhibit oscillatory response to external forcing, denoted as the damping ratio $\zeta$, as well as the Reynolds number, $Re$, which determines the flow regime. The definitions of the mass ratio and the damping ratio are also not uniform in the literature but depend on whether the added fluid mass is taken into account. In this work, we have selected a set of non-dimensional parameters listed in table \ref{tab:parameters}. It should be noted that we define the reduced velocity using the natural frequency of the system in vacuum, $f_n=(1/2\upi)\sqrt{k/m}$, in common with most previous numerical studies of vortex-induced vibration.
\begin{table}
\begin{center}
\begin{tabular}{cccccc}
\thead{Normalized\\amplitude} & \thead{Normalized\\frequency} &
\thead{Reduced\\velocity} & \thead{Mass ratio} &
\thead{Damping ratio} & \thead{Reynolds number} \\
$\displaystyle A^* = \frac{A}{D}$ & $\displaystyle f^* = \frac{fD}{U_\infty}$ &
$\displaystyle U^* = \frac{U_\infty}{f_nD}$ &
$\displaystyle m^* = \frac{m}{\frac{1}{4}\upi\rho D^2}$ &
$\displaystyle \zeta = \frac{c}{2\sqrt{km}}$ &
$\displaystyle Re = \frac{\rho U_\infty D}{\mu}$ \vspace{3pt} \\
\end{tabular}
\caption{\label{tab:parameters} Definitions of non-dimensional parameters employed in the present study.}
\end{center}
\end{table}
Typically, the response amplitude of cylinder vibration is magnified in distinct ranges of the reduced velocity, which mimic the classical resonance of a single degree-of-freedom oscillator to external harmonic forcing. These distinct regions of high-amplitude response have been given various names such as `instability regions' \citep{King1977}, `excitation regions' \citep{Naudascher1987}, or `response branches' \citep{Williamson2004}.
It has been established as early as the 1970's that there exist two distinct excitation regions of free in-line vibration: the first one appears at $U_a^*\lesssim 2.5$ and has been associated with symmetrical shedding of vortices simultaneously from both sides of the cylinder, whereas the second one appears at $U_a^*\gtrsim2.5$ and has been associated with alternating shedding of vortices from each side of the cylinder \citep{Wootton1972,King1977,Aguirre1977}. The value of $U_a^*\approx2.5$ corresponds to $1/(2S)$ assuming a Strouhal number of 0.20, at which reduced velocity the frequency of vortex shedding from a stationary cylinder becomes equal to half the natural frequency of the structure in still fluid, i.e.\ $f_{v0}\approx \frac{1}{2}f_{n,a}$. A factor of 2 arises in the denominator from the fact that two vortices shed from alternate sides of the cylinder each one induces a periodic oscillation of the fluid force in the streamwise direction. It should be remembered that the Strouhal number is a function of the Reynolds number, defined as $Re=\rho U_\infty D/\mu$, where $\rho$ is the density and $\mu$ is the dynamic viscosity of the fluid. Therefore, the above value of $U_a^*\approx2.5$ should generally be replaced by $U_a^*\approx 1/(2S)$ in dealing with response data at different Reynolds numbers.
In experimental studies with elastically mounted rigid cylinders, the structural frequency depends on the oscillating mass and the stiffness of the supporting springs, which provide elastic restoring forces.
In an early study, \cite{Aguirre1977} conducted more than a hundred tests in a water channel to investigate vortex-induced in-line vibration. He concluded that the structural mass and the stiffness affected the cylinder response independently and in different ways: for a given value of $f/f_{v0}$, the density ratio (i.e.\ the mass ratio here) did not affect the response amplitude when normalized with the cylinder diameter, nor did the stiffness affect the normalized frequency of vibration $f/f_{n,a}$, where $f$ is the actual vibration frequency. Beyond that early study, the effects of the structural mass and stiffness, which are embodied in the mass ratio and the reduced velocity values, have not been systematically addressed in more recent studies. Yet, \cite{Okajima2004} found that the response amplitude of in-line oscillation decreases in both excitation regions with increasing the reduced mass-damping, or Scruton number - a non-dimensional parameter that is proportional to the product $m^*\zeta$ and is often employed to compile peak amplitude data as a function of a single parameter. The later findings possibly illustrate the influence of structural damping alone since the mass ratio was constant in those tests.
The existence of two distinct response branches of free in-line vibration and corresponding modes of vortex shedding were confirmed in more recent experimental studies at Reynolds numbers in the range approximately from $10^3$ to $3.5\times10^4$ \citep{Okajima2004,Cagney2013a}. The drop in response amplitude in-between the two branches, i.e.\ at $U_a^*\approx2.5$, has been attributed to the phasing of alternating vortex shedding, which provides a positive-damping, or negative-excitation force with respect to the oscillation of the cylinder \citep{Konstantinidis2005,Konstantinidis2014}. More recently, a mixed mode of combined symmetric and alternating vortex shedding was also reported to exist in-between the two branches \citep{Gurian2019}.
\subsection{Energy transfer and harmonic approximation}
For self-excited vibrations to be possible, energy must be transferred from the fluid to the structural motion in an average cycle so as to sustain the oscillations, i.e.\ $E\geqslant0$ where $ E=\oint{F_x\mathrm{d}x_c}$; ${F_x}$ is the instantaneous fluid force driving the body motion and $x_c$ is the displacement of the body. A rationale is to determine the energy transfer from forced vibrations of the body in order to predict whether free vibrations can occur for the corresponding case where the cylinder is elastically constrained. This is typically based on the approximation that both the motion of the cylinder $x_c(t)$ and the driving fluid force per unit length $F_x(t)$ can be expressed as single-harmonic functions of time $t$, e.g.\
\begin{eqnarray}
x_c(t) & = & X_0 + A\cos{(2\upi f t)},\label{eq:Xharmonic}\\
F_x(t) & = & F_{x0} + F_{x1}\cos{(2\upi ft+\phi_x)},\label{eq:Fharmonic}
\end{eqnarray}
where $X_0$ is the mean streamwise displacement of the cylinder, $A$ is the amplitude and $f$ is the frequency of body oscillation, $F_{x0}$ and $F_{x1}$ respectively are the magnitudes of the mean and unsteady in-line fluid forces, and $\phi_x$ is the phase lag between the displacement and the driving force at the oscillation frequency.
Using the above harmonic approximations, it can be readily shown that
\begin{equation}
E = \upi A F_{x1}\sin\phi_x.
\end{equation}
Hence, free vibration is possible only if $\sin\phi_x\geqslant0$, or equivalently the phase lag be within the range $0^\circ\leqslant\phi_x<180^\circ$.
Two independent studies employing forced harmonic in-line vibrations at fixed amplitudes of oscillation in the range from 0.1 to 0.28 diameters peak-to-peak, have shown that energy is transferred from the fluid to the cylinder motion in two excitation regions separated by approximately $U^*_{r} \approx2.5$ for $Re$ values higher than $10^3$ \citep{Tanida1973,Nishihara2005}. Here, $U^*_{r} =U_\infty/fD$ is the reduced velocity based on the actual frequency of forced oscillation. The use of $U^*_{r}$ is also critical in correlating forced-vibration studies with the response from free-vibration studies, in which case the vibration frequency is not necessarily equal to the structural frequency \citep{Williamson2004,Konstantinidis2014}. Overall, predictions using forced harmonic vibration agree well with the excitation regions found in free vibration at relatively high Reynolds numbers, including the wake modes responsible for free vibration \citep{Tanida1973,Nishihara2005}. On the contrary, \citeauthor{Tanida1973} found that energy transfer was always negative for all reduced velocities at $Re=80$.
They stated that results obtained for $Re=80$ were representative over the range $40\leqslant Re \leqslant150$, which indicated that free vibration may not be possible in that range Reynolds numbers.
More recently, detailed results from two-dimensional numerical simulations for a cylinder placed in an oscillating free stream and the equivalent case of a cylinder oscillating in-line with a steady free stream both showed that $E<0$ for all reduced velocities at fixed Reynolds numbers of $Re=150$ \citep{Konstantinidis2017AOR} and $Re=100$ \citep{Kim2019}. The lowest amplitudes of forced oscillation in those studies were 0.1 and 0.05 diameters, respectively. The findings from both studies have also indicated that free in-line vibration may not be feasible for Reynolds numbers in the laminar regime, which is consistent with the earlier experimental study of \cite{Tanida1973}. Nevertheless, it is plausible that energy transfer may become positive at lower amplitudes of oscillation than those employed in previous studies, which has not received attention to date since free in-line vibration has scarcely been studied at low Reynolds numbers; the authors are aware of only few numerical studies where in-line vibration of circular cylinder rotating at prescribed rates was investigated at $Re=100$ \citep{Bourguet2015,LoJacono2018}. These studies showed that an elastically-mounted rotating cylinder can be excited into large-amplitude galloping-type vibrations as the reduced velocity increases. However, the response amplitudes were negligible in the case of a non-rotating cylinder compared to the rotating cases and the former results were not discussed.
Apart from addressing the question whether in-line vortex-induced vibration is possible at low Reynolds numbers, a more fundamental issue is to clarify what are the flow physics causing variations in the amplitude and frequency of response when self-excited vibration does occur, not only at low Reynolds numbers. This issue impacts our understanding of vortex-induced vibration as well as its modelling and prediction using semi-empirical codes in industrial applications. To address that issue, it is essential to formulate a theoretical framework with the aid of which results can be interpreted. In this study, we maintain that the in-line free vibration offers a convenient test case because it allows different dynamical effects, which are associated with fluid inertia, fluid damping, and fluid excitation from the unsteady wake, to be segregated.
\subsection{Previous theoretical-empirical approaches}
A long-standing approach is to represent the in-line force per unit length $F_x(t)$ based on the equation proposed by \cite{Morison1950}. For a cylinder of circular cross section oscillating in-line with a steady free stream, the equation can be written as
\begin{equation}\label{eq:Morison1}
F_x(t) = \frac{1}{2}\rho D C_{dh}\left|U_\infty - \dot{x}_c\right|\left(U_\infty - \dot{x}_c\right) - \frac{1}{4}\upi\rho D^2C_{mh} \ddot{x}_c,
\end{equation}
where $\rho$ is the density of the fluid. The coefficients $C_{dh}$ and $C_{mh}$ are often referred to as drag and added mass (or inertia) coefficients, respectively, and their values are empirically determined from measurements or simulations. Inherent to this approach is the harmonic approximation since the coefficients $C_{dh}$ and $C_{mh}$ are often determined from tests where the cylinder is forced to vibrate harmonically. Even when the cylinder motion is self-excited, it is still necessary to characterize the vibration in terms of the least number of appropriate non-dimensional parameters for compiling fluid forcing data; this usually boils down to the use of two parameters, i.e.\ the normalized amplitude and normalized frequency of oscillation, which can fully characterize only single-harmonic oscillations.
Another similar approach is to decompose the fluctuating part of the in-line force into harmonic components in-phase with the displacement (or alternately acceleration) and in-phase with the velocity of the oscillating cylinder, in addition to a steady term for the mean drag.
By using harmonic approximations, the steady-state response can be predicted as we present in appendix \ref{app:harmonic}. This approach is analogous to using Morison \etal's equation and linearising the drag term as $\left|U_\infty - \dot{x}_c\right|\left(U_\infty - \dot{x}_c\right) \approx U_\infty^2 - 2U_\infty\dot{x}_c$. However, Morison \etal's equation comprises two force coefficients whereas the harmonic approximation comprises three force coefficients. The lack of an independent term for the mean drag may explain, at least partially, why Morison \etal's equation reconstructs the in-line force unsatisfactorily in the case of vibrations in-line with a steady free stream. However, it should be noted that the addition of a third term for steady drag in Morison's equation did not considerably improve the empirical-fit results \citep[see][]{Konstantinidis2017AOR}.
\citet{Sarpkaya2001} discussed some limitations of Morison \etal's equation to represent the in-line force on a cylinder placed perpendicular to zero-mean oscillatory flow. Recently, \cite{Konstantinidis2017AOR} demonstrated the inability of the thus reconstructed force acting on a cylinder in non-zero-mean oscillatory flows to capture fluctuations due to vortex shedding in the drag-dominated regime, where the vortex shedding and the cylinder motion (the wave motion in that study) are not synchronized. When the primary mode of sub-harmonic synchronization occurs, Morison's equation provides a fairly accurate fit to the in-line force but subtle differences still exist, which may have detrimental effects when the model equation is used for predicting the free response. A disadvantage of previous approaches based on Morison \etal's equation as well as on the harmonic representation, is that the values of the force coefficients show some dependency on the best-fitting method, e.g.\ Fourier averaging \textit{vs.} least-squares method \citep[see][]{Konstantinidis2017AOR}. Another disadvantage, more important, is that force coefficients are empirically determined and as a consequence it is difficult to decipher the flow physics from the variation of the force coefficients, which is our primary goal in this work.
\subsection{Force decomposition and added mass}
Despite the empirical use of Morison \etal's equation, its inventors stated that it originates from the summation of a quasi-steady drag force and the added-mass force resulting from `wave theory' \citep{Morison1950}. A comprehensive discussion of the theory can be found in \citet{Lighthill1986}. For a body accelerating rectilinearly within a fluid medium, there is an ideal `potential' force acting on the body, which can be expressed as $F_{x,\mathrm{potential}}=-C_am_d\ddot{x}_c$, where $C_a$ is the added mass coefficient, $m_d$ is the mass of fluid displaced by the body, and $\ddot{x}_c$ is the acceleration of the body. Thus, the body behaves as if it has a total mass of $m+m_a$, where $m_a=C_am_d$ is the added mass of fluid. According to the theory of inviscid flow, in which case the velocity field can be defined by the flow potential, the added mass coefficient $C_a$ of any body is assumed to depend exclusively on the shape of the body. For a circular cylinder, $C_a=1$. Free-decay oscillation tests in quiescent fluid have shown that $C_a$ is quite close to the ideal value of unity. However, the applicability of the ideal $C_a$ value in general flows, including cylinders oscillating normal to a free stream, has been criticized \citep{Sarpkaya1979,Sarpkaya2001,Sarpkaya2004}. On the other hand, \citet{Khalak1996} argued that removing the ideal added-mass force from the total force will leave a viscous force that may still comprise a component in-phase with acceleration, i.e.\ the decomposition does not have to separate all of the acceleration-depended forces as done in empirical approaches. Ever since the separation of `potential' (inviscid) and `vortex' (viscous) components has been widely employed to shed light into the vortex dynamics around oscillating bodies transversely to a free stream \citep[see, e.g.,][]{Govardhan2000,Carberry2005,Morse2009b,Zhao2014,Zhao2018,Soti2018}.
For body oscillations in-line with a free stream, one may also split the streamwise force as
\begin{equation}
F_x(t) = F_{x,\mathrm{potential}}(t) + F_{x,\mathrm{vortex}}(t),
\end{equation}
in order to explore the link between fluid forcing and vortex dynamics.
This was previously done for the case of a fixed cylinder placed normal to a free stream with small-amplitude sinusoidal oscillations superimposed on a mean velocity \citep{Konstantinidis2011}. This case is kinematically equivalent to the forced vibration of the cylinder in-line with a steady free stream. In that study, large-eddy simulations corresponding to $Re=2150$ showed that alternating vortex shedding provides positive energy transfer for $U^*_r>2.5$, in very good agreement with previous experimental studies discussed earlier. It was also observed that the streamwise vortex force diminished in magnitude while the instantaneous phase of the vortex force with respect to the imposed oscillation drifted continuously near the middle of the wake resonance (synchronization) region. This was considered to be inconsistent with the flow physics in the following sense: within the synchronization region the vortex shedding and the oscillation are strongly phase-locked and the corresponding wake fluctuations are resonantly intensified. Therefore, the magnitude of the vortex force would have been expected to increase and its instantaneous phase to remain fairly constant in this region. The irregular phase dynamics observed in that study indicates that the vortex force remaining from subtracting the ideal inertial force from the total force may not fully represent the effect of the unsteady vortex motions on the fluid forcing.
\subsection{New theory and outline of the present approach}
In this paper, we develop a new theoretical model for representing the streamwise force on a cylinder oscillating in-line with a free stream. The model stems from some recent observations. In particular, it was recently shown that Morison \etal's equation based on the sum of a quasi-steady viscous drag force and an inviscid inertial force represents the in-line force with comparable accuracy as does the equation with best-fitted coefficients over a wide range of parameters from the inertia to drag-dominated regimes \citep{Konstantinidis2017AOR}. However, neither method could capture fluctuations at the vortex shedding frequency in the drag-dominated regime, as noted earlier. Thus, the idea here is to introduce an independent force term $F_{dw}$, i.e.\ to express the total force as
\begin{equation}\label{eq:Morison2}
F_x(t) = \frac{1}{2}\rho D C_{d}\left|U_\infty - \dot{x}_c\right|\left(U_\infty - \dot{x}_c\right) - \frac{1}{4}\upi\rho D^2C_a \ddot{x}_c + F_{dw}(t),
\end{equation}
where the first term represents the quasi-steady drag, the second term represents the inviscid added-mass force, and the third term represents the unsteady force due to periodic vortex formation in the wake. In this new approach, there are two viscous contributions: the quasi-steady drag, which is an `instantaneous' reaction force, and the wake drag, which represents the `memory' effect in a time-dependent flow. These contributions may be thought of as originating from the vorticity in the thin boundary and free shear layers, and from the vorticity in the near-wake region, respectively. Both are affected by the rate of diffusion of the vorticity, which is finite. However, at sufficiently high Reynolds numbers for which separation occurs, the diffusion within the thin vortex layers occurs fast enough, almost `instantaneously'. Then, the force required to supply the rate of increase of the kinetic energy of the rotational motion in these regions may be taken to be proportional to the square of the relative velocity $U_\infty-\dot{x}_c$, which gives rise to a quasi-steady drag \citep{Lighthill1986}. On the other hand, the diffusion of vorticity at the back of the cylinder is a very complex process involving its cross-annihilation as oppositely-signed vortices roll-up close together in the formation region \citep[see, e.g.,][]{Konstantinidis2016}; as a consequence the resulting fluid force acting on the body depends on the history of the vortex motions in the near wake.
The splitting of the viscous drag to quasi-steady and wake components is consistent with the contribution of vorticity in distinguishable flow regions around a cylinder to the fluid forces as shown in the work of \citet{Fiabane2011}. They revealed these separable contributions by displaying force-density distributions based on the volume-integral expression proposed by \citet{Wu2007}. \citeauthor{Fiabane2011} were able to separate an `external-flow' region containing the thin vortex structures in the attached and free shear layers, which contributed 90\% of the mean drag, and a `back-flow' region between these two vortex layers behind the cylinder, which contributed almost all the drag fluctuations. Moreover, they found that the intensification of the vortex roll-up closer to the cylinder with increasing Reynolds number in the range $Re=50-400$ resulted an increase of the drag fluctuations imposed by the back-flow region. Their findings suggest that -- when the cylinder is oscillating -- the `external flow' is at the origin of the quasi-steady drag whereas the contribution from the `back flow' is captured by the wake drag. Interestingly, \citet{Wu2007} also considered the flow around a circular cylinder in a steady free stream as a test case in their study; they remarked that while a concentrated vortex after its feeding sheet is cut off makes little direct contribution to the fluid force, it plays an indirect but major role through its induced effect on unsteadiness of the boundary-layer separation and the motion of separated shear layers, which implies that wake vortices influence the phasing of the fluid forces.
Assuming that a single periodic mode of vortex shedding occurs, the wake-induced force can be further modelled as a single-harmonic function of time, i.e.
\begin{equation}\label{eq:Fvharmonic}
F_{dw}(t) = \frac{1}{2}\rho U_\infty^2DC_{dw}\cos{(2\upi f_{dw}t+\phi_{dw})},
\end{equation}
where the coefficient $C_{dw}$ represents the magnitude of the unsteady wake drag, and $\phi_{dw}$ the phase between the wake drag and the displacement of the cylinder. The frequency $f_{dw}$ excited by the unsteady wake depends on the vortex-shedding mode, e.g.\ $f_{dw}=2f_{vs}$ for the alternating mode, whereas $f_{dw}=f_{vs}$ for the symmetrical mode.
Equation~(\ref{eq:Fvharmonic}) serves as a reduced-order model with the aid of which the fluid dynamics of vortex-induced vibration can be analysed more thoroughly than possible heretofore by employing previous semi-empirical approaches from the literature.
In this study, we conducted numerical simulations of the flow-structure interaction of a circular cylinder elastically constrained so as to oscillate only in-line with a free stream. The main objectives are: (\textit{a}) use the numerical data from simulations to obtain the variations of the model parameters $C_{dw}$ and $\phi_{dw}$ as functions of the problem parameters, and (\textit{b}) use the expressions derived to calculate the model parameters in conjunction with the equation of cylinder motion to develop a theoretical framework for interpreting the phenomenology of vortex-induced in-line vibration. Simulations were restricted to the two-dimensional laminar regime at low Reynolds numbers to keep computer time within reason so as to examine the influence of the reduced velocity and the mass ratio over wide ranges and with a good resolution. The numerically produced sets of data for flow fields and induced fluid forces and their interaction with the resulting free motion of the cylinder allowed us to address the issues raised in the foregoing paragraphs and hopefully make a contribution to the understanding of the complex flow physics.
\section{Methodology}
\subsection{Governing equations\label{sec:equations} }
The flow is assumed incompressible and two-dimensional while physical properties of the fluid are constant. The fluid motion is governed by the momentum (Navier--Stokes) and continuity equations, which can be written in non-dimensional form using the pressure-velocity formulation as
\begin{eqnarray}
\frac{\partial u_x}{\partial t} + u_x\frac{\partial u_x}{\partial x} + u_y\frac{\partial u_x}{\partial y} & = & - \frac{\partial p}{\partial x} + \frac{1}{Re} \left(\frac{\partial^2 u_x}{\partial x^2} + \frac{\partial^2u_x}{\partial y^2} \right) - \ddot{x}^*_{c}, \label{eq:Xmomentum} \\
\frac{\partial u_y}{\partial t} + u_x\frac{\partial u_y}{\partial x} + u_y\frac{\partial u_y}{\partial y} &=& - \frac{\partial p}{\partial y} +
\frac{1}{Re} \left(\frac{\partial^2 u_y}{\partial x^{2}} + \frac{\partial^2u_y}{\partial y^{2}} \right),\\
\frac{\partial u_x}{\partial x} + \frac{\partial u_y}{\partial y}&=& 0.\label{eq:continuity}
\end{eqnarray}
Coordinates are normalized with $D$, fluid velocities with $U_\infty$, time with $D/U_\infty$, and pressure with $\rho U_\infty^2$. The acceleration of the cylinder $\ddot{x}^*_c$ appears on the right-hand side of equation (\ref{eq:Xmomentum}) because the Navier--Stokes equations are applied in a non-inertial frame of reference that moves with the vibrating cylinder. Instead of explicitly enforcing the continuity equation, the pressure field was computed by solving the following Poisson equation at each time step,
\begin{equation}\label{eq:poisson}
\nabla^2p = 2\left( \frac{\partial u_x}{\partial x}\frac{\partial u_y}{\partial y} - \frac{\partial u_x}{\partial y}\frac{\partial u_y}{\partial x} \right)
-\frac{\partial \mathcal{D}}{\partial t},
\end{equation}
where $\mathcal{D}=\partial u_x/\partial x + \partial u_y/\partial y$ is the dilation. Although the dilation is zero by default in incompressible flows (Eq.~\ref{eq:continuity}), the term $\partial\mathcal{D}/\partial t$ is kept in Eq. (\ref{eq:poisson}) to avoid the propagation of numerical inaccuracies \citep{Harlow1965}.
On the cylinder surface, the no-slip boundary condition gives
\refstepcounter{equation}
$$
u_x = 0, \qquad u_y = 0, \eqno{(\theequation{\mathit{a},\mathit{b}})}\label{eq35}
$$
and the following condition for the normal pressure gradient at the wall
\begin{equation}
\frac{\partial p}{\partial n} = \frac{1}{Re}\nabla^2 u_n - \ddot{x}^*_{c,n},
\end{equation}
where $n$ refers to the component normal to the cylinder surface pointing to the fluid side.
At the far field, a potential flow field is assumed so that
\refstepcounter{equation}
$$
u_x = u_{x,pot} - \dot{x}_c^*, \qquad u_y = u_{y,pot},
\eqno{(\theequation{\mathit{a},\mathit{b}})} \label{eq:BCfar}
$$
where $u_{x,pot}$ and $u_{y,pot}$ is the velocity field from the known potential of irrotational flow. The corresponding condition for the far-field pressure is
\begin{equation}
\frac{\partial p}{\partial n} = 0. \label{eq:BCfar2}
\end{equation}
The initial field corresponds to the potential flow around a circular cylinder. The dimensionless form of the equations illustrates that the fluid motion depends solely on the Reynolds number given the boundary and initial conditions.
The displacement of a cylinder elastically constrained so that it can oscillate only in-line with a uniform free stream is governed by Newton's law of motion, which can be written in non-dimensional form as
\begin{equation}\label{eq:motion}
\ddot{x}_c^* + \frac{4\upi\zeta}{U^*}\dot{x}_c^* + \left(\frac{2\upi}{U^*}\right)^2 x_c^* = \frac{2C_x(t)}{\upi m^*},
\end{equation}
where $x_c^*$, $\dot{x}_c^*$, and $\ddot{x}_c^*$ respectively are the non-dimensional displacement, velocity, and acceleration of the cylinder, normalized using $D$ and $U_\infty$ as length and velocity scales; $C_x(t)$ is the sectional fluid force on the cylinder normalized by $0.5\rho U^2_\infty D$. Here, the fluid forcing is provided by the ambient flow through balancing the normal and shear stresses on the cylinder surface. The cylinder is initially at rest. The dimensionless form of equation (\ref{eq:motion}) illustrates that the cylinder motion depends on the reduced velocity $U^*$, the mass ratio $m^*$, and the damping ratio $\zeta$. The full set of independent dimensionless parameters of the problem comprises $Re,\,U^*,\,m^*$, and $\zeta$.
Because of the two-dimensional approach, simulations were limited up to $Re = 250$.
Our main simulations are at $Re = 180$, a value which is slightly lower than the threshold for which mode-A spanwise instability occurs in the wake of a stationary cylinder, i.e.\ $Re_c \approx190$ \citep[see][]{Williamson1996,Barkley1996}, and may provide some indication of the corresponding threshold of three-dimensional transition for oscillating cylinders. Yet, the in-line oscillation of the cylinder causes wake synchronization, which has been shown to suppress the mode-A spanwise instability into two-dimensional laminar flow with strong K\'arm\'an vortices at $Re=220$ \citep{Kim2009}. Therefore, the flow may well be expected to remain strictly two dimensional for our main simulations at $Re=180$.
For simulations at the highest value of $Re=250$, it is plausible that some three-dimensional instability might exist. Three dimensionality usually appears first in the form of weak coherent structures of streamwise vorticity with specific wavelength riding on the primary spanwise vorticity that remains in-phase along the length of freely-vibrating cylinders \citep[see][]{LoJacono2018,Bourguet2020}. Under such flow conditions, the spanwise vorticity component is much higher than the other two components by one order and the direct effect of the three-dimensionality of the vortical structures on the force is weak as noted by \citet{Wu2007}. Thus, the plausible existence of three-dimensional vortex structures in the cylinder wake may be reasonably expected to not directly influence the magnitude and phase of the streamwise and transverse fluid forces, which are primarily determined by the two-dimensional wake instability, nor influence the cylinder response at the highest Reynolds number of 250 at which we conducted simulations to check the trends of results in the laminar regime.
\subsection{Numerical code}
An in-house code based on the finite difference method was used to solve the equations of fluid motion \citep{Baranyi2008}. The flow domain is enclosed between two concentric circles: the inner circle is the boundary fitted to the cylinder surface while the outer circle represents the far field boundary. The polar physical domain is mapped into a rectangular computational domain using linear mapping functions. The computational mesh of the `physical domain' is fine in the vicinity of the cylinder and coarse in the far field while the corresponding mesh of the transformed domain is equidistant. Space derivatives are approximated using a fourth order finite-difference scheme except for the convective terms for which a third-order modified upwind difference scheme is employed. The pressure Poisson equation is solved using the successive over-relaxation (SOR) method and the continuity equation is implicitly satisfied at each time step. The Navier-Stokes equations are integrated explicitly using the first-order Euler method and the fourth-order Runge--Kutta scheme is employed to integrate the equation of cylinder motion in time. At each time step the fluid forces acting on the cylinder are calculated by integrating the pressure and shear stresses around the cylinder surface, which are obtained from the flow solver. The streamwise force is supplied to the right-hand side of Eq.~(\ref{eq:motion}) which is integrated to advance the cylinder motion. At the next time step, the cylinder acceleration is updated and the equations of fluid motion are integrated to complete the fluid-solid coupling. For all simulations reported here, the cylinder is initially at rest and the initial field around the cylinder satisfies the potential flow.
\subsection{Domain size, grid resolution and time-step dependence studies}
In this section, we present results from preliminary simulations to check the dependence of main output parameters on a) the size of the computational domain in terms of the radius ratio $R_2/R_1$, b) the grid resolution $\xi_{max}\times\eta_{max}$, and c) the dimensionless time step $\Delta t$. Here $\xi_{max}$ and $\eta_{max}$ are the number of grid points in peripheral and radial directions, respectively. During these computations, the Reynolds number, the reduced velocity, the mass ratio, and the structural damping ratio were fixed at $Re=180$, $U^*=2.55$, $m^*=10$, and $\zeta=0$, respectively. The main output parameters of interest are the amplitude $A^*$ and frequency $f^*$ of cylinder response (normalized with $D$ and $U_\infty$) as well as the standard deviations of the in-line and transverse fluid forces (normalized with $0.5\rho U_\infty^2D$), which are respectively denoted $C'_x$ and $C'_y$.
\begin{table}
\begin{center}
\def~{\hphantom{0}}
\begin{tabular}{lccllll}
domain & $R_2/R_1$ & $\xi_{max}\times\eta_{max}$ & \mbox{$A^*$} & \mbox{$f^*$} & \mbox{$C'_x$} & \mbox{$C'_y$} \\
small &120 & $360\times274$ & 0.01082 & 0.3730 & 0.06974 & 0.4226 \\
medium &160 & $360\times291$ & 0.01079 & 0.3726 & 0.07051 & 0.4231 \\
large & 200 & $360\times304$ & 0.01077 & 0.3725 & 0.07094 & 0.4235 \\
\end{tabular}
\caption{\label{tab:Indep_Area} Results of the domain dependence studies at $(U^*, m^*, Re)=(2.55, 10, 180)$}
\end{center}
\end{table}
\begin{table}
\begin{center}
\def~{\hphantom{0}}
\begin{tabular}{lcllll}
grid & $\xi_{max}\times\eta_{max}$ & \mbox{$A^*$} & \mbox{$f^*$} & \mbox{$C'_x$} & \mbox{$C'_y$} \\
coarse & 300$\times$242 & 0.01074 & 0.3725 & 0.07087 & 0.4235 \\
medium & 360$\times$291 & 0.01079 & 0.3726 & 0.07051 & 0.4231 \\
fine & 420$\times$339 & 0.01082 & 0.3727 & 0.07033 & 0.4229 \\
\end{tabular}
\caption{\label{tab:Indep_Grid} Results of the grid dependence studies at $(U^*, m^*, Re)=(2.55, 10, 180)$}
\end{center}
\end{table}
\begin{table}
\begin{center}
\def~{\hphantom{0}}
\begin{tabular}{cllll}
$\Delta t$ & \mbox{$A^*$} & \mbox{$f^*$} & \mbox{$C'_x$} & \mbox{$C'_y$} \\
0.0004 & 0.01082 & 0.3727 & 0.07060 & 0.4233 \\
0.0002 & 0.01079 & 0.3726 & 0.07051 & 0.4231 \\
0.0001 & 0.01078 & 0.3726 & 0.07049 & 0.4231 \\
\end{tabular}
\caption{\label{tab:Indep_Time} Results of dimensionless time step dependence studies at $(U^*, m^*, Re)=(2.55, 10, 180)$}
\end{center}
\end{table}
First, three different values of the radius ratio $R_2/R_1$ of the inner and outer circles defining the computational domain were tested and the corresponding results are shown in table~\ref{tab:Indep_Area}. The number of circumferential nodes $\eta_{max}$ of the physical domain was adjusted in each case in order to keep the grid equidistant in the transformed domain. For these computations, the dimensionless time step was fixed at $\Delta t=10^{-4}U^*\cong0.0002$. Table~\ref{tab:Indep_Area} shows that the most sensitive quantity on the domain size is $C'_x$ displaying a relative difference of 1.7\% between the small and large domains, whereas the corresponding differences for $A^*$, $f^*$ and $C'_y$ are below 0.7\%. The relative differences for all quantities of interest become less than 0.6\% between the medium and large domains. Thus, the medium-sized domain with a radius ratio of $R_2/R_1=160$ was chosen for the rest of the computations.
Next, we tested three grids with different resolutions $\xi_{max}\times\eta_{max}$ where the number of peripheral and radial grid points was increased so that the grid remains equidistant in the transformed plane. For these computations, the radius ratio and dimensionless time step values were fixed at $R_2/R_1=160$ and $\Delta t=10^{-4}U^*\cong0.0002$, respectively. The results of these tests are shown in table~\ref{tab:Indep_Grid}. Again, $C'_x$ is the most sensitive quantity displaying a relative difference of 0.7\% between coarse and fine grids. All quantities of interest display relative differences less than 0.3\% between medium and fine grids. Thus, the medium-resolution grid was chosen for further computations.
Finally, we tested the dependence on the dimensionless time step, $\Delta t$ for three different values corresponding to $\Delta t\cong2\times10^{-4}U^*, 10^{-4}U^*$ and $5\times10^{-5}U^*$. For these computations, the radius ratio and grid resolution values were fixed at $R_2/R_1=160$ and $360\times292$, respectively. The results are shown in table~\ref{tab:Indep_Time}. In this tests, the quantity that is most sensitive to the time step is $A^*$, which shows a relative difference of 0.4\% between the largest and smallest time steps whereas the corresponding relative differences for $f^*$, $C'_x$ and $C'_y$ are all below 0.15\%. The relative differences of all quantities of interest are below 0.1\% between the intermediate and the smallest time steps. Thus, a dimensionless time step of $\Delta t=10^{-4}U^*$ was chosen for the main computations.
\subsection{Code validation}
\begin{figure}
\centerline{\includegraphics[width=0.95\textwidth]{Fig2.eps}}
\caption{Comparison of results obtained in the present study (open circles) against the study of \cite{Bourguet2015} (filled squares) pertinent to purely in-line free vibration in terms of the variation of normalized amplitude $A^*$ and normalized frequency $f^*$ of cylinder response and the standard deviations of the normalized forces in-line and transverse to the free stream, $C'_x$ and $C'_y$ respectively, as functions of the reduced velocity $U^*$ for $(Re,\,m^*,\,\zeta)=(100,\,4/\upi,\,0)$.} \label{fig:comparison}
\end{figure}
The computational code used in the present study was previously employed in several studies of flows about stationary and oscillating cylinders and results have been extensively compared against data from the literature \citep[see][]{Baranyi2008,Dorogi2018,Dorogi2019}.
For instance, \cite{Baranyi2008} found good agreement with the study of \cite{Al-Mdallal2007} in terms of the time history of the lift coefficient and Lissajous patterns of lift \textit{vs.} cylinder displacement at comparable situations for the case of a cylinder forced to oscillate in the streamwise direction. In addition, the extended code handling flow-structure interaction has been validated for the case of a cylinder undergoing vortex-induced vibration with two degrees of freedom with equal natural frequencies in the streamwise and transverse directions against results from \cite{Prasanth2008} for $m^*=10$, $\zeta=0$, and Reynolds numbers in the range from 60 to 240 \citep[for details see][]{Dorogi2018}. Furthermore, \cite{Dorogi2019} showed that results obtained with the present code compare well with published results in \cite{Navrose2017} for purely transverse free vibration, as well as in \cite{Prasanth2011} and \cite{Bao2012} for free vibration with two degrees of freedom with equal or unequal, respectively, natural frequencies in the streamwise and transverse directions, for similar conditions in each case.
In addition to the validation tests presented in previous studies, we compare in figure \ref{fig:comparison} results obtained with the present code against the study of \cite{Bourguet2015} for purely in-line free vibration. Here, we employed a finer step in the reduced velocity to resolve the maximum in $A^*$ as well as the minimum in $C'_x$.
There is excellent agreement of results for $A^*$ and $C'_x$ but there are some minor deviations for $f^*$ and $C'_y$ of 0.2\% and 1.2\%, respectively, which might be attributable to different numerical methods employed in those studies (finite difference \textit{vs.} spectral element). Overall, previous and present validation tests show that the numerical code employed in the present study provides accurate solutions.
\section{Results and discussion}
\subsection{Effect of Reynolds number on in-line response at a fixed mass ratio}
\begin{figure}
\centerline{\includegraphics[width=0.68\textwidth]{Fig3.eps}}
\caption{The in-line response amplitude $A^*$ and frequency $f^*$ with reduced velocity $U^*$ at different Reynolds numbers (see legend); $(m^*,\,\zeta)=(5,\,0)$. The $f^*$ values are divided by the corresponding $S$ values to take into account the effect of Reynolds number on the Strouhal number.} \label{fig:Re_response}
\end{figure}
To start with, we consider effect of the Reynolds number on the in-line response of a cylinder with a mass ratio of $m^*=5$. The structural damping was set to zero so as to allow for the highest possible amplitude response to take place.
Figure~\ref{fig:Re_response} (top plot) shows that there exists a single excitation region in which the response amplitude $A^*$ displays a marked peak for all Reynolds numbers considered. The $U^*$ value at which peak amplitudes occur decreases with $Re$, which can be attributable to the corresponding increase of the Strouhal number since peak amplitudes occur at approximately $U^*_a\approx 1/(2S)$. The peak amplitude over the entire $U^*$ range, denoted as $A^*_\mathrm{max}$, increases from 0.002 at $Re=100$ to 0.024 at $Re=250$, i.e.\ an increase in $Re$ by a factor of 2.5 results in an increase in $A^*_\mathrm{max}$ by a factor of 12. This is a remarkable increase of the order of magnitude, which contrasts the constancy of peak amplitudes of purely transverse free vibration in the corresponding range of Reynolds numbers \citep[a compilation of $A^*_\mathrm{max}$ data as a function of $Re$ from several studies can be found in][]{Govardhan2006}.
For all simulations conducted in this study, the vortex shedding remained synchronized at half the frequency of the cylinder oscillation. As shown in the bottom plot in figure~\ref{fig:Re_response} the normalized response frequency $f^*$ is approximately twice the corresponding Strouhal number at each Reynolds number. However, $f^*$ displays a trough within the excitation region that becomes more pronounced as the Reynolds number increases; the trough can be hardly discerned for $Re=100$. The variations of $A^*$ and $f^*$ appear to be strongly correlated. The characteristics of free response remain similar for all Reynolds numbers considered here. However, a sudden drop in $A^*$ appears just after the peak amplitude for $Re=250$, a feature which is not present for lower Reynolds numbers. This might indicate the onset of branching behaviour similar to that observed in vortex-induced vibration purely transverse to the free stream \citep[see][]{Leontini2006}.
\begin{figure}
\centerline{\includegraphics[width=0.96\textwidth]{Fig4.eps}}
\caption{\label{fig:Re_wake} Snapshots of the distribution of vorticity around a cylinder undergoing free in-line vibration for $(m^*,\,\zeta)=(5,0)$; (\emph{a}) $(Re,\,U^*)=100,\,2.75)$, (\emph{b}) $(Re,\,U^*)=(180,\,2.44)$, (\emph{c}) $(Re,\,U^*)=(250,\,2.40)$. $U^*$ values correspond to peak response amplitudes in figure \ref{fig:Re_response}. Each snapshot corresponds to a random phase of the cylinder oscillation. Contour levels of normalized vorticity at $\pm0.1,\pm0.5,\pm0.9,\ldots$.}
\end{figure}
In comparison to previous experimental studies, which typically correspond to Reynolds numbers above $10^3$, we did not observe another excitation region associated with symmetrical vortex shedding. In contrast, we observed only the alternating mode of vortex shedding in the present simulations corresponding to the laminar wake regime. Figure~\ref{fig:Re_wake} shows vorticity distributions in the wake at $U^*$ values corresponding to peak amplitudes for different Reynolds numbers. In all cases, the vorticity distributions display the familiar von K\'arm\'an vortex street similar to the wake of a stationary cylinder.
As the Reynolds number is increased, the contours of individual vortices become more concentrated and peak vorticity values within them increase. This might be partly attributable to the increase in the amplitude of cylinder oscillation with Reynolds number, which accrues the generation of vorticity on the cylinder surface \citep{Konstantinidis2016}. In addition, the streamwise spacing between the centres of subsequent vortices decreases due to the increase of the normalized frequency of cylinder oscillation $f^*$ with Reynolds number. The absence of the other excitation region associated with the symmetrical vortex shedding may be attributable to the fact that, as has been shown in several previous studies where the cylinder is forced to oscillate in the streamwise direction at correspondingly low Reynolds numbers, the onset of this mode occurs at relatively high amplitudes above 0.1 diameters \citep{Al-Mdallal2007,Marzouk2009,Kim2019}. Since streamwise amplitudes of free vibration are much lower than that threshold, it is not surprising that the mode of symmetrical shedding and the corresponding excitation region were not observed in the present study.
\subsection{Effect of mass ratio on in-line response at a fixed Reynolds number}
Next, we concentrate on the effect of mass ratio on the in-line response at a fixed Reynolds number of $Re=180$, at which the flow is expected to remain laminar and strictly two dimensional. The structural damping was set to zero $(\zeta=0)$ to allow the highest possible amplitudes.
\subsubsection{Cylinder response }
Figure~\ref{fig:response} shows the variations of $A^*$ and $f^*$ with $U^*$ for four $m^*$ values. It can be seen that $A^*$ displays a single excitation region with peak amplitudes of approximately 1\% of the cylinder diameter, irrespectively of $m^*$. At high reduced velocities, i.e.\ $U^*>4$, the response amplitude gradually drops off down to a level that depends on the mass ratio, with $A^*$ becoming lower as $m^*$ increases.
The response frequency $f^*$ initially decreases within the excitation region reaching a minimum value of approximately 0.372 for all mass ratios. As $U^*$ is increased beyond the point of minimum, $f^*$ increases asymptotically towards the value corresponding to twice the Strouhal number for a stationary cylinder, i.e.\ $f^*\approx2S=0.384$. It is interesting to note that $f^*<2S$ over the entire $U^*$ range for all $m^*$, i.e.\ the cylinder always oscillates at a frequency slightly lower than twice the frequency of vortex shedding from a stationary cylinder. This is consistent with forced harmonic vibration studies, which show that vortex-induced vibration due to alternating vortex shedding occurs for $f<2f_{v0}$ \citep{Nishihara2005,Konstantinidis2011}.
\begin{figure}
\centerline{\includegraphics[width=0.72\textwidth]{Fig5.eps}}
\caption{\label{fig:response} The variation of normalized amplitude $A^*$ and normalized frequency $f^*$ of cylinder response as functions of the reduced velocity $U^*$ for $Re=180$ and different mass ratios $m^*$ (see the symbol legend for $m^*$ values). }
\end{figure}
The variations of $A^*$ and $f^*$ shown in figure~\ref{fig:response} appear to be strongly correlated, i.e.\ $A^*$ increases as $f^*$ decreases and \textit{vice versa}. However, it should be noted that peak amplitudes occur at marginally lower $U^*$ values than those that correspond to the minimum in $f^*$. The $U^*$ values at which peak amplitudes occur depend on $m^*$ quite substantially. In fact, peak amplitudes occur approximately at the point where the frequency of cylinder oscillation approaches the natural frequency of the system in still fluid, i.e.\ $f\approx f_{n,a}$. This condition can also be expressed in dimensionless form as:
\begin{equation}\label{eq:Upeak}
\mathrm{Normalized~frequency~at~point~of~peak~amplitude:}\qquad f^*\approx\frac{1}{U^*}\sqrt{\frac{m^*}{m^*+C_a}}.
\end{equation}
The above condition can be verified in table~\ref{tab:Apeak}, which summarizes the response characteristics at peak amplitudes for different mass ratios. Effectively, we see that the reduced velocity at which the peak amplitude occurs primarily depends on the mass ratio but the peak response amplitude does not depend on the mass ratio.
\begin{table}
\begin{center}
\small\addtolength{\tabcolsep}{10pt}
\begin{tabular}{cllll}
$m^*$ & \mbox{$U^*$} & \mbox{$A_\mathrm{max}^*$} & \mbox{$f^*$} & $\frac{1}{U^*}\sqrt{\frac{m^*}{m^*+1}}$ \\
2 & 2.17 & 0.01081 & 0.3725 & 0.3763\\
5 & 2.44 & 0.01082 & 0.3724 & 0.3741\\
10&2.55 & 0.01078 & 0.3726 & 0.3739\\
20& 2.614 & 0.01081& 0.3727 & 0.3733\\
\end{tabular}
\caption{\label{tab:Apeak} Response characteristics at peak amplitude for different mass ratios ($Re=180)$. }
\end{center}
\end{table}
The drop in normalized frequency $f^*$ within the excitation region in fact illustrates the tendency of the oscillation to `lock-in' at the natural frequency of the structure in still fluid, i.e.\ $f\approx f_{n,a}$, over a range of reduced velocities. It is important to distinguish this `lock-in' tendency, which only occurs in free vibration, from `vortex lock-in' (a.k.a.\ `vortex lock-on'), which regards the synchronization of the vortex shedding and the cylinder oscillation and may occur in both forced and free vibration \citep{Konstantinidis2014}. In forced vibration, there is no natural frequency of the structure so the lock-in relationship $f=f_{n,a}$ is meaningless. It should also be noted that for all simulations reported in figure~\ref{fig:response} (i.e.\ for all $U^*$ and $m^*$ values), the vortex shedding was synchronized with the cylinder oscillation so that $f=2f_{vs}$, where $f_{vs}$ is the frequency of vortex shedding from the freely vibrating cylinder. This is tantamount to the sub-harmonic vortex lock-on in the context of forced oscillations where alternating vortex shedding synchronizes at half the frequency of cylinder oscillation, i.e.\ $f_{vs}=\frac{1}{2}f$ \cite[see, e.g.,][]{Kim2019}. However, the conventional lock-in $f\approx f_{n,a}$ occurs over a narrow range of reduced velocities in the region of peak-amplitude response.
\subsubsection{Magnitude and phase of fluid forces}
\begin{figure}
\centerline{\includegraphics[width=0.74\textwidth]{Fig6.eps}}
\caption{The variations of the mean drag coefficient $\overline{C}_x$ and the standard deviations of the normalized fluid forces in-line with and transversely to the free stream, respectively $C'_x$ and $C'_y$, as functions of the reduced velocity $U^*$ for different mass ratios, $m^*$ at $Re=180$ (see the legend for $m^*$ values). Dashed lines indicate constant values that correspond to the stationary cylinder.}\label{fig:forces}
\end{figure}
In figure \ref{fig:forces} we present the variations of the mean drag coefficient $\overline{C}_x$ and the standard deviations of the unsteady forces in-line with and transversely to the free stream, $C'_x$ and $C'_y$ respectively, as functions of $U^*$. The dashed lines indicate constant values corresponding to a stationary cylinder at $Re=180$ \cite[taken from][]{Qu2013}.
All three quantities exhibit similar variations with $U^*$ for all mass ratios, i.e.\ initially they increase above the corresponding fixed-cylinder values, subsequently they decrease steeply, and finally they gradually increase thereafter as $U^*$ is increased.
The maximum $\overline{C}_x$ is 0.5\% greater and the minimum $\overline{C}_x$ is 1.6\% smaller than the mean drag coefficient for the stationary cylinder, independently of the $m^*$ value. However, the maxima and minima in $\overline{C}_x$ occur at different $U^*$ values depending on the $m^*$ value.
The middle plot in figure~\ref{fig:forces} shows that $C'_x$ reaches a peak value at approximately the point of peak amplitude response, which is nearly three times the value corresponding to a stationary cylinder; a substantial increase despite the low amplitude of peak oscillation of only 1\% of the cylinder diameter.
Following the peak there is a steep decrease within a narrow $U^*$ range at the end of which $C'_x$ tends to zero. Interestingly, for all mass ratios this occurs at $U^*=2.625$, a point at which the response frequency coincides with the natural frequency of the structure in vacuum, i.e.\ $f=f_n$ (or in non-dimensional parameters $f^*=1/{U^*}$). Hereafter, this special operating point will be referred to as the `coincidence point' and its ramifications will be discussed in more detail in Sect.~\ref{sec:superharmonic}. Beyond the coincidence point, $C'_x$ gradually reaches to a plateau at a value that is proportional to $m^*$. On the other hand, the peak and trough $C'_y$ values are merely 6\% higher and 5\% lower, respectively, than the value corresponding to a stationary cylinder (see the bottom plot in figure~\ref{fig:forces}). For all $m^*$, a maximum value of $C'_y=0.45$ exactly is attained at the point of peak response amplitude. In addition, at the coincidence point, i.e.\ $U^*=2.625$, $C'_y$ attains exactly the same value of 0.418 for all $m^*$, which is just slightly below the value corresponding to a stationary cylinder. Since there is no body acceleration in the transverse direction, changes of $C'_y$ can be related to changes in the vortex dynamics around the oscillating cylinder \citep{Leontini2013}. Then, the small variation in the $C'_y$ magnitude illustrates that the process of vortex formation and shedding does not substantially change over the entire range of reduced velocities, even though the cylinder can be oscillating with different but generally small amplitudes.
The phase angle $\phi_x$ between the driving force $F_x(t)$ and cylinder displacement $x_c(t)$, as defined by harmonic approximations in equations (\ref{eq:Xharmonic}) and (\ref{eq:Fharmonic}), is often considered to be a useful parameter in studies of vortex-induced vibration. In addition to $\phi_x$, we also compute here the phase angle $\phi_y$ between the unsteady force acting in the transverse direction $F_y(t)$ and the displacement, assuming that $F_y(t)$ can also be approximated as a single-harmonic function of time. The calculation method of the phase angle is described in appendix \ref{app:phases}. It should be noted that the harmonic approximations work very well at low Reynolds numbers considered here except for the time history of $F_x(t)$ near the coincidence point as will be discussed in detail further below.
Figure \ref{fig:phases} shows the variations of the phase angles $\phi_x$ and $\phi_y$ as functions of $U^*$.
\begin{figure}
\centerline{\includegraphics[width=0.74\textwidth]{Fig7.eps}}
\caption{The variation of the phase angles $\phi_x$ and $\phi_y$ between in-line and transverse forces, respectively, and the cylinder displacement as functions of the reduced velocity $U^*$ at $Re=180$ and different mass ratios $m^*$ (see the symbol legend for $m^*$ values).}\label{fig:phases}
\end{figure}
In figure \ref{fig:phases}, it can be seen that $\phi_x$ jumps suddenly from $0^\circ$ to $180^\circ$ across the coincidence point at $U^*=2.625$ for all $m^*$. This behaviour can be predicted from the steady-state harmonic solution, which is given in appendix \ref{app:harmonic}, as follows. Equation (\ref{eq:Hsine}) requires that the condition $\sin\phi_x=0$ be satisfied when the structural damping is null $(\zeta=0)$. This constrains $\phi_x$ to be either $0^\circ$ or $180^\circ$. This taken in tandem with the requirement per equation (\ref{eq:Hcosine}) that $\cos\phi_x$ must change from positive to negative as $U^*$ increases through $f^*U^*=1$ translates to the jump in $\phi_x$ from $0^\circ$ to $180^\circ$ appearing at the coincidence point. We would like to stress that the equation of cylinder motion constraints $\phi_x$, which makes it impossible to infer any changes in the flow from the variation of $\phi_x$ as the reduced velocity is varied.
In contrast to $\phi_x$, there is no analogous constraint on $\phi_y$, which instead of a jump displays a smooth variation with $U^*$, as can be seen in figure \ref{fig:phases}. For all mass ratios, $\phi_y$ increases from an initial value of approximately $20^\circ$ at low reduced velocities to a terminal value of approximately $105^\circ$ at high reduced velocities, with precise values being slightly depended on $m^*$. The $U^*$ range over which $\phi_y$ changes rapidly is consistent with the range of relatively high-amplitude response, which broadens as $m^*$ decreases (see figure \ref{fig:response}). The variation of $\phi_y$ clearly suggests a gradual change in the vortex dynamics as $U^*$ is varied over the prescribed range. Based on evidence from previous studies \citep[see][]{Konstantinidis2005,Konstantinidis2011}, our main hypothesis is that the variation of the phase angle $\phi_y$ can be linked to a gradual shift in the timing of vortex shedding with respect to the cylinder oscillation as $U^*$ is increased.
This hypothesis will be verified in the following subsection by inspection of vorticity distributions at different phases of the cylinder oscillation.
\subsubsection{Vorticity distributions}
We have selected three $U^*$ values, which are listed in table~\ref{tab:phase}, for presenting vorticity distributions in the wake. In purpose, the normalized frequency of cylinder oscillation $f^*$ is nearly the same in all three cases so that the vorticity patterns are easier to describe since the streamwise spacing of the shed vortices scales with $f^*$ \citep{Griffin1978}. However, the frequency ratio $f/f_n=U^*f^*$ and the phase angle $\phi_y$ both increase with $U^*$, at different degrees, as also shown in table~\ref{tab:phase}. The two higher $U^*$ values are just before and after the `coincidence point', $U^*f^*=1$.
For each $U^*$ value, figure~\ref{fig:vortex_phase} shows instantaneous vorticity distributions at two phases corresponding to the maximum and the subsequent zero displacement of the cylinder as indicated in the time traces (see bottom plots in figure~\ref{fig:vortex_phase}). The significant change of the phase angle $\phi_y$ as a function of $U^*$ can be readily inferred from the time traces of the displacement and the transverse force, which have been appropriately normalized with their maximum amplitudes for better visualisation of their waveforms.
As $U^*$ is increased, the position of individual vortices at corresponding instants appears to have shifted slightly downstream. The shift in the streamwise position of individual vortices fits very well with the change in $\phi_y$ as $U^*$ is varied; when $U^*$ changes from 2.35 to 2.6, $\phi_y$ more than doubles and a large shift is observed; when $U^*$ changes from 2.6 to 2.7, $\phi_y$ changes by few degrees and the shift is hardly perceptible.
We made a quantitative estimation of the relative phase shift for pairs of $U^*$ values from the corresponding spatial shift in the position of corresponding vortex centres at the instant of maximum displacement, with vortex centres extracted by locating the points of peak vorticity in each individual vortex.
When $U^*$ changes from 2.35 to 2.6 the above method yields a relative phase shift of $54^\circ$, which is very close to $\Delta\phi_y=55.4^\circ$ computed directly from the phase shift of the transverse force (see table \ref{tab:phase}). Similar inferences can be made by looking into the stage of formation of vortices just behind the cylinder. For instance, a stripe of negative vorticity (in blue colour) connecting the second vortex to the third vortex taken from right to left, i.e.\ as time progresses, can be observed at maximum displacement for $U^*=2.35$. However, the corresponding vortices are no longer connected by a stripe of vorticity for $U^*=2.6$ and 2.7, which illustrates that vortex shedding is at a progressed stage in these cases.
It should be noted that although this visual feature, which we could make clear by selecting a minimum level of the normalized vorticity of 0.1, depends on this minimum level, it does provide a quantitative comparison of the process of vortex formation at different reduced velocities since the same contour levels were employed for producing all vorticity distributions; in a similar manner the space occupied by individual vortices provides a measure of their strength, which allows quantitative comparisons when the same contour level is employed. It is also interesting to observe that nothing special changes in the vortex patterns between $U^*=$ 2.6 and 2.7 but a marginal phase shift, although $\phi_x$ jumps by $180^\circ$ on crossing over the coincidence point.
The above observations firmly support that the variation of $\phi_y$ as a function of $U^*$, in contrast to the variation of $\phi_x$, is directly linked to changes in the timing of vortex shedding from the cylinder.
\begin{table}
\begin{center}
\small\addtolength{\tabcolsep}{22pt}
\begin{tabular}[1.8\textwidth]{rlll}
\mbox{$U^*$} & 2.35 & 2.60 & 2.70 \\
\mbox{$f^*$} & 0.378 & 0.380 & 0.381\\
\mbox{$U^*f^*$} & 0.888 & 0.989 & 1.030\\
\mbox{$\phi_y$} (degrees) & 40.6 & 96.0 & 99.2\\
\end{tabular}
\caption{\label{tab:phase} Important characteristics at three reduced velocities $(Re,\,m^*)=(180,\,5)$. }
\end{center}
\end{table}
\begin{figure}
\centerline{\includegraphics[width=0.99\textwidth]{Fig8.eps}}
\caption{The top two rows show instantaneous distributions of the vorticity at maximum and subsequent zero displacement during a cycle of cylinder oscillation for three reduced velocities in each column; $(Re,\,m^*)=(180,\,5)$.
The bottom row shows time traces of the normalized displacement $\hat{x}_c(t)$ (dashed lines) and the normalized transverse force $\hat{C}_y(t)$ (solid lines) where circle and square symbols, respectively, mark the instants for which vorticity distributions are shown above. Contour levels of normalized vorticity: $\pm0.1,\,\pm0.5,\,\pm0.9, \ldots$}
\label{fig:vortex_phase}
\end{figure}
\subsubsection{Variable added mass}
Changes in the frequency of cylinder response are often correlated with variations in the inertial force due to the added mass . This follows from the relationship
\begin{equation}\label{eq:CGW}
\frac{f}{f_{n,a}}=\sqrt{\frac{m^*+C_a}{m^*+C_{EA}}}.
\end{equation}
Here, $C_a$ is the ideal added mass coefficient from potential flow theory and $C_{EA}$ is a variable added mass coefficient \citep{Aguirre1977}. The latter coefficient has become known as the `effective added mass' in the context of transverse free vibration \citep{Khalak1996,Williamson2004}. Equation (\ref{eq:CGW}) is equivalent to (\ref{eq:Hcosine}) of the harmonic approximation solution where the component of the force in-phase with displacement has been substituted by
\begin{equation}
C_{x1}\cos\phi_x = 2\upi^3f^{*2}A^*C_{EA},
\end{equation}
and the ratio of the natural structural frequency in vacuum to that in still fluid is given by
\begin{equation}\label{eq:fn_ratio}
\frac{f_n}{f_{n,a}}=\sqrt{\frac{m^*+C_a}{m^*}}.
\end{equation}
\begin{figure}
\centerline{\includegraphics[width=0.75\textwidth]{Fig9.eps}}
\caption{The variation of the effective added mass $C_{EA}$ with the reduced velocity $U^*$ for different mass ratios $m^*$ (symbol legend as in figure \ref{fig:forces}).} \label{fig:CEA}
\end{figure}
In figure \ref{fig:CEA}, we present the variation of $C_{EA}$ as a function of $U^*$ for each $m^*$ investigated. $C_{EA}$ values were computed \textit{via} equation (\ref{eq:CGW}) and the known frequency ratio; note that $f/f_{n,a}=f^*U_a^*=f^*U^*(f_n/f_{n,a})$.
It can be seen that $C_{EA}$ decreases continuously from positive to negative values with $U^*$; $C_{EA}$ decreases from as high as 40.8 to as low as $-13.3$ for $m^*=20$. At some point, $C_{EA}$ takes the theoretical $C_a$ value of unity, which is indicated by the dashed line in the inset of figure \ref{fig:CEA}. Interestingly, this occurs at approximately the reduced velocity of peak amplitude response, which is given by equation~(\ref{eq:Upeak}).
Furthermore, for $m^*=20$ a gap separating operating points with $C_{EA}$ values above and below unity appears at $U^*=2.614$, which is not present at lower $m^*$values. The discontinuous variation remained in place even though an extremely fine step of $\Delta U^*=0.002$ was employed around this point.
When the frequency of cylinder oscillation approaches the natural frequency of the structure in vacuum, equation~(\ref{eq:CGW}) yields $C_{EA}=0$, which occurs at the coincidence point $U^*=2.625$ for all mass ratios.
The very wide variation of $C_{EA}$ values as a function of $U^*$ is improbable to represent some physical change, e.g.\ due to inertial effects, as we have already seen that flow physics remain fairly robust over the entire range of reduced velocities. We point out this simply because we would like to decipher the true effect due to the added mass, which seems impossible to do through the effective added mass concept. On the other hand, it is shown further below with the aid of the new theory that a constant value of the added mass coefficient aligns very well the observations from the simulations.
Nonetheless, the empirical values of $C_{EA}$ as defined above are valid.
\subsubsection{Super-harmonic fluid forcing at the coincidence point}\label{sec:superharmonic}
As already pointed out, the standard deviation of the unsteady in-line force tends to zero level as the frequency of cylinder vibration approaches the natural frequency of the structure in vacuum, which occurs at $U^*\approx2.625$ for all mass ratios.
This can be predicted from the harmonic solution as follows. Equations (\ref{eq:Hsine}) and (\ref{eq:Hcosine}) given in appendix \ref{app:harmonic} can be combined so as to eliminate the phase angle $\phi_x$. In the case of zero structural damping $(\zeta=0)$, the following relationship is obtained
\begin{equation}\label{eq:Ctotal}
C_{x1} = 2\upi^3\frac{m^*A^*}{{U^*}^2}\left| 1 - \left(f^*U^*\right)^2 \right|.
\end{equation}
The above relationship shows that if the vibration frequency approaches the natural frequency of the structure in vacuum, which can be expressed in non-dimensional form as $f^*U^*\rightarrow1$, then the only feasible solution is that $C_{x1}\rightarrow0$. In the present simulations we have found that for all mass ratios $f$ approaches $f_n$ within less than 0.1\%, which was made possible by using very fine steps of $\Delta U^*=0.01$ or even less.
\begin{figure}
\centerline{\includegraphics[width=0.99\textwidth]{Fig10.eps}}
\caption{The top-row plots show time series of the cylinder displacement (dashed lines) and the unsteady streamwise force (solid lines) for different reduced velocities around the coincidence point; $(Re,\,m^*)=(180,2)$. Here, the cylinder displacement $\hat{x}_c(t)$ and the in-line force $\hat{C}_x(t)$ have been normalized with their corresponding maximum amplitudes in order to reveal their relative waveforms.
The bottom-row plots show the spectra of the streamwise force for the corresponding cases shown in the top row. } \label{fig:m2traces}
\end{figure}
Figure \ref{fig:m2traces} shows time series of the cylinder displacement and the fluctuating part of the in-line force at different reduced velocities. The displacement and the force were normalized with their corresponding maxima in order to reveal their relative waveforms because absolute values of the force are extremely small. Although the displacement remains remarkably harmonic, the force starts to deviate from pure harmonic as $U^*$ is varied with a fine step around the coincidence point. Spectra of the force show that a secondary peak appears at the first super-harmonic of the vibration frequency. At $U^*=2.63$, the main spectral peak appears at the first super-harmonic of the vibration frequency, whereas a very small peak remains at the main frequency of vibration; the latter corresponds to an extremely small magnitude of $C_{x1}$, which is however sufficient to sustain the vibration at the main harmonic of the vibration.
The emergence of a dominant super harmonic accompanies the phase jump by $180^\circ$ of the main harmonic occurring as the coincidence point is crossed over, which can be also seen from the time traces in figure \ref{fig:m2traces}. The predominance of the first super harmonic was also observed in flow-induced transverse vibration of a rotating cylinder at corresponding coincidence points \citep{Bourguet2014}. It seems that the phase-jump mechanism through a super harmonic could be a generic feature.
\section{Development of new linear theory}\label{sec:theory}
In this section, we develop a theoretical framework based on the triple decomposition of the total in-line force $F_x(t)$ in equation~(\ref{eq:Morison2}) and the reduced-order model of the wake force $F_{dw}(t)$ in equation~(\ref{eq:Fvharmonic}). Initially, we derive analytical expressions from which the model parameters $C_d$, $C_{dw}$ and $\phi_{dw}$ can be calculated using numerical data from the simulations. Then, we present the variations of the model parameters with reduced velocity and mass ratio to elucidate the fluid dynamics of vortex-induced in-line vibration. In addition, we combine the analytical expressions for drag components with the equation of cylinder motion in order to derive further expressions that allow us to interpret the flow physics at play.
\subsection{Linearised force}
We begin with linearising the quasi-steady drag term in equation~(\ref{eq:Morison2}) and substituting the harmonic approximations in equations (\ref{eq:Xharmonic}) and (\ref{eq:Fvharmonic}) to obtain the linearised force, which we can express explicitly as a function of time $t$ as
\begin{eqnarray}
F_x(t) = \frac{1}{2}\rho U_\infty^2D \left[ C_d +2 \left(\frac{\omega A}{U_\infty}\right) C_d \sin{(\omega t)} + C_{dw} \cos{(\omega t+\phi_{dw})} \right] \nonumber\\ + \frac{1}{4}\upi\rho D^2C_a \omega^2A\cos{(\omega t)}, \label{eq:Morison3}
\end{eqnarray}
where the angular frequency $\omega=2\upi f$ has been introduced for brevity and $f_{dw}$ has been replaced by $f$ due to the synchronization of the wake and the cylinder vibration.
The first term on the right-hand side of equation (\ref{eq:Morison3}) represents the steady part of the in-line force whereas the remaining terms represent the unsteady part, which comprises the combination of cosine and sine functions at the frequency of cylinder oscillation. Equating the steady parts in equations (\ref{eq:Fharmonic}) and (\ref{eq:Morison3}) yields
\begin{equation}\label{eq:meanCd}
C_d = \overline{C}_x,
\end{equation}
i.e.\ $C_d$ is equal to the mean drag coefficient. Equating the cosine and sine terms in equations (\ref{eq:Fharmonic}) and (\ref{eq:Morison3}) yields the following relationships:
\begin{eqnarray}
C_{dw}\sin{\phi_{dw}} & = & C_{x1}\sin\phi_x + 4\upi f^*A^* C_d, \label{eq:Cdvsin} \\
C_{dw}\cos{\phi_{dw}} & = & C_{x1}\cos\phi_x - 2\upi^3f^{*2}A^*C_a. \label{eq:Cdvcos}
\end{eqnarray}
The set of equations (\ref{eq:meanCd}-\ref{eq:Cdvcos}) establishes relationships between fluid forcing components as functions of $A^*$ and $f^*$ alone, i.e.\ the structural parameters do not appear. Therefore, the same values of the fluid forcing components also hold when the cylinder is forced to oscillate at the same operating points in the $A^*:f^*$ parameter space as the free vibration. We use the above set of relationships in order to determine $C_{dw}$ and $\phi_{dw}$ appearing in the new theoretical model from the cylinder response $(A^*,\,f^*)$ and the fluid forcing $(C_{x1},\,\phi_{x})$ data obtained from the simulations. It should be remembered here that $C_a=1$ is the known ideal added-mass coefficient of unity.
\subsection{Magnitude and phase of wake drag}
Figure \ref{fig:vortex_drag} shows the variation of $C_{dw}$ and $\phi_{dw}$ as functions of the parameter $U^*_af^*$. We have employed this parameter because peak amplitudes occur at $U^*_af^*\approx1$. As can be seen in the top plot, $C_{dw}$ also displays a peak at the same point with a maximum value of 0.067 for all $m^*$. The mutual amplification of the response amplitude and the wake drag are representative of resonance. Hence, we refer to the condition $U^*_af^*=1$ as the `resonance point'. Away from resonance $C_{dw}$ tends asymptotically to 0.036, which corresponds to the amplitude of the fluctuating drag coefficient for a non-vibrating cylinder at $Re=180$.
\begin{figure}
\centerline{\includegraphics[width=0.7\textwidth]{Fig11.eps}}
\caption{The variation of the wake drag magnitude $C_{dw}$ and phase $\phi_{dw}$ as functions of $U^*_af^*$ for different mass ratios $m^*$ at $Re=180$ (see the symbol legend for $m^*$ values).} \label{fig:vortex_drag}
\end{figure}
In the bottom plot in figure~\ref{fig:vortex_drag} can be seen that $\phi_{dw}$ follows an S-type increase as a function of $U^*_af^*$ from approximately $0^\circ$ to $180^\circ$. At $U^*_af^*=1$, $\phi_{dw}$ passes through $90^\circ$ for all $m^*$.
Overall, the variation of $\phi_{dw}$ is very similar to that of $\phi_y$ shown earlier; in fact there is a direct relationship between them, which is illustrated in figure \ref{fig:phaselink}. Since the variation of $\phi_y$ is invariably linked to the vortex dynamics, in particular to the timing of vortex shedding as discussed earlier, it follows that $\phi_{dw}$ appropriately captures related changes as the reduced velocity is varied.
\begin{figure}
\centerline{\includegraphics[width=0.74\textwidth]{Fig12.eps}}
\caption{Relationship between phase angles $\phi_{dw}$ and $\phi_{y}$ for different mass ratios $m^*$ (see the symbol legend for $m^*$ values).} \label{fig:phaselink}
\end{figure}
\subsection{Peak response at resonance point}
The new hydrodynamic model can be combined with the equation of cylinder motion in order to illustrate the interaction between the fluid and the structural dynamics.
The steady-state solution can be obtained following a similar procedure as described in appendix \ref{app:harmonic}, which results in the following two relationships:
\begin{equation}
C_{dw}\sin{\phi_{dw}} = 4\upi f^*A^* \left(C_d + \frac{\upi^2m^*\zeta}{U^*}\right), \label{eq:Cdvsin2}\\
\end{equation}
\begin{equation}
C_{dw}\cos{\phi_{dw}} = 2\upi^3\frac{m^*A^*}{U^{*2}}\left[1 - \left(\frac{f}{f_{n,a}}\right)^2 \right]. \label{eq:Cdvcos2}
\end{equation}
When the structural damping is zero $(\zeta=0)$, the solution of equation (\ref{eq:Cdvsin2}) apparently does not depend on $m^*$. It should be noted however that the feasible operating points in the $A^*:f^*$ space are limited to those that satisfy equation~(\ref{eq:Cdvsin2}), which correspond to the contour of zero energy transfer $(\sin\phi_x=0)$.
In addition, equation (\ref{eq:Cdvcos2}) becomes indeterminate at the resonance point where $f=f_{n,a}$ and $\phi_{dw}=90^\circ$ simultaneously. It should be noted that there is no \textit{a priori} guarantee that the `resonance point' is attainable since the solution pursued here is open form. Nevertheless, as we can see from the simulations the resonance point is attained and corresponds to the maximum amplitude, in which case we can write equation (\ref{eq:Cdvsin2}) as
\begin{equation}\label{eq:peak}
A^*_\mathrm{max} = \frac{C_{dw}}{4\upi f^*C_d}.
\end{equation}
This result shows that the steady-state solution at the resonance point does not depend on $m^*$; rather the operating point is solely determined by the fluid dynamics.
When the structural damping is finite positive $(\zeta>0)$, equation~(\ref{eq:peak}) may still be used to estimate approximately the peak amplitude for very low values of the mass and the damping such that $\upi^2m^*\zeta/U^* \ll C_d$, on the condition that the resonance point is attained. The early experiments of \citet{Aguirre1977} support the above inference; he observed little variation of the peak amplitude in the second excitation region, i.e.\ the one associated with alternating vortex shedding, which occurred at $U_a^*\approx 1.2/(2S)$ for $m^*=1.2$ and 4.3 (see figure 43 of his thesis). The mass--damping was reported in terms of a stability parameter $k'_{s0}$ to be approximately 0.5, which was estimated from free-decay tests in still water, i.e.\ it includes contributions from hydrodynamic damping from the cylinder as well as its mountings. From his data, we estimate here that the corresponding $k_s$ value in air was significantly lower than the one measured in water so that we can assume $\upi^2m^*\zeta/U^*<0.05$; this is small compared to the mean drag coefficient $C_d$. Thus, we argue that the present analysis is consistent with the experimental facts at Reynolds numbers higher than considered in this study.
At the resonance point ($\phi_{dw}=90^\circ$), equations (\ref{eq:Cdvsin}) and (\ref{eq:Cdvcos}) reduce to
\begin{eqnarray}
C_{dw} & = & 4\upi f^*A^* C_d, \label{eq:Cdvsin3}\\
C_{x1} & = & 2\upi^3f^{*2}A^*C_a. \label{eq:Cdvcos3}
\end{eqnarray}
The above set of equations establishes relationships between the wake, mean, and total drag coefficients at resonance point, which involve variations in $A^*$ and $f^*$.
Considering that all three force coefficients, $C_{dw}$, $C_{d}$, and $C_{x1}$, may be `uniquely' specified in the parameter space $A^*:f^*$, this could suggest a single operating point inside this space at which a steady-state harmonic solution is feasible, which implies that the cylinder peak response is drawn towards this operating point irrespectively of the mass ratio. Nonetheless, it should be cautioned that bimodal dynamics can also appear at particular regions of the parameter space $A^*:f^*$ \citep{Cagney2013pof,Cagney2013a,Gurian2019}. Assuming that a unique solution exists at the resonance point, equation~(\ref{eq:Cdvsin3}) illustrates that the component of the wake drag in-phase with the cylinder velocity counterbalances the quasi-steady drag, whereas equation~(\ref{eq:Cdvcos3}) illustrates that the only contribution of the fluid force in-phase with the cylinder acceleration is the inviscid added-mass force. This is commensurate to the situation where an elastically mounted cylinder oscillates freely within a fluid medium with no net viscous forces; in such a situation, it would be anticipated that the frequency of cylinder oscillation is equal to the natural frequency of the structure in an inviscid fluid, i.e.\ $f=f_{n,a}$. This is as if the fluid was phenomenologically interacting with the cylinder motion only through its ideal inertia. It should be emphasized that in the case of a viscous fluid, the zero net viscous force results from the cancelling out of the wake drag and the quasi-steady drag, which is brought about by a gradual change in the phasing of the vortex shedding, thereby in $\phi_{dw}$, as the reduced velocity is varied.
The above changes can be viewed from another perspective, which is more insightful. As $U^*$ increases, the phasing of vortex shedding gradually shifts to follow the changes in the frequency of oscillation with the reduced velocity. Since these variations occur in a continuous manner for the low-$Re$ cases considered here, a point is reached where the timing of vortex shedding induces a wake drag exactly in-phase with the cylinder velocity, $\phi_{dw}=90^\circ$. At this point, the net viscous force must become null (Eq.~\ref{eq:Cdvsin3}), whereas only the added mass contributes to the force in-phase with the cylinder acceleration (Eq.~\ref{eq:Cdvcos3}). It is important to note again that these changes are brought in entirely by the fluid dynamics.
\subsection{Infinitesimal net force at coincidence point}
At first glance, it seems counter-intuitive that the cylinder experiences almost no net force at the coincidence point. However, this can be explained by the cancelling out of the fluid-force components. It should be remembered that when the structural damping is zero $(\zeta=0)$ the linearised solution of the equation of cylinder motion (Eq.~\ref{eq:Ctotal}) shows that the component of the streamwise force at the main harmonic is null $(C_{x1}=0)$ at the coincidence point $(f^*U^*=1)$. In this case, equations (\ref{eq:Cdvsin}) and (\ref{eq:Cdvcos}) reduce to
\begin{eqnarray}
C_{dw}\sin{\phi_{dw}} & = & 4\upi f^*A^* C_d, \label{eq:Cdvsin4} \\
-C_{dw}\cos{\phi_{dw}} & = & 2\upi^3f^{*2}A^*C_a. \label{eq:Cdvcos4}
\end{eqnarray}
From these equations, we can see that the wake drag in-phase with the velocity of the cylinder balances the quasi-steady drag while the wake drag in-phase with acceleration balances the inviscid inertia. This is physically possible as these force contributions originate from different flow structures. Furthermore, the cylinder must be oscillating for this case to be realized; the terms on the right-hand side of equations (\ref{eq:Cdvsin4}) and (\ref{eq:Cdvcos4}) are null for a non-oscillating cylinder. Although the above scenario is idealised as some damping will be present in a real situation, extending the analysis for a system with very low structural damping shows that the cylinder will experience a very low net in-line force at the coincidence point, if this point can still be attained. It should be noted that we approached very close to the coincidence point by employing a very fine step around this region, but exact coincidence was unattainable. A similar observation was made by \citet{Shiels2001} who showed that the net transverse force $C_y(t)$ exerted on a cylinder undergoing free vibration in the transverse direction becomes null at limiting values of the structural parameters, $m=c=k=0$. Here, we show that a null net streamwise force $C_x(t)$ on a cylinder undergoing free in-line vibration is physically possible at the coincidence point for finite $m$ and $k$ values.
\subsection{$A^*_\mathrm{max}$ as a function of $Re$ in the laminar regime}
\begin{table}
\begin{center}
\def~{\hphantom{0}}
\small\addtolength{\tabcolsep}{16pt}
\begin{tabular}{lcc}
$Re$ & $SC_{d}$ & $(f^*/2)C_d$\\
100 & 0.217 & 0.216 \\
180 & 0.252 & 0.244 \\
\end{tabular}
\caption{\label{tab:SCd} Compilation of data for stationary and vibrating circular cylinders in the laminar regime. Data for the stationary cylinder are from the study of \citet{Qu2013} and for the freely vibrating cylinder are from the present simulations at peak amplitude.}
\end{center}
\end{table}
The variation of the peak amplitude as a function of the Reynolds number can be estimated using the following assumptions. The mean drag coefficient $C_d$ and the inverse of the Strouhal number $1/S$ for stationary bluff bodies display similar variations as functions of $Re$ so that the product $SC_d$ remains almost constant \citep{Alam2008}. For vibrating cylinders, there is also evidence that the corresponding product $(f^*/2)C_d$ remains constant \citep{Griffin1978}. \citet{Alam2008} showed that $SC_d\approx0.25$ for stationary bluff bodies of various cross-sectional shapes throughout the sub-critical regime. However, $SC_d$ appears to increase slightly with $Re$ in the laminar regime as seen in their data compilation (see their figure 5). We have confirmed this dependency on $Re$ in the laminar regime for the stationary as well as the freely-vibrating circular cylinder as shown in table~\ref{tab:SCd}. Nevertheless, the variation of $SC_d$ and $(f^*/2)C_d$ with $Re$ is small, and in order to keep some generality in the result, we assume that $f^*C_d$ remains approximately constant. In this case, equation (\ref{eq:peak}) shows that $A^*_\mathrm{max}\propto C_{dw}$. In addition, we assume $C_{dw}$ is magnified by some constant factor at resonance so that $A^*_\mathrm{max}\propto C_{dw0}$, where $C_{dw0}$ is the fluctuating drag coefficient for a stationary cylinder. Then, peak amplitudes at different $Re$ can be estimated from data at one particular Reynolds number $Re_0$, i.e.
\begin{equation}\label{eq:Apredict}
A^*_\mathrm{max}(Re) = \left( \frac{A^*_\mathrm{max}}{C_{dw0}} \right)_{Re_0} C_{dw0}(Re).
\end{equation}
Figure \ref{fig:Amax} shows predictions of $A^*_\mathrm{max}$ using $C_{dw0}$ data at different $Re$ \citep[taken from][]{Qu2013} and the known response at $Re=180$. It can be seen that $A^*_\mathrm{max}$ predictions fit well with the amplitude from the simulations at $Re=100$ while they also extrapolate to the result at $Re=250$. The enhancement of $A^*_\mathrm{max}$ with $Re$ is close to quadratic, which is quite marked compared to the nearly constant amplitude of purely transverse vortex-induced vibration in the laminar regime \citep[see][]{Govardhan2006}.
\begin{figure}
\centerline{\includegraphics[width=0.82\textwidth]{Fig13.eps}}
\caption{Comparison of $A^*_\mathrm{max}$ as a function of $Re$ obtained from the simulations (open symbols) and predicted from equation (\ref{eq:Apredict}) (full symbols).} \label{fig:Amax}
\end{figure}
\section{Conclusions}
In this study we have developed a theoretical model for the fluid force acting on a circular cylinder vibrating in-line with a free stream. The streamwise fluid force comprises the sum of three components: an inviscid added-mass force opposing the inertia of the cylinder, a quasi-steady drag force opposing the velocity of the cylinder, and a wake drag force.
The key element here is the splitting of the viscous force into quasi-steady and wake components, which stem from the contribution of vorticity in fairly-separable flow regions, i.e.\ in the boundary/free shear layers and in the near wake, respectively.
The theory enabled us to decipher three important fluid-dynamical aspects. First, the phase angle of the vortex force with respect to the cylinder displacement $\phi_{dw}$ increases smoothly with $U^*$ due to a gradual shift in the timing of vortex shedding. This contrasts the variation of the phase angle between the total force and the cylinder displacement, which must remain fixed at $\phi_x=0^\circ$ for $U^*f^*<1$ and at $\phi_x=180^\circ$ for $U^*f^*>1$, with a sudden jump at the coincidence point $(U^*f^*=1)$; this constraint is imposed by the equation of cylinder motion when the structural damping is null and has no bearing on the fluid dynamics. Second, the added mass coefficient for inviscid flow is also applicable for viscous flow about a cylinder oscillating in-line with a free stream. Third, the wake drag is amplified in the excitation region of relatively high-amplitude response, which supports the classification of vortex-induced in-line vibration as a resonance phenomenon.
The above findings could be confirmed because, unlike previous works dealing with vortex-induced vibration that primarily consider the fluid force in the direction of cylinder motion, here we calculated the phase angle between the force transverse to the direction of motion and the cylinder displacement, $\phi_y$. This allowed us to illustrate that the smooth variation of $\phi_y$ as a function of $U^*$, in contrast to $\phi_x$ that displays a sudden jump by $180^\circ$, is directly linked to a gradual shift in the timing of vortex shedding.
The theory developed in this investigation is linear, which means that all force components can be well approximated by single-harmonic functions of time, although the fluid dynamics is non linear by default. Departures from linearity can arise at much higher amplitudes and/or frequencies of cylinder vibration because of the quadratic relative velocity term. In addition, other non linearities may arise due to competing requirements posed by the structural dynamics and the fluid dynamics. In fact, we have pinpointed in the present study that such non-linear effects become apparent at the `coincidence point' where the vibration frequency becomes equal to the natural frequency of the structure in vacuum. At this point, the equation of cylinder motion requires that the component of the in-line force at the main harmonic of the vibration becomes almost null, which results from balances between the quasi-steady drag and the wake drag in-phase with the velocity, and between the inviscid inertia and the wake drag in-phase with acceleration. In this case, the component at the first super-harmonic of the vibration frequency dominates the driving in-line force.
An important result predicted from the theory is that the response does not depend on $m^*$ at the point where $\phi_{dw}=90^\circ$. The simulations show that this occurs at the `resonance point' where the same peak amplitude is attained for all $m^*$ values, in accord with the theory. It should be noted however that the theory cannot predict that the maximum amplitude occurs when $\phi_{dw}=90^\circ$, or that this operating point does materialize, since the solution is given in open form.
On the other hand, the flow physics suggest that at some point $\phi_{dw}=90^\circ$ due to the gradual shift in the timing of vortex shedding, which accompanies the continuous variation of the vibration frequency as the reduced velocity is varied. Nevertheless, we have also observed that discontinuities may arise by changing some parameter, e.g.\ $Re$ or $m^*$. Further study over wider ranges of these parameters than considered here may worth undertaking in the future.
The simulations show that Reynolds number has a remarkable effect on the maximum amplitudes $A^*_\mathrm{max}$ attained over the entire reduced velocity range; increasing $Re$ from 100 to 250 results in a 12-fold increase of $A^*_\mathrm{max}$. This stands in sharp contrast to the free vibration transversely to a free stream, in which case peak amplitudes remain fairly constant in the laminar regime. This may be attributable to variations of the added mass coefficient in the latter configuration \citep{Konstantinidis2013prsa}. In view of this complexity, we consider that free in-line vibration offers a more convenient test case to uncover the fluid dynamics.
The fluid dynamics could be elucidated because in-line response amplitudes remain very small for the low Reynolds numbers investigated in the present study. As a consequence, the fluid excitation comes solely from the primary wake instability associated with alternating vortex shedding, which remains robust and similar as in the wake of a non-vibrating cylinder. It has been well established in the published literature that other instabilities can be excited, even at small amplitudes, with increasing the Reynolds number, such as the symmetrical mode of vortex shedding. There may also exist competition between different modes. Further instabilities due to gradual transition to turbulence will inevitably perplex the phenomenology of vortex-induced vibration. However, we maintain that the theory developed here remains valid and can be used to analyse the more complex phenomena at higher Reynolds numbers, possibly with some adjustments for different fluid excitation mechanisms.
\vspace{3pt}
{\bfseries Acknowledgements.}
This research was supported by the European Union and the Hungarian State, co-financed by the European Regional Development Fund in the framework of the GINOP-2.3.4-15-2016-00004 project, aimed to promote the cooperation between the higher education and the industry.
\vspace{3pt}
{\bfseries Declaration of Interests.} The authors report no conflict of interest.
| 2e5881871b0c8485777f07976e8ec9466e139ef7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Inflation in the early universe (see e.g.~\cite{Sato:2015dga} for a review) practically realizes isotropic universe even if we start from high degree of anisotropy.
This is guaranteed by Wald's cosmic no hair theorem of the universe with a positive cosmological constant
which applies under several conditions such as the strong and dominant energy conditions~\cite{Wald:1983ky}.
To realize anisotropic inflation, one must therefore introduce anisotropic matter content breaking
the strong energy condition such as a vector field~\cite{Watanabe:2009ct}.
The above is the case of general relativity,
and the story is totally different in generalized scalar-field theories with nonminimal derivative coupling to the metric
which is provided by the generalized Galileon~\cite{Deffayet:2011gz}
or the Horndeski theory~\cite{Horndeski:1974wa,Kobayashi:2011nu,Kobayashi:2019hrl}.
Their higher-order curvature term allows the system to approach a nontrivial attractor
and to have anisotropic expansion rates~\cite{Tahara:2018orv}.
The universe can be spontaneously anisotropized in the presence of such a higher-order curvature term.
Expansion rates of two spatial dimensions vanish in the limit of large coefficient of the higher-order curvature term.
The purpose of this letter is to show that such an anisotropic attractor also exists in higher dimensional models
and it can be used to stabilize the extra dimensions.
Higher dimensions may play an important role in the unification of fundamental forces.
A pioneering attempt using a higher dimension has been given by Kaluza and Klein~\cite{Kaluza:1984ws, Klein:1926tv}.
Superstring theory also requires ten-dimensional spacetime to
provide a consistent theory of quantum gravity~\cite{Green:1987sp, Green:1987mn}.
Observationally the dynamics of extra dimensionas would manifest itself in time variation of fundamental constants
such as Newton's gravitaional constant and the fine structure constant, which have been severely constrained~\cite{Uzan:2010pm}.
Hence the extra dimensions must be stabilized in sensible higher-dimensional theories.
In superstring theories, flux compactification is used for this purpose (\textit{e.g.}~\cite{Douglas:2006es}).
Here we will propose a different way to \textit{freeze} the extra dimensions with higher-order curvature term
in the Lovelock theory~\cite{Lovelock:1971yv},
which is a special case of the Horndeski or generalized Galileon theories.
The Lovelock theory is the most general gravity theory which consists only of metric degrees of freedom.
Its evolution equations contain up to the second derivative but it is not always linear.
In four dimensions, the Lovelock theory is reduced to general relativity with a cosmological constant.
There have been several studies on the cosmological dynamics in higher dimensions in the Lovelock theory~\cite{MuellerHoissen:1985mm, Deruelle:1989fj, MenaMarugan:1992ig, Maeda:2004hu,
Akune:2006dg, Canfora:2008iu, Chirkov:2014nua, Chirkov:2015kja, Kastor:2015sxa, Canfora:2016umq, Chirkov:2019qbe},
and they have not manifestly pointed out the presense of nontrivial attractors.
\section{Anisotropic attractor}
\subsection{Lovelock theory}
We start with the Lovelock action up to the cubic order with matter Lagrangian $\mathcal{L}_\textrm{mat}$
\bal
&S + S_\mathrm{mat} = \int d^{(D+1)}x \sqrt{-g} \left( \mathcal{L} + \mathcal{L}_\mathrm{mat} \right) \label{LovelockAction}
\\
& { \mathcal{L } \equiv -\Lambda + \sum_{m=1}^{3} \frac{2m-1}{(2m)!} \kappa_m^{-1} {\mathcal{L}_m} }
\\
&\mathcal{L}_m \equiv 2^{-m} \delta^{\mu_1 \mu_2 \ldots \mu_{2m}}_{\nu_1 \nu_2 \ldots \nu_{2m}} \prod_{i=1}^{m} R_{\mu_{2i-1} \mu_{2i}}^{\nu_{2i-1} \nu_{2i}}
\eal
where $R_{\mu_{2i-1} \mu_{2i}}^{\nu_{2i-1} \nu_{2i}}$ is the Riemann curvature tensor
and $\delta^{\mu_1 \cdots \mu_{N} }_{\nu_1 \cdots \nu_{N} }$
is the generalized Kronecker delta.
We have defined $\kappa_m$ as constants with mass dimension $2m-D-1$.
We assume that the space is flat and use the Kasner metric
\begin{align}
ds^2= - dt^2 + a_{(i)}^2 \delta_{ij} dx^i dx^j ,
\label{metricinhigherdim}
\end{align}
where $a_{(i)}=a_{(i)}(t)$ is the scale factor in the direction of $x^i$.
This leads to $\mathcal{L}=-\Lambda-\sum_{m} \kappa_m^{-1} s_{2m}$,
where $s_{2m}$ is the $2m$-th order symmetric polynomial of $H_{(i)}$'s.
From Eqs.~\eqref{LovelockAction} to~\eqref{metricinhigherdim},
we obtain the Hamiltonian constraint and evolution equations
\bal
&\rho + \Lambda - \sum_{m=1}^{3} (2m-1) \kappa_{m}^{-1} s_{2m} = 0,
\label{Hamiltonianeqwithmatter}
\eal
\bal
{p_{(i)}} + \mathcal{L}
+ \frac{1}{V} \frac{d}{dt} \left\{ V \sum_{m=1}^{3} \kappa_{m}^{-1} \frac{\partial s_{2m}}{\partial H_{(i)}} \right\} = 0
\nonumber \\
\qquad \textrm{~for~} {i=1,\ldots,D},
\label{evolutioneqwithmatter}
\eal
where
$V\equiv\prod_{i} a_{(i)}$ is the volume factor,
$H_{(i)} \equiv {\dot a_{(i)}}/{a_{(i)}}$ is the expansion rate along $x^i$-axis,
and $\rho$ and $p_{(i)}$ are defined by
\bal
\rho \equiv \frac{2 g_{00}}{\sqrt{-g}} \frac{\delta S_\mathrm{mat}}{\delta g_{00}},
\quad
p_{(i)} \equiv \frac{2 g_{ii}}{\sqrt{-g}} \frac{\delta S_\mathrm{mat}}{\delta g_{ii}}.
\eal
{In this letter, we consider the cases that the stress-energy tensor is diagonal,
which is represented only by $\rho$ and $p_{(i)}$.}
The $(2m)$-th symmetric polynomial $s_{2m}$ is explicitly defined as
\begin{align}
&s_2 = \sum_{1 \le i < j \le D} H_{(i)} H_{(j)} ,
\\
&s_4 = \sum_{1 \le i < j < k < l \le D} H_{(i)} H_{(j)} H_{(k)} H_{(l)} ,
\\
&s_6 = \sum_{1 \le i < j < k < l < m < n \le D} H_{(i)} H_{(j)} H_{(k)} H_{(l)} H_{(m)} H_{(n)}.
\end{align}
\subsection{Attractors under isotropic pressure}
We will show that there exist attractors
if the spacetime is filled with energy contents with isotropic pressure, which means $p_{(i)} = p$ for all $i$.
Subtracting different components of Eqs.~\eqref{evolutioneqwithmatter}, we yield
\begin{align}
\frac{d}{dt} \left\{ V \left( H_{(j)} - H_{(k)} \right)
\sum_{m=1}^{3} \kappa_{m}^{-1} \frac{\partial^2 s_{2m}}{\partial H_{(j)} \partial H_{(k)}} \right\} = 0
\nonumber \\
\textrm{~~~~for $1 \le j < k \le D$} .
\label{subeq}
\end{align}
We integrate these and get their solutions
\begin{align}
\left( H_{(j)} - H_{(k)} \right)
\sum_{m=1}^{3} \kappa_{m}^{-1} \frac{\partial^2 s_{2m}}{\partial H_{(j)} \partial H_{(k)}} = \frac{ \mathcal{A}_{jk} }{V}
\nonumber\\
\textrm{~~~~for $1 \le j < k \le D$} ,
\label{subeq2}
\end{align}
where $\mathcal{A}_{jk}'s$ are integration constants.
In principle, full set of Eqs.~\eqref{Hamiltonianeqwithmatter} and \eqref{evolutioneqwithmatter}
describes evolution of the system in the phase space $\{ a_{(i)} , H_{(i)} \}$,
but we try to see what the subset \eqref{subeq} or \eqref{subeq2} indicates.
The left-hand-side of Eq.~\eqref{subeq2} is a function of only $H_{(i)}$'s, and if we know how $V$ evolves,
we can track the evolution of the system in a phase space $\{ H_{(i)} \}$.
When $V$ is an increasing function of time,
the system approaches the regions where the following equations are satisfied.
\begin{align}
\left( H_{(j)} - H_{(k)} \right) & \sum_{m=1}^{3} \kappa_{m}^{-1} \frac{\partial^2 s_{2m}}{\partial H_{(j)} \partial H_{(k)}} = 0 .
\label{fixedpointeq}
\end{align}
A similar equation has been derived in \cite{Chirkov:2015kja} under the assumption that all $H_{(i)}$'s are constant.
In deriving Eq.~\eqref{fixedpointeq}, however, we do not need to restrict $H_{(i)}$'s to be constant, so that we can trace the entire evolution
of the universe as we will see below.
We name those regions ``{(an)isotropic attractors}'',
although we must carefully analyze the system to see whether they really act as an attractor.
\subsection{Classification of attractors}
Here we classify the roots of Eq.~\eqref{fixedpointeq} into several types.
On these roots, some of $H_{(i)}$'s have the same value, and thus we label the attractors with ${N}_\textrm{diff}$,
which denotes the number of different values of $H_{(i)}$'s.
\paragraph{Isotropic case $(N_{\rm diff}=1)$}
When all of $H_{(i)}$'s are the same, all of the equations \eqref{fixedpointeq} are trivially satisfied.
The universe exhibits isotropic expansion and we just call it isotropic attractor.
\paragraph{Anisotropic case $(N_{\rm diff} \ge 2)$}
For the simplest departure from isotropy, we consider the case
in which $H_{(i)}$'s take two different values, $\alpha$ and $\beta$.
\begin{align}
H_{(i)} &= \alpha \textrm{~~~~for~} 1\le i \le d, \label{subalpha}
\\
H_{(i)} &= \beta \textrm{~~~~for~} d+1 \le i \le D , \label{subbeta}
\end{align}
where $d$ is any integer satisfying $2 \le d \le D-1$.
This corresponds to the case for $N_{\rm diff} = 2$.
Then Eq.~\eqref{fixedpointeq} gives the relation between $\alpha$ and $\beta$
\begin{align}
\sum_{m=1}^{3} \kappa_{m}^{-1} {Q}_{m}(\alpha,\beta) = 0 ,
\label{alphabeta}
\end{align}
where
\begin{align}
{Q}_{m}(\alpha,\beta) &\equiv \left. \frac{\partial^2 s_{2m}}{\partial H_{(j)} \partial H_{(k)}} \right|_{H_{(1\le j \le d)}=\alpha, H_{(d+1 \le k \le D)}=\beta }
\nonumber \\
&= \sum_{l=0}^{2m-2} \left(\begin{array}{@{}c@{}} d-1 \\ l \end{array} \right)
\left(\begin{array}{@{}c@{}} D-d-1 \\ 2m-2-l \end{array} \right) \alpha^l \beta^{2m-2-l}.
\end{align}
The discussion above can be generalized straightforwardly to the cases with larger $N_\textrm{diff}<D$.
In this letter, we focus on the case for $N_\textrm{diff}=2$ for simplicity.
\section{Large universe and frozen extra dimensions} \label{secFreezing}
\subsection{Evolution scenario}
In order to compactify the extra dimensions successfully,
we have to explain both why they are small and stable.
Otherwise, we could observe Kaluza-Klein particle in a particle accelerator or varying Newton's constant~\cite{Uzan:2010pm}.
We try to explain those properties by realizing $\beta \ll \alpha$ in the whole cosmic history from inflation,
where we regard $\alpha$ and $\beta$ as the expansion rate of $(d+1)$-dimensional universe and the extra $(D-d)$ dimensions.
We consider the standard history of inflationary cosmology.
That is, we set the initial condition at the beginning of inflation,
followed by reheating regime dominated by coherent field oscillation of the inflaton
in the case of standard potential-driven models or by kinetic energy in the case of
k- or G-inflation models~\cite{ArmendarizPicon:1999rj, Kobayashi:2010cm}.
Then the universe turns to be radiation dominant with a reheating temperature $T_R$
to trace the standard thermal history of the universe in Big Bang cosmology toward matter domination
and dark energy domination.
We set the topology of space as $D$-dimensional tori and assume that the size or period of each dimension
is of the same order of magnitude given by $\ell$, simply from democratic viewpoint,
at the beginning of inflation of $d$-dimensional space with the Hubble parameter $\alpha$.
We assume that inflation and subsequent reheating are realized by an appropriate scalar-field model
which has a practically homogeneous configuration over $T^D$ initially and
anisotropic inflation is realized in a similar manner as in the previous paper~\cite{Tahara:2018orv}.
This initial condition is no less natural than required in conventional three dimensional inflation models
if $\ell$ is of the order of $\alpha^{-1}$ or smaller.
With $\beta \ll \alpha$ for a sufficiently long time $\tau \gg \alpha^{-1}$,
the $d$-dimensional space becomes exponentially large while
freezing the extra dimensions practically.
After inflation the universe will be eventually dominated by radiation through reheating process
intrinsic to each inflation model.
For simplicity, we consider the case that reheating temperature $T_R$ is much smaller than
the inverse size of the extra dimension, $\ell^{-1}$, so that equation-of-state parameter $w$ is
anisotropic in this regime, namely, $p_{(i)}/\rho=1/d$ in large $d$ dimensions,
and $p_{(i)}=0$ in the small extra dimensions.
As the universe becomes matter ($w=0$) or dark-energy ($w\approx -1$) dominant,
the equation of state becomes isotropic again.
We will concentrate on the case for ten-dimensional spacetime with $D=9$ and $d=3$, since it is of our most interest.
First, we consider condition for freezing extra dimensions $\beta \ll \alpha$ when pressure is isotropic.
We need other treatment in the radiation-dominated era and we will find that the radiation also makes
expansion rate of the extra dimensions be strongly suppressed.
\subsection{Under isotropic pressure} \label{subsecIsotropic}
First we consider an anisotropic attractor under isotropic pressure.
The anisotropic attractor is allowed to exist when the pressure is isotropic.
Let us investigate the roots of Eq.~\eqref{alphabeta} for $\beta$
and show some of them can satisfy $\beta \ll \alpha$.
When the root $\beta=\beta(\alpha)$ is much smaller than $\alpha$,
$\beta(\alpha)$ is given by solving
\bal
\kappa_1^{-1} + \kappa_2^{-1} \alpha^2 + 10 \kappa_3^{-1} \alpha^2 \beta^2 = 0,
\label{alphabetaAsymptotic}
\eal
which we get by neglecting higher-order terms of $\beta$ from Eq.~\eqref{alphabeta}.
We immediately obtain
\bal
\beta(\alpha) = \pm \sqrt{ - \frac{ \kappa_1^{-1} + \kappa_2^{-1} \alpha^2 }{ 10 \kappa_3^{-1} \alpha^2 } },
\label{BetaOfAlpha}
\eal
where $\kappa_3 < 0$ is assumed.
We now define lower bound of $\alpha$ as $\alpha_\textrm{min}$,
since our Universe seems to expand exponentially today.
We require $\beta(\alpha) \ll \alpha$ to be satisfied for any $\alpha \ge \alpha_\textrm{min}$
for consistency.
From Eq.~\eqref{BetaOfAlpha}, we get
\bal
- \kappa_3/\kappa_1 \ll \alpha_\textrm{min}^4 , \qquad - \kappa_3/\kappa_2 \ll \alpha_\textrm{min}^2.
\label{ConsistencyCondition}
\eal
Substituting $\beta=\beta(\alpha)$ into the Hamiltonian constraint~\eqref{Hamiltonianeqwithmatter},
we get the effective constraint at the leading order
\bal
{\rho} + \Lambda - 3 \kappa_{1}^{-1} \alpha^2
- 8 \kappa_{2}^{-1} \alpha^3 \beta(\alpha) = 0. \label{EffectiveConstraint}
\eal
This corresponds to the Friedmann equation with an additional term.
It brings us a new effect, which may matter in the very early universe unless $\kappa_2^{-1} = 0$.
When $\kappa_1^{-1}$ and $\kappa_2^{-1} \alpha \beta(\alpha)$ are comparable and $\rho + \Lambda$ is negligible,
the universe expands with almost constant rate
\bal
\alpha = \mp \sqrt{ - \frac{ 45 \kappa_2^{3} }{ 32 \kappa_1^{2} \kappa_3 } } \times \mathrm{sign}(\kappa_1 \kappa_2).
\eal
If we can neglect the additional term for any $\alpha$ of interest, the observed universe always obeys
the same Friedmann equation as in general relativity.
In this case, since $\beta \ll \alpha$,
only $\alpha$ affects the evolution of density of energy contents and nothing feels extra dimensions.
In the limit $\kappa_ 3\to 0$, we have $\beta(\alpha) \to 0$ for fixed $\alpha$.
This means that it is possible to realize arbitrarily small $\beta$.
In the limit $\alpha \to \infty$, $\beta(\alpha)$ converges to a finite value,
and thus we do not suffer from divergent behavior for larger $\alpha$.
Therefore, we can slow down the extra dimensions arbitrarily even during inflation,
as long as isotropic pressure dominates the higher-dimensional spacetime.
\subsection{Radiation-dominated universe} \label{sectionExtraFreezeAnisotropic}
Next, we consider a radiation-dominated universe,
where we no longer use the roots \eqref{BetaOfAlpha}.
We continue using $\alpha$ and $\beta$ instead of $H_{(i)}$,
since the expantion rates tend to take two values as in Eqs.~\eqref{subalpha} and \eqref{subbeta}
even after entering the radiation-dominated era.
For $D=9$ and $d=3$, if $\kappa_2^{-1}=0$ then radiation allows $\beta=\dot \beta=0$ to be a solution, that is,
Eqs.~\eqref{Hamiltonianeqwithmatter} and \eqref{evolutioneqwithmatter} yield
\bal
\dot \beta(\alpha,\beta)|_{\beta=0} = 0 ,
\qquad
\left. \frac{\partial }{\partial \beta} \dot \beta(\alpha,\beta) \right|_{\beta=0} = -3\alpha .
\eal
The minus sign of the first derivative shows this solution is stable as long as the universe is expanding.
If $\kappa_2^{-1} \ne 0$, the solution shifts like ($\beta=\beta_0(\alpha)$, $ \dot \beta=0$).
In general, it is difficult to find an analytic form of $\beta_0$,
but its approximate value can be estimated as
\bal
\beta_0(\alpha) &\approx - \left.{\dot \beta(\alpha,\beta)}/{\tfrac{\partial }{\partial \beta} \dot \beta(\alpha,\beta) }
\right|_{\beta=0}
\nonumber\\
&= - \frac{\kappa_1^{-1}\kappa_2^{-1} \alpha^3
(4\kappa_1^{-2} + 3\kappa_1^{-1}\kappa_2^{-1}\alpha^2+9\kappa_2^{-2}\alpha^4) }
{ 16\kappa_1^{-4} + 56\kappa_1^{-3}\kappa_2^{-1}\alpha^2 - 9\kappa_1^{-2}\kappa_2^{-2}\alpha^4
+ 18 \kappa_1^{-1}\kappa_2^{-3}\alpha^6 + 27 \kappa_2^{-4}\alpha^8
+ 20 \kappa_1^{-2}\kappa_2^{-1}\kappa_3^{-1} \alpha^6}.
\eal
This expression typically leads to $\beta_0(\alpha) = \mathcal{O} (\beta(\alpha)^2/\alpha)$,
where $\beta(\alpha)$ is the function defined in Eq.~\eqref{BetaOfAlpha}.
This means that in the radiation-dominated era, the expansion rate of the extra dimensions is suppressed more strongly
than in the isotropic-pressure-dominated case with the same value of $\rho$.
Taking this suppression into account,
we now succeed in freezing the extra dimensions for the whole standard history.
\section{Discussion}\label{secDiscussion}
As we have seen in the case under isotroic pressure,
small $\kappa_3$ plays a key role to freeze extra dimensions.
On the other hand, nonvanishing $\kappa_2^{-1}$ yields the additional term in Eq.~\eqref{EffectiveConstraint}.
It can cause exponential expansion in the very early universe
where the propagation speed of gravitational waves can be varied.
This is worth considering but it is beyond the scope of this letter
and we avoid this exponential expansion by taking $\kappa_2 \gg \kappa_1 \alpha_\textrm{max}^2$
in the following discussion,
where $\alpha_\textrm{max}$ is the maximum of $\alpha$ of interest.
The $d$-dimensional effective gravitational constant depends on $\beta$ as $\dot G/G = (d-D) \beta$.
From the constraint $|\dot G/G| = (4\pm 9) \times 10^{-13} ~\textrm{yr}^{-1} $ given by the Lunar laser ranging experiment~\cite{Williams:2004qba} implies
that we have to require that in the late universe
\bal
| \beta | \lesssim 10^{-13} ~\textrm{yr}^{-1} \sim 10^{-3} H_0,
\eal
where $H_0$ is the present Hubble constant.
From Eq.~\eqref{BetaOfAlpha}, we set $\alpha=H_0$ and get
\bal
- \kappa_3/\kappa_1 \lesssim 10^{-{5}} H_0^4.
\eal
Note that although the effective gravitational constant in the Friedmann equation is
different in principle from Newton's gravitational constant in the Poisson equation of the gravitational potential
which is constrained in~\cite{Williams:2004qba},
those constants are approximately the same if the extra dimensions are frozen $\beta \ll \alpha$
and if their size is much smaller than the scale of observation.
When $\kappa_3$ is small, one might think that such a large higher-order term could
decrease the energy scale of the unitarity bound of the model.
To show that it is not the case around the anisotropic attractor,
we give a simple estimation of interaction of gravitons $h_{ij}$
on the unperturbed metric $\bar{g}_{\mu\nu}$ which is equal to the metric in Eq.~\eqref{metricinhigherdim}.
We compare the following coefficients of the perturbed Riemann tensor $\delta {R}_{\mu_1\mu_2}^{ \nu_1\nu_2}$
in the Lagrangian
\bal
\kappa_1^{-1} \delta^{\mu_1\mu_2}_{\nu_1\nu_2}
, \qquad
\kappa_3^{-1} \delta^{\mu_1\mu_2 \cdots \mu_6}_{ \nu_1\nu_2 \cdots \nu_6} \bar{R}_{\mu_3\mu_4}^{ \nu_3\nu_4} \bar{R}_{\mu_5\mu_6}^{ \nu_5\nu_6},
\eal
where $\bar{R}_{\mu_1\mu_2}^{ \nu_1\nu_2}$ is the unperturbed Riemann tensor.
The former are just of the order of $\kappa_1^{-1}$.
Since $h_{ij}$ appears in $\delta R_{i_1 i_2}^{j_1 j_2}$ and $\delta R_{0i}^{0j}$,
$\kappa_3^{\!-1} \! \delta^{0 i i_1 i_2 M_1 M_2}_{0 j j_1j_2 N_1 N_2} \bar{R}_{i_1 i_2}^{j_1 j_2} \bar{R}_{M_1 M_2}^{ N_1 N_2}$ and
$\kappa_3^{\!-1} \! \delta^{0 i i_1 i_2 M_1 M_2}_{0 j j_1j_2 N_1 N_2} \bar{R}_{0 i}^{0 j} \bar{R}_{M_1 M_2}^{ N_1 N_2}$ are largest of all possibilities, where $M,N,\ldots$ denote the indices for the extra dimensions.
These are of the order of $\kappa_3^{-1} \alpha^2 \beta^2$, which is of the same order of $\kappa_1^{-1} $
as long as $\kappa_2 \gg \kappa_1 \alpha^2$.
Therefore the self-couplings of gravitons is not so larger than in the general relativity.
Perturbation theory does not break down until around the Planck scale.
We have also checked that propagation speed of gravitons is almost the same as speed of light,
and we can avoid inconsistency with observations by taking small $\kappa_3$.
Small extra dimensions let
the Kaluza-Klein modes obtain large masses
in both our compactification and the flux compactification.
In the latter case, the zero modes called moduli also obtain large mass,
while in our case they are still massless.
In our calculation, we notice that the coefficient of their kinetic term is much larger
if inequalities~\eqref{ConsistencyCondition} are satisfied.
This fact leads to their strongly supressed coupling to matters via canonicalization of variables.
For this reasons, we expect that small coupling constant $\kappa_3$ suppresses production of
those massless degrees of freedom and astrophysical and cosmological observation
is consistent with prediction of general relativity.
\section{Conclusion}\label{secConclusion}
We have investigated the spatially flat homogeneous model in the Lovelock theory up to the cubic order.
Under isotropic pressure,
subtractions of the evolution equations yield the conservation law \eqref{subeq}.
It implies that as the total volume expands, the system converge on the roots of Eq.~\eqref{fixedpointeq},
which we call (an)isotropic attractors.
We have focused especially on the anisotropic attractor with two different expansion rates.
We have considered the possibility to apply the anisotropic attractor to freeze extra dimensions.
According to the standard cosmology, the universe is dominated by two kinds of energy contents.
One is energy contents having isotropic pressure, e.g.,
potential and kinetic energy of a scalar field, cold matter, and the cosmological constant.
In this isotropic case, we can use the ansiotropic attractor to suppress growth of the extra dimensions
compared to the lower-dimensional universe
if inequalities~\eqref{ConsistencyCondition} are satisfied.
The other has anisotropic pressure, such as radiation of sufficiently low energy scale.
Even though we cannot use the anisotropic attractor, radiation allows the expansion rate of extra dimensions
to become smaller than in the isotropic case.
Therefore, we conclude that in the whole of standard cosmic history,
extra dimensions are frozen by using the Lovelock theory.
We can straightforwardly extend this mechanism to the Lovelock theory
or generalized Galileon up to an arbitrary higher order.
In the generalized Galileon, odd-power terms of expansion rate appears in the Hamiltonian constraint,
whereas the Lovelock theory contains only even-power terms.
The details will be studied in our forthcoming paper.
\begin{acknowledgments}
We thank Taizan Watari and Hideyuki Tagoshi for useful comments.
HWHT was supported by the Grant-in-Aid for JSPS Research Fellow and the Advanced Leading Graduate Course for Photon Science.
TK was partially supported by JSPS KAKENHI Grant Nos.~JP20H04745 and JP20K03936.
JY was partially supported by JSPS KAKENHI Grant Nos.~JP15H02082, JP20H00151, and Grant-in-Aid for Scientific Research on Innovative Areas JP20H05248.
\end{acknowledgments}
| 0b0bd3409e805f57ce41c98bbbabbb3db751bb2c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The theory containing coupling between the scalar fields and gravity was firstly considered by Fisher who found a static and spherically symmetric solution of the Einstein massless scalar field equations \cite{Fisher1948}. Whereafter, variety of theories {emerge} with the introduction of scalar fields, such as dimensional reductions {and} low-energy limit of string theories, various models of supergravity, etc. The low energy limit of the string theory has the Einstein action which is supplemented by fields, such as axion, gauge fields, and dilaton coupled in a nontrivial way to other fields. One of the most important theories of the low-energy limit of string theories is the {dilaton} gravitation theory \cite{green:1987}. It is a very interesting theory because the dilaton field can naturally couple to gauge fields, and the properties of the black hole change drastically due to the dilaton field. The presence of the dilaton field can change the causal structure and asymptotic behavior of the spacetime as well as the thermodynamics and stability of the electrically charged black holes \cite{pro1}. Therefore the dilaton black holes have attracted many attention in recent decades.
On the other hand, anti-de Sitter space({AdS}) and conformal field theory correspondence (AdS/CFT) \cite{ADSCFT1} is a significant method to unify quantum fields and gravitations. This correspondence between the gravity in an AdS spacetime and conformal field theory (CFT) {leaking} on to the boundary of the AdS spacetime suggests that physics of the gauge field defined on the AdS-boundary can be derived from gravitations in the bulk space and vice versa \cite{ADSCFT2}. Stimulated by AdS/CFT, many works are devoted to find the asymptotically (anti)-de Sitter black hole solution in dilaton gravity with various self-interacting potential of dilaton. Exact solutions of charged dilaton black holes have been constructed by many authors. Some black holes are asymptotically flat \cite{gib:1988,gar:1991,flat} where the dilaton changes the causal structure of the black hole and leads to the curvature singularities at finite radii. Meanwhile, some solutions are asymptotically neither flat nor (anti)-de Sitter \cite{nor}. Finally, the asymptotically (anti)-de Sitter versions are found in \cite{gao:2004,gao:2005a,Hajkhalili:2019,Hendi:2010,Hajkhalili:2018,Sheykhi2008,Yamazaki2001,anti}. These solutions reveal that the cosmological constant proposed in the Einstein theory is coupled to Liouville-type dilaton scalar potential.
With the building of these asymptotically (anti)-de Sitter dilaton black holes, many interesting physical phenomena have been explored, such as the black hole thermodynamics \cite{Sheykhi:2016,Sheykhi:2010,Hendi:2010}, holographic thermalization \cite{Zhang2015}, black hole phase transition in {extended} phase space \cite{Li2018} etc.
The purpose of this paper is, starting from the known dilaton black holes in de Sitter or anti-de Sitter universe, to construct new exact asymptotically (anti)-de Sitter black hole solutions in the Einstein-Maxwell-scalar theory which is the generalization of Einstein-Maxwell-dilaton theory. It is known that the Einstein-Maxwell-scalar theory can emerge naturally in physics, for example, in the contexts of Kaluza-Klein models
\cite{Ka1921}, supergravity/string theory \cite{van1981} and cosmology \cite{mar2008}. In the last two years, a remarkable phenomenon of spontaneous scalarisation of charged black holes is discovered by \cite{her2018,fer2019}. It has motivated vast studies on various Einstein-Maxwell-scalar models (see \cite{bla2020} and references therein). This is one of the motivations for this research. On the other hand, according to the no-hair theorem, there are at most three physical parameters (mass, charge and angular momentum) for a black hole. However, by {giving up} the corresponding assumptions, one can violate the no-hair theorem in the Einstein-Maxwell-scalar theory (EMS). For example, dilaton charge $D=-\frac{Q^{2}}{2M}$ emerges in diaton gravitation theory. One finds that the spacetime of dilaton black hole and {its de Sitter counterpart} are all affected by the dilaton charge, but not the electrical charge. Therefore, dilaton charge can be viewed as a new hair. In view of these points and inspired by the expression of dilaton black holes in de Sitter or anti-de Sitter universe, we find two new exact black hole solutions in the Einstein-Maxwell-scalar theory. These solutions are the extensions of Reissner-Nordstrom-(anti) de Sitter and dilaton-(anti) de Sitter solutions. We find that the new black hole spacetime in (anti)-de Sitter universe {is} directly affected by the electrical charge. {This} is different from the dilaton black hole solution.
The paper is organized as follows. In Sec. II, we derive the equations of motion in the Einstein-Maxwell-scalar theory. In Sec. III, we present the first electrically charged dilaton black hole in (anti)-de Sitter spacetime and give the analysis on its horizons. In Sec. IV, we study its thermodynamics which covers the construction of Smarr formula and first law of thermodynamics. Then the second electrically charged dilaton black hole in (anti)-de Sitter spacetime with arbitrary coupling constant $\alpha$ is found in Sec. V and the corresponding analysis of thermodynamics are given in Sec. VI. Finally, we give the conclusion and discussion in Sec. VII. Throughout the paper, we adopt the system of units in which $G=c=\hbar=1$ and the metric signature
$(-,\ +,\ +,\ +)$.
\section{The equations of motion}
The action of Einstein-Maxwell-scalar theory is given by
\begin{equation}
S=\int d^4x\sqrt{-g}\left[R-2\nabla_{\mu}\phi\nabla^{\mu}\phi-K\left(\phi\right)F^2-V\left(\phi\right)\right]\;,
\end{equation}
where $R$ is the Ricci scalar curvature, $F^2\equiv F_{\mu\nu}F^{\mu\nu}$ comes from the Maxwell field, $K\left(\phi\right)$ is the coupling function between scalar field and Maxwell field. $V\left(\phi\right)$ is the scalar potential. When $K=1$ and $V=2\lambda$, it is the Einstein-Maxwell theory with cosmological constant $\lambda$. The theory gives the Reissner-Nordstrom-de Sitter solution. When $K=e^{2\phi}$ and $V=0$, it gives the dilaton black hole solution \cite{gib:1988,gar:1991}. When $K=e^{2\phi}$ and
\begin{equation}\label{eq:pot0}
V=\frac{\lambda}{3}\left(e^{2\phi}+4+e^{-2\phi}\right)\;,
\end{equation}
it gives the dilaton black hole in de Sitter universe \cite{gao:2004}. Except for these solutions, there are other important solutions with different $K$ and $V$ \cite{gao:2005a,Hajkhalili:2019,Hendi:2010,Hajkhalili:2018,Sheykhi2008,Yamazaki2001,anti}. In principle, once the expressions of $K\left(\phi\right)$ and $V\left(\phi\right)$
are given, the gravity theory and the corresponding black hole solution are {determined}. This is the {conventional method}. But in this paper, we shall presume not $K\left(\phi\right)$ and $V\left(\phi\right)$, but the black hole solution in advance. Then $K\left(\phi\right)$ and $V\left(\phi\right)$ are subsequently {derived}. {This is the solution-generating method \cite{cadoni:2011,Stephani:book}.} We require that the parameters such as the mass $M$ and electric charge $Q$ don't appear in the coupling function $K\left(\phi\right)$ and the scalar potential $V(\phi)$.
Now we derive the equations of motion. Varying the action with respect to the metric, Maxwell and dilaton fields, respectively, yields
\begin{eqnarray}
R_{\mu\nu}=2\nabla_{\mu}\phi\nabla_{\nu}\phi+\frac{1}{2}g_{\mu\nu}V+2K\left(F_{\mu\alpha}F^{\alpha}_{\nu}-\frac{1}{4}g_{\mu\nu}F^2\right)\;,
\end{eqnarray}
\begin{eqnarray}
\nabla_{\mu}\left(K F^{\mu\nu}\right)=0\;,
\end{eqnarray}
\begin{eqnarray}
\nabla_{\mu}\nabla^{\mu}\phi-\frac{1}{4}\left(V_{,\phi}+K_{,\phi}F^2\right)=0\;,
\end{eqnarray}
where ``$,\phi$'' denotes the derivative with respect to $\phi$.
The metric for static and spherically symmetric black hole solution can always be written as
\begin{equation}
ds^2=-U\left(r\right)dt^2+{U\left(r\right)}^{-1}dr^2+f\left(r\right)\left(d\theta^2+\sin^2\theta d\phi^2\right)\;.
\end{equation}
In this spacetime, the non-vanishing components of four-vector $A_{\mu}$ is uniquely $A_0(r)$.
So the equations of motion turn out to be
\begin{equation}\label{eom:1}
2ff^{''}+4f^2\phi^{'2}-f^{'2}=0\;,
\end{equation}
\begin{equation}\label{eom:2}
\left(fKA_0^{'}\right)^{'}=0\;,
\end{equation}
\begin{eqnarray}\label{eom:3}
&& 2f^2U^{''}+2fUf^{''}+2fU^{'}f^{'}-Uf^{'2}+4Uf^2\phi^{'2}\nonumber\\&&-4f^2KA_0^{'2}+2f^2V=0\;,
\end{eqnarray}
\begin{equation}\label{eom:4}
fU\phi^{''}+Uf^{'}\phi^{'}+fU^{'}\phi^{'}+\frac{1}{2}fK_{,\phi}A_0^{'2}-\frac{1}{4}fV_{,\phi}=0\;.
\end{equation}
Here prime denotes the derivative with respect to $r$. Observing these equations, we find six unknown functions. They are $U,\ f,\ \phi,\ A_0,\ K,\ V$. But we have only four equations of motion. Therefore, the equations of motion are not closed. In principle, one can presume any two of them in advance. Then the equations of motion are closed. In the next section, we {shall employ the solution-generating techniques \cite{cadoni:2011,Stephani:book} to construct new black holes }.
\section{The first solution}
We recall that when $K=e^{2\phi}$ and $V=0$, the theory gives the dilaton black hole solution
with \cite{gib:1988,gar:1991}
\begin{eqnarray}
U&=&1-\frac{2M}{r}\;,\ \ \ \ f=r\left(r-\frac{Q^2}{M}\right)\;,\nonumber\\
\phi &=&-\frac{1}{2}\ln\left(1-\frac{Q^2}{M{r}}\right)\;.
\end{eqnarray}
Here $M$ and $Q$ have the physical meaning of ADM mass and electric charge of the black hole, respectively. This spacetime is asymptotically flat. On the other hand,
when $K=e^{2\phi}$ and $V$ of Eq.~(\ref{eq:pot0}), the theory gives the dilaton black hole in de Sitter universe with \cite{gao:2004}
\begin{eqnarray}\label{eq:gao}
U&=&1-\frac{2M}{r}-\frac{1}{3}\lambda f\;,\ \ \ \ f=r\left(r-\frac{Q^2}{M}\right)\;,\nonumber\\
\phi &=&-\frac{1}{2}\ln\left(1-\frac{Q^2}{M{r}}\right)\;.
\end{eqnarray}
Here $\lambda$ is a constant. When $Q=0$, it goes back to the Schwarzschild-de Sitter solution. In order to construct a new solution, we observe the expression of $U$ and find that the $\lambda$ term is {not} proportional to $r^2$, but $f$. Inspired by this fact, we presume the new solution
\begin{equation}\label{eq:uf1}
U=1-\frac{2M}{r}+\frac{\beta Q^2}{f}-\frac{1}{3}\lambda f\;,
\end{equation}
\begin{equation}\label{eq:uf2}
f=r\left(r+\gamma\frac{Q^2}{M}\right)\;.
\end{equation}
Here $\beta$ and $\gamma$ are dimensionless constants. The $\beta$
term is a new term and it is similar to the $Q^2/r^2$ term in Reissner-Nordstrom-de Sitter solution.
Then substituting equations (\ref{eq:uf1}) and (\ref{eq:uf2}) into the equations of motion (\ref{eom:1},\ref{eom:2},\ref{eom:3},\ref{eom:4}), we obtain
\begin{equation}
\phi=-\frac{1}{2}\ln\left(1+\frac{\gamma Q^2}{Mr}\right)\;,
\end{equation}
\begin{equation}
A_0=\frac{\gamma Q}{r}-\frac{\beta Q}{2r}-\frac{\beta M Q}{2\left(Mr+\gamma Q^2\right)}\;,
\end{equation}
\begin{equation}
K=\frac{2e^{2\phi}}{-2\gamma+\beta+\beta e^{4\phi}}\;,
\end{equation}
\begin{equation}\label{eq:pot1}
V=\frac{\lambda}{3}\left(e^{2\phi}+4+e^{-2\phi}\right)\;.
\end{equation}
We notice that the scalar potential, Eq.~(\ref{eq:pot1}) is exactly that in Eq.~(\ref{eq:pot0}). Two dimensionless constants $\gamma$ and $\beta$ appear in the coupling function $K$. Up to this point, equations from Eq.~(\ref{eq:uf1}) to Eq.~(\ref{eq:pot1}) constitute the first black hole solution in this paper.
When $\gamma=-1,\ \beta=0$, the solution returns to Eq.~(\ref{eq:gao}).
It seems there are two dimensionless constants $\gamma$ and $\beta$ in the coupling function $K$. We find this not the case. Actually, if we rescale $A_{\mu}\rightarrow A_{\mu}\sqrt{-\gamma}$ and $\beta\rightarrow-\beta\gamma$, the coupling constant $\gamma$ disappears (or is gauged to minus one). Then we are left with unique coupling constant $\beta$. Therefore, in the next, we adopt $\gamma=-1$. In TABLE.~I, we list the three solutions for {comparison}.
When $\beta=0$, the solution returns to Eq.~(\ref{eq:gao}).
To understand the role of coupling constant $\beta$, we plot the evolution of $K$ with respect to $\phi$ for different $\beta$ in Fig.~{\ref{FigK}}. Fig.~{\ref{FigK}} shows that with the increasing of $\beta$, the effect of Maxwell invariant becomes smaller and smaller. When $\beta\rightarrow +\infty$, gravity takes over electromagnetic interaction and the electromagnetic field can be safely neglected. This means the value of $\beta$ determines the strength of interaction between Maxwell field and the scalar field. The solution, as given by Eq.~(15)-(18), is purely electric. We remember that in standard dilaton theory there is an electric-magnetic duality transformation
\begin{equation}
\{P\rightarrow Q,\ \ Q\rightarrow P\} \ \ \textrm{and}\ \ \ \ {K\left(\phi\right)\rightarrow\frac{1}{K\left(\phi\right)}}\;,\nonumber\\
\end{equation}
which turns the electric field into a magnetic field and changes the sign of $\phi$ without modifying the spacetime metric. We find this does not apply to the solution except for $\beta=0$ and $\beta=2\gamma$.
Now let's consider {the} situation of $\lambda=0$ for the moment. We can make a translation transformation $\phi\rightarrow \phi+\phi_0$ with $\phi_0$ a constant on the coupling function $K(\phi)$ such that $K(\phi)\propto 1/\cosh{2\phi}$. Then there is a maximum for $K(\phi)$ at $\phi=0$. At the maximum, we have $dK/d\phi=0$. In this case, $\phi=0$ solves the field equations and the Reissner-Nordstrom (RN) spacetime is a solution. However, the RN solution is not unique to the theory. As we show, Eq.~(13)-(17) constitute the second set of black hole solutions, with the nontrivial scalar field profile given by Eq.~(15). It is the corresponding scalarised counterparts of RN black holes.
The effective mass squared of the scalar field ($\lambda=0$) is
\begin{equation}
\mu^2_{\textrm{eff}}=\frac{F^2}{4}\frac{d^2 K}{d\phi^2}|_{\phi=0}>0\;,
\end{equation}
because we have $F^2<0$ for the pure electric field. Therefore, there is no the tachyonic instability. In other words, the RN solution is stable to the scalar perturbations. According to the classification of Ref.~\cite{asto:2019}, this scalarised solution belongs to the Subclass IIB or scalarised-disconnected-type. In this case, the scalarised black holes do not bifurcate from RN black holes, and do not continuously reduce to the latter when $\phi=0$.
\begin{table}[]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline \hline
&$U(r)$& $f(r)$&$K(r)$&$V(r)$&$\phi(r)$& $A(r)$\\\hline
Dilaton &$1-\frac{2M}{r}$&$r^{2}(1-\frac{Q^{2}}{Mr})$&$e^{2\phi}$&$0$&$-\frac{1}{2}\text{ln}(1-\frac{Q^{2}}{Mr})$& $-\frac{Q}{r}$\\\hline
Dilaton&&&&&&\\
de-Sitter&$1-\frac{2M}{r}-\frac{1}{3}\lambda f$&$r^{2}(1-\frac{Q^{2}}{Mr})$&$e^{2\phi}$&$\frac{\lambda}{3}\left(e^{2\phi}+4+e^{-2\phi}\right)$&$-\frac{1}{2}\text{ln}(1-\frac{Q^{2}}{Mr})$& $-\frac{Q}{r}$\\\hline
First&&&&&&\\
solution&$1-\frac{2M}{r}+\frac{\beta Q^{2}}{f}-\frac{1}{3}\lambda f$&$r^{2}(1-\frac{Q^{2}}{Mr})$&$\frac{2e^{2\phi}}{\beta e^{4\phi}+\beta+2}$&$\frac{\lambda}{3}\left(e^{2\phi}+4+e^{-2\phi}\right)$&$-\frac{1}{2}\text{ln}(1-\frac{Q^{2}}{Mr})$& $-\frac{Q}{r}-\frac{\beta Q}{2}\left(\frac{1}{r}+\frac{r}{f}\right)$\\
\hline \hline
\end{tabular}
\label{table111}
\end{center}
\end{table}
The condition of $dK/d\phi|_{\phi=0}=0$ guarantees the existence of RN solution, but not the general scalarised black hole solutions. In order to guarantee there are scalarised black hole solutions, two Bekenstein-type identities should be satisfied \cite{asto:2019}.
The first identity is given by
\begin{equation}
\int\sqrt{-g}d^4x\left(K_{\phi\phi}\nabla_{\mu}\phi\nabla^{\mu}\phi+\frac{K_{\phi}^2}{4}F^2\right)=0\;.
\end{equation}
For a purely electric field, one has $F^2<0$. This implies
\begin{equation}
K_{,\phi\phi}>0\;,
\end{equation}
should be satisfied in some region of $r$. Otherwise, the two terms of the integrand
will have always the same sign which makes the identity holds if and only if $\phi=0$.
The second identity is given by
\begin{equation}
\int\sqrt{-g}d^4x\left(\nabla_{\mu}\phi\nabla^{\mu}\phi+\frac{{\phi K_{,\phi}}}{4}F^2\right)=0\;.
\end{equation}
This reveals that for a pure electric field, the potential should satisfy the condition
\begin{equation}
\phi K_{,\phi}>0\;,
\end{equation}
in some region of $r$. In Fig.~\ref{fig:bek}, we present the plots of $K_{,\phi\phi}$ and $\phi K_{,\phi}$ with $K=1/\cosh{2\phi}$ and $Q^2/M=0.5$. It is apparent we always have $K_{,\phi\phi}>0$ and $\phi K_{,\phi}>0$ in some range of $r$. Then we conclude that the theory has the scalarised black hole solution with pure electric field. The conclusion is in agreement with the solution given by Eq.~(15-17) for $\lambda=0$.
\begin{figure}[h]\label{FigK}
\begin{center}
\includegraphics[width=9cm]{Fig0K.eps}
\caption{The evolution of coupling function $K(\phi)$. $K(\phi)$ is regular in the entire field space. The plots correspond to $\beta=0,\ \ 0.1,\ \ 0.2,\ \ 0.3,\ \ 0.5,\ \ 0.9$, respectively, from top down.}\label{FigK}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=9cm]{bek.eps}
\caption{The plots of $K_{,\phi\phi}$ and $\phi K_{,\phi}$ ($K=1/cosh{2\phi}$) with respect to $r$, respectively. In some region of $r$, one always have $K_{,\phi\phi}>0$ and $\phi K_{,\phi}>0$. Therefore, the theory has the scalarised black hole solution with pure electric field.}\label{fig:bek}
\end{center}
\end{figure}
In the following, we shall compute the scalar charge $D$, the electric charge $Q$ and the ADM mass $M$ of the black hole.
The scalar charge $D$ can be computed by
\begin{equation}
D=\frac{1}{4\pi}\oint \nabla_{\mu}\phi \ dS^{\mu}=-\frac{Q^2}{2M}\;.
\end{equation}
Here the integration is taken over a spacelike surface enclosing the origin. The scalar charge is a conserved value and thus it does not depend on the choice of the surface. It is clear that the scalar charge is determined by the mass $M$ {and} electric charge $Q$.
The electric charge of the black hole is shown to be
\begin{equation}\label{charge}
Q=\frac{1}{4\pi}\oint K \left({\phi}\right)\nabla_{\mu}A_{0}\ dS^{\mu}\;.
\end{equation}
The electric charge is also a conserved value and it does not depend on the choice of the surface.
The ADM mass \cite{ADM:1,ADM:2,ADM:3} satisfies
\begin{equation}
M=\frac{1}{16\pi}\lim_{S^{\mu}\rightarrow i^{0}}\oint g^{\alpha\beta}\left(g_{\alpha\mu,\beta}-g_{\alpha\beta,\mu}\right)dS^{\mu}\;,
\end{equation}
where $i^{0}$ is the spacetlike infinity. The Ricci scalar $R$ and the Maxwell invariant $F_{\mu\nu}F^{\mu\nu}$ of the spacetime are give by
\begin{eqnarray}
R&=&-\frac{4 f^{''} U f+2U^{''} f^{2}-f^{'2} U+4U^{'}f^{'} f-4 f}{2f^{2}}\\\;
&=&\frac{2}{r^{3}\left(r+2D\right)^{3}}\left(2 \lambda r^{6}+12 D \lambda r^{5}+25 D^{2} \lambda r^{4}+20 D^{3} \lambda r^{3}\right. \;, \nonumber\\&&
+\left.\left(4 D^{4} \lambda+D^{2}\right) r^{2}+\left(2 D^{3}-2 D^{2} M\right) r+D^{2} Q^{2} \beta-4 D^{3} M\right)\;,
\end{eqnarray}
\begin{eqnarray}\label{negative}
F_{\mu\nu}F^{\mu\nu}&=&-2 A0^{'2}=-2\left[\frac{ Q}{r^2}+\frac{\beta Q}{2 r^2}+\frac{\beta Q}{2\left(r+2D\right)^2}\right]^2\;.
\end{eqnarray}
From the definition of $f$ in Eq.~({\ref{eq:uf2}}) we know $r$ should satisfy
\begin{eqnarray}
r\geq-2D>0\;.
\end{eqnarray}
We conclude {that} both the Ricci scalar and the Maxwell invariant are divergent at $r=-2D$. Therefore, the position of $r=-2D$ is the physical singularity of this spacetime. As for the black hole horizons, we find that, under the condition {of} $\beta>0$, there are at most three horizons for positive $\lambda$. They are cosmic horizon, black hole event horizon and black hole inner horizon, respectively, as shown in Fig.~\ref{fig:ur}. For negative $\lambda$, there are at most two horizons. They are black hole event horizon and black hole inner horizon. As a result, in the presence of cosmological constant $\lambda$, the causal structure (Penrose diagram) of the black hole spacetime is exactly the same as the Reissner-Nordstrom-de Sitter (or anti-de Sitter) spacetime.
\begin{figure}[h]
\centering
\subfigure[]{\label{fig:ur}\includegraphics[width=0.47\textwidth]{Fig1f.eps}}
\subfigure[]{\label{fig:urp0}\includegraphics[width=0.47\textwidth]{u-r-p0.eps}}
\subfigure[]{\label{fig:urn}\includegraphics[width=0.47\textwidth]{u-r-n.eps}}
\caption{{Variation of $U$ with respect to} $r$. (a) There are four cases for the scenario of horizons. The curves correspond to 1-horizon, 2-horizon, 3-horizon and 2-horizon from down to up, respectively. We fix $M=1,\ Q=1$ and choose $\beta=0.1,\ \lambda=0.5$; $\beta=0.1,\ \lambda=0.184$; $\beta=0.2,\ \lambda=0.14$ and $\beta=0.277,\ \lambda=0.12$, respectively. (b) There are three cases for the scenario of horizons {with vanishing} $\beta$ term. The curves correspond to 0-horizon, 0-horizon, 1-horizon and 2-horizon from down to up, respectively. We fix $M=1,\ Q=1$ and choose $\lambda=0.5$, $ \lambda=0.3$, $\lambda=0.168$ and $\lambda=0.12$, respectively. (c) There are three cases for the scenario of horizons. The curves correspond to 0-horizon, 1-horizon, 2-horizon and 1-horizon from up to down, respectively. We fix $M=1,\ Q=1,\ \lambda=-0.4$ and choose $\beta=0.0\ $, $\beta=0.1\ $, $\beta=0.2\ $ and $\beta=0.3$, respectively. }
\end{figure}
We find that it is rather involved to study the variation of horizons in the presence of cosmological constant. So here we shall focus on the case of $\lambda=0$. In order to analyse the {scenario} of horizons, we make coordinate transformation as follows
\begin{eqnarray}
r=\frac{Q^2+\sqrt{Q^4+4M^2 x^2}}{2M}\;.
\end{eqnarray}
Then the metric is presented in the Schwarzschild coordinate system
\begin{eqnarray}
ds^2-Udt^2+\frac{4M^2x^2}{U\left(Q^4+4M^2x^2\right)}dx^2+x^2d \Omega^2\;.
\end{eqnarray}
with
\begin{eqnarray}
U=1-\frac{4M^2}{Q^2+\sqrt{Q^4+4M^2x^2}}+\frac{\beta Q^2}{x^2}\;.
\end{eqnarray}
Now the physical singularity is at $x=0$ and the positions of horizons are obtained by solving $U=0$,
\begin{eqnarray}
x_{+}=\sqrt{-Q^2-\beta Q^2+2M^2+\sqrt{Q^4+4M^4-4Q^2M^2-4\beta Q^2M^2}}\;,
\end{eqnarray}
\begin{eqnarray}
x_{-}=\sqrt{-Q^2-\beta Q^2+2M^2-\sqrt{Q^4+4M^4-4Q^2M^2-4\beta Q^2M^2}}\;.
\end{eqnarray}
$x_+$ and $x_-$ represent the event horizon and the inner Cauchy horizon, respectively.
Defining the ratio $\epsilon$ of charge to mass {as} $\epsilon=Q/M$, we have
\begin{eqnarray}
x_{+}=M\sqrt{2-\epsilon^2-\beta\epsilon^2+\sqrt{\left(2-\epsilon^2\right)^2-4\epsilon^2\beta}}\;,
\end{eqnarray}
\begin{eqnarray}
x_{-}=M\sqrt{2-\epsilon^2-\beta\epsilon^2-\sqrt{\left(2-\epsilon^2\right)^2-4\epsilon^2\beta}}\;.
\end{eqnarray}
Without the loss of generality, we assume $Q>0$. Then we have the following conclusions.
(1) When
\begin{eqnarray}
\epsilon\geq \sqrt{2}\;,\ \ \ \beta>0\;,
\end{eqnarray}
there are no horizons and the central singularity is naked. This reveals the maximum charge of the black hole should be smaller than $\sqrt{2}M$ in order that the singularity is dressed with horizon.
(2) When
\begin{eqnarray}
\epsilon=0\;,
\end{eqnarray}
it reduces to the {Schwarzschild} solution.
(3) When \begin{eqnarray}
0<\epsilon<\sqrt{2}\;,\ \ \textrm{and} \ \ \ \beta>\frac{\left(2-\epsilon^2\right)^2}{4\epsilon^2}\;,
\end{eqnarray}
there are no horizons and the central singularity is naked. This reveals that the singularity can be naked
if the coupling constant $\beta$ is sufficiently large no matter how small the charge is.
(4) When \begin{eqnarray}
0<\epsilon<\sqrt{2}\;,\ \ \textrm{and} \ \ \ \beta<\frac{\left(2-\epsilon^2\right)^2}{4\epsilon^2}\;,
\end{eqnarray}
there are two horizons.
(5) When \begin{eqnarray}
0<\epsilon<\sqrt{2}\;,\ \ \textrm{and} \ \ \ \beta=\frac{\left(2-\epsilon^2\right)^2}{4\epsilon^2}\;,
\end{eqnarray}
the two horizons coincide and the so-called extreme black hole is achieved. This tells us the black hole can be extreme
if the coupling constant $\beta$ {meets above condition} no matter how small the charge is.
In the next section, we shall study the thermodynamics of this spacetime.
\section{black hole thermodynamics}
In order to explore the black holes thermodynamics, we start from the calculation of its temperature. The black hole temperature is well defined in quite general setting due to its strict geometrical basis \cite{Gib:1977}. One can employ the so-called surface gravity which is defined as follows:
\begin{equation}
\kappa^2=-\frac{1}{2}\nabla_{\mu}\chi_{\nu}\nabla^{\mu}\chi^{\nu}\;,
\end{equation}
where $\chi_{\mu}$ is a Killing vector field and it is null on the horizon. We can chose $\chi^{\nu}=\partial/\partial t$ because spacetime is the static. As a result, the black hole temperature takes the form
\begin{eqnarray}\label{temperature}
T&=&\frac{{\kappa}}{2\pi}=\frac{1}{4\pi}U^{'}\left(r_{+}\right)\;.
\end{eqnarray}
Substituting the first black hole solution metric Eq.(\ref{eq:uf1}), we get
\begin{eqnarray}\label{eq:t1}
T&=&\frac{M}{2\pi r_+^2}-\frac{\beta Q^2}{4\pi r_+^2\left(r_++2D\right)}-\frac{\beta Q^2}{4\pi r_+\left(r_++2D\right)^2}\nonumber\\&&-\frac{\lambda r_{+}}{6\pi}+\frac{\lambda Q^2}{12\pi M}\;,
\end{eqnarray}
where $r_+$ represents the black hole event horizon which is determined by $U\left(r_+\right)=0$. In Fig.~\ref{Fig2f} and Fig.~\ref{Fig4f} we plot the black hole temperature $T$ as the function of black hole event horizon $r_+$ with running $\beta$ and $Q$, respectively. In Fig.~{4a}, we put $M=1,\ Q=1$ and $\beta=0.2,\ 0.24,\ 0.28,\ 0.32,\ 0.335$. Given $M,\ Q,\ \beta$, then $\lambda$ can be expressed as the function of $\lambda(M,Q,\beta,r_{+})$ by using the event horizon equation. Eventually, the temperature $T$ becomes the parametric equations of $r_{+}$. Fig.~\ref{Fig2f} shows there are two phases of black holes with the same temperature, the so-called small and large black holes, respectively. When $\beta=0.335$, there is uniquely one phase and its temperature is zero. With the increasing of $\beta$, the black holes make phases transition from 2-phase to 1-phase. In Fig. 4(b), we put $\lambda=-1.5,\ \beta=0.14$ and $Q=0.069,\ 0.063,\ 0.057,\ 0.051,\ 0.044$. Given $\lambda,\ Q,\ \beta$, then $M$ can be expressed as the function of $M(\lambda,Q,\beta,r_{+})$ by using the event horizon equation. Eventually, temperature $T$ becomes the parametric equations of $r_{+}$. Fig.~\ref{Fig4f} shows that there are three phases of black holes with the same temperature. They are large, middle and small black holes, respectively. With the decreasing of $Q$, the black holes make phases transition from 1-phase to 3-phase.
\begin{figure}[htbp]
\centering
\subfigure[]{ \label{Fig2f}
\includegraphics[width=8cm,height=6cm]{Fig2f.eps}}\subfigure[]{\label{Fig4f}
\includegraphics[width=8cm,height=6cm]{Fig4f.eps}}\\
\caption{Black hole temperature $T$ as the function of black hole event horizon $r_+$. (a) The curves correspond to $\beta=0.2,\ 0.24,\ 0.28,\ 0.32,\ 0.335$, from up to down, respectively. We put $M=1,\ Q=1$. There are two phases of black holes with the same temperature, the so-called small (solid lines) and large black holes (dotted lines), respectively. When $\beta=0.335$, there is uniquely one phase and its temperature is zero. (b) The curves correspond to $Q=0.076,\ 0.069,\ 0.063,\ 0.057,\ 0.051,\ 0.044$, from down to up, respectively. We put $\lambda=-1.5,\ \beta=0.14$. With the decreasing of $Q$, the black holes make phases transition from 1-phase to 3-phase which consists of small (solid line), middle (dash-dot line) and large (dotted line) black holes.}
\end{figure}
The entropy of black holes generally satisfies the area law which states that the entropy is a quarter of the area of black hole event horizon \cite{beck:1973}.
The law applies to almost all kinds of black holes including Einstein-Maxwell-scalar black holes \cite{haw:1999}. Therefore we have the entropy of the black hole
\begin{equation}
S=\frac{A}{4}=\pi r_{+}\left(r_{+}+2D\right)\;.
\end{equation}
According to definition of \cite{cve:1999}, the electric potential measured by the distant observer is
\begin{equation}\label{eq:p1}
\Phi=\int_{r_{+}}^{+\infty}A_0^{'}dr=-A_0\left(r_{+}\right)\;.
\end{equation}
The study on the thermodynamic phase structure of AdS black holes \cite{kas:2009} tells us the cosmological constant acts as a thermodynamic pressure
\begin{equation}\label{eq:press}
P=-\frac{\lambda}{8\pi}\;.
\end{equation}
The variation of thermal pressure requires the presence of a conjugate thermal volume in the first law of thermodynamics, which can lead to a variety
of novel thermodynamic behaviour, for example, triple points \cite{alt:2014}, reentrant phase transitions
\cite{alt:2013}, the emergence of polymer-like phase structure \cite{fra:2014}, the superfluid-like phase structure \cite{hen:2017} and
the Van der Waals transition \cite{kub:2012}.
We find the conjugate thermodynamic volume is
\begin{equation}\label{eq:v1}
\mathfrak{V}\equiv\left(\frac{\partial M}{\partial P}\right)_{S,Q}=\frac{4\pi}{3}r_{+}\left(r_{+}+2D\right)\left(r_{+}+D\right)\;.
\end{equation}
Then the Smarr formula
\begin{equation}\label{Smarr}
M=2TS-2\mathfrak{V}P+Q\Phi\;,
\end{equation}
is satisfied.
In the following, we shall make an examination whether the thermal quantities satisfy the requirement of the first law of thermodynamics. From the equation of horizon, we obtain
\begin{equation}\label{eq:lamb}
\lambda=\frac{3}{f_+}\left(1-\frac{2M}{r_+}+\frac{\beta Q^2}{f_+}\right)\;,
\end{equation}
where $f_+=f\left(r=r_+\right)$. So in view of Eq.~(\ref{eq:press}), the pressure $P$ can be written as
\begin{equation}\label{eq:prs}
P=-\frac{3}{8\pi f_+}\left(1-\frac{2M}{r_+}+\frac{\beta Q^2}{f_+}\right)\;.
\end{equation}
We treat the pressure $P$, the entropy $S$ as the function of $r_+,\ M,\ Q$. Then we have
\begin{equation}
dP=P_{,M}dM+P_{,Q}dQ+P_{,r_{+}}dr_+\;,
\end{equation}
\begin{equation}
dS=S_{,M}dM+S_{,Q}dQ+S_{,r_{+}}dr_+\;.
\end{equation}
By using the equations Eq.~(\ref{eq:t1}), Eq.~(\ref{eq:p1}) and Eq.~(\ref{eq:v1}), we find the first law of thermodynamics
\begin{equation}\label{first law}
dM=TdS+\Phi dQ+\mathfrak{V}dP\;,
\end{equation}
is indeed satisfied.
It is pointed \cite{kas:2009,gun:2012} that once the thermodynamic pressure $P$ and thermal volume $\mathfrak{V}$ are introduced, the ADM mass $M$ should be understood as enthalpy. Then the more convenient function to analyze thermodynamic behaviour of a system, especially in the case that there is some critical behaviour, is the Gibbs free energy which is defined in the following way
\begin{equation}
G=M-TS\;.
\end{equation}
The Gibbs free energy $G$ is understood to depend on $M,\ Q,\ \beta,\ r_+$.
In Fig.~{\ref{Fig3f}} and Fig.~{\ref{Fig5f}}, we plot the $G(T)$ relations with running $\beta$ and $Q$, respectively. In Fig.~{\ref{Fig3f}}, the curves correspond to $\beta=0.2,\ 0.24,\ 0.28,\ 0.32,\ 0.335$, from outer to inner, respectively. We put $M=1,\ Q=1$. Given $M,\ Q,\ \beta$, then $\lambda$ can be expressed as the function of $\lambda(M,Q,\beta,r_{+})$ by using the event horizon equation. Eventually, the Gibbs free energy $G$ becomes the parametric equations of $r_{+}$. It shows with the increasing of $\beta$, the black holes make phases transition from 2-phase to 1-phase. It also tells us the Gibbs free energy of large black hole is decreasing with the increasing of Hawking temperature. On the other hand, the Gibbs free energy of small black hole first decreases and then increases with the increasing of Hawking temperature. As is known, the specific heat is $C_{Q,P}=-\partial^2G/\partial T^2$. Therefore, the thermodynamically stable and unstable phases have the concave downward and upward $G(T)$ curves, respectively. Then we conclude that the large black holes are thermodynamically stable while the small black holes are unstable.
In Fig.~\ref{Fig5f}, the curves correspond to $Q=0.069,\ 0.063,\ 0.057,\ 0.051,\ 0.044$, from up to down, respectively. We put $\lambda=-1.5,\ \beta=0.14$. Given $\lambda,\ Q,\ \beta$, then $M$ can be expressed as the function of $M(\lambda,Q,\beta,r_{+})$ by using the event horizon equation. Eventually, the Gibbs free energy $G$ {becomes} the parametric equations of $r_{+}$. It shows that with the decreasing of $Q$, the black holes make phases transition from 1-phase to 3-phase. In this case, we let $Q$ be running. Then the Gibbs free energy of large, middle and small black holes all decreases with the increasing of Hawking temperature. But only large
and small black holes are thermodynamically stable while the middle black holes are unstable.
\begin{figure}[htbp]
\centering
\subfigure[]{ \label{Fig3f}
\includegraphics[width=8cm,height=6cm]{Fig3f.eps}}\subfigure[]{\label{Fig5f}
\includegraphics[width=8cm,height=6cm]{Fig5f.eps}}\\
\caption{Black hole Gibbs free energy $G$ as the function of temperature $T$. (a) The curves correspond to $\beta=0.2,\ 0.24,\ 0.28,\ 0.32,\ 0.335$, from outer to inner, respectively. We put $M=1,\ Q=1$. There are two phases of black holes (dotted lines for small black hole and solid lines for large black holes) with the same temperature. When $\beta=0.335$, the Gibbs free energy is $G=M=1$. (b) The curves correspond to $Q=0.069,\ 0.063,\ 0.057,\ 0.051,\ 0.044$, from up to down, respectively. We put $\lambda=-1.5,\ \beta=0.14$. With the decreasing of $Q$, the black holes make phases transition from 1-phase to 3-phase. }
\end{figure}
\section{second solution}
In this section, we construct the second solution in Einstein-Maxwell-scalar theory. We remember that the dilaton black hole in de Sitter universe for arbitrary coupling constant $\alpha$ is \cite{gao:2004}
\begin{eqnarray}\label{eq:gao2}
U=\left(1-\frac{b}{r}\right)\left(1-\frac{a}{r}\right)^{\frac{1-\alpha^2}{1+\alpha^2}}-\frac{1}{3}\lambda f\;,
\end{eqnarray}
with
\begin{equation}
f=r^2\left(1-\frac{ a}{r}\right)^{\frac{2\alpha^2}{1+\alpha^2}}\;.
\end{equation}
Here $a,\ b$ are two constants which are determined by the black hole mass $M$, charge $Q$ and coupling constant $\alpha$. The corresponding coupling function $K$ and scalar potential $V$ in the action are
\begin{equation}
K=e^{2\alpha\phi}\;,
\end{equation}
and
\begin{eqnarray}\label{eq:2a}
V&=&\frac{2\lambda}{3\left(1+\alpha^2\right)^2}\left[\alpha^2\left(3\alpha^2-1\right)e^{2\phi/\alpha}+\left(3-\alpha^2\right)e^{-2\alpha\phi}\right.\nonumber\\&&\left.+8\alpha^2 e^{-\phi\alpha+\phi/\alpha}\right]\;.
\end{eqnarray}
Observing the $\lambda$ term in Eq.~(\ref{eq:gao2}), we find it is proportional to $f$. So we presume
\begin{equation}\label{eq:u2a}
U=\left(1-\frac{b}{r}\right)\left(1-\frac{a}{r}\right)^{\frac{1-\alpha^2}{1+\alpha^2}}+\frac{\beta Q^2}{f}-\frac{1}{3}\lambda f\;,
\end{equation}
\begin{equation}\label{eq:u2b}
f=r^2\left(1-\frac{a}{r}\right)^{\frac{2\alpha^2}{1+\alpha^2}}\;.
\end{equation}
We notice that when the coupling constant $\alpha=1$, the solution restores to the first solution in section III.
Substituting Eqs.~(\ref{eq:u2a}), (\ref{eq:u2b}) into the equations of motion (\ref{eom:1},\ref{eom:2},\ref{eom:3},\ref{eom:4}), we obtain
\begin{equation}\label{eq:2b}
\phi=-\frac{\alpha}{1+\alpha^2}\ln\left(1-\frac{a}{r}\right)\;,\ \ \ \ \ Q^2=\frac{ab}{\alpha^2+1}\;,
\end{equation}
\begin{equation}\label{eq:2c}
A_0=-\frac{Q}{r}-\frac{1}{1+\alpha^2}\frac{\beta Q}{r}-\frac{\alpha^2}{1+\alpha^2}\frac{\beta Q}{r-a}\;,
\end{equation}
\begin{equation}\label{eq:2d}
K=\frac{e^{2\alpha \phi}(\alpha^2+1)}{\alpha^2+\beta+1+\alpha^2 \beta e^{\frac{2\phi (\alpha^2+1)}{\alpha}}},
\end{equation}
with the scalar potential $V$ of the same expression in Eq.~(\ref{eq:2a}). When $\alpha=1$, they reduce to the counterparts in the first solution of section III.
So equations from (\ref{eq:2a}) to Eq.~(\ref{eq:2d}) constitute the second black hole solution in this study.
The scalar charge $D$ can be computed by
\begin{equation}
D=-\frac{\alpha a}{\alpha^2 +1}\;.
\end{equation}
Here the integration is taken over the surface of spacetlike infinity, $i^{0}$. It tells us the scalar charge is determined by $m$, electric charge $Q$ {and} the coupling constant $\alpha$.
The electric charge of the black hole is the same as Eq.(\ref{charge}).
Here the integration is taken over a spacelike surface enclosing the origin. The electric charge is a conserved value and it does not depend on
the choice of the surface.
We find the ADM mass $M$ of the spacetime is
\begin{eqnarray}
M=\frac{1}{2}\left(b+\frac {- {\alpha}^{2}+1}{{\alpha}^{2}+1} a\right).
\end{eqnarray}
The curvature singularity of this spacetime is at $r=a$. Similar to the first solution, there are at most three horizons in this spacetime and the corresponding causal structure is exactly the same as the Reissner-Nordstrom-de Sitter (or anti-de Sitter) spacetime. Therefore, we need not plot the Penrose diagram anymore. In the next section, we study the black hole thermodynamics.
\section{thermodynamics}
Following the procedure in section IV, we figure out the thermodynamic quantities for the second solution. We find the temperature, the electric potential and the thermal pressure are the same as Eq.(\ref{temperature}), Eq.(\ref{eq:p1}) and Eq.(\ref{eq:prs}), respectively.
The entropy and the thermal volume are
\begin{equation}
S=\pi r_+^2\left(1-\frac{a}{r_+}\right)^{\frac{2\alpha^2}{1+\alpha^2}}\;,
\end{equation}
\begin{equation}\label{eq:v2}
\mathfrak{V}=\frac{4}{3}\,{\frac {\pi\, \left( -{\alpha}^{2}r_{+}+a-r_{+} \right) }{ \left( {\alpha
}^{2}+1 \right) r_{+} \left( -r_{+}+a \right) }}
f^2\;.
\end{equation}
Then the Smarr formula Eq.(\ref{Smarr}) can be obtained.
In the following, we check whether these thermodynamic quantities satisfy the first law of thermodynamics.
From the equation of horizon, we obtain
\begin{equation}\label{eq:lamb}
\lambda=\frac{3}{f_+}\left[\left(1-\frac{b}{r_+}\right)\left(1-\frac{a}{r_+}\right)^{\frac{1-\alpha^2}{1+\alpha^2}}+\frac{\beta Q^2}{f_+}\right]\;,
\end{equation}
where $f_+=f\left(r=r_+\right)$. So the pressure $P$ can be written as
\begin{equation}
P=-\frac{3}{8\pi f_+}\left[\left(1-\frac{b}{r_+}\right)\left(1+\frac{a}{r_+}\right)^{\frac{1-\alpha^2}{1+\alpha^2}}+\frac{\beta Q^2}{f_+}\right]\;.
\end{equation}
We take the mass $M$, the pressure $P$, the entropy $S$ and the charge $Q$ as the function of $r_+,\ a$ and $b$. Then we derive the total differential of $P$, $S$, $Q$, $M$. By using the equations Eq.~(\ref{eq:t1}), Eq.~(\ref{eq:p1}) and Eq.~(\ref{eq:v2}), we arrive at the first law of thermodynamics Eq.(\ref{first law}).
\begin{figure}[htbp]
\begin{center} \label{PVdiagram}
\includegraphics[width=9cm]{PVdiagram.eps}
\caption{$P$ versus $V$ for $\alpha=0.1$, $Q=0.4$, $T=0.08$, and $\beta=0$ (red line), $\beta=1$ (green line), $\beta=2$ (blue line). }
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}\label{GTdiagram}
\includegraphics[width=9cm]{GTdiagram.eps}
\caption{$G$ versus $T$ for $\alpha=0.1$, $Q=0.1$, $\lambda=-2.5$, and $\beta=0$ (red line), $\beta=1$ (green line), $\beta=3$ (blue line). }
\end{center}
\end{figure}
To investigate the effect of $\beta$ on the black hole thermodynamical properties, we plot the $P-V$ diagram and $G-T$ diagram for fixed temperature and charge in Fig.~6 and Fig.~7. Fig.~6 shows that the isothermal curves are raised with the increasing of $\beta$. This means the temperature for Hawking-Page transition (critical temperature) decreases with the increasing of $\beta$. On the other hand, Fig.~7 shows that the temperature for Hawking-Page transition decreases with the increasing of $\beta$. Therefore, they are consistent with each other.
It has been shown by Sheykhi et al \cite{Sheykhi:2010} that for fixed values of the parameters in the dilaton-de Sitter black hole spacetime, there exists
a maximum value of the coupling constant $\alpha_{max}$ such that the black hole becomes thermally unstable. Compared to the dilaton-de Sitter solution, an extra coupling constant $\beta$ is in the presence in the second solution. Thus one may wonder if such a maximum value of the coupling constant $\alpha_{max}$ still remains. {It is the problem of local thermal stability for the system. For this point, following the method of Sheykhi et al, we investigate the variation of $\frac{\partial ^{2}M}{\partial S^{2}}$ with respect to $\alpha$ for different $\beta$. Fig.~(8) shows that for fixed $\beta$, the maximum value of the $\alpha_{max}$ remains. If $\alpha>\alpha_{max}$, $\frac{\partial ^{2}M}{\partial S^{2}}$ would become negative which means the system is locally and thermodynamically instable. But $\alpha_{max}$ increases with the increasing of $\beta$.
\begin{figure}[htbp]
\begin{center}\label{capacity}
\includegraphics[width=9cm]{capacity.eps}
\caption{$\frac{\partial ^{2}M}{\partial S^{2}}$ versus $\alpha$ for $a=0.2$, $\lambda=-3$, $r_{+}=0.4$ and $\beta=0$ (black line), $\beta=0.4$ (purple line), $\beta=0.8$ (red line). }
\end{center}
\end{figure}
\section{{Conclusion and discussion}}
In conclusion, we construct two exact and electrically charged black holes in the Einstein-Maxwell-scalar theory. The two solutions are all asymptotically de Sitter or anti-de Sitter. They are the extensions of dilaton black holes in de Sitter or anti-de Sitter universe. The first solution is for the case of $\alpha=1$ and the second solution is for arbitrary $\alpha$. In this sense, the second solution is the extension of the first one. An interesting fact is that the corresponding scalar potential $V(\phi)$ is the same as the dilaton counterpart. The only difference from the Einstein-Maxwell-dilaton theory is the presence of coupling function $K(\phi)$ in front of the Maxwell invariant. With the presence of $K(\phi)$, the well-known Reissner-Nordstrom-(anti) de Sitter solution is included in these solutions.
We calculate the ADM mass $M$, the electric charge $Q$ and the scalar charge $D$ of the black holes. The scalar charge $D$ is not independent, but dependent on $M$ and $Q$. This is the same as the dilaton black holes. Then we compute the Hawking temperature, the entropy, the electric potential and the thermal volume. Afterwards, both the Smarr formulae and the first law of thermodynamics are constructed. The evolution of Hawking temperature and Gibbs free energy shows that these black holes have two situations of phase structure, the scenario of two phases and three phases. In the scenario of two phases, we let the coupling constants $\beta$ and $\alpha$ be running. In this case, the Gibbs free energy of large black hole is decreasing with the increasing of Hawking temperature. On the other hand, the Gibbs free energy of small black holes first decreases and then increases with the increasing of Hawking temperature. This shows the large black hole is thermodynamically stable and the small black hole is unstable. In the scenario of three phases, we let $Q$ be running. In this case, the Gibbs free energy of large, middle and small black holes all decreases with the increasing of Hawking temperature. But only the large and small black holes are thermodynamically stable while the middle black holes are unstable. Sheykhi et al \cite{Sheykhi:2010} have shown that, for fixed values of the parameters in the dilaton-de Sitter black hole spacetime, there exists
a maximum value of the coupling constant $\alpha_{max}$ such that the black hole becomes thermally unstable. Compared to the dilaton-de Sitter solution, an extra coupling constant $\beta$ is in presence in our new solution. We find that such a maximum value of the coupling constant $\alpha_{max}$ still remains in the new solution. But $\alpha_{max}$ increases with the increasing of $\beta$.
{Finally, since these black hole solutions are asymptotically anti-de Sitter, it is worth studying the corresponding ADS/CFT correspondence. On the other hand, the extension of the static black holes to the rotating and higher dimensional cases {can be a topic} worthy of study.}
\section*{ACKNOWLEDGMENTS}
We thank the anonymous referees for constructive comments. This work is partially supported by the Strategic Priority Research Program ``Multi-wavelength Gravitational Wave Universe'' of the CAS, Grant No. XDB23040100 and the NSFC under grants 11633004, 11773031.
\newcommand\ARNPS[3]{~Ann. Rev. Nucl. Part. Sci.{\bf ~#1}, #2~ (#3)}
\newcommand\AL[3]{~Astron. Lett.{\bf ~#1}, #2~ (#3)}
\newcommand\AP[3]{~Astropart. Phys.{\bf ~#1}, #2~ (#3)}
\newcommand\AJ[3]{~Astron. J.{\bf ~#1}, #2~(#3)}
\newcommand\GC[3]{~Grav. Cosmol.{\bf ~#1}, #2~(#3)}
\newcommand\APJ[3]{~Astrophys. J.{\bf ~#1}, #2~ (#3)}
\newcommand\APJL[3]{~Astrophys. J. Lett. {\bf ~#1}, L#2~(#3)}
\newcommand\APJS[3]{~Astrophys. J. Suppl. Ser.{\bf ~#1}, #2~(#3)}
\newcommand\JHEP[3]{~JHEP.{\bf ~#1}, #2~(#3)}
\newcommand\JMP[3]{~J. Math. Phys. {\bf ~#1}, #2~(#3)}
\newcommand\JCAP[3]{~JCAP {\bf ~#1}, #2~ (#3)}
\newcommand\LRR[3]{~Living Rev. Relativity. {\bf ~#1}, #2~ (#3)}
\newcommand\MNRAS[3]{~Mon. Not. R. Astron. Soc.{\bf ~#1}, #2~(#3)}
\newcommand\MNRASL[3]{~Mon. Not. R. Astron. Soc.{\bf ~#1}, L#2~(#3)}
\newcommand\NPB[3]{~Nucl. Phys. B{\bf ~#1}, #2~(#3)}
\newcommand\CMP[3]{~Comm. Math. Phys.{\bf ~#1}, #2~(#3)}
\newcommand\CQG[3]{~Class. Quantum Grav.{\bf ~#1}, #2~(#3)}
\newcommand\PLB[3]{~Phys. Lett. B{\bf ~#1}, #2~(#3)}
\newcommand\PRL[3]{~Phys. Rev. Lett.{\bf ~#1}, #2~(#3)}
\newcommand\PR[3]{~Phys. Rep.{\bf ~#1}, #2~(#3)}
\newcommand\PRd[3]{~Phys. Rev.{\bf ~#1}, #2~(#3)}
\newcommand\PRD[3]{~Phys. Rev. D{\bf ~#1}, #2~(#3)}
\newcommand\RMP[3]{~Rev. Mod. Phys.{\bf ~#1}, #2~(#3)}
\newcommand\SJNP[3]{~Sov. J. Nucl. Phys.{\bf ~#1}, #2~(#3)}
\newcommand\ZPC[3]{~Z. Phys. C{\bf ~#1}, #2~(#3)}
\newcommand\IJGMP[3]{~Int. J. Geom. Meth. Mod. Phys.{\bf ~#1}, #2~(#3)}
\newcommand\IJMPD[3]{~Int. J. Mod. Phys. D{\bf ~#1}, #2~(#3)}
\newcommand\IJMPA[3]{~Int. J. Mod. Phys. A{\bf ~#1}, #2~(#3)}
\newcommand\GRG[3]{~Gen. Rel. Grav.{\bf ~#1}, #2~(#3)}
\newcommand\EPJC[3]{~Eur. Phys. J. C{\bf ~#1}, #2~(#3)}
\newcommand\PRSLA[3]{~Proc. Roy. Soc. Lond. A {\bf ~#1}, #2~(#3)}
\newcommand\AHEP[3]{~Adv. High Energy Phys.{\bf ~#1}, #2~(#3)}
\newcommand\Pramana[3]{~Pramana.{\bf ~#1}, #2~(#3)}
\newcommand\PTP[3]{~Prog. Theor. Phys{\bf ~#1}, #2~(#3)}
\newcommand\APPS[3]{~Acta Phys. Polon. Supp.{\bf ~#1}, #2~(#3)}
\newcommand\ANP[3]{~Annals Phys.{\bf ~#1}, #2~(#3)}
\newcommand\RPP[3]{~Rept. Prog. Phys. {\bf ~#1}, #2~(#3)}
| be44afe81273770b2b042382395c2c9b166a3967 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{secintro}
A central goal of modern cosmology is to discover whether the dark
energy that appears to fill the Universe is associated with its
matter-energy content, laws of gravity, or some alternative physics.
A compelling means of distinguishing between these scenarios is to
analyse the different observational signatures that are present in the
clumpy, inhomogeneous Universe, which powerfully complements
measurements of the expansion history of the smooth, homogeneous
Universe
\citep[e.g.][]{Linder05,Wang08,Guzzo08,Weinberg13,Huterer15}.
Two important observational probes of the inhomogeneous Universe are
the peculiar velocities induced in galaxies by the gravitational
collapse of large-scale structure, which are statistically imprinted
in galaxy redshift surveys as redshift-space distortions
\citep[e.g.][]{Hamilton98,Scoccimarro04,Song09}, and the gravitational
lensing of light by the cosmic web, which may be measured using cosmic
shear surveys \citep[e.g.][]{Bartelmann01,Kilbinger15,Mandelbaum18}.
These probes are complementary because they allow for the
differentiation between the two space-time metric potentials which
govern the motion of non-relativistic particles, such as galaxy
tracers, and the gravitational deflection of light. The difference or
`gravitational slip' between these potentials is predicted to be zero
in general relativity, but it may be significant in modified gravity
scenarios
\citep[e.g.][]{Uzan01,Zhang07,Jain10,Bertschinger11,Clifton12}.
Recent advances in weak gravitational lensing datasets, including the
Kilo-Degree Survey \citep[KiDS,][]{Hildebrandt20}, the Dark Energy
Survey \citep[DES,][]{Abbott18} and the Subaru Hyper Suprime-Cam
Survey \cite[HSC,][]{Hikage19}, have led to dramatic improvements in
the quality of these observational tests. Gravitational lensing now
permits the accurate determination of, and combinations of, important
cosmological parameters such as the matter density of the Universe and
normalisation of the matter power spectrum, and thereby detailed
comparisons with other cosmological probes such as galaxy clustering
\citep{Alam17b} and the cosmic microwave background radiation
\citep{Planck18}. Some of these comparisons have yielded intriguing
evidence of `tension' on both small and large scales
\citep[e.g.][]{Joudaki17,Leauthaud17,Lange19,Hildebrandt20,Asgari20a},
which are currently unresolved.
In this paper we perform a new study regarding this question using the
latest weak gravitational lensing dataset from the Kilo-Degree Survey,
KiDS-1000 \citep{Kuijken19}, in conjunction with overlapping galaxy
spectroscopic redshift survey data from the Baryon Oscillation
Spectroscopic Survey \citep[BOSS,][]{Reid16} and the 2-degree Field
Lensing Survey \citep[2dFLenS,][]{Blake16b}. In particular, we focus
on a simple implementation of the lensing-clustering test which
compares the amplitude of gravitational lensing around foreground
galaxies (commonly known as galaxy-galaxy lensing), tracing
low-redshift overdensities, with the amplitude of galaxy velocities
induced by these overdensities and measured by redshift-space
distortions, which constitutes an amplitude-ratio test. This
diagnostic was first proposed by \citet{Zhang07} as the
$E_{\mathrm{G}}$ statistic and implemented in its current form by
\citet{Reyes10} using data from the Sloan Digital Sky Survey. These
measurements have subsequently been refined by a series of studies
\citep{Blake16a,Pullen16,Alam17a,delaTorre17,Amon18,Singh19,Jullo19}
which have used new datasets to increase the accuracy of the
amplitude-ratio determination, albeit showing some evidence of
internal disagreement.
The availability of the KiDS-1000 dataset and associated calibration
samples has allowed us to perform the most accurate existing
amplitude-ratio test, on projected scales up to $100 \, h^{-1}$ Mpc,
including rigorous systematic-error control. As part of this
analysis, we use these datasets and representative simulations to
study the efficacy of different corrections for the effects of source
photometric redshift errors, comparing different galaxy-galaxy lensing
estimators and the relative performance of angular and projected
statistics. Our analysis sets the stage for future per-cent level
implementations of these tests using new datasets from the Dark Energy
Spectroscopic Instrument \citep[DESI,][]{DESI16}, the 4-metre
Multi-Object Spectrograph Telescope \citep[4MOST,][]{deJong19}, the
Rubin Observatory Legacy Survey of Space and Time
\citep[LSST,][]{Ivezic19}, and the {\it Euclid} satellite
\citep{Laureijs11}.
This paper is structured as follows: in Sect.\ \ref{sectheory} we
review the theoretical correlations between weak lensing and
overdensity observables, on which galaxy-galaxy lensing studies are
based. In Sect.\ \ref{secest} we summarise the angular and projected
galaxy-galaxy lensing estimators derived from these correlations, with
particular attention to the effect of source photometric redshift
errors. In Sect.\ \ref{secamp} we introduce the amplitude-ratio test
between galaxy-galaxy lensing and clustering observables, constructed
from annular differential surface density statistics, and in
Sect.\ \ref{seccov} we derive the analytical covariances of these
estimators in the Gaussian approximation, including the effects of the
survey window function. We introduce the KiDS-1000 weak lensing and
overlapping Luminous Red Galaxy (LRG) spectroscopic datasets in
Sect.\ \ref{secdata}. We create representative survey mock catalogues
in Sect.\ \ref{secmocks}, which we use to verify our cosmological
analysis in Sect.\ \ref{secmocktests}. Finally, we describe the
results of our cosmological tests applied to the KiDS-LRG datasets in
Sect.\ \ref{secdatatests}. We summarise our investigation in
Sect.\ \ref{secsummary}.
\section{Theory}
\label{sectheory}
In this section we briefly review the theoretical expressions for the
auto- and cross-correlations between weak gravitational lensing and
galaxy overdensity observables, which form the basis of galaxy-galaxy
lensing studies.
\subsection{Lensing convergence and tangential shear}
The observable effects of weak gravitational lensing, on a source
located at co-moving co-ordinate $\chi_{\mathrm{s}}$ in sky direction
$\ho$, can be expressed in terms of the lensing convergence $\kappa$
\citep[for reviews, see][]{Bartelmann01,Kilbinger15,Mandelbaum18}.
The convergence is a weighted integral over co-moving distance $\chi$
of the matter overdensity $\delta_{\mathrm{m}}$ along the
line-of-sight, which we can write as
\begin{equation}
\kappa(\chi_{\mathrm{s}},\ho) = \frac{3 \Omega_{\mathrm{m}} H_0^2}{2
c^2} \int_0^{\chi_{\mathrm{s}}} d\chi \, \frac{\chi \,
(\chi_{\mathrm{s}} - \chi)}{\chi_{\mathrm{s}}} \,
\frac{\delta_{\mathrm{m}}(\chi,\ho)}{a(\chi)} ,
\label{eqkappa1}
\end{equation}
assuming (throughout this paper) a spatially-flat Universe, where
$\Omega_{\mathrm{m}}$ is the matter density as a fraction of the
critical density, $H_0$ is the Hubble parameter, $c$ is the speed of
light, and $a = 1/(1+z)$ is the cosmic scale factor at redshift $z$.
We can conveniently write Eq.\ \ref{eqkappa1} in terms of the critical
surface mass density at a lens plane at co-moving distance
$\chi_{\mathrm{l}}$,
\begin{equation}
\Sigma_{\mathrm{c}}(\chi_{\mathrm{l}},\chi_{\mathrm{s}}) =
\frac{c^2}{4 \mathrm{\pi} G}
\frac{\chi_{\mathrm{s}}}{(1+z_{\mathrm{l}}) \, \chi_{\mathrm{l}} \,
(\chi_{\mathrm{s}} - \chi_{\mathrm{l}})} ,
\end{equation}
where $G$ is the gravitational constant, and $\chi_{\mathrm{s}} >
\chi_{\mathrm{l}}$. Hence,
\begin{equation}
\kappa(\chi_{\mathrm{s}},\ho) = \overline{\rho_{\mathrm{m}}}
\int_0^{\chi_{\mathrm{s}}} d\chi \,
\Sigma_{\mathrm{c}}^{-1}(\chi,\chi_{\mathrm{s}}) \,
\delta_{\mathrm{m}}(\chi,\ho) ,
\label{eqkappa2}
\end{equation}
where $\overline{\rho_{\mathrm{m}}}$ is the mean matter
density.\footnote{We refer the reader to \citet{Dvornik18} Appendix C
for a full discussion of the different definitions of
$\Sigma_{\mathrm{c}}$ that have been adopted in the literature.}
Suppose that the overdensity is associated with an isolated lens
galaxy at distance $\chi_{\mathrm{l}}$ in an otherwise homogeneous
Universe. In this case, Eq.\ \ref{eqkappa2} may be written in the
form
\begin{equation}
\kappa(\chi_{\mathrm{l}},\chi_{\mathrm{s}},\ho) \approx
\overline{\rho_{\mathrm{m}}} \,
\Sigma_{\mathrm{c}}^{-1}(\chi_{\mathrm{l}},\chi_{\mathrm{s}}) \,
\int_0^{\chi_{\mathrm{s}}} d\chi \, \delta_{\mathrm{m}}(\chi,\ho) .
\label{eqkappa3}
\end{equation}
Eq.\ \ref{eqkappa3} motivates that the weak lensing observable can be
related to the projected mass density around the lens, $\Sigma = \int
\rho_{\mathrm{m}} \, d\chi$, where $\delta_{\mathrm{m}} =
\rho_{\mathrm{m}}/\overline{\rho_{\mathrm{m}}} - 1$. The convergence
may be written in terms of this quantity as
\begin{equation}
\kappa(\chi_{\mathrm{l}},\chi_{\mathrm{s}},\ho) \approx
\Sigma_{\mathrm{c}}^{-1}(\chi_{\mathrm{l}},\chi_{\mathrm{s}}) \left(
\Sigma - \overline{\Sigma} \right) ,
\end{equation}
where $\overline{\Sigma} = \int \overline{\rho} \, d\chi$ represents
the average background, emphasising that gravitational lensing traces
the increment between the mass density and the background.
The average tangential shear $\gamma_{\mathrm{t}}$ at angular
separation $\theta$ from an axisymmetric lens is related to the
convergence as
\begin{equation}
\langle \gamma_{\mathrm{t}}(\theta) \rangle = \langle
\overline{\kappa}(<\theta) \rangle - \langle \kappa(\theta) \rangle
,
\label{eqgt}
\end{equation}
where $\overline{\kappa}(<\theta)$ is the mean convergence within
separation $\theta$. At the location of the lens, angular separations
are related to projected separations as $R = \chi(z_{\mathrm{l}}) \,
\theta$. Defining the differential projected surface mass density
around the lens as a function of projected separation,
\begin{equation}
\Delta \Sigma(R) = \overline{\Sigma}(<R) - \Sigma(R) ,
\label{eqdsigdef}
\end{equation}
where,
\begin{equation}
\overline{\Sigma}(<R) = \frac{2}{R^2} \int_0^R R' \, \Sigma(R') \, dR'
,
\end{equation}
we find that for a single source-lens pair at distances
$\chi_{\mathrm{l}}$ and $\chi_{\mathrm{s}}$ (omitting the angled
brackets),
\begin{equation}
\gamma_{\mathrm{t}}(\theta) =
\Sigma_{\mathrm{c}}^{-1}(\chi_{\mathrm{l}},\chi_{\mathrm{s}}) \,
\Delta\Sigma(R) .
\label{eqgtdsig}
\end{equation}
\subsection{Galaxy-convergence cross-correlation}
For an ensemble of sources with distance probability distribution
$p_{\mathrm{s}}(\chi)$ (normalised such that $\int
p_{\mathrm{s}}(\chi) \, d\chi = 1$), the total convergence in a given
sky direction is
\begin{equation}
\begin{split}
\kappa(\ho) &= \int d\chi_{\mathrm{s}} \,
p_{\mathrm{s}}(\chi_{\mathrm{s}}) \, \kappa(\chi_{\mathrm{s}},\ho)
\\ &= \overline{\rho_{\mathrm{m}}} \int_0^\infty d\chi \,
\overline{\Sigma_{\mathrm{c}}^{-1}}(\chi) \,
\delta_{\mathrm{m}}(\chi,\ho) ,
\end{split}
\end{equation}
where,
\begin{equation}
\overline{\Sigma_{\mathrm{c}}^{-1}}(\chi) = \int_\chi^\infty
d\chi_{\mathrm{s}} \, p_{\mathrm{s}}(\chi_{\mathrm{s}}) \,
\Sigma_{\mathrm{c}}^{-1}(\chi,\chi_{\mathrm{s}}) ,
\label{eqavesigc}
\end{equation}
with the lower limit of the integral applying because
$\Sigma_{\mathrm{c}}^{-1}(\chi_{\mathrm{l}},\chi_{\mathrm{s}}) = 0$
for $\chi_{\mathrm{s}} < \chi_{\mathrm{l}}$. We consider forming the
angular cross-correlation function of this convergence field with the
projected number overdensity of an ensemble of lenses with distance
probability distribution $p_{\mathrm{l}}(\chi)$,
\begin{equation}
\delta_{\mathrm{g,2D}}(\ho) = \int d\chi \, p_{\mathrm{l}}(\chi) \,
\delta_{\mathrm{g}}(\chi,\ho) .
\end{equation}
The galaxy-convergence cross-correlation function at angular
separation $\vt$ is,
\begin{equation}
\omega_{\mathrm{g\kappa}}(\vt) = \langle \kappa(\ho) \,
\delta_{\mathrm{g,2D}}(\ho + \vt) \rangle .
\end{equation}
Expressing the overdensity fields in terms of their Fourier components
we find, after some algebra,
\begin{equation}
\omega_{\mathrm{g\kappa}}(\vt) = \int \frac{d^2\vl}{(2\mathrm{\pi})^2} \,
C_{\mathrm{g\kappa}}(\vl) \, \mathrm{e}^{-\mathrm{i}\vl\cdot\vt} ,
\label{eqwgk}
\end{equation}
where $\vl$ is a 2D Fourier wavevector, and the corresponding angular
cross-power spectrum $C_{\mathrm{g\kappa}}(\ell)$ is given by
\citep{Guzik01,Hu04,Joachimi10},
\begin{equation}
C_{\mathrm{g\kappa}}(\ell) = \overline{\rho_{\mathrm{m}}} \int d\chi
\, p_{\mathrm{l}}(\chi) \,
\frac{\overline{\Sigma_{\mathrm{c}}^{-1}}(\chi)}{\chi^2} \,
P_{\mathrm{gm}} \left( \frac{\ell}{\chi},\chi \right) ,
\label{eqclgk}
\end{equation}
where $P_{\mathrm{gm}}(k,\chi)$ is the 3D galaxy-matter cross-power
spectrum at wavenumber $k$ and distance $\chi$. Taking the azimuthal
average of Eq.\ \ref{eqwgk} over all directions $\vt$, the complex
exponential integrates to a Bessel function of the first kind,
$J_0(x)$, such that,
\begin{equation}
\omega_{\mathrm{g\kappa}}(\theta) = \int
\frac{d^2\vl}{(2\mathrm{\pi})^2} \, C_{\mathrm{g\kappa}}(\vl) \,
J_0(\ell \theta) = \int \frac{d\ell \, \ell}{2\mathrm{\pi}} \,
C_{\mathrm{g\kappa}}(\ell) \, J_0(\ell \theta) .
\end{equation}
Using Eq.\ \ref{eqgt} and Bessel function identities, we can then
obtain an expression for the statistical average tangential shear
around an ensemble of lenses,
\begin{equation}
\gamma_{\mathrm{t}}(\theta) =
\overline{\omega_{\mathrm{g\kappa}}}(<\theta) -
\omega_{\mathrm{g\kappa}}(\theta) = \int \frac{d\ell \,
\ell}{2\mathrm{\pi}} \, C_{\mathrm{g\kappa}}(\ell) \, J_2(\ell
\theta) .
\label{eqgtmod}
\end{equation}
Likewise, we can generalise Eq.\ \ref{eqgtdsig} to apply to broad
source and lens distributions:
\begin{equation}
\gamma_{\mathrm{t}}(\theta) = \int d\chi \, p_{\mathrm{l}}(\chi) \,
\overline{\Sigma_{\mathrm{c}}^{-1}}(\chi) \, \Delta\Sigma(R,\chi) .
\label{eqgtdsigbroad}
\end{equation}
Comparing the formulations of Eqs.\ \ref{eqgtmod} and
\ref{eqgtdsigbroad} allows us to demonstrate that,
\begin{equation}
\Sigma(R) = \overline{\rho_{\mathrm{m}}} \int_{-\infty}^\infty d\Pi
\, \left[ 1 + \xi_{\mathrm{gm}}(R,\Pi) \right] ,
\label{eqsig}
\end{equation}
in terms of the 3D galaxy-matter cross-correlation function
$\xi_{\mathrm{gm}}(R,\Pi)$ at projected separation $R$ and
line-of-sight separation $\Pi$, where the constant term `$1+$'
cancels out in the evaluation of the observable $\Delta \Sigma$.
After some algebra we find,
\begin{equation}
\Delta \Sigma(R) = \overline{\rho_{\mathrm{m}}} \int_0^\infty dr \,
W(r,R) \, \xi_{\mathrm{gm}}(r) ,
\end{equation}
where,
\begin{equation}
W(r,R) = \frac{4r^2}{R^2} - \left[ \frac{4r \sqrt{r^2-R^2}}{R^2} +
\frac{2r}{\sqrt{r^2 - R^2}} \right] \, H(r-R) ,
\end{equation}
where $H(x) = 0$ if $x<0$ and $H(x) = 1$ if $x>0$ is the Heaviside
step function. The relations in this section make the approximations
of using the Limber equation \citep{Limber53} and neglecting
additional effects such as cosmic magnification \citep{Unruh20} and
intrinsic alignments \citep{Joachimi15}.
\subsection{Auto-correlation functions}
In order to determine the analytical covariance in
Sect.\ \ref{seccov}, we also need expressions for the auto-correlation
functions of the convergence, $\omega_{\mathrm{\kappa\kappa}}(\vt) =
\int \frac{d^2\vl}{(2\mathrm{\pi})^2} \,
C_{\mathrm{\kappa\kappa}}(\vl) \,
\mathrm{e}^{-\mathrm{i}\vl\cdot\vt}$, and the galaxy overdensity,
$\omega_{\mathrm{gg}}(\vt) = \int \frac{d^2\vl}{(2\mathrm{\pi})^2} \,
C_{\mathrm{gg}}(\vl) \, \mathrm{e}^{-\mathrm{i}\vl\cdot\vt}$. Given
two source populations with distance probability distributions
$p_{\mathrm{s},1}(\chi)$ and $p_{\mathrm{s},2}(\chi)$, and associated
integrated critical density functions
$\overline{\Sigma^{-1}_{\mathrm{c},1}}$ and
$\overline{\Sigma^{-1}_{\mathrm{c},2}}$, the angular power spectrum of
the convergence is given by,
\begin{equation}
C_{\mathrm{\kappa\kappa}}(\ell) = \overline{\rho_{\mathrm{m}}}^2
\int d\chi \, \frac{\overline{\Sigma_{\mathrm{c},1}^{-1}}(\chi) \,
\overline{\Sigma_{\mathrm{c},2}^{-1}}(\chi)}{\chi^2} \,
P_{\mathrm{mm}}\left( \frac{\ell}{\chi},\chi \right) ,
\end{equation}
where $P_{\mathrm{mm}}(k,\chi)$ is the 3D (non-linear) matter power
spectrum at wavenumber $k$ and distance $\chi$. Likewise, for two
projected galaxy overdensity fields with distance probability
distributions $p_{\mathrm{l},1}(\chi)$ and $p_{\mathrm{l},2}(\chi)$,
the angular power spectrum is,
\begin{equation}
C_{\mathrm{gg}}(\ell) = \int d\chi \, \frac{p_{\mathrm{l},1}(\chi) \,
p_{\mathrm{l},2}(\chi)}{\chi^2} \, P_{\mathrm{gg}}\left(
\frac{\ell}{\chi},\chi \right) ,
\label{eqclgg}
\end{equation}
where $P_{\mathrm{gg}}(k,\chi)$ is the 3D galaxy power spectrum.
\subsection{Bias model}
\label{secbias}
We computed the linear matter power spectrum $P_{\mathrm{L}}(k)$ in
our models using the CAMB software package \citep{Lewis00}, and
evaluated the non-linear matter power spectrum $P_{\mathrm{mm}}(k)$
including the `halofit' corrections \citep[][we define the fiducial
cosmological parameters used for the simulation and data analysis in
subsequent sections]{Smith03,Takahashi12}. We adopted a model for
the non-linear galaxy-galaxy and galaxy-matter 2-point functions,
appearing in Eqs.\ \ref{eqclgk} and \ref{eqclgg}, following
\citet{Baldauf10} and \citet{Mandelbaum13}. This model assumes a
local, non-linear galaxy bias relation via a Taylor expansion of the
galaxy density field in terms of the matter overdensity,
$\delta_{\mathrm{g}} = b_{\mathrm{L}} \, \delta_{\mathrm{m}} +
\frac{1}{2} b_{\mathrm{NL}} \, \delta_{\mathrm{m}}^2 + ...$, defining
a linear bias parameter $b_{\mathrm{L}}$ and non-linear bias parameter
$b_{\mathrm{NL}}$. The auto- and cross-correlation statistics in this
model can be written in the form \citep{McDonald06,Smith09},
\begin{equation}
\begin{split}
\xi_{\mathrm{gg}} &= b_{\mathrm{L}}^2 \, \xi_{\mathrm{mm}} + 2 \,
b_{\mathrm{L}} \, b_{\mathrm{NL}} \, \xi_{\mathrm{A}} +
\frac{1}{2} \, b_{\mathrm{NL}}^2 \, \xi_{\mathrm{B}} ,
\\ \xi_{\mathrm{gm}} &= b_{\mathrm{L}} \, \xi_{\mathrm{mm}} +
b_{\mathrm{NL}} \, \xi_{\mathrm{A}} ,
\end{split}
\end{equation}
where $\xi_{\mathrm{mm}}$ is the correlation function corresponding to
$P_{\mathrm{mm}}(k)$, and $\xi_{\mathrm{A}}$ and $\xi_{\mathrm{B}}$
are obtained by computing the Fourier transforms of,
\begin{equation}
\begin{split}
A(k) &= \int \frac{d^3\vq}{(2\mathrm{\pi})^3} \, F_2(\vq,\vk-\vq)
\, P_{\mathrm{L}}(q) \, P_{\mathrm{L}}(|\vk-\vq|) , \\ B(k) &=
\int \frac{d^3\vq}{(2\mathrm{\pi})^3} \, P_{\mathrm{L}}(q) \,
P_{\mathrm{L}}(|\vk-\vq|) ,
\end{split}
\end{equation}
which depend on the mode-coupling kernel in standard perturbation
theory,
\begin{equation}
F_2(\vq_1,\vq_2) = \frac{5}{7} + \frac{1}{2} \frac{\vq_1
. \vq_2}{q_1 q_2} \left( \frac{q_1}{q_2} + \frac{q_2}{q_1} \right)
+ \frac{2}{7} \left( \frac{\vq_1 . \vq_2}{q_1 q_2} \right)^2 .
\end{equation}
We evaluated these integrals using the {\tt FAST} software package
\citep{McEwen16} and note that $\xi_{\mathrm{B}} =
\xi_{\mathrm{L}}^2$, where $\xi_{\mathrm{L}}$ is the correlation
function corresponding to $P_{\mathrm{L}}(k)$. This model is only
expected to be valid on scales exceeding the virial radius of dark
matter haloes, since it does not address halo exclusion, the
distribution of galaxies within haloes, or other forms of stochastic
or non-local effects \citep{Asgari20b}. However, this 2-parameter
bias model is adequate for our large-scale analysis, which we verify
using representative mock catalogues in Sect.\ \ref{secmocktests}.
\section{Estimators}
\label{secest}
In this section we specify estimators that may be used to measure
$\gamma_{\mathrm{t}}(\theta)$ and $\Delta \Sigma(R)$ from ensembles of
sources and lenses, and discuss how estimates of $\Delta \Sigma(R)$
are affected by uncertainties in source distances.
\subsection{Average tangential shear $\gamma_{\mathrm{t}}(\theta)$}
We can estimate the average tangential shear of a set of sources (s)
around lenses (l) by evaluating the following expression
\citep{Mandelbaum06}, which also utilises an unclustered random lens
catalogue (r) with the same selection function as the lenses:
\begin{equation}
\hgt(\theta) = \frac{\sum\limits_{\mathrm{ls}} w_{\mathrm{l}} \,
w_{\mathrm{s}} \, e_{\mathrm{t,ls}} - \sum\limits_{\mathrm{rs}}
w_{\mathrm{r}} \, w_{\mathrm{s}} \,
e_{\mathrm{t,rs}}}{\sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \,
w_{\mathrm{s}}} .
\label{eqgtest}
\end{equation}
The sums in Eq.\ \ref{eqgtest} are taken over pairs of sources and
lenses with angular separations within a bin around $\theta$, $w_i$
are weights applied to the different samples (normalised such that
$\sum_{\mathrm{l}} w_{\mathrm{l}} = \sum_{\mathrm{r}}
w_{\mathrm{r}}$), and $e_{\mathrm{t}}$ indicates the tangential
ellipticity of the source, projected onto an axis normal to the line
joining the source and lens (or random lens).
Eq.\ \ref{eqgtest} involves the random lens catalogue in two places.
First, the tangential shear of sources around random lenses is
subtracted from the data signal. The subtracted term has an
expectation value of zero, but significantly decreases the variance of
the estimator at large separations \citep{Singh17}. Second, the
estimator is normalised by a sum over pairs of sources and random
lenses, rather than data lenses. This ensures that the estimator is
unbiased: the alternative estimator, $\hgt = \sum_{\mathrm{ls}}
w_{\mathrm{l}} \, w_{\mathrm{s}} \, e_{\mathrm{t,ls}} /
\sum_{\mathrm{ls}} w_{\mathrm{l}} \, w_{\mathrm{s}}$, is biased by any
source-lens clustering (if the angular cross-correlation function
$\omega_{\mathrm{ls}}(\theta) \ne 0$), which would modify the
denominator of the expression but not the numerator. The magnitude of
this effect is sometimes known as the `boost' factor
\citep{Sheldon04},
\begin{equation}
B(\theta) = \frac{\sum\limits_{\mathrm{ls}} w_{\mathrm{l}} \,
w_{\mathrm{s}}}{\sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \,
w_{\mathrm{s}}} ,
\end{equation}
where the sums are again taken over source-lens pairs with angular
separations within a given bin. We note that $\langle B(\theta)
\rangle = 1 + \omega_{\mathrm{ls}}(\theta)$ for unity weights.
\subsection{Projected mass density $\Delta\Sigma(R)$}
Assuming the source and lens distances are known, each source-lens
pair may be used to estimate the projected mass density around the
lenses by inverting Eq.\ \ref{eqgtdsig}:
\begin{equation}
\hds(R) = e_{\mathrm{t}}(R/\chi_{\mathrm{l}}) \,
\Sigma_{\mathrm{c}}(\chi_{\mathrm{l}},\chi_{\mathrm{s}}) .
\end{equation}
For an ensemble of sources and lenses, the mean projected mass density
may then be estimated by an expression analogous to Eq.\
\ref{eqgtest} \citep{Singh17},
\begin{equation}
\begin{split}
\hds(R) = \frac{\sum\limits_{\mathrm{ls}} w_{\mathrm{l}} \, w_{\mathrm{s}}
\, w_{\mathrm{ls}} \, e_{\mathrm{t,ls}}(R/\chi_{\mathrm{l}}) \,
\Sigma_{\mathrm{c}}(\chi_{\mathrm{l}},\chi_{\mathrm{s}})}{\sum\limits_{\mathrm{rs}}
w_{\mathrm{r}} \, w_{\mathrm{s}} \, w_{\mathrm{rs}}} \\ -
\frac{\sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \, w_{\mathrm{s}} \,
w_{\mathrm{rs}} \, e_{\mathrm{t,rs}}(R/\chi_{\mathrm{r}}) \,
\Sigma_{\mathrm{c}}(\chi_{\mathrm{r}},\chi_{\mathrm{s}})}{\sum\limits_{\mathrm{rs}}
w_{\mathrm{r}} \, w_{\mathrm{s}} \, w_{\mathrm{rs}}} ,
\end{split}
\label{eqdsigest1}
\end{equation}
where we have allowed for an additional pair weight between sources
and lenses, $w_{\mathrm{ls}}$, and random lenses, $w_{\mathrm{rs}}$.
Assuming a constant shape noise in $e_{\mathrm{t}}$, the noise in the
estimate of $\Delta \Sigma(R) = e_{\mathrm{t}} \, \Sigma_{\mathrm{c}}$
from each source-lens pair is proportional to $\Sigma_{\mathrm{c}}$,
hence the optimal inverse-variance weight is $w_{\mathrm{ls}} \propto
\Sigma_{\mathrm{c}}^{-2}$, and the weighted estimator may be written,
\begin{equation}
\begin{split}
\hds(R) = \frac{\sum\limits_{\mathrm{ls}} w_{\mathrm{l}} \, w_{\mathrm{s}}
\, e_{\mathrm{t,ls}}(R/\chi_{\mathrm{l}}) \,
\Sigma^{-1}_{\mathrm{c,ls}}}{\sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \,
w_{\mathrm{s}} \, \Sigma^{-2}_{\mathrm{c,rs}}} \\ -
\frac{\sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \, w_{\mathrm{s}} \,
e_{\mathrm{t,rs}}(R/\chi_{\mathrm{r}}) \,
\Sigma^{-1}_{\mathrm{c,rs}}}{\sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \,
w_{\mathrm{s}} \, \Sigma^{-2}_{\mathrm{c,rs}}} .
\end{split}
\label{eqdsigest2}
\end{equation}
\subsection{Photo-$z$ dilution correction for $\Delta\Sigma(R)$}
\label{secphotoz}
The difficulty faced when determining $\Delta \Sigma$ is that source
distances are typically only accessible through photometric redshifts
and may contain significant errors, leading to a bias in the estimate
through incorrect scaling factors $\Sigma_{\mathrm{c}}$ (we assume in
this discussion that spectroscopic lens distances are available). For
example, sources may apparently lie behind lenses according to their
photometric redshift, whilst in fact being positioned in front of the
lenses and contributing no galaxy-galaxy lensing signal, creating a
downward bias in the measurement.
For a single source-lens pair, the estimated value of
$\Sigma_{\mathrm{c}}$ for the pair based on the source photometric
redshift, $\Sigma_{\mathrm{c,lp}}$, may differ from its true value
based on the source spectroscopic redshift, $\Sigma_{\mathrm{c,ls}}$,
\begin{equation}
\hds = e_{\mathrm{t}} \, \Sigma_{\mathrm{c,lp}} = \left( \frac{\Delta
\Sigma^{\mathrm{true}}}{\Sigma_{\mathrm{c,ls}}} \right)
\Sigma_{\mathrm{c,lp}} = \left(
\frac{\Sigma_{\mathrm{c,lp}}}{\Sigma_{\mathrm{c,ls}}} \right) \Delta
\Sigma^{\mathrm{true}} .
\end{equation}
Combining many source-lens pairs allowing for a pair weight
$w_{\mathrm{ls}}$ we find,
\begin{equation}
\hds = \frac{\sum\limits_{\mathrm{ls}} w_{\mathrm{ls}} \left(
\frac{\Sigma_{\mathrm{c,lp}}}{\Sigma_{\mathrm{c,ls}}} \right)
\Delta \Sigma^{\mathrm{true}}}{\sum\limits_{\mathrm{ls}} w_{\mathrm{ls}}}
.
\end{equation}
Using the optimal weight $w_{\mathrm{ls}} \propto
\Sigma^{-2}_{\mathrm{c,lp}}$ this expression may be written,
\begin{equation}
\hds = \frac{\sum\limits_{\mathrm{ls}} \Sigma^{-1}_{\mathrm{c,lp}} \,
\Sigma^{-1}_{\mathrm{c,ls}} \, \Delta
\Sigma^{\mathrm{true}}}{\sum\limits_{\mathrm{ls}}
\Sigma^{-2}_{\mathrm{c,lp}}} .
\end{equation}
The estimated value of $\Delta \Sigma$ hence contains a multiplicative
bias, $\Delta \Sigma^{\mathrm{true}} = f_{\mathrm{bias}} \, \langle
\hds \rangle$ where,
\begin{equation}
f_{\mathrm{bias}} = \frac{\sum\limits_{\mathrm{ls}}
\Sigma^{-2}_{\mathrm{c,lp}}}{\sum\limits_{\mathrm{ls}}
\Sigma^{-1}_{\mathrm{c,lp}} \, \Sigma^{-1}_{\mathrm{c,ls}}} =
\frac{\sum\limits_{\mathrm{ls}} w_{\mathrm{ls}}}{\sum\limits_{\mathrm{ls}}
w_{\mathrm{ls}} \, \Sigma_{\mathrm{c,lp}} \,
\Sigma^{-1}_{\mathrm{c,ls}}} .
\label{eqfbias}
\end{equation}
This multiplicative correction factor may be estimated at each lens
redshift from a representative subset of sources with complete
spectroscopic and photometric redshift information, by evaluating the
sums in the numerator and denominator of Eq.\ \ref{eqfbias}
\citep{Nakajima12}.
An alternative formulation of the photo-$z$ dilution correction may be
derived from the statistical distance distribution of the sources.
Provided that the lens distribution is sufficiently narrow,
Eq.\ \ref{eqgtdsigbroad} indicates that an unbiased estimate of
$\Delta\Sigma$ from each lens-source pair is,
\begin{equation}
\hds(R) = e_{\mathrm{t}}(R/\chi_{\mathrm{l}}) \, \left[
\overline{\Sigma_{\mathrm{c}}^{-1}}(\chi_{\mathrm{l}})
\right]^{-1} ,
\end{equation}
where $\overline{\Sigma_{\mathrm{c}}^{-1}}$ is evaluated from
Eq.\ \ref{eqavesigc} using the source distribution
$p_{\mathrm{s}}(\chi)$. This motivates an alternative estimator
mirroring Eq.\ \ref{eqdsigest2} \citep{Sheldon04,Miyatake15,Blake16a},
\begin{equation}
\begin{split}
\hds(R) = \frac{\sum\limits_{\mathrm{ls}} w_{\mathrm{l}} \, w_{\mathrm{s}}
\, e_{\mathrm{t,ls}}(R/\chi_{\mathrm{l}}) \,
\overline{\Sigma^{-1}_{\mathrm{c,ls}}}}{\sum\limits_{\mathrm{rs}}
w_{\mathrm{r}} \, w_{\mathrm{s}} \, \left(
\overline{\Sigma^{-1}_{\mathrm{c,rs}}} \right)^2} \\ -
\frac{\sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \, w_{\mathrm{s}} \,
e_{\mathrm{t,rs}}(R/\chi_{\mathrm{r}}) \,
\overline{\Sigma^{-1}_{\mathrm{c,rs}}}}{\sum\limits_{\mathrm{rs}}
w_{\mathrm{r}} \, w_{\mathrm{s}} \, \left(
\overline{\Sigma^{-1}_{\mathrm{c,rs}}} \right)^2} .
\end{split}
\label{eqdsigest3}
\end{equation}
The accuracy of these potential photo-$z$ dilution corrections must be
assessed via simulations, which we consider in
Sect.\ \ref{secmocktests}. We trialled both point-based and
distribution-based correction methods in our analysis.
\section{Amplitude-ratio test}
\label{secamp}
In this section we construct test statistics which utilise the
relative amplitudes of galaxy clustering and galaxy-galaxy lensing to
test cosmological models. We first define the input statistics for
these tests.
\subsection{Projected clustering $w_{\mathrm{p}}(R)$}
The amplitude of galaxy-galaxy lensing is sensitive to the
distribution of matter around lens galaxies, projected along the
line-of-sight. We can obtain an analogous projected quantity for lens
galaxy clustering by integrating the 3D galaxy auto-correlation
function, $\xi_{\mathrm{gg}}$ along the line-of-sight,
\begin{equation}
w_{\mathrm{p}}(R) = \int_{-\infty}^\infty d\Pi \,
\xi_{\mathrm{gg}}(R,\Pi) ,
\end{equation}
where $\Pi$ is the line-of-sight separation. This formulation has the
additional feature of reducing sensitivity of the clustering
statistics to redshift-space distortions, which modulate the apparent
radial separations $\Pi$ between galaxy pairs.
We can estimate $w_{\mathrm{p}}(R)$ for a galaxy sample by measuring
the galaxy correlation function in $(R,\Pi)$ separation bins, and
summing over the $\Pi$ direction in the range $0 < \Pi <
\Pi_{\mathrm{max}}$:
\begin{equation}
\hwp(R) = 2 \sum_{{\mathrm{bins}} \; i} \Delta \Pi_i \,
\hxi_{\mathrm{gg}}(R,\Pi) .
\label{eqwpest}
\end{equation}
\subsection{The Upsilon statistics, $\Upsilon_{\mathrm{gm}}(R)$ and $\Upsilon_{\mathrm{gg}}(R)$}
Eq.\ \ref{eqdsigdef} demonstrates that the amplitude of
$\Delta\Sigma(R)$ around lens galaxies depends on the surface density
of matter across a range of smaller scales from zero to $R$, and hence
on the galaxy-matter cross-correlation coefficient at these scales.
Given that this cross-correlation is a complex function which is
difficult to model from first principles, it is beneficial to reduce
this sensitivity to small-scale information using the annular
differential surface density statistic
\citep{Reyes10,Baldauf10,Mandelbaum13},
\begin{equation}
\begin{split}
\Upsilon_{\mathrm{gm}}(R,R_0) &= \Delta\Sigma(R) - \frac{R_0^2}{R^2}
\Delta\Sigma(R_0) \\ &= \frac{2}{R^2} \int_{R_0}^R dR' \, R' \,
\Sigma(R') - \Sigma(R) + \frac{R_0^2}{R^2} \Sigma(R_0) ,
\end{split}
\label{equpsgm}
\end{equation}
which is defined such that $\Upsilon_{\mathrm{gm}} = 0$ at some
small-scale limit $R = R_0$, chosen to be large enough to reduce the
main systematic effects (typically, $R_0$ is somewhat larger than the
size scale of dark matter haloes). In this sense, the cumulative
effect from the cross-correlation function at scales $R < R_0$ is
cancelled, although it is not the case that this small-scale
suppression translates to Fourier space
\citep{Baldauf10,Asgari20b,Park20}. In any case, the efficacy of
these statistics and choice of the $R_0$ value must be validated using
simulations, as we consider below.
The corresponding quantity suppressing the small-scale contribution to
the galaxy auto-correlations is \citep{Reyes10},
\begin{equation}
\begin{split}
& \Upsilon_{\mathrm{gg}}(R,R_0) = \rho_{\mathrm{c}} \\ & \left[
\frac{2}{R^2} \int_{R_0}^R dR' \, R' \, w_{\mathrm{p}}(R') -
w_{\mathrm{p}}(R) + \frac{R_0^2}{R^2} \, w_{\mathrm{p}}(R_0)
\right] ,
\end{split}
\label{equpsgg}
\end{equation}
where $\rho_{\mathrm{c}}$ is the critical matter density. We note
that if $w_{\mathrm{p}}$ is defined as a step-wise function in bins
$R_i$ (with bin limits $R_{i,\mathrm{min}}$ and $R_{i,\mathrm{max}}$)
then Eq.\ \ref{equpsgg} may be written in the useful form,
\begin{equation}
\Upsilon_{\mathrm{gg}}(R,R_0) = \frac{\rho_{\mathrm{c}}}{R^2}
\sum_{i=j}^k C_i \, w_{\mathrm{p}}(R_i) ,
\label{equpsgg2}
\end{equation}
where $(k,j)$ are the bins containing $(R,R_0)$, and
\begin{equation}
C_i = \begin{cases} R_{i,\mathrm{max}}^2 & i = j
\\ R_{i,\mathrm{max}}^2 - R_{i,\mathrm{min}}^2 & j < i < k
\\ -R_{i,\mathrm{min}}^2 & i = k \end{cases}
\end{equation}
For convenience we chose $R_0$ to coincide with the centre of a
separation bin, such that we could use the direct measurements of
$\Delta\Sigma(R_0)$ and $w_{\mathrm{p}}(R_0)$ in Eqs.\ \ref{equpsgm}
and \ref{equpsgg} without interpolation between bins (we will show
below that our results are not sensitive to the choice of $R_0$).
\subsection{The $E_{\mathrm{G}}$ test statistic}
\label{seceg}
The relative amplitudes of weak gravitational lensing and the rate of
assembly of large-scale structure depend on the `gravitational slip'
or difference between the two space-time metric potentials. This
signature is absent in general relativity but may be significant in
modified gravity scenarios
\citep{Uzan01,Zhang07,Jain10,Bertschinger11,Clifton12}.
\citet{Zhang07} proposed that these amplitudes might be compared by
connecting the velocity field and lensing signal generated by a given
set of matter overdensities, probed via redshift-space distortions and
galaxy-galaxy lensing, respectively. \citet{Reyes10} implemented this
consistency test by constructing the statistic,
\begin{equation}
E_{\mathrm{G}}(R) = \frac{1}{\beta} \,
\frac{\Upsilon_{\mathrm{gm}}(R,R_0)}{\Upsilon_{\mathrm{gg}}(R,R_0)}
,
\label{eqeg}
\end{equation}
where $\beta = f/b_{\mathrm{L}}$ is the redshift-space distortion
parameter which governs the observed dependence of the strength of
galaxy clustering on the angle to the line-of-sight, in terms of the
linear growth rate of a perturbation, $f = d\ln{\delta}/d\ln{a}$.
Eq.\ \ref{eqeg} is independent of the linear galaxy bias
$b_{\mathrm{L}}$ and the amplitude of matter clustering $\sigma_8$,
given that $\beta \propto 1/b_{\mathrm{L}}$, $\Upsilon_{\mathrm{gm}}
\propto b_{\mathrm{L}} \, \sigma_8^2$ and $\Upsilon_{\mathrm{gg}}
\propto b_{\mathrm{L}}^2 \, \sigma_8^2$. The prediction of linear
perturbation theory for general relativity in a $\Lambda$CDM Universe
is a scale-independent value $E_{\mathrm{G}}(z) =
\Omega_{\mathrm{m}}(z=0)/f(z)$, although see \citet{Leonard15} for a
detailed discussion of this approxmation.
\section{Covariance of estimators}
\label{seccov}
In this section we present analytical formulations in the Gaussian
approximation for the covariance of estimates of
$\gamma_{\mathrm{t}}(\theta)$ and $\Delta \Sigma(R)$, and model how
this covariance is modulated by the presence of a survey mask (that
is, by edge effects). Our covariance determination hence neglects
non-Gaussian and super-sample variance components. This is a
reasonable approximation in the context of the current analysis as
these terms are subdominant (we refer the reader to Joachimi et
al.\ (in prep.) for more details on the relative amplitude of the
different covariance terms in the context of KiDS-1000).
\subsection{Covariance of average tangential shear}
\label{seccovgt}
In Appendix \ref{seccovgtap} we derive the covariance of
$\gamma_{\mathrm{t}}$ averaged within angular bins $\theta_m$ and
$\theta_n$:
\begin{equation}
\mathrm{Cov}[\gamma_{\mathrm{t}}^{ij}(\theta_m),\gamma_{\mathrm{t}}^{kl}(\theta_n)]
= \frac{1}{\Omega} \int \frac{d\ell \, \ell}{2\mathrm{\pi}} \,
\sigma^2(\ell) \, \overline{J_{2,m}}(\ell) \,
\overline{J_{2,n}}(\ell) ,
\label{eqgtcov}
\end{equation}
where $\gamma_{\mathrm{t}}^{ij}$ denotes the average tangential shear
of source sample $j$ around lens sample $i$, $\Omega$ is the total
survey angular area in steradians, and $\overline{J_{2,n}}(\ell) =
\int_{\theta_{1,n}}^{\theta_{2,n}} \frac{2\mathrm{\pi}\theta \,
d\theta}{\Omega_n} \, J_2(\ell \theta)$, where the integral is
between the bin limits $\theta_1$ and $\theta_2$ and $\Omega_n$ is
angular area of bin $n$ (i.e.\ the area of the annulus between the bin
limits). The variance $\sigma^2(\ell)$ is given by the expression for
Gaussian random fields
\citep[e.g.][]{Hu04,Bernstein09,Joachimi10,Krause17,Blanchard19},
\begin{equation}
\sigma^2(\ell) = C_{\mathrm{g\kappa}}^{il}(\ell) \,
C_{\mathrm{g\kappa}}^{kj}(\ell) + \left[
C_{\mathrm{\kappa\kappa}}^{jl}(\ell) + N_{\mathrm{\kappa\kappa}}^j
\delta^{\mathrm{K}}_{jl} \right] \, \left[
C_{\mathrm{gg}}^{ik}(\ell) + N_{\mathrm{gg}}^i
\delta^{\mathrm{K}}_{ik} \right] ,
\label{eqgtvar}
\end{equation}
where $\delta^{\mathrm{K}}_{ij}$ is the Kronecker delta. The angular
auto- and cross-power spectra appearing in Eq.\ \ref{eqgtvar} may be
evaluated using the expressions in Sect.\ \ref{sectheory}, and the
noise terms are $N_{\mathrm{\kappa\kappa}}^i =
\sigma_{\mathrm{e}}^2/\overline{n}_{\mathrm{s}}^i$ and
$N_{\mathrm{gg}}^i = 1/\overline{n}_{\mathrm{l}}^i$, where
$\sigma_{\mathrm{e}}$ is the shape noise and
$\overline{n}_{\mathrm{l}}^i$ and $\overline{n}_{\mathrm{s}}^i$ are
the angular lens and source densities of sample $i$ in units of per
steradian.
\subsection{Covariance of projected mass density}
\label{seccovdsig}
The covariance of $\Delta \Sigma$ may be deduced from the covariance
of $\gamma_{\mathrm{t}}$ using $\Delta \Sigma(R) =
\gamma_{\mathrm{t}}(\theta) / \overline{\Sigma_{\mathrm{c}}^{-1}}$,
and by scaling angular separations to projected separations at an
effective lens distance $\chi_{\mathrm{l}}$ using $\theta =
R/\chi_{\mathrm{l}}$ \citep{Singh17,Dvornik18,Shirasaki18}. We can
map multipoles $\ell$ to the projected wavevector $k =
\ell/\chi_{\mathrm{l}}$ such that,
\begin{equation}
\mathrm{Cov}[\Delta \Sigma^{ij}(R), \Delta \Sigma^{kl}(R')] =
\frac{1}{\Omega} \int \frac{dk \, k}{2\mathrm{\pi}} \, \sigma^2(k)
\, \overline{J_2}(k R) \, \overline{J_2}(k R') ,
\label{eqdsigcov}
\end{equation}
where we now express the variance in terms of projected power spectra,
\begin{equation}
\sigma^2(k) = P^{il}_{\mathrm{g\kappa}}(k) \,
P^{kj}_{\mathrm{g\kappa}}(k) + \left[
P^{jl}_{\mathrm{\kappa\kappa}}(k) + N^j_{\mathrm{\kappa\kappa}}
\delta^{\mathrm{K}}_{jl} \right] \left[ P^{ik}_{\mathrm{gg}}(k) +
N^i_{\mathrm{gg}} \delta^{\mathrm{K}}_{ik} \right] .
\end{equation}
The power spectra are given by the following relations:
\begin{equation}
P_{\mathrm{g\kappa}}(k) = \frac{\chi_{\mathrm{l}}^2 \,
C_{\mathrm{g\kappa}}(k\chi_{\mathrm{l}})}{\left[\overline{\Sigma_{\mathrm{c}}^{-1}}(\chi_{\mathrm{l}})\right]^2}
\approx \overline{\rho_{\mathrm{m}}} \int d\chi \,
p_{\mathrm{l}}(\chi) \, P_{\mathrm{gm}}(k,\chi) ,
\end{equation}
\begin{equation}
\begin{split}
& P_{\mathrm{\kappa\kappa}}(k) = \chi_{\mathrm{l}}^2 \,
C_{\mathrm{\kappa\kappa}}(k\chi_{\mathrm{l}}) \\ & =
\chi_{\mathrm{l}}^2 \, \overline{\rho_{\mathrm{m}}}^2 \int d\chi \,
\left[ \frac{\overline{\Sigma_{\mathrm{c},1}^{-1}}(\chi) \,
\overline{\Sigma_{\mathrm{c},2}^{-1}}(\chi)}{\overline{\Sigma_{\mathrm{c},1}^{-1}}(\chi_{\mathrm{l}})
\, \overline{\Sigma_{\mathrm{c},2}^{-1}}(\chi_{\mathrm{l}})}
\right] \left( \frac{\chi_{\mathrm{l}}^2}{\chi^2} \right) \,
P_{\mathrm{mm}} \left( \frac{k \, \chi_{\mathrm{l}}}{\chi},\chi
\right) ,
\end{split}
\end{equation}
\begin{equation}
P_{\mathrm{gg}}(k) = \chi_{\mathrm{l}}^2
C_{\mathrm{gg}}(k\chi_{\mathrm{l}}) \approx \int d\chi \, p_1(\chi) \,
p_2(\chi) \, P_{\mathrm{gg}}(k,\chi) ,
\end{equation}
and the noise terms are,
\begin{equation}
N_{\mathrm{\kappa\kappa}} = \frac{\sigma_{\mathrm{e}}^2 \,
\chi_{\mathrm{l}}^2}{\overline{n}_{\mathrm{s}} \, \left[
\overline{\Sigma_{\mathrm{c}}^{-1}}(\chi_{\mathrm{l}}) \right]^2}
, \; \; \; N_{\mathrm{gg}} =
\frac{\chi_{\mathrm{l}}^2}{\overline{n}_{\mathrm{l}}} .
\end{equation}
\subsection{Covariance of remaining statistics}
The expression for the analytical covariance of $w_{\mathrm{p}}(R)$
may be derived as \citep[see also,][]{Singh17},
\begin{equation}
\begin{split}
&\mathrm{Cov}[w_{\mathrm{p}}(R), w_{\mathrm{p}}(R')] = \\ &\frac{2
L_\parallel \Pi_{\mathrm{max}}}{\Omega} \int \frac{dk \,
k}{2\mathrm{\pi}} \, \sigma^2(k) \, J_0(k R) \, J_0(k R') ,
\end{split}
\end{equation}
where $L_\parallel$ is the total co-moving depth of the lens redshift
slice and the expression for the variance is,
\begin{equation}
\sigma^2(k) = \left[ P_{\mathrm{gg}}(k) + N_{\mathrm{gg}} \right]^2 ,
\end{equation}
where $P_{\mathrm{gg}}(k)$ and $N_{\mathrm{gg}}$ are the 2D projected
power spectra and noise as defined in Sect.\ \ref{seccovdsig}.
We determined the analytical covariance of
$\Upsilon_{\mathrm{gm}}(R,R_0)$ from the covariance of
$\Delta\Sigma(R)$:
\begin{equation}
\begin{split}
& \mathrm{Cov}[\Upsilon_{\mathrm{gm}}(R,R_0) ,
\Upsilon_{\mathrm{gm}}(R',R_0)] = \mathrm{Cov}[\Delta\Sigma(R) ,
\Delta\Sigma(R')] \\ &- \frac{R_0^2}{R'^2}
\mathrm{Cov}[\Delta\Sigma(R) , \Delta\Sigma(R_0) ] -
\frac{R_0^2}{R^2} \mathrm{Cov}[\Delta\Sigma(R') , \Delta\Sigma(R_0)
] \\ &+ \frac{R_0^4}{R^2 R'^2} \mathrm{Var}[\Delta\Sigma(R_0)] .
\end{split}
\label{equpsgmcov}
\end{equation}
For the case of $\Upsilon_{\mathrm{gg}}(R,R_0)$, we propagated the
covariance using Eq.\ \ref{equpsgg2}:
\begin{equation}
\begin{split}
&\mathrm{Cov}[\Upsilon_{\mathrm{gg}}(R,R_0) ,
\Upsilon_{\mathrm{gg}}(R',R_0)] =
\\ &\frac{\rho_{\mathrm{c}}^2}{R^2 \, R'^2} \sum_i \sum_j C_i \, C_j
\, \mathrm{Cov}[w_{\mathrm{p}}(R_i) , w_{\mathrm{p}}(R_j)] .
\end{split}
\label{equpsggcov}
\end{equation}
We evaluated the covariance of the $E_{\mathrm{G}}$ statistic, where
required, by assuming small fluctuations in the variables in
Eq.\ \ref{eqeg} with respect to their mean, neglecting any
correlations between the measurements:
\begin{equation}
\begin{split}
&\frac{\mathrm{Cov}[E_{\mathrm{G}}(R) \,
E_{\mathrm{G}}(R')]}{E_{\mathrm{G}}(R) \, E_{\mathrm{G}}(R')} =
\frac{\mathrm{Cov}[\Upsilon_{\mathrm{gm}}(R,R_0),\Upsilon_{\mathrm{gm}}(R',R_0)]}{\Upsilon_{\mathrm{gm}}(R,R_0)
\, \Upsilon_{\mathrm{gm}}(R',R_0)} \\ &+
\frac{\mathrm{Cov}[\Upsilon_{\mathrm{gg}}(R,R_0),
\Upsilon_{\mathrm{gg}}(R',R_0)]}{\Upsilon_{\mathrm{gg}}(R,R_0)
\, \Upsilon_{\mathrm{gg}}(R',R_0)} +
\frac{\sigma_\beta^2}{\beta^2} ,
\end{split}
\label{eqegcov}
\end{equation}
where $\sigma_\beta$ is the error in the measurement of $\beta$. This
neglect of correlations is an approximation, justified in the case of
our dataset by the fact that the sky area used for the galaxy
clustering measurement is substantially different to the sub-sample
used for galaxy-galaxy lensing (see Joachimi et al.\ (in prep.) for a
detailed justification of this approximation), and that the projected
lens clustering measurement ($\Upsilon_{\mathrm{gg}}$) is largely
insensitive to redshift-space distortions ($\beta$) owing to the
projection over the line-of-sight separations. We note that in our
fiducial fitting approach, we determined the scale-independent
statistic $\langle E_G \rangle$ through direct fits to
$\Upsilon_{\mathrm{gm}}$ and $\Upsilon_{\mathrm{gg}}$ as discussed in
Sect.\ \ref{secdatatests}, without requiring the covariance of
$E_G(R)$.
\subsection{Modification of noise term}
We can replace the noise terms in Sects.\ \ref{seccovgt} and
\ref{seccovdsig} with a more accurate computation using the survey
source and lens distributions. Neglecting the random lens term (which
is not important on the small scales for which the noise term is
significant), we find that the variance associated with the
$\gamma_{\mathrm{t}}$ estimator in Eq.\ \ref{eqgtest} is
\citep[e.g.][]{Miyatake19},
\begin{equation}
\mathrm{Var}[\gamma_{\mathrm{t}}(\theta)] = \frac{\sum\limits_{\mathrm{ls}}
w_{\mathrm{l}}^2 \, w_{\mathrm{s}}^2 \,
\sigma_{\mathrm{e}}^2}{\left( \sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \,
w_{\mathrm{s}} \right)^2} .
\end{equation}
Likewise, the variance associated with the $\Delta \Sigma$ estimator
in Eq.\ \ref{eqdsigest1} is,
\begin{equation}
\mathrm{Var}[\Delta\Sigma(R)] = \frac{\sum\limits_{\mathrm{ls}}
w_{\mathrm{l}}^2 \, w_{\mathrm{s}}^2 \, w_{\mathrm{ls}}^2 \,
\sigma_{\mathrm{e}}^2 \left( \Sigma_{\mathrm{c,ls}}
\right)^2}{\left( \sum\limits_{\mathrm{rs}} w_{\mathrm{r}} \,
w_{\mathrm{s}} \, w_{\mathrm{rs}} \right)^2} .
\end{equation}
We adopted these noise terms in our covariance model.
\subsection{Modification for survey window}
Eqs.\ \ref{eqgtcov} and \ref{eqdsigcov} for the analytical covariance
are modified by the survey window function. We can intuitively
understand the need for this modification by considering that, whilst
Fourier transforms assume periodic boundary conditions, the boundaries
of the survey restrict the number of source-lens pairs on scales that
are a significant fraction of the survey dimensions.
In Appendix \ref{seccovwinap} we derive how the covariance of a
cross-correlation function $\xi(r)$ between two Gaussian fields is
modified by the window function of the fields, $W_1(\vx)$ and
$W_2(\vx)$ \citep[see also,][]{Beutler17}. We find,
\begin{equation}
\begin{split}
& \mathrm{Cov}[\xi(r),\xi(s)] \approx \frac{A_3(r,s)}{A_2(r) \,
A_2(s)} \\ & \frac{1}{2\mathrm{\pi}} \int dk \, k \, \left[
P_{11}(k) \, P_{22}(k) + P^2_{12}(k) \right] \, J_0(kr) \, J_0(ks)
,
\end{split}
\end{equation}
where $P_{11}$, $P_{22}$ and $P_{12}$ are the auto- and cross-power
spectra of the fields and the pre-factors are given by,
\begin{equation}
\begin{split}
A_2(r) &= \int_{\mathrm{bin} \, r} d^3\vr \int d^2\vx \, W_1(\vx) \,
W_2(\vx+\vr) \\ A_3(r,s) &= \int_{\mathrm{bin} \, r} d^3\vr
\int_{\mathrm{bin} \, s} d^3\vs \int d^2\vx \, A_{12}(\vx,\vr) \,
A_{12}(\vx,\vs) ,
\end{split}
\end{equation}
where the integrals over $\vr$ and $\vs$ are performed within the
separation bin, and we have written $A_{12}(\vx,\vr) = W_1(\vx) \,
W_2(\vx+\vr)$. We hence approximated the dependence of the covariance
on the survey window by replacing the survey area in
Eqs.\ \ref{eqgtcov} and \ref{eqdsigcov} by the expression $A_2(r) \,
A_2(s) / A_3(r,s)$.
We calculated the terms $A_2$ and $A_3$ using the mean and covariance
of the pair count $R_{\mathrm{s}}R_{\mathrm{l}}(r)$ between random
source and lens realisations \citep{Landy93}, which have respective
densities $\overline{n}_{\mathrm{s}}$ and $\overline{n}_{\mathrm{l}}$.
The mean pair count in a separation bin at scale $r$ (between $r_1$
and $r_2$), containing bin area $A_{\mathrm{bin}}(r) = \mathrm{\pi}
(r_2^2 - r_1^2)$, is
\begin{equation}
\langle R_{\mathrm{s}}R_{\mathrm{l}}(r) \rangle =
\overline{n}_{\mathrm{s}} \, \overline{n}_{\mathrm{l}} \,
A_{\mathrm{bin}}(r) \, A_2(r),
\end{equation}
which allows us to find $A_2(r)$, given that the other variables are
known. The covariance of the pair count between separation bins $r$
and $s$ is,
\begin{equation}
\begin{split}
& \mathrm{Cov}[ R_{\mathrm{s}}R_{\mathrm{l}}(r) ,
R_{\mathrm{s}}R_{\mathrm{l}}(s) ] = \\ & \overline{n}_{\mathrm{s}}
\, \overline{n}_{\mathrm{l}} \, A_{\mathrm{bin}}(r) \left[ A_2(r) \,
\delta^K_{rs} + \left( \overline{n}_{\mathrm{s}} +
\overline{n}_{\mathrm{l}} \right) A_{\mathrm{bin}}(s) \, A_3(r,s)
\right] ,
\end{split}
\end{equation}
which allows us to determine $A_3(r,s)$. For all the source-lens
configurations and separation bins considered in this study, we find
that the area correction factor differs from 1.0 by less than $10\%$.
\subsection{Propagation of errors in multiplicative corrections}
\label{seccovprop}
Galaxy-galaxy lensing measurements are subject to multiplicative
correction factors arising from shape measurement calibration (see
Sect.\ \ref{secdatakids}) and, in the case of $\Delta\Sigma$, owing to
photometric redshift dilution (see Sect.\ \ref{secphotoz}). We
propagated the uncertainties in these correction factors, which are
correlated between different source and lens samples, into the
analytical covariance of the measurements. Taking $\Delta\Sigma$ as
an example and writing a general amplitude correction factor as
$\alpha$, the relation between the corrected and analytical statistics
(denoted by the superscripts `corr' and `ana', respectively) is,
\begin{equation}
\Delta\Sigma^{\mathrm{corr}}_{ijk} = \frac{\alpha_{ij} \,
\Delta\Sigma^{\mathrm{ana}}_{ijk}}{\langle \alpha_{ij} \rangle} ,
\end{equation}
which is normalised such that $\langle
\Delta\Sigma^{\mathrm{corr}}_{ijk} \rangle =
\Delta\Sigma^{\mathrm{ana}}_{ijk}$, where $i$ denotes the lens sample,
$j$ the source sample and $k$ the separation bin. We hence find,
\begin{equation}
\begin{split}
\mathrm{Cov}[ \Delta\Sigma^{\mathrm{corr}}_{ijk} ,
\Delta\Sigma^{\mathrm{corr}}_{lmn} ] &= \mathrm{Cov}[
\Delta\Sigma^{\mathrm{ana}}_{ijk} ,
\Delta\Sigma^{\mathrm{ana}}_{lmn} ] \left( 1 + C_{ij,lm} \right)
\\ &+ \langle \Delta\Sigma^{\mathrm{ana}}_{ijk} \rangle \langle
\Delta\Sigma^{\mathrm{ana}}_{lmn} \rangle \, C_{ij,lm} ,
\end{split}
\label{eqcovprop}
\end{equation}
where,
\begin{equation}
C_{ij,lm} = \frac{\langle \alpha_{ij} \alpha_{lm} \rangle - \langle
\alpha_{ij} \rangle \langle \alpha_{lm} \rangle}{\langle
\alpha_{ij} \rangle \langle \alpha_{lm} \rangle} =
\frac{\mathrm{Cov}[ \alpha_{ij}, \alpha_{lm}]}{\langle \alpha_{ij}
\rangle \langle \alpha_{lm} \rangle} .
\label{eqcovprop2}
\end{equation}
We describe our specific implementation of these equations in the case
of the KiDS dataset in Sect.\ \ref{secdatatests}.
\section{Data}
\label{secdata}
\subsection{KiDS-1000}
\label{secdatakids}
The Kilo-Degree Survey is a large optical wide-field imaging survey
optimised for weak gravitational lensing analysis, performed with the
OmegaCAM camera on the VLT Survey Telescope at the European Southern
Observatory's Paranal Observatory. The survey covers two regions of
sky each containing several hundred square degrees, KiDS-North and
KiDS-South, in four filters $(u,g,r,i)$. The companion VISTA-VIKING
survey has provided complementary imaging in near-infrared bands
$(Z,Y,J,H,K_{\mathrm{s}})$, resulting in a deep, wide, nine-band
imaging dataset.
Our study is based on the fourth public data release of the project,
KiDS-1000 \citep{Kuijken19}, which comprises $1 \, 006$ deg$^2$ of
multi-band data, more than doubling the previously-available coverage.
We used an early-science release of the KiDS-1000 shear catalogues,
which was created using the exact pipeline version and PSF modelling
strategy implemented in \citet{Hildebrandt17} for the KiDS-450
release. We note that these catalogues have not undergone any
rigorous assessment for the presence of cosmic shear systematics, but
they are sufficient for the galaxy-galaxy lensing science presented in
this paper, as this is less susceptible to systematic errors in the
lensing catalogues. The raw pixel data was processed by the {\tt
THELI} and {\tt ASTRO\_WISE} pipelines \citep{Erben13,deJong15}, and
source ellipticities were measured using {\tt lensfit}
\citep{Miller13}, assigning an optimal weight for each source, and
calibrated by a large suite of image simulations \citep{Kannawadi19}.
Photometric redshifts $z_{\mathrm{B}}$ were determined from the
nine-band imaging for each source using the Bayesian code {\tt BPZ}
\citep{Benitez00}, calibrated using spectroscopic sub-samples
\citep{Hildebrandt20}, and used to divide the sources into tomographic
bins according to the value of $z_{\mathrm{B}}$.
\subsection{BOSS}
The Baryon Oscillation Spectroscopic Survey \citep[BOSS,][]{Dawson13}
is the largest existing galaxy redshift survey, which was performed
using the Sloan Telescope between 2009 and 2014. BOSS mapped the
distribution of $1.5$ million Luminous Red Galaxies (LRGs) and quasars
across $\sim 10 \, 000$ deg$^2$, inspiring a series of cosmological
analyses including the most accurate existing measurements of baryon
acoustic oscillations and redshift-space distortions in the galaxy
clustering pattern \citep{Alam17b}. The final (Data Release 12)
large-scale structure catalogues are described by \citet{Reid16}; we
used the combined LOWZ and CMASS LRG samples in our
study.\footnote{The BOSS large-scale structure samples are available
for download at the link
\url{https://data.sdss.org/sas/dr12/boss/lss/}.}
\subsection{2dFLenS}
The 2-degree Field Lensing Survey \citep[2dFLenS,][]{Blake16b} is a
galaxy redshift survey performed at the Australian Astronomical
Observatory in 2014-2015 using the 2-degree Field spectroscopic
instrument, with the goal of extending spectroscopic-redshift coverage
of gravitational lensing surveys in the southern sky, particularly the
KiDS-South region. The 2dFLenS sample covers an area of 731 deg$^2$
and includes redshifts for $40 \, 531$ LRGs in the redshift range $z <
0.9$, selected by applying BOSS-inspired colour-magnitude cuts to the
VST-ATLAS imaging data.\footnote{The 2dFLenS dataset is publicly
available at the link \url{http://2dflens.swin.edu.au}.} The
2dFLenS dataset has already been utilised in conjunction with the
KiDS-450 lensing catalogues to perform a previous implementation of
the amplitude-ratio test \citep{Amon18}, a combined cosmological
analysis of cosmic shear tomography, galaxy-galaxy lensing and galaxy
multipole power spectra \citep{Joudaki18} and to determine photometric
redshift calibration by cross-correlation
\citep{Johnson17,Hildebrandt20}. In our study we utilised the 2dFLenS
LRG sample which overlapped with the KiDS-1000 pointings in the
southern region. Fig.\ \ref{figoverlap} illustrates the overlaps of
the KiDS-1000 source catalogues in the north and south survey regions
with the BOSS and 2dFLenS LRG catalogues.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{overlap_kids1000.png}
\caption{KiDS source and LRG lens catalogues within the KiDS-N region
(top panel) and KiDS-S region (bottom panel). The grey squares
represent the KiDS-1000 pointings, and the fluctuating background is
the gridded number density of 2dFLenS (blue) and BOSS (red)
galaxies. The open rectangles outline the footprint of the full
KiDS survey, indicating that the LRG overlap will continue to
increase as the survey is completed.}
\label{figoverlap}
\end{figure*}
\section{Mocks}
\label{secmocks}
We used the {\tt MICECATv2.0} simulation
\citep{Fosalba15a,Crocce15,Fosalba15b} to produce representative KiDS
lensing source catalogues and LRG lens catalogues for testing the
estimators, models and covariances described above. The Marenostrum
Institut de Ciencias de l'Espai (MICE) catalogues cover an octant of
the sky ($0 < \mathrm{RA} < 90^\circ$, $0 < \mathrm{Dec} < 90^\circ$)
for redshift range $z < 1.4$. We used boundaries at constant RA and
Dec to divide this area into 10 sub-samples, each of area 516 deg$^2$.
The fiducial set of cosmological parameters for the mock is
$\Omega_{\mathrm{m}} = 0.25$, $h = 0.7$, $\Omega_{\mathrm{b}} =
0.044$, $\sigma_8 = 0.8$ and $n_{\mathrm{s}} = 0.95$.
\subsection{Mock source catalogue}
We constructed the representative mock source catalogue by applying
the following steps (see van den Busch et al.\ (in prep.) for a full
description of the MICE KiDS source mocks). The MICE catalogue is
non-uniform across the octant: the region $\mathrm{Dec} < 30^\circ$
AND [($\mathrm{RA} < 30^\circ$) OR ($\mathrm{RA} > 60^\circ$)] has a
shallower redshift distribution than the remainder. Firstly, we
homogenised the catalogue with the cut ${\tt
des\_asahi\_full\_i\_true} < 24$, such that we could construct mocks
using the complete octant. The MICE catalogue shears $(\gamma_1,
\gamma_2)$ are defined by the position angle relative to the
declination axis. Given the MICE system for mapping 3D positions to
(RA, Dec) co-ordinates, the KiDS conventions can be recovered by the
following transformations: $\mathrm{RA} \rightarrow 90^\circ -
\mathrm{RA}$, $\gamma_1 \rightarrow -\gamma_1$ ($\gamma_2$ is
effectively negated twice and therefore unchanged).
We constructed a KiDS-like photometric realisation based on the galaxy
sizes and shapes, median KiDS seeing and limiting magnitudes,
including photometric noise (see van den Busch et al.\ in prep.). We
ran BPZ photometric redshift estimation \citep{Benitez00} on the mock
source magnitudes and sizes, assigning $z_{\mathrm{B}}$ values for
each object. We used a KDTree algorithm to assign weights to the mock
sources on the basis of a nearest-neighbour match to the data
catalogue in magnitude space, and randomly sub-sampled the catalogue
to match the KV450 effective source density. We produced noisy shear
components $(e_1, e_2)$ as $e = (\gamma + n)/(1 + n \gamma^*)$
\citep{Seitz97} where $\gamma = \gamma_1 + \gamma_2 \, \mathrm{i}$, $e
= e_1 + e_2 \, \mathrm{i}$ and $n = n_1 + n_2 \, \mathrm{i}$, where
$n_1$ and $n_2$ are drawn from Gaussian distributions with standard
deviation $\sigma_{\mathrm{e}} = 0.288$ \citep{Hildebrandt20}. The
redshift distribution estimates of the KiDS data and MICE mock source
tomographic samples are displayed in the left panel of
Fig.\ \ref{fignzmock}, illustrating the reasonable match between the
two catalogues.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{nz_mock_vs_data.png}
\caption{{\it Left panel:} Redshift distribution estimates of the KiDS
source catalogue (solid lines) and MICE mock source catalogue
(dashed lines) in the five source tomographic bins split by
photometric redshift. {\it Right panel:} Number density as a
function of redshift of the BOSS LRG dataset (black solid line), the
MICE mock lens catalogue with the original BOSS colour selection
(red dashed line), and the MICE mock lens catalogue with the
adjusted BOSS colour selection (blue dotted line).}
\label{fignzmock}
\end{figure*}
\subsection{Mock lens catalogue}
We constructed the representative mock LRG lens catalogue from the
MICE simulation as follows. We used the galaxy magnitudes {\tt
sdss\_g\_true}, {\tt sdss\_r\_true}, {\tt sdss\_i\_true} and first
applied the MICE evolution correction to these magnitudes as a
function of redshift, $m \rightarrow m - 0.8*[\mathrm{arctan}(1.5*z) -
0.1489]$ \citep{Crocce15}. We then constructed the LRG lens
catalogues using the BOSS LOWZ and CMASS colour cuts in terms of the
variables,
\begin{equation}
\begin{split}
& c_\parallel = 0.7 \, (g-r) + 1.2 \, (r-i-0.18) , \\
& c_\perp = (r-i) - (g-r)/4 - 0.18 , \\
& d_\perp = (r-i) - (g-r)/8 .
\end{split}
\end{equation}
Applying the original BOSS colour-magnitude selection cuts
\citep{Eisenstein11} to the MICE mock did not reproduce the BOSS
redshift distribution (which is unsurprising, since this mock has not
been tuned to do so; the BOSS data is also selected from noisy
observed magnitudes). Our approach to resolve this issue, following
\citet{Crocce15}, was to vary the colour and magnitude selection cuts
to minimise the deviation between the mock and data redshift
distributions. We applied the following LOWZ selection cuts (where we
indicate our changed values in bold font, and the previous values
immediately following in square brackets):
\begin{equation}
\begin{split}
& 16.0 < r < \mathbf{20.0} [19.6] , \\
& r < \mathbf{13.35} [13.5] + c_\parallel/0.3 , \\
& |c_\perp| < 0.2 .
\end{split}
\end{equation}
We applied the following CMASS selection cuts:
\begin{equation}
\begin{split}
& 17.5 < i < \mathbf{20.06} [19.9] , \\
& r - i < 2 , \\
& d_\perp > 0.55 , \\
& i < \mathbf{19.98} [19.86] + 1.6 \, (d_\perp-0.8) .
\end{split}
\end{equation}
The resulting redshift distributions of the MICE lens mock (original,
and after adjustment of the colour selection cuts) and the BOSS data
are shown in the right panel of Fig.\ \ref{fignzmock}, illustrating
that our modified selection produced a much-improved representation of
the BOSS dataset. The clustering amplitude of the MICE LRG mock
catalogues was consistent with a galaxy bias factor $b \approx 2$,
although did not precisely match the clustering of the BOSS dataset,
since it was not tuned to do so. However, these representative
catalogues nonetheless allowed us to test our analysis procedures.
\section{Simulation tests}
\label{secmocktests}
In this section we analyse the representative source and lens
catalogues constructed from the MICE mocks described in Sect.\
\ref{secmocks}. Our specific goals are to:
\begin{itemize}
\item Test that the non-linear galaxy bias model specified in
Sect.\ \ref{secbias} is adequate for modelling the galaxy-galaxy
lensing and clustering statistics across the relevant scales.
\item Test that the approaches to the photo-$z$ dilution correction
of $\Delta\Sigma$ described in Sect.\ \ref{secphotoz} recovered
results consistent with those obtained using source spectroscopic
redshifts.
\item Use the multiple mock realisations and jack-knife techniques
to test that the covariance of the estimated statistics is
consistent with the analytical Gaussian covariance specified in
Sect.\ \ref{seccov}.
\item Test that the $E_{\mathrm{G}}$ test statistics constructed
from the mock as described in Sect.\ \ref{seceg} are consistent
with the theoretical expectation, and determine the degree to
which this result depends on the choice of the small-scale cut-off
parameter $R_0$ (see Eq.\ \ref{equpsgm}).
\item Test that, given our galaxy bias model, the galaxy-galaxy
lensing and clustering statistics may be jointly described by a
normalisation parameter $\sigma_8$ that is consistent with the
mock fiducial cosmology, and use this test to assess the relative
precision of angular and projected estimators.
\end{itemize}
Consistent with our subsequent data analysis, we divided the source
catalogues into five different tomographic samples by the value of the
BPZ photometric redshift, with divisions $z_{\mathrm{B}} = [0.1, 0.3,
0.5, 0.7, 0.9, 1.2]$ \citep[following][]{Hildebrandt20}. We divided
the lens catalogue into five slices of spectroscopic redshift
$z_{\mathrm{l}}$ of width $\Delta z_{\mathrm{l}} = 0.1$ in the range
$0.2 < z_{\mathrm{l}} < 0.7$. This narrow spectroscopic slicing
minimises systematic effects due to redshift evolution across the lens
slice \citep{Leauthaud17,Singh19}.
\subsection{Measurements}
\label{secmicemeas}
We measured the following three statistics, which form the fundamental
set of measurements from which the associated statistics are derived.
Firstly, we measured the average tangential shear
$\gamma_{\mathrm{t}}(\theta)$ between all tomographic pairs of source
and lens samples, in 15 logarithmically-spaced angular bins in the
range $0.005^\circ < \theta < 5^\circ$, using the estimator of
Eq.\ \ref{eqgtest}. This measurement is displayed in
Fig.\ \ref{figgttom} as the mean of the 10 individual mock
realisations (which each have area 516 deg$^2$), for each of the five
source samples against the five lens redshift slices.
Secondly, we measured the projected mass density $\Delta\Sigma(R)$
between all tomographic pairs of source and lens samples, in 15
logarithmically-spaced projected separation bins in the range $0.1 < R
< 100 \, h^{-1}$ Mpc. The mock mean measurement is displayed in
Fig.\ \ref{figdsigtom} in units of $h \, M_\odot$ pc$^{-2}$, for each
of the five source samples against the five lens redshift slices.
When performing a $\Delta\Sigma$ measurement between source and lens
samples we only included individual source-lens galaxy pairs with
$z_{\mathrm{B}} > z_{\mathrm{l}}$, for which the source photometric
redshift lies behind the lens spectroscopic redshift (adopting an
alternative cut $z_{\mathrm{B}} > z_{\mathrm{l}} + 0.1$ did not change
the results significantly). We applied the photo-$z$ dilution
correction $f_{\mathrm{bias}}$ computed using Eq.\ \ref{eqfbias} based
on the point photo-$z$ values, and we study the efficacy of this
correction in Sect.\ \ref{secmicephotoz} below.
Thirdly, we measured the projected clustering $w_{\mathrm{p}}(R)$ of
the lens samples in the same projected separation bins as above, using
the estimator of Eq.\ \ref{eqwpest} with $\Pi_{\mathrm{max}} = 100 \,
h^{-1}$ Mpc. The mock mean measurement of $w_{\mathrm{p}}(R)$ is
displayed as the third row of Fig.\ \ref{figstatmice}, for each of the
five redshift slices. Each of these measurements will be compared
with cosmological model predictions as described in the subsections
below.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{gttom_kids_vs_mice.png}
\caption{Measurements of the average tangential shear,
$\gamma_{\mathrm{t}}(\theta)$, for the KiDS-LRG dataset (black
points) and representative MICE mocks (red points), between all
pairs of lens spectroscopic redshift slices (rows) and source
tomographic samples split by photometric redshift (columns). The
errors are derived from the diagonal elements of the full analytical
covariance matrix (where we note that measurements are correlated
across scales and samples). The overplotted model is not a fit to
this dataset, but rather a prediction based on the galaxy bias
parameter fits to the $\Upsilon_{\mathrm{gg}}$ statistic of each
lens redshift slice, which is inaccurate on small scales. The
$y$-values are scaled by a factor of $\theta$ (in degrees) for
clarity.}
\label{figgttom}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{delsigtom_kids_vs_mice.png}
\caption{Measurements of the projected mass density,
$\Delta\Sigma(R)$, for the KiDS-LRG dataset (black points) and
representative MICE mocks (red points), displayed in the same style
as Fig.\ \ref{figgttom}. No measurements are possible for the lower
left-hand set of panels, owing to the adopted cut in source-lens
pairs, $z_{\mathrm{B}} > z_{\mathrm{l}}$. The units of
$\Delta\Sigma$ are $h \, M_\odot$ pc$^{-2}$.}
\label{figdsigtom}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{allstat_mice.png}
\caption{Measurements of a series of galaxy-galaxy lensing and
clustering statistics (rows) for each different lens redshift slice
(columns) of the MICE mocks. The first row displays the projected
mass density $\Delta\Sigma(R)$, where measurements corresponding to
the different source tomographic samples have been optimally
combined following the procedure described in Appendix
\ref{seccombap}. The second row shows the corresponding
measurements of $\Upsilon_{\mathrm{gm}}(R)$, derived from
$\Delta\Sigma(R)$ using Eq.\ \ref{equpsgm} and assuming $R_0 = 2 \,
h^{-1}$ Mpc. The third and fourth rows display the galaxy
clustering statistics $w_{\mathrm{p}}(R)$ and
$\Upsilon_{\mathrm{gg}}(R)$ for each lens redshift slice, where
$w_{\mathrm{p}}(R)$ is the projected correlation function
measurement, and $\Upsilon_{\mathrm{gg}}(R)$ is the associated
quantity suppressing contributions from small scales, derived using
Eq.\ \ref{equpsgg} and again assuming $R_0 = 2 \, h^{-1}$ Mpc. The
fifth row shows the combined-probe statistic $E_{\mathrm{G}}(R)$,
derived from the $\Upsilon_{\mathrm{gm}}(R)$ and
$\Upsilon_{\mathrm{gg}}(R)$ measurements using Eq.\ \ref{eqeg}.
Errors are displayed using the diagonal elements of the analytical
covariance matrix, propagating errors where appropriate. The
overplotted models are determined using the galaxy bias factors
fitted to the $\Upsilon_{\mathrm{gg}}$ measurements for each lens
redshift slice, and $\chi^2$ statistics between the mock mean data
and model are displayed in each panel.}
\label{figstatmice}
\end{figure*}
We computed the covariance matrix for each statistic using the
analytical Gaussian covariance specified in Sect.\ \ref{seccov}, where
we initially used a fiducial lens linear bias factor $b_{\mathrm{L}} =
1.8$, and iterated this value following a preliminary fit to the
projected lens clustering. Fig.\ \ref{figdelsigerrmice} compares
three different determinations of the error in $\Delta\Sigma(R)$ for
each individual 516 deg$^2$ realisation of the MICE mocks: using the
analytical covariance, using a jack-knife analysis, and derived from
the standard deviation of the 10 realisations. For the jack-knife
analysis, we divided the sample into $7 \times 7$ angular regions
using constant boundaries in $\mathrm{RA}$ and $\mathrm{Dec}$, such
that each region contained the same angular area $10.5$ deg$^2$. In
Fig.\ \ref{figdelsigerrmice} we display the comparison as a ratio
between the jack-knife or realisations error, and the analytical
error.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{delsigerr_mice.png}
\caption{Comparison of different estimates of the error in
$\Delta\Sigma(R)$ for individual 516 deg$^2$ realisations of the
MICE mocks, between all pairs of lens spectroscopic redshift slices
(rows) and source tomographic samples (columns). The black solid
line shows the ratio between the error determined by a jack-knife
analysis of the data and the error derived from the diagonal
elements of the analytical covariance matrix, and the red dashed
line is the ratio between the standard deviation across 10
realisations and the analytical error. No measurements are possible
for the lower left-hand set of panels, owing to the adopted cut in
source-lens pairs, $z_{\mathrm{B}} > z_{\mathrm{l}}$.}
\label{figdelsigerrmice}
\end{figure*}
We find that in the range $R > 1 \, h^{-1}$ Mpc, where the model
provides a reasonable description of the measurements, the average
(fractional) absolute difference between the analytical and jack-knife
errors is $15\%$, and between the analytical and realisation scatter
is $21\%$ (which is the expected level of difference given the error
in the variance for 10 realisations). Small differences between these
error estimates may arise due to the Gaussian approximation in the
analytical covariance, the exact details of the survey modelling, or
the scale of the jack-knife regions.
Fig.\ \ref{figdelsigcovfullmice} displays the full analytical
covariance matrix of $\Delta\Sigma(R)$ -- spanning five lens redshift
slices, five source tomographic samples and 15 bins of scale -- as a
correlation matrix with $375 \times 375$ entries. We note that there
are significant off-diagonal correlations between measurements
utilising the same lens or source sample, and between different
scales. The covariance matrix is reduced in dimension if source-lens
sample pairs with $z_{\mathrm{B}} < z_{\mathrm{l}}$ are excluded, as
illustrated by the missing panels in Figs.\ \ref{figdsigtom}.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{delsigcovfull_mice.png}}
\caption{Full analytical covariance matrix $C_{ij}$ of the projected
mass density $\Delta\Sigma(R)$ for the MICE mocks, spanning five
lens redshift slices, five source tomographic samples and 15 bins of
scale -- i.e.\ $375 \times 375$ entries -- ordered where each
subsequent bin changes separation, then source sample, then lens
sample, such that the vertical and horizontal solid lines demarcate
different lens redshift slices. We display the results as a
correlation matrix $r = C_{ij}/\sqrt{C_{ii} \, C_{jj}}$, and note
that there are significant off-diagonal correlations between
measurements utilising the same lens or source sample, and between
different scales. We note that the colour bar is saturated at
$r=0.1$ to reveal low-amplitude cross-correlation more clearly.}
\label{figdelsigcovfullmice}
\end{figure}
We combined the correlated $\Delta\Sigma$ measurements for each
individual lens redshift slice, averaging over the five different
source tomographic samples, using the procedure described in Appendix
\ref{seccombap}. The resulting combined $\Delta\Sigma$ measurement
for each lens redshift sample (again corresponding to a mock mean) is
shown as the first row in Fig.\ \ref{figstatmice}.
We then used the $\Delta\Sigma(R)$ and $w_{\mathrm{p}}(R)$
measurements to infer the Upsilon statistics,
$\Upsilon_{\mathrm{gm}}(R,R_0)$ and $\Upsilon_{\mathrm{gg}}(R,R_0)$,
using Eqs.\ \ref{equpsgm} and \ref{equpsgg} respectively, adopting a
fiducial value $R_0 = 2 \, h^{-1}$ Mpc (we consider the effect of
varying this choice below). These measurements are shown in the
second and fourth rows of Fig.\ \ref{figstatmice}. We determined the
covariance of $\Upsilon_{\mathrm{gm}}(R,R_0)$ and
$\Upsilon_{\mathrm{gg}}(R,R_0)$ using error propagation following
Eqs.\ \ref{equpsgmcov} and \ref{equpsggcov}, respectively.
Finally, we determined the $E_{\mathrm{G}}(R)$ statistic for each lens
redshift slice using Eq.\ \ref{eqeg} where, for the purposes of these
tests focussed on galaxy-galaxy lensing, we assumed a fixed input
value for the redshift-space distortion parameter $\beta =
f(z)/b_{\mathrm{L}}(z)$, where we evaluated $f(z) =
\Omega_{\mathrm{m}}(z)^{0.55}$ using the fiducial cosmology of the
MICE simulation -- we note that the exponent $0.55$ is an excellent
approximation to the solution of the differential growth equation in
$\Lambda$CDM cosmologies \citep{Linder05} -- and $b_{\mathrm{L}}(z)$
is the best-fitting linear bias parameter to the
$\Upsilon_{\mathrm{gg}}$ measurements for each lens redshift slice
$z$. Hence, systematic errors associated with redshift-space
distortions lie beyond the scope of this study, and in our subsequent
data analysis we will infer the required $\beta$ values from existing
literature. We propagated errors in $E_{\mathrm{G}}$ using
Eq.\ \ref{eqegcov} (and assuming no error in $\beta$ in the case of
the mocks). Our $E_{\mathrm{G}}$ measurements are shown as the fifth
row in Fig.\ \ref{figstatmice}.
We generated fiducial cosmological models for these statistics using a
non-linear matter power spectrum $P(k,z)$ corresponding to the
fiducial cosmological parameters of the MICE simulation listed in
Sect.\ \ref{secmocks}. We determined the best-fitting linear and
non-linear galaxy bias parameters $(b_{\mathrm{L}}, b_{\mathrm{NL}})$
by fitting to the $\Upsilon_{\mathrm{gg}}$ measurements for each lens
redshift slice for scales $R > 5 \, h^{-1}$ Mpc, and applied these
same bias parameters to the galaxy-galaxy lensing models. The models
plotted in Figs.\ \ref{figgttom}, \ref{figdsigtom} and
\ref{figstatmice} do not otherwise contain any free parameters. In
Fig.\ \ref{figstatmice} we display corresponding $\chi^2$ statistics
between the models and mock mean data, demonstrating a satisfactory
goodness-of-fit in general. We evaluated the $\chi^2$ statistics for
$R > 5 \, h^{-1}$ Mpc for $\Delta\Sigma$, $w_{\mathrm{p}}$ and
$\Upsilon_{\mathrm{gg}}$, and using all scales for
$\Upsilon_{\mathrm{gm}}$ and $E_{\mathrm{G}}$. We conclude that our
lensing and clustering measurements from the MICE mocks generally
agree with the underlying $\Lambda$CDM cosmology, which we further
explore via cosmological parameter fitting in
Sect.\ \ref{secmicecosmo}.
\subsection{Photo-$z$ dilution correction}
\label{secmicephotoz}
Within our mock analysis we considered three different implementations
of the photo-$z$ dilution correction necessary for the
$\Delta\Sigma(R)$ measurements, as described in
Sect.\ \ref{secphotoz}. Firstly, we used the source spectroscopic
redshift values (which are available given that this is a simulation)
to produce a baseline $\Delta\Sigma$ measurement free of photo-$z$
dilution. Secondly, for our fiducial analysis choice, we used the
source photometric redshift point values in the estimator of
Eq.\ \ref{eqdsigest2}, adopting a source-lens pair cut $z_{\mathrm{B}}
> z_{\mathrm{l}}$ and correcting for the photo-$z$ dilution using the
$f_{\mathrm{bias}}$ factor of Eq.\ \ref{eqfbias}. We also considered
the same case, excluding the $f_{\mathrm{bias}}$ correction factor.
Thirdly, we used the redshift probability distributions for each
source tomographic sample to determine
$\overline{\Sigma_{\mathrm{c}}^{-1}}$ relative to each lens redshift
using Eq.\ \ref{eqavesigc}, and then estimated $\Delta\Sigma$ using
Eq.\ \ref{eqdsigest3}. We refer to this as the $P(z)$
distribution-based method.
The results of these $\Delta\Sigma$ analyses are compared in
Fig.\ \ref{figmicephotoz} for each lens redshift slice, where
measurements corresponding to the different source tomographic samples
have been optimally combined. We find that, other than in the case
where the $f_{\mathrm{bias}}$ correction is excluded, both the
point-based and distribution-based photo-$z$ dilution corrections
produce $\Delta\Sigma$ measurements which are statistically consistent
with the baseline measurements using the source spectroscopic
redshifts. We further verify in Sect.\ \ref{secmicecosmo} that these
analysis choices do not create significant differences in cosmological
parameter fits.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{delsig_photz_mice.png}
\caption{Measurements of the projected mass density $\Delta\Sigma$ for
each lens redshift slice of the MICE mocks, where measurements
corresponding to the different source tomographic samples have been
optimally combined following the procedure described in Appendix
\ref{seccombap}. Results are shown for four cases: using the
spectroscopic redshifts of the sources (black points, fiducial
measurement with perfect source redshifts), using the photometric
redshifts of the sources but without a correction for the photo-$z$
dilution factor $f_{\mathrm{bias}}$ (red points, photo-$z$ dilution
remains uncorrected), including the dilution correction (green
points), and using the redshift probability distribution for each
source tomographic slice (blue points). We only include individual
source-lens pairs with $z_{\mathrm{B}} > z_{\mathrm{l}}$ in the
measurement. We find that, other than for the case where the
$f_{\mathrm{bias}}$ correction is excluded, both the point-based and
distribution-based photo-$z$ dilution corrections produce
$\Delta\Sigma$ measurements which are consistent with those obtained
using the source spectroscopic redshifts.}
\label{figmicephotoz}
\end{figure*}
\subsection{Recovery of cosmological parameters}
\label{secmicecosmo}
Finally, we verified that our analysis methodology recovered the
fiducial cosmological parameters of the MICE simulation within an
acceptable statistical accuracy. In this study we focus only on the
amplitudes of the clustering and lensing statistics, keeping all other
cosmological parameters fixed. In particular we test the recovery of
the $E_{\mathrm{G}}$ statistics, and the recovery of the $\sigma_8$
normalisation, marginalising over galaxy bias parameters.
First, we determined a scale-independent $E_{\mathrm{G}}$ value (which
we denote $\langle E_{\mathrm{G}} \rangle$) for each lens redshift
slice from the MICE mock mean statistics displayed in
Fig.\ \ref{figstatmice}. We considered two approaches to this
determination. In one approach, we fitted a constant value to the
$E_{\mathrm{G}}(R)$ measurements shown in the fifth row of
Fig.\ \ref{figstatmice} using the corresponding analytical covariance
matrix, that is, varying a vector of five parameters,
\begin{equation}
\vec{p} = \left[ E_G(z_1) , E_G(z_2) , E_G(z_3) , E_G(z_4) ,
E_G(z_5) \right] .
\end{equation}
This approach has the disadvantage that it is based on the ratio of
two noisy quantities $\Upsilon_{\mathrm{gm}}/\Upsilon_{\mathrm{gg}}$,
which may result in a biased or non-Gaussian result.
Our second approach avoided this issue by including $\langle
E_{\mathrm{G}} \rangle$ as an additional parameter in a joint fit to
the $\Upsilon_{\mathrm{gm}}$ and $\Upsilon_{\mathrm{gg}}$ statistics
for each lens redshift slice, where $\langle E_{\mathrm{G}} \rangle$
changed the amplitude of $\Upsilon_{\mathrm{gm}}$ relative to
$\Upsilon_{\mathrm{gg}}$. Specifically, we fitted the model,
\begin{equation}
\begin{split}
\Upsilon_{\mathrm{gm}}(R) &= A_{\mathrm{E}} \, b_{\mathrm{L}} \,
\Upsilon_{\mathrm{gm}}(R,\sigma_8=0.8,b_{\mathrm{L}},b_{\mathrm{NL}})
\\ \Upsilon_{\mathrm{gg}}(R) &=
\Upsilon_{\mathrm{gg}}(R,\sigma_8=0.8,b_{\mathrm{L}},b_{\mathrm{NL}})
,
\end{split}
\label{eqaefit}
\end{equation}
in terms of an amplitude parameter $A_{\mathrm{E}}$ and galaxy bias
parameters $b_{\mathrm{L}}$ and $b_{\mathrm{NL}}$, and then determined
$\langle E_{\mathrm{G}} \rangle = A_{\mathrm{E}} \, b_{\mathrm{L}} \,
E_{G,\mathrm{fid}}$ for each lens redshift slice, where
$E_{G,\mathrm{fid}}(z) = \Omega_{\mathrm{m}}/f(z)$ in terms of the
fiducial matter density parameter of the MICE mocks,
$\Omega_{\mathrm{m}}$, and the theoretical growth rate of structure
based on this matter density, $f(z)$. Hence we vary a vector of 15
parameters,
\begin{equation}
\vec{p} = \left[ A_{\mathrm{E},1}, b_{\mathrm{L,1}},
b_{\mathrm{NL,1}}, ..., A_{\mathrm{E},5}, b_{\mathrm{L,5}},
b_{\mathrm{NL,5}} \right] .
\end{equation}
We note that the $b_{\mathrm{L}}$ factor in Eq.\ \ref{eqaefit} for
$\Upsilon_{\mathrm{gm}}$ arises as a consequence of our treatment of
$\beta$ as a fixed input parameter as described in
Sect.\ \ref{secmicemeas}, and ensures that $A_{\mathrm{E}}$ is
constrained only by the relative ratio
$\Upsilon_{\mathrm{gm}}/\Upsilon_{\mathrm{gg}}$, and not the absolute
amplitude of these functions.
Fig.\ \ref{figmiceegvsz} displays the different determinations of
$\langle E_{\mathrm{G}} \rangle$ in each lens redshift slice. The
left panel compares measurements using the four different treatments
of photo-$z$ dilution shown in Fig.\ \ref{figmicephotoz}, confirming
that these methods produce consistent $E_{\mathrm{G}}$ determinations
(other than the case in which $f_{\mathrm{bias}}$ is excluded; our
fiducial choice is the direct photo-$z$ pair counts with
$z_{\mathrm{B}} > z_{\mathrm{l}}$). The middle panel compares
$\langle E_{\mathrm{G}} \rangle$ fits varying the small-scale
parameter $R_0$ (where our fiducial choice is $R_0 = 2.0 \, h^{-1}$
Mpc, and we also considered choices corresponding to the adjacent
separation bins $1.2$ and $3.1 \, h^{-1}$ Mpc). The right panel
alters the method used to determine $\langle E_{\mathrm{G}} \rangle$,
comparing the default choice using the non-linear bias model, a linear
model where we fix $b_{\mathrm{NL}} = 0$, and a direct fit to the
scale-dependent $E_{\mathrm{G}}(R)$ values. Reassuringly, all these
methods yielded very similar results.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{eg_vs_z_mice.png}
\caption{Comparison of the scale-independent values $\langle
E_{\mathrm{G}} \rangle$ determined for each lens redshift slice of
the MICE mocks, varying the galaxy-galaxy lensing analysis
assumptions and methodology. Our fiducial analysis adopted a
point-based photo-$z$ correction with $z_{\mathrm{B}} >
z_{\mathrm{l}}$, $R_0 = 2.0 \, h^{-1}$ Mpc and a model fitted to
$(\Upsilon_{\mathrm{gm}}, \Upsilon_{\mathrm{gg}})$ including
non-linear galaxy bias. The left panel compares determinations of
$\langle E_{\mathrm{G}} \rangle$ varying the photo-$z$ dilution
correction method studying the same four cases described in
Fig.\ \ref{figmicephotoz}, the middle panel varies the small-scale
parameter $R_0$, and the right panel alters the fitting method to
only include linear galaxy bias, and to use a direct fit to the
$E_{\mathrm{G}}(R)$ values. The model line in each case is the
prediction $E_{\mathrm{G}}(z) = \Omega_{\mathrm{m}}/f(z)$, where
$\Omega_{\mathrm{m}}$ is the fiducial matter density parameter for
the MICE mocks.}
\label{figmiceegvsz}
\end{figure*}
We compared these determinations to the model prediction
$E_{\mathrm{G}}(z) = \Omega_{\mathrm{m}}/f(z)$ shown in
Fig.\ \ref{figmiceegvsz}. Other than for the case where the
$f_{\mathrm{bias}}$ correction is excluded, both the point-based and
distribution-based photo-$z$ dilution corrections produce
determinations of $\langle E_{\mathrm{G}} \rangle$ which recover the
fiducial value. This conclusion holds independently of the chosen
value of $R_0$, although higher $R_0$ values produce slightly
increased error ranges. The different modelling approaches also
produce consistent results.
Next, we utilised our mock dataset to perform a fit of the
cosmological parameter $\sigma_8$ to the joint lensing and clustering
statistics, marginalising over different bias parameters
$(b_{\mathrm{L}}, b_{\mathrm{NL}})$ for each redshift slice such that
we vary a vector of 11 parameters,
\begin{equation}
\vec{p} = \left[ \sigma_8, b_{\mathrm{L},1}, b_{\mathrm{NL},1}, ...,
b_{\mathrm{L},5}, b_{\mathrm{NL},5} \right] .
\end{equation}
We fixed the remaining cosmological parameters, and performed our
parameter fit using a Markov Chain Monte Carlo method implemented
using the {\tt emcee} package \citep{ForemanMackey13}. We used wide,
uniform priors for each fitted parameter.
As above, we adopted for our fiducial analysis the point photo-$z$
dilution correction using $f_{\mathrm{bias}}$, and we performed fits
to the $\Upsilon_{\mathrm{gm}}(R)$ and $\Upsilon_{\mathrm{gg}}(R)$
statistics with $R_0 = 2 \, h^{-1}$ Mpc, considering the same analysis
variations as above. Following the scale cuts mentioned above, the
data vector contains eight scales for $\Upsilon_{\mathrm{gm}}(R)$ and
seven scales for $\Upsilon_{\mathrm{gg}}(R)$ for each of the five lens
redshift slices, comprising a total of 75 data points.
For this fiducial case, we obtained a measurement $\sigma_8 = 0.779
\pm 0.019$, consistent with the MICE simulation cosmology $\sigma_8 =
0.8$. The $\chi^2$ statistic of the best-fitting model is $69.9$ for
64 degrees of freedom (d.o.f.), that is, 75 data points minus the 11
fitted parameters. Fig.\ \ref{figmicesig8} displays the dependence of
the $\sigma_8$ measurements on the analysis choices. All
methodologies using the non-linear bias model recovered the fiducial
$\sigma_8$ value, with the exception of excluding the
$f_{\mathrm{bias}}$ correction. Adopting a linear bias model instead
produced a significantly poorer recovery.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{sig8fit_mice.png}}
\caption{Measurements of the $\sigma_8$ parameter marginalising over
the galaxy bias factors, resulting from fits to the
$\Upsilon_{\mathrm{gm}}$ and $\Upsilon_{\mathrm{gg}}$ statistics
measured for the MICE lens redshift slices. We adopted a
2-parameter galaxy bias model $(b_{\mathrm{lin}}, b_{\mathrm{NL}})$
for each redshift slice, such that the fit varied 11 parameters in
total. The left-hand point resulted from our fiducial analysis
choice, assuming $R_0 = 2.0 \, h^{-1}$ Mpc and a point-based
photo-$z$ dilution correction. The second point from the left
compares the result of a fit using only the linear galaxy bias
parameter (setting $b_{\mathrm{NL}} = 0$). We also show results
varying the value of $R_0$ and the photo-$z$ dilution correction.
The horizontal red line indicates the MICE simulation fiducial
cosmology $\sigma_8 = 0.8$.}
\label{figmicesig8}
\end{figure}
We also considered fitting to different pairs of lensing-clustering
statistics: $\Delta\Sigma(R)$ and $w_{\mathrm{p}}(R)$ for $R >
R_{\mathrm{min}} = 5 \, h^{-1}$ Mpc, compared to
$\gamma_{\mathrm{t}}(\theta)$ and $w_{\mathrm{p}}(R)$, where we
applied a minimum-scale cut in $\theta$ which matches
$R_{\mathrm{min}}$ in each lens redshift slice. These alternative
statistics also successfully recovered the fiducial value of
$\sigma_8$, with errors of $0.018$ (for $\Delta\Sigma$) and $0.022$
(for $\gamma_{\mathrm{t}}$). According to this analysis, the
projected statistics produced a $\sim 20\%$ more accurate $\sigma_8$
value than the angular statistics, in agreement with the results of
\citet{Shirasaki18}.
We conclude this section by noting that the application of our
analysis pipeline to the MICE lens and source mocks successfully
recovered the fiducial $E_{\mathrm{G}}$ and $\sigma_8$ parameters of
the simulation, and is robust against differences in photo-$z$
dilution correction, choice of the small-scale parameter $R_0$, and
choice of statistic included in the analysis
[$\gamma_{\mathrm{t}}(\theta)$, $\Delta \Sigma(R)$ or
$\Upsilon_{\mathrm{gm}}(R)$].
\section{Results}
\label{secdatatests}
\subsection{Measurements}
We now summarise the galaxy-galaxy lensing and clustering measurements
we generated from the KiDS-1000 and overlapping LRG datasets. We cut
these catalogues to produce overlapping subsets for our galaxy-galaxy
lensing analysis, by only retaining sources and lenses within the set
of KiDS pointings which contain BOSS or 2dFLenS galaxies. The
resulting KiDS-N sample comprised $15 \, 150 \, 250$ KiDS shapes and
$47 \, 332$ BOSS lenses within 474 KiDS pointings with total unmasked
area $366.0$ deg$^2$, and the KiDS-S sample consisted of $16 \, 994 \,
252$ KiDS shapes and $18 \, 903$ 2dFLenS lenses within 478 KiDS
pointings with total unmasked area $382.1$ deg$^2$. We also utilised
BOSS and 2dFLenS random catalogues in our analysis, with the same
selection cuts and size 50 times bigger than the datasets, sub-sampled
from the master random catalogues provided by \citet{Reid16} and
\citet{Blake16b}, respectively.
We split the KiDS-1000 source catalogue into five different
tomographic samples by the value of the BPZ photometric redshift,
using the same bin divisions $z_{\mathrm{B}} = [0.1, 0.3, 0.5, 0.7,
0.9, 1.2]$ adopted in Sect.\ \ref{secmocktests}. The effective
source density of each tomographic sample is $n_{\mathrm{eff}} =
[0.88, 1.33, 2.04, 1.49, 1.26]$ arcmin$^{-2}$ \citep{Hildebrandt20},
estimated using the method of \citet{Heymans12}. We divided the BOSS
and 2dFLenS LRG catalogues into five spectroscopic redshift slices of
width $\Delta z_{\mathrm{l}} = 0.1$ in the range $0.2 < z_{\mathrm{l}}
< 0.7$.
We measured the average tangential shear $\gamma_{\mathrm{t}}(\theta)$
and projected mass density $\Delta\Sigma(R)$ between all pairs of
KiDS-1000 tomographic source samples and LRG redshift slices in the
north and south regions, using the same estimators and binning as
utilised for the MICE mocks in Sect.\ \ref{secmicemeas} and applying a
multiplicative shear bias correction for each tomographic sample
\citep{Kannawadi19}. For the $\Delta\Sigma$ measurement, we again
restricted the source-lens pairs such that $z_{\mathrm{B}} >
z_{\mathrm{l}}$, and (in our fiducial analysis) applied a point-based
photo-$z$ dilution correction.
We generated an analytical covariance matrix for each measurement,
initially using a fiducial lens linear bias factor $b_{\mathrm{L}} =
2$, and iterating this value following a preliminary fit to the
projected lens clustering. We tested that the analytical error
determination agreed sufficiently well with a jack-knife error
analysis where the regions were defined as the KiDS pointings; the
results of this test and the overall analytical covariance are
visually similar to their equivalents for the MICE mocks shown in
Figs.\ \ref{figdelsigerrmice} and \ref{figdelsigcovfullmice}, and we
do not repeat these figures. We used the KV450 spectroscopic
calibration sample with DIR weights \citep{Hildebrandt20} to estimate
the redshift distribution of each tomographic source sample for use in
the analytical covariance matrix, in the modelling of
$\gamma_{\mathrm{t}}(\theta)$ and in the distribution-based correction
to $\Delta\Sigma(R)$ for photo-$z$ dilution, and to determine the
$f_{\mathrm{bias}}$ values for the point-based photo-$z$ dilution
correction.
We propagated the uncertainties in the multiplicative correction
factors due to the shear calibration bias and photometric redshift
dilution using the method described in Sect.\ \ref{seccovprop}.
Regarding the multiplicative shear calibration, we followed
\citet{Hildebrandt20} in adopting an error $\sigma_{\mathrm{m}} =
0.02$ that is fully correlated across all samples, such that
$\mathrm{Cov}[ \alpha_{ij}, \alpha_{lm} ] = \sigma_{\mathrm{m}}^2$ in
Eq.\ \ref{eqcovprop2}.
The sample variance in the spectroscopic training set can be
characterised by an uncertainty in mean spectroscopic redshift which
varies for each tomographic sample in the range $\sigma_{\mathrm{z}} =
0.011 \rightarrow 0.039$ \citep[see][Table 2]{Hildebrandt20}. We
propagated these errors into the determination of $f_{\mathrm{bias}}$
by re-evaluating Eq.\ \ref{eqfbias} shifting all the spectroscopic
redshifts by a small amount to determine the derivatives $\partial
f_{\mathrm{bias},ij}/\partial z_j$, where $i$ denotes the lens sample
and $j$ the source sample. Using error propagation, we then scaled
the derivatives by the errors $\sigma_{\mathrm{z},j}$ to find the
covariance matrix of the uncertainties,
\begin{equation}
\mathrm{Cov}[ f_{\mathrm{bias},ij} \, f_{\mathrm{bias},lm} ] =
\frac{\partial f_{\mathrm{bias},ij}}{\partial z_j} \, \frac{\partial
f_{\mathrm{bias},lm}}{\partial z_m} \, \sigma^2_{\mathrm{z},j} \,
\delta^{\mathrm{K}}_{jm} ,
\end{equation}
where the final Kronecker delta $\delta^{\mathrm{K}}_{jm}$ indicates
that these uncertainities are correlated for different lens samples
corresponding to the same source sample, but uncorrelated between
source samples (we refer the reader to Joachimi et al.\ (in prep.) for
further investigation of this point). This uncertainty can be
propagated into the analytical covariance matrix using
Eqs.\ \ref{eqcovprop} and \ref{eqcovprop2} with $\mathrm{Cov}[
\alpha_{ij}, \alpha_{lm} ] = \mathrm{Cov}[ f_{\mathrm{bias},ij} \,
f_{\mathrm{bias},lm} ]$.
We used the analytical covariance matrices to combine the separate
KiDS-N and KiDS-S measurements into a single joint estimate of the
galaxy-galaxy lensing statistics and associated covariance, which we
utilised in the remainder of this study (we test the consistency of
the individual BOSS and 2dFLenS results in Sect.\ \ref{seckidssys}).
We display the KiDS $\gamma_{\mathrm{t}}(\theta)$ and
$\Delta\Sigma(R)$ galaxy-galaxy lensing measurements in the different
tomographic combinations in Figs.\ \ref{figgttom} and
\ref{figdsigtom}. We note again that there are some differences
between the galaxy-galaxy lensing signals measured in the mocks and
data, given that the mocks have not been tuned to reproduce the BOSS
and 2dFLenS clustering properties. These differences are particularly
evident on the smallest scales, owing to an inconsistent halo
occupation. Our study does not require the mocks to precisely
replicate the data in order to test our analysis framework.
We obtained the most accurate measurement of the projected correlation
function $w_{\mathrm{p}}(R)$ of each lens redshift slice using the
full BOSS DR12 dataset, combining the LOWZ and CMASS selections and
spanning $9 \, 376$ deg$^2$ \citep{Reid16}. We adopted the same
spatial separation bins as for the MICE mocks, again assuming
$\Pi_{\mathrm{max}} = 100 \, h^{-1}$ Mpc. When analysing the BOSS
sample we included completeness weights but excluded `FKP' weights
\citep{Feldman94}, which are designed to optimise the clustering
signal-to-noise ratio but may not be appropriate in the case of
galaxy-galaxy lensing. The full 2dFLenS dataset is too small to offer
a competitive measurement of $w_{\mathrm{p}}(R)$, although given that
it was selected using BOSS-inspired colour-magnitude cuts, we assumed
that the BOSS clustering is representative of the combined LRG sample
(and we test this approximation in Sect.\ \ref{seckidssys}). Since
the overlap of the KiDS-N source catalogue and full BOSS sample is
small ($4\%$ of BOSS), we also assumed that the galaxy-galaxy lensing
and clustering measurements are uncorrelated.
We combined the correlated $\Delta\Sigma$ measurements for each lens
redshift slice, averaging over the different source samples.
Fig.\ \ref{figstatkids} displays these measurements, together with the
projected clustering $w_{\mathrm{p}}(R)$ of the full BOSS sample, the
corresponding $\Upsilon_{\mathrm{gm}}(R,R_0)$ and
$\Upsilon_{\mathrm{gg}}(R,R_0)$ statistics assuming a fiducial choice
$R_0 = 2 \, h^{-1}$ Mpc (we consider the impact of varying this choice
in Sect.\ \ref{seckidssys}), and the direct $E_{\mathrm{G}}(R)$
estimate using Eq.\ \ref{eqeg}.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{allstat_kids.png}
\caption{Measurements of a series of galaxy-galaxy lensing and
clustering statistics (rows) for each different LRG redshift slice
(columns) correlated with the KiDS-1000 source sample, displayed in
the same style as Fig.\ \ref{figstatmice}. We note that the first
and third rows -- $\Delta\Sigma(R)$ and $w_{\mathrm{p}}(R)$ --
represent the original galaxy-galaxy lensing and projected
clustering measurements, and the remaining rows --
$\Upsilon_{\mathrm{gm}}(R)$, $\Upsilon_{\mathrm{gg}}(R)$ and
$E_{\mathrm{G}}(R)$ -- represent statistics derived from these
original measurements.}
\label{figstatkids}
\end{figure*}
We generated fiducial cosmological models for these statistics using a
non-linear matter power spectrum corresponding to the best-fitting
`TTTEEE+lowE+lensing' {\it Planck} cosmological parameters
\citep{Planck18}. When producing the models overplotted in
Figs.\ \ref{figgttom}, \ref{figdsigtom} and \ref{figstatkids}, we
determined best-fitting linear and non-linear galaxy bias parameters
by fitting to the $\Upsilon_{\mathrm{gg}}$ measurements for each lens
redshift slice for the separation range $R > 5 \, h^{-1}$ Mpc, and
applied these same bias parameters to the galaxy-galaxy lensing
models. Values of the $\chi^2$ statistic for each statistic and lens
redshift slice, produced using these models and the analytical
covariance, are displayed in each panel of Fig.\ \ref{figstatkids},
and indicate that the measurements are consistent with the model.
\subsection{Redshift-space distortion inputs}
We adopted values of the redshift-space distortion parameters $\beta$
for the BOSS sample as a function of redshift by interpolating the
literature analysis of \citet{Zheng19}, who provide RSD measurements
in narrow redshift slices \citep[which also agree with the compilation
of results in][]{Alam17b}. In order to interpolate these
measurements to our redshift locations, which slightly differ from the
bin centres of \citet{Zheng19}, we created a Gaussian process model
for $\beta(z)$ and its errors, which lie in the range $12-20\%$ as a
function of redshift, using the sum of a Matern kernel and white noise
kernel. We note that the error in $\beta$ makes up roughly half the
variance budget for $E_{\mathrm{G}}$ in the lowest lens redshift slice
(i.e.\ increases the total error by $\sim \sqrt{2}$), but is
subdominant for the other redshift slices.
\subsection{Amplitude-ratio test $E_{\mathrm{G}}$}
\label{secegkids}
We used the KiDS-1000 and LRG clustering and galaxy-galaxy lensing
measurements, with the previously-published values of $\beta$, to
determine a scale-independent value of the amplitude ratio, $\langle
E_{\mathrm{G}} \rangle$. We adopted the same fiducial analysis method
as for the MICE mocks: we performed a joint fit to the
$\Upsilon_{\mathrm{gm}}$ and $\Upsilon_{\mathrm{gg}}$ measurements for
each redshift slice, varying $A_{\mathrm{E}}$ and the bias parameters
$b_{\mathrm{L}}$ and $b_{\mathrm{NL}}$ as in Eq.\ \ref{eqaefit},
treating $A_{\mathrm{E}}$ as an additional amplitude parameter for
$\Upsilon_{\mathrm{gm}}$. We then deduced $\langle E_{\mathrm{G}}
\rangle = A_{\mathrm{E}}/\beta$, propagating the errors in $\beta$
assuming all these statistics are independent (please see
Sect.\ \ref{seceg} for a note on this approximation). We used the
analytical covariance matrices for these statistics, assumed $R_0 = 2
\, h^{-1}$ Mpc, and fitted the model to $R > 5 \, h^{-1}$ Mpc for
$\Upsilon_{\mathrm{gg}}$ and to all scales for
$\Upsilon_{\mathrm{gm}}$. We consider the effect of varying these
analysis choices in Sect.\ \ref{seckidssys}. In particular, we note
that fitting the directly-determined $E_{\mathrm{G}}(R)$ values (shown
in the fifth row of Fig.\ \ref{figstatkids}) produced results which
were entirely consistent with our fiducial analysis.
Our resulting fits for $\langle E_{\mathrm{G}} \rangle$ were $[0.43
\pm 0.09, 0.45 \pm 0.07, 0.33 \pm 0.06, 0.38 \pm 0.07, 0.34 \pm
0.08]$ for redshifts $z = [0.25, 0.35, 0.45, 0.55, 0.65]$. The
measurements have a small degree of correlation, owing to sharing a
common source sample, and the analytical covariance matrix is listed
in Table \ref{tabegcov}. We plot these measurements in
Fig.\ \ref{figegall}, together with a literature compilation
\citep{Reyes10,Blake16a,Pullen16,Alam17a,delaTorre17,Amon18,Singh19,Jullo19}.
The thickness of the purple shaded stripe in Figs.\ \ref{figegall} and
\ref{figegscale} illustrates the $68\%$ confidence range of the
prediction of the {\it Planck} `TTTEEE+lowE+lensing' parameter chain
at each redshift, assuming a flat $\Lambda$CDM Universe.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{egall.png}
\caption{Compilation of measurements of a scale-independent $\langle
E_{\mathrm{G}} \rangle$ as a function of redshift. The solid black
data points illustrate the results of this study, in five redshift
bins. These are compared to literature values from: \citet{Reyes10}
(green), \citet{Blake16a} (red), \citet{Pullen16} (magenta),
\citet{Alam17a} (orange), \citet{delaTorre17} (purple),
\citet{Amon18} (blue), \citet{Singh19} (cyan) and \citet{Jullo19}
(brown). We note that these previous measurements typically
correspond to significantly wider lens redshift ranges than our
study, and are also correlated in some cases owing to overlapping
datasets. The thickness of the purple shaded stripe illustrates the
$68\%$ confidence range of the prediction of the {\it Planck}
`TTTEEE+lowE+lensing' parameter chain at each redshift assuming a
flat $\Lambda$CDM Universe, and the dashed lines are the predictions
of the model $E_{\mathrm{G}}(z) = \Omega_{\mathrm{m}}/f(z)$ for
$\Omega_{\mathrm{m}} = [0.2, 0.3, 0.4]$.}
\label{figegall}
\end{figure*}
\begin{table}
\caption{The covariance matrix corresponding to our measurements
$E_{\mathrm{G}} = [0.43, 0.45, 0.33, 0.38, 0.34]$ at $z = [0.25,
0.35, 0.45, 0.55, 0.65]$. Each entry has been multiplied by
$10^4$ for clarity of display. The cross-correlation between
different redshift slices is small.}
\label{tabegcov}
\centering
\begin{tabular}{cccccc}
\hline
Redshift & $0.25$ & $0.35$ & $0.45$ & $0.55$ & $0.65$ \\
\hline
$0.25$ & $76.821$ & $3.477$ & $3.251$ & $3.330$ & $3.178$ \\
$0.35$ & $3.477$ & $52.101$ & $1.559$ & $1.633$ & $1.509$ \\
$0.45$ & $3.251$ & $1.559$ & $34.674$ & $1.088$ & $1.017$ \\
$0.55$ & $3.330$ & $1.633$ & $1.088$ & $43.179$ & $0.893$ \\
$0.65$ & $3.178$ & $1.509$ & $1.017$ & $0.893$ & $67.585$ \\
\hline
\end{tabular}
\end{table}
Our measurements provide the best existing determination of the
lensing-clustering amplitude ratio (noting that previous measurements
displayed in Figure \ref{figegall} typically correspond to
significantly wider lens redshift ranges than our study), which is
consistent with matter density values $\Omega_{\mathrm{m}} \sim 0.3$.
Varying the $\Omega_{\mathrm{m}}$ parameter within a flat $\Lambda$CDM
cosmological model, assuming $E_{\mathrm{G}} = \Omega_{\mathrm{m}}/f =
\Omega_{\mathrm{m}}(0)/\Omega_{\mathrm{m}}(z)^{0.55}$, we find
$\Omega_{\mathrm{m}} = 0.27 \pm 0.04$ (with a minimum $\chi^2 = 1.2$
for four d.o.f.). In principle in linear theory, this measurement is
insensitive to the other cosmological parameters in a flat
$\Lambda$CDM scenario. The resulting error in $\Omega_{\mathrm{m}}$
is, naturally, somewhat larger than that provided by analyses
utilising the full shape of the cosmic shear and clustering functions
\citep[e.g.][]{Troester20}, albeit requiring fewer model assumptions.
We tested for the scale dependence of the $E_{\mathrm{G}}(R)$
measurements (in the fifth row of Fig.\ \ref{figstatkids}) by jointly
fitting an empirical six-parameter model $E_{\mathrm{G}}(R,z_i) = A_i
\left[ 1 + \alpha \, \log_{10}(R) \right]$ to all the redshift slices,
where $\alpha$ quantifies the fractional variation in $E_{\mathrm{G}}$
per decade in projected scale $R$ (in $h^{-1}$ Mpc) and $A_i$ is a
free amplitude for each of the five lens redshift slices. We obtained
a $68\%$ confidence region $\alpha = 0.17 \pm 0.26$ (with a minimum
$\chi^2 = 35.9$ for 34 d.o.f.), which is consistent with no scale
dependence, as predicted in the standard gravity scenario. These fits
are displayed in Fig.\ \ref{figegscale}.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{eg_vs_scale_kids.png}
\caption{Measurements of the $E_{\mathrm{G}}(R)$ statistic for the
KiDS-1000 source sample in each LRG lens redshift slice. Errors are
derived from the diagonal elements of the analytical covariance
matrix, propagating errors where appropriate. The thickness of the
purple horizontal stripes illustrates the $68\%$ confidence range of
the prediction of the {\it Planck} `TTTEEE+lowE+lensing' parameter
chain at each redshift assuming a flat $\Lambda$CDM Universe, and
the red bands illustrate the $68\%$ confidence ranges of a
scale-dependent model $E_{\mathrm{G}}(R,z_i) = A_i \left[ 1 + \alpha
\, \log_{10}(R) \right]$ discussed in Sect.\ \ref{secegkids}.}
\label{figegscale}
\end{figure*}
\subsection{Systematics tests}
\label{seckidssys}
We now consider the effect on our cosmological fits of varying our
fiducial analysis choices. Fig.\ \ref{figkidsegvsz} is a compilation
of different determinations of $\langle E_{\mathrm{G}} \rangle$ in
each lens redshift slice. The upper-left panel compares the fits
varying the photo-$z$ dilution correction for $\Delta\Sigma$ between
the point-based and distribution-based approaches, and the upper-right
panel shows measurements varying the small-scale parameter $R_0$. The
lower-left panel considers separate determinations based on the BOSS
and 2dFLenS samples (using the BOSS clustering measurements in both
cases). The lower-right panel alters the fitting method to only use a
linear-bias model (set $b_{\mathrm{NL}} = 0$), and to use a direct fit
to the $E_{\mathrm{G}}(R)$ measurements presented in the fifth row of
Fig.\ \ref{figstatkids}, as opposed to our fiducial fits to
$\Upsilon_{\mathrm{gm}}$ and $\Upsilon_{\mathrm{gg}}$. In all cases,
the systematic variation of the recovered $E_{\mathrm{G}}$ values is
negligible compared to the statistical errors (noting that the BOSS
and 2dFLenS comparison is also subject to sample variance error). We
find that varying the photo-$z$ dilution correction, choice of $R_0$
and fitting method produce a systematic variation of $0.08$-$\sigma$,
$0.24$-$\sigma$ and $0.38$-$\sigma$ respectively, when expressed as a
fraction of the statistical error $\sigma$.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{eg_vs_z_kids.png}
\caption{Comparison of the scale-independent values $\langle
E_{\mathrm{G}} \rangle$ determined for each LRG redshift slice
correlated with the KiDS-1000 source sample, varying the
galaxy-galaxy lensing analysis assumptions, sample and methodology.
Our fiducial analysis adopted a point-based photo-$z$ correction
with $z_{\mathrm{B}} > z_{\mathrm{l}}$, $R_0 = 2.0 \, h^{-1}$ Mpc
and a model fitted to $(\Upsilon_{\mathrm{gm}},
\Upsilon_{\mathrm{gg}})$ including non-linear galaxy bias. The
upper-left panel compares determinations of $\langle E_{\mathrm{G}}
\rangle$ varying the photo-$z$ dilution correction method, the
upper-right panel varies the small-scale parameter $R_0$, the
lower-left panel varies the lens sample to consider BOSS and 2dFLenS
separately, and the lower-right panel alters the fitting method to
only include linear galaxy bias, and to use a direct fit to the
$E_{\mathrm{G}}(R)$ values. The model line in each case is the
prediction $E_{\mathrm{G}}(z) = \Omega_{\mathrm{m}}/f(z)$, where
$\Omega_{\mathrm{m}} = 0.3$.}
\label{figkidsegvsz}
\end{figure*}
\section{Summary}
\label{secsummary}
We have used the latest weak gravitational lensing data release from
the Kilo-Degree Survey, KiDS-1000, together with overlapping galaxy
redshift survey data from BOSS and 2dFLenS, to perform cosmological
tests associated with the relative amplitude of galaxy-galaxy lensing
and galaxy clustering statistics. We quantified our results using the
$E_{\mathrm{G}}(R)$ statistic, which we were able to measure up to
projected separations of $100 \, h^{-1}$ Mpc, recovering a
scale-independent value of $\langle E_{\mathrm{G}} \rangle$ with
accuracies in the range 15-20\% in five lens redshift slices of width
$\Delta z = 0.1$. The scale-dependence and redshift-dependence of
these measurements are consistent with the theoretical expectation of
general relativity in a Universe with matter density
$\Omega_{\mathrm{m}} \sim 0.3$. The measurements are consistent with
a scale-independent model for $E_{\mathrm{G}}$, and constrain allowed
variation within $25\%$ (1-$\sigma$) error per decade in projected
scale. Fitting our $E_{\mathrm{G}}$ dataset with a flat $\Lambda$CDM
model, we find $\Omega_{\mathrm{m}} = 0.27 \pm 0.04$.
We demonstrated that our results are robust against different analysis
methodologies. In particular, we showed that:
\begin{itemize}
\item Source photometric redshift errors cause a significant dilution
of the inferred projected mass density $\Delta\Sigma$, by causing
unlensed foreground sources to appear in the background of lenses.
We demonstrated that this dilution may be corrected either by
modifying the estimator for $\Delta\Sigma$ to include the redshift
probability distribution of the sources, or by using a source
spectroscopic calibration sample to compute the multiplicative bias,
producing consistent results. When applied to the mock catalogue,
these estimation methods recovered the $\Delta\Sigma$ measurement
obtained using source spectroscopic redshifts.
\item A Gaussian analytical covariance for the galaxy-galaxy lensing
statistics, with suitable modifications for the survey selection
function and small-scale noise terms, predicted errors which agreed
within $20\%$ with estimates from a jack-knife procedure and from
the variation across mock realisations.
\item Our amplitude-ratio test, based on the annular differential
density statistics $\Upsilon_{\mathrm{gm}}$ and
$\Upsilon_{\mathrm{gg}}$, is insensitive to the small-scale
parameter $R_0$ adopted in these statistics, producing consistent
results for choices in the range $1 < R_0 < 3 \, h^{-1}$ Mpc. We
also obtained consistent results analysing BOSS and 2dFLenS
separately.
\end{itemize}
We performed an additional series of tests jointly fitting an overall
amplitude $\sigma_8$ to the galaxy-galaxy lensing and clustering
statistics in our mocks, marginalising over linear and non-linear bias
parameters. For our scenario, the projected galaxy-galaxy lensing
measurements produced a slightly more accurate determination of
$\sigma_8$ than the angular statistics after matching of scales,
although the results are fully consistent.
Our analysis sets the stage for upcoming, increasingly accurate,
cosmological tests using amplitude-ratio statistics, as gravitational
lensing and galaxy redshift samples continue to grow. These datasets
will continue to offer rich possibilities for placing tighter
constraints on allowed gravitational physics.
\begin{acknowledgements}
We thank the anonymous referee for useful comments on the paper. CB
is grateful to Rossana Ruggeri, Alexie Leauthaud, Johannes Lange and
Sukhdeep Singh for valuable discussions on galaxy-galaxy lensing
measurements. \\
We acknowledge support from: the Australian Research Council Centre of
Excellence for All-sky Astrophysics (CAASTRO) through project number
CE110001020 (CB, KG); the European Research Council under grant
numbers 647112 (MA, BG, CH, TT), 770935 (AD, HH, JLvdB, AHW) and
693024 (SJ); the Polish Ministry of Science and Higher Education
through grant DIR/WK/2018/12 and the Polish National Science Center
through grant 2018/30/E/ST9/00698 (MB); the Max Planck Society and the
Alexander von Humboldt Foundation in the framework of the Max
Planck-Humboldt Research Award endowed by the Federal Ministry of
Education and Research (CH); Heisenberg grant Hi 1495/5-1 of the
Deutsche Forschungsgemeinschaft (HH); the Beecroft Trust (SJ); Vici
grant 639.043.512 financed by the Netherlands Organisation for
Scientific Research (AK); the Alexander von Humboldt Foundation (KK);
the NSFC of China grant 11973070, the Shanghai Committee of Science
and Technology grant 19ZR1466600 and Key Research Program of Frontier
Sciences grant ZDBS-LY-7013 (HYS); and the European Union’s Horizon
2020 research and innovation programme under the Marie
Sk{l}odowska-Curie grant 797794 (TT). \\
The 2dFLenS survey is based on data acquired through the Australian
Astronomical Observatory, under program A/2014B/008. It would not have
been possible without the dedicated work of the staff of the AAO in
the development and support of the 2dF-AAOmega system, and the running
of the AAT. \\
Funding for SDSS-III has been provided by the Alfred P. Sloan
Foundation, the Participating Institutions, the National Science
Foundation, and the U.S. Department of Energy Office of Science. The
SDSS-III web site is http://www.sdss3.org/. SDSS-III is managed by
the Astrophysical Research Consortium for the Participating
Institutions of the SDSS-III Collaboration including the University of
Arizona, the Brazilian Participation Group, Brookhaven National
Laboratory, Carnegie Mellon University, University of Florida, the
French Participation Group, the German Participation Group, Harvard
University, the Instituto de Astrofisica de Canarias, the Michigan
State/Notre Dame/JINA Participation Group, Johns Hopkins University,
Lawrence Berkeley National Laboratory, Max Planck Institute for
Astrophysics, Max Planck Institute for Extraterrestrial Physics, New
Mexico State University, New York University, Ohio State University,
Pennsylvania State University, University of Portsmouth, Princeton
University, the Spanish Participation Group, University of Tokyo,
University of Utah, Vanderbilt University, University of Virginia,
University of Washington, and Yale University. \\
We have used {\tt matplotlib} \citep{Hunter07} for the generation of
scientific plots, and this research also made use of {\tt astropy}, a
community-developed core Python package for Astronomy
\citep{Astropy13}. \\ \\
{\it Author contributions:} All authors contributed to the development
and writing of this paper. The authorship list is given in two groups:
the lead author (CB), followed by an alphabetical group who have
made a significant contribution to either the data products or to the
scientific analysis.
\end{acknowledgements}
\bibliographystyle{aa}
| 9adb75a517802ddb4a1d2f4a89479384b0265755 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Evaluating whether a given configuration of a robot manipulator is in a feasible position, such as for motion planning applications, typically involves a collision detector that produces a binary output: in-collision or collision-free \cite{Das2020}. However, it is often far more useful to know the distance to collision in the workspace, as closer proximities may suggest that the robot is more at risk to colliding with an obstacle. This distance to collision is useful in defining robot control rules for safe robot motion in potentially unpredictable, partially observed, or nonstationary environments \cite{Flacco2012}, such as in cases where the robots are working in close proximity with humans \cite{Schiavi2009}. In addition to collision avoidance for robots \cite{Flacco2012}, applications requiring accurate representation of the distance between two objects include contact resolution in simulation (where objects must respond to collisions realistically) \cite{Pan2013} and force feedback in haptic rendering (where forces are generated for a human manipulating a virtual or teleoperated tool) \cite{Lin2003}.
\begin{figure}[t]
\centering
\subfloat[Workspace\label{fig:distWorkspace}] {
\includegraphics[width=0.46\linewidth, trim={3cm 2cm 2.2cm 1.5cm}, clip]{images/distExampleWorkspace}}
\hfill
\subfloat[Distances in C-space\label{fig:distCspace}] {
\includegraphics[width=0.51\linewidth, trim={0cm 0cm 0cm 0cm}, clip]{images/distExampleCspace}}
\caption{(a) A 2 DOF robot and a square obstacle. (b) A C-space representation where the workspace distance to collision for each configuration is shown. The axes represent each DOF of the robot.}
\label{fig:distExample}
\end{figure}
A robot's distance to collision is typically defined to be the minimum distance between all of the robot's links and all of the workspace obstacles \cite{Flacco2012,Han2019}. Assuming that the $i^{th}$ link of a robot in configuration $\vec{x}$ is represented as ${\mathcal{R}}_i(\vec{x})$ and the $j^{th}$ workspace obstacle is $o_j$, we define $d_{\mathcal{R}}(\cdot)$ as
\begin{equation}
d_{\mathcal{R}}(\vec{x}) = \min_{i,j}dist({\mathcal{R}}_i(\vec{x}), o_j)
\end{equation}
where $dist(\cdot, \cdot)$ provides a positive scalar distance between two non-intersecting bodies. When the two bodies are in collision, $dist(\cdot, \cdot)$ provides the penetration depth (the minimum amount intersecting bodies must move to be out of collision \cite{Lin2003}) as a negative value. Fig. \ref{fig:distWorkspace} shows a 2 degrees-of-freedom (DOF) robot with an obstacle, and its corresponding configuration space (or C-space) representation in Fig. \ref{fig:distCspace}. For each point in the C-space, the workspace distance to collision is shown.
As with collision detection, distance estimation is a time-consuming operation, with time complexity as high as $\mathcal{O}(m^6n^6)$ for generalized distance calculations, where $m$ and $n$ are the number of vertices in two different bodies \cite{Pan2012}. Furthermore, reliable collision avoidance often requires the extensive use of external sensors \cite{Flacco2012}. The sensors that provide obstacle information, such as cameras or depth sensors, are a source of uncertainty in the distance measurements \cite{Kyriakopoulos1990}, and thus allow us to acquire only noisy evaluations of distance, which we represent as $\tilde{d}_\mathcal{R}(\cdot)$. Accounting for this uncertainty is a necessity in distance estimation, yet few strategies exist that consider these imperfections in distance sensing. As there are many applications that utilize distances, in this work, we seek to create a model to efficiently and accurately estimate $d_{\mathcal{R}}(\cdot)$ while accounting for potential uncertainty in the distance measurements.
\subsection{Contributions}
In this work, we describe a modeling approach to estimate a robot manipulator's distance to collision given noisy measurements. More specifically, we employ Gaussian process (GP) regression with the forward kinematics (FK) kernel, a similarity function designed for robot manipulators, to estimate $d_{\mathcal{R}}(\vec{x})$.
We demonstrate that the GP model with the FK kernel has significantly better modeling accuracy than other comparable regression techniques and show that querying the proposed GP is orders of magnitude faster than the standard geometric approach for distance evaluation. We utilize our proposed technique in trajectory optimization applications and realize a 9 times improvement in optimization time compared to a noise-free geometric approach yet with similar final results. Finally, we show how a hybrid model, which considers the confidence of the model, can switch from GP-based predictions to sensor-based distance estimation in areas of low confidence.
\subsection{Related Work}
The standard approach to calculating distances involves first using sensors to obtain the locations of obstacles, and then approximating the robot geometry and workspace obstacles to actually compute the distance, which is a considerably expensive process if there are many pairwise checks to perform \cite{Han2019,Safeea2017}. Some techniques attempt to reduce this computational effort through approximate methods, such as by using values from a depth sensor to approximate a distance or by creating a proxy machine learning model using exemplar data. This section describes some geometric, depth perception-based, and machine learning techniques.
\subsubsection{Geometric Approaches}
A standard approach to calculating distance to collision is to approximate the objects and use a geometric distance method between them. For example, Han et al. \cite{Han2019} estimate $d_\mathcal{R}(\cdot)$ by first representing each link of a robot arm as a cylinder and converting point clouds acquired from a depth camera into multiple small cubes, and then using the GJK algorithm \cite{Gilbert1988} to compute the distances. While their technique can accurately represent distances for a robot and obstacles observed with a depth camera, breaking obstacles into numerous small polyhedra may increase the potential number of pairwise checks that need to be performed, and there is no measure of uncertainty with this approach.
\subsubsection{Perception-Based Distance Fields}
Flacco et al. create a method to estimate distance to collision for robot manipulators based on depth sensors \cite{Flacco2012}. For any point in the workspace, distance equations are defined using a depth map, which is an image with depth values evaluated for each pixel.
The distances are used to create artificial repulsive forces to drive the robot away from obstacles. The authors claim that the distance values computed with their technique are satisfactory for the purpose of collision avoidance but are not to be considered as actual Cartesian distances. This limitation precludes the usage of this technique beyond the purpose of defining repulsive forces.
Furthermore, their distance estimation technique has no measure of sensor or model noise. However, as the primary purpose of their technique is to define repulsive forces away from obstacles, the authors take an average over all obstacles' repulsive fields to reduce any potential noise originating from the depth sensor.
\subsubsection{Support Vector Machines}
Pan et al. use a support vector machine (SVM) to model the collision region in C-space between two bodies. Binary collision checks are performed on random configurations of one object relative to the other to form a decision boundary. By determining the distance between a query configuration and the decision boundary, their SVM model is used to estimate distance. Since there is no closed-form solution for the distance to the boundary for a nonlinear SVM, sequential quadratic programming, a constrained nonlinear optimization technique, is used to find this distance \cite{Pan2013}. As distances in the C-space do not generally equate to distances in the workspace, a different distance metric is used (the displacement distance metric), which involves determining application-specific weighting terms \cite{Pan2013}. Additionally, this approach requires a unique SVM model between each pair of obstacles, which may make applying this technique to robot manipulators or environments with multiple obstacles more computationally expensive. Finally, this technique assumes the state of each object is perfectly known, which means there is no measure of uncertainty in this model.
\subsubsection{Neural Networks}
RelaxedIK estimates self-collision costs by using line segments to represent the links. Distances between the links are used to approximate the distances. These distance computations are further approximated using a neural network \cite{Rakita2018}. Their method successfully estimates these self-collision costs which enables usage in optimization problems. However, as with many neural network approaches, this technique requires a large amount of labeled data and a significant amount of time to train, precluding this approach from being applied to instances beyond distance to self-collision.
ClearanceNet is a promising neural network-based approach to estimate $d_{\mathcal{R}}(\cdot)$. The input to this network is the robot configuration and a parameterized representation of the environment. Their approach achieves over $90\%$ accuracy if the network's output is used to predict the collision status \cite{Kew2019}. Neural networks, however, are of fixed size and complexity, often require a large amount of data to train, and since the locations of the workspace obstacles are parameterized, generalizing the implementation to new types of environments may require a restructuring of this specific neural network implementation.
\section{Background}
\subsection{Kernel Functions}
A kernel function $k:\mathbb{R}^D \times \mathbb{R}^D\rightarrow \mathbb{R}$ is a function that produces a similarity score between two objects, such as $D$-dimensional robot configurations. A kernel function is a fundamental tool for learning-based algorithms including binary classification, probability density estimation, and dimensionality reduction \cite{Hofmann2008}. Kernel functions are often employed in nonparametric models, which are models that make no strong assumptions on the structure of the model and whose complexity scales with the data \cite{Altman1992}, which is desired for modeling distance fields given that many of the environments in which a manipulator is used is nonstationary.
A kernel matrix $\vec{K}\in\mathbb{R}^{N\times N}$ can be defined for a set of $N$ points $\mathcal{X}$ where $\vec{K}_{ij} = k(\mathcal{X}_i,\mathcal{X}_j)$. A positive definite kernel is one that yields a positive definite $\vec{K}$. Positive definiteness is an important property because it guarantees nonsingularity for the kernel matrix, represents an implicit mapping to some higher-dimensional feature space according to Mercer's theorem \cite{Minh}, and is a requirement in some machine learning algorithms including Gaussian processes \cite{Rasmussen2006}.
There are various positive definite kernel functions, and the one employed in the model should be selected to capture meaningful similarities between the data \cite{BelancheMunoz2013}. Two kernels that we will use in this paper are the Gaussian kernel and the forward kinematics kernel, which are described below.
\begin{figure}[!t]
\centering
\subfloat[Gaussian Kernel\label{fig:gaussianKernel}]{%
\includegraphics[width=0.32\linewidth, trim={0 0 0 0}, clip]{images/gaussianKernel}}
\hfill
\subfloat[FK Kernel\label{fig:fkKernel}]{%
\includegraphics[width=0.32\linewidth, trim={0 0 0 0}, clip]{images/fkKernel}}
\hfill
\subfloat[2 DOF Robot\label{fig:controlPointPlacement}]{%
{\includegraphics[width=0.32\linewidth, trim={1cm 1.1cm 1.2cm 0.2cm}, clip]{images/controlPointPlacement}}}
\caption[Different levels of confidence of the GP models for $d_{\mathcal{R}}(\cdot)$ being greater than 0 for a 2 DOF robot are shown.]{Visualizations of (a) the Gaussian kernel and (b) the FK kernel for the configuration of the 2 DOF robot specified in (c). The FK kernel is based on the control point locations marked with red circles.}
\label{fig:kernelShapes}
\end{figure}
\subsubsection{Gaussian Kernel}
A common positive definite kernel function to use in kernel-based models is the Gaussian kernel $k_G(\vec{x},\vec{x'}) = \exp\left(-\gamma\|\vec{x} - \vec{x'}\|^2\right)$, where $\gamma$ is a parameter dictating the width of the kernel \cite{Rasmussen2006}. Fig. \ref{fig:gaussianKernel} shows a topological visualization of the Gaussian kernel for the 2 DOF robot in Fig. \ref{fig:controlPointPlacement}. The Gaussian kernel is often considered the default kernel to use if there is limited prior knowledge of the underlying structure of the data \cite{Smola1998}.
\subsubsection{Forward Kinematics Kernel}
The forward kinematics (FK) kernel was designed to compare robot manipulators \cite{Das2020a}. Rather than comparing the joint angles representing the robot configuration directly, the locations of $M$ control points along the arm are compared. This avoids measuring distances directly in joint space, where proximal joint displacement has far greater effect than distal joint displacement to the Cartesian shape of the robot. The FK kernel is defined as $k_{FK}(\vec{x}, \vec{x'}) = \sum_{m=1}^{M}{\left(1+\frac{\gamma}{2}\|FK_m(\vec{x}) - FK_m(\vec{x'})\|^2\right)}^{-2}$, where $FK_m(\vec{x})$ is the location of the $m^{th}$ control point in the workspace for configuration $\vec{x}$ and $\gamma$ is a parameter dictating kernel width. Fig. \ref{fig:fkKernel} shows a topological visualization of the FK kernel for a 2 DOF robot with control points placed at the end of each link. Significant improvement to classification performance was realized with the FK kernel \cite{Das2020a} over other kernels \cite{Das2020,Das2017}, and thus the FK kernel is a worthwhile avenue to explore when creating a model for distance estimation.
\subsection{Regression}
Regression models can produce a continuous scalar output given a possibly multi-dimensional input. A common perspective of regression analysis is to view the true output as a sum of its input-dependent expected value and some error term, and using the expected value as a predictor of the output for new queries \cite{Fahrmeir2013}. In the context of distance evaluation, we can assume that the measured, sensor-based distance is a sum of the true distance and a noise term: $\tilde{d}_\mathcal{R}(\vec{x}) = d_{\mathcal{R}}(\vec{x}) + \epsilon$, where $\epsilon \sim \mathcal{N}(0,\eta^2)$ and $\eta^2$ is the measurement noise variance.
Choosing nonparametric models such as kernel regression and Gaussian processes allow for distance maps to scale to the complexity of the environment and to account for changing environments over time. This is the ideal scenario as many environments a robot must work in are nonstationary and may include other agents.
\subsubsection{Kernel Regression}
Kernel regression (KR) \cite{Nadaraya1964,Watson1964} is one example of a nonparametric regression technique, where a probability distribution is estimated and used to compute an expected value. A training set of $N$ input samples $\mathcal{X}$ and their corresponding outputs $\vec{y}$ can be used to estimate a probability distribution of the outputs conditioned on the inputs. This probability distribution may then be used to calculate the expected value of query inputs by taking a weighted sum of all training datapoints to predict the output of a query point. The equation for prediction for a single query point $\vec{x}$ is
\begin{equation}
\mu_{KR}\left(\vec{x}\right) = \frac{\sum_{i=1}^Nk(\mathcal{X}_i,\vec{x})\vec{y}_i}{\sum_{i=1}^Nk(\mathcal{X}_i,\vec{x})} \label{eq:kr}
\end{equation}
where $k(\cdot, \cdot)$ is some kernel function. KR does not need to learn any weights to estimate the expected value of $d_{\mathcal{R}}(\cdot)$, so a training procedure is not required. KR typically uses a kernel that has equal influence in each direction, such as the Gaussian kernel \cite{Rasmussen2006}. The FK kernel is not a function of the distance between its inputs and changes shape depending on where the kernel is evaluated, so it is not well suited for the KR formulation in Eq. \ref{eq:kr}
\subsubsection{Gaussian Process Regression}
\label{sec:gp}
Gaussian processes (GPs) can be used for regression problems. GPs are nonparametric kernel-based models that fit a Gaussian distribution over the functions that can potentially fit a set of training data \cite{Rasmussen2006}.
A GP is completely defined by a mean and covariance (or kernel) function. The prior distribution considers the space of all possible functions that can be defined using a prior mean (which is sometimes, but not always, 0) and a positive definite kernel function $k(\cdot, \cdot)$. A posterior distribution is defined given training data to condition the prior distribution. Assuming that $\mathcal{X}$ is a training dataset, $\vec{y}$ is a vector of corresponding training outputs, the posterior distribution for the regression output $y$ for a query $\vec{x}$ is
\begin{equation}
\begin{split}
P&\left(y|\mathcal{X},\vec{y},\vec{x}\right) = \mathcal{N}\left(k(\vec{x}, \mathcal{X})[k(\mathcal{X}, \mathcal{X}) +\eta^2\vec{I}]^{-1}\vec{y}, \right.\\
&\left. k(\vec{x}, \vec{x}) - k(\vec{x}, \mathcal{X})[k(\mathcal{X}, \mathcal{X}) + \eta^2\vec{I}]^{-1}k(\mathcal{X},\vec{x})\right)
\end{split} \label{eq:gpEquation}
\end{equation}
where $\eta^2$ is the measurement noise variance in the measurements of $\vec{y}$ \cite{Rasmussen2006}. This noise term can either be selected prior to model fitting if it is known or estimated in advance, or it can be treated as a hyperparameter that can be estimated based on the training dataset \cite{Rasmussen2006}.
For a single query $\vec x$, the posterior mean according to Eq. \ref{eq:gpEquation}, $\mu(\vec x)$, can serve as the regression output. A benefit of GPs is some measure of confidence may be obtained with the variance of the posterior distribution in Eq. \ref{eq:gpEquation}, represented now as $\sigma^2(\vec x)$. A confidence intervals (CI) is a set of bounds containing the true output with some probability. The bounds of a two-sided CI are $\mu(\vec x) \pm z\,\sigma(\vec x)$, where $z$ corresponds to a number of standard deviations from the mean. $z$ can be selected to contain a certain proportion of the area under a standard Gaussian distribution and may be determined from Gaussian distribution tables \cite{Altman1988} or statistics software. For example, if $z=1.96$, the two-sided CI contains 95\% of the area, which intuitively means the CI has a probability of 95\% of containing the true regression output \cite{Altman1988}. Selecting larger $z$ values to increase the probability of containing the true output will have a tradeoff of looser bounds on the CI
The inversion that occurs in Eq. \ref{eq:gpEquation} may be costly for larger training datasets. This issue has been addressed by reduced rank approximations or using subsets of the training dataset \cite{Rasmussen2006}, which are handled by many GP software libraries, and for which robot-specific GP modeling strategies have been proposed \cite{Wilcox2020}.
\section{Methods}
In this section, we define a model $f : \mathbb{R}^D\rightarrow \mathbb{R}$ that accurately estimates $d_{\mathcal{R}}(\vec{x})$ for a given robot configuration $\vec{x}$ and accounts for potential uncertainty in the estimation.
\subsection{Gaussian Process Regression Model Fitting}
We propose to fit a GP model to a training set of configurations $\mathcal{X}$ because of a GP's flexibility in modeling arbitrary datasets and its capability in accounting for uncertainty. We further propose the use of the FK kernel in the GP due to the boost in performance this kernel has proven to provide in binary proxy collision checking applications \cite{Das2020a}.
Uniformly randomly sampled robot configurations are used as the training set $\mathcal{X}$. Each configuration in the set is labeled with a noisy distance measurement (calculated using the GJK/EPA \cite{Gilbert1988,VandenBergen2001} geometric methods in our implementation), and the labels are stored in vector $\vec{y}$. To simulate noisy, sensor-based measurements, each label in $\vec{y}$ is created by taking a sum of the true distance and Gaussian noise: $\tilde{d}_\mathcal{R}(\vec{x}) = d_{\mathcal{R}}(\vec{x}) + \epsilon$, where $\epsilon \sim \mathcal{N}(0,\eta^2)$.
All methods were written in MATLAB. In the implementation of the FK kernel, a control point is placed at the distal end of each link in the manipulator as was done in \cite{Das2020a}.
\subsection{Confidence-Based Hybrid Model}
\label{sec:hybrid}
As described in Section \ref{sec:gp}, two-sided CIs may be defined based on the posterior distribution standard deviation. Similarly, a one-sided CI may be defined, e.g., $[\mu(\vec x) - z\,\sigma(\vec x), \infty)$ is the CI with a lower bound on what the true regression output is for configuration $\vec{x}$ with some degree of confidence. In this case, $z=1.64$ means that there is 95\% probability that the true regression output is greater than $\mu(\vec x) - 1.64\,\sigma(\vec x)$. A one-sided lower-bounded CI may be used when a bound tighter than in the two-sided CI case is desired and an upper bound on the true value is not as important as a lower bound.
\begin{figure}[!t]
\centering
\subfloat[GP (Gaussian)\label{fig:ciRBF}]{%
\includegraphics[width=0.49\linewidth, trim={0 0 0 0}, clip]{images/ciRBF}}
\hfill
\subfloat[GP (FK)\label{fig:ciFK}]{%
\includegraphics[width=0.49\linewidth, trim={0 0 0 0}, clip]{images/ciFK}}
\caption[Different levels of confidence of the GP models for $d_{\mathcal{R}}(\cdot)$ being greater than 0 for a 2 DOF robot are shown.]{The colored regions represent different levels of confidence of the GP models for $d_{\mathcal{R}}(\cdot)$ being greater than 0 for a 2 DOF robot, where each axis represents one DOF. The ground truth is shown in Fig. \ref{fig:distExample}. The dashed black curve is where $d_{\mathcal{R}}(\cdot)=0$ using GJK. The GPs use (a) the Gaussian kernel and (b) the FK kernel.}
\label{fig:ci}
\end{figure}
In Fig. \ref{fig:ci}, we visualize the levels of confidence for a 2 DOF robot having $d_{\mathcal{R}}(\vec{x})\ge 0$ when using either GP (Gaussian) or GP (FK). The dashed black curve is where $d_{\mathcal{R}}(\vec{x}) = 0$. We can see that the confidence regions for GP (FK) closely align with the C-space obstacle contour, while GP (Gaussian) does not match as closely. In fact, GP (Gaussian) has pockets of inappropriate confidence, such as decreased confidence in the corners of the shown C-space or high confidence at the bottom of the C-space obstacle. On the other hand, GP (FK) does not have these issues and has increasing confidence of $d_{\mathcal{R}}(\vec{x})$ truly being greater than 0 for configurations farther away from the C-space obstacle. GP (FK)'s more appropriate assignment of confidence suggests that we can create a hybrid model that switches distance estimation methods depending on the GP's level of confidence
Using the lower-bounded CI, we define a hybrid distance estimation technique that uses the GP mean as the predicted distance in areas with a high confidence of being collision-free and a more expensive method in other areas:
\begin{equation}
\mu^{(P)}(\vec{x}) =
\begin{cases}
\mu(\vec{x}), & \text{if }\mu(\vec{x}) - z\,\sigma(\vec{x}) \ge 0 \\
\frac{1}{N_{\tilde{d}}}\sum_{i=1}^{N_{\tilde{d}}}\tilde{d}_\mathcal{R}(\vec{x}), & \text{otherwise}
\end{cases} \label{eq:hybrid}
\end{equation}
where $P$ denotes the desired level of confidence associated with the number of standard deviations $z$ (e.g., $P=95\%$ for $z=1.64$), and $N_{\tilde{d}}$ is the number of samples to collect from $\tilde{d}_\mathcal{R}(\cdot)$ to potentially reduce the effects of sensor noise. The lower bound in Eq. \ref{eq:hybrid} may of course be nonzero if a more conservative model is desired. This hybrid model may be useful when cheaper distance evaluations from the GP may be used when working far from the obstacle, but more meticulous methods may be required when in close proximity with obstacles.
\section{Results}
We evaluate the performance of our proposed GP-based distance estimation technique against three other approaches: the geometric approach GJK/EPA (ground truth), a GP with a typical Gaussian kernel, and KR.
We begin by comparing the accuracy and timing of each technique before applying these techniques to path optimization. In each experiment, a 7 DOF planar robot manipulator is modeled with unit length link and a rectangular link profile, and a polygonal workspace obstacle is randomly placed in the manipulator's reachable workspace. We set the measurement noise to be $\eta = 0.05$ when creating each dataset. To get an idea of the level of this noise, for an obstacle one link length away from the robot (i.e., $d_{\mathcal{R}}(\vec{x}) = 1$), 99\% of the measured values could be anywhere in the range $[0.85,1.15]$. We assume the scale of the robot and environments are arbitrary, so units for any measurements of length are omitted.
\subsection{Accuracy and Query Timing}
In this section, we evaluate the GP model's performance in terms of its accuracy and computation timing. Metrics that we utilize to evaluate performance include mean squared error (MSE) and model query time. MSE is defined as
\begin{equation}
MSE = \textstyle \frac{1}{\widehat{N}}\sum_{i=1}^{\widehat{N}}{\left|\widehat{\vec{y}}_i-\mu\left(\widehat{\mathcal{X}}_i\right)\right|^2}
\label{eq:mse}
\end{equation}
where the test set $\widehat{\mathcal{X}}$ has $\widehat{N}$ elements, $\widehat{\vec{y}}$ is the set of true labels for $\widehat{\mathcal{X}}$, and $\mu(\cdot)$ provides the predicted distance using one of the approximation techniques. Note that while each model is trained on noisy distance evaluations (i.e., $\tilde{d}_\mathcal{R}(\cdot)$), we use noise-free distance evaluations (i.e., $d_{\mathcal{R}}(\cdot)$) to obtain $\widehat{\vec{y}}$ when computing the MSE.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth, trim={.5cm 0cm 1.4cm 0.1cm}, clip]{images/errorHist7DOF}
\caption{Histogram of absolute errors in distance estimation of $d_{\mathcal{R}}(\cdot)$ when using KR and GP models with the Gaussian and FK kernel. Models with a higher proportion of lower error are better.}
\label{fig:errorHist}
\end{figure}
Because the volumes of the in-collision and collision-free regions in the C-space may not be equal, the overall MSE may misrepresent the MSE within each region. Thus, we also evaluate MSE on subsets of the test data for which $\widehat{\vec{y}}$ is either positive (collision-free) or negative (in-collision), which we refer to as true positive MSE (TPMSE) and true negative MSE (TNMSE), respectively. TPMSE and TNMSE are defined as
\begin{align}
TPMSE = \textstyle \frac{1}{\widehat{N}_+}\sum_{i:\widehat{\vec{y}}_i>0}{\left|\widehat{\vec{y}}_i-\mu\left(\widehat{\mathcal{X}}_i\right)\right|^2} \\
TNMSE = \textstyle \frac{1}{\widehat{N}_-}\sum_{i:\widehat{\vec{y}}_i<0}{\left|\widehat{\vec{y}}_i-\mu\left(\widehat{\mathcal{X}}_i\right)\right|^2}
\end{align}
where $\widehat{N}_+$ and $\widehat{N}_-$ denote the sizes of the subsets for TPMSE and TNMSE, respectively.
Table \ref{table:gp7DOF} shows the performance of the models in terms of MSE and query time of each of the methods. Note that the average distance to collision is 1.7. We can see the MSEs when using the GP method with the FK kernel are significantly lower (i.e., closer to the GJK/EPA ground truth values) than the other two approximation methods. The FK kernel gives a large boost in modeling accuracy compared to the Gaussian kernel, as was expected \cite{Das2020a}, due to the FK kernel's strong relation with locations in the workspace. The KR method performs worse than GP (Gaussian) in terms of MSE, probably because all points in $\mathcal{X}$ have equal importance with the KR method while GPs assign weights to each point. Fig. \ref{fig:errorHist} shows histograms of absolute errors in estimation of $d_{\mathcal{R}}(\cdot)$ for GP (Gaussian), GP (FK), and KR. We can see that GP (FK) has a larger proportion of lower errors compared to the other methods.
\begin{table}[!t]
\centering
\begin{tabular*}{0.95\columnwidth}{lllll}
\toprule
Method & MSE & TPMSE & TNMSE & Query Time ($\mu$s) \\
\midrule
GJK/EPA & --- & --- & --- & 466.9 \\
KR & 0.78 & 0.71 & 2.76 & 10.4 \\
GP (Gaussian) & 0.50 & 0.41 & 2.92 & \textbf{1.8} \\
GP (FK) & \textbf{0.06} & \textbf{0.05} & \textbf{0.12} & 6.7 \\
\bottomrule
\end{tabular*}
\caption{Performance in terms of MSE and query time for noise-free GJK/EPA, KR, GP (Gaussian), and GP (FK).}
\label{table:gp7DOF}
\end{table}
\begin{figure*}[!b]
\centering
\subfloat[GJK/EPA]{
\includegraphics[width=0.18\linewidth, trim={0.8cm 1cm 0.6cm .7cm}, clip]{images/Traj7DOFMethodGJK}
}
\hfill
\subfloat[GJK/EPA (Noisy)]{
\includegraphics[width=0.18\linewidth, trim={0.8cm 1cm 0.6cm .7cm}, clip]{"images/Traj7DOFMethodGJKNoisy"}
}
\hfill
\subfloat[KR]{
\includegraphics[width=0.18\linewidth, trim={0.8cm 1cm 0.6cm .7cm}, clip]{"images/Traj7DOFMethodKR_RBF"}
}
\hfill
\subfloat[GP (Gaussian)]{
\includegraphics[width=0.18\linewidth, trim={0.8cm 1cm 0.6cm .7cm}, clip]{"images/Traj7DOFMethodGP_RBF"}
}
\hfill
\subfloat[GP (FK)]{
\includegraphics[width=0.18\linewidth, trim={0.8cm 1cm 0.6cm .7cm}, clip]{"images/Traj7DOFMethodGP_FK"}
}
\caption[Examples of end effector paths before and after optimization when using GJK, GP with Gaussian kernel, GP with FK kernel, and KR for distance estimation.]{Examples of end effector paths before (black curves) and after (colored curves) optimization of Eq. \ref{eq:trajOptimization} (clearance constraint of 0.2) when using (a) GJK/EPA, (b) GJK/EPA with noise, (c) KR, (d) GP with Gaussian kernel, (e) GP with FK kernel for distance estimation. The configuration along the path with the worst distance to collision and the amount of clearance required from the obstacle are also shown.}
\label{fig:gpPaths}
\end{figure*}
GP (FK) is about 3.7 times slower to query than GP (Gaussian) because the FK kernel involves an extra step of evaluating the forward kinematics of the robot and thus takes longer to evaluate than the Gaussian kernel. Also, while the KR method does not require any training time, its query times are longer because the entire training set is used during querying while only a subset of the training set needed for the GP methods. The GP and KR methods used datasets of size $N = 500$, which took approximately 229 ms to create. The training time for the GP (Gaussian) method is about 2.2 seconds, which is approximately 1 second faster than when using the FK kernel.
While GP (FK) is not as fast for training and querying compared to the GP (Gaussian), its MSE is significantly lower. GP (FK) has better modeling performance most likely because the FK kernel has a shape that more appropriately fits the structure of the data, as can be seen by comparing Fig. \ref{fig:fkKernel} to Fig. \ref{fig:distCspace}. Due to the greater accuracy of the GP (FK) model over GP (Gaussian) and KR and the fact that the query time of GP (FK) is still 70 times faster than the ground truth method, GP (FK) is a strong candidate for accurately and efficiently modeling $d_{\mathcal{R}}(\cdot)$.
\subsection{Path Optimization with Clearance Constraint}
\label{sec:gpOptimizationPerformance}
In this section, we utilize the distance estimation techniques described above in a path optimization application. One purpose of path optimization is to minimize a cost, such as path length, while satisfying constraints, such as keeping a certain distance from obstacles, given a motion plan generated by some other method \cite{Schulman2013}. In our experiment, we structure our optimization problem similar to the trajectory optimization formulation in TrajOpt \cite{Schulman2013}. More specifically, the cost is defined as a sum of distances between waypoints under the inequality constraint of distance to collision.
In our implementation, we minimize the manipulator's end effector trajectory rather than the C-space trajectory in order to avoid trivial solutions such as rotating the manipulator the other way around an obstacle. In this section, we represent the Cartesian position of the end effector in configuration $\vec{x}$ as $FK_{EE}(\vec{x})$. Our optimization problem to find a sequence of $T$ configurations $\Theta=[\theta_1,\ldots,\theta_T]$ where $\theta_t \in \mathbb{R}^D$ is thus
\begin{equation}
\begin{aligned}
\Theta^* = \argmin_{\Theta} \quad & L\left(\Theta\right)\\
\qquad \text{subject to}\quad & \theta_1 = \theta_{start}, \\
\quad & \theta_T = \theta_{goal}, \\
\quad & \|\theta_{t+1} - \theta_t\| \le \Delta\theta\ \forall t \in [1, \ldots, T-1], \\
\quad & \min_{t\in[1,\ldots,T]}{d_{\mathcal{R}}(\theta_{t})} \ge d_{min}
\end{aligned} \label{eq:trajOptimization}
\end{equation}
where $\Theta^*$ is the optimized motion plan, $L\left(\Theta\right)=\sum_{t=1}^{T-1}\|FK_{EE}(\theta_{t+1}) - FK_{EE}(\theta_t)\|$ is the distance traveled by the end effector when executing motion plan $\Theta$, $\theta_{start}$ and $\theta_{goal}$ are desired start and goal configurations, $\Delta\theta$ prevents large joint changes, and $d_{min}$ is a desired amount of clearance to maintain from a workspace obstacle. $d_{\mathcal{R}}(\theta_t)$ in the constraint is replaced by each model's estimate.
\begin{table}[!b]
\footnotesize
\centering
\begin{tabular}{llll}
\toprule
Method & Optim. Time (s) & $L(\Theta)$ & $\min d_{\mathcal{R}}(\Theta) - d_{min}$ \\
\midrule
GJK/EPA & 864.1 & 14.4 & 0.16 \\
GJK/EPA (Noisy) & 240.4 & 16.5 & 0.23 \\
KR & \textbf{29.1} & 26.9 & \textbf{0.29} \\
GP (Gaussian) & 49.2 & 16.1 & -0.12 \\
GP (FK) & 106.8 & \textbf{13.7} & 0.06 \\
\bottomrule
\end{tabular}
\caption{Average optimization time, end effector path length, and minimum satisfaction of clearance constraint after optimization of Eq. \ref{eq:trajOptimization} (clearance constraint) when using GJK/EPA, GJK/EPA with noise, GP (Gaussian kernel), GP (FK kernel), and KR to estimate $d_{\mathcal{R}}(\cdot)$.}
\label{table:gpOpt7DOF}
\end{table}
The constrained optimization problem solver $\mathrm{fmincon}()$ in MATLAB is used to solve Eq. \ref{eq:trajOptimization}. $\theta_{start}$ and $\theta_{goal}$ are randomized on either side of the obstacle to require obstacle avoidance in the motion plan. The initial motion plan is acquired using a Rapidly-Exploring Random Tree motion planner \cite{LaValle2008}, which is then used as the initial seed for the optimization. $d_{min}$ is set to be 0.2 in these experiments.
Table \ref{table:gpOpt7DOF} shows optimization times, optimized path lengths, and the worst satisfaction of the clearance constraint $\min d_{\mathcal{R}}(\Theta)-d_{min}$. Note that $\min d_{\mathcal{R}}(\Theta)-d_{min}$ is positive if the constraint is satisfied. While optimization with GP (Gaussian) and KR is faster than when optimizing with GP (FK), the optimized result is often either in collision or has a long end effector path length. All learning-based methods are significantly faster than GJK/EPA, but using the GP (FK) provides the safest results.
Fig. \ref{fig:gpPaths} shows various examples of the optimized paths using the various methods to approximate $d_{\mathcal{R}}(\cdot)$ in Eq. \ref{eq:trajOptimization}, including a case where the noisy distance measurements are used directly. The original unoptimized end effector path is shown as the dotted curve while the optimized paths are the thicker, colored curves. The configuration along each optimized path with the lowest distance to collision (according to GJK) is displayed, and the amount of desired clearance is represented as the dotted box around the workspace obstacle. It is evident that the optimized trajectories when using the GP (FK) is fairly similar to those when using GJK/EPA, while the other methods yield less optimal paths such as ones with collisions or with longer path lengths.
\begin{figure*}[t]
\centering
\subfloat[GJK/EPA]{
{\includegraphics[width=0.18\linewidth, trim={0.7cm 2.2cm 0.4cm 1.9cm}, clip]{images/Traj7DOFMethodGJKNewOpt}}
}
\hfill
\subfloat[GJK/EPA (Noisy)]{
\includegraphics[width=0.18\linewidth, trim={0.7cm 2.2cm 0.4cm 1.9cm}, clip]{"images/Traj7DOFMethodGJKNoisyNewOpt"}
}
\hfill
\subfloat[KR]{
\includegraphics[width=0.18\linewidth, trim={0.7cm 2.2cm 0.4cm 1.9cm}, clip]{"images/Traj7DOFMethodKR_RBFNewOpt"}
}
\hfill
\subfloat[GP (Gaussian)]{
\includegraphics[width=0.18\linewidth, trim={0.7cm 2.2cm 0.4cm 1.9cm}, clip]{"images/Traj7DOFMethodGP_RBFNewOpt"}
}
\hfill
\subfloat[GP (FK)]{
\includegraphics[width=0.18\linewidth, trim={0.7cm 2.2cm 0.4cm 1.9cm}, clip]{"images/Traj7DOFMethodGP_FKNewOpt"}
}
\caption[Examples of end effector paths before and after optimization when using GJK, GP with Gaussian kernel, GP with FK kernel, and KR for distance estimation.]{Examples of end effector paths before (black curves) and after (colored curves) optimization of Eq. \ref{eq:trajOptimization2} (clearance maximization) when using (a) GJK/EPA, (b) GJK/EPA with noise, (c) KR, (d) GP with Gaussian kernel, (e) GP with FK kernel for distance estimation. The configuration along the path with the worst distance to collision is also shown.}
\label{fig:gpPaths2}
\end{figure*}
\begin{figure*}[!b]
\centering
\subfloat[Optimized Path\label{fig:hybridModel}]{%
{\includegraphics[width=0.48\linewidth, trim={0.6cm 0.6cm 0cm 0.1cm}, clip]{images/hybridModel}}}
\hfill
\subfloat[Distances\label{fig:hybridDistances}]{%
{\includegraphics[width=0.48\linewidth, trim={0cm 0.1cm 1cm 0.6cm}, clip]{images/hybridDistances}}}
\caption{(a) The optimized path determined using the hybrid model $\mu^{(95)}_{FK}(\cdot)$ is shown, where the GP's confidence level that the true distance to collision is greater than 0 is represented by the color of the optimized path. The plotted robot configuration marks the point where the hybrid model switches from the GP-based predictions to the sensor-based method. (b) The distances based on the GP model with the FK kernel ($\mu_{FK}(\cdot)$), the hybrid model ($\mu^{(95)}_{FK}(\cdot)$), and the noise-free geometric method $d_{\mathcal{R}}(\cdot)$ for the trajectory $\Theta$ are shown. The shaded region represents the lower-bounded CI for the GP model, and the vertical line shows where the hybrid model switches from GP-based to sensor-based predictions.}
\label{fig:hybrid}
\end{figure*}
\subsection{Path Optimization with Clearance Maximization}
In this section, we once again apply the distance estimation techniques to path optimization. However, rather than utilizing the distance to collision as a constraint, we include distance in the objective function to increase a trajectory's clearance. We weight the path length objective function used in Eq. \ref{eq:trajOptimization} with an exponentiated distance, which changes the optimization problem to
\begin{equation}
\begin{aligned}
\Theta^* = \argmin_{\Theta} \quad & C\left(\Theta\right)L\left(\Theta\right)\\
\qquad \text{subject to}\quad & \theta_1 = \theta_{start}, \\
\quad & \theta_T = \theta_{goal}, \\
\quad & \|\theta_{t+1} - \theta_t\| \le \Delta\theta\ \forall t \in [1, \ldots, T-1]
\end{aligned} \label{eq:trajOptimization2}
\end{equation}
where $C\left(\Theta\right) = \exp(-\mind_{\mathcal{R}}(\Theta))$. $d_{\mathcal{R}}(\Theta)$ in $C(\Theta)$ is replaced by each distance method's estimate. The exponentiation makes $C(\Theta)$ large for trajectories causing egregious penetration into an obstacle and small for trajectories with larger clearances.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{llll}
\toprule
Method & Optim. Time (s) & $L(\Theta)$ & $\min d_{\mathcal{R}}(\Theta)$ \\
\midrule
GJK/EPA & 481.9 & \textbf{4.3} & \textbf{1.73} \\
GJK/EPA (Noisy) & 56.4 & 32.3 & 0.47 \\
KR & 28.7 & 9.3 & 0.03 \\
GP (Gaussian) & \textbf{15.7} & 4.5 & -0.25 \\
GP (FK) & 51.0 & 4.9 & 1.24 \\
\bottomrule
\end{tabular}
\caption[Results for optimal paths using GJK \cite{Gilbert1988,VandenBergen2001}, GP (Gaussian kernel) \cite{Rasmussen2006}, GP (FK kernel) \cite{Rasmussen2006, Das2020a}, KR \cite{Nadaraya1964,Watson1964} to calculate $d_{\mathcal{R}}(\cdot)$ are presented. We include the average times to achieve the optimized plan, the average final cost, and the average worst distance to collision in each optimized path.]{Average optimization time, end effector path length, and minimum distance to collision after optimization of Eq. \ref{eq:trajOptimization2} (clearance maximization) when using GJK/EPA, GJK/EPA with noise, GP (Gaussian kernel), GP (FK kernel), and KR to estimate $d_{\mathcal{R}}(\cdot)$.
\label{table:gpOpt7DOF2}
\end{table}
Table \ref{table:gpOpt7DOF2} shows average optimization times, path lengths, and minimum distances to collision for each method. GP (FK) has a higher amount of clearance than GP (Gaussian), KR, and noisy GJK/EPA methods, and in fact achieves results closer to the noise-free geometric method both in terms of clearance and path length despite being trained on noisy measurements. The noisy version of GJK/EPA yielded suboptimal solutions with large end effector path lengths and smaller clearances than GP (FK). GP (Gaussian) and KR have shorter optimization times because the optimization solver often quickly converges to an invalid solution (as can be seen from the lower minimum distances to collision). The paths in Fig. \ref{fig:gpPaths2} exemplify these results.
\subsection{Confidence-Based Hybrid Model}
We apply our proposed hybrid model to a scenario involving a narrow passage, where the manipulator must reach between two nearby obstacles. Inaccuracy in the distance estimation model may inadvertently lead to collision between these obstacles, so the hybrid model's predictions can potentially handle this scenario by switching to the geometric method if the GP model is not confident.
Fig. \ref{fig:hybridModel} shows an example optimized trajectory achieved with the hybrid model $\mu^{(95)}_{FK}(\cdot)$. The plotted robot configuration denotes the transition point where the hybrid model switches from GP-based predictions to noisy, sensor-based distance evaluations $\tilde{d}_\mathcal{R}(\cdot)$. We used $N_{\tilde{d}}=5$ evaluations of $\tilde{d}_\mathcal{R}(\cdot)$ per configuration when the hybrid model did not use a GP-based prediction. The colors of the path denote the confidence level of the GP that the true distance to collision is greater than 0, which we determine via MATLAB's $\mathrm{normcdf}()$. We can observe that the GP model is fairly confident when far from the obstacles, but the confidence level rapidly drops when the manipulator needs to move into the narrow passage.
Fig. \ref{fig:hybridDistances} shows the GP-based distance estimates $\mu_{FK}(\Theta)$, the hybrid estimates $\mu^{(95)}_{FK}(\Theta)$, and the true distance to collision $d_{\mathcal{R}}(\Theta)$. The shaded region represents the lower-bounded CI for the GP model. We can see that $\mu_{FK}(\Theta)$ deviates from $d_{\mathcal{R}}(\Theta)$ as the distance decreases, but the hybrid model $\mu^{(95)}_{FK}(\Theta)$ manages to produce accurate measurements of distance by switching to the noisy distance measurements when the GP loses confidence.
\section{Conclusion}
In this paper, we propose a method to create a model for evaluating a robot manipulator's distance to collision. Our approach utilizes GP regression and the FK kernel, a similarity function designed to compare configurations of a robot manipulator. Noisy distance measurements are used to train the models.
Querying the GP model with the FK kernel obtained up to 70 times faster distance estimates than the geometric method. The accuracy of this GP model significantly surpassed that of other regression techniques with up to 13 times lower MSE, even when trained on noisy distance values. The GP model with the FK kernel achieved optimized paths 9 times faster but of similar quality as was achieved by a noise-free geometric method.
We also introduce a confidence-based hybrid model that uses the GP mean as a predicted distance in areas of high confidence and an averaging method over sensor-based distances in areas of lower confidence. We demonstrate the capability of this technique in a narrow passage environment and realize the hybrid model switches to the sensor-based method as it gets closer to the obstacles.
Using GP regression with the FK kernel is thus a promising approach for creating a model for distance evaluation, even when working with noisy measurements. More efficient sampling, online training and updating, and parallelization are left for future work.
\bibliographystyle{IEEEtran}
| cfe863e1844cba3d3b0ca51fd88b3d99e1eacca3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The phase behavior of two-dimensional (2D-) ice is the subject of recent experimental and
theoretical interest, which is still controversial~\cite{nat2015,prl,prb,ACS}.
Although, recently the report on the observation of monolayer, bilayer and trilayer ice using transmission electron microscopy (TEM)~\cite{nat2015} was challenged later~\cite{commentnat2015}, however, several theoretical studies based on both classical force fields and ab-initio simulations revealed the exciting possibility of exploring 2D-ice structures at specific conditions~\cite{prl,prb,mario,zangi}.
In particular, both classical force fields and ab-initio simulations predict that water molecules form ordered flat square lattice (f-SQ) structure while they are trapped in a few angstrom size slit~\cite{prl,mario}. The confinement width needed for the formation of stable monolayer ice is approximately h$\simeq$5-7\AA~~\cite{nat2015,zangi,mario}.
In contrast, the structural properties of 2D-ice (and bulk ice), the electronic, dielectric and optical properties of 2D-ice and nanoconfined water is not well understood. Therefore, a solid theoretical background for the effects of size reduction on the dielectric properties of ice and confined water is highly demanded. This can be helpful to understand the measured anomalous dielectric properties of confined water~\cite{science}: recently, by using electrostatic force detection of atomic force microscope (AFM), unexpected variation in the out-of-plane dielectric constant of confined water between graphene and hexagonal boron nitride (h-BN) has been observed~\cite{science}. The presence of an interfacial water layer (having ice phase) with vanishingly small polarization is the
reason for such small out-of-plane dielectric constant ($\simeq$2) for channels with size of $h<$15\AA.~Indeed the dielectric constant of nanoconfined water was found to be about 2, which is above the high frequencies dielectric constant of water, i.e. 1.8.
In the past few decades, molecular dynamics simulations (MDS) and monte carlo simulations (MCS) have been used to calculate the dipolar dielectric constant of water and confined water.
For instance the variation of the dipolar dielectric constant with temperature
and pressure for the ices Ih, III, V, VI, and VII was studied by Aragones et. al.~\cite{aragon}. In addition to MDS and MCS methods, mean field theory (such as Kirkwood's theory) was also used and
yielded valuable insights into the H-bonding effects on water dielectric constant~\cite{daniel}. Notwithstanding the existing MD based theory studies in the past few years, the questions about the ionic and electronic contributions of dielectric constant of 2D-ice is still unanswered. Most of the previous studies reported the dipolar dielectric constant of water confined at the nanoscale channels~\cite{ref7,ref8,ito}. Thus, one naturally expects to quantify the dipolar, electronic and ionic dielectric constants of crystalline 2D-ice and corresponding frequency dependence. This will provide a solid theoretical support for the recent experiment~\cite{science}.
Here, we conducted a systematic study for detrmining the electronic, dielectric and optical properties of 2D-ice using multi-scale approach including first principles and molecular dynamics simulations. Both the out-of-plane and in-plane components of ionic, electronic and dipolar dielectric constants of stable structure of 2D-ices including flat square (f-SQ), buckled square (b-SQ), buckled rhombic (b-RH) and hexagonal structure (HEX) are investigated~\cite{prl}. We also reported results for the electronic, dielectric and optical properties of the crystalline bulk ice and bulk water.
Our work provides benchmark theoretical data for the electronic, dielectric and optical properties of crystalline 2D- and bulk ices.
\begin{figure*}[]
\includegraphics[width=.08\linewidth]{1111.png}
\includegraphics[width=.220\linewidth]{f-SQ.png}
\includegraphics[width=.220\linewidth]{b-SQ.png}
\includegraphics[width=.220\linewidth]{b-RH.png}
\includegraphics[width=.220\linewidth]{HEX.png}
\includegraphics[width=.08\linewidth]{11111.png}
\includegraphics[width=.220\linewidth]{8.png}
\includegraphics[width=.220\linewidth]{11.png}
\includegraphics[width=.220\linewidth]{ic.png}
\includegraphics[width=.220\linewidth]{ih.png}
\caption{ The lattice structure of two-dimensional and bulk ices. The top and side views of (a) flat square (f-SQ), (b)
buckled square (b-SQ), (c) buckled rhombic (b-RH), and (d) hexagonal (HEX) 2D-ice structures. The blue squares show the
corresponding supercells. (e) ice VIII, (f) ice XI, (g) ice Ic, and (h) ice Ih. The blue cubes show the corresponding unitcells.} \label{fig1}
\end{figure*}
\section{Two-dimensional ice}
The 2D-ice structures considered in this paper are flat square (f-SQ), buckled square (b-SQ), buckled rhombic (b-RH) and hexagonal (HEX). Their optimized structures were reported in Ref.~[2]. The f-SQ and b-RH (HEX) structures have square (rectangular) unitcell containing 4 (8) water molecules (see Figs.~\ref{fig1}(a-d)). In order to eliminate the interaction between periodic images in ab-initio calculations, we set a large vacuum with height ``$c$" for calculating each 2D-ice (see section ``Dielectric constant"). The lattice parameters of the simulation supercells ($a,~b,~c$) are listed in Table 1. The input lattice structures were extracted from the optimized structures of Ref.~[2]. Using van der Waals diameter of oxygen atom, one can approximate the effective thickness ($t$) of a given 2D-ice lattice: considering flat structure of f-SQ, $t_{f-SQ}=t_0=2\times R_v$, where $R_v$ is the van der Waals radius of oxygen atom. For b-SQ and HEX structures, there is 1 angstrom difference between top row and bottom row of O atoms (see side view of b-SQ and HEX structures in Fig. 1(b,d)), thus $t_{b-SQ,HEX}=1+t_0$. For the b-RH, there is a large (4\AA)~ distance between the top row and bottom row of O atoms, i.e $t_{b-b-RH}=4+t_0$. All the structure parameters of the studied systems and the important findings of this papaer are tabulated in Table 1.
\section{Crystalline bulk ice}
In order to compare electronic, dielectric and optical properties of above mentioned 2D-ices with the high pressure phase of ice, we also calculated the dielectric constant of four crystalline ice, i.e. cubic (ice VIII and ice Ic) and hexagonal (ice XI and ice Ih) bulk ices.\\
The ice XI, ice Ic and ice Ih structures have 8 water molecules per orthorhombic unitcell (see Figs.~\ref{fig1}(e-h)). The lattice parameters of the crystalline ices are listed in Table I. The ice Ih has the hexagonal crystalline form of ordinary ice, which is stable at temperature 273\,K (down to few Kelvin) and pressures up to 200\,MPas. The ice Ic is one of the metastable cubic crystalline form of ice, which is stable at temperatures between 130 and 220 K. The ice VIII with 8 water molecules per unitcell has a tetragonal crystalline form and is stable under high pressures about 3\,GPa below 278\,K. The ice XI is a hydrogen-ordered form of ice Ih containing 8 molecules per unitcell and is stable at temperature 5\,K and pressures of about 100\,MPas.
We also calculated the dielectric constant of bulk water using 17 water molecules inside a cubic unitcell with size 7.93$\times$7.93$\times$7.93\,\AA$^3$. We have used 20 different relaxed MDS configurations as inputs for the density functional theory (DFT) calculations of bulk water. For the bulk water and bulk crystalline ices, we applied periodic boundary conditions in all directions, though for the 2D-ices, a vacuum space must be set (see section V. A). In order to obtain more insights and for comparison purposes, we put in order the dielectric constant of some common 2D-materials such as monolayer h-BN and monolayer of three transition metal dichalcogenides (TMDs)~\cite{2Dmater}, i.e. MoS$_{2}$, WS$_{2}$, and WSe$_{2}$.
\section{Dielectric constant}
For the polar systems, the total dielectric constant tensor includes three main contributions, i.e. electronic, ionic and dipolar:
\begin{equation}\label{Eq1}
\varepsilon_{total}^{\mu\nu}=\varepsilon_{el}^{\mu\nu}+\varepsilon_{ion}^{\mu\nu}+\varepsilon_{dip}^{\mu\nu},
\end{equation}
The indexes $\mu$ and $\nu$ run over the three spatial directions. At zero Kelvin the dipolar term vanishes. By increasing temperature of 3D-material, the dipolar term becomes important and should be taken into account. Note that different bulk ices are stable at diffrent tempratures~\cite{Martin}. In Table I, the relevant temperatures are listed.
\section{The methods}
In this study, density functional theory (DFT) has been implemented for electronic band structure calculations. We have used density functional perturbation theory (DFPT) to obtain the electronic and ionic dielectric constants at zero Kelvin. Moreover, molecular dynamics simulations (MDS) has been used to find the dipolar dielectric constant at a finite temperature for 2D-(bulk) ices. In order to calculate the optical dielectric function, the random-phase approximation (RPA)~\cite{Brener1975} based on DFT ground-state calculations has been conducted. In the following sections, we briefly explain the different used methods.
\subsection{Density functional theory: Electronic band structure}
We have calculated the electronic band structure of 2D-ices using
density functional theory (DFT) as implemented in the Quantum-ESPRESSO (QE) package~\cite{Giannozzi2009}. We have used ultrasoft pseudopotentials to treat the interaction between the ion cores and valence electrons and applied the generalized gradient approximation (GGA) for the exchange-correlation interactions. We also have studied the effect of nonlocal correlations using the van der Waals density functional (vdW-DF) of Dion et al~\cite{Dion2011}. Accuracy of forces on each atom has been considered about 0.1 mRy/bohr for the variable-cell optimization, relaxing the cell parameters and atomic positions. in order to accurate calculation of structural and electronic properties, the kinetic energy cut-offs of $150~Ry$ and $1500~Ry$ were found to be sufficient for the wavefunctions and the charge densities, respectively, where the k-point grid for 2D- (bulk) ices was set to $6\times6\times1$ ($6\times6\times6$). The k-point grid for 2D- (bulk) ices was chose $24\times24\times1$ ($24\times24\times24$) for non-self-consistent calculation in partial density of state (PDOS) analysis. The smearing parameter of 0.01 Ry has been used for PDOS analysis. In addition, total energy convergence threshold was set to $10^{-12}eV$. The Coulomb cutoff technique~\cite{Rozzi2006,Sohier2017} was used to reduce interactions between periodic images and cost of the ab-initio calculations for 2D-ice structures.
\subsection{Density functional perturbation theory: Electronic and ionic dielectric constant}
The dielectric properties of the 2D-ices were determined using the DFPT approach. In DFPT, the dielectric constant tensor is defined as a linear response to the perturbative electric field~\cite{baroni2001,Gnoze1997} and the ionic displacement are considered as a perturbation to the equilibrium system. The response of the electronic charge density to the perturbative electric field in the linear response regime, employed to determine the electronic contribution of the dielectric tensor, i.e. the high-frequency dielectric constant ($\varepsilon_{el}$). Subsequently, the static dielectric constant ($\varepsilon_0^{\mu\nu}$) is the summation of the electronic and the ionic parts of the system to the applied electric field:
\begin{equation}\label{Eq2}
\varepsilon_0^{\mu\nu}=\varepsilon_{el}^{\mu\nu}+\varepsilon_{ion}^{\mu\nu},
\end{equation}
where
\begin{equation}\label{Eq3}
\varepsilon_{ion}^{\mu\nu}=\frac{4\pi}{\Omega}\\\sum_{m}\frac{S_m,\mu\nu}{\omega_m^2}.
\end{equation}
Here $S_{m,μν}$ is the mode oscillator strength tensor, defined in terms of the Born effective charges $Z^*$ , the atomic masses $M_i$, and the normalized eigenvectors $u_{i,m\mu}$ of the i$^{th}$ ion along a given direction $\mu$ for a particular mode $m$. Also $\omega_m$ is the phonon mode frequency and $\Omega$ is the unitcell volume. Thus, in order to compute $\varepsilon_0$, the knowledge of all the phonon frequencies at the
zone center of Brillouin zone is needed. The latter requires the solution of the dynamical matrix at the zone center.
\begin{figure*}[!ht]
\includegraphics[width=.49\linewidth]{a.png}
\includegraphics[width=.49\linewidth]{b.png}
\includegraphics[width=.49\linewidth]{c.png}
\includegraphics[width=.49\linewidth]{d.png}
\includegraphics[width=.49\linewidth]{e.png}
\includegraphics[width=.49\linewidth]{f.png}
\includegraphics[width=.49\linewidth]{g.png}
\includegraphics[width=.49\linewidth]{h.png}
\caption{ The electronic band structure of (a) f-SQ, (b) b-SQ, (c) b-RH, (d) HEX, (e) ice VIII, (f) ice XI, (g) ice Ic,
and (h) ice Ih and corresponding partial density of states. The arrows denote the type of band gap (direct or indeirect) and the insets show
the corresponding Brillouin zones. } \label{fig2}
\end{figure*}
\subsection{Molecular dynamics simulations: Dipolar dielectric constant}
By means of equilibrium molecular dynamics
simulations employing the large scale atomic/molecular massively
parallel simulator LAMMPS~\cite{{lammps}}, we computed the molecular (dipolar) dielectric constant of 2D-ice at 80 K. Our simulated systems contain 2D-ices, confined between two walls, separated by a distance~\cite{prl} $h=2\,t$ (these walls are used to produce the confining potential). The number of water
molecules are 400, 100 and 128 for f-SQ, R-SQ, and HEX, respectively\cite{1810}.
The TIP4P model was employed to describe the water molecules~\cite{TIP4P}. The NVT ensemble
(Nose-Hoover thermostat) is used to keep the temperature at 80 K for 2D- and bulk ices excluding ice Ih (200 K). We modified TIP4P model to incorporate
the lattice structure of the optimized structures by using DFT: we changed bond lengths and bond angles of water molecules relevant to the DFT outputs~\cite{prl}. Also, the charges of O and H atoms were modified such that the large dipole moment ($\sim$ 3D) for single water molecule is reached~\cite{aragon}. In order to verify this modification, we calculated the variation of dipolar dielectric constant of bulk water with time using the modified TIP4P model, see Fig. 3(a). It is seen that the dielectric constant of bulk water lies in acceptable range, i.e. it is between dielectric constant of bulk water using TIP4P and TIP4P/2005 models.
Periodic boundary conditions are applied along $x$, $y$ directions
and the confinement was along the $z$-direction. The
particle-particle particle-mesh method was used to compute the
long-range Coulomb interaction with a relative accuracy of
10$^{-4}$. Water bond lengths and bond angles were fixed by the SHAKE
algorithm~\cite{shake}. In all MDS a time step of 1\,fs was
chosen. Following the system relaxation (1\,ns), the thermodynamical
sampling was done up to 8\,ns to ensure the smoothness of the
converged dielectric constant.
The temperature in MDS for bulk ices (water) set to be 80 K (300 K); because all of the studied bulk ices are stable at this temperature.
A microscopic picture of dielectric properties of 2D-ices could be found by calculating the fluctuations
of the total polarization of a system, $\vec{M}$ at finite temperature. By calculating different components of the total dipole moment i.e. $M_x$, $M_y$, and $M_z$ after equilibration, one obtains different components of the molecular dielectric constant tensor as~\cite{zhang}
\begin{equation}
\varepsilon_{dip}^{\mu\nu}=\varepsilon_{\infty}+\frac{\sigma^2_{\mu\nu}}{\varepsilon_0 k_B T
V},\label{eq1}
\end{equation}
where $\varepsilon_{\infty}$ is the optical dielectric constant and taken to be 1. Also, $\sigma_{\mu\nu}^2$ $=$ $\langle
M_{\mu}M_{\nu}\rangle-\langle M_{\mu}\rangle \langle
M_{\nu}\rangle$ while $\mu,\nu=x,y,z$ and $V$ is the volume of
the system. Here, the time averaging was taken for more than 5\,ns when in-plane dielectric constant is converged. Note that Eq. (4) can only be used
for a homogeneous systems~\cite{ref7}.
\subsection{Random phase approximation: The frequency dependent optical dielectric constant}
The optical dielectric function $\varepsilon_{el}(\omega)$ was calculated using norm-conserving pseudopotentials, in the energy range of 0 to 30\,eV. In order to increase accuracy for the dielectric functrion calculations, we used the k-point grid for 2D- (bulk) ices as $12\times12\times1$ ($12\times12\times12$).
The optical dielectric constant was calculated (the frequency dependent of the electronic dielectric constant or equivalently the electronic part of the dielectric function) within the framework of the random-phase approximation (RPA) based on DFT ground-state calculations. The mentioned dielectric function consists of frequency dependent real ($\varepsilon_{el}^r(\omega)$) and imaginary part ($\varepsilon_{el}^i(\omega)$). It is represented as:
\begin{equation}\label{Eq3}
\varepsilon_{el}(\omega)=\varepsilon_{el}^r(\omega)+i \varepsilon_{el}^i(\omega)
\end{equation}
The imaginary part of dielectric function ($\varepsilon^i(\omega)$) can be calculated using Kubo–Greenwood formalism~\cite{Morgan1988}.
Once we know the imaginary part, the real part can be obtained using the Kramers–Kronig relations (for more details see Ref.~[29]). Note that the optical dielectric constant extracted from RPA (at zero frequency) is compatible to that obtained from DFPT.
\section{Results and discussion}
\subsection{Electronic band structure}
\begin{figure}[]
\includegraphics[width=0.95\linewidth]{FigR1.png}
\caption{(a) The variation of dielectric constant of bulk water with time using the modified TIP4P model in this study. (b) The in-plane radial distribution function of f-SQ 2D-ice. Two arrows indicate the emergence of two peaks in f-SQ which is absent in bulk ices. (c) The radial distribution function of three typical bulk ice. } \label{fig03}
\end{figure}
\begin{figure}[]
\includegraphics[width=0.85\linewidth]{gap.png}
\caption{ The energy gap of bulk phase and 2D- phase of various materials.
The data for transition metal dichalcogenides were taken from Refs.~[30,31].} \label{fig3}
\end{figure}
\begin{table*}[!ht]
\centering
\caption{Simulation parameters used for the calculation of
dielectric constant of 2D-ices and bulk ices: ionic dielectric constant ($\varepsilon_{ion}$), the electronic dielectric constant ($\varepsilon_{el}$), the dipolar dielectric constant ($\varepsilon_{dip}$), total dielectric constant ($\varepsilon_{tot}$)\cite{Notice}, the energy gap ($\Delta(eV)$), and the effective thickness ($t$). For comparison purposes we show the corresponding data for bulk water, TMDs and h-BN. The used temperatures in the molecular dynamics simulations are given too.}\label{table1}
\begin{center}
\begin{tabular}{c|cccccc|ccc|ccc|ccc|ccc|cc}
\hline
\hline
& \multirow{3}{*}{structure} & \multirow{3}{*}{a} &\multirow{3}{*}{b} & \multirow{3}{*}{c}& \multirow{3}{*}{t}& \multirow{3}{*}{$\Delta(eV)$} &\multicolumn{6}{c|}{$\varepsilon_{0}$}&& \multirow{2}{*}{$\varepsilon_{dipolar}$}&&&\multirow{2}{*}{$\varepsilon_{total}$}&& \multirow{3}{*}{T(K)}& \multirow{3}{*}{Reference}\\
\cline{8-13}
& & & & & & & \multicolumn{3}{c|}{$\varepsilon_{el}$} & \multicolumn{3}{c|}{$\varepsilon_{ion}$}&& &&& & && \\
\cline{8-19}
& & & & & & & xx & yy & zz& xx & yy & zz & xx & yy & zz & xx & yy & zz & &\\
\hline
\multirow{4}{*}{\rotatebox{90}{2D ice}}
& f-SQ & 5.84 & 5.84 & 15 & ($t_0$=)3& 5.49& 2.0 &2.0 & 1.7& 1.0 &1.0 & 2.0 & 1.55& 1.54 & 2.41 & 3.55& 3.54& 5.11& 80 & present work\\
& b-SQ & 5.78 & 5.78 & 20 &4& 5.55& 1.8 & 1.8 & 1.49 &3.85 &3.9 & 0.01 & -&- &-& -& -& -& Ref.~[24] & present work \\
& b-RH & 4.47 & 5.76 .& 20 & 8 & 5.55 & 1.45 & 1.45 & 1.3 &10.75& 0.95 & 1.0 &1.49 & 1.52& 1.92 & 12.69& 2.92& 3.22& 80 & present work \\
& HEX & 8.7 & 7.72 & 15 &4.0 &5.12 & 1.71 & 1.71 & 1.45 & 2.0 & 1.90 & 3.53 & 3.02 & 4.32 & 2.49& 5.73& 6.93& 6.47& 80& present work \\
\hline
\multirow{4}{*}{\rotatebox{90}{Bulk ice}}
& ice VIII & 4.73 & 4.73 & 6.85&- & 5.37 & 2.36 & 2.36 & 2.34 & 3.96 &3.96 & 0.26& 8.71 & 8.54 &1.03& 14.03& 13.86& 2.63& 80 & present work \\
& ice XI & 4.45 & 7.7 & 7.27 &-& 5.06 & 1.83 & 1.82 & 1.82 & 1.81 & 0.83 & 0.93& 2.06 & 1.31& 1.86 & 4.7& 2.96& 3.61& 80 & present work \\
& ice Ic & 6.48 & 6.48 & 6.48 &- & 4.99 & 1.77 & 1.77 & 1.76 & 1.21 & 1.21 & 0.16 & 2.1 & 2.1 & 1.01 &4.08& 4.08& 1.93& 80 & present work \\
& ice Ih & 7.81 & 7.38 & 4.52 &- & 5.05 & 1.80 & 1.80 & 1.80 &1.14 & 1.0 & 0.96 & 1.94 & 1.92 & 1.5 & 3.88& 3.72& 3.26& 200 & present work \\
\hline
\rotatebox{90}{\parbox{0.7cm}{Bulk}}& water &7.93 & 7.93 & 7.93&-& 4.99& 1.85& 1.84& 1.86& 1.28&1.16&1.29& 72&72&72& 74.13& 74& 74.15& 300 & present work \\
\hline
& h-BN & 2.51 & 2.51 & 25.1 & 3.17 & 5.97 & 4.97 & 4.97 & 2.89 &1.85 & 1.85 & 0.4 & - & - & - & -& -& -& & Ref.~[14] \\
\hline
\multirow{3}{*}{\rotatebox{90}{TMDs}}
& MoS$_{2}$ & 3.21 & 3.21 & 32.1 & 6.12 & $<2$ & 15.1 & 15.1 & 6.4 &0.2 & 0.2 & 0.0 & -& -& - & -& -& -& & Ref.~[14] \\
& WS$_{2}$ & 3.21 & 3.21 & 32.1 & 6.14 & $<2$ & 13.6 & 13.6 & 6.3 & 0.1 & 0.1 & 0.0 & -& -& -& -& -& -& & Ref.~[14] \\
& WSe$_{2}$ & 3.34 & 3.34 & 33.4 & 6.52 & $<2$ & 15.1 & 15.1 & 7.5 & 0.2 & 0.2 & 0.0 & -& -& -& -& -& -& & Ref.~[14] \\
\hline
\hline
\end{tabular}
\end{center}
\end{table*}
The electronic band structure and partial density of states (PDOS) of 2D- and bulk ices are shown in Fig.~\ref{fig2}.
We found the direct band gap of energy about 5.49, 5.55 and 5.12 $eV$ for f-SQ, b-SQ and HEX 2D-ices and 5.37 and 4.99 $eV$ for ice VIII and ice Ic, respectively.
The electronic band structure for b-RH 2D-ice, ice XI and ice Ih confirm the indirect band gap. The corresponding energy gap for b-RH, ice XI and ice Ih are 5.55, 5.06 and 5.08 $eV$, respectively. The PDOS are shown in the right side of the panels of Fig.~\ref{fig2}. It is seen that in contrast to the conduction bands, the O (H) atoms contribution in the valance band is larger (negligible). Notice that the difference between the energy gap of 2D-ices (and 2D-h-BN) and bulk water (bulk h-BN) is small, i.e. 2$\%$ (4$\%)$, as compared to the difference between bulk TMDs and 2D-TMDs, see Fig.~\ref{fig3}. This is due to the weak hydrogen bond between water molecules in bulk and 2D-ices as well as insulating feature in all phases of water and h-BN. The TMDs have large reductions of energy gap while transiting from 3D to 2D. The obtained energy gap, the electronic band structure and PDOS for ice VIII, ice XI,and ice Ih are in agreement with the results of Refs.~[32-35].
Independent of water phase (including ice), the nearest neighbor distance between oxygen atoms of water molecules are almost the same, i.e. 2.8-3\AA~ i.e. all the ice phases formed under very high pressures satisfy the so-called Bernal-Fowler ice rules where each water molecule has four hydrogen-bonded neighbors with a quasi-tetrahedral configuration with two short O-H distances (the donated protons) and two long ones (the accepted protons). Transiting from bulk into 2D phase of ice, the crystalline structure with larger density can be formed where the nearest neighbor distance are more or less the same but, the preferential tetrahedral bonding geometry is different. In Figs. 3(b,c) we depict the in-plane radial distribution function for O-O distance in f-SQ and 3D-radial distribution function of three bulk ice (Ice XI, ice Ic and ice Ih). As can be seen, as expected the nearest neighbor O-O distance (first peaks) for all structures are equal. However, the second nearest neighbors (shown by arrows in Fig. 3(b)) are different. The latter causes a larger density of f-SQ~\cite{mario} (1.36gcm$^{-3}$) as compared to bulk crystalline ices i.e. $~$ 0.92 gcm$^{-3}$ ice Ih and $~$ 0.92 gcm$^{-3}$ ice XI.\\
On the other hand, the most well-known electrical conductivity mechanism for all phases of water/ice is the Grotthuss mechanism (GM) also known as proton jumping. In GM an excess proton or proton defect diffuses through the hydrogen bond network of water molecules and a covalent bond between neighboring molecules are formed and broken continuously. Because of the same nearest neighbor distance in bulk ice (bulk water) and those of 2D-ice, the GM mechanism should be valid for 2D-ices. Subsequently only an infinitesimal increase in the band gap 2D-ices as compared to the bulk ices is seen. This was confirmed by our ab-initio simulations.\\
Moreover, h-BN is a wide band gap semiconductor with high thermal and chemical stability. In both bulk and monolayer h-BN, N atoms and B atoms are hybridized with sp$^2$ at the interlayer forming a honeycomb structure. In bulk h-BN there is weak interactions between each layer of h-BN, such as electrostatic interactions and Van der Waals interactions, which causes the electrical band gap of both bulk h-BN and 2D-h-BN become more or less equal~\cite{Today}. Transition from bulk into 2D-TMDs also obeys the same general role, i.e. the band gap of a 2D-TMD is larger than that of bulk.
\subsection{Dielectric constant}
Here, we turn our attention to the main focus of this paper. The dielectric constants extracted from QE represent the combined dielectric constant of 2D-ice and surroundings large vacuum with a height ``$c$". In order to distill the dielectric constant of 2D-ices, we
eliminate the contribution of the vacuum using a capacitance model~\cite{2Dmater}.
In fact, in the out-of-plane (in-plane) direction, the capacitance of the supercell extracted directly from QE code ($\epsilon_{SC}$) is the series (parallel) combination of vacuum capacitance
and the 2D-ice capacitance. This helps us to find the out-of-plane and in-plane relative dielectric constant of 2D-ices using below equations~\cite{2Dmater}:
\begin{figure*}[]
\includegraphics[width=0.45\linewidth]{e-el.png}
\includegraphics[width=0.45\linewidth]{e-ion.png}
\caption{ The electronic (a) and ionic (b) dielectric constant of 2D-ices, bulk ices and TMDs. The data for TMDs were taken
from Ref.~[14]. }\label{fig4}
\end{figure*}
\begin{figure}[]
\includegraphics[width=0.85\linewidth]{e-tot.png}
\caption{The total dielectric constant ($\varepsilon_{total}$) of 2D-ices, bulk ices and bulk water~\cite{Notice}.}
\label{fig5}
\end{figure}
\begin{equation}\label{EqSc}
\varepsilon^{zz}=[1+\frac{c}{t}(\frac{1}{\varepsilon_{SC}}-1)]^{-1},~
\varepsilon^{xx,yy}=1+\frac{c}{t}(\varepsilon_{SC}-1),
\end{equation}
``t" is the effective thickness of 2D-ice.
In order to make sure that the used ``c'' values are sufficiently large and the obtained $\varepsilon^{xx,yy,zz}$ were converged well, we performed several additional calculations and found corresponding $\varepsilon_{SC}$. The results show that when ``c'' values change from 12\AA~ to 40\AA,~ the $\varepsilon^{xx,yy,zz}$ values only increase about 5$\%$. Thus, the used vacuum size (see Table I) are sufficiently large.
Using the aforementioned correction, the results for electronic ($\varepsilon_{el}$), ionic ($\varepsilon_{ion}$), and dipolar ($\varepsilon_{dip}$) dielectric constant for 2D-ices are listed in Table I. The main findings are illustrated in Fig.~\ref{fig4} and are listed as:
\begin{itemize}
\item $\varepsilon_{ion}^{xx}\simeq\varepsilon_{ion}^{yy}$ for 2D-ices except for b-RH ice in $\varepsilon_{ion}^{xx}\simeq10 \varepsilon_{ion}^{yy}$, which is due to the remarkable in-plane anisotropic in the lattice structure of b-RH.
\item $\varepsilon_{el}^{xx}\simeq\varepsilon_{el}^{yy}$ for 2D- (bulk) ices and larger (equal to) than $\varepsilon_{el}^{zz}$. The larger the $\varepsilon_{el}^{xx}$ and $\varepsilon_{el}^{yy}$, the more flatness of the structures. This is due to the confined electron clouds in quasi 2D-space.
\item $\varepsilon_{el}^{xx}$ and $\varepsilon_{el}^{yy}$ of f-SQ ice are larger than those for other 2D-ices. This might be due to the confining electron in 2D-plane and stronger response of system to an in-plane electric field, where for other 2D-ices there is a small buckling in their geometry.
\item Because of the crystalline structure of all the studied crystalline 2D-ices and bulk ices their dipolar contributions are remarkably smaller than that of bulk water.
\item The electronic dielectric constants of 2D- and bulk ices are smaller than for other 2D-materials such as MoS$_2$, WSe$_2$, and WS$_2$~\cite{2Dmater}. The latter is attributed to the semiconducting nature of the TMDs rather than insulating nature of 2D-ices.
\end{itemize}
The items aforementioned are represented in Figs.~\ref{fig6}(a,b).
Consequently the total dielectric constant (Fig.~\ref{fig5}) of the studied ices (except for b-RH and ice VIII) are smaller than 10.
\subsection{Optical dielectric function}
In Fig.~\ref{fig6}, the real ($\varepsilon^r$) and imaginary ($\varepsilon^i$) parts of the optical dielectric function are shown for 2D- and bulk ices and bulk water. The results show that:
\begin{itemize}
\item For 2D-ices, $\varepsilon^{xx} \ne \varepsilon^{zz}$ in both real and imaginary parts which is expectable.
\item Because of isotropic (anisotropic) lattice of the f-SQ and b-SQ ices (b-RH and HEX) for both real and imaginary parts satisfy. $\varepsilon_{el}^{xx}(\omega)=\varepsilon_{el}^{yy}(\omega) (\varepsilon_{el}^{xx}(\omega)\neq\varepsilon_{el}^{yy}(\omega)$).
\item Despite the bulk cubic ice structures (ice VIII and ice Ic), the bulk hexagonal ice structures (ice XI and ice Ih), the dielectric functions are almost equal, i.e. $\varepsilon_{el}^{xx}(\omega)\simeq\varepsilon_{el}^{yy}(\omega)\simeq\varepsilon_{el}^{zz}(\omega)$, for both real and imaginary parts.
\end{itemize}
The absorption behavior of ices can be understood by analysing the imaginary part of dielectric function. Also, the absorption edge gives the optical gap ($\Delta_o$). Moreover, the peaks in the $\varepsilon^{i}(\omega)$ function is related to interband transitions in electronic band structure. For instance, the first peak corresponds to the energy gap ($\Delta$). Indeed, the first critical point calculated in $\varepsilon_{el}^i(\omega)$ is related to the transition from valence band maximum to the conduction band minimum, i.e. the energy band gap.
The green-solid and pink-dashed vertical arrows in Fig.~\ref{fig6} denote the optical gap and energy gap. In general, the optical gap is equal to the electronic band gap minus the exciton binding energy. In other words, the optical gap and electronic gap should be equal if the many-body perturbation theory is not taken into account. Therefore, our obtained difference between optical gap and electronic band gap is due to the smearing applied in the implemented Kubo-Greenwood approach~\cite{APL2018}. Moreover, note that $\varepsilon^r(\omega=0)$ in Figs. 7 and $\varepsilon_{el}^0$ listed in Table I, apparently, are different. However, the difference originates in the two different methods that we employed to calculate them. The first one is the microscopic dielectric function that obtained by using RPA and the second one is the macroscopic dielectric function which was obtained by using DFPT. Thus, they coincide if the same method are used to obtain them.
To compare the optical properties of 2D- and bulk ices, we demonstrated in Fig.~\ref{fig7} (Fig.~\ref{fig8}) the real (imaginary) dielectric functions.
The results show that:
\begin{itemize}
\item The real part of dielectric function has larger values in bulk ices as compared to 2D-ices in low energy region. This is due to the fact that electrons are distributed in 3D space in bulk ice. Moreover, the peaks are larger in 2D-ices as compared to the bulk ices.
\item The negative values of real dielectric function of bulk ices correspond to the largest electromagnetic wave reflection.
In fact the inductive properties will dominate in this range of energies with the negative values of real dielectric function. In simple words, at energy ranges (e.g. in Ice VIII~the energy range 15-20\,eV has the negative dielectric function) the electric displacement vector and the electric field vector have opposite directions.
\item There is a redshift (move toward lower energy region of the major peaks of $\varepsilon_{el}^i(\omega)$ for bulk ices in comparison to 2D-ices. In all studied ice, the prominent peaks in $\varepsilon_{el}^i(\omega)$ correspond to optical transmission which are mainly due to the interband transition from the $p$ valence bands to $s$ conduction bands (i.e. the O-$p$ orbital below Fermi level -valence band- and H-s orbital above Fermi level -conduction band-). PDOS of each system is shown in the RHS panels of Fig. 2.
\item The absorption energy ranges for 2D- and bulk ices are in the ultraviolet spectra ($>3.2 eV$) and visible spectra (between 2 and 3.2 eV), respectively.
\end{itemize}
The obtained dielectric functions for ice VIII, ice XI,ice Ic, ice Ih, and bulk water are consistent with the results of Refs.~[32,39], Ref.~[35], and Refs.~[40,41], respectively. Note that $\varepsilon^r(\omega=0)$ in Figs. 7 and $\varepsilon_{el}^0$ listed in Table I, apparently, are different. However, the difference originates in the two different methods that we employed to calculate each of them. The first one corresponds to the microscopic dielectric function which was obtained by using RPA and the second one corresponds to the macroscopic dielectric function which was obtained by using DFPT.
\begin{figure*}[]
\includegraphics[width=0.45\linewidth]{aep.png}
\includegraphics[width=0.45\linewidth]{bep.png}
\includegraphics[width=0.45\linewidth]{cep.png}
\includegraphics[width=0.45\linewidth]{dep.png}
\includegraphics[width=0.45\linewidth]{eep.png}
\includegraphics[width=0.45\linewidth]{fep.png}
\includegraphics[width=0.45\linewidth]{gep.png}
\includegraphics[width=0.45\linewidth]{hep.png}
\includegraphics[width=0.45\linewidth]{wep.png}
\caption{The real and imaginary part of dielectric function for
(a) f-SQ, (b) b-SQ, (c) b-RH, (d) HEX, (e) ice VIII, (f) ice XI, (g) ice Ic, (h) ice Ih, (i) bulk water. The green-solid (pink-dashed) arrows refer to
the optical gap (energy gap). }
\label{fig6}
\end{figure*}
\begin{figure*}[]
\includegraphics[width=0.9\linewidth]{rcompare.png}
\caption{The real dielectric function of 2D- and bulk ices: (a) in-plane ($xx,yy$ components) and (b) out-of-plane ($zz$ component).}
\label{fig7}
\end{figure*}
\begin{figure*}[]
\includegraphics[width=0.9\linewidth]{icompare.png}
\caption{The imaginary dielectric function of 2D- and bulk ices: (a) in-plane ($xx,yy$ components) and (b) out-of-plane ($zz$ component).}
\label{fig8}
\end{figure*}
\subsection{Mechanical stiffness of 2D-ice}
It is insightful to calculate mechanical stiffness of a typical 2D-ice (the results are selectively presented for f-SQ). We performed an additional calculation for determining the Young's modulus and Poisson's ratio of f-SQ ice. We applied both uniaxial and biaxial strains found the corresponding energies $E_u$ and $E_b$, for uniaxial and biaxial strained systems. Next, the best fits on two equations $E_b=2b u^2_b$ and $E_u=\frac{1}{2}(b+\mu)u^2_u$, helped us to obtain 2D-bulk modulus ``b" and 2D-shear modulus ``$\mu$". Here $u_u$ and $u_b$ are the applied uniaxial and biaxial strains, respectively. Consequently, the Young's modulus ($Y_{2d}$) and Poisson's ratio ($\nu_{2d}$) can be computed: ~\cite{zakhar}
\begin{equation}\label{Young}
Y_{2d}=\frac{4b\mu}{b+\mu} ~~and~~ \nu_{2d}=\frac{b-\mu}{b+\mu}
\end{equation}
The results are $Y_{f-SQ}$=3.6\,N/m and $\nu=0.6$. The Young's modulus of 2D-ice is two orders of magnitude smaller than the one for graphene, i.e. 340\,N/m~\cite{Lee}. This is due to the weak hydrogen bonds in 2D-ices compared to the strong planar covalent bonds in graphene. The corresponding Poisson's ratio of 2D-ice is larger than graphene ($~$0.3). The obtained Young's modulus (Poisson's ratio) for 2D-ice is more or less equal to bulk modulus of bulk ice. In Table II, we listed Young's modulus/bulk modulus and Poisson's ratio of f-SQ/bulk crystalline ice. To convert N/m unit in $Y_{2d}$ to Pas unit, one may use the simple relation $Y_{3d}=Y_{2d}/t_0$ where $t_0$=3\AA~is the effective thickness of f-SQ ice (see Table I).
\begin{table}
\centering
\caption{The Young's modulus and Poisson's ratio of f-SQ and bulk modulus (B) of three bulk crystalline ice.}\label{table3}
\begin{center}
\begin{tabular}{p{2cm}|p{3cm}|p{2cm}}
\hline
ice& \makebox[3cm][c]{$Y(GPa)$ ($B^*(Gpa))$} & \makebox[2cm][c]{$\nu$} \\[2.5mm]
\hline
f-SQ &\makebox[3cm][c]{12} &\makebox[2cm][c]{0.6} present work \\[2mm]
\hline
ice VIII~\cite{Klotz} &\makebox[3cm][c]{18$^{*}$}&\makebox[2cm][c]{-} \\[1.5mm]
ice VII~\cite{Klotz} &\makebox[3cm][c]{13$^{*}$} &\makebox[2cm][c]{0.28\cite{Shaw}} \\[1.5mm]
ice Ih~\cite{Klotz} &\makebox[3cm][c]{8.5$^{*}$} & \makebox[2cm][c]{0.325\cite{Shaw}}\\[1.5mm]
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\section{conclusions}
In summary, we found that the energy gap in f-SQ, b-SQ, and HEX 2D-ice structures and cubic bulk ices (ice VIII and ice Ic) is direct, whereas b-RH 2D-ice and hexagonal bulk ices (ice XI and ice Ih) have indirect band gap. Underlying lattice structure, symmetry significantly influences the ionic and dipolar terms, but its effects on the electronic dielectric constant are negligible.
We found the total out-of-plane dielectric constant is larger than 2 for all the studied 2D-ices (except b-SQ) and bulk ices, i.e. $\varepsilon_{total}^{zz}>$2.0 (see Table I). This clearly shows that the lattice structure of the confined water in recent experiment~\cite{science} is none of the lattice structure of the studied 2D-ices here, and has likely random structure. On the other hand the small out-of-plane dielectric constant of about $\simeq$2.1 (for nanoconfined water in channels
with heights $h$ $\sim$10\AA)~\cite{science}, should not correspond to a monolayer water.~Beyond $\sim$15\AA~a nonlinear increase (up to the bulk value) in the dielectric constant was found~\cite{science}. Therefore, we do not expect to recover experimental data when studying monolayer crystalline ice. It is also interesting to know that there is no reliable experimental data for the in-plane dielectric constant of confined water at sub-nanometer scale slit. Equivalently, the density of trapped water may be much lower than the bulk density~\cite{science}. This motivated us to determine the dielectric properties of amorphous 2D-ice in an ongoing study.\\The optical gap of 2D-ices is found to be larger than that of bulk ices. The absorption energy ranges for 2D- and bulk ices are in the ultraviolet spectra ($>3.2 eV$) and visible spectra (between 2 and 3.2\,eV), respectively. Generally, in bulk materials due to the presence of large number of atoms and merging bunch of adjacent energy levels results in the well-known energy conduction and valance bands. In 2D-materials, due to the smaller number of atoms, the number of energy level decreases giving the narrower energy bands. As a result, energy band gap will increase (the difference between valance band and conduction band). Also, the larger band gap in 2D-ice will cause a shift of absorption spectrum toward lower wave length (larger energies). In other words, there is redshift in the peaks of $\varepsilon^i$ of bulk ices in comparison to that of 2D-ices~\cite{Yiling}. \\We believe that our findings not only provide a theoretical background for understanding the different aspects of dielectric properties of confined water, but also gives insights into the light absorption mechanism and corresponding absorption energy range of confined water which might be necessary for further experimental characterizations of 2D-ices.\\
\\
\section{Acknowledgments}
We acknowledge Nassim Kangarlou for critically reading our paper and giving us useful comments. We would like to acknowledge high-performance computing support from Shahid Rajaee TT-University sponsored by Iran National Science Foundation (INSF).
| a93b95f88352500cb105f03444a4dad2dbc6c4bc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The dangerous dynamics of the coronavirus spread throughout the world gives rise to numerous studies in a wide scientific spectrum. Improving known epidemic models and developing new models are complicated tasks because the lack of verified statistics on the infection spread and disease dynamics. Unstable predictability of infection, ambiguity with drugs \cite{Corey:2020}, uncertainty regarding the immunity of disease \cite{Nature:2020}, and other factors (such as a viral mutation) make it difficult to predict the dynamics of pandemic.
This may relay to some mathematical models, which depend on a significant number of free (statistically-driven) parameters. The known models of the SIR family give solutions in a form of smooth functions\cite{Choisy:2007} (solutions of differential equations) that only indirectly include important random factors.
Naturally, that in such a situation, most statistically reliable forecasts are obtained by methods based on the direct application of the central limit theorem with a predicted error of $1/\sqrt N$, see \cite{Barmparis:2020a},\cite{Eugene:2020a},\cite{Fandaou:2020a} and references therein.
In this paper, we propose the use of the dynamic Monte Carlo (DMC) method that self-consistently includes various dynamic random factors. Such a technique was previously used to study the processes associated with aggregation, viscous flow properties, the formation of biological structures, and allows to scale the associated geometric and dynamic quantities that characterize these phenomena \cite{Vicsek:1995},\cite{Solon:2015},\cite{Solon:2015a}.
In our study, as a control (free) parameter, we use the generalized risk factor $\beta$, which includes some of the factors mentioned above in an integral form. In our 2D toy model, the transmission of infection occurs due to contacts of randomly moving individuals, that determines the complex dynamics of the infection spread and various critical aspects of such a dynamics.
The paper is organized as follows. In Section 2, we formulate our approach and examine the behavior of infected individuals (order parameter $A(t,\beta)$), which, as it turns out latter, critically depends on the value $\beta$. It also is discussed the similarity of the asymptotic behavior of the infection dynamics with the critical phase transition in a two-dimensional (2D) percolation system. In the next Section, we analyze the dynamic properties of the extended system, where we deal with two additional parameters which allow to on/off the quarantine state. The next Section, contains the study of dynamics of the infection spreading and the formation of immunity for infected individuals. The last Section summarizes our conclusions.
\section{Dynamic Monte Carlo simulations}
First we explain which the properties of dynamic Monte Carlo (DMC) approach we deal with. In order to study the infection dynamic in epidemiological system (that is far from equilibrium) the DMC method is used, that allows investigation both temporal and spatial properties by the numerical simulations.
As a toy model, we choose a 2D $L\times L$ (where $L$ is
size) bounded system that contains a disordered population of $N$ individuals.
Following the classifications commonly known from SIR model \cite{Choisy:2007} in our DMC model we divide the host population into a set of distinct categories,
according to its epidemiological status, that are susceptible (S), currently infectious (I ), and recovered (R). The total size of the host population is then $ N = S + I + R $ and all the individuals are born in the susceptible
category\cite{Choisy:2007}. Following the actual situations we assume that initially the maternally derived immunity is clear. (The effect of immunity is studied in the following sections). Upon contact with infectious individuals, the susceptibles may get infected and move into the infectious category. To apply the DMC approach it is constructed the Person class (individual, alias object) that encapsulates properties of a randomly placed and moving individual and contains the following significant attribute
\begin{equation}
\{x,y,v_{x},v_{y},I,M\},\label{attrib
\end{equation}
where $x,y$ are the components of position, $v_{x},v_{y}$ are the components
of velocity, the parameters $I,M$ describe the states: infected/uninfected
and immunized/non-immunized, respectively. The list of Persons that represents a total host population is used in our DMC simulations.
One of the underlying reasons why epidemiological systems exhibit variation is due to a
different way that the individuals in a population have contact with each other.
In our DMC simulation we assume that the spreading (transmission) of the infection occurs
because of random contacts for moving individuals.
To do that in DMC simulations we use the following strategy. (i) Any contact can occur only between two nearest individuals. (ii) At any contact, the state of an infected transmits to the other contact person. But the infected one can still be recovered with probability $1-\beta$ (recall that $\beta \in [0,1]$ is a risk factor). This means that if $\beta \lesssim 1$, the probability to a recovery is small.
To take an advantage of the visualizations at applying the DMC technique, we allow each object to have a visual representation, which is a yellow circle (non-immunized individual), a green circle (immunized but not infected individual) and a red circle (infected individual), see Fig.\ref{Pic_Fig1a}.
We used the interaction radius $r$ as the unit scale to measure distances (used $r = 6$, see Fig.\ref{Pic_Fig1a}) between the individuals, while the time unit $\Delta t = 1$ was the
time interval between two updatings of the directions and positions.
In our simulations we used the
simplest initial conditions: at time $t=0$ the positions and velocities for all the $N$ individuals are randomly distributed.
We use the velocity scale such that random $v_x, v_y \in [2,10]$ for which the
individuals always interact with their actual neighbors and move fast enough to change the configuration after a few
updates. According to our simulations, the variation of
actual interval of values of $v_x, v_y$ does not affect the results.
We also investigated the cases when the basic parameter of the model, the density
$\rho = N/L^2$ is slightly varied.
\begin{figure}[ptb
\centering
\includegraphics[
width=0.45\textwidth
{Fig1a.png
\caption{(Color on line.) The snapshots ($N=1000, \beta=0.95$, $L=400$) show the dynamics of the infection spreading at: (a) t=10, (b) t=30, (c) t=50, (d) t=70. We observe that for shown case at $t=70$ nearly all the individuals are infected.
\label{Pic_Fig1a
\end{figure}
When the simulation time runs a lot contacts occur between nearest randomly moving persons that leads to fast and uncontrollable transfer of infection between many individuals, see Fig. \ref{Pic_Fig1a}.
\begin{figure}[ptb
\centering
\includegraphics[
width=0.5\textwidth
{Fig2a.png
\caption{(Color on line.) The dynamics of the infections spreading
coefficient (order parameter) $A=I/N$ for times $t<8$ at different values of the risk factor (control parameter) $\beta$. [In this figure the abscissa axis shows the fitting time $0.01t$]
The blue
lines (arrows A) show the numerical simulations (DMC) data, while red lines
(arrows B) display the fitting function Eq.(\ref{fitting}), where only $a_{0}$ coefficient changes considerably at $\beta$\ variation: (a) shows
case $\beta=0.99$, (b) $\beta=0.90$, (c) $\beta=0.80$, (d) $\beta = 0.60$. At
$\beta<0.60$ the DMC solution rapidly converges to 0 ($A=0$). This means that for
$\beta<\beta_{c}\approx0.60$ all the infected individuals will be recovered up to $t=8$.
\label{Pic_Fig2a
\end{figure}
It is of great interest to investigate the temporal infection dynamics at
various risk factors $\beta$. Such a dynamics of the infections
spreading (coefficient $A(t)=I(t)/N$) as function of time $t$ is displayed in Fig.\ref{Pic_Fig2a}.
Since $A(t)$ is a random-valued function we will fit (see \cite{Press:2002}) $A(t)$ by a suitable
fitting function that is chosen as
\begin{equation}
f(t)=a_{0}t^{a_{1}}\tanh(t^{a_{2}}),\label{fitting
\end{equation}
where $a_{0,1,2}$ are the fitting coefficients. We found that $a_{1}$ is very small $\sim 10^{-5}$
and $a_{2}\simeq2$ for all the cases, but the amplitude $a_{0}$ changes considerably at
the $\beta$\ variation. In Fig.\ref{Pic_Fig2a} the blue lines
show the numerical simulations (DMC) data, while the red lines display the
fitting function $f(t)$. Fig.\ref{Pic_Fig2a} (a) shows the case $\beta=0.99$, (b) $\beta=0.90$,
(c) $\beta=0.80$, (d) $\beta=0.60$. We indicate a \textit{remarkable} observation that for $\beta<0.60$ the system asymptotically converges to a
trivial solution with $A\simeq a_0=0$ already for $t\simeq 6$
\begin{figure}[ptb
\centering
\includegraphics[
width=0.5\textwidth
{Fig3a.png
\caption{(Color on line.) The comparison of the order parameter functions for the infection spreading $A(\beta)$ and the order parameter for 2D percolating $P(p)$, where $p$ is is the occupation probability of defect state\cite{Grimmett:1999a}. Red line shows the dependence of (slightly tailored) amplitude parameter the fitting function $a_0$ corresponding
to $A=I/N$ at fixed time $t=8$ as function
of the risk factor $\beta$. Blue line shows the dependence of the order
parameters $P(p)$ for 2D percolating material as function of the occupation probability $p$. We observe that both curves are very close and the phase transitions to
infected/percolating state occurs similarly at $\beta_{c}\simeq p_{c}\simeq0.6$ for both cases.
\label{Pic_Fig3a
\end{figure}
Such observation leads to an interesting assumption that the studied dynamics of the infection spread can (asymptotically) be associated with a critical transition in the two-dimensional (2D) percolation system, that occurs when the occupation probability of defects is $p_{c}=0.594$ \cite{Isichenko:1992a},\cite{Grimmett:1999a},\cite{Stauffer:1992},\cite{Burlak:2009a},\cite{Burlak:2015} see Fig.\ref{Pic_Fig3a}. Such an assumption is studied in the following Section.
\section{Critical value of the risk factor}
Fig. \ref{Pic_Fig3a} displays a comparison of above mentioned dependencies. In Fig.
\ref{Pic_Fig3a} the red line shows the dependence $a_0(\beta)$ (see Fig.\ref{Pic_Fig2a}) associated with the infecting parameter $A(\beta)=I/N$, and the blue
line shows the percolating order parameters $P(p)$ as function of the
occupation defect probability $p$. We observe that both dependencies are in excellent agreement and at $\beta_{c}\simeq p_{c}\simeq0.6$ the phase transition to infected/percolating state occurs. From Fig. \ref{Pic_Fig3a} we can assume that the parameter
$A(\beta)$ can be mentioned further as an order parameter (similarly $P(p)$).
This results that the formalism of the percolation critical percolating phase transition \cite{Isichenko:1992a},\cite{Grimmett:1999a},\cite{Stauffer:1992}
can be applied to investigation the asymptotic of infection spreading at various $\beta$.
On the other hand, good agreement between the results of DMC modeling and the critical transition in 2D percolation system shows the general applicability of DMC approach to analyze the dynamics of infection spread in the epidemiological system.
\section{The extension of model}
\subsection{Quarantine regime}
Mass infection shown in Fig. \ref{Pic_Fig1a} is an extremely dangerous and highly undesirable scenario for the development of epidemiological situation. This Section discusses the extension of the model, which in principle allows
suspending this trend. One of the simple solutions proposed recently is introducing the quarantine by localizing of infected individuals in order to significantly reduce the number of contacts that leads to the transmission of infection. This can be modeled by setting $v_x=v_y=0$
for infected individuals and ignoring all the contacts with them in our approach.
We call such a regime of simulation as a quarantine mode. In order to do this we
introduce two new parameters into the model, $A_{\max}$ (the infection level
when the quarantine is automatically turned on), and $A_{\min}$ (the infection
level when the quarantine is turned off).
\begin{figure}[ptb
\centering
\includegraphics[
width=0.5\textwidth
{Fig4a.png
\caption{(Color on line.) The dynamics of infection parameter $A$ in quarantine mode with $A_{\max}=0.7$, $A_{\min}=0.36$ at moderate values $\beta$, (a) $\beta=0.78$, ( b) $\beta=0.80$, and (c) $\beta=0.82$; panel (d) shows the typical dynamics $A$ at initial times. We observe the generation of irregular oscillations of $A$ with large amplitudes between $A_{\max}$ and $A_{\min}$. We calculated (by the method \cite{Alanwolf:1985a}) that the Lyapunov exponent for such irregular oscillations is about $0.3$.
\label{Pic_Fig4a
\end{figure}
Fig. \ref{Pic_Fig4a} shows the dynamics of infections $A(t)$ in quarantine mode with $A_{\max}=0.7$, $A_{\min}=0.36$ at moderate values $\beta$, (a) $\beta=0.78$, (b) $\beta=0.80$, and (c) $\beta=0.82$; panel (d) shows the typical dynamics $A$ at initial times. For such parameters from Fig.\ref{Pic_Fig4a} we observe that the system transmits to unexpected dynamic state: the generation of irregular oscillations of $A$ with large amplitudes between $A_{\max}$ and $A_{\min}$. We have calculated (by the method \cite{Alanwolf:1985a}) that the Lyapunov exponent for such irregular oscillations is about $0.3$.
This means that if quarantine is turned off too early, the growth of infections suppressed, but the system goes into dynamic mode with irregular oscillations. In this case, a significant number of infected and recovered individuals can be re-infected, therefore, a full recovery does not occur.
\begin{figure}[ptb
\centering
\includegraphics[
width=0.5\textwidth
{Fig5a.png
\caption{(Color on line.) The same quarantine case as in Fig.\ref{Pic_Fig4a} but for large the risk factor $\beta$: (a) $\beta=0.88$, (b) $\beta=0.90$, (c) $\beta=0.92$, (d) $\beta=0.88$ for small times. We observe that for large $\beta$ the evolution of the infections has monotonic shape (with small random variations) but without the large oscillations shown in Fig.\ref{Pic_Fig4a}. One can see that the dynamics $A(t)$ in (b) for $\beta=0.90$ is already suppressed and considerably differs with respect to situation without the quarantine shown in Fig.\ref{Pic_Fig2a} (b).
\label{Pic_Fig5a
\end{figure}
Fig.\ref{Pic_Fig5a} shows the infections dynamics $A(t)$ in the quarantine mode but for large the risk factors $\beta$: (a) $\beta=0.88$, (b) $\beta=0.90$, (c) $\beta=0.92$, (d) $\beta=0.88$. We observe that for large $\beta$ the evolution of the infections has monotonic shape (with small random variations) but without the oscillations as in Fig.\ref{Pic_Fig4a}. However the dynamics $A(t)$ in (b) for $\beta=0.90$ is already suppressed and strongly differs with respect to a case without the quarantine shown in Fig.\ref{Pic_Fig2a} (b).
\subsection{The immunity}
Although actually there are no reliable statistics for the congenital or acquired immunity for persons (for animals see Ref.\cite{Chandrashekar:2020}), in this Section we analyze this aspect in framework of our model. Following the Ref.\cite{Choisy:2007}, we assume that in the host population there is no innate immunity to the virus. But we suppose that the persons (at least a large majority) will acquire this immunity, as is usually the case. To this end, in our model we use the parameter $M$, see Eq. (\ref{attrib}). Following \cite{Choisy:2007}, we assume that this parameter acquires a non-zero value (the presence of immunity) only after first infection and recovery. Re-infection no longer occurs even at contacts with infected persons.
Fig. \ref{Pic_Fig6a} shows the dynamics of recovery at presence of immunity in
the quarantine mode for fixed parameters $\beta=0.94$ and $A_{\max}=0.24$ and
different $A_{\min}=0.17, 0.1, 0.05, 0.02, 0.01$. One can see that now the
oscillations shown in Fig.\ref{Pic_Fig6a} acquire shape of strongly damped
picks that results the number of infected (order parameter $A$) to rapidly decrease. This
allows to predicts that after the first high pick of infection (that has nearly fixed amplitude for all the cases) may occur a second pick but with lesser amplitude and then the complete recover may become
\begin{figure}[ptb
\centering
\includegraphics[
width=0.5\textwidth
{Fig6a.png
\caption{(Color on line.) The dynamics of recovery in presence of immunity
in the quarantine mode for the parameters $\beta = 0.94 $, $ A_{\max} = 0.24 $ for different $A {\min} = 0.17, 0.1, 0.05, 0.02, 0.01$. One can see that after the high peak, the oscillations rapidly decay that leads to a decrease of infections (the order parameter $A(t) $ rapidly decreases).
\label{Pic_Fig6a
\end{figure}
Now we compare the effects of quarantine and immunity factors for
recovery. Fig.\ref{Pic_Fig7a} shows the dynamics of infections (order parameter $A$) for different
values of the risk factor $\beta=0.99, 0.94, 0.9, 0.8$ at situation without the quarantine
when only the personal immunity $M>0$ presents (see Eq.(\ref{attrib})). This
simulation shows that in such case the complete recover can occur even for a lesser time comparing to Fig.\ref{Pic_Fig6a}.
\begin{figure}[ptb
\centering
\includegraphics[
width=0.5\textwidth
{Fig7a.png
\caption{(Color on line.) The dynamics of infections (order parameter $A$) for different
values of the risk factor $\beta=0.99,0.94,0.9,0.8$ for situation when only the effective
personal immunity $M>0$ presents (without the quarantine), see Eq.(\ref{attrib}). This
simulation shows that in such case the complete recover can occur for a lesser time comparing to Fig.\ref{Pic_Fig6a}.
\label{Pic_Fig7a
\end{figure}
\section{Discussion and Conclusion}
We studied the dynamics of the infection spread at various values of the risk factors $\beta$ (control parameter) using the dynamic Monte Carlo method (DMC). In our model, it is accepted that the infection is transmitted through the contacts of randomly moving individuals. We show that the behavior of recovered individuals critically dependents on
the value $\beta$. For sub-critical values $\beta<\beta_{c}\sim 0.6$, the number of infected cases (the order parameters $A(t)$) asymptotically converges to zero, so that at moderate risk factor, the infection can quickly disappear.
\begin{figure}[ptb
\centering
\includegraphics[
width=0.5\textwidth
{Fig8a.png
\caption{(Color on line.) The fraction of infections $A(t,\beta)$ as function of time $t$ at different risk factors $\beta$ near the critical transition at $\beta_{c} \sim 0.6$ for $N=1000$ and initial number of infections $I_0=100$, with $\beta$: (a) 0.56,(b) 0.57,(c) 0.58,(d) 0.581. We observe that for $\beta \lesssim 0.58$ the number of infections rapidly reach zero. However for $\beta>0.8$ the process of recovering may be long.}
\label{Pic_Fig8a
\end{figure}
However such a nontrivial behavior has to be confirmed by direct calculation. Fig. \ref{Pic_Fig8a} shows the dynamics of infections fraction $A(t,\beta)$ with time for different risk factors $\beta$ near the critical transition $\beta \sim \beta_{c}=0.6$ for $N=1000$ and rather large the initial number of infections $I_0=100$. We observe that really for $\beta \lesssim 0.58$ the number of infections rapidly reach zero. We also analyzed the extended system, which currently is widely used to prevent the spread of the virus. In our approach such a system includes two additional parameters on/off the quarantine state. It was revealed that early exit from the quarantine leads to irregular oscillating dynamics (with positive Lyapunov exponent) of the infection. However when the lower limit of the quarantine off is sufficiently small, the infection dynamics acquires a characteristic nonmonotonic shape with several damped peaks. The dynamics of the infection spread in case of individuals with immunity was studied too. Our comparison of the quarantine and the immunity factors on a recovery shows that in case of stable immunity a complete recovery occurs faster than in a quarantine mode.
\section{Acknowledgment}
This work was supported in part by CONACYT (M\'{e}xico) under the grant No. A1-S-9201.
| 5567f06ffc4eefc1678d85b947e8691ecaa1026f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
At least 1.5 million preventable adverse drug events (ADE) occur each year in the U.S.~\cite{IoM2006}.
Medication errors related to ADEs can occur while writing or filling prescriptions, or even when taking or managing medications~\cite{aronson2009medication}. For example, pharmacists must verify thousands of pills dispensed daily in the pharmacy, a process that is largely manual and prone to error. Chronically-ill and elderly patients often separate pills from their original prescription bottles, which can lead to confusion and misadministration. %
Many errors are preventable; however, pills with visually similar characteristics are difficult to identify or distinguish, increasing error potential.%
The pill images available for vision-based approaches fall into two categories: reference and consumer images (Figure~\ref{fig:ref_cons_examples}). The reference images are taken with controlled lighting and backgrounds, and with professional equipment. For most of the pills, one image per side (two images per pill type) is available from the National Institutes of Health (NIH) Pillbox dataset~\cite{pillbox}. The consumer images are taken with real-world settings including different lighting, backgrounds, and equipment. Building pill-image datasets, especially for prescription medications, is costly and has additional regulatory obstacles as prescription medications require a clinician's order to be dispensed. A labeled dataset is scarce and even unlabeled images are hard to collect. This setup requires the model to learn representations from very few training examples. Existing benchmarks are either not publicly available or smaller scale, which are not suitable for developing real-world pill identification systems.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.8\linewidth]{ref_cons_images-crop.pdf}
\end{center}
\caption{\textbf{Representative examples of reference images (top row) and corresponding consumer images (bottom row)} Both sides (front and back) of the two pill types (NDC:00093-7180 and NDC:50111-0441, respectively) are shown.
}
\label{fig:ref_cons_examples}
\end{figure}
Over the years, deep learning has achieved unprecedented performance on image recognition tasks with efforts from both large-scale labeled datasets~\cite{imagenet,zhou2014learning} and modeling improvements~\cite{ResNet50CVPR2015,krizhevsky2012imagenet,simonyan2014very}; however, fine-grained visual categorization (FGVC) is still a challenging task, which needs to distinguish subtle differences within visually similar categories. FGVC tasks mainly include natural categories (e.g., birds~\cite{van2015building,wah2011caltech}, dogs~\cite{khosla2011novel} and plants~\cite{nilsback2008automated,wegner2016cataloging}) and man-made objects (e.g., cars~\cite{krause20133d,yang2015large} and airplanes~\cite{maji2013fine}). In this work, we target the pill identification task, which is an under-explored FGVC task yet important and frequent in healthcare settings. The distributions of pill shape and color are skewed to certain categories (Figure~\ref{fig:shape_color_heatmap}), highlighting the importance of distinguishing subtle differences such as materials, imprinted text and symbols.
The main contribution of this paper is introducing ePillID, a new pill identification benchmark with a real-world low-shot recognition setting. Leveraging two existing NIH datasets, our benchmark is composed of 13k images representing 9804 appearance classes (two sides for 4902 pill types). This is a low-shot fine-grained challenge because (1) for most of the appearance classes there exist only one image and (2) many pills have extremely similar appearances. %
Furthermore, we empirically evaluate various approaches with the benchmark to serve as baselines.
The baseline models include standard image classification approaches and metric learning based approaches.
Finally, we present error analysis to motivate future research directions.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\linewidth]{pill_shape_color_heatmap-crop.pdf}
\caption{\textbf{The distribution of shape and color of pills.} For pills registered with multiple colors, we grouped them as OTHER. The combinations less than $0.1\%$ are omitted.}
\label{fig:shape_color_heatmap}
\end{figure}
\section{Related Work}
\paragraph{Image-based Pill Identification:}
In 2016, the NIH held a pill image recognition challenge~\cite{yaniv2016national}, and released a dataset composed of 1k pill types. The challenge winner~\cite{zeng2017mobiledeeppill} proposed a deep similarity-based approach~\cite{DeepRankingCVPR2014} and developed a mobile-ready version using knowledge-distillation techniques~\cite{hinton2015distilling}. Following the competition, classification-based approaches were applied, with higher recognition performance reported~\cite{delgado2019fast, szegedy2015going, wong2017development}. Aside from deep neural networks, various approaches with feature engineering have been proposed~\cite{caban2012automatic,cunha2014helpmepills,lee2012pill,neto2018cofordes,wong2017development}. For instance, the Hu moment~\cite{hu1962visual} was applied~\cite{cunha2014helpmepills,lee2012pill}, because of its rotation-invariant properties. Other methods~\cite{chen2013new,suntronsuk2016pill,yu2014pill,yu2015accurate} were proposed to generate imprint features, recognizing the importance of imprints for pill identification. The past methods achieved remarkable success; however, the lack of benchmarks prevents us from developing methods for real-world low-shot settings. Wong \textit{et al.}~\cite{wong2017development} created a 5284-image dataset of 400 pill types with a QR-like board to rectify geometric and color distortions. Yu \textit{et al.}~\cite{yu2014pill} collected 12,500 images of 2500 pill categories. Unfortunately, neither of these datasets are publicly available.
\paragraph{Fine-Grained Visual Categorization (FGVC):}
In many FGVC datasets~\cite{bossard2014food,beery2019iwildcam,kaur2019foodx,cui2018large}, the number of categories is not extremely large, often less than 200. Recent large-scale datasets ~\cite{imatproduct,fieldguide,van2018inaturalist} offer large numbers of categories with many images (e.g., \num{675170} training images for 5089 categories~\cite{van2018inaturalist}) with challenging long-tailed distributions. Compared to other FGVC benchmarks, the data distribution on the ePillID benchmark imposes a low-shot setting (one image for most of the classes) with a large number of classes (8k appearance classes). Among many algorithms~\cite{yang2018learning,wang2018learning,zheng2019looking} proposed for FGVC tasks, bilinear models~\cite{kong2017low,li2018towards,zheng2019learning} achieved remarkable performances by capturing higher-order interactions between feature channels. B-CNN~\cite{lin2015bilinear} is one of the first approaches, which obtains full bilinear features by calculating outer product at each location of the feature map, followed by a pooling across all locations; however, the full bilinear features can be very high dimensional (e.g., over 250k when the input has 512 channels). Compact Bilinear Pooling (CBP)~\cite{gao2016compact} addresses the dimensionality issue by approximating bilinear features with only a few thousand dimensions, which was shown to outperform B-CNN in few-shot scenarios.
Another line of work is metric learning~\cite{qian2015fine,cui2016fine,sun2018multi}, where an embedding space that captures semantic similarities among classes and images is learned. Metric learning has been also successfully used in few- and low-shot settings~\cite{wang2019cvpr,sungflood2018cvpr}, making it suitable for our ePillID benchmark. %
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.8\linewidth]{ePillID_diagram-crop.pdf}
\end{center}
\caption{\textbf{The distribution of reference and consumer images on the ePillID benchmark.} The reference images are available for 4902 pill types, while the consumer images are only available for 960 pill types. The dataset includes 9804 reference images (two images, front and back, for each pill type) and 3728 consumer images.}
\label{fig:ePillID}
\end{figure}
\section{ePillID Benchmark}
We construct a new pill identification benchmark, ePillID, by leveraging the NIH challenge dataset~\cite{NIHpillcomp} and the NIH Pillbox dataset~\cite{pillbox}. We use the challenge dataset, which offers consumer images, as a base dataset and extend it with the reference images from the Pillbox dataset. In total, the ePillID dataset includes 3728 consumer images for 1920 appearance classes (two sides for 960 pill types) and 9804 reference images (two sides for 4902 pill types). This requires a fine-grained low-shot setup, where models have access to one reference image for all the 9804 appearance classes; however, there only exists a few consumer images for 1920 appearance classes (Figure~\ref{fig:ePillID}).
\paragraph{Experimental Setup:}
The consumer images are split into 80\% training and 20\% holdout sets in such a way that the pill types are mutually exclusive. The training set is further split on pill types for 4-fold cross-validation. The models have access to reference images for all of the 4902 pill types, but consumer images are unavailable for most pill types during training. To evaluate the performance in situations where both front- and back-sides are available as input, we construct two-sided queries by enumerating all possible consumer image pairs for each pill type.
\paragraph{Evaluation Metrics:}
For each query, a model calculates an ordered list of pill types with confidence scores.
Note that, for the experiments with both sides of a pill, a query consists of a pair of front and back images.
We consider Mean Average Precision (MAP) and Global Average Precision (GAP) for the model performance evaluation. %
For MAP, the average precision score is calculated separately for each query, and the mean value is calculated. MAP measures the ability to predict the correct pill types given queries. For GAP, all the query and pill-type pairs are treated independently, and the average precision score is calculated globally. GAP measures both the ranking performance and the consistency of the confidence scores i.e.\ the ability to use a common threshold across different queries~\cite{perronnin2009family}. We also calculate MAP@1 and GAP@1, where only the top pill type per query is considered. %
\begin{table}[!t]
\centering
\scalebox{0.65}{
\begin{tabular}{lcccc}
\toprule
Model & GAP & GAP@1 & MAP & MAP@1 \\
\toprule
\multicolumn{5}{c}{\bf{Plain Classification: Both-sides input}} \\
\midrule
DenseNet121 & $30.05 \pm 2.04$ & $38.53 \pm 4.42$ & $60.03 \pm 1.05$ & $41.21 \pm 1.89$ \\
DenseNet201 & $33.30 \pm 1.27$ & $42.64 \pm 3.40$ & $64.38 \pm 0.85$ & $46.61 \pm 1.11$ \\
DenseNet161 & $34.78 \pm 2.80$ & $44.47 \pm 4.26$ & $64.87 \pm 2.27$ & $46.85 \pm 2.72$ \\
DenseNet161 B-CNN & $36.35 \pm 1.67$ & $47.08 \pm 2.16$ & $67.07 \pm 1.36$ & $50.84 \pm 2.47$ \\
DenseNet161 BCP & $39.47 \pm 1.90$ & $48.72 \pm 1.92$ & $69.77 \pm 1.01$ & $52.72 \pm 1.42$ \\
\midrule
ResNet18 & $24.13 \pm 1.99$ & $33.13 \pm 2.34$ & $54.70 \pm 2.13$ & $35.70 \pm 2.79$ \\
ResNet34 & $29.07 \pm 2.45$ & $38.53 \pm 2.54$ & $58.88 \pm 2.26$ & $40.34 \pm 2.01$ \\
ResNet50 & $35.16 \pm 3.46$ & $45.43 \pm 4.22$ & $65.79 \pm 1.85$ & $48.32 \pm 2.34$ \\
ResNet50 B-CNN & $38.15 \pm 2.73$ & $48.13 \pm 4.17$ & $68.68 \pm 2.23$ & $51.92 \pm 3.23$ \\
ResNet50 CBP & $38.97 \pm 2.74$ & $48.63 \pm 3.94$ & $68.35 \pm 3.08$ & $51.60 \pm 4.24$ \\
ResNet101 & $36.15 \pm 3.21$ & $46.96 \pm 5.29$ & $65.50 \pm 2.95$ & $48.20 \pm 5.02$ \\
ResNet152 & $39.57 \pm 1.23$ & $49.97 \pm 1.58$ & $68.51 \pm 1.09$ & $51.64 \pm 2.31$ \\
ResNet152 B-CNN & $44.75 \pm 2.11$ & $55.14 \pm 3.67$ & $72.65 \pm 1.69$ & $56.75 \pm 2.51$ \\
ResNet152 CBP & $ 41.45 \pm 2.15$ & $52.43 \pm 2.07$ & $69.12 \pm 2.10$ & $52.64 \pm 2.29$ \\
\midrule
\multicolumn{5}{c}{\bf{Multi-head Metric Learning: Both-sides input}} \\
\midrule
DenseNet121 & $74.04 \pm 2.32$ & $90.15 \pm 2.63$ & $93.42 \pm 0.42$ & $87.58 \pm 0.87$ \\
DenseNet201 & $74.19 \pm 2.08$ & $87.69 \pm 2.58$ & $92.52 \pm 0.91$ & $86.42 \pm 1.56$ \\
DenseNet161 & $76.60 \pm 2.35$ & $88.86 \pm 2.88$ & $93.41 \pm 0.72$ & $87.78 \pm 1.55$ \\
DenseNet161 B-CNN & $75.27 \pm 4.65$ & $88.24 \pm 2.83$ & $93.11 \pm 0.81$ & $87.22 \pm 1.57$ \\
DenseNet161 BCP & $77.15 \pm 2.89$ & $89.11 \pm 2.36$ & $94.03 \pm 1.03$ & $88.98 \pm 1.78$ \\
\midrule
ResNet18 & $66.58 \pm 1.91$ & $87.87 \pm 2.86$ & $91.16 \pm 0.75$ & $84.38 \pm 1.36$ \\
ResNet34 & $73.61 \pm 3.17$ & $89.22 \pm 3.35$ & $92.74 \pm 1.95$ & $87.10 \pm 2.96$ \\
ResNet50 & $78.27 \pm 2.35$ & $90.71 \pm 2.55$ & $94.02 \pm 0.66$ & $89.06 \pm 1.05$ \\
ResNet50 B-CNN & $79.95 \pm 1.96$ & $90.26 \pm 1.56$ & $94.59 \pm 0.23$ & $89.66 \pm 0.62$ \\
ResNet50 CBP & $78.44 \pm 1.48$ & $91.36 \pm 2.83$ & $95.27 \pm 0.43$ & $91.01 \pm 0.97$ \\
ResNet101 & $79.83 \pm 2.11$ & $92.23 \pm 1.12$ & $94.99 \pm 0.40$ & $90.65 \pm 0.60$ \\
ResNet152 & $78.63 \pm 2.75$ & $90.54 \pm 1.96$ & $95.71 \pm 0.83$ & $91.93 \pm 1.56$ \\
ResNet152 B-CNN & $80.45 \pm 1.09$ & $89.61 \pm 1.83$ & $95.01 \pm 0.43$ & $90.65 \pm 0.81$ \\
ResNet152 CBP & $81.20 \pm 1.47$ & $91.19 \pm 0.28$ & $95.76 \pm 0.40$ & $92.01 \pm 0.67$ \\
\midrule
\multicolumn{5}{c}{\bf{Multi-head Metric Learning: Single-side input}} \\
\midrule
ResNet34 & $54.70 \pm 2.58$ & $78.69 \pm 2.37$ & $80.52 \pm 1.76$ & $70.61 \pm 2.11$ \\
ResNet50 & $61.75 \pm 1.40$ & $81.63 \pm 2.15$ & $82.34 \pm 0.79$ & $72.56 \pm 1.35$ \\
ResNet101 & $63.93 \pm 1.71$ & $84.51 \pm 1.68$ & $84.18 \pm 1.23$ & $75.72 \pm 1.48$ \\
ResNet152 & $64.61 \pm 2.45$ & $82.67 \pm 1.17$ & $85.25 \pm 0.91$ & $76.42 \pm 1.54$ \\
\bottomrule
\end{tabular}
}
\caption{\textbf{Recognition results on the ePillID benchmark.} Mean and standard deviations of the holdout metrics from the 4-fold cross-validation are reported in percentages. }
\label{tab:results}
\end{table}
\section{Experiments}
\label{sec:experiments}
We first introduce our baseline approaches, then present quantitative and qualitative results.
\subsection{Baseline Models}
We use ResNet~\cite{ResNet50CVPR2015} and DenseNet~\cite{densenet} as base networks pretrained on ImageNet~\cite{imagenet} for initial weights. In addition to the global average pooling layer as features, we evaluate two bilinear methods, B-CNN~\cite{lin2015bilinear} and CBP~\cite{gao2016compact}, applied to the final pooling layer of the base network. For CBP, we use their Tensor Sketch projection with 8192 dimensions, which was suggested for reaching
close-to maximum accuracy. We insert a 1x1 convolutional layer before the pooling layer to reduce the dimensionality to 256.
\paragraph{Plain Classification:}
As a first set of baselines, we train the models with the standard softmax cross-entropy loss. For regularization, we add a dropout layer~\cite{srivastava2014dropout} with probability 0.5 before the final classification layer. We use the appearance classes as target labels during training, and take the max of the softmax scores for calculating pill-type confidence scores. For the two-sides evaluation, the mean confidence score is used for a score between a two-sided input and a pill type.
\paragraph{Multi-head Metric Learning:}\label{sec:metric_losses}
As another set of baselines, we employ a combination of four losses to learn an embedding space optimized for fine-grained low-shot recognition.
The set up is a multi-task training procedure:
\begin{equation}
L_{final} = \lambda_{SCE} L_{SCE} + \lambda_{\eta} L_{\eta} + \lambda_{\rho} L_{\rho} + \lambda_{\Gamma} L_{\Gamma},
\end{equation}
where $L_{SCE}$ indicates softmax cross-entropy, $L_{\eta}$ cosine-softmax loss (ArcFace~\cite{deng2019cvpr}), $L_{\rho}$ triplet loss~\cite{schroff2015cvpr}, and $L_{\Gamma}$ contrastive loss~\cite{chopra2005learning}.
The loss weights, $\lambda_{SCE}$, $\lambda_{\eta}$, $\lambda_{\rho}$, and $\lambda_{\Gamma}$
are chosen empirically with a ResNet50 model~(Section~\ref{sec:impl}).
In order to compute the loss for every mini-batch, the triplet and contrastive loss requires additional sampling and pairing procedures. We apply online hard-example mining to find informative negatives for the triplets and pairs respectively evaluated by these losses. %
The trained model is used for generating embeddings for the query consumer images and all the reference images. We calculate cosine similarities between the query and reference embeddings and use them as confidence scores.
\paragraph{Implementation Details:}\label{sec:impl}
The images are cropped and resized to $224 \times 224$. Extensive data augmentation is employed to mimic consumer-image-like variations, including rotation and perspective transformation. The Adam optimizer~\cite{adam2014} is used with an initial learning rate of $\num{1e-4}$. The learning rate is halved whenever a plateau in the validation GAP score is detected. The model hyper-parameters are chosen to optimize the average validation GAP score using the 4-fold cross-validation. The loss weights for the metric learning are determined based on the ResNet50 experiment~(Table~\ref{tab:loss_weights}). The training is done with the mini-batch of 48 images in a machine equipped with Intel Xeon E5-2690, 112GB RAM, one NVIDIA Tesla P40, CUDA 8.0 and PyTorch 0.4.1.
\begin{table}[t!]
\centering
\scalebox{0.65}{
\begin{tabular}{ccccc}
\toprule
Contrastive~($\lambda_{\Gamma}$) & Triplet~($\lambda_{\rho}$) & SCE~($\lambda_{SCE}$) & CSCE~($\lambda_{\eta}$) & Validation GAP \\
\midrule
1.0 & 1.0 & 1.0 & 0.1 & $77.18 \pm 1.90$ \\
1.0 & 1.0 & 1.0 & 0.3 & $74.07 \pm 2.29$ \\
1.0 & 1.0 & 1.0 & 0.0 & $75.48 \pm 1.02$ \\
1.0 & 0.0 & 1.0 & 0.0 & $74.40 \pm 2.93$ \\
0.0 & 1.0 & 1.0 & 0.0 & $73.89 \pm 2.16$ \\
1.0 & 1.0 & 0.0 & 0.1 & $71.38 \pm 4.52$ \\
0.0 & 0.0 & 1.0 & 0.0 & $61.78 \pm 3.24$ \\
\bottomrule
\end{tabular}
}
\caption{\textbf{Validation GAP scores from the corss-validation for comparing the metric-learning loss weights.}
The ResNet50 baseline model is used.}
\label{tab:loss_weights}
\end{table}
\subsection{Quantitative Results}
In Table~\ref{tab:results}, we report the baseline results on the ePillID benchmark. The plain classification baselines performed much worse than the metric learning baselines, suggesting the difficulty of the low-shot fine-grained setting. The performance differences among the models are consistent with the ImageNet pretraining performance~\cite{bianco2018benchmark}. The multi-head metric leraning models performed remarkably well, achieving over 95\% MAP and 90\% GAP@1. In most of the cases, bilinear pooling methods outperformed the global average pooling counterpart, showing the representation power of the bilinear features. As an ablation study, we report single-side experiment results i.e.\ only one image per query. The ResNet152 metric learning approach achieved over 85\% MAP and 82\% GAP@1; however, the results indicate that both sides are required for accurate identification.
\subsection{Qualitative Results}
Figure~\ref{fig:qualitative} depicts qualitative comparisons from examples of the ePillID holdout dataset with confidence scores. In (a), the plain classification approaches misclassified, whereas the metric learning approaches identified successfully, even with the challenging lighting and background variations in the query images. In (b), only the metric learning with CBP approach identified correctly, suggesting CBP was effective for capturing the small difference in the imprinted text. For (c) and (d), all the four models failed to identify the correct types. In (c), the consumer images are affected by the lighting variations with the shiny pill material. In (d), the pill types share extremely similar appearances, except the one character difference in the imprinted text.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.9\linewidth]{qualitative_vl3-crop.pdf}
\end{center}
\caption{\textbf{Qualitative results from the ePillID holdout dataset.} For each query, top confidence scores are shown from the four ResNet152 baselines. The reference images of \textbf{correct} and \underline{wrong} pill types are shown with the input consumer images in the left.}
\label{fig:qualitative}
\end{figure}
\section{Conclusion}
We introduced ePillID, a low-shot fine-grained benchmark on pill image recognition. To our knowledge, this is the first publicly available benchmark that can be used to develop pill identification systems in a real-world low-shot setting. We empirically evaluated various baseline models with the benchmark. The multi-head metric learning approach performed remarkably well; however, our error analysis suggests that these models still cannot distinguish confusing pill types reliably.
In the future, we plan to integrate optical character recognition~(OCR) models. OCR integration has been explored for storefronts and product FGVC tasks~\cite{karaoglu2016words,bai2018integrating} and recent advances in scene text recognition are promising~\cite{karatzas2015icdar,gomez2017icdar2017,liu2018fots}; however, existing OCR models are unlikely to perform reliably on pills as they stand. Challenging differences include low-contrast imprinted text, irregular-shaped layouts, lexicon, and pill materials such as capsules and gels.
Finally, we plan to extend the benchmark further with more pill types and images as we collect more data. By releasing this benchmark, we hope to support further research in this under-explored yet important task in healthcare.
{\small
\bibliographystyle{ieee_fullname}
| 72a6670dc0a764892662b5d1a0eac1fad556c45e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{introduction}
The chemically peculiar (CP) stars of the upper main sequence (spectral types early B to early F) are traditionally characterised by the presence of certain absorption lines of abnormal strength or weakness that indicate peculiar surface abundances \citep{preston74}. For most groups of CP stars, current theories ascribe the observed chemical peculiarities to the interplay between radiative levitation and gravitational settling (atomic diffusion) \citep{michaud70,richer00}: whereas most elements sink under the force of gravity, those with numerous absorption lines near the local flux maximum are radiatively accelerated towards the surface. Because CP stars are generally slow rotators and boast calm radiative atmospheres, atomic diffusion processes are able to significantly influence the chemical composition of the outer stellar layers.
Following \citet{preston74}, CP stars are traditionally divided into the following four main groups: CP1 stars (the metallic-line or Am/Fm stars), CP2 stars (the magnetic Bp/Ap stars), CP3 stars (the Mercury-Manganese or HgMn stars), and CP4 stars (the He-weak stars). Although the chemical composition within a group may vary considerably, each group is characterised by a distinct set of peculiarities. The CP1 stars show underabundances of Ca and Sc and overabundances of the iron-peak and heavier elements. CP2 stars exhibit excesses of elements such as Si, Sr, Eu, or the rare-earth elements. The CP3 stars are characterised by enhanced lines of Hg and Mn and other heavy elements, whereas the main characteristic of the CP4 stars is anomalously weak He lines. Further classes of CP stars have been described, such as the He strong stars -- early B stars with anomalously strong He lines --, the $\lambda$ Bootis stars \citep{murphy17}, which boast unusually low surface abundances of iron-peak elements, or the barium stars, which are characterised by enhancements of the \textit{s}-process elements Ba, Sr, Y, and C \citep{bidelman51}. Generally, with regard to the strength of chemical peculiarities, a continuous transition from normal to peculiar stars is observed \citep{loden87}.
Most of the CP2 and He-peculiar stars possess stable and globally organised magnetic fields with strengths of up to several tens of kG \citep{babcock47,auriere07}, the origin of which is still a matter of some controversy \citep{moss04}. However, a body of evidence has been built up that strongly favours the fossil field theory, which states that the magnetic field is a relic of the 'frozen-in' interstellar magnetic field (e.g. \citealt{braithwaite04}). These stars are often referred to as magnetic chemically peculiar (mCP) stars in the literature -- a convention which we will adhere to throughout this paper. The magnetic field affects the diffusion processes in such a way that mCP stars show a non-uniform distribution of chemical elements (chemical spots or belts) on their surfaces, which can be studied in detail via the technique of Doppler imaging \citep{kochukhov17}. As the magnetic axis is oblique to the rotation axis (oblique rotator model; \citealt{stibbs50}), mCP stars show strictly periodic light, spectral, and magnetic variations with the rotation period. The photometric variability arises because flux is redistributed in the abundance patches (e.g. \citealt{wolff71,molnar73,krticka13}).
The mCP stars show vastly differing abundance patterns. Some of the most peculiar objects belong to this group, such as the extreme lanthanide star HD\,51418 \citep{jones74} or Przybylski's star HD\,101065 (\citealt{przybylski66}), which is widely regarded as the most peculiar star known. Excesses of several orders of magnitude are commonly observed in these objects. \citet{morgan33} already recognised a relationship between a CP2 star's temperature and the predominant spectral peculiarities and showed that the CP2 stars can thus be sorted into subgroups. Since then, many authors have proposed corresponding classification schemes with varying levels of detail (cf. the discussions in \citealt{wolff83} and \citealt{gray09}). It is generally useful to at least differentiate between the 'cool' CP2 stars mostly characterised by Sr, Cr and Eu peculiarities and the 'hot' CP2 stars that generally show Si overabundances, although considerable overlap exists.
The mCP stars are important to astrophysics in several respects. Their complex atmospheres lend themselves perfectly to the investigation of such diverse phenomena as atomic diffusion, magnetic fields, stellar rotation and their interplay. They furthermore provide important testing grounds for the evaluation of model atmospheres \citep{krticka09} and, through their characteristic light variability, allow the derivation of rotational periods with great accuracy and comparatively little effort.
\begin{figure}
\includegraphics[width=\columnwidth]{5200A_spectra_comparison.eps}
\caption{4800\,\AA\ to 5700\,\AA\ region of (from top to bottom) the non-CP A0 V star LAMOST\,J194655.00+402559.5 (HD\,225785), a synthetic spectrum with ${T}_{\rm eff}$\,=\,9750\,K, $\log g$\,=\,4.0, [M/H]\,=\,0.0 and a microturbulent velocity of 2\,km\,s$^{-1}$, and the newly-identified Si-strong mCP star LAMOST\,J025951.09+540337.5 (\#78; TYC\,3701-157-1). The position of the characteristic 5200\,\AA\ depression and the Si II lines at 5041\,\AA\ and 5055/56\,\AA\ are indicated. LAMOST spectra have been taken from DR4.}
\label{fig_5200A_depression}
\end{figure}
The most up-to-date collection of CP stars -- the most recent version of the General Catalogue of CP Stars -- was published a decade ago \citep{RM09}. It contains about 3500 mCP stars or candidates ($\sim$2000 confirmed mCP stars and $\sim$1500 candidate mCP stars). Since then, several studies have identified new mCP stars on a minor scale (e.g. \citealt{huemmerich18,Scholz2019,sikora19}) but no large scale spectroscopic surveys have been conducted during the past recent decades that aim specifically at the identification of new mCP stars 'en masse'.
The works of \citet{hou15} and \citet{qin19} warrant special mention as they exploited spectra collected by the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) of the Chinese Academy of Science. \citet{hou15} presented a list of 3537 candidate CP1 stars from LAMOST Data Release (DR) 1. Building on this work, \citet{smalley17} investigated pulsational properties versus metallicism in this subgroup of CP stars. \citet{qin19} searched for CP1 stars in the low-resolution spectra of LAMOST DR5 and compiled a catalogue of 9372 CP1 stars. They identified CP2 stars as a contaminant and searched for corresponding candidates in their sample of CP1 candidates, identifying 1131 candidate CP2 stars in this process.
Here we present our efforts aimed at identifying new mCP stars using spectra from the publicly available LAMOST DR4, which have led to the discovery of 1002 mCP stars. With this work, we significantly increase the sample size of known Galactic mCP stars, paving the way for future in-depth statistical studies. Spectroscopic data and target selection process are discussed in Section \ref{dataanalysis}. Spectral classification workflow and results are detailed in Section \ref{spectral_classification} and discussed, together with other interesting information on our sample of stars, in Section \ref{discussion}. We conclude our findings in Section \ref{conclusion}.
\section{Spectroscopic data and target selection} \label{dataanalysis}
This section contains a description of the employed spectral archive and the MKCLASS code and details the process of target selection.
\subsection{The Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST)} \label{LAMOST}
The LAMOST telescope \citep{lamost1,lamost2}, also called the Guo Shou Jing\footnote{Guo Shou Jing (1231–1316) was a Chinese astronomer, hydraulic engineer, mathematician, and politician of the Yuan Dynasty.} Telescope, is a reflecting Schmidt telescope located at the Xinglong Observatory in Beijing, China. It boasts an effective aperture of 3.6$-$4.9\,m and a field of view of 5$\degr$. Thanks to its unique design, LAMOST is able to take 4000 spectra in a single exposure with spectral resolution R\,$\sim$\,1800, limiting magnitude r\,$\sim$\,19\,mag and wavelength coverage from 3700 to 9000\,\AA. LAMOST is therefore particularly suited to survey large portions of the sky and is dedicated to a spectral survey of the entire available northern sky. LAMOST data products are released to the public in consecutive data releases and can be accessed via the LAMOST spectral archive.\footnote{\url{http://www.lamost.org}} With about 10 million stellar spectra contained in DR6, the LAMOST archive constitutes a real treasure trove for researchers.
\begin{figure}
\includegraphics[width=\columnwidth]{5200A_spectra_instrumental_flux2.eps}
\caption{4700\,\AA\ to 5700\,\AA\ region of (from top to bottom) the LAMOST DR4 spectra of the mCP stars LAMOST J034458.31+464848.7 (\#139; TYC\,3313-1279-1), LAMOST J040642.34+454640.8 (\#180; HD\,25706), LAMOST J072118.92+223422.7 (\#792; TYC\,1909-1687-1), and the late-type 'impostor' LAMOST J001159.88+435908.5 (GSC\,02794-00977). The position of the characteristic 5200\,\AA\ depression is indicated.}
\label{fig_5200A_depression2}
\end{figure}
\subsection{The MKCLASS code} \label{MKCLASS}
MKCLASS is a computer program written to classify stellar spectra on the Morgan-Keenan-Kellman (MKK) system \citep{gray14}. It has been designed to emulate the process of classifying by a human classifier. First, a rough spectral type is assigned, which is then refined by direct comparison with spectra from standard star libraries.
Currently, MKCLASS is able to classify spectra in the violet-green region (3800-5600\,\AA) in either rectified or flux-calibrated format. Several studies (e.g. \citealt{gray14,gray16,huemmerich18}) have shown that, providing input spectra of sufficient signal-to-noise (S/N), the results of MKCLASS compare well with the results of manual classification (precision of 0.6 spectral subclass and half a luminosity class according to \citealt{gray14}).
MKCLASS comes with two libraries of MKK standard spectra, which have been acquired with the Gray/Miller (GM) spectrograph on the 0.8\,m reflector of the Dark Sky Observatory in North Carolina, USA. \textit{libr18} contains rectified spectra in the spectral range from 3800–4600\,\AA\ and a resolution of 1.8\,\AA\ that were obtained with a 1200\,g\,mm$^{-1}$ grating. \textit{libnor36} boasts flux-calibrated and normalised spectra in the spectral range from 3800–5600\,\AA\ and a resolution of 3.6\,\AA\ obtained with a 600\,g\,mm$^{-1}$ grating. MKCLASS allows for the use of additional spectral libraries tailored to the specific needs of the researcher.
An interesting feature of the MKCLASS code is its ability to identify a set of spectral peculiarities, such as found in CP1 and CP2 stars, barium stars, carbon-rich giants etc. For more information on the MKCLASS code, we refer the reader to \citet{gray14} and the corresponding website.\footnote{\url{http://www.appstate.edu/~grayro/mkclass/}}
\subsection{Target selection criteria} \label{target}
To select suitable mCP star candidates, we specifically searched for the presence of the tell-tale 5200\,\AA\ depression in the LAMOST DR4 spectra of early-type stars. In the following, we provide background information and detail our selection criteria and the methods employed in the construction of the present sample of stars.
\subsubsection{The flux depressions in mCP stars} \label{5200A_depression}
The first to notice significant flux depressions at 4100\,\AA, 5200\,\AA, and 6300\,\AA\ in the spectrum of the mCP star HD\,221568 was \citet{1969ApJ...157L..59K}. Similar features in the ultraviolet region at 1400\,\AA, 1750\,\AA, and 2750\,\AA\ were later identified and investigated \citep{1977A&A....56..413J,1978A&A....70..379J}. It was found that these spectral features solely occur in mCP stars. \citet{khan07} showed that Fe is the principal contributor to the 5200\,\AA\ depression for the whole range of $T_{\rm eff}$ of mCP stars, while Cr and Si play a role primarily in the low $T_{\rm eff}$ region. Figure \ref{fig_5200A_depression} shows the 4800\,\AA\ to 5700\,\AA\ region of the spectra of a non-CP star, a corresponding synthetic spectrum and the newly-identified mCP star LAMOST J025951.09+540337.5 (\#78\footnote{The numbers given behind the identifiers refer to the internal identification number and facilitate easy identification in the tables.}; TYC\,3701-157-1), illustrating the 5200\,\AA\ depression in the latter object.
\begin{figure*}
\includegraphics[width=\textwidth]{F0V_standard_spectra.eps}
\caption{Blue-violet region of (from top to bottom) the F0 V standard spectra from the \textit{liblamost}, \textit{libsynth}, \textit{libnor36}, and \textit{libr18} standard libraries.}
\label{fig_F0V_standards}
\end{figure*}
To investigate the flux depression at 5200\,\AA, \citet{1976A&A....51..223M} introduced the narrow-band three-filter $\Delta a$ system, which samples the depth of this depression by comparing the flux at the center (5220~{\AA}, $g_{\rm 2}$) with the adjacent regions (5000~{\AA}, $g_{\rm 1}$ and 5500~{\AA}, $y$) using a band-width of 130~{\AA} for $g_{\rm 1}$ and $g_{\rm 2}$ and 230~{\AA} for the Str{\"o}mgren $y$ filter. The index was introduced as: $$ a = g_{\rm 2} - (g_{\rm 1} + y)/2. $$ This quantity is slightly dependent on temperature in the sense that it increases towards lower temperatures. Therefore, the intrinsic peculiarity index had to be defined as: $$ \Delta a = a - a_{\rm 0}(g_{\rm 1} - y), $$ that is, the difference between the individual $a$ value and the $a$ value of non-peculiar stars of the same colour. The locus of the $a_{\rm 0}$ values for non-peculiar objects was termed the normality line. Virtually all mCP stars have positive $\Delta a$ values up to $+75$\,mmag \citep{2005A&A...441..631P}. Only extreme cases of CP1 and CP3 stars can exhibit marginally positive $\Delta a$ values. Be/Ae, B[e] and $\lambda$ Bootis stars exhibit significant negative values. In summary, it has been shown that the $\Delta a$ system is an efficient and reliable means of identifying mCP stars.
\subsubsection{Sample selection} \label{sample_selection}
In the present study, we concentrated on the publicly-available spectra from LAMOST DR4 \citep{lamost1,DR4}. As first step, the complete catalogue was cross-matched with the $Gaia$ DR2 catalogue \citep{2018A&A...616A...1G}. To identify suitable targets, we exploited the $G$ versus $(BP-RP)$ diagram to set a corresponding limit on the investigated spectral type range (hotter than mid F, i.e. $(BP-RP)$\,$<$\,0.45\,mag). From the remaining objects, apparent supergiants were excluded. As this approach is bound to miss highly-reddened hot objects, we searched the spectral types listed in the DR4 VizieR online catalogue\footnote{\url{http://cdsarc.u-strasbg.fr/viz-bin/cat/V/153}} \citep{DR4} for additional early-type (B-, A-, and F-type) targets, which were also included into the analysis.
Only spectra with a S/N of more than 50 in the Sloan $g$ band were considered for further analysis. This cut was deemed necessary because a lower S/N renders the detection of mCP star features difficult \citep{2011AN....332...77P}. If more than one spectrum was available for a single object, only the spectrum with the highest Sloan $g$ band S/N was included into the analysis.
From the remaining spectra, suitable candidates were selected by the presence of the tell-tale 5200\,\AA\ depression. To calculate synthetic $\Delta a$ values, all spectra were normalised to the flux at 4030\,\AA. This guarantees that the large absolute flux differences introduced by the apparent visual magnitude do not cause any numerical biases in the final magnitudes. The filter curves of $g_1$, $g_2$, and $y$ as defined in \citet{2003MNRAS.341..849K} were then folded with the spectra and the corresponding magnitudes calculated. All objects with a significant positive $\Delta a$ index were visually inspected for the presence of a 5200\,\AA\ depression in order to sort out glitches in the spectra or contamination by other objects such as cool stars with strong features in the 5200\,\AA\ range.
Figure \ref{fig_5200A_depression2} illustrates this process by providing sample LAMOST DR4 spectra of mCP stars showing 5200\,\AA\ flux depressions of various strengths. Also shown is the 'impostor' LAMOST J001159.88+435908.5 (GSC\,02794-00977). This object is actually a mid to late K star whose 5200\,\AA\ region is dominated by absorption lines of the \ion{Mg}{i} triplet at 5167\,\AA, 5173\,\AA, and 5184\,\AA, which leads to a significantly positive $\Delta a$ value and highlights the need for setting a limit on the investigated spectral type range via the above described colour-colour cut.
In this way, a list of 1002 mCP star candidates was compiled. This collection of stars is referred to in the following as the 'final sample'. We here emphasise that our sample is obviously biased towards mCP stars with conspicuous flux depressions at 5200\,\AA. However, not all mCP stars show significant 5200\,\AA\ depressions, in particular in low-resolution spectra, and such objects will have been missed by the imposed selection criteria. On the other hand, early-type stars with significant 5200\,\AA\ depressions are nearly always mCP stars; therefore, the chosen approach should be well suited to collecting a pure sample of mCP stars.
\begin{figure*}
\includegraphics[width=\textwidth]{showcase1.eps}
\caption{Showcase of three newly identified 'hot' mCP stars, illustrating the blue-violet region of the LAMOST DR4 spectra of (from top to bottom) LAMOST J035046.03+363648.2 (\#151; Gaia DR2 220081859486642816), LAMOST J195631.74+253407.8 (\#929; Gaia DR2 2026771741029840128), and LAMOST J062529.84-032411.9 (\#576; TYC 4789-2924-1). MKCLASS final types and, where available, manual types derived in the present study are indicated. Some prominent lines of interest are identified. The asterisk marks the position of a 'glitch' in the spectrum of LAMOST J062529.84-032411.9.}
\label{spectra_early_types}
\end{figure*}
\section{Spectral classification} \label{spectral_classification}
Spectral classification is an important tool in astrophysics, which allows for the easy identification of astrophysically interesting objects. Furthermore, it enables to place stars in the Hertzsprung-Russell diagram, thus enabling the derivation of physical parameters. However, in the era of large survey projects such as LAMOST, RAVE, or SDSS/SEGUE, which produce a multitude of stellar spectra, human manual classification is no longer able to cope with the amount of data and the need for automatic classification has arisen. For the present study, we chose to employ a modified version of the MKCLASS code (cf. Section \ref{MKCLASS}) that has been specifically tailored to the needs of our project. More details are provided in this section, which details the spectral classification workflow.
\subsection{Spectral classification with the MKCLASS code} \label{spectral_classification_MKCLASS}
As deduced from an investigation of the program code, the current version of the MKCLASS code (v1.07) is able to identify the following spectral features, which are important in the detection of CP2 stars: the blend at 4077\,\AA\ (which may contain contributions from \ion{Si}{ii} 4076\,\AA, \ion{Sr}{ii} 4077\,\AA, and \ion{Cr}{ii} 4077\,\AA), the blend at 4130\,\AA\ (due to enhanced \ion{Si}{ii} 4128/30\,\AA\ and/or \ion{Eu}{ii} 4130\,\AA), and the \ion{Eu}{ii} 4205\,\AA\ line. On a significant detection of these features, the following output is created: 'Sr' (4077\,\AA), 'Si' (4130\,\AA), 'Eu' (4205\,\AA). As other elements besides Sr and Si contribute to the blends at 4077\,\AA\ and 4130\,\AA, the output may be misleading in some cases. Nevertheless, this allows a robust detection of mCP stars (e.g. \citealt{huemmerich18}) and is a good starting point for further investigations.
To suit the special needs of our project, which is solely concerned with the identification and classification of mCP stars among a sample of early-type candidate stars, we opted to refine the MKCLASS peculiarity identification routine. The code was therefore altered to probe several additional lines, with the advantage that the new version is now able to more robustly identify traditional mCP star peculiarities. In addition, we enabled the identification of Cr peculiarities and, to some extent, He peculiarities, which are relevant to the classification of mCP stars. The choice of lines was dictated by the resolution and quality of the input material (i.e. LAMOST DR4 spectra), in particular concerning the availability of neighbouring continuum flux to probe a certain line and the numerous line blends due to the low resolution. We therefore stress that the resulting version of the MKCLASS code, which is referred to hereafter as MKCLASS\_mCP, has been created specifically for identifying and classifying mCP stars in LAMOST low-resolution spectra. Applying the code to spectra of other resolutions will require a corresponding update of the peculiarity classification routine and, perhaps, an update and enlargement of the employed standard star libraries (see below). Table \ref{table_lines} lists the lines and blends identified by MKCLASS\_mCP, as well as the spectral range in which the corresponding features are searched for. We note that at the resolution of the employed LAMOST spectra, all these lines are, to some extent, blended with other absorption lines. Nevertheless, the listed ions generally constitute the main contributors to these blends in mCP stars.
A sample output of MKCLASS\_mCP is provided in column five of Table \ref{table_MKCLASS_manual}. Further information on the interpretation of this output is provided below; a discussion of the accuracy of the derived classifications is provided in Section \ref{evaluation}.
\begin{table}
\caption{Absorption lines and blends identified by the modified version of the MKCLASS code (MKCLASS\_mCP) and used in the identification and classification of mCP stars in the present study. The columns denote: (1) Blend/line. (2) Wavelength (\AA). (3) Spectral range in which the corresponding feature was probed.}
\label{table_lines}
\begin{center}
\begin{tabular}{cll}
\hline
\hline
(1) & (2) & (3) \\
\textbf{Blend/line} & \textbf{Wavelength (\AA)} & \textbf{SpT\_range} \\
\hline
(\ion{Si}{ii}/\ion{Cr}{ii}/\ion{Sr}{ii}) & 4076/77 & B7$-$F5 \\
\hline
(\ion{Si}{ii}/\ion{Eu}{ii}) & 4128/30 & B3$-$F2 \\
\hline
\ion{Si}{ii} & 3856 & B3$-$F2 \\
\ion{Si}{ii} & 4200 & B3$-$A2 \\
\ion{Si}{ii} & 5041 & B3$-$F2 \\
\ion{Si}{ii} & 5056 & B3$-$F2 \\
\ion{Si}{ii} & 6347 & B3$-$F2 \\
\ion{Si}{ii} & 6371 & B3$-$F2 \\
\hline
\ion{Cr}{ii} & 3866 & B7$-$F2 \\
\ion{Cr}{ii} & 4172 & B7$-$F2 \\
\hline
\ion{Sr}{ii} & 4216 & B7$-$F2 \\
\hline
\ion{Eu}{ii} & 4205 & B7$-$F2 \\
\hline
\ion{He}{i} & 4009 & B0$-$A0 \\
\ion{He}{i} & 4026 & B0$-$A0 \\
\ion{He}{i} & 4144 & B0$-$A0 \\
\ion{He}{i} & 4387 & B0$-$A0 \\
\hline
\end{tabular}
\end{center}
\end{table}
We note that the \ion{Si}{ii} line at 5041\,\AA\ increases significantly with temperature type; therefore, different detection limits were applied depending on the investigated temperature range. Furthermore, the \ion{Si}{ii} 6347/71\,\AA\ lines were found to show a significant scatter in MKK standard stars. However, at the resolution of the LAMOST spectra, the red \ion{Si}{ii} lines are a readily detectable and outstanding feature of strong Si stars and contribute significantly to an unambiguous detection of Si peculiarity, in particular as the corresponding lines in the blue-violet region are difficult to detect because of continuum flux issues (3856/62\,\AA) and blending issues (4076\,\AA\ and 4128/30\,\AA).
\begin{figure*}
\includegraphics[width=\textwidth]{showcase2.eps}
\caption{Showcase of three newly identified 'cool' mCP stars, illustrating the blue-violet region of the LAMOST DR4 spectra of (from top to bottom) LAMOST J034854.70+521413.1 (\#150; Gaia DR2 251609324623302400), LAMOST J052816.11-063820.1 (\#344; TYC 4765-708-1), and LAMOST J062221.82+595613.0 (\#561; TYC 3776-269-1). MKCLASS final types and, where available, manual types derived in the present study are indicated. Some prominent lines of interest are identified.}
\label{spectra_late_types}
\end{figure*}
As has been desribed in Section \ref{MKCLASS}, MKCLASS is a computer program that emulates the workflow of a human classifier in the traditional MKK spectral classification process, which involves comparing the input spectrum to a set of standard star spectra. It is therefore imperative to carefully select standard star libraries that match the input spectra in resolution and calibrationwise. MKCLASS comes with the two standard libraries \textit{libr18} and \textit{libnor36} (cf. Section \ref{MKCLASS}), which, unfortunately, do not match the spectral resolution of the LAMOST low-resolution spectra. Furthermore, as far as we are aware of, a standard library based on LAMOST spectra does not exist.
\begin{table*}
\caption{Spectral classifications derived by manual classification and the MKCLASS\_mCP code. The columns denote: (1) Internal identification number. (2) LAMOST identifier. (3) Spectral type derived by manual classification. (4) MKCLASS\_mCP final type. (5) MKCLASS\_mCP output using the standard star libraries \textit{libr18}, \textit{libnor36}, \textit{libsynth}, and \textit{liblamost}.}
\label{table_MKCLASS_manual}
\scriptsize{
\begin{center}
\begin{tabular}{|l|l|l|l|l|}
\hline
(1) & (2) & (3) & (4) & (5) \\
\hline
\textbf{No.} & \textbf{LAMOST ID} & \textbf{SpT\_manual} & \textbf{SpT\_MKCLASS\_mCP} & \textbf{Output using \textit{libr18}/\textit{libnor36}/\textit{libsynth}/\textit{liblamost}} \\
\hline
142 & J034541.53+275631.8 & kA3hA6mA6 SrCrEu & A6 V SrCrEu & kA4hA4mA8 bl4077 bl4130 Sr4216 Cr4172 Eu4205 \\
& & & & A6 V bl4077 bl4130 Sr4216 Eu4205 \\
& & & & A6 V bl4077 bl4130 Sr4216 Eu4205 \\
& & & & A5 IV$-$V bl4077 bl4130 Sr4216 \\
\hline
151 & J035046.03+363648.2 & B8 V Si & B8 IV Si & B6 IV$-$V Si5041 Si5056 Si6347 bl4130 \\
& & & & B8 IV Si6347 bl4130 \\
& & & & B8 IV Si5041 Si5056 Si6347 bl4130 \\
& & & & B7 IV Si5041 Si5056 Si6347 bl4077 bl4130 \\
\hline
232 & J043201.64+471447.8 & A2 V SrCrEu(Si) & A1 V SrCrEu & kA2hA6mA8 bl4077 Sr4216 Cr4172 Eu4205 \\
& & & & A1 IV$-$V bl4077 bl4130 Sr4216 \\
& & & & kA2hA3mA6 bl4077 bl4130 Sr4216 Cr4172 \\
& & & & A1 V bl4077 bl4130 Sr4216 \\
\hline
306 & J051844.95+380605.3 & B9 V SrCrEuSi & B9 IV SrCr & kB9hA0mA2 bl4077 bl4130 Sr4216 Cr4172 \\
& & & & B9.5 III$-$IV bl4077 bl4130 Sr4216 Cr4172 \\
& & & & kB9.5hA1mA3 bl4077 bl4130 Sr4216 Cr3866 Cr4172 \\
& & & & B9 IV bl4077 bl4130 Sr4216 Cr4172 \\
\hline
344 & J052816.11-063820.1 & kA1hA9mA9 SrCrEu & kA1hA8mA9 SrEu & A1 II$-$III bl4077 bl4130 Sr4216 Eu4205 \\
& & & & kA2hA5mA9 bl4077 bl4130 Sr4216 Eu4205 \\
& & & & kA1hA9mA8 bl4077 bl4130 Sr4216 Eu4205 \\
& & & & kA1hA8mA9 Si3856 bl4077 bl4130 Sr4216 Eu4205 \\
\hline
561 & J062221.82+595613.0 & kA5hA7:mF2 SrCrEuSi: & A7 V SrCrEu & A8 mA4 metal weak \\
& & & & kA6hA8mF3 bl4077 bl4130 Sr4216 Cr3866 \\
& & & & kA5hA7mF1 bl4077 bl4130 Sr4216 Cr3866 Cr4172 Eu4205 \\
& & & & A7 V bl4077 bl4130 Sr4216 Cr4172 \\
\hline
596 & J062909.51+023823.8 & B8 V Si & B8 IV$-$V Si & B8 IV$-$V Si5041 Si5056 Si6347 Si6371 bl4130 \\
& & & & B8 IV$-$V Si5041 Si5056 Si6347 Si6371 bl4130 \\
& & & & B8 IV Si5041 Si5056 Si6347 Si6371 bl4130 \\
& & & & B8 IV$-$V Si5041 Si5056 Si6347 Si6371 bl4077 bl4130 \\
\hline
724 & J065511.76+115158.3 & A7 V SrCrEu & A7 V SrEu & A8 V bl4077 bl4130 Sr4216 Eu4205 \\
& & & & A9 V bl4077 bl4130 Sr4216 Eu4205 \\
& & & & A7 V Si6347 bl4077 bl4130 Sr4216 Eu4205 \\
& & & & A7 V bl4077 bl4130 Sr4216 \\
\hline
732 & J065647.94+242958.8 & A0 V SiSrCr & B9.5 V Sr & kB9hA7mA5 Sr4216 Cr3866 Eu4205 \\
& & & & B9.5 IV$-$V bl4077 bl4130 \\
& & & & kB9.5hA3mA3 bl4077 Sr4216 \\
& & & & B9.5 V bl4077 bl4130 \\
\hline
929 & J195631.74+253407.8 & B8 V SiCr & B8 IV$-$V Si & B8 V Si5041 Si5056 Si6347 Si6371 bl4077 bl4130 \\
& & & & B8 IV$-$V Si5056 Si6347 Si6371 bl4077 bl4130 \\
& & & & B8 IV$-$V Si5041 Si5056 Si6347 Si6371 bl4077 bl4130 \\
& & & & A0 II$-$III Si5056 Si6347 Si6371 bl4077 bl4130 \\
\hline
\end{tabular}
\end{center}
}
\end{table*}
In the framework of the LAMOST-$Kepler$ project, \citet{gray16} presented MKK spectral classifications of more than 100\,000 LAMOST spectra of about 80\,000 objects situated in the $Kepler$ field. The authors solved the above-mentioned issue by degrading the flux-calibrated LAMOST spectra to a resolution of R\,$\sim$\,1100 and truncating them to the 3800$-$5600\,\AA\ region in order to enable the use of MKCLASS with the flux-calibrated \textit{libnor36} library (cf. also the MKCLASS documentation). Here we follow their approach, but we also searched for alternative methods, as degrading spectra obviously results in loss of information. This, however, is detrimental to the identification of the often subtle chemical peculiarities of our sample stars.
We synthesised a library of spectra using the program SPECTRUM\footnote{\url{http://www.appstate.edu/~grayro/spectrum/spectrum.html}} \citep{SPECTRUM} and ATLAS9 model atmospheres \citep{ATLAS9}, which is referred to hereafter as the \textit{libsynth} library. Only dwarf spectra (luminosity class V) were synthesised because no models were available to reproduce the subtle differences in surface gravity among early-type giant stars. Furthermore, we collected a set of LAMOST standard star spectra (the \textit{liblamost} library) by carefully choosing a grid of suitable high S/N spectra from the list presented by \citet{gray16}. Only dwarf and giant spectra were chosen, as not enough suitable spectra of higher luminosity class objects were available to build up a corresponding grid. We note, however, that it has been well confirmed that mCP stars are generally main-sequence objects (cf. \citealt{netopil17}, and references therein, and Sections \ref{introduction}, \ref{evaluation}, and \ref{evolutionary_status}); therefore, we do not expect that the absence of spectra of higher luminosity class objects in these two libraries significantly affects our results -- in particular as the \textit{libr18} and \textit{libnor36} libraries boast corresponding spectra of all luminosity classes.
The stars and corresponding LAMOST spectrum identifiers of the \textit{liblamost} library are given in the Appendix in Table \ref{table_liblamost}. Although it contains a very suitable grid of dwarf spectra, we note that the \textit{liblamost} library is far from being a perfect set of standard star spectra (a corresponding quality flag that estimates the suitability of a spectrum as a standard is also provided in Table \ref{table_liblamost}). It contains a fast rotator and some spectra show 'impurities' not expected in MKK standards. These shortcomings will lead to increased uncertainties in the derivation of the temperature and luminosity classes. However, the library consists of spectra obtained with the same instrument -- and hence, importantly, of the same resolution as our input spectra -- and was extremely valuable in the identification of chemical peculiarities. We nevertheless explicitly caution against using the \textit{liblamost} library as a standard star library out of the context of the present investigation. We also emphasise the need for a standard star library based on LAMOST low-resolution spectra, which will greatly facilitate further research based on this excellent data source. The \textit{liblamost} library may very well serve as a starting point; this, however, is beyond the scope of the present investigation.
In the following, an overview over the employed spectral libraries is presented. As an example, Figure \ref{fig_F0V_standards} illustrates the F0V standard spectra from the corresponding libraries. We note that the \textit{libsynth} and \textit{liblamost} libraries only contain spectra in the approximate spectral type range of our sample stars.
\begin{itemize}
\item \textbf{\textit{libr18}}: spectral range from 3800–4600\,\AA, resolution of 1.8\,\AA\ (R\,$\sim$\,2200), normalised and rectified spectra; all luminosity classes (Ia-V)
\item \textbf{\textit{libnor36}}: spectral range from 3800–5600\,\AA, resolution of 3.6\,\AA\ (R\,$\sim$\,1100), flux-calibrated and normalised spectra; all luminosity classes (Ia-V)
\item \textbf{\textit{libsynth}}: spectral range from 3800–4600\,\AA, smoothed to a resolution of 3.0\,\AA\ and an output spacing of 0.5\,\AA, flux-calibrated and normalised synthetic spectra; only dwarf spectra (luminosity class V); spectral types B5 to F5
\item \textbf{\textit{liblamost}}: spectral range from 3800–5600\,\AA, resolution R\,$\sim$\,1800, flux-calibrated and normalised spectra; only dwarf and giant spectra (luminosity classes V and III); spectral types B3 to G0
\end{itemize}
mCP stars may exhibit peculiar \ion{Ca}{ii} K profiles and line strengths \citep{faraggiana87,gray09,ghazaryan18} as well as generally enhanced metal-lines. Several authors have therefore adopted a notation that indicates separate spectral types as derived from the \ion{Ca}{ii} K line (the k-type), the hydrogen lines (the h-type), and the general strength of the metal-lines (the m-type), in the same way as is usually done for CP1 stars. The MKCLASS code also assigns k/h/m-types in cases where discrepant spectral types are derived from the corresponding features. As mCP stars are prone to exhibiting marked Ca and He deficiencies (e.g. \citealt{gray09,ghazaryan18}) and often enhanced metal-lines, the hydrogen-line profile is a better indicator of the actual effective temperature \citep{gray09}. Where they have been derived by the code (or by manual classification), k/h/m types are listed in the present study.
For most stars, only minor differences in temperature and luminosity types were found between the results from the different spectral libraries. In cases where the same spectral type was derived more than once, the most common spectral type was adopted (cf. Table \ref{table_MKCLASS_manual}). If no common classifications existed, spectral types were favoured in the order \textit{liblamost} > \textit{libsynth} > \textit{libnor36} > \textit{libr18}. In the case of strong differences between the derived classifications, the corresponding spectra were visually inspected and the best fitting type was chosen.
To determine significant chemical peculiarities from the 'raw' MKCLASS\_mCP output, the number of detections $N$\textsubscript{det} of the peculiar strength of a given line with the different standard star libraries (0\,$\le$\,$N$\textsubscript{det}\,$\le$\,4) was counted, which provides an estimation of significance. Obviously, $N$\textsubscript{det}\,=\,4 is a very robust detection; $N$\textsubscript{det}\,<\,2 detections, on the other hand, have to be viewed with caution. Furthermore, we required that the identification of overabundances cannot be based on a single strong line (with the exception of the \ion{Cr}{ii} 4172\,\AA\ line in the identification of a Cr peculiarity; see below).
To come up with an approach that forms a compromise between spurious detections and overly high thresholds required a good amount of experimentation and experience in comparing the results of manual and automatic classification. Table \ref{table_peculiarities} lists the conditions found to work best with the input material and our methodological approach. $N$\textsubscript{det}($\lambda$) is the number of detections of a peculiarly strong line at the specified wavelength (\AA). For instance, a Cr peculiarity was flagged when (a) a strong \ion{Cr}{ii} 4172\,\AA\ line was detected with a least two different standard star libraries \textbf{or} (b) a strong \ion{Cr}{ii} 3866\,\AA\ line \textbf{and} a strong \ion{Cr}{ii} 4172\,\AA\ line were detected at least once \textbf{or} (c) a strong blend at 4077\,\AA\ was detected at least twice \textbf{and} a strong \ion{Cr}{ii} 4172\,\AA\ line was detected at least once.
Following the conventions of the MKK system, the peculiarity types 'Si', 'Cr', 'Sr', and 'Eu' were flagged according to the conditions given in Table \ref{table_peculiarities} and attached to the temperature and luminosity types in the final spectral classification. Several stars in our sample that were not assigned any of the above mentioned peculiarity types nevertheless show strong blends at 4077\,\AA\ and/or 4130\,\AA. In these cases, we decided to add the non-standard suffixes 'bl4077' and 'bl4130' to the derived spectral types (e.g. 'B8 IV bl4130') if the corresponding blends had been detected at least twice. In these objects, apart from the strong blends, the peculiarities are either too subtle to have passed our significance criteria, no other significant features are present, or the code failed to identify them for some reason. Manual classification is necessary to throw more light on this matter (cf. Section \ref{peculiarity_types}).
In addition, we opted to probe the \ion{He}{i} lines at 4009\,\AA, 4026\,\AA, 4144\,\AA, and 4387\,\AA\ to identify CP2 stars with weak \ion{He}{i} lines and He-peculiar objects. The corresponding detection thresholds for all four standard star libraries were determined using the He lines of 626 apparently chemically-normal B stars with spectra boasting S/N\,$>$\,100. The number of detections as 'weak' or 'strong' of the He lines with the different standard star libraries was counted and the results across all lines and libraries were added up to yield $N$\textsubscript{det}(He-wk) and $N$\textsubscript{det}(He-st). He-weakness and He-overabundance were assumed when $N$\textsubscript{det}(He-wk)\,$>$\,2 and $N$\textsubscript{det}(He-st)\,$>$\,2, respectively. In this way, we identified 55 mCP stars with weak \ion{He}{i} lines and three mCP stars with apparently strong \ion{He}{i} lines. As expected, these are mostly B7$-$B9 Si CP2 stars, which are notorious for their weak He lines, and mid-B type stars (likely He-peculiar objects). Interestingly, in three mid-B type stars, both weak and strong \ion{He}{i} lines were identified, which strongly suggests He peculiarity. The corresponding suffixes 'He-wk' and 'He-st' were added to the derived spectral types. Several He-peculiar objects are discussed in Section \ref{Hepec}.
In this way, peculiarities were identified in all but 36 stars from our sample, which highlights the efficiency of the chosen approach. The (mostly low S/N) spectra of the remaining objects were investigated manually and searched for the presence of chemical peculiarities. Most of these objects show subtle or complicated peculiarities that failed to meet the imposed significance criteria. Corresponding peculiarity types were manually added to the final spectral types. The remaining objects are a 'mixed bag' containing stars with enhanced metal-lines and strong flux depressions that nevertheless lack the traditional Si, Cr, Sr, Eu peculiarities and several He-peculiar objects. Appropriate comments were added to the tables, in which manually-derived peculiarity types and all further additions that are not directly based on the MKCLASS\_mCP output are highlighted.
\begin{table}
\caption{Conditions employed to flag the presence of an overabundance from the 'raw' MKCLASS\_mCP output. The columns denote: (1) Overabundant ion. (2) Condition(s) required to be met for a detection to be deemed significant.}
\label{table_peculiarities}
\begin{tabular}{cll}
\hline
\hline
(1) & (2) & \\
\textbf{Ion} & \textbf{condition(s)} & \\
\hline
\textbf{\ion{Si}{ii}} & ($N$\textsubscript{det}(3856)\,$+$\,$N$\textsubscript{det}(4200)\,$+$\,$N$\textsubscript{det}(5041)\,$+$ & \\
& $N$\textsubscript{det}(5056)\,$+$\,$N$\textsubscript{det}(6347)\,$+$\,$N$\textsubscript{det}(6371))\,$>$\,1 & \\
\hline
\textbf{\ion{Cr}{ii}} & $N$\textsubscript{det}(4172)\,$>$\,1 & OR \\
& ($N$\textsubscript{det}(3866)\,$>$\,0\,AND\,$N$\textsubscript{det}(4172)\,$>$\,0) & OR \\
& ($N$\textsubscript{det}(4077)\,$>$\,1 AND $N$\textsubscript{det}(4172)\,$>$\,0) & \\
\hline
\textbf{\ion{Sr}{ii}} & ($N$\textsubscript{det}(4077)\,$>$\,0 AND $N$\textsubscript{det}(4216)\,$>$\,0) & \\
\hline
\textbf{\ion{Eu}{ii}} & ($N$\textsubscript{det}(4130)\,$>$\,2 AND $N$\textsubscript{det}(4205)\,$>$\,0) & \\
\hline
\end{tabular}
\end{table}
\subsection{Evaluation} \label{evaluation}
As a test of the validity of our approach, we manually classified a sample of ten randomly chosen stars and compared our results with the final spectral types derived in the described manner from the MKCLASS\_mCP output, which, for convenience, are termed hereafter the 'MKCLASS final types'. In addition, we visually inspected the spectra of about 100 further stars to check for the presence of peculiarly strong lines and evaluate the reliability of the classifications. The results from the manual classification are shown in Table \ref{table_MKCLASS_manual} and highlight the good agreement between the manually- and automatically-derived (hydrogen-line) temperature types. In general, we estimate the uncertainty of the derived temperature types to be $\pm$1\,subclass. This, however, increases to about $\pm$2\,subclasses towards later and more peculiar mCP stars for which the classification is notoriously difficult and, for the most extreme objects, unreliable. Figures \ref{spectra_early_types} and \ref{spectra_late_types} showcase, respectively, three 'hot' mCP stars and three 'cool' mCP stars, which have been newly identified as such in the present study. MKCLASS final types and, where available, manual types from Table \ref{table_MKCLASS_manual} are indicated.
\begin{figure}
\includegraphics[width=\columnwidth]{comparison1.eps} \\
\includegraphics[width=\columnwidth]{comparison3.eps}
\caption{Upper panel shows a comparison of the blue-violet spectra of the mCP star LAMOST J065647.94+242958.8 (\#732; TYC\,1898-1408-1; manual type: A0 V SiSrCr; MKCLASS final type: B9.5 V Sr) to the \textit{liblamost} A0 V standard spectrum. Lower panel illustrates a comparison of the blue-violet spectra of the mCP star LAMOST\,J052816.11-063820.1 (\#344; TYC\,4765-708-1; manual type: kA1hA9mA9 SrCrEu; MKCLASS final type: kA1hA8mA9 SrEu) to the \textit{liblamost} F0 V standard spectrum. Some prominent lines of interest are indicated. We note the weak \ion{Ca}{ii} K line with a peculiar profile and the unusual profile of the H$\epsilon$ line in LAMOST\,J052816.11-063820.1.}
\label{spectra_comparison}
\end{figure}
There is also a generally good agreement in regard to the derived peculiarity types. However, with our approach, we obviously missed the presence of peculiarities in several objects (Table \ref{table_MKCLASS_manual}). A further investigation of these stars shows that this has been mostly due to either the imposed significance criteria, weak or complicated peculiarities, or the absence of continuum flux to probe certain lines (or a combination thereof). A good case in point is LAMOST J065647.94+242958.8 (\#732; TYC\,1898-1408-1), whose blue-violet spectrum is shown in the upper panel of Figure \ref{spectra_comparison}. It shows enhanced \ion{Si}{II} lines at 3856/62\,\AA, 4128/31\,\AA, and 4200\,\AA. Furthermore, strong \ion{Cr}{II} lines are present at 3866\,\AA, 4111\,\AA, 4172\,\AA, 4559\,\AA, and 4588\,\AA. The blend at around 4077\,\AA\ is notoriously difficult to interpret and can contain contributions from \ion{Si}{II}, \ion{Cr}{II}, and \ion{Sr}{II}. The strong \ion{Sr}{II} line at 4216\,\AA, however, indicates the presence of a Sr peculiarity. Consequently, we have classified this star as A0 V SiSrCr. The MKCLASS\_mCP code missed the rather subtle Si and Cr peculiarities (MKCLASS final type: B9.5 V Sr).
The bottom panel of Figure \ref{spectra_comparison} illustrates the case of LAMOST J052816.11-063820.1 (\#344; TYC\,4765-708-1; also shown in Figure \ref{spectra_late_types}), a cool CP2 star that may serve as a warning and an example illustrating the difficulty of classifying the more extreme mCP stars, which may possess peculiarities that render their spectra difficult to match to any standard star \citep{gray09}. The star exhibits a weak \ion{Ca}{ii} K line with a peculiar profile. While the hydrogen-line profile points to a late A-type star, the broad K line is that of an early A-type star and reasonably matched by that of an A1 V standard. As CP2 stars are prone to exhibiting marked Ca deficiencies (e.g. \citealt{gray09,ghazaryan18}), the hydrogen-line profile is a better indicator of the actual effective temperature than the K line strength \citep{gray09}, hence LAMOST J052816.11-063820.1 is obviously a late-type mCP star. However, while the H$\gamma$, H$\delta$, H8, and H9 lines are reasonably well matched by that of an A9 V star, the H$\epsilon$ line exhibits an unusual profile, which makes it difficult to match the hydrogen-line profile to that of any standard star. We also note the unusually weak Mg II line at 4481\,\AA. The MKCLASS\_mCP code duly assigned final k/h/m-types of A1, A8, and A9; we prefer k/h/m-types of A1, A9, and A9.
The spectrum of LAMOST J052816.11-063820.1 shows strong blends at 4077\,\AA\ and 4130\,\AA, which -- judging from the strong lines at 4216\,\AA\ (\ion{Sr}{II}), 4205\,\AA\ (\ion{Eu}{II}), and 4435\,\AA\ (\ion{Eu}{II}) -- are mainly caused by overabundances of Sr and Eu. The strong \ion{Cr}{II} lines at 3866\,\AA\ and 4111\,\AA\ (the bump in the red wing of H$\delta$) indicate a Cr overabundance; a corresponding enhanced line at 4172\,\AA, however, is notably absent. We have classified this star as kA1hA9mA9 V SrCrEu. The MKCLASS\_mCP code duly identified the main peculiarities (MKCLASS final type: kA1hA8mA9 SrEu).
The examples show that the peculiarity types derived in the present investigation are in many cases not exhaustive but rather denote the main peculiarities present. They are still very useful for first orientation and an excellent starting point for more detailed investigations; they are furthermore suited to statistical studies (cf. Section \ref{peculiarity_types}).
Some cautionary words are necessary in regard to luminosity classification. It is well known that problems with the luminosity classification may arise by the confusion of luminosity criteria and mCP star characteristica or peculiarities. In the early A-type stars, luminosity classification is primarily based on hydrogen-line profiles. The mCP stars, in general, are slow rotators that display narrow lines and hydrogen-line profiles that are easily misinterpreted as belonging to stars of higher luminosity. Indeed, the hydrogen-line profiles of many late B- and early A-type Si stars of our sample are best matched by standards of luminosity class III although there is no further indication that these star are in fact giant stars. Additional confusion may arise due to peculiarly strong lines that are also used in luminosity classification. \ion{Si}{ii} lines, for example, are enhanced in giants and supergiants as well as in several types of mCP stars, which might lead to corresponding misclassifications \citep{loden89}. This holds especially true for classifications based on photographic plates or low S/N spectra. In regard to CP1 stars, the term 'anomalous luminosity effect' has been coined, which describes the perplexing situation that luminosity criteria from different regions of the spectrum indicate different luminosities. This also applies at least partly to mCP stars, which may show strong general enhancements of metal-lines in their spectra that are reminiscent of much cooler stars. Furthermore, while obtaining the hydrogen-line type is fairly straightforward for most mCP stars, the more peculiar objects show distorted atmospheres and unusual and peculiar hydrogen-line profiles that may not match any standard star \citep{gray09}, which may result in classification errors. Abnormal hydrogen-line profiles are especially observed in cool mCP stars \citep{kochukhov02}.
\begin{figure}
\includegraphics[width=\columnwidth]{histograms.eps}
\caption{Histograms of the $G$ magnitudes (upper panel) and distances from the Sun (lower panel). For the construction of the latter, only stars with absolute parallax errors less than 25\,\% were employed.}
\label{histograms}
\end{figure}
These issues also impact automatic classification routines such as the MKCLASS code and will surely be at the root of the high luminosity classification of many of our sample stars, which should be regarded with caution. It has been well confirmed that mCP stars are generally main-sequence objects (e.g. \citealt{netopil17} and references therein) and the results from the colour-magnitude diagram (CMD) (Section \ref{evolutionary_status}) fully support this finding. A detailed analysis of this issue is necessary but beyond the scope of the present investigation.
At this point, we would like to also recall that at least half of the mCP stars are spectroscopic variables, that is, the observed line strengths may vary considerably over the rotation period \citep[e.g.][]{gray09}, which should always be kept in mind when working with mCP star spectra.
Table \ref{table_master1} in the Appendix presents the MKCLASS final types along with essential data for our sample stars.
\section{Discussion} \label{discussion}
This section discusses properties of our final sample such as magnitude distribution, distances from the Sun, evolutionary status, and their distribution in space, and discusses interesting objects such as the eclipsing binary system LAMOST J034306.74+495240.7.
\subsection{Magnitudes and distances from the Sun} \label{magnitudes_and_distances}
In Figure \ref{histograms}, we present the histograms of the $G$ magnitudes and the distances from the Sun of our sample stars. The magnitude distribution peaks between 11th and 12th magnitude, which corresponds to a distance of about 1\,kpc for the investigated range of spectral types. While our sample contains only a few new mCP stars within 500\,pc around the Sun, there is a significant number of objects beyond 2\,kpc, which might help to shed more light on the Galactic radial metallicity gradient \citep{2016A&A...585A.150N} and its influence on the formation and evolution of CP stars.
In summary, our sample is a perfect extension to the mCP stars listed in the catalogue of \citet{RM09}, which are on the average closer and brighter, peaking at 9th magnitude.
\subsection{Evolutionary status} \label{evolutionary_status}
In the following sections, we investigate the evolutionary status of our sample stars in the $(BP - RP)_0$ versus $M_{\mathrm{G,0}}$ and mass versus fractional age on the main sequence spaces. We caution that, due to the imposed selection criteria (cf. Section \ref{sample_selection}), our sample is biased towards stars showing a conspicuous 5200\,\AA\ flux depression in the low-resolution LAMOST spectra. Therefore, our results, while being based on a statistically significant sample size, have to be viewed with caution and their general validity needs to be tested by a more extended sample selected via different methodological approaches.
\subsubsection{Colour-magnitude diagram} \label{evolutionary_status_CMD}
To investigate the astrophysical properties of our sample stars in a CMD, we employed the homogeneous Gaia DR2 photometry from \citet{gaia3}. Most of our sample stars are situated within the Galactic disk farther than 500\,pc from the Sun; therefore, interstellar reddening (absorption) cannot be neglected. Because hardly any objects have Str{\"o}mgren-Crawford indices available \citep{paunzen15b}, we relied on the published reddening map by \citet{Green2018}. To interpolate within this map, parallaxes were directly converted to distances. To limit the error of the absorption values to 0.1\,mag, only objects with relative parallax errors of at most 25\,\% (942 in total) were used. The transformation of the reddening values was performed using the relations:
\begin{equation}
E(B - V) = 0.76E(BP - RP) = 0.40A_G.
\end{equation}
These relations already take into account the conversion to extinction in different bands using the coefficients as listed in \citet{Green2018}.
In Figure \ref{CMD}, we present the CMD of our sample stars together with PARSEC isochrones \citep{bressan12} for solar metallicity [Z]\,=\,0.020. We favour this value because it has been shown to be consistent with recent results of Helioseismology \citep{Vagnozzi2019}. Also included is the reddening vector according to an uncertainty of 0.1\,mag for $E(B-V)$. About 20 stars are situated below the zero-age main sequence (ZAMS) to such an extent that it cannot be explained by errors in the reddening estimation. Inconsistent photometry or binarity might possibly have led to the observed positions but, with the available data, we are unable to shed more light on this matter.
The stars LAMOST J061609.42+265703.2 (\#537; $M_{\mathrm{G,0}}$\,=\,+7.78\,mag, $(BP-RP)_0$\,=\,+0.840\,mag), LAMOST J064757.48+105648.2 (\#678; $M_{\mathrm{G,0}}$\,=\,+7.75\,mag, $(BP-RP)_0$\,=\,+1.721\,mag), and LAMOST J202943.73+384756.6 (\#943; $M_{\mathrm{G,0}}$\,=\,+3.61\,mag, $(BP-RP)_0$\,=\,+1.180\,mag) are not plotted in the CMD because they lie outside the chosen boundaries. The available spectra clearly confirm that they are mCP stars. We double-checked the identifications in the Gaia DR2 and LAMOST catalogues and searched for nearby objects on the sky that might have influenced the photometry, albeit with negative results. In the case of LAMOST J202943.73+384756.6, we strongly suspect that binarity might be at the root of the observed outlying position. Its spectrum has an unusual black-body curve, and its spectral energy distribution (SED) looks like the superposition of two objects, with a clearly visible infrared excess. However, we are unable to explain the reasons behind the observed inconsistent photometry for the other two stars.
\begin{figure}
\begin{center}
\includegraphics[width=80mm, clip]{CMD.eps}
\caption{$(BP - RP)_0$ versus $M_{\mathrm{G,0}}$ diagram of our sample stars, together with PARSEC isochrones for solar metallicity [Z]\,=\,0.020 (listed are the logarithmic ages). The arrow indicates the reddening vector for the maximum expected error due to the employed reddening map and the parallax error.}
\label{CMD}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{plot_Z.eps}
\caption{Distribution of [Z] values for the CP2 stars from the \citet{ghazaryan18} catalogue with a least three measurements of C, N, O and S.}
\label{plot_Z}
\end{center}
\end{figure}
From the distribution of stars in Figure \ref{CMD}, we conclude that the majority of our sample stars is between 100\,Myr and 1\,Gyr old. There are only a few very young stars and an accumulation of objects older than 300\,Myr.
The here employed isochrones have been calculated for stars of standard composition, that is, chemically normal stars, assuming solar metallicity [Z]\,=\,0.020. As the chemical bulk composition of mCP stars is unknown, however, the choice of the right chemical composition for the theoretical tracks remains an open question (cf. e.g. the discussion in \citealt{bagnulo06}). The main question is whether the apparent overabundances encountered at the surface are representative for the whole stellar interior. If diffusion is assumed as the main mechanism (in line with most theoretical studies), the overall abundance will be close to solar because corresponding underabundances are expected in the stellar interior. All current isochrone calculations are based on assuming a [Z] value for the whole star; it is currently not possible to consider different [Z] values for different layers of the stellar atmosphere. A detailed discussion of the influence of [Z] on the determination of age on the main sequence is provided in the following section.
\begin{figure}
\includegraphics[width=\columnwidth]{mass_age.eps}
\caption{Mass versus fractional age on the main sequence ($\tau$) distribution assuming solar metallicity [Z]\,=\,0.020 for the 903 sample stars fulfilling the imposed accuracy criteria. Upper panel shows a density plot for masses up to 4\,M$_\odot$. The position of the spectral types has been based on the information given in \citet{pecaut13}.}
\label{plot_age_mass}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{mass_age_errors.eps}
\caption{Distributions of errors for the derived fractional ages on the main sequence ($\tau$; upper panel) and masses (lower panel) assuming solar metallicity [Z]\,=\,0.020.}
\label{plot_age_mass_errors}
\end{figure}
\subsubsection{Mass versus age on the main sequence} \label{evolutionary_status_mass_age}
To examine the evolutionary status of our sample stars in more detail, we investigated the distribution of mass ($M$) versus age on the main sequence. Age on the main sequence ($\tau$) is here defined as the fraction of life spent on the main sequence, with the ZAMS corresponding to $\tau$\,=\,0\,\% and the terminal-age main sequence (TAMS) to $\tau$\,=\,100\,\%. Only objects with absolute parallax errors less than 25\,\% were considered in this process. We again used PARSEC isochrones \citep{bressan12} for solar metallicity [Z]\,=\,0.020 and 7.0\,$<$\,log\,$t$\,$<$\,10.0 (step size: 0.025). Bearing in mind the observational errors, the chosen step size is sufficient not to run into numerical inaccuracies due to the grid being too coarse. We did not interpolate within the grid but always selected the point with minimal Euclidean distance to the observed value in the $(BP-RP){_0}$ versus $M_{\mathrm{G,0}}$ space. Only grid points representing the main sequence (flagged ``1'' in the corresponding isochrones) were used. As next step, we discarded all data points with a distance of more than 0.05\,mag between observed value and theoretical grid point, which guarantees the exclusion of points below the ZAMS and above the TAMS. In this way, masses and ages were derived for 903 sample stars fulfilling the imposed accuracy criteria.
As final step, the lifetime on the main sequence was calculated for a given mass using the upper envelope of the isochrone grid. With this parameter, the fractional lifetime of a star on the main sequence can be easily calculated. To compute upper and lower limits for $\tau$ and $M$, the full error ellipse was taken into account. This procedure is described in more detail in \citet{kochukhov06}.
Table \ref{table_age_mass} lists the derived masses and fractional ages on the main sequence for solar metallicity [Z]\,=\,0.020, which are graphically represented in Figure \ref{plot_age_mass}. The density plot in the upper panel clearly shows that there are only very few stars in our sample younger than $\tau$\,<\,20\,\%. Most stars have a relative age of $\tau$\,$\ge$\,60\,\%.
To check the reliability of our results, we have investigated the error distribution of the age and mass estimates in detail. This analysis was done for solar metallicity. We have investigated the influence of the assumed metallicity in a second step. For this, masses were binned in 0.2\,M$_\odot$ and ages in 10\,\% intervals. Sizes have been chosen to guarantee the availability of a significant number of data points in each bin. Figure \ref{plot_age_mass_errors} illustrates the corresponding histograms. The absolute errors of the masses increase linearly until $M$\,=\,3.4\,M$_\odot$ and then flatten out, which means that the relative error stays constant over the whole investigated mass range. The situation is different for the derived ages; up to $\tau$\,$\le$\,90\,\%, absolute errors remain almost constant. Relative errors obtained for younger stars, therefore, are significantly larger than for old ones. This, however, does not impact our conclusions (see below). The significant increase of the errors for the last age bin is due to the higher density of isochrones in this region; furthermore, taking the error ellipse into account, some stars may be located above the TAMS.
In order to evaluate the effect of the chosen metallicity on our results, we have investigated the [Z] distribution for CP2 stars from the \citet{ghazaryan18} catalogue with a least three measurements of C, N, O, and S. These light elements were chosen because they contribute the most to the derived [Z] values. They appear significantly underabundant in most CP2 stars, which therefore show lower [Z] values than chemically normal stars (Figure \ref{plot_Z}). For most objects, we find [Z] values in the range from about 0.008 to 0.060. From the reference source, we were not able to estimate the errors of the derived [Z] values, which mainly depend on the errors of the individual abundance determinations; these, however, are mostly not available.
We emphasise that all available isochrones use scaled abundances according to the abundance pattern of the Sun. Whether this approximation can also be applied to CP stars remains at the present time unknown (cf. Figure 1 of \citealt{ghazaryan18}). On the basis of open cluster members, \citet{bagnulo06} investigated the influence of the overall metallicity on the error of the age determination for CP2 stars, using metallicities of [Z]\,=\,0.020 (solar) and [Z]\,=\,0.008, as derived from the corresponding host clusters. As clusters with [Z]\,>\,0.020 are very rare in the Milky Way \citep{2016A&A...585A.150N}, no isochrones for metallicities exceeding solar metallicity were considered in their study.
When estimating $\tau$, we have to consider two effects, which are illustrated in Figure \ref{plots_sequences_Zs}. As is well known, lines of constant $\tau$ are not distributed equally across the main sequence (Figure \ref{plots_sequences_Zs}, upper panel). For a constant value of $(BP-RP){_0}$, they are much tighter in terms of $M_{\mathrm{G,0}}$ for the first $\sim$70\,\% of the main-sequence lifetime. The total interval of $M_{\mathrm{G,0}}$ from ZAMS to TAMS spans about 2.6\,mag, whereas the intervals covered to $\tau$\,=\,25\,\% and $\tau$\,=\,50\,\% amount to only 0.3\,mag and 0.7\,mag, respectively.
The lower panel of Figure \ref{plots_sequences_Zs} explores the impact of isochrones of different metallicity on the positions of the ZAMS and TAMS. The magnitude differences between the positions of the ZAMS are nearly constant and amount to 0.4\,mag, which corresponds to an age difference of about 30\,\% at the ZAMS. This means
that a ZAMS star of solar metallicity ([Z]\,=\,0.020) would have already spent 30\,\% of its main-sequence lifetime for [Z]\,=\,0.008 but would be situated 0.4\,mag below the ZAMS for [Z]\,=\,0.060. This illustrates the dilemma ellicited by the lack of knowledge of the overall metallicity of mCP stars and the resulting loss of accuracy, as clearly demonstrated by \citet{bagnulo06}. The uncertainty is largest for stars near the ZAMS and has to be considered together with the distribution of errors for a given distinct metallicity (Figure \ref{plot_age_mass_errors}). However, because individual [Z] values and their errors are unavailable for our sample stars, we are not able to provide reliable estimations of the contribution to the computed errors.
Figure \ref{plot_different_metallicities} illustrates the mass versus $\tau$ distributions for all three investigated isochrones. The majority of our sample stars is situated in the rather narrow spectral range from B8 to A0 (cf. Section \ref{peculiarity_types}). However, there is also a lesser but significant amount of stars with spectral types between A5 and F0. Any calibrated mass distribution should represent these results to some extent. The lower-metallicity isochrone ([Z]\,=\,0.008) yields mainly old ($\tau$\,$>$75\%) and cooler (later than spectral type A0) stars; no young stars are present. On the other hand, adopting the isochrone for [Z]\,=\,0.060 results in a quite homogeneous age and mass distribution. However, there is a lack of stars cooler than A5, in conflict with the observations. Overall, as expected, none of the employed isochrones is suitable to reproduce the observed distribution of spectral types. Nevertheless, assuming solar metallicity offers the best compromise, with most stars situated in the late B-type realm and a tail of objects extending down to spectral type F0.
To further tackle this important problem, a modern and detailed abundance analysis of the light elements is needed. The current available data are rare and mainly based
on the assumption of local thermodynamical equilibrium \citep{1990ApJS...73...67R}. For the relevant spectral type domain, almost all suitable spectral lines (i.e. lines of sufficient strength) are situated in the spectral region redwards of 6000\,\AA. Unfortunately, the medium-resolution spectra of the LAMOST survey, which should be sufficient in terms of resolution, do not cover a significant amount of the specified spectral region \citep{zhang20}.
Finally, diffusion calculations for light elements are needed to estimate the influence of the magnetic field and to what extent the observed surface abundances are representative of the stellar composition. Until now, however, because of the lack of corresponding observations, such calculations are not available \citep{2012MNRAS.425.2715S}. Therefore, in the following, we have adopted the results for solar metallicity ([Z]\,=\,0.020) as best approximation. Assuming isochrones for lower [Z] values would lead to the derivation of older fractional ages (cf. Figure \ref{plot_different_metallicities}).
\citet{hubrig00} put forth the hypotheses that mCP stars with masses $M$\,$<$\,3\,M$_\odot$ are concentrated towards the centre of the main-sequence band and that magnetic fields only appear in stars that have completed about 30\,\% of their lifetime on the main sequence. In their investigation of the evolutionary status of mCP stars, \citet{kochukhov06} demonstrated that mCP stars with $M$\,$>$\,3\,$M_\odot$ are distributed homogeneously among the main sequence. They further identified 22 young ($\tau$\,$<$\,30\,\%) mCP stars among their sample stars with $M$\,$\le$\,3\,$M_\odot$, thereby rejecting the proposal of \citet{hubrig00} that all observably magnetic low-mass CP stars have completed a significant fraction of their main-sequence lifetime. That very young (ZAMS to 25\% on the main sequence) mCP stars do exist has been unambiguously demonstrated by several studies on the basis of members of open clusters (cf. e.g. \citealt{bagnulo03}, \citealt{poehnl03}, \citealt{landstreet07}, and \citealt{landstreet08}).
Nevertheless, \citet{kochukhov06} also find an uneven distribution for mCP stars with masses of $M$\,$<$\,3\,M$_\odot$, in particular for stars with $M$\,$\le$\,2\,$M_\odot$, which tend to cluster in the centre of the main-sequence band. Confirming their previous results, \citet{Hubrig2007} again found that mCP stars with $M$\,$<$\,3\,M$_\odot$ are concentrated towards the centre of the main-sequence band.
As is obvious from Figure \ref{plot_age_mass}, most of our sample stars are situated in the 2\,$\le$\,$M_\odot$\,$\le$3 bin. In agreement with the results of the aformentioned studies, we also find an uneven distribution of the fractional lifetime; however, our sample stars boast a mean fractional age of $\tau$\,=\,63\,\% (standard deviation of 23\,\%; cf. Figure \ref{plots_age}). Young mCP stars, while undoubtedly present, are conspicuously underrepresented in our sample.
In summary, our results strongly suggest an inhomogeneous age distribution among low-mass ($M$\,$<$\,3\,M$_\odot$) mCP stars as hinted at by previous studies. However, we stress that our sample is biased towards mCP stars showing a conspicuous 5200\,\AA\ flux depression in the low-resolution LAMOST spectra and has not been selected on the basis of an unbiased, direct detection of a magnetic field. Therefore, our results have to be viewed with caution and their general validity needs to be tested by a more extended sample selected via different methodological approaches. It remains to be sorted out in what way the occurrence of the 5200\,\AA\ depression is connected with this result, in particular why this feature is apparently much more prominent in older stars. Several studies have shown that the 5200\,\AA\ depression increases with magnetic field strength and the atmospheric metal content (e.g. \citealt{2005A&A...433..671K}; \citealt{2006A&A...448.1153K}).
Our analysis has been based on a statistically significant sample of mCP stars. Furthermore, due to the applied methods, it is not impacted by potential error sources that have been proposed to influence the results of former studies, such as a displacement of stars from the ZAMS by the application of negative Lutz-Kelker corrections or incorrect ${T}_{\rm eff}$ values caused by the anomalous flux distributions of mCP stars (cf. e.g. the discussions in \citealt{kochukhov06} and \citealt{netopil08}). Even if the here derived error margins had been significantly underestimated, which we see no reason to believe, the general conclusion would hold. However, we caution that individual [Z] values and their errors are not available for our sample stars and that the influence of [Z] on the derived fractional ages is large.
\begin{figure}
\includegraphics[width=\columnwidth]{plot_sequences_Zs.eps}
\caption{Lines of constant fractional ages on the main sequence ($\tau$) for solar metallicity [Z]\,=\,0.020 (upper panel). Lower panel shows the positions of the ZAMS and TAMS for isochrones with [Z]\,=\,0.008, 0.020, and 0.060. Values have been chosen to cover the main range of [Z] values found for CP2 stars (Figure \ref{plot_Z}).}
\label{plots_sequences_Zs}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{plot_different_metallicities.eps}
\caption{Mass versus fractional age on the main sequence ($\tau$) distributions for isochrones with [Z]\,=\,0.008, 0.020, and 0.060, illustrating the differences in the derived mass and age distributions.}
\label{plot_different_metallicities}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{plots_age.eps}
\caption{Distribution of fractional ages on the main sequence ($\tau$) among the 903 sample stars fulfilling our accuracy criteria.}
\label{plots_age}
\end{figure}
\subsection{Space distribution} \label{space_distribution}
\begin{figure}
\includegraphics[width=\columnwidth]{map.eps}
\caption{Distribution of the 942 stars with absolute parallax errors less than 25\% in the [XY] plane. Stars were divided in probable members of the thin and thick disk according to the scale heights given in \citet{2001MNRAS.322..426O,2017MNRAS.470.2113A}. Ten stars have Z values larger than 1200\,pc and might be Halo objects.}
\label{map_3D}
\end{figure}
\begin{table*}
\caption{Kinematic and astrometric data for the ten stars of our sample with a height larger than 1200\,pc above the Galactic plane. The columns denote: (1) Internal identification number. (2) LAMOST ID. (3) MKCLASS final type. (4) X-coordinate towards the Galactic centre. (5) Y-coordinate in direction of Galactic rotation. (6) Z-coordinate towards the north Galactic pole. (7) Radial velocity ($R_{\mathrm{V}}$) taken from LAMOST DR4. (8) Standard deviation of $R_{\mathrm{V}}$. (9) Total spatial velocity ($v_{\mathrm{tot}}$). (10) Standard deviation of $v_{\mathrm{tot}}$.}
\label{space_3D}
\begin{center}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{llllllllll}
\hline
\hline
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) \\
No. & LAMOST\_ID & SpT\_final & X (pc) & Y (pc) & Z (pc) & $R_{\mathrm{V}}$ (km\,s$^{-1}$) & $\sigma R_{\mathrm{V}}$ (km\,s$^{-1}$) &
$v_{\mathrm{tot}}$ (km\,s$^{-1}$) & $\sigma v_{\mathrm{tot}}$ (km\,s$^{-1}$) \\
\hline
816 & J073950.01+201812.7 & B9 IV$-$V SrCrEu & $-$3305 & $-$1170 & +1239 & +71 & 4 & 74 & 7 \\
859 & J091053.70+285032.7 & A9 V SrCrEu & $-$1592 & $-$496 & +1487 & $-$24 & 9 & 44 & 11 \\
870 & J104323.95+045214.3 & A1 IV$-$V CrEu & $-$440 & $-$875 & +1265 & +23 & 3 & 30 & 4 \\
873 & J114130.23+403822.7 & B9 IV$-$V CrEuSi & $-$480 & +125 & +1380 & $-$43 & 5 & 98 & 16 \\
875 & J122139.23+383309.5 & kA2hA4mA7 bl4077 bl4130 & $-$346 & +202 & +1732 & $-$42 & 1 & 133 & 10 \\
876 & J122746.05+113635.3 & B8 IV Si (He-wk) & +142 & $-$665 & +2298 & +182 & 5 & 292 & 39 \\
879 & J140422.54+044357.9 & kA4hA7mF0 SrCrEu & +703 & $-$202 & +1355 & $-$55 & 15 & 200 & 23 \\
880 & J150331.87+093125.4 & A8 V SrCrEu & +1233 & +213 & +1736 & $-$38 & 5 & 312 & 33 \\
881 & J155549.85+401144.4 & B8 IV Si & +1491 & +3070 & +4063 & +90 & 5 & 96 & 10 \\
895 & J181156.38+523411.4 & B7 IV$-$V Si & +635 & +3869 & +2002 & $-$6 & 5 & 147 & 31 \\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table*}
To investigate the location of our sample stars in the Galactic [XYZ] plane, the corresponding coordinates were obtained from a conversion of spherical Galactic coordinates (latitude and longitude) to Cartesian coordinates using the distance $d$ from \citet{2018AJ....156...58B}. In this work, the positive X-axis points towards the Galactic centre, the Y-axis is positive in the direction of Galactic rotation and the positive Z-axis points towards the north Galactic pole. Only objects with absolute parallax errors less than 25\,\% were considered in this process. 942 stars satisfied this criterion.
We divided our sample in candidate members of the thin disk (scale height of 350\,pc) and the thick disk (1200\,pc). The scale heights were taken from \citet{2001MNRAS.322..426O} and \citet{2017MNRAS.470.2113A}. Our results are shown in Figure \ref{map_3D}. From the sample of 942 stars, 797 objects likely belong to the thin disk and 135 objects to the thick disk. The remaining ten stars qualify for being members of the halo population and are thus worth a closer look.
As a first step, we checked the spectra of the halo star candidates and confirmed that all objects exhibit the typical spectral features of mCP stars and a clearly visible flux depression. For the calculation of the total spatial velocity $v_{\mathrm{tot}}$, the radial velocity (RV) is needed. Data from the LAMOST survey include automatically measured RV information of the spectra \citep{anguiano18}. We checked the reliability of these values for mCP stars, that is, the spectral range from late B- to early F-type objects, by searching for common entries with the RV catalogue of \citet{2007AN....328..889K}. In total, 11 stars were found that are common to both our sample and this catalogue. Some of these objects boast more than one spectrum in LAMOST DR4; in these cases, mean RV values were calculated. Comparing the RV values from both sources, we find a mean difference of +2.4\,km\,s$^{-1}$, which lends confidence that the LAMOST RVs are useful in a statistical sense. However, we caution that an external uncertainty we cannot account for is introduced by the spot-induced RV variations of mCP stars that can reach up to $\pm$50\,km\,s$^{-1}$ \citep{1999A&A...351..283P}.
The space velocities were calculated following the formulae of \citet{1987AJ.....93..864J}. The final values are listed in Table \ref{space_3D}. Stars of the halo population show $v_{\mathrm{tot}}$\,>\,180\,km\,s$^{-1}$ as compared to the local standard of rest \citep{2004AJ....128.1177V}. We therefore conclude that the stars LAMOST J122746.05+113635.3 (\#876; Gaia DR2 3907547639444408064) and LAMOST J150331.87+093125.4 (\#880; Gaia DR2 1167894108493926016) are kinematically true halo objects, which is of considerable interest as no halo CP2 stars have been discovered so far. Considering the error, the star LAMOST J140422.54+044357.9 (\#879) does not satisfy this criterion.
\citet{1996ApJS..103..433B} identified LAMOST J122746.05+113635.3 as candidate field horizontal-branch star. Its spectrum, however, is that of a Si CP2 star (Figure \ref{comparison_halo}). There are very strong \ion{Si}{ii} lines at 3856/62\,\AA, 4128/31\,\AA, 4200\,\AA, 5041/56\,\AA, and 6347/71\,\AA. In addition, the \ion{He}{i} lines are weak, which is commonly observed in CP2 stars. It has consequently been classified as B8 IV Si (He-wk) by the MKCLASS\_mCP code. Almost all blue horizontal-branch stars, on the other hand, are metal-weak and their spectra rather resemble that of $\lambda$ Bootis stars \citep{gray09}. We feel therefore safe in rejecting the proposed horizontal-branch classification.
In summary, according to the available evidence, LAMOST J122746.05+113635.3 and LAMOST J150331.87+093125.4 are bona-fide CP2 stars whose distances and kinematical properties are in agreement with halo stars. If confirmed, they would be the first CP2 halo objects known and therefore of great interest.
LAMOST J155549.85+401144.4 (\#881; Gaia DR2 1382933122321062912) is another interesting object because it is listed in the catalogue of hot subdwarfs by \citet{2017A&A...600A..50G}. The location in the $(BP - RP)_0$ versus $M_{\mathrm{G,0}}$ diagram (Figure \ref{CMD}), however, does not support this classification. The same is true for the spectrum, which is that of a classical Si CP2 star (MKCLASS final type: B8 IV Si). We note that the occurrence of abundance anomalies in hot subdwarfs has been well established; for example, \citet{2018MNRAS.473.4021W} identified two hot subdwarfs with effective temperatures of about 37\,000\,K and enrichments of 1.5 to 3\,dex in heavy metals. This, however, is very different from what we see in LAMOST J155549.85+401144.4, which is significantly cooler than that ($\sim$14\,000\,K) and shows the abundance pattern of a CP2 star. The current evidence, therefore, points to it being no subdwarf but a Si CP2 star.
Although the LAMOST survey is avoiding dense regions such as star clusters, we searched for possible cluster members among our sample stars. To this end, the positions, diameter, proper motions, distances and their errors of star clusters from \citet{2013A&A...558A..53K} and \citet{2018A&A...618A..93C} were employed and we searched for matches within 3\,$\sigma$ of these parameters. In total, seven matches in six open clusters were found, which are listed in Table \ref{OCs}. Judging from a comparison of the ages derived from Figure \ref{CMD} and the cluster ages, all objects seem to be true cluster members.
\begin{table}
\caption{Open cluster members among our sample stars. The columns denote: (1) Internal identification number. (2) LAMOST ID. (3) Open cluster.}
\label{OCs}
\begin{center}
\begin{tabular}{lll}
\hline
\hline
(1) & (2) & (3) \\
No. & LAMOST\_ID & Open cluster \\
\hline
203 & J041641.15+511253.2 & NGC 1528 \\
269 & J045933.63+395715.8 & Alessi 2 \\
317 & J052059.29+351123.5 & Gulliver 8 \\
497 & J060815.12+045107.6 & NGC 2168 \\
502 & J060827.84+204832.0 & NGC 2168 \\
675 & J064741.02+072458.8 & Collinder 115 \\
868 & J095855.77-044413.8 & Collinder 359 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\includegraphics[width=\columnwidth]{comparison_halo.eps}
\caption{Blue-violet spectra of the proposed halo stars LAMOST\,J122746.05+113635.3 (\#876; MKCLASS final type B8 IV Si (He-wk); upper spectrum) and LAMOST\,J150331.87+093125.4 (\#880; MKCLASS final type A8 V SrCrEu; lower spectrum). Some prominent lines of interest are indicated.}
\label{comparison_halo}
\end{figure}
\subsection{Peculiarity type distribution} \label{peculiarity_types}
\begin{figure*}
\includegraphics[width=\textwidth]{peculiarity_types.eps}
\caption{Fractional distribution of chemical peculiarities versus hydrogen-line spectral type for the 876 stars with unambiguous peculiarity type identifications. The numbers above the bars indicate the number of objects in the corresponding spectral type bin. Because a single object may have multiple peculiarities, fractions may exceed 1.}
\label{peculiarity_type_distribution}
\end{figure*}
Figure \ref{peculiarity_type_distribution} explores the distribution of Si, Cr, Sr, and Eu peculiarities versus hydrogen-line spectral type for the 876 stars of our sample with unambiguous identifications (i.e. without the stars in which only strong 4077\,\AA\ and/or 4130\,\AA\ blends or no traditional peculiarities were identified). Stars with hydrogen-line types of B9.5 are not listed separately but included under spectral type B9. The number of stars per spectral type bin varies considerably, with the B8$-$A0 stars forming the vast majority of our sample. Nevertheless, some tentative trends can be identified, although the interpretation towards the low temperature end is severely hampered by the small number statistics for objects later than A9:
\begin{itemize}
\item Si peculiarities are present between spectral types B4 and F0. They play a dominant role in stars with spectral types earlier than B9, strongly decreasing in importance in later-type objects. Except for He peculiarites (which are not shown in the plot), Si peculiarities are the only chemical peculiarities identified in objects earlier than B8.
\item Cr peculiarities set in at spectral type B8 and form an important part of the peculiarity mix between spectral types B9 and A9.
\item Sr and Eu peculiarities set in at spectral type B8 and increase in strength towards later types.
\end{itemize}
This is in good agreement with the expectations and the literature. It is well known that Si peculiarities are present throughout a wide range of effective temperatures in mCP stars \citep[e.g.][]{RM09}. The hottest stars with Eu peculiarities from the \citet{RM09} catalogue are of spectral type B8, which holds true also for the vast majority of stars with Cr and Sr peculiarities, with the exception of only three objects (HD 35502, spectral type B6 Sr Cr Si; HD 167288, spectral type B7 Si Cr; HD 213918, spectral type B7 Si Sr). Likewise, the work of \citet{ghazaryan18} contains atomic data for Eu, Cr, and Sr from effective temperatures of, respectively, 12\,900\,K ($\sim$B8), 14\,700\,K ($\sim$B6), and 13\,300\,K ($\sim$B8) downwards. The good agreement of the peculiarity type 'blue borders' between the present work and the literature provides independent proof of the reliability of the here derived spectral types.
\begin{figure}
\includegraphics[width=0.5\textwidth]{bl4077bl4130.eps}
\caption{Distribution of stars in which only strong blends at 4077\,\AA\ and/or 4130\,\AA\ were identified.}
\label{bl4077bl4130}
\end{figure}
Figure \ref{bl4077bl4130} illustrates the distribution of stars in which only strong blends at 4077\,\AA\ and/or 4130\,\AA\ were identified. Again, stars with hydrogen-line types of B9.5 are included under spectral type B9. These stars were not assigned Si, Cr, Sr, and Eu types with the here employed workflow because, apart from the strong blends, the peculiarities are either too subtle to have passed our significance criteria, no other significant features are present or the code failed to identify them for some reason (cf. Section \ref{spectral_classification_MKCLASS}).
Manual classification is necessary to throw more light on what elements contribute to the observed blends. Nevertheless, the distribution of the 4130\,\AA\ blend identifications, in particular for objects earlier than B9, is in general agreement with the distribution of Si peculiarities. We therefore expect that most of the 'bl4130' stars in our sample will turn out to be Si stars. No similar predictions can be made for the 'bl4077' stars from the available data.
\subsection{Comparison with samples from the literature} \label{discussion_literature}
The following sections compare our results with the works of \citet{RM09}, \citet{Skiff}, and \citet{qin19}. We further note that 22 of our sample stars are contained in the sample of strongly magnetic Ap stars of \citet{Scholz2019}. As the authors do not list spectral types, a direct comparison of results was not possible. The stars common to both samples are identified in Table \ref{table_master1}.
\subsubsection{Comparison with the compilations of \citet{RM09} and \citet{Skiff}} \label{discussion_RM09}
Our final sample contains 59 mCP stars or candidates that are also included in the catalogue of \citet{RM09}. This low level of coincidence (6.65\,\%) is expected because the \citet{RM09} sample mostly consists of bright stars (peaking at around 9th magnitude) for which there are no LAMOST spectra available.
Table \ref{table_comparison_RM09} gives a comparison of the spectral types from the present study, the RM09 catalogue, and the compilation of \citet{Skiff}. For most stars, there is a good general agreement between the different sources. For example, for the 46 stars that have at least one other detailed literature classification listing the temperature subtype, the determined hydrogen-line types agree within $\pm$2\,subclasses, which seems reasonable considering the inhomogeneous source material behind the literature classifications and the inherent difficulties in classifying mCP stars. For several stars, our results provide a first detailed classification; furthermore, we confirm several doubtful objects as mCP stars and show that some suspected CP1 stars are in fact mCP stars. A more detailed investigation into this matter will be the topic of an upcoming study that will be concerned with a new classification of stars in the RM09 catalogue based on homogeneous spectroscopic material.
\subsubsection{Comparison with the sample of \citet{qin19}} \label{discussion_Qin}
\citet{qin19} searched for CP1 stars in low resolution spectra of early-type stars from LAMOST DR5 and compiled a catalogue of 9372 CP1 stars. Because cooler CP2 stars may exhibit similar spectral features (Ca deficiency, overabundance of Fe-group elements), the authors expect a contamination of their sample by these objects. To identify potential CP2 stars among the CP1 star candidates, they used the 4077\,\AA\ blend as reference line, which may contain contributions from \ion{Si}{ii}, \ion{Cr}{ii}, and \ion{Sr}{ii}. Synthetic spectra with overabundances of Sr, Cr, Eu, and Si of 2.0 dex were computed for different effective temperatures and the equivalent widths of the 4077\,\AA\ blend were calculated and compared for both the templates and the observed spectra. If the equivalent width of the observed 4077\,\AA\ feature ($EW$\textsubscript{4077\_obs}) exceeded that of the corresponding templates ($EW$\textsubscript{4077\_temp}), a star was flagged as a CP2 star candidate. In this way, \citet{qin19} flagged 1131 stars within their sample of CP1 star candidates as CP2 star candidates. From a cursory investigation of about 20 randomly chosen objects, several bona-fide CP2 stars have indeed been found among the objects with high values of ($EW$\textsubscript{4077\_obs}\,$-$\,$EW$\textsubscript{4077\_temp}), in line with the expectations of \citet{qin19}. However, the incidence of CP2 stars seems to drop rather sharply towards lower values of ($EW$\textsubscript{4077\_obs}\,$-$\,$EW$\textsubscript{4077\_temp}). We assume that this is because a strong 4077\,\AA\ feature alone, while often helpful, is an insufficient criterion for identifying CP2 stars.
Only 45 objects are common to both our sample and the \citet{qin19} catalogue. Of these stars, about 70\,\% (31 objects) were flagged as CP2 star candidates. Table \ref{table_comparison_Qin} compares the \citet{qin19} k/h/m spectral types to the final types derived in the present study for all objects common to both lists. In general, the agreement between the derived spectral types is poor. With the here employed methodology, different k/h/m-types were assigned to only 21 of these stars. A detailed investigation of the stars common to both lists, in particular a comparison of the automatically derived classifications to manually derived spectral types and an investigation of the source(s) of the observed discrepancies, is beyond the scope of the present paper but might be beneficial to both studies and help with the refinement of the employed algorithms.
On first glance, the low level of coincidence between the \citet{qin19} catalogue and the here presented sample of mCP seems surprising. We have identified several reasons for this, which are related to the different approaches and goals of both studies.
\citet{qin19} explicitly searched for CP1 stars. CP2 stars were identified as a contaminant and corresponding candidates were subsequently identified. Because they searched for CP2 star candidates within their sample of CP1 star candidates, that is, among objects with pronounced differences between k and h spectral types, their sample will not contain any CP2 stars that do not share these characteristics. However, early-type CP2 stars generally do not show significant (if any) differences between k and h types; this phenomenon is mostly restricted to late-type CP2 stars. Thus, we assume that the \citet{qin19} subsample of CP2 star candidates is biased towards late-type CP2 stars.\footnote{Although not statistically significant, it is interesting to note that all stars from the \citet{qin19} candidate sample we were able to confirm as CP2 stars were indeed late-type CP2 stars.} This is further corroborated by the fact that, in agreement with the expectations for identifying CP1 stars, \citet{qin19} constrained their search to objects with 6500\,K\,<\,${T}_{\rm eff}$\,<\,11\,000\,K. In fact, their final catalogue contains only 11 objects with ${T}_{\rm eff}$\,>\,10\,000\,K. Thus, only very few stars hotter than spectral type A0 are present in their sample.
Our study follows a very different approach and is concerned with the identification of mCP stars that were selected among early-type targets by the presence of a significant 5200\,\AA\ flux depression. The spectra of CP1 stars, on the other hand, generally do not show this feature \citep{2005A&A...441..631P}. However, not all mCP stars distinctly show a 5200\,\AA\ depression either, in particular in low-resolution spectra. Therefore, we will have missed any such mCP stars, which might have found their way into the \citet{qin19} candidate sample. Most important, however, is that the majority of our sample stars is situated in the spectral range of B8 to A0 (9700\,K\,<\,${T}_{\rm eff}$\,<\,12\,500\,K), which renders them mostly incompatible with the candidate sample of \citet{qin19}.
In summary, the above mentioned issues, combined with the fact that the incidence of CP2 stars seems to drop rather sharply towards lower values of ($EW$\textsubscript{4077\_obs}\,$-$\,$EW$\textsubscript{4077\_temp}) in the CP2 star candidate subsample of \citet{qin19} and the different source material ($\sim$9.0 million spectra in DR5 vs. $\sim$7.6 million spectra in DR4), illustrate that no significant overlap between both samples is to be expected. We would like to again stress that it never was the intention of \citet{qin19} to collect a pure sample of CP2 stars. Their subsample of candidates, however, is a valuable starting point for further investigations. As a spin-off of the present study, we intend to investigate the \citet{qin19} CP2 star candidates to confirm or reject their status as mCP stars. It is clear that investigations based on a broader analysis of LAMOST early-type spectra (i.e. also including stars without conspicuous 5200\,\AA\ flux depressions) will lead to the discovery of many more mCP stars in the future.
\begin{table*}
\caption{Comparison of the spectral types derived in this study, the RM09 catalogue, and the compilation of \citet{Skiff}. The columns denote: (1) Internal identification number. (2) Identification number from the RM09 catalogue. (3) LAMOST identifier. (4) Spectral type, as derived in this study. (5) Spectral type from the RM09 catalogue. An asterisk (*) or a question mark (?) in parantheses behind the spectral type denote, respectively, well-known CP2 stars and CP2 stars of doubtful nature. (6) Spectral type from the compilation of \citet{Skiff}. (7) Corresponding reference from the compilation by \citet{Skiff}.}
\label{table_comparison_RM09}
\begin{center}
\begin{adjustbox}{max width=0.9\textwidth}
\begin{tabular}{cclllll}
\hline
\hline
(1) & (2) & (3) & (4) & (5) & (6) & (7) \\
\textbf{No.} & \textbf{ID\_RM09} & \textbf{ID\_LAMOST} & \textbf{SpT\_final} & \textbf{SpT\_RM09} & SpT\_Skiff & Ref\_Skiff \\
\hline
14 & 1236 & J004947.16+525208.2 & B9 V CrEuSi (He-wk) & B9 Si & ApSi & \citet{1998PASP..110..270B} \\
19 & 1543 & J010007.84+045339.3 & kB8hA3mA6 Cr & A0 Si Cr (?) & A0/2 & \citet{hou15} \\
22 & 1680 & J010651.35+154426.9 & kA1hA7mA7 SrCrEu & A3 & A3p & \citet{1955ApJ...121...32N} \\
33 & 2210 & J013055.87+452155.7 & kA1hA3mA4 SrCr & A0 Sr & A0pSrCr? & \citet{1975AAS...21...25F} \\
38 & 2750 & J015059.58+540259.1 & B8 IV bl4130 & B8 Si & B8IIIpSi & \citet{1999AAS..137..451G} \\
61 & 3660 & J022120.02+280415.6 & A8 V SrEuSi & A2 Sr Eu & ApSrEu & \citet{1983AJ.....88.1182B} \\
77 & 4580 & J025945.31+541941.5 & kB7hA7mA6 CrEuSi & A2 Si Cr Eu & B9pSi & \citet{1999AAS..137..451G} \\
80 & 4740 & J030350.21+463718.3 & A8 V SrCrEuSi & Sr Eu & ApSrEu Sr vstrong & \citet{1985AJ.....90..341B} \\
83 & 4820 & J030633.66+025615.7 & B9.5 IV$-$V CrEu & A2 Cr Eu (?) & A2pCrEu Sr 4077A weak & \citet{1973AJ.....78...47D} \\
130 & 5750 & J034000.57+444858.4 & kA0hA1mA3 (Si) & A0 & n/a & n/a \\
132 & 5800 & J034112.38+453031.7 & A1 V SrCrEu & Sr Eu & ApSrEu & \citet{1985AJ.....90..341B} \\
134 & 5850 & J034229.41+353820.9 & A0 IV$-$V bl4077 bl4130 & A1 Sr & A1pSr & \citet{1977AJ.....82..598G} \\
138 & 5860 & J034417.13+494336.6 & B8 IV$-$V CrSi & A0 Cr Eu (?) & ApCrEu: & \citet{1985AJ.....90..341B} \\
180 & 6530 & J040642.34+454640.8 & kB9hA2mA6 CrSi & A0 Si & ApSi & \citet{1985AJ.....90..341B} \\
228 & 7210 & J042736.18+063643.1 & A2 IV$-$V SrCrEu & A0 Sr Cr Eu & ApSrCrEu & \citet{1973AJ.....78..687B} \\
241 & 7740 & J044407.32-005639.0 & F0 V SrEuSi & F0 Sr Eu & FpSrEu & \citet{1973AJ.....78..687B} \\
252 & 7960 & J045121.11+093555.8 & A0 V SrCrEuSi & A0 Sr Cr Eu & ApSrCrEu & \citet{1973AJ.....78..687B} \\
275 & 8140 & J050210.72+464600.0 & B8 III$-$IV Si & Si & ApSi & \citet{1985AJ.....90..341B} \\
318 & 8872 & J052118.97+320805.7 & B4 Vpn & B8 Si Sr & B9 & \citet{1931AnHar.100...61C} \\
325 & 8938 & J052259.54+343944.9 & B9.5 V Cr & B9 Si Cr Sr & A0/2 & \citet{hou15} \\
334 & 9082 & J052616.48+331544.2 & B8 IV$-$V bl4130 & A0 Si Sr & A2 & \citet{1931AnHar.100...61C} \\
343 & 9200 & J052812.16+415006.4 & kA0hA3mA7 Si & B5 Si & ApSi & \citet{1983AJ.....88.1182B} \\
351 & 9295 & J053025.32+332639.6 & A6 V SrEu & A1- & A7 & \citet{1931AnHar.100...61C} \\
360 & 9350 & J053239.91+434307.5 & B8 IV Si & B9 Si & B8pSi & \citet{1975AAS...21...25F} \\
371 & 9740 & J053504.75-012406.5 & kA0hA2mA4 CrEu & A2 Cr Eu & A0pSi & \citet{1983AA...118..102G} \\
393 & 10107 & J054102.12+332331.1 & B7 IV Si (He-wk) & B9 Si (?) & n/a & n/a \\
403 & 10323 & J054630.44+273518.1 & B9 V CrEu & B8 Si (?) & n/a & n/a \\
411 & 10375 & J054819.92+333516.9 & kA4hA9mF1 SrCrEuSi & A5 Si Sr & A7 & \citet{1931AnHar.100...61C} \\
409 & 10385 & J054757.12+235011.8 & B9.5 II$-$III EuSi & B9 Si Sr & B8 & \citet{1931AnHar.100...61C} \\
418 & 10447 & J055002.80+234023.9 & B9 IV bl4130 & B9 Si Cr & B9 & \citet{1931AnHar.100...61C} \\
428 & 10460 & J055121.05+420610.5 & B8 IV Si & A Si & ApSi & \citet{1983AJ.....88.1182B} \\
436 & 10536 & J055237.95+274922.8 & A6 V SrCrEu & A2- & n/a & n/a \\
445 & 10602 & J055422.76+305401.8 & B8 III bl4130 & A0- & A2 & \citet{1931AnHar.100...61C} \\
470 & 10900 & J060045.94-035344.3 & A0 IV$-$V Si & A0 Cr Eu & A0VpSi & \citet{1999AAS..137..451G} \\
476 & 10915 & J060227.33+282943.9 & B8 III$-$IV EuSi (He-wk) & A0 Si & A3 & \citet{1931AnHar.100...61C} \\
475 & 10917 & J060225.92+244628.5 & B8 IV Si & B9 Si & B8/9.5IV/VSi: & \citet{1979RA......9..479C} \\
560 & 11800 & J062155.55+001812.2 & B8 III Si & A0 Si Sr & A0pSi & \citet{1999AAS..137..451G} \\
564 & 11810 & J062257.61+231625.8 & A1 IV$-$V SrCrEu & A2 Sr & ApSr & \citet{1983AJ.....88.1182B} \\
597 & 12200 & J062914.34+004257.0 & B7 III$-$IV Si & B9 Si & ApSi & \citet{1973AJ.....78..687B} \\
604 & 12300 & J063035.50+035245.3 & A1 IV$-$V SrCr & A2 Sr Eu & ApSrEu & \citet{1973AJ.....78..687B} \\
611 & 12360 & J063218.45+032146.3 & B9 III$-$IV Si (He-wk) & B9 Si & ApSi & \citet{1973AJ.....78..687B} \\
627 & 12630 & J063744.29+195655.1 & kA1hA7mF4 SrCrSi & A Sr Eu & ApSrEu & \citet{1983AJ.....88.1182B} \\
630 & 12690 & J063752.90+091516.7 & B8 IV Si & B9 Si & ApSi & \citet{1973AJ.....78..687B} \\
715 & 13790 & J065403.63+221545.2 & A6 IV SrCrEu & A2 Sr Eu & ApSrEu & \citet{1972PASP...84..446B} \\
721 & 13980 & J065458.31+040826.9 & kA1hA3mA6 SrCrEu & A2 Sr Cr Eu & FpSrCrEu & \citet{1973AJ.....78..687B} \\
753 & 14520 & J070252.77+023700.0 & kB9hA9mA7 SrSi & A2 Si & ApSi & \citet{1973AJ.....78..687B} \\
763 & 14740 & J070617.23+101601.6 & B9 IV EuSi & A0 Si & ApSi & \citet{1973AJ.....78..687B} \\
774 & 15123 & J071337.30+040720.7 & B8 IV CrEuSi & B9 Si & n/a & n/a \\
792 & 15650 & J072118.92+223422.7 & B9 V bl4077 & Sr (?) & ApSr & \citet{1983AJ.....88.1182B} \\
827 & 17490 & J074851.40+001619.1 & kB8hA3mA3 CrEu & Cr Eu & ApCrEu & \citet{1973AJ.....78..687B} \\
829 & 17540 & J074959.61+013517.8 & kA1hA3mA7 SrCrEuSi & Sr Cr Eu & ApSrCrEu & \citet{1973AJ.....78..687B} \\
830 & 17630 & J075041.80-060338.3 & A2 IV SrCrEu & A0 Cr Eu & FpCrEu & \citet{1973AJ.....78..687B} \\
838 & 18380 & J080339.87-082141.0 & kB9hA3mA8 SrCrEu & A0 Cr Eu & ApCrEu & \citet{1973AJ.....78..687B} \\
867 & 24620 & J095644.95-021719.5 & kA2hA3mA6 SrCrEu & A2 Sr Eu Cr & ApSrCrEu & \citet{1981AJ.....86..553B} \\
873 & 29280 & J114130.23+403822.7 & B9 IV$-$V CrEuSi & A0- (?) & A0m: & \citet{1959AAHam...5..105S} \\
877 & 31550 & J122855.36+255446.3 & B9 V CrEu & A0 Sr Cr Eu (*) & B8pSiCrSr & \citet{1990AAS...85.1069S} \\
967 & 59010 & J222549.96+343851.0 & B8 IV EuSi & A0 Si (?) & ApSi: & \citet{1985AJ.....90..341B} \\
980 & 60185 & J230905.79+523711.2 & B8 IV EuSi & B8 Si & ApSi & \citet{1998PASP..110..270B} \\
1001 & 61520 & J235740.51+470001.7 & B9 IV CrSi & B8 (?) & n/a & n/a \\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table*}
\begin{table}
\caption{Comparison of the spectral types derived in this study to the catalogue of \citet{qin19}. The columns denote: (1) Internal identification number. (2) LAMOST identifier. (3) Spectral type, as derived in this study. (4) Spectral type from \citet{qin19}. (5) Ap\_flag from \citet{qin19}. A value of 1 indicates that the star is a CP2 star candidate.}
\label{table_comparison_Qin}
\begin{center}
\begin{adjustbox}{max width=0.5\textwidth}
\begin{tabular}{clllc}
\hline
\hline
(1) & (2) & (3) & (4) & (5) \\
\hline
\textbf{No.} & \textbf{ID\_LAMOST} & \textbf{SpT\_final} & \textbf{SpT\_Qin} & \textbf{Ap\_flag} \\
2 & J000834.32+321205.4 & A2 IV SrCrEu & kA2hA7mA7 & 1 \\
12 & J003425.61+452108.7 & kA1hA7mA9 SrEuSi & kA2hA3mA5 & 1 \\
25 & J011435.00+534757.2 & A1 IV$-$V SrCrEuSi & kA1hA5mA5 & 1 \\
29 & J012028.54+480545.6 & kA4hA7mA8 SrCrEu & kA5hA6mA7 & 1 \\
30 & J012122.31+442826.2 & kA3hA7mA8 CrEu & kA5hA5mA7 & 1 \\
35 & J014508.87+322430.3 & kA1hA6mA7 SrCrEu & kA1hA5mA5 & 1 \\
68 & J023936.01+540059.5 & kA2hA5mA7 CrEu & kA6hA5mA5 & 0 \\
70 & J024243.57+453720.4 & A1 V SrCrEuSi & kA6hA6mA7 & 1 \\
106 & J032333.76+510336.6 & A2 V CrEuSi & kA3hA6mA7 & 0 \\
218 & J042235.95+411448.9 & A1 IV SrCrEuSi & kA6hA5mA7 & 1 \\
235 & J043752.46+533259.6 & B9.5 V CrSi & kA0hA1mA5 & 1 \\
244 & J044446.24+513129.2 & A9 V SrEu & kA5hA5mF0 & 0 \\
246 & J044713.47+540515.5 & A2 IV CrEu & kA6hA6mF0 & 0 \\
252 & J045121.11+093555.8 & A0 V SrCrEuSi & kA2hA1mA6 & 1 \\
259 & J045508.24+204943.7 & kA2hA4mA7 SrCrEu & kA6hA5mA7 & 1 \\
286 & J050925.66+512444.9 & A3 III$-$IV CrEu & kA5hA6mA7 & 0 \\
330 & J052459.98-064651.8 & A0 IV$-$V Si & kA3hA5mA7 & 1 \\
332 & J052552.24+344817.3 & A0 V CrEu & kA1hA1mA5 & 1 \\
344 & J052816.11-063820.1 & kA1hA9mA9 SrCrEu & kA2hA7mA6 & 1 \\
387 & J053840.17+413754.0 & kA1hA7mA9 SrCrEuSi & kA3hA6mA7 & 1 \\
423 & J055045.30+372809.0 & kA2hA4mA7 SrCrEu & kA6hA5mA7 & 1 \\
427 & J055110.67+360517.0 & B9.5 V SrCr & kA1hA1mA5 & 1 \\
441 & J055333.92+180202.9 & F2 V SrEu & kA7hA7mF5 & 0 \\
483 & J060356.06+213033.2 & A8 V SrEu & kA5hA6mF0 & 0 \\
542 & J061743.05+595315.3 & A1 IV Eu & kA3hA6mA7 & 1 \\
570 & J062449.08+190854.0 & kA2hA3mA7 SrCrEu & kA3hA5mF0 & 1 \\
571 & J062449.21+161325.7 & A0 V Cr & kA1hA1mA5 & 0 \\
627 & J063744.29+195655.1 & kA1hA7mF4 SrCrSi & kA2hA6mA7 & 1 \\
643 & J064158.41+223927.2 & A5 IV SrCrEu & kA5hA6mA7 & 0 \\
689 & J064907.51+114600.1 & kA1hA7mA8 SrCr & kA2hA6mA5 & 0 \\
690 & J064947.96+202510.8 & kA3hA5mA7 bl4077 bl4130 & kA7hA7mA7 & 0 \\
770 & J070907.08+441114.7 & B9.5 V CrEu & kA1hA6mA5 & 1 \\
773 & J071258.59+065952.3 & A2 IV SrCrEu & kA3hA6mA7 & 1 \\
775 & J071413.88+142449.5 & A3 III$-$IV SrCrEu & kA6hA5mA7 & 1 \\
779 & J071550.38+063655.9 & A1 IV SrCrEuSi & kA2hA5mA7 & 1 \\
814 & J073548.83+123225.1 & kA1hA9mA8 SrEu & kA6hA7mA7 & 1 \\
828 & J074919.49+051551.8 & kA1hA9mA9 SrCrEu & kA2hA7mA5 & 1 \\
831 & J075220.93+113710.6 & kA5hA6mA9 SrCrEu & kA5hA6mA7 & 0 \\
848 & J082137.47+064401.0 & A1 IV SrCrEu & kA2hA5mA5 & 1 \\
850 & J082326.74+072116.4 & kA3hA7mF1 SrCrEu & kA2hA6mA7 & 0 \\
853 & J083539.12+002150.4 & kA2hA3mA6 SrCrEu & kA3hA1mA7 & 0 \\
861 & J092233.23+072519.4 & kA3hA6mA9 SrCrEu & kA6hA5mA7 & 1 \\
879 & J140422.54+044357.9 & kA4hA7mF0 SrCrEu & kA6hA6mA7 & 1 \\
899 & J185127.99+262012.1 & A0 IV$-$V Cr & kA0hA1mA7 & 1 \\
909 & J192524.14+431911.5 & kA3hA7mA9 SrCrEu & kA6hA5mA5 & 1 \\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table}
\begin{figure*}
\includegraphics[width=\textwidth]{hepec.eps}
\caption{Showcase of five newly identified peculiar mid-B type stars, illustrating the blue-violet region of the LAMOST DR4 spectra of (from top to bottom) LAMOST J014940.99+534134.2 (\#37; TYC 3684-1139-1), LAMOST J052118.97+320805.7 (\#318; HD 242764), LAMOST J055023.89+261330.2 (\#421; TYC 1866-861-1), LAMOST J062307.91+264642.0
(\#565; Gaia DR2 3432273606513132544), and LAMOST J062348.46+034201.1 (\#567; HD 256582). MKCLASS final types and manual types derived in the present study are indicated. Some prominent lines of interest are identified. The asterisk marks the position of a 'glitch' in the spectrum of LAMOST J014940.99+534134.2.}
\label{spectra_hepec}
\end{figure*}
\subsection{The mid-B type mCP stars - He-peculiar objects?} \label{Hepec}
23 stars of our sample have MKCLASS final types earlier than B7. More than half of these objects were classified as showing peculiarly weak \ion{He}{i} lines ('He-wk'). Interestingly, three stars were identified as showing both weak and strong ('He-st') \ion{He}{i} lines, which strongly suggests He peculiarity. Besides that, only Si overabundances and strong 4130\,\AA\ blends were identified in several of these objects. As He-rich stars are generally hotter than spectral type B4 \citep{gray09} and therefore not expected to contribute to our sample, we consider these objects good candidates for He-weak (CP4) stars.
Figure \ref{spectra_hepec} showcases the spectra of five mid-B type stars. The hydrogen-line profiles and the prominent \ion{C}{ii} 4267\,\AA\ lines corroborate the classifications, although we manually derived slightly different temperature types in two cases. LAMOST J014940.99+534134.2 (\#37; TYC 3684-1139-1) and LAMOST J062348.46+034201.1 (\#567; HD 256582) are He-weak stars with Si overabundances. LAMOST J062307.91+264642.0 (\#565; Gaia DR2 3432273606513132544) is also certainly He-weak but does not fit any of the standard subclasses of the He-weak stars (the 'hot' Si stars, i.e. Si stars with hotter temperatures than the classical Si CP2 stars; the P Ga stars; the Sr Ti stars; cf. \citealt{gray09}). It is here classified as B4 V HeB7 (R. O. Gray, personal communication).
LAMOST J055023.89+261330.2 (\#421; TYC 1866-861-1) boasts a rather noisy spectrum ($g$ band S/N of 79) that is, apart from the hydrogen lines, basically a 'smattering' of metal-lines, without any particularly outstanding features -- except for the strong C II 4267\,\AA\ and the weak Mg II 4481\,\AA\ lines that support its classification as a mid B-type object. This is also supported by the colour index $(BP-RP){_0}$\,=\,$-0.182$\,mag. The He I lines, then, seem to be curiously absent from its spectrum, so the star may be related to the CP4 stars, although the metal-lines seem way too strong to support this interpretation. We here tentatively classify it as B4: V HeB9. The star shows a strong flux depression at 5200\,\AA, and, according to data from the SuperWASP archive \citep{butters10}, is a periodic photometric variable with a period of about 11.5\,d. It certainly merits a closer look -- this, however, is beyond the scope of the present investigation.
The He lines of LAMOST J052118.97+320805.7 (\#318; HD 242764) do not look weak for its temperature type but have broad profiles suggesting rapid rotation. This, however, is not supported by the hydrogen-line profile, which almost exactly matches that of the B4 V standard. The presence of a number of unidentified metal-lines (which do show evidence for rotational broadening, but not to the extent the He I lines imply) and the conspicuous 5200\,\AA\ depression suggest that this star is indeed chemically peculiar, although it also does not fit into any of the standard mid-B peculiarity subtypes. It has been classified as B4 Vpn in the present study (R. O. Gray, personal communication).
Additional proof that this star is indeed chemically peculiar comes from its periodic photometric variability with a period of about 5.1\,d in SuperWASP data and the conspicuous 5200\,\AA\ flux depression in its spectrum. Incidentally, the star is listed with a spectral type of B8 Si Sr in the \citet{RM09} catalogue, which is not supported by the available LAMOST spectrum. We were unable to get at the root of this classification; however, the star is listed as B9 in the Henry Draper extension \citep{1931AnHar.100...61C} and was classified as B5p by \citet{chargeishvili88}. Further He-peculiar objects and candidates can be gleaned from Table \ref{table_master1}.
\subsection{The eclipsing binary system LAMOST J034306.74+495240.7} \label{eclipser}
The star LAMOST J034306.74+495240.7 (\#135; TYC\,3321-881-1) was identified as an eclipsing binary system in ASAS-SN data \citep{jayasinghe19}. It is listed in the International Variable Star Index of the AAVSO (VSX; \citealt{VSX}) under the designation ASASSN-V J034306.74+495240.8 and with a period of 5.1431\,d. We have analyzed the available ASAS-SN data for this star and derive a period of 5.1435$\pm$0.0012\,d and an epoch of primary minimum at HJD 2457715.846$\pm$0.002. The light curve is shown in Figure \ref{eclipser_lightcurve} and illustrates that the orbit is slightly eccentric, the secondary minimum occurs at an orbital phase of $\varphi$\,=\,0.46. In addition, there is evidence for out-of-eclipse variability in agreement with rotational modulation on one component of the system.
Two spectra are available for this star in LAMOST DR4. The first spectrum ('spectrum1') was obtained on 23 October 2015 (MJD 57318; observation median UTC 17:33:00; $g$ band S/N: 258), which corresponds to an orbital phase of $\varphi$\,=\,0.890. The second spectrum ('spectrum2') was taken on 19 January 2016 (MJD 57406; observation median UTC 11:46:00; $g$ band S/N: 207), which corresponds to an orbital phase of $\varphi$\,=\,0.952. Therefore, both spectra were taking during maximum light, and we find no significant difference between them. Both show a strong flux depression at 5200\,\AA\ and enhanced \ion{Si}{ii} lines at 3856/62\,\AA, 4128/31\,\AA, 4200\,\AA, 5041/56\,\AA, and 6347/71\,\AA. We have analyzed the spectrum with the highest S/N (spectrum1) and derive a spectral type of B9 III Si. Figure \ref{eclipser_spectra} compares the blue-violet part of both spectra to the liblamost B9 III standard, whose hydrogen-line profile provides a good fit to the observed ones.
In summary, we conclude that at least one component of the LAMOST J034306.74+495240.7 system is a Si CP2 star. It is, therefore, of great interest because mCP stars in eclipsing binaries are exceedingly rare \citep{RM09,niemczura17,kochukhov18,skarka19} and accurate parameters for the components can be derived via an orbital solution of the system. We strongly encourage further studies of this interesting object.
\begin{figure}
\includegraphics[width=\columnwidth]{eclipser_lightcurve.eps}
\caption{ASAS-SN light curve of the eclipsing binary system LAMOST J034306.74+495240.7 (\#135; TYC\,3321-881-1). The data have been folded with the orbital period of $P$\textsubscript{orb}\,=\,5.1435$\pm$0.0012\,d.}
\label{eclipser_lightcurve}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{comparison_spectra_eclipser.eps}
\caption{Comparison of the blue-violet spectra of the eclipsing binary system LAMOST J034306.74+495240.7 (\#135; TYC\,3321-881-1) to the \textit{liblamost} B9 III standard spectrum (upper spectrum). Some prominent lines of interest are indicated.}
\label{eclipser_spectra}
\end{figure}
\subsection{The SB2 system LAMOST J050146.85+383500.8} \label{SB2_system}
\begin{figure}
\includegraphics[width=\columnwidth]{SB2.eps}
\caption{Comparison of the blue-violet spectra of the propsed SB2 system LAMOST J050146.85+383500.8 (\#272; HD 280281; MKCLASS final type B8 V Si) to the \textit{libsynth} B8 V standard spectrum (upper spectrum). Some prominent lines of interest are indicated.}
\label{SB2}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{SB2_Hgam.eps}
\caption{Close-up view of the H$\gamma$ region of the proposed SB2 system LAMOST J050146.85+383500.8 (blue spectrum) and the \textit{libsynth} B8 V standard (black spectrum), illustrating the peculiar profile of the H$\gamma$ line indicative of binarity.}
\label{SB2_Hgam}
\end{figure}
Figure \ref{SB2} illustrates the peculiar spectrum of LAMOST J050146.85+383500.8 (\#272; HD 280281), which we suspect to be a blend of two different stars. This becomes especially obvious in the profile of the H$\gamma$ line (Figure \ref{SB2_Hgam}).
To further investigate this matter, we employed the VO Sed Analyzer tool VOSA\footnote{\url{http://svo2.cab.inta-csic.es/theory/vosa/}} v6.0 \citep{2008A&A...492..277B} to fit the SED to the available photometry. For comparison, we used a Kurucz ODFNEW/NOVER model \citep{1997A&A...318..841C} with $T_\mathrm{eff}$\,=\,12\,500\,K, which corresponds to a spectral type of B8. We emphasise that a change of $T_\mathrm{eff}$ of about 2000\,K in either direction will not impact our conclusions. Figure \ref{fig_SED} illustrates the results of the fitting process. The flux model was fitted to either match the ultraviolet or the optical wavelength region. In any case, the discrepancies are readily visible and it is obvious that the observed flux distribution of LAMOST J050146.85+383500.8 cannot be fitted with a single star flux model. We note that it is well known that CP2 stars show a 'blueing' effect \citep{1980A&A....89..230M}, which leads to observed flux discrepancies due to stronger absorption in the ultraviolet than in chemically normal stars. However, a slight shift in the ultraviolet region will not alter our conclusions.
Because the features of the companion star are readily visible in the available LAMOST spectrum, we conclude that its absolute magnitude must be similar to the B-type main-sequence component. We therefore assume that it is a supergiant star with a progenitor of higher mass. Such a combination of components is quite unusual among mCP stars; in order to put further constraints on this spectroscopic binary system, orbital elements or the analysis of light-travel time effects are needed. LAMOST J050146.85+383500.8, therefore, is an interesting target for follow-up studies.
\begin{figure}
\begin{center}
\includegraphics[width=85mm]{SED.eps}
\caption[]{Comparison of the SED of LAMOST J050146.85+383500.8 (red dots) to a Kurucz ODFNEW/NOVER model with $T_\mathrm{eff}$\,=\,12\,500\,K (black squares). The model was forced to either fit the ultraviolet (lower model) or optical flux (upper model). The discrepancies are clearly visible, the star's SED cannot be fitted with a single star flux model.}
\label{fig_SED}
\end{center}
\end{figure}
\section{Conclusions} \label{conclusion}
We carried out a search for mCP stars in the publicly available spectra of LAMOST DR4. Suitable candidates were selected by searching for the presence of the characteristic 5200\,\AA\ flux depression. In consequence, our sample is biased towards mCP stars with conspicuous flux depressions at 5200\,\AA. Spectral classification was carried out with a modified version of the MKCLASS code (MKCLASS\_mCP) and, for a subsample of stars, by manual classification. We evaluated our results by spot-checking with manually derived spectral types and comparison to samples from the literature.
The main findings of the present investigation are summarised in the following:
\begin{itemize}
\item We identified 1002 mCP stars, most of which are new discoveries. There are only 59 common entries with the catalogue of \citealt{RM09}. With our work, we significantly increase the sample size of known Galactic mCP stars, paving the way for future in-depth statistical studies.
\item To suit the special needs of our project, we updated the current version (v1.07) of the MKCLASS code to probe several additional lines, with the advantage that the new version (here termed MKCLASS\_mCP) is now able to more robustly identify traditional mCP star peculiarities, including Cr peculiarities and, to some extent, He peculiarities.
\item mCP star peculiarities (Si, Cr, Sr, Eu, strong blends at 4077\,\AA\ and/or 4130\,\AA) were identified in all but 36 stars of our sample, highlighting the efficiency of the chosen approach and the peculiarity identification routine. The remaining objects (mostly mCP stars with weak or complicated peculiarities and He-peculiar objects) were manually searched to locate the presence of peculiarities.
\item Comparisons between manually derived spectral types and the MKCLASS\_mCP final types indicate a good agreement between the derived temperature and peculiarity types. This is further corroborated by a comparison with spectral types from the \citet{RM09} and \citet{Skiff} catalogues and the good agreement of the peculiarity type versus spectral type distribution between this study and the literature. However, with our approach, we missed the presence of certain peculiarities in several objects. The peculiarity types presented here are therefore not exhaustive. They nevertheless form a sound basis for statistical and further studies.
\item Our sample stars are between 100\,Myr and 1\,Gyr old, with the majority having masses between 2\,$M_\odot$ and 3\,$M_\odot$. We investigated the evolutionary status of 903 mCP stars, deriving a mean fractional age on the main sequence of $\tau$\,=\,63\,\% (standard deviation of 23\,\%). Young mCP stars, while undoubtedly present, are conspicuously underrepresented in our sample. Our results could be considered as strong evidence for an inhomogeneous age distribution among low-mass ($M$\,$<$\,3\,M$_\odot$) mCP stars, as hinted at by previous studies. However, we caution that our sample has not been selected on the basis of an unbiased, direct detection of a magnetic field. Therefore, our results have to be viewed with caution and their general validity needs to be tested by a more extended sample selected via different methodological approaches.
\item The mCP stars LAMOST J122746.05+113635.3 (\#876) and LAMOST J150331.87+093125.4 (\#880) boast distances and kinematical properties in agreement with halo stars. If confirmed, they would be the first CP2 halo objects known and therefore of special interest.
\item We identified LAMOST J034306.74+495240.7 (\#135; TYC\,3321-881-1) as an eclipsing binary system ($P$\textsubscript{orb}\,=\,5.1435$\pm$0.0012\,d) hosting a Si CP2 star component (spectral type B9 III Si). This is of great interest because mCP stars in eclipsing binaries are exceedingly rare.
\item The star LAMOST J050146.85+383500.8 was identified as an SB2 system likely comprising of a Si CP2 star and a supergiant.
\end{itemize}
Future investigations will be concerned with an in-depth study of the new mCP stars identified in this work, particularly with regard to their photometric variability, along with further development and refinement of the approach for identifying and classifying mCP stars in large spectroscopic databases using the MKCLASS code.
\begin{acknowledgements}
We thank the referee for his thoughtful report that helped to significantly improve the paper. This work has been supported by the DAAD (project No. 57442043). The Guo Shou Jing Telescope (the Large Sky Area Multi-Object Fiber Spectroscopic Telescope, LAMOST) is a National Major Scientific Project built by the Chinese Academy of Sciences. Funding for the project has been provided by the National Development and Reform Commission. LAMOST is operated and managed by National Astronomical Observatories, Chinese Academy of Sciences. This work presents results from the European Space Agency (ESA) space mission Gaia. Gaia data are being processed by the Gaia Data Processing and Analysis Consortium (DPAC). Funding for the DPAC is provided by national institutions, in particular the institutions participating in the Gaia MultiLateral Agreement (MLA). The Gaia mission website is https://www.cosmos.esa.int/gaia. The Gaia archive website is https://archives.esac.esa.int/gaia.
\end{acknowledgements}
\clearpage
\bibliographystyle{aa}
| 4812a2c0d9028eee5c57376804367e2adff4023c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Measurements of heavy quark bound states in hadronic collisions provide
unique tools for testing quantum chromodynamics (QCD). Charmonium, the
bound state of a charm and anticharm quark, is produced predominantly
via gluon fusion at Relativistic Heavy Ion Collider (RHIC) kinematics.
The $J/\psi$ is a colorless neutral meson with spin 1 and decays with
considerable branching ratio into lepton pairs ($\approx 6\%$ each for
dielectrons and dimuons). Various theoretical models have been developed
to describe the $J/\psi$ production cross section and polarization as a
function of transverse momentum, but none can describe both
simultaneously. All approaches assume a factorization between the
production of the heavy-quark pair, $Q\bar{Q}$, and its hadronization
into a meson. Different approaches differ in the treatment of the
hadronization. The color-evaporation model that is based on quark-hadron
duality and the color-singlet model (CSM) that only allows hadronization
without gluon emissions are the most popular earlier approaches; see
Refs.~\cite{Brambilla:2010cs}~and~\cite{Lansberg:2019adr} for recent
reviews of theoretical developments and phenomenological work.
One theoretical approach to $J/\psi$ production is within the rigorous
framework of nonrelativistic QCD (NRQCD), which is an effective theory
that describes the dynamics of heavy quark bound states at
nonrelativistic scales
($\nu=v/c\,{\ll}1$)~\cite{Lepage:1992nrqcd,Bodwin:1995hv,Cho:1996nrqcd1,Cho:1996nrqcd2}. The large heavy-quark mass scale relative to the
hadronization scale, $m_{Q} \gg \Lambda_{QCD}$, factorizes the $J/\psi$
production process into the quark-antiquark ($Q\bar{Q}$) pair production
at short-distance and subsequent formation of the heavy quark meson at
long-distance. In the former regime, process-dependent short-distance
coefficients are calculated perturbatively, and in the latter
nonperturbative regime, the behavior is encoded in long-distance matrix
elements (LDMEs). The $Q\bar{Q}$ intermediate states are allowed to have
quantum numbers different from those of the final-state meson. The
leading-order relativistic corrections, for instance, put the $Q\bar{Q}$
either in the color-singlet (CS) state $^{3}S^{[1]}_{1}$ or one of the
color-octet (CO) states, $^{1}S^{[8]}_{0}$, $^{3}S^{[8]}_{1}$ or
$^{3}P^{[8]}_{J}$ (J=0,1,2). The relative importance of states with
different color and angular momentum quantum numbers is estimated by
$\nu$-scaling rules~\cite{Bodwin:1995hv}. Including CO states is found
to be crucial as their leading-power contributions $\approx
\mathcal{O}(\frac{1}{p_T})^4$ show up at leading order (LO), in contrast
to the CS terms that only appear at next-to-next-to-leading order
(NNLO). The LDMEs can only be determined from experimental data by
fitting them via a global analysis. Several groups who performed global
analyses~\cite{Ma:2011glo,Chao_2012,BUTENSCHOEN:2011g,Gong_2013}
successfully described world data for transverse momentum ($p_T$)
spectra, while consistent predictions of the spin alignment measurements
for quarkonia remain a challenge.
The spin alignment of a positively charged lepton from a $J/\psi$ decay,
commonly known as ``polarization," has been measured at the
Tevatron~\cite{Abulencia:2007cdf},
RHIC~\cite{Adare:2010tj,Adare:2016jta}, and the Large Hadron
Collider~\cite{Abelev:2012alice,Chatrchyan:2013cms2,Aaij:2013lhcb,Acharya:2018alice}.
Measuring spin alignment provides additional tests for the theory and
understanding dominant quarkonium production mechanisms in different
kinematic regimes. The $J/\psi$ polarization is measured by fitting the
angular distribution of a positively charged lepton, shown in
Eq.~(\ref{eq:angdist}), to data and extracting decay angular
coefficients.
\begin{equation}\label{eq:angdist}
\frac{dN}{d\Omega} \approx 1+\lambda_{\theta} \cos^{2}{\theta} + \lambda_{\theta\phi}\sin^2{\theta}\cos{2\phi}+\lambda_{\phi}\sin{2\theta}\cos{\phi}, \quad
\end{equation}
\noindent where the coefficients $\lambda_{\theta}$,
$\lambda_{\theta\phi}$, and $\lambda_{\phi}$ are determined most
commonly in the helicity (HX) frame~\cite{JACOB:1959helicity},
Collins-Soper~(C-S) frame~\cite{Collins:1977iv} and
Gottfried-Jackson~(G-J) frame~\cite{Jackson:1964gj} defined in the
$J/\psi$ production plane. Invariant variables are constructed using
SO(2) symmetry in choosing the $z$-axis in the production plane.
Physical interpretation is straightforward only with these invariant
quantities, making direct comparison possible between experimental
results in different kinematic regimes. Equation~(\ref{eq:rotinvvar})
shows the two frame-invariant variables defined in
Refs.~\cite{Faccioli_2010,Faccioli_2011} to characterize the decay
angular distribution.
\begin{equation}\label{eq:rotinvvar}
\tilde{\lambda} = \frac{\lambda_{\theta} + 3 \lambda_{\phi}}{1- \lambda_{\phi}}, \quad
F = \frac{1 + \lambda_{\theta} + 2 \lambda_{\phi}}{3 + \lambda_{\theta}}
\end{equation}
Recently, general methods have been developed for finding all
independent invariants under rotations for particles with various spin
quantum numbers~\cite{Ma:2017rotinv,Martens:2017cvj,Gavrilova:2019rotinv}.
There are also new theoretical developments for less inclusive observables
that include looking at the polarization of quarkonia produced in
jets~\cite{Baumgart:2014upa,Kang:2017yde,Dai:2017cjq,Bain:2017wvk}. \\
Previously a PHENIX $J/\psi$ measurement at $\sqrt{s}$ = 510~GeV in $p$$+$$p$
at forward rapidity showed largely longitudinal net polarization
(negative angular coefficients)~\cite{Adare:2016jta}, while a prior
midrapidity PHENIX measurement at $\sqrt{s}$ = 200~GeV was consistent
with no strong polarization~\cite{Adare:2010tj}. The present analysis
for midrapidity $J/\psi$ production in $\sqrt{s}$ = 510~GeV collisions
is complementary to both previous measurements.
\section{Experimental setup}
In 2013, the PHENIX experiment at RHIC collected data from
longitudinally polarized $p$$+$$p$ collisions at $\sqrt{s}$ = 510~GeV. An
integrated luminosity of 136 pb$^{-1}$ was used for J/$\psi$
polarization measurements at midrapidity.
The PHENIX detector is described in detail in Ref.~\cite{Adcox:2003zm}.
The PHENIX central arms comprise the east and west arms and cover a
pseudorapidity range $|\eta| <0.35$ and an azimuthal coverage of
$\Delta{\phi} = \frac{\pi}{2}$ for each arm. The PHENIX detector
elements used in this analysis include the drift chamber (DC), pad
chambers (PC), ring-imaging \v{C}erenkov (RICH) detector and
electromagnetic calorimeter (EMCal). Charged particle tracks are
reconstructed with the DC and PC tracking system. These detectors also
provide the momentum information of the tracks. The data sample used for
this analysis is the sum of events triggered with three different energy
deposit thresholds applied in the EMCal. Triggers used in this study are
described in more detail in Section~\ref{sec:analysis} and in
Ref.~\cite{phenix_online}. The RICH was used for electron
identification. Two sets of 64 quartz-crystal radiators attached to
photomultipliers at $z$ positions of $\pm$ 144 cm and pseudorapidity
$3.1<|\eta|<3.9$ were used to trigger hard collision events and to
evaluate the collision vertex position. These beam-beam-counters (BBCs)
and zero-degree calorimeters were used together to evaluate and
compare the luminosities seen by PHENIX. In addition, the silicon vertex
detector (VTX) was placed in the west arm around the beam pipe at
nominal radii of 2.6, 5.1, 11.8, 16.7 cm with an acceptance of $|\eta|
<1$ and $\Delta{\phi} = 0.8\pi$. The total material budget expressed as
a percentage of a radiation length is 13.42\%. As the VTX was not in
operation in 2013, this created a large source of electron background
from conversions of direct and decay photons.
\section{Analysis procedure}\label{sec:analysis}
\subsection{Event Selection}
Events were triggered by a minimum energy in any 2$\times$ 2-tower group
of the EMCal and associated hits in the RICH, in coincidence with the
minimum-bias trigger condition. The various EMCal-RICH trigger (ERT)
thresholds for energy deposited in the EMCal were 2.2, 3.7, and 4.7 GeV.
Detailed description of the PHENIX ERT system can be found in
Ref.~\cite{Adare:2010tj}. Different scale-down factors were used for
each threshold, that is, different fractions of triggered events were
written to tape. An OR of all these triggers was used for polarization
measurements and only the lowest threshold trigger was required for
cross section measurements. The $J/\psi$ candidates are triggered by
either or both members of the electron-positron pair. The transverse
momentum of electrons is determined by the bending of the track in the
magnetic field before the DC. For electron identification, information
from both the RICH and the EMCal was used. At least one photomultiplier
tube in the RICH, associated with the track, is required to have fired.
The energy $E$ deposited in the EMCal was used for electron
identification by requiring an expected $(E-p)/p$ ratio within 3 sigma,
where the measured uncertainty is parameterized as a function of $E/p$.
The electron identification efficiency is approximately 90--95\%.
Hadrons misidentified as electrons and conversion electrons are
subtracted statistically as combinatorial background. Due to the
presence of nonoperational silicon vertex detector material located very
close to the beam pipe, conversion electrons mimic prompt electrons and
thus direct tagging of conversion electrons would have a very limited
efficiency; therefore, tagging is not used.
Due to the partial azimuthal coverage on each side of the PHENIX Central
Arm detector, different arm combinations in the measurements of decay
pair products give access to different pair $p_T$ regions. For instance,
high-$p_T$ enhancement is achieved in the sample of $J/\psi$ mesons
whose decay dielectrons are detected in the same arm due to their small
opening angle in the lab frame. The dielectrons detected in the east arm
offer a relatively clean high-$p_T$-enhanced $J/\psi$ sample. The pairs
within the west arm were not used in this analysis due to excessive
background consisting of $\gamma$ to $e^+e^-$ conversion caused by
presence of the VTX that was not operational. The pairs with one lepton
in each arm were used to obtain $J/\psi$ mesons with lower $p_T$. The
number of $J/\psi$ events is determined by fitting the invariant mass of
lepton pairs in the data and counting within the mass range between 2.8
and 3.3~GeV/$c^2$. Combinatorial background attributed to decay
electrons from hadrons is eliminated statistically before performing the
fit using the like-sign method~\cite{Adler:2003qs} established in
PHENIX. To describe the combinatorial background in the data, the ratio
of like-sign mass distributions between the data in the same event and
the mixed-event data is used in normalizing the like-sign mass
distribution in the mixed-event data. Unlike-sign pairs cannot be used
for normalization because their mass distributions contain correlated
background, as described below. For this reason, the like-sign method
is also referred to as the mixed event method.
The sum of the signal and residual background components, described by a
crystal ball function~\cite{CBF} and an exponential function,
respectively, is used to further subtract correlated electron background
coming from open-heavy-flavor and Drell-Yan decay processes. Invariant
mass distributions for dielectrons measured in the same arm and opposite
arms are shown in Fig.~\ref{fig:invmass} along with fit results.
\begin{figure}[!htb]
\includegraphics[width=1.0\linewidth]{fig1a.png}
\includegraphics[width=1.0\linewidth]{fig1c.png}
\caption{\label{fig:invmass}
(a) Example of invariant mass distribution for the high $p_T$ (3.0 $\leq
p_T <$ 10.0~GeV/$c$) $J/\psi$ sample. Solid red line indicates a total
fit and dashed line shows correlated background from the fit. (b)
Example of invariant mass distribution for the low-$p_T$ (0.0 $< p_T <$
3.0~GeV/$c$) sample. Green histogram shows combinatorial background
calculated using the mixed event method and subtracted before fitting.
}
\end{figure}
\subsection{Efficiency and acceptance corrections}\label{effacccorr}
The MC pseudodata were generated to estimate the effects of detector
acceptance and efficiency. Events containing a single $J/\psi$ per event
were generated with zero polarization, a flat $p_{T}$ distribution, and
a flat $z$-vertex distribution along the beam direction, using Monte
Carlo techniques. The $J/\psi$'s subsequently were forced to decay into
dielectrons within the detector acceptance. The yields were then
reweighted, as described in more detail below, to emulate effects that
are present in the data such as smeared vertex distribution along the
beam axis, the shape of the $J/\psi$ $p_T$ distribution and the ERT
trigger efficiencies.
Polarization measurements are particularly susceptible to inconsistency
between the simulation and data as the acceptance corrections accounting
for the zero polarization baseline are solely dependent on the
simulation. A comparison of the polar and azimuthal angular
distributions between the data and MC simulation, after the effects of
dead areas, shows good agreement, as seen in Fig.~\ref{fig:simdatacomp}.
\begin{figure*}[!thb]
\includegraphics[width=0.49\linewidth]{fig2a.png}
\includegraphics[width=0.49\linewidth]{fig2b.png}
\caption{\label{fig:simdatacomp}
Comparison of azimuthal angle $\phi$ (a) and polar angle $\theta$ (b)
distributions in data (red points) and MC (blue histogram) for
electrons.
}
\end{figure*}
\begin{figure}[!th]
\includegraphics[width=1.0\linewidth]{pttrigs.pdf}
\caption{\label{fig:pttrigs}
Comparison of $J/\psi$ $p_T$ shape in the data with MC for events
triggered with three different ERT settings [(a) Type A, (b) Type B, (c)
Type C] and (d) for all settings inclusive Type A+B+C. For direct
comparisons, trigger efficiency corrections are applied to MC to
describe the data.
}
\end{figure}
The shape of $\frac{dN (J/\psi)}{dp_T} $ can affect the polarization
extraction in a nontrivial manner due to the limited azimuthal
acceptance of the PHENIX Central Arms. Opposite-sign pairs bending
toward each other into the detector populate different phase space than
the ones bending away from each other. For this reason, the $p_T$
differential $J/\psi$ cross section has been measured in data. The
measured cross section was fit with a Kaplan function, defined in
Eq.~\ref{eq:kaplan}, and the fit results were then used to reweight the
MC pseudodata for the acceptance correction.
\begin{equation}\label{eq:kaplan}
\frac{d\sigma}{dp_T} = \frac{A \cdot p_T}{\left[1+\left(p_T/b\right)^2\right]^{n}}
\end{equation}
As the data were sampled from an OR of three triggers with different
energy thresholds, a method was developed to emulate the trigger
efficiency effect on the $p_T$ shape of the data. The efficiency of
triggering on $J/\psi$ events for each type of ERT was determined from
the minimum-bias dataset by checking the trigger condition. The
acceptance-corrected pseudodata were then processed by randomly sampling
the distributions of ERT-measured efficiencies and prescale factors. The
detector area masked due to being ineffective or suffering from
excessive electron background was taken into account in this method.
The shape of the raw $J/\psi$ yield as a function of $p_T$ in data
before corrections is in excellent agreement with the efficiency
corrected normalized yield in simulation for each trigger type as well
as for all possible trigger combinations (see Fig.~\ref{fig:pttrigs}).
\subsection{Cross Sections}
For cross section measurements, the raw $J/\psi$ yield is corrected for
trigger efficiencies as well as track reconstruction efficiencies and
acceptance. The EMCal trigger efficiencies were measured using data and
the electron reconstruction efficiencies and acceptance corrections were
obtained from MC simulation as described in the previous section. An
integrated luminosity of 136 pb$^{-1}$ was sampled in the analysis and
the global normalization of the cross section was determined by the
$p$+$p$ cross section sampled by the BBC trigger, which is found to be
32.5 $\pm$ 3.0 (stat) $\pm$ 1.2 (syst) mb, based on van-der-Meer scan
results. The effects of multiple collisions per beam crossing due to
high luminosity at $\sqrt{s}$ = 510~GeV, which were estimated to be of
order of 20\%, were also taken into account in counting $J/\psi$ yields.
\subsection{Angular coefficients}
To account for acceptance and efficiency effects, the raw
angular distribution in data is divided by the reweighted single
$J/\psi$ MC pseudodata generated under an assumption of no polarization
as described in Section~\ref{effacccorr}. This procedure closely follows
methods described in Ref.~\cite{Adare:2016jta} and additionally adopts
the trigger emulator described in Section~\ref{effacccorr}. The
$\cos{\theta}-\phi$ distributions of the decay positron in three
different polarization frames in MC pseudodata are shown in
Fig.~\ref{fig:angdist}.
\begin{figure}[thb]
\includegraphics[width=1.0\linewidth]{acceff.pdf}
\caption{\label{fig:angdist}
Polar-azimuthal angular distributions in three different polarization
frames [(a)(b) Helicity, (c)(d) Collins-Soper, (e)(f) Gottfried-Jackson]
for decay positrons in MC pseudodata generated with a flat $J/\psi$
$p_T$ and flat z-vertex distribution. The distributions are reweighted
for $p_T$ and z-vertex distribution in the data and corrected for
trigger efficiencies. The $p_T$ ranges are (a)(c)(e)
$0.0<p_T<3.0$~GeV/$c$ and (b)(d)(f) $3.0\leq p_T<10.0$~GeV/$c$.
}
\end{figure}
Angular coefficients were extracted with three different methods; the
$\chi^2$ method, the maximum log-likelihood (MLL) method and the
parametric bootstrap method~\cite{Hastie_01}. The $\chi^2$ fitting
results are displayed in contours at 68.3\%, 95.4\% and 99.7\%
confidence level (CL) in Fig.~\ref{fig:fitcontours}. One can see that
some of the coefficients are correlated in one frame and uncorrelated in
other frames. The helicity frame is known to be orthogonal to the C-S
frame and similar to the G-J frame at midrapidity. This is also seen in
our results, i.e.~the correlation pattern between $\lambda_{\theta}$ and
$\lambda_{\phi}$, represented by the orientation of elliptic fit
contours, in the helicity frame appears to be better aligned with the
ones in the G-J frame. The MLL method treats low statistics bins more
properly than the $\chi^2$ method, while it is prone to underestimating
uncertainties as the signal-to-background ratio becomes smaller. In
fitting methods, the uncertainties on the invariants that are a function
of primary angular coefficients are propagated from the ones on
coefficients, taking into account the correlation matrix. The parametric
bootstrap method properly estimates uncertainties on the invariant
variables. In this method, a Gaussian noise term is added to each
measurement point at each sampling trial with a standard deviation of
the corresponding measured uncertainty. After a large number of sampling
procedures, the uncertainties are defined at a 68\% CL.
\begin{figure}[!thb]
\includegraphics[width=1.0\linewidth]{fitcontours.pdf}
\caption{\label{fig:fitcontours}
The best-fit results for the three angular coefficients are shown with a
central value in a dot along with contours at 68.3\%, 95.4\% and 99.7\%
confidence levels. The polarization frames are (a)(b)(c) HX, (d)(e)(f)
C-S, and (g)(h)(i) G-J. The plot is for the high $p_T$ bin (3.0$\leq
p_T<$ 10.0~GcV/$c$).
}
\end{figure}
The results are consistent among different methods within 0.5$\sigma$ in
all frames and the central measured values and statistical uncertainties
from the best-fit results were taken as the final results.
At lower $p_T$ ($<$3.0~GeV/$c$), the decay electron and positron are
detected in opposite PHENIX central arms. A silicon vertex detector
(VTX) was installed in the PHENIX west arm, but was not operational
during the data taking. This resulted in a high level of irreducible
electron background originating from photon conversion in the VTX
material. These random combinations of background electrons were
subtracted using an event mixing technique, but resulted in higher
statistical and additional systematic uncertainty for the low $p_T$
measurement. The PHENIX acceptance in $\cos{\theta}-\phi$ space is very
different for decay electrons and positrons detected in opposite PHENIX
Central Arms compared to the case when both the electron and positron
are detected in the same arm. As a result, the acceptance for the C-S
frame turned out to be very limited, and a polarization measurement in
this frame was not done at low $p_T$. The polarization coefficients at
low $p_T$ were determined using only the $\chi^2$ fitting method.
\subsection{Systematic uncertainties}
Sources of systematic uncertainties on the polarization and cross
section measurement include uncertainties on acceptance, tracking and
electron identification efficiency, trigger efficiencies, uncertainties
on the global normalization derived from the cross section seen by the
BBC, and the trigger rate dependence of $J/\psi$ yield counting. The two
latter sources do not affect the polarization measurements at all, but
they are the main sources of systematic uncertainty for $J/\psi$ cross
section determination.
During RHIC run 2013, the proton beam luminosity was very high, which
resulted in a nonnegligible probability of multiple collisions per beam
crossing at high trigger rates. The correction to the $J/\psi$ cross
section due to multiple collisions per crossing was calculated by
dividing the whole data set into five run groups with different trigger
rates, calculating the cross section for each group, and extrapolating
to zero trigger rate. The correction was estimated to be $0.80 \pm
0.20$. The large uncertainty of this estimate is due to extrapolating
from very high trigger rates to zero and lack of statistics. This is the
largest systematic uncertainty of the $J/\psi$ cross section.
The primary source of systematic uncertainties on the polarization
measurements in the central-arm detectors is the shape of the $J/\psi$
$p_T$ spectra. The $p_T$ shape is affected by statistical uncertainties
on the raw measured yields as well as uncertainties on the ERT
efficiencies. The uncertainties on the ERT efficiencies were estimated
using a sampling method. In each sampling trial, parameters were
extracted from fitting the efficiency curve for each trigger type, and
the analysis was repeated with those parameters to estimate the combined
ERT efficiencies. The quadratic sum of the statistical and systematic
uncertainties was assigned as a total uncertainty at each measurement
point. The systematic uncertainties on angular coefficients that are
attributed to $p_T$ shape were estimated using the parametric bootstrap
method with a Gaussian noise term corresponding to the aforementioned
uncertainties. Uncertainties estimated in this method depend on the
$p_T$ of $J/\psi$, the type of $\lambda$ coefficient and the
polarization frame. Resulting uncertainties from this source are larger
(as high as $\approx$0.5 on $\lambda_{\theta}$) at $p_T <$ 3.0~GeV/$c$ and
$<$0.1 in most cases at $p_T \geq$ 3.0~GeV/$c$. In addition, the
differences between fitting methods, estimated to be less than 0.5 times
the statistical uncertainties in all angular coefficients, were added as
systematic uncertainties at high $p_T$.
At low $p_T$ (below $\approx$3.0 GeV/$c$), the PHENIX acceptance in
$\cos{\theta} - \phi$ space is very different from that at higher $p_T$.
Relative contributions from different sources of systematic
uncertainties also change, while the dependence on exact $p_T$ shape
remains a dominant factor. An additional systematic uncertainty at low
$p_T$ is caused by the much larger combinatorial background, which had
to be subtracted using a mixed-event method. This uncertainty was
estimated by varying the mixed-event background normalization by
$\pm$1$\sigma$ and was found smaller than the two uncertainties
mentioned above. Systematic uncertainties of this measurement are shown
in Table~\ref{table:polresults}.
\section{Results}\label{result}
The measured $p_T$-differential $J/\psi$ cross section is shown in
Fig.~\ref{fig:jpsixsection} with a fit to a Kaplan function. Due to the
higher collision energy the $p_T$ spectrum is harder than previously
published PHENIX results for $\sqrt{s}$ = 200~GeV \cite{Adare:2010tj},
where the fit parameter $b$ that determines the hardness of the spectrum
for a given $n$ was estimated to be smaller at 3.41 $\pm$ 0.21 and the
parameter $n$ was comparable at 4.6 $\pm$ 0.4.
Figure~\ref{fig:jpsixsection} also shows a comparison of the measured
$J/\psi$ differential cross section with a theory prediction based on
full NRQCD at NLO with leading relativistic corrections that includes CS
and CO states, provided by Butensch\"on et al.~\cite{Butenschoen:2011c}.
The sources of theory uncertainties include variations of theory scale
and LDMEs. Within its uncertainties, the theory calculation is in
agreement with the experimental results within its valid range of $p_T
\gtrsim 2$~GeV/$c$, justifying use of the theory model for predictions
of polarization measurements in this kinematic range.
The $p_T$-integrated cross section times branching ratio is shown in
Fig.~\ref{fig:xsecworld} along with the previous PHENIX results at
$\sqrt{s}$ = 200~GeV~\cite{PHENIXcs} and the world results from the
Large Hadron Collider~\cite{ALICEcs1,ALICEcs2} and
Tevatron~\cite{CDFcs}. A simple logarithmic dependence on the collision
energy is seen for $J/\psi$ production at midrapidity, making estimates
of $J/\psi$ yield at any $\sqrt{s}$ easy, and inviting the theory
community to model the trend. The $p_T$-integrated cross section times
branching ratio was found to be 97.6 $\pm$ 3.6 (stat) $\pm$ 5.1 (syst)
$\pm$ 9.8 (global) $\pm$ 19.5 (multiple collision) nb.
\begin{figure}[thb]
\includegraphics[width=1.0\linewidth]{new_jpsi_crossection_withtheory.png}
\caption{\label{fig:jpsixsection}
The measured differential $J/\psi$ cross section times branching ratio
as a function of transverse momentum. Fit parameters are $A = 37.6 \pm
2.2$~nb/(GeV/$c$), $b = 4.33 \pm 0.28$~GeV/$c$, and $n = 4.61 \pm 0.32$.
Open rectangles show systematic point-by-point uncertainties. Global
normalization uncertainty of 10.1\% is not shown. Green curve shows
theory prediction based on full NRQCD at NLO with leading relativistic
corrections that include CS and CO states, provided by Butensch\"on et
al.~\cite{Butenschoen:2011c}. Light green band indicates theory
uncertainty.
}
\end{figure}
\begin{figure}[!tbh]
\includegraphics[width=1.0\linewidth]{world.png}
\caption{\label{fig:xsecworld}
The PHENIX results of $p_T$-integrated cross section times branching
ratio for $J/\psi$ production at midrapidity, shown with world data. All
systematic errors for CDF and ALICE experiments were added in
quadrature. The fit parameters are C = 70.4 nb, and a = 9.27~TeV$^{-1}$.
The pink band shows the one-sigma fit uncertainty.
}
\end{figure}
The final results of the three primary angular coefficients are shown in
Fig.~\ref{fig:lamthe}, Fig.~\ref{fig:lamphi}, and
Fig.~\ref{fig:lamthephi}. Uncertainties on measurements in different
frames are correlated. Due to limited detector acceptance in $\eta$ at
midrapidity, $\lambda_{\theta}$ (Fig.~\ref{fig:lamthe}) is poorly
constrained compared to $\lambda_{\phi}$ (Fig.~\ref{fig:lamphi}).
Theory predictions based on full NRQCD at NLO with leading relativistic
corrections that includes CS and CO states, provided by Butensch\"on et
al.~\cite{Butenschoen:2012}, are overlaid with the measurements. The
uncertainty bands on the NRQCD predictions account for the scale
uncertainties and uncertainties on the LDMEs. The LDMEs were obtained in
a global analysis of unpolarized data that excludes measurements from
hadroproduction with $p_T<$3.0~GeV/$c$. To improve consistency with the
data, the feed-down from heavier charmonium states was subtracted in the
theory prediction. The fraction of $J/\psi$ events from $b$-flavored
hadron decays is negligible at RHIC. In PHENIX, the unpolarized yield
measurements~\cite{Adare:2010tj} are well described down to 1~GeV/$c$,
justifying the comparison to the polarized measurements.
\begin{figure}[!ht]
\includegraphics[width=1.0\linewidth]{lamthe_th_nrqcd.pdf}
\caption{\label{fig:lamthe}
$\lambda_{\theta}$ measured in $J/\psi$ transverse momentum bins of 0.0
$< p_T <$ 3.0~GeV/$c$ and 3.0 $\leq p_T <$ 10.0~GeV/$c$ overlaid with
NRQCD predictions in the Helicity and Collins-Soper frames. The points
for different frames are shifted for visual clarity.
}
\end{figure}
\begin{figure}[!ht]
\includegraphics[width=1.0\linewidth]{lamphi_th_nrqcd.pdf}
\caption{\label{fig:lamphi}
Angular coefficient $\lambda_{\phi}$ measured in $J/\psi$ transverse
momentum bins of 0.0 $< p_T <$ 3.0~GeV/$c$ and 3.0 $\leq p_T <$
10.0~GeV/$c$ overlaid with NRQCD predictions in the Helicity and
Collins-Soper frames. The points for different frames are shifted for
visual clarity.
}
\end{figure}
\begin{figure}[!th]
\includegraphics[width=1.0\linewidth]{lamthephi_th_nrqcd.pdf}
\caption{\label{fig:lamthephi}
Angular coefficient $\lambda_{\theta\phi}$ measured in $J/\psi$
transverse momentum bins of 0.0 $< p_T <$ 3.0~GeV/$c$ and 3.0 $\leq p_T
<$ 10.0~GeV/$c$ overlaid with NRQCD predictions in the Helicity and
Collins-Soper frames. The points for different frames are shifted for
visual clarity.
}
\end{figure}
\begin{figure}[!thb]
\includegraphics[width=1.0\linewidth]{lamtil_th_hx.pdf}
\caption{\label{fig:lamtil}
$\tilde{\lambda}$ measured in $J/\psi$ transverse momentum bins of 0.0
$< p_T <$ 3.0~GeV/$c$ and 3.0 $\leq p_T <$ 10.0~GeV/$c$ overlaid with
CSM and NRQCD predictions in the Helicity frame. Predictions for this
frame-invariant variable in the other two frames are consistent with the
one in the Helicity frame. The points for different frames are shifted
for visual clarity.
}
\end{figure}
\begin{figure}[!thb]
\includegraphics[width=1.0\linewidth]{invf_th_hx.pdf}
\caption{\label{fig:invf}
Invariant F measured in $J/\psi$ transverse momentum bins of 0.0 $< p_T
<$ 3.0~GeV/$c$ and 3.0 $\leq p_T <$ 10.0~GeV/$c$ overlaid with CSM and
NRQCD predictions in the Helicity frame. Predictions for this
frame-invariant variable in the other two frames are consistent with the
one in the Helicity frame. The points for different frames are shifted
for visual clarity.
}
\end{figure}
In Figs.~\ref{fig:lamtil} and \ref{fig:invf}, the results are also shown
in terms of frame-invariant observables $\tilde{\lambda}$ and F, defined
in Eq.~\ref{eq:rotinvvar}. Measuring these invariant variables provides
a direct test for the underlying production mechanisms. Comparing
results in different frames can additionally serve as robust tools to
address systematic uncertainties. Consistency between different frames
shown for $\tilde{\lambda}$ in Fig.~\ref{fig:lamtil} and for F in
Fig.~\ref{fig:invf} indicates that systematic effects are under good
control. Frame-invariant variables are interpreted as the total net
polarization summed over all production mechanisms. Different production
mechanisms can lead to natural polarization in different frames. In the
case of $\tilde{\lambda}$, transverse polarization corresponds to the
value of 1 and longitudinal polarization to the value of -1. The zero
value seen in $\tilde{\lambda}$ is interpreted as no net polarization.
The other variable F also carries similar meaning. Unlike
$\tilde{\lambda}$, F is mathematically bounded between 0 and 1. The zero
polarization corresponds to $\frac{1}{3}$ and transverse and
longitudinal polarization to $\frac{2}{3}$ and 0, respectively.
Two scenarios were considered for these frame-invariant results, NRQCD
and the Color Singlet Model, which is equivalent to the $v\rightarrow 0$
limit of NRQCD. The uncertainties on CSM predictions include the scale
uncertainty. The CSM prediction is qualitatively opposite to the NRQCD
predictions~\cite{Butenschoen:2012,Butenschoen:2013}. NRQCD predicts
transverse polarization while the CSM predicts longitudinal
polarization, as is shown in Fig.~\ref{fig:lamtil} and
Fig.~\ref{fig:invf}. The data does not conclusively exclude either
strong transverse or longitudinal polarization.
\begin{figure}[!thb]
\includegraphics[width=1.0\linewidth]{lamthe_midrap_comp.pdf}
\caption{\label{fig:lamthemidrapcomp}
Angular coefficient $\lambda_{\theta}$ of midrapidity $J/\psi$
production at $\sqrt{s}$= 200~GeV and 510~GeV, shown as a function of
$x_T = \frac{2 p_{T}}{\sqrt{s}}$. Data points at the lowest $p_T$ have
been shifted against each other for visual clarity.
}
\end{figure}
\begin{figure*}[!thb]
\begin{minipage}{0.48\linewidth}
\includegraphics[width=0.99\linewidth]{lamtherap_comp.png}
\caption{\label{fig:lamtherapcomp}
Angular coefficient $\lambda_{\theta}$ of $J/\psi$ production at
$\sqrt{s}$ = 510~GeV shown as a function of rapidity in three
polarization frames. The points for different frames are shifted for
clarity. Forward rapidity points are from Ref.~\protect\cite{Adare:2016jta}.
}
\end{minipage}
\hspace{0.4cm}
\begin{minipage}{0.48\linewidth}
\includegraphics[width=0.99\linewidth]{lamtilrap_comp.png}
\caption{\label{fig:lamtilrapcomp}
Rapidity dependence of $\tilde{\lambda}$ of $J/\psi$ production at
$\sqrt{s}$ = 510~GeV in three polarization frames. The points for
different frames are shifted for clarity. Forward rapidity points are
from Ref.~\protect\cite{Adare:2016jta}.
}
\end{minipage}
\end{figure*}
\begingroup
\begin{table*}[tbh]
\caption{\label{table:polresults}
Results with statistical $\oplus$ total systematic uncertainties.
}
\begin{ruledtabular}
\renewcommand{\arraystretch}{2}
\begin{tabular}{ccccccc}
Frame & $p_T$ & $\lambda_{\theta}$ & $\lambda_{\phi}$ &
$\lambda_{\theta\phi}$ & $\tilde{\lambda}$ & F \\
\hline
Helicity
& 0.0$<p_T<$3.0
& 0.01$^{+(0.44 \oplus 0.24) }_{-(0.24 \oplus 0.31)}$
& 0.06$^{+(0.20 \oplus 0.23)}_{ -(0.35\oplus 0.27)}$
& 0.23 $^{+(0.12 \oplus 0.09)}_{-(0.08\oplus 0.11)}$
& 0.20 $^{+(0.32 \oplus 0.48)}_{-(0.35 \oplus 0.52)}$
& 0.38 $^{+(0.06\oplus 0.06)}_ {-(0.09 \oplus 0.10)}$
\\
& 3.0$\leq p_T<$10.0
& -0.12$^{+(0.18\oplus 0.04)}_{-(0.16\oplus 0.06)}$
& 0.06$^{+(0.11\oplus 0.07)}_{-(0.13\oplus 0.07)}$
& 0.12$^{+(0.19\oplus 0.04)}_{-(0.19 \oplus 0.04)}$
&0.01$^{+(0.30\oplus 0.20 )}_{-(0.37 \oplus 0.19)}$
&0.34 $^{+(0.07\oplus 0.04)}_{-(0.08 \oplus 0.05)}$
\\
Collins-Soper
& 3.0$\leq p_T<$10.0
& -0.08$^{+(0.40\oplus 0.10)}_{-(0.38 \oplus 0.16)}$
& -0.01$^{+(0.09\oplus 0.03)}_{-(0.10 \oplus 0.12)}$
& 0.02$^{+(0.18 \oplus 0.01 )}_{-(0.18 \oplus 0.07)}$
& -0.11$^{+(0.47 \oplus 0.22)}_{-(0.47 \oplus 0.17)}$
& 0.31$^{+(0.11 \oplus 0.05 )}_{-(0.11 \oplus 0.05)}$
\\
Gottfried-Jackson
& 0.0$<p_T<$3.0
& 0.13 $^{+(0.52 \oplus 0.35)}_{-(0.59\oplus 0.52)}$
& 0.01 $^{+(0.06 \oplus 0.07)}_{-(0.05 \oplus 0.10)}$
& -0.01 $^{+(0.25\oplus 0.06)}_{-(0.24\oplus 0.07)}$
& 0.17$^{+(0.42\oplus 0.29)}_{-(0.42\oplus 0.42)}$
&0.37$^{+(0.07\oplus 0.08) }_{-(0.12 \oplus 0.11)}$
\\
& 3.0$\leq p_T<$10.0
& 0.03$^{+(0.24 \oplus 0.13)}_{-(0.21\oplus 0.05) }$
& -0.02$^{+(0.12\oplus 0.05)}_{-(0.15 \oplus 0.06)}$
& 0.13$^{+(0.20\oplus 0.08)}_{-(0.18 \oplus 0.09)}$
& -0.04$^{+(0.34\oplus 0.23)}_{-(0.40\oplus 0.17)}$
& 0.33$^{+(0.08 \oplus 0.05)}_{-(0.09 \oplus 0.04)}$
\\
\end{tabular} \end{ruledtabular}
\end{table*}
\endgroup
Results are also compared to previous PHENIX measurements at a different
beam energy~\cite{Adare:2010tj} and rapidity~\cite{Adare:2016jta}.
Figure~\ref{fig:lamthemidrapcomp} compares $\lambda_{\theta}$ in the
helicity frame for different collision energies. At midrapidity, the
decay lepton spin alignment is consistent with no polarization both at
$\sqrt{s}$ = 200~GeV (1-dimensional analysis) and 510~GeV (2-dimensional
analysis). The rapidity dependence of $\lambda_{\theta}$ and
$\tilde{\lambda}$ is shown in Fig.~\ref{fig:lamtherapcomp} and
Fig.~\ref{fig:lamtilrapcomp}, respectively. While midrapidity data
indicate no polarization, moderate polarization is seen at forward
rapidity. At forward rapidity $\tilde{\lambda}$ was measured to be
largely negative, indicating longitudinal polarization. This is in stark
contrast to this result that sees no preferred polarization direction,
shown in Fig.~\ref{fig:lamtherapcomp} and Fig.~\ref{fig:lamtilrapcomp}.
No strong polarization was seen in other experiments at higher $p_T$ and
higher beam energies in general, and the discrepancy between
measurements and theory predictions is still being studied. Results of
polarization measurements are summarized in
Table~\ref{table:polresults}.
\section{Summary}
The PHENIX experiment measured the $J/\psi$ polarization at midrapidity
in $\sqrt{s}=$ 510 GeV $p$$+$$p$ collisions by reconstructing the
hadronized charmonium state in the dielectron decay channel. The
midrapidity cross section at $\sqrt{s}=$ 510 GeV in the same channel has
been newly measured and is consistent with NRQCD calculations above $p_T
\gtrsim 2$~GeV/$c$. The results show the expected hardening of the
$J/\psi$ $p_T$ spectrum as compared to the measurement at $\sqrt{s}=$
200 GeV. At both low $p_T$ and high $p_T$, the net polarization has
been observed to be consistent with zero within uncertainties. This is
in contrast to the measurements made at forward rapidity. The new
results do not rule out either the CSM or the NRQCD $J/\psi$ production
models. The new measurements from the 2-dimensional analysis show
consistency in $\lambda_\theta$ with the results from a previous
1-dimensional midrapidity analysis at $\sqrt{s}$ = 200~GeV.
\begin{acknowledgments}
We thank the staff of the Collider-Accelerator and Physics
Departments at Brookhaven National Laboratory and the staff of
the other PHENIX participating institutions for their vital
contributions. We acknowledge support from the
Office of Nuclear Physics in the
Office of Science of the Department of Energy,
the National Science Foundation,
Abilene Christian University Research Council,
Research Foundation of SUNY, and
Dean of the College of Arts and Sciences, Vanderbilt University
(U.S.A),
Ministry of Education, Culture, Sports, Science, and Technology
and the Japan Society for the Promotion of Science (Japan),
Conselho Nacional de Desenvolvimento Cient\'{\i}fico e
Tecnol{\'o}gico and Funda\c c{\~a}o de Amparo {\`a} Pesquisa do
Estado de S{\~a}o Paulo (Brazil),
Natural Science Foundation of China (People's Republic of China),
Croatian Science Foundation and
Ministry of Science and Education (Croatia),
Ministry of Education, Youth and Sports (Czech Republic),
Centre National de la Recherche Scientifique, Commissariat
{\`a} l'{\'E}nergie Atomique, and Institut National de Physique
Nucl{\'e}aire et de Physique des Particules (France),
Bundesministerium f\"ur Bildung und Forschung, Deutscher Akademischer
Austausch Dienst, and Alexander von Humboldt Stiftung (Germany),
J. Bolyai Research Scholarship, EFOP, the New National Excellence
Program ({\'U}NKP), NKFIH, and OTKA (Hungary),
Department of Atomic Energy and Department of Science and Technology
(India),
Israel Science Foundation (Israel),
Basic Science Research and SRC(CENuM) Programs through NRF
funded by the Ministry of Education and the Ministry of
Science and ICT (Korea).
Physics Department, Lahore University of Management Sciences (Pakistan),
Ministry of Education and Science, Russian Academy of Sciences,
Federal Agency of Atomic Energy (Russia),
VR and Wallenberg Foundation (Sweden),
the U.S. Civilian Research and Development Foundation for the
Independent States of the Former Soviet Union,
the Hungarian American Enterprise Scholarship Fund,
the US-Hungarian Fulbright Foundation,
and the US-Israel Binational Science Foundation.
\end{acknowledgments}
| 4416a65eb67dbcb0fe798fc9a80465b192c667b8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Human visual attention comes in two forms. One is bottom-up, where prioritization is based solely on processing of the visual input. The other is top-down, where prioritization is based on a myriad of top-down information sources (the object context of a scene, the semantic relationships between objects, etc.~\cite{wolfe2017five,eckstein2011visual,nakayama2011situating}). When your food arrives at a restaurant, among your very first movements of attention will likely be to the fork and knife (\Fref{fig:teaser}), because these objects are important to your goal of having dinner. Goal-directed attention control underlies all the tasks that we {\bf try} to do, thereby making its prediction a vastly more challenging and important problem than predicting the bottom-up control of attention by a visual input. Perhaps the strongest form of top-down attention control is in the definition of a target goal, with visual search being arguably the simplest goal-directed tasks---there is a target goal and the task is to find it. Humans are highly efficient and flexible in the image locations that they choose to fixate while searching for a target-object goal, making the prediction of human search behavior important for both behavioral and computer vision, and in particular for robotic visual systems such as robotic search~\cite{elder2006attentive, narayanan2016perch}.
In this paper, we introduce Inverse Reinforcement Learning as a computational model of human attention in visual search.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{images/teaser2.pdf}
\vskip -0.1in
\caption{{\bf Predicting fixations in a visual search task}. Left: behavioral scanpath shown in yellow. Right: predicted scanpath in green. The search target is the fork, shown in the yellow bounding box. }
\label{fig:teaser}
\vskip -0.2in
\end{figure}
\myheading{Gaze prediction in visual search}.
The aim of gaze prediction is to predict the patterns of fixations made by people during image viewing. These patterns can be either spatial (fixation density maps) or spatial+temporal (scanpaths). Most fixation-prediction models are in the context of a free-viewing task, where the prioritization of locations for fixation is believed to be controlled by bottom-up saliency. Since the seminal work by \citet{IttiPAMI98}, the use of saliency maps to predict free-viewing fixations has grown into a sizeable literature \cite{borji2013state, kummerer2017understanding, kruthiventi2017deepfix, huang2015salicon, kummerer2014deep, cornia2018predicting,jiang2015salicon,jetley2016end, borji2015salient, masciocchi2009everyone, berg2009free}. However, the predictions of saliency models do not generalize to fixations made in goal-directed attention tasks, such as the search for a target object~\cite{henderson2007visual, koehler2014saliency}.
A critical difference between a search task and a free-viewing task is that search fixations are guided towards a target-object goal, whereas in free viewing there are no explicit goals. Target guidance during search was first quantified using simple targets having simple features that were known to the searcher~\cite{wolfe1994guided}. This work was followed by computational models that used images of objects and scenes as inputs~\cite{ehinger2009modelling, zelinsky2008theory, zelinsky2013modeling}, which further expanded to include target spatial relationships~\cite{aydemir2011search} and global scene context~\cite{torralba2006contextual}. Target spatial relationships and scene context are important factors that substantially affect human search efficiency. When searching for objects in scenes, scene context can guide attention to the target. To date, there have been very few deep network models attempting to predict human search fixations \cite{zhang2018finding, adeli2018deep,Wei-etal-NIPS16}. What all of these models have in common, however, is that they use some algorithm, and knowledge about a particular source of information (target features, meaning, context, etc), to prioritize image locations for fixation selection.
\myheading{Inverse Reinforcement Learning}. Our approach to search-fixation prediction is exactly the opposite. Rather than using an algorithm to prioritize locations in an image, here we use Inverse Reinforcement Learning (IRL)~\cite{Ng-Russell-ICML00, Abbeel-etal-ICML04,Ziebart-etal-AAAI08,ho2016generative,fu2017learning} to learn sequences of search fixations by treating each as a potential source of reward. IRL, a form of imitation learning, refers to the recovery of an expert's underlying reward function through repeated observation. Most IRL algorithms \cite{Ziebart-etal-AAAI08,fu2017learning,wulfmeier2015maximum} simultaneously learn an optimal policy and the reward function on which the policy is optimized. Although early IRL algorithms \cite{Ng-Russell-ICML00,Ziebart-etal-AAAI08} were often restricted to problems having low-dimensional state spaces, \citet{wulfmeier2015maximum} proposed deep maximum entropy IRL to handle raw image inputs. Recent work~\cite{ho2016generative,fu2017learning} applies adversarial training \cite{Goodfellow-etal-NIPS14} to learn the underlying reward function and the policy, treating each as (part of) the discriminator and the generator in adversarial training, respectively. The discriminator assigns high reward to an expert's behavior and low reward to a non-expert's behavior, where behavior is represented as state-action pairs. The generator/policy is optimized using a reinforcement learning algorithm to get higher reward by behaving more like the expert. In this work, we use the GAIL (generative adversarial imitation learning) algorithm~\cite{ho2016generative}, given its ability to imitate behaviors in complex and high-dimensional environments \cite{ho2016generative}. Using this approach, a unified information-maximization framework is defined for combining diverse sources of information for the purpose of selecting maximally-rewarding locations to fixate, thus increasing both the accuracy and applicability of predicting human search fixations.
\begin{table}[]
\begin{center}
\setlength{\tabcolsep}{2pt}
\begin{tabular}{lcrrrr}
\toprule
Dataset & Search & Image & Class & Subj/img & Fixation\\
\midrule
SALICON \cite{jiang2015salicon} & \ding{55} & 10000 & - & 60 & 4600K* \\
POET \cite{papadopoulos2014training} & \ding{55} & 6270 & 10 & 5 & 178K \\
People900 \cite{ehinger2009modelling} & \ding{51} & 912 & 1 & 14 & 55K \\
MCS \cite{zelinsky2019benchmarking} & \ding{51} & 2183 & 2 & 1-4 & 16K \\
PET \cite{gilani2015pet} & \ding{51} & 4135 & 6 & 4 & 30K \\
COCO-Search18 & \ding{51} & 6202 & 18 & 10 & 300K \\
\bottomrule
\end{tabular}
\vskip -0.1in
\caption{{\bf Comparison of fixation datasets}. Previous datasets either did not use a search task, or had very few target-object classes, subjects, or fixations. *: Fixations are approximated by mouse clicks.
}
\label{table: dataset cmp}
\end{center}
\vskip -0.3in
\end{table}
\myheading{Search Fixation Datasets}. Another significant contribution of our work is the introduction of the COCO-Search18 dataset, which is currently the world's largest dataset of images that have been annotated with the gaze fixations made by humans during search. COCO-Search18 is needed because the best models of goal-directed attention will likely be models trained on goal-directed behavior. Using free-viewing as an example, the currently best model of free-viewing fixations is DeepGaze II \cite{kummerer2017understanding}, which is a deep network trained on SALICON \cite{jiang2015salicon}. SALICON is a crowd-sourced dataset consisting of images that were annotated with human mouse clicks indicating attentionally salient image locations. Without SALICON, DeepGaze II and models like it, would not have been possible, and our understanding of free-viewing behavior, widely believed to reflect bottom-up attention control, would be greatly diminished.
There is nothing comparable to SALICON for training models of goal-directed attention. Moreover, those suitably large datasets that do exist, each suffer from some weakness that limits their usefulness (\Tref{table: dataset cmp}), with the most common weakness being that the task used to collect the fixation behavior was not visual search (as in \cite{mathe2014actions, papadopoulos2014training}). Other datasets did use a search task, but either had people search for multiple targets simultaneously~\cite{gilani2015pet} or used only one target category (people; \cite{ehinger2009modelling}) or two (microwaves and clocks; \cite{zelinsky2019benchmarking}). All these inadequacies demand a new, larger, and higher-quality dataset of search fixations for model training. We use multiple fixation-based behavioral search metrics to interrogate COCO-Search18, which we attempt to predict using our IRL model and other state-of-the-art methods.
\myheading{Contributions}. This study makes several important contributions:
\textbf{(1)} We apply Inverse Reinforcement Learning (GAIL) to the problem of predicting fixation scanpaths during visual search, the first time this has been done for a goal-directed attention.
\textbf{(2)} In order to apply IRL to scanpath prediction we needed to integrate changes in fixation location with changes in the state representation, a problem that we solved using Dynamic Contextual Beliefs. DCB is a novel state encoder that updates beliefs about peripherally-viewed objects (an object context) based on the movements of a simulated fovea. This is a technical novelty of our method.
\textbf{(3)} We introduce COCO-Search18, a large-scale, high-quality dataset of COCO images annotated with the fixations of 10 people searching for 18 target-object categories. COCO-Search18 makes possible the deep network modeling of goal-directed attention.
\textbf{(4)} We show through model comparison and with multiple metrics that our IRL model outperforms other state-of-the-art methods in predicting search scanpaths. We also show that the IRL model (i) learns an object's scene context; (ii) generalizes to predict the behavior of new subjects, and (iii) needs less data to achieve good performance compared to other models.
\textbf{(5)} Finally, we learned how to quantify a reward function for the fixations in a search task. This will make possible a new wave of experimental investigation that will ultimately result in a better understanding of goal-directed attention.
\section{Scanpath Prediction Framework}
We propose an IRL framework (Fig.~\ref{fig:irl_pipeline}) to model human visual search behavior.
A person performing a visual search task can be considered a goal-directed agent, with their fixations being a sequential decision process of the agent. At each time step, the agent attends to (fixates) a specific location within the image and receives a version of the image that is blurred to approximate the human viewer's visual state, what we call a retina-transformed image. This is an image that has high-resolution (non-blurred) information surrounding the attended location, and lower-resolution information outside of this central simulated fovea~\cite{perry2002gaze}. The state of the agent is determined by the sequence of visual information that accumulates over fixations toward the search target (\Sref{sec:state_rep}), with each action of the agent depending on the state at that time during the evolving state representation. The goal of the agent is to maximize internal rewards through changes in gaze fixation. While it is difficult to behaviorally measure how much reward is received from these fixations, with IRL this reward can be assumed to be a function of the state and the action, and this function can be jointly learned using the imitation policy (\Sref{sec:learning}).
\begin{figure}[t]
\centering
\includegraphics[width=.95\linewidth]{images/IRL_pipe.pdf}
\vskip -0.1in
\caption{{\bf Overview of the IRL framework}. The generator (policy) generates fake state-action pairs $\{B^f_t, a^f_t\}$ by sampling eye movements from given images and tasks. The discriminator (reward function) is trained to differentiates real human state-action pairs $\{B^r_t, a^r_t\}$ from the generated ones and provides reward to train the generator. The states $B^f_t$ and $B^r_t$ use DCB representations.}
\label{fig:irl_pipeline}
\vskip -0.15in
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.95\linewidth]{images/DCB_pipe.pdf}
\vskip -0.1in
\caption{{\bf Overview of DCB}. First, an input image and its low-res image counterpart are converted into the high-res beliefs and low-res beliefs. The initial state $B_0$ is set as the low-res belief. At each fixation which is discretized into a binary fixation mask $M_t$ with 1's at the fixation location and 0's elsewhere, a new state is generated by applying \Eref{eq:heuristic-update}.}
\label{fig:belief_pipeline}
\vskip -0.2in
\end{figure}
\subsection{State Representation Modeling}\label{sec:state_rep}
To model the state of a viewer we propose a novel state representation for accumulating information through fixations that we term a \textbf{Dynamic-Contextual-Belief (DCB)}. As shown in \Fref{fig:belief_pipeline}, DCB is composed of three components:~1) Fovea, which receives a high-resolution visual input only from the image region around the fixation location; 2) Contextual beliefs, which represent a person's gross ``what'' and ``where'' understanding of a scene in terms of level of class confidence; and 3) Dynamics, which actively collects information with each fixation made during search. We discuss each component in greater detail below.
\myheading{Fovea:}
The primate visual system has a fovea, which means that high-resolution visual information is available only at a central fixated location. To accumulate information from the visual world, it is therefore necessary to selectively fixate new image locations. Visual inputs outside of the fovea have lower resolution, with the degree of blur depending on the distance between the peripherally-viewed input and the fovea. Rather than implementing a full progressive blurring of an image input (i.e., a complete retina-transformed image, as in \cite{zelinsky2019benchmarking}), for computational efficiency here we use a local patch from the original image as the high-resolution foveal input and a blurred version of the entire image to approximate low-resolution input from peripheral vision.
\myheading{Contextual Belief:}
Attention is known to be guided to target (and target-like) objects during search, but more recently it has been suggested that attention is also guided to ``anchor objects'' \cite{boettcher2018anchoring, vo2019reading}, defined as those objects having a learned spatial relationship to a target that can help in the efficient localization of that target. For example, people often look at the wall when searching for a TV because TVs are often found hanging on the wall. Inspired by this, we propose to model, not only the target features (as in \cite{zhang2018finding}), but also other objects and background information in the state.
We hypothesize that people have an internal scene parser that takes an image input and generates belief maps for that image based on all the objects and background classes in that person's knowledge structure. We believe these belief maps also guide movements of the fovea for the purpose of capturing high-resolution information and forming better beliefs. We approximate these belief maps using a Panoptic-FPN~\citep{kirillov2019panopticfpn} for panoptic segmentation \cite{kirillov2019panoptic}. Given an image, Panoptic-FPN generates a pixel-level mask for each ``thing'' class (object) and each ``stuff'' class (background) in the image. There are 80 ``thing'' categories (including a single ``other'' class for the 80 ``thing'' classes) and 54 ``stuff'' categories \cite{lin2014microsoft, caesar2018coco, kirillov2019panoptic}. We create a mask for each category by grouping all mask instances belonging to the same category and use the belief maps of the 134 categories as the primary component of the state representation. We term these belief maps {\it contextual beliefs} because the collective non-target beliefs constitute a context of spatial cues that might affect the selection of fixations during the search for a target.
\myheading{Dynamics} refers to the change in the state representation that occurs following each fixation.
We propose a simple yet effective heuristic to model state dynamics (see \Fref{fig:belief_pipeline}). Initially, the state is based on the contextual beliefs on the low-resolution image corresponding to a peripheral visual input. For each fixation by the searcher, we update the state by replacing the portion of the low-resolution belief maps with the corresponding high-resolution portion obtained at the new fixation location. The state is updated as follows:
\begin{align}
B_0 = L \ \textrm{and } B_{t+1} = M_t\odot H + (1-M_t)\odot B_t, \label{eq:heuristic-update}
\end{align}
where $B_t$ is the belief state after $t$ fixations, $M_t$ is the circular mask generated from the $t^{th}$ fixation, $L$ and $H$ are the belief maps of ``thing'' and ``stuff'' locations for low-resolution and high-resolution images, respectively. Humans have different search behaviors on the same image given different search targets. To capture this, we augment the state by concatenating it with a one-hot task vector. Please refer to the supplementary material for more detail.
\subsection{Reward and Policy Learning}\label{sec:learning}
We learn the reward function and the policy for visual search behavior using Generative Adversarial Imitation Learning (GAIL) \cite{ho2016generative}. As shown in \Fref{fig:irl_pipeline}, GAIL is an adversarial framework with a discriminator and a generator. The policy is the generator that aims to generate state-action pairs that are similar to human behavior. The reward function maps a state-action pair to a numeric value, and this function is formulated as the logarithm of the discriminator output. The policy and reward functions are obtained by training the generator and discriminator with an adversarial optimization framework.
Let $D$ and $G$ denote the discriminator and the generator, respectively. The discriminator aims to differentiate human state-action pairs from fake state-action pairs generated by the policy. This corresponds to maximizing the following objective function:
\begin{align}
\mathcal{L}_{D} = & \mathbb{E}_r [\log(D(S, a))] + \mathbb{E}_f [\log(1-D(S,a))] \nonumber \\
& - \lambda \mathbb{E}_r [\left\| \nabla D(S, a))\right\|^2].
\end{align}
In the above objective function, $\mathbb{E}_r$ denotes the expectation over the distribution of real state-action pairs, while $\mathbb{E}_f$ denotes the expectation over the fake samples generated by the generator (i.e., the policy). The last term of the above objective is the expected squared norm of the gradients, which is added for faster convergence \cite{roth2017stabilizing}. The reward function is defined based on the discriminator:
\begin{equation}
r(S, a) = \log (D(S, a)).
\label{eq:rwd_fun}
\end{equation}
The generator aims to fool the discriminator, and its objective is to maximize the log likelihood of the generated state-action pairs, i.e., to maximize: $\mathcal{L}_{G} = \mathbb{E}_f [\log(D (S, a))] = \mathbb{E}_f[r(S, a)]. $
The generator is an RL policy, hence its objective can be equivalently reformulated as an RL objective and optimized by Proximal Policy Optimization \cite{schulman2017proximal}:
\begin{align}
&\mathcal{L}_{\pi} = \mathbb{E}_{\pi} [\log(\pi(a|S)) A(S, a)] + H(\pi).
\end{align}
We use GAE \cite{schulman2015high} to estimate the advantage function $A$ which measures the gain of taking action $a$ over the policy’s default behavior. $H({\pi}) = - \mathbb{E}_{\pi} [\log(\pi(a| S))]$, the entropy in max-entropy IRL \cite{Ziebart-etal-AAAI08}.
\section{COCO-Search18 Dataset}
COCO-Search18 is a large-scale and high-quality dataset of search fixations obtained by having 10 people viewing~6202 images in the search for each of 18 target-object categories. Half of these images depicted an instance of the designated target object and the other half did not, meaning that we adopted a standard target-present (TP) or target-absent (TA) search task. All images in COCO-Search18 were selected from the COCO trainval set \cite{lin2014microsoft}. Five criteria were imposed when selecting the TP images:
\textbf{(1)} No images depicting a person or an animal (to avoid known strong biases to these categories that might skew our measures of attention control~\cite{cerf2008predicting, judd2009learning}).
\textbf{(2)} The image should include one and only one instance of the target.
\textbf{(3)} The size of the target, measured by the area of its bounding box, must be $>$1\% and \textless10\% of the area of the search image.
\textbf{(4)} The target should not be at the center of the image, enforced by excluding an image if the target bounding box overlapped with the center cell of a 5x5 grid.
\textbf{(5)} The original image ratio (width/height) must be between 1.2 and 2.0 to accommodate the display screen ratio of 1.6. After applying these exclusion criteria, and excluding object categories that had less than 100 images of exemplars, we were left with 32 object categories (out of COCO's 80) to use as search targets. To exclude images in which the target was highly occluded or otherwise difficult to recognize, we trained a patch-based classifier for target recognition and only selected images in which the cropped target-object patch had a classification confidence in the top 1\%. Finally, we manually excluded images depicting digital clocks from the clock target category (because the features of analog and digital clocks are very different and this would be expected to reduce data quality by creating variability in the search behavior), as well as images depicting objectionable content. This left us with 3101 TP images over 18 target categories. To select the same number of TA images for each of these 18 categories, we randomly sampled COCO trainval images with the following constraints:
\textbf{(1)} The image should not depict an instance of the target, and \textbf{(2)} The image must include at least two instances of the target's siblings, as defined in COCO. For example, a microwave sibling can be an oven, a toaster, a refrigerator, or a sink, which are under the parent category of appliance. We did this to discourage TA responses from being based on scene type (e.g., a city street scene would be unlikely to contain a microwave).
\begin{figure}
\centering
\includegraphics[width=1.\linewidth]{images/human_matrix_new2.pdf}
\vskip -0.1in
\caption{Normalized gaze data [0,1] on response error, reaction time, number of fixations (NumFix), time to target (T2T), and number of fixations to target (NumFix2T) averaged over 10 subjects searching for 18 categories in TP images. Redder color indicates harder search targets, bluer color indicates easier search.}
\label{fig:gaze_data}
\vskip -0.2in
\end{figure}
Each of the 10 student participants (6 males, age range 18-30, normal or corrected-to-normal vision) viewed all 6202 images, and their eye position throughout was sampled at 1000Hz using an EyeLink 1000 eyetracker (SR Research) in tower-mount configuration under controlled laboratory conditions. For each subject, data collection was distributed over six sessions in six days, with each session having equal number TP trials and TA trials ($\sim$500 each) randomly interleaved. Each session required $\sim$2 hours. For each image, subjects made a TP or TA judgment by pressing a `yes' or `no' button on a game pad. They searched all the images for one target category before preceding to next category. A total of 299,037 fixations were extracted from the eye position data, over the 10 subjects, although only data from the TP fixations will be reported here (\Fref{fig:gaze_data}). TP fixations occurring on error trials, or after fixation on the target, were discarded. This left 100,232 TP search fixations to use for training and testing.
All model evaluations are based on 70\% training, 10\% validation, and 20\% test, random splits of COCO-Search18, within each target category.
\section{Experiments}
We evaluate the proposed framework and its constituent components in multiple experiments. We first compare the scanpath predictions by the IRL-based algorithm to predictions from various heuristic methods and behavior cloning methods using ConvNets and convolutional LSTM. We then study the algorithm's ability to generalize to new human subjects. Finally, we analyze a context effect, the value of having more training data, and report the results of an ablation study. Here we used only the target-present trials from COCO-Search18, leaving analyses of the target-absent data for future study.
\subsection{Comparing Scanpath Prediction Models}\label{sec:policy_cmpr}
\myheading{Comparison methods}. We compare the IRL algorithm for predicting scanpaths to several baselines, heuristics, and behavior cloning methods: (1) \textbf{Random scanpath}: we predict the scanpath for an input image by randomly selecting a human scanpath for the same search target but in a different input image. (2) \textbf{Detector}: we sample a sequence of fixation locations based on the detection confidence map over the image. Regions with higher confidence scores are more likely to be sampled. (3) \textbf{Fixation heuristics}: rather than sampling from a detector's confidence map, here we generate fixations by sampling from a fixation density map produced by a ConvNet trained on human fixation density maps.
(4) {\bf BC-CNN} is a behavior cloning method, where we train a ConvNet to predict the next fixation location from the DCB state representation. Note that this state representation and network structure are identical to the one used by the IRL policy described in \Sref{sec:state_rep}. (5) {\bf BC-LSTM} is a behavior cloning method similar to BC-CNN, but the state representation and update are done with a convolutional LSTM. Instead of having the simple predefined update rule used by both IRL and BC-CNN, as shown in \Eref{eq:heuristic-update}, BC-LSTM aims to learn a recurrent update rule automatically with an LSTM: $B_{t+1} = ConvLSTM(B_t, I_t)$, where $ConvLSTM$ denotes a convolutional LSTM cell~\cite{ballas2015delving}, $B_t$ is the hidden state of the LSTM cell and also the searcher's belief state after $t$ fixations. $I_t$ is the input to the LSTM at time $t$, and it is defined as $I_t = M_t\odot H + (1-M_t)\odot L$. Recall that $M_t$ is the circular mask generated from the $t^{th}$ fixation, $L$ and $H$ are the predicted maps from the Panoptic-FPN \cite{kirillov2019panoptic} for the 80 COCO objects and 54 ``stuff'' classes for low- and high-resolution input images, respectively.
\begin{figure}[t]
\centering
\includegraphics[width=.9\linewidth]{images/search_efficiency.pdf}
\vskip -0.1in
\caption{{\bf Cumulative probability of fixating the
target} for human searchers and all predictive methods. X-axis is the number of fixations until the fovea moves to the target object; Y-axis is the percentage of scanpaths that succeed in locating the target. Means and standard errors are first computed over target categories, and then over searchers.}
\label{fig:cdf}
\vskip -0.2in
\end{figure}
\myheading{Results}. \label{sec:results}
Fig. \ref{fig:cdf} shows the cumulative probability of gaze landing on the target after each of the first 6 fixations made by humans and the algorithms in our model comparison. First, note that even the most predictive models have a performance ceiling lower than that of humans, whose ceiling over this range is nearly 1. These lower ceilings likely reflect a proportion of trials in which the models search was largely unguided. Second, note the steep increase in target fixation probability after the first and second fixations. The slopes of these functions indicate strong target guidance. The target was fixated in the very first movement on about half of the images, with the IRL model replicating human search guidance slightly better than its nearest competitors: the Detector and BC-CNN models.
We quantify the patterns from Fig. \ref{fig:cdf} using several metrics. Two of these metrics follow directly from Fig. \ref{fig:cdf} and capture aggregate measures combining search guidance and accuracy. The first of these computes the area under the cumulative probability of the target fixation curve, a metric we refer to as {\bf Target Fixation Probability AUC} or TFP-AUC. Second, we compute the sum of the absolute differences between the human and model cumulative probability of target fixation in a metric that we refer to as {\bf Probability Mismatch}. We also report the {\bf Scanpath Ratio}, which is a widely used metric for search efficiency. It is computed by the ratio of Euclidean distance between the initial fixation location and the center of the target to the summed Euclidean distances between fixations to the target~\citep{hout2015target}. Finally, we compute two metrics for scanpath prediction success, that both capture the scanpath similarity between fixation sequences generated by humans and sequences generated by the model. The first of these computes a {\bf Sequence Score} by first converting a scanpath into a string of fixation cluster IDs and then use a string matching algorithm \cite{Needleman-Wunsch-JMB70} to measure similarity between two strings. Finally, we use {\bf MultiMatch}~\cite{anderson2015comparison, dewhurst2012depends} to measure the scanpath similarity at the pixel level. MultiMatch measures five aspects of scanpath similarity: shape, direction, length, position, and duration. We exclude the duration metric because the studied models do not predict fixation duration. Unless otherwise specified, each model generates 10 scanpaths of maximum length 6 (excluding the first fixation) for each testing image by sampling from the predicted action map at each fixation, with the results averaged over scanpaths.
As seen from \Tref{table:all_results}, the IRL algorithm outperforms the other methods on all metrics. The performance of IRL is closest to {\it Human}, an oracle method where the scanpath of a subject is used to predict the scanpath of another subject for the same input image. \Fref{fig:rwd_maps} also shows that {\bf reward maps} recovered by the IRL model depend greatly on the category of the search target. In the top row, higher reward was assigned to the laptop when searching for a mouse, while for the same image greater reward was expected from fixating on the monitor when searching for a tv. Similarly, the search for a car target in the bottom image resulted in the expectation of reward from the other cars on the road but almost not at all from the highly-salient stop sign, which becomes intensely prioritized when the stop sign is the target.
\begin{table*}[t]
\begin{center}
\begin{tabular}{l|c|c|c|c|cccc}
& \multirow{2}{*}{TFP-AUC $\bm{\uparrow}$} & \multirow{2}{1.8cm}{Probability Mismatch $\bm{\downarrow}$} & \multirow{2}{1.5cm}{Scanpath Ratio $\bm{\uparrow}$} & \multirow{2}{1.5cm}{Sequence Score $\bm{\uparrow}$} & \multicolumn{4}{c}{MultiMatch $\bm{\uparrow}$} \\ \cline{6-9}
& & & & & shape & direction & length & position \\ \Xhline{1.2pt}
Human & 5.200 & - & 0.862 & 0.490 & 0.903 & 0.736 & 0.880 & 0.910 \\ \hline
Random scanpath & 0.795 & 4.407 & - & 0.295 & 0.869 & 0.558 & 0.849 & 0.849 \\
Detector & 4.046 & 1.166 & 0.687 & 0.414 & 0.877 & 0.676 & 0.853 & 0.863 \\
Fixation heuristic & 2.154 & 3.046 & 0.545 & 0.342 & 0.873 & 0.614 & \textbf{0.870} & 0.850 \\
BC-CNN & 3.893 & 1.328 & 0.706 & 0.409 & 0.880 & 0.669 & 0.865 & 0.874 \\
BC-LSTM & 1.702 & 3.497 & 0.406 & 0.324 & 0.834 & 0.567 & 0.818 & 0.770 \\
IRL(Ours) & \textbf{4.509} & \textbf{0.987} & \textbf{0.826} & \textbf{0.422} & \textbf{0.886} & \textbf{0.695} & 0.866 & \textbf{0.885}
\end{tabular}
\end{center}
\vskip -0.2in
\caption{ {\bf Comparing scanpath prediction algorithms} (rows) using multiple scanpath metrics (columns) on the COCO-Search18 test dataset. In the case of Sequence Score and Multimatch, \enquote{Human} refers to an oracle method where one searcher's scanpath is used to predict another searcher's scanpath; \enquote{Human} for all other metrics refers to observed behavior.}
\label{table:all_results}
\vskip -.2in
\end{table*}
\myheading{Implementation details.}
We resize each input image to $\htimesw{320}{512}$ and obtain a low-resolution image by applying a Gaussian filter with standard deviation \mbox{$\sigma = 2$}. To compute the contextual beliefs, we use a Panoptic-FPN with backbone network ResNet-50-FPN pretrained on COCO2017 \cite{kirillov2019panoptic}. Panoptic-FPN outputs a feature map of 134 channels, corresponding to 80 object categories and 54 background classes in COCO, and it is resized to $\htimesw{20}{32}$ spatially.
For IRL and BC-CNN, we use the same policy network architecture: a network composed of four convolutional (conv) layers and a softmax layer.
IRL model has two additional components---critic network and discriminator network. The \textbf{critic network} has two convolutional layers and two fully-connected (fc) layers. The discriminator network shares the same sturcture with the IRL policy network except the last layer which is a sigmoid layer.
Each conv layer and fc layer in BC-CNN and IRL is followed by a ReLU layer.
BC-LSTM has the same policy network as the BC-CNN, with the difference being the use of a convolutional LSTM~\cite{ballas2015delving} to update the states. BC-CNN and BC-LSTM use the KL divergence between predicted spatial distribution and ground truth as loss.
The prediction of both behavior cloning models and IRL is conditioned on the search target. We implement the target conditioning by introducing an additional bias term based on the search task to the input features at each layer \cite{perez2018film}. The human visual system employs Inhibition-of-Return~(IOR) to spatially tag previously attended locations with inhibition to discourage attention from returning to a region where information has already been depleted~\cite{wang2010searching}. To capture this mechanism, we enforce IOR on the policy by setting the predicted probability map to 0 at each attended location using a \htimesw{3}{3} grid.
See the supplementary for more detail.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.48\textwidth]{images/rwd_maps_diff_targets.pdf}
\end{center}
\vskip -0.2in
\caption{Initial {\bf reward maps} learned by the IRL model for two different search targets in two test images. Top row: original image (left), mouse target (middle), and tv target (right). Bottom row: original image (left), car target (middle), stop sign target (right). Redder color indicates the expectation of higher reward for fixating a location.
}
\label{fig:rwd_maps}
\vskip -.2in
\end{figure}
\subsection{Group Model vs Individual Model}
The previous subsection described the IRL model's ability to predict a searcher's scanpath on unseen test images, but how well can this model predict the scanpaths of a new unseen searcher without training on that person's scanpaths? To answer this question, we perform ten leave-one-subject-out experiments, with each experiment corresponding to a test subject. For every subject we train two models: (1) a group model using the scanpaths of the 9 other subjects; and (2) an individual model using the scanpaths of the test subject on the training images. We evaluate the performance of these models on the scanpaths of each test subject on the unseen test images. \Fref{fig:group vs indv} shows that both models perform well, with an insignificant performance gap between them. This suggests that there is good agreement between group and individual behaviors, and that a group model can generalize well to new searchers.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=.48\textwidth]{images/group_indv2.pdf}
\end{center}
\vskip -0.25in
\caption{No significant differences were found between a group model (solid), trained with 9 subjects, and an individual model (striped), trained with one subject. }
\label{fig:group vs indv}
\vskip -0.2in
\end{figure}
\subsection{Context Effects}\label{sec:context}
\myheading{Search efficiency}. With DCB we can ask how an object from category $A$ affects the search for a target from category~$B$. This effect can either increase (guidance) or decrease (distraction) search efficiency. To study this, we first zero out the belief map of category $A$ in the DCB state representation and then measure the TFP-AUC (see \Sref{sec:results}) on test images for category $B$. We compute the difference between the TFP-AUC obtained with and without switching off the belief map for category $A$ in DCB. A positive value indicates that an object in category $A$ helps to guide search for a target in category $B$, while a negative value indicates the opposite (that the object is distracting). We did this for the 134 COCO objects and stuff non-target categories $A$ and the 18 target categories $B$. \Fref{fig:ctx_effect} shows the six most guiding and distracting objects for the knife and car searches. Note that the fork was highly distracting when searching for a knife, likely because the two look similar in periphery vision, but that the cake facilitated the knife search. Similarly for the car search, pavement provided the strongest guidance whereas trucks were the most distracting.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{images/context_bar_new.pdf}
\vskip -0.1in
\caption{{\bf Context effect}. The six most influential context objects (grey bars) for knife and car search tasks. The y-axis is the context object category and the x-axis is a measure of how much the belief map for a context object contributed to search efficiency, as measured by TFP-AUC. Larger positive values mean that the context object improved search guidance to the target, more negative values mean that the object distracted attention from the search.}
\label{fig:ctx_effect}
\vskip -0.1in
\end{figure}
\myheading{Directional Prior}. Can an object from category $A$ serve as a directional spatial cue in the search for a target from category $B$? Suppose $M$ is the probability map produced by the policy network of our IRL model, and let $M'$ be the modified probability map from the policy network but with the belief map of category $A$ in the DCB state representation being switched off. By computing the difference between~$M'$ and~$M$ which we call a \textit{context map} (as depicted in the top row of \Fref{fig:spatial}), we can see the spatial relationship between the context object $A$ and the target object $B$ (see \Fref{fig:spatial} for examples).
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{images/spatial.pdf}
\vskip -0.1in
\caption{{\bf Spatial relations} between context and target objects learned by the model. Top row shows individual context maps for a dining table (left) and a refrigerator (middle) in a microwave search, and a keyboard (right) in a mouse search. Bottom row are the belief maps of the corresponding context objects. Gaze is guided to the top of the dinning table and refrigerator when searching for a microwave, and to the right of the keyboard when searching for a mouse.}
\vskip -0.15in
\label{fig:spatial}
\end{figure}
\subsection{Ablation Study on State Representation}\label{sec:state_cmpr}
DCB is a rich representation that uses top-down, bottom-up, and history information. Specifically, it consists of 136 belief maps, divided into five factor groups: target object (1 map), context objects (79 maps), ``stuff'' (54 maps), saliency (1 map, extracted using DeepGaze2 \cite{kummerer2017understanding}), and history (1 binary map for the locations of previous fixations). To understand the contribution of a factor group, we remove the group from the full state representation and measure the effect on performance. As we describe in the supplementary material, the most important factor groups were target and context objects, followed by stuff, whereas saliency and history weakly impacted model performance. In addition,
an alternative state representation to DCB is the Cumulative Foveated Image (CFI)~\cite{zelinsky2019benchmarking}, but replacing DCB with CFI degrades the performance of IRL
Full results can be found in the supplementary material.
\subsection{Data Efficiency}
\Fref{fig: data_eff} shows IRL and BC-CNN performance as we vary the number of training images per object category. Both methods use DCB as the state representation. IRL is more data efficient than BC-CNN, achieving comparable or better results using less training data. A likely reason for this is that the GAIL-based~\cite{ho2016generative} IRL method includes an adversarial component that generates augmented training data, leading to a less prone to overfitting policy network. Data efficiency is crucial for training for new categories, given the time and cost of collecting human fixations.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{images/data_efficiency3.pdf}
\vskip -0.1in
\caption{Performance of IRL (solid line) and BC-CNN (dashed line) as the number of training images per category increases from 5 to 20. IRL is more data efficient than BC-CNN, likely due to an adversarial data generator.}
\label{fig: data_eff}
\vskip -0.2in
\end{figure}
\section{Conclusions}
We proposed a new model for predicting search fixation scanpaths that uses IRL to jointly recover the reward function and policy used by people during visual search. The IRL model uses a novel and highly explainable state representation, \textit{dynamic contextual beliefs (DCB)}, which updates beliefs about objects to obtain an object context that changes dynamically with each new fixation. To train and test this model we also introduced COCO-Search18, a large-scale dataset of images annotated with the fixations of people searching for target-object goals. Using COCO-Search18, we
showed that the IRL model outperformed comparable models in predicting search scanpaths.
Better predicting human search behavior means better robotic search applications and human-computer systems that can interact with users at the level of their attention movements \cite{park2019advancing}. It may also be possible to use reward maps from the IRL model to annotate and index visual content based on what is likely to attract a person's attention. Finally, our work impacts the behavioral vision literature, where the visual features guiding human goal-directed attention are still poorly understood for real images \cite{zelinsky2013modeling}.
\myheading{Acknowledgements}. This project is supported by US National Science Foundation Award IIS-1763981, the Partner University Fund, the SUNY2020 Infrastructure Transportation
Security Center, and a gift from Adobe.
{\small
\setlength{\bibsep}{0pt}
\bibliographystyle{plainnat}
\section{Appendix}
This section provides further details about the COCO-Search18 dataset (\Sref{sec:dataset}), Dynamic Contextual Beliefs (\Sref{sec:DCB}), and implementation (\Sref{sec:impl_details}). We also include additional results from experiments and ablation studies, and interpretation (\Sref{sec:results}).
\subsection{Details about COCO-Search18 Dataset}\label{sec:dataset}
\myheading{Data source:}
The COCO-Search18 dataset annotates COCO \cite{lin2014microsoft} with human gaze fixations made during a standard target-present (TP) or target-absent (TA) search task, where on each trial the search image either depicted the target (TP) or it did not (TA). All of the images were selected from the {\it trainval} set, and detailed descriptions of TP and TA image selection and gaze collection methods are provided below.
\myheading{Target present image selection:}
In addition to the exclusion criteria described in the main text, we also excluded images in which the target was highly occluded or otherwise difficult to recognize. Specifically, we only selected images in which the cropped target-object patch had a classification confidence $>$.99.
To train this classifier, we cropped the target in each image (by bounding box) and used these image patches as positive samples. Same-sized image patches of non-target objects were used as negative samples. Negative samples were constrained to intersect with the target by 25\% (area of intersection divided by area of target) so that they could serve as hard negatives for specific targets. More than 1 million cropped patched were collected and resized to 224x224 pixels, while keeping the original aspect ratio by padding. The classifier is fine-tuned from an ImageNet-pretrained ResNet-50 model with the last fully connected layer changed from 1000 outputs to 33 (32+``Negative''). Images with a classification score for the cropped target patch that was \textless.99 were excluded. This resulted in 18 categories with at least 100 images in each category, and 3131 images in total. As described in the main text, we conducted a final manual checking of the dataset to exclude images depicting digital clocks (5 images), so as to make the clock target category specific to analog clocks, and to remove images depicting content that participants might find objectionable. This latter criterion resulting in the exclusion of 30 images, 22 of which were from the toilet category.
\begin{table}
\begin{center}
\begin{tabular}{lcccc}
\toprule
Category & TP images & ACC & TA images & ACC \\
\midrule
bottle & 166 & 0.84 & 166 & 0.92 \\
bowl & 141 &0.80 & 141 & 0.90\\
car & 104 &0.89 & 104 & 0.91\\
chair & 253 &0.89 & 253 & 0.64\\
clock & 119 &0.99 & 119 & 0.97\\
cup & 276 &0.92 & 276 & 0.76\\
fork & 230 &0.96 & 230 & 0.98\\
keyboard & 184 &0.92 & 184 & 0.98\\
knife & 141 &0.89 & 141 & 0.97\\
laptop & 123 &0.95 & 123 & 0.95\\
microwave & 156 &0.97 & 156 & 0.95\\
mouse & 109 &0.97 & 109 & 0.97\\
oven & 101 &0.91 & 101 & 0.93\\
potted plant & 154 &0.84 & 154 & 0.95\\
sink & 279 &0.97 & 279 & 0.94\\
stop sigh & 126 &0.95 & 126 & 0.99\\
toilet & 158 &0.99 & 158 & 1.00\\
tv & 281 &0.96 & 281 & 0.93\\
\midrule
total/mean & 3101 & 0.92 & 3101 & 0.92\\
\bottomrule
\end{tabular}
\caption{Number of images and response accuracy (ACC) for TP and TA images grouped by target category.}
\label{table: image numbers}
\end{center}
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=1.\linewidth]{images/gaze_example.pdf}
\vskip -0.05in
\caption{Examples of human scanpaths during target-present (top 3 rows) and target-absent (bottom 3 rows) visual search. From left to right and top to bottom, the 18 target categories are: bottle, bowl, car, chair, clock, cup, fork, keyboard, knife, laptop, microwave, mouse, oven, potted plant, sink, stop sign, toilet, and tv. Each yellow line represents the scanpath of one behavioral searcher, with numbers indicating fixation order. }
\label{fig:scanpath_example}
\end{figure*}
After implemented all exclusion criteria, we selected 3101 target-present images from 18 categories: bottle, bowl, car, chair, clock, cup, fork, keyboard, knife, laptop, microwave, mouse, oven, potted plant, sink, stop sign, toilet, tv. See Table \ref{table: image numbers} for the specific number of images in each category and the average response accuracy (ACC). There were an equal number of TA images (for a total of 6202 images), which were all resized and padded to fit the ${1050}\times{1680}$ resolution of the display monitor.
\myheading{Gaze data collection procedure:}
Ten university undergraduate and graduate students (6 males, age range 18--30) with normal or corrected to normal vision participated in this study, which was approved by the Institutional Review Board. They were naive with respect to experimental question and design, and were compensated with course credits or money for their participation. Informed consent was obtained at the beginning of the experiment, and every participant read and understood the consent form before signing it.
The 6202 images were divided into six days of experiment sessions with each session consisting of $\sim$500 TP images and the same number of TA images, randomly interleaved. Images for a given target category were grouped and presented sequentially in an experiment block (i.e., target type was blocked). Preceding each block was a calibration procedure needed to map eye position obtained from the eye-tracker to screen coordinates, and a calibration was not accepted until the average calibration error was $\leq$.51 and the maximal error was $\leq$.94. Each trial began with a fixation dot appearing at the center of the screen. To start a trial, the subject should press the ``X'' button on a gamepad while carefully looking at the fixation dot. A scene would then be displayed and their task was to answer ``yes'' or ``no'' whether an exemplar of the target category for that block appears in the displayed scene. The subject registered a ``yes'' target judgment by pressing the right rigger of the gamepad, and a ``no'' judgment by pressing the left trigger. They were told that there were equal number of target present and absent trials, and that they should respond as quickly as possible while remaining accurate. Participants were allowed to take multiple breaks between and within each block.
Image presentation and data collection was controlled by Experiment Builder (SR research Ltd., Ottawa, Ontario, Canada). Images were presented on a 22-inch LCD monitor (resolution: 1050x1680), and subjects viewed these stimuli in a distance of 47cm from the monitor, enforced by both chin rest and head rest. Eye movements were recorded using an EyeLink 1000 eye tracker in tower-mount configuration (SR research Ltd., Ottawa, Ontario, Canada). The experiment was conducted in a quiet and dimmed laboratory room. \Fref{fig:scanpath_example} shows some TP and TA images from the 18 object categories, with overlaid human scanpaths.
\subsection{Detailed Description of DCB}\label{sec:DCB}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\linewidth]{images/DCB_detail.pdf}
\vskip -0.1in
\caption{{\bf Detailed illustration of Dynamic-Contextual-Belief}. First, an input image and its low-res image counterpart are converted into high-res beliefs and low-res beliefs. At each fixation, which is discretized into a binary fixation history map with 1's around the fixation location and 0's elsewhere, a new state is generated by concatenating the output of \Eref{eq:dcb} with a one-hot task embedding (best viewed in color).}
\label{fig:dcb_detail}
\vskip -0.2in
\end{figure*}
\myheading{DCB:}
An input image is resized to $\htimesw{320}{512}$ for computational efficiency (the original image is \htimesw{1050}{1680}), while the blurred image is obtained by applying a Gaussian filter on the original image with the standard deviation $\sigma = 2$. Both images are passed through a Panoptic-FPN with backbone network ResNet-50-FPN pretrained on COCO2017 \cite{kirillov2019panoptic}. The output of the Panoptic-FPN has 134 feature maps, consisting of 80 ``thing'' categories (objects) and 54 ``stuff'' categories (background) in COCO. Feature maps are then resized to
$\htimesw{20}{32}$ spatially, same as the discretization of fixation history. At a given time step $t$, feature maps $H$ for the original image and feature maps $L$ for the blurred image are combined for DCB:
\begin{equation}\label{eq:dcb}
B_t = M_t\odot H + (1-M_t)\odot L
\end{equation}
where $\odot$ is element-wise product and $M_t$ is the mask generated from fixation history and repeated over feature channels (see \Fref{fig:dcb_detail}).
Note that the above equation is equivalent to Eq. (1) in the main paper which is written in a recurrent form.
\myheading{Encoding the target object category:} The task embedding used in our model is the one-hot encoding maps which spatially repeat the one-hot vector. To make predictions conditioned on the task, inputs of each convolutional layer are concatenated with this embedding. This is equivalent to adding a task-dependent bias term for every convolutional layer.
\subsection{Implementation details}\label{sec:impl_details}
\myheading{Action Space}.
Our goal is to predict the pixel location where the person is looking in the image during visual search. To reduce the complexity of prediction, we discretize the image into a $\htimesw{20}{32}$ grid, with each patch corresponding to $\htimesw{16}{16}$ pixels in the original image coordinates. This descretized grid defines the action space for all models tested in this paper. At each step, the policy chooses one out of 640 patches and the center location of that selected patch in the original image coordinates is used as an action. The maximum approximation error due to this discretization procedure is $1.75$ degrees visual angle.
\myheading{IRL}.
The IRL model is composed of three components---the policy network, the critic network and the discriminator network.
The \textbf{policy network} consists of four convolutional layers whose kernel sizes are 5, 3, 3, 1 with padding 2, 1, 1, 0 and output channels are 128, 64, 32 and 1, and a softmax layer to output a final probability map.
The \textbf{critic network} has two convolutional layers of kernel size 3 and two fully-connected (fc) layers whose output sizes are 64 and 1. The convolutional layers have output sizes 128 and 256, respectively, and each is followed by a max pooling layer of kernel size 2 to compress the feature maps into a vector. Then this feature vector is regressed to predict the value of the state through two fc layers . The \textbf{discriminator network} shares the same structure with the IRL policy network except that the last layer is a sigmoid layer.
Note that all convolutional layers and fully-connected layers are followed by a ReLU layer except the output layer.
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=.9\textwidth]{images/seq_score_spvis.pdf}
\end{center}
\vskip -0.2in
\caption{Left: cumulative distribution of the sequence scores of the proposed IRL scanpath prediction method. Right: Four qualitative examples. Human scanpaths are colored in yellow, and the IRL-generated scanpaths are in green. The sequence score for the generated scanpaths are 0.33, 0.40, 0.50, and 0.75, from top to bottom.}
\label{fig:seq_score_cdf}
\end{figure*}
The critic network is jointly trained with the policy network to estimate the value of a state (i.e., expected return) using smoothed $L_1$ loss. The estimated value is used to compute the advantage $A(S, a)$ (note that the state $S$ is represented by the proposed DCB in our approach) in Eq. (4) of the main paper using the Generalized Advantage Estimation (GAE) algorithm \cite{schulman2015high}. At each iteration, the policy network first generates two scanpaths by sampling fixations from the current policy outputs for each image in a batch. Second, we break the generated scanpaths into state-action pairs and sample the same number of state-action pairs from ground-truth human fixations to train the discriminator network which discriminates the generated fixation from behavioral fixations. Lastly, we update the policy and critic network jointly using the PPO algorithm \cite{schulman2017proximal} by maximizing the total expected rewards which are given by the discriminator (see Eq. (3) of the main paper).
\myheading{Training:}
The IRL model was trained for 20 epochs with an image batch size of 128. The batch sizes used for training the discriminator and policy networks were 64. For the PPO algorithm, the reward discount factor, the clip ratio and number of epochs were set to 0.99, 0.2, and 10, respectively. The extra discount factor in the GAE algorithm was set to 0.96. Both the policy network and the discriminator network were trained with a learning rate of $0.0005$. It took approximately 40 minutes to train the proposed IRL model (for 20 epochs) on a single NVIDIA Tesla V100 GPU. The training procedure consumed about 5.6GB GPU memory. Note that the segmentation maps used to construct the DCB state representation had been computed beforehand.
\myheading{Additional details on two baseline methods}. \textbf{Detector}: The detector network consists of a feature pyramid network (FPN) for feature extraction (1024 channels) with a ResNet50 pretrained on ImageNet as the backbone and a convolution layer for detection of 18 different targets. The detector network predicts a 2D spatial probability map of the target from the image input and is trained using the ground-truth location of the target. Another similar baseline is \textbf{Fixation Heuristics}.
This network shares exactly the same network architecture with the detector baseline but it is trained with behavioral fixations in the form of spatial fixation density map (FDM), which is generated from 10 subjects on the training images.
\myheading{Scanpath Generation}.
When generating scanpaths, a fixation location is sampled from the probability map that the models have produced and Inhibition-of-Return is applied to prevent revisiting previously attended locations. All predictive methods including IRL, behavior cloning, and heuristic methods, generate a new spatial probability map at every step, while the predicted probability map is fixed over all steps for the Detector and Fixation Heuristic baselines.
\begin{table*}[t!]
\begin{center}
\begin{tabular}{l|c|c|c|c|cccc}
\multirow{2}{2.2cm}{State Representation} & \multirow{2}{1.3cm}{\centering Sequence Score $\bm{\uparrow}$} & \multirow{2}{1.3cm}{\centering Scanpath Ratio $\bm{\uparrow}$} & \multirow{2}{1.2cm}{\centering TFP- AUC $\bm{\uparrow}$} & \multirow{2}{1.7cm}{\centering Probability Mismatch $\bm{\downarrow}$} & \multicolumn{4}{c}{MultiMatch $\bm{\uparrow}$} \\
\cline{6-9}
& & & & & shape & direction & length & position\\
\Xhline{1.2pt}
DCB & {\bf 0.422} & {\bf 0.826} & {\bf 4.509} & {\bf 0.987} & {\bf 0.886} & {\bf 0.695} & 0.866 & {\bf 0.885}\\
CFI & 0.402 & 0.619 & 3.412 & 1.797 & 0.875 & 0.666 & 0.864 & 0.857\\
\end{tabular}
\end{center}
\vskip -0.2in
\caption{{\bf Dynamic contextual belief (DCB) vs. cumulative foveated image (CFI)} under the framework of IRL.}
\label{table:cfi_cmpr}
\end{table*}
\begin{table*}[t!]
\begin{center}
\begin{tabular}{l|c|c|c|c|cccc}
\multirow{2}{1.8cm}{State Representation} & \multirow{2}{1.3cm}{\centering Sequence Score $\bm{\uparrow}$} & \multirow{2}{1.3cm}{\centering Scanpath Ratio $\bm{\uparrow}$} & \multirow{2}{1.2cm}{\centering TFP- AUC $\bm{\uparrow}$} & \multirow{2}{1.7cm}{\centering Probability Mismatch $\bm{\downarrow}$} & \multicolumn{4}{c}{MultiMatch $\bm{\uparrow}$} \\
\cline{6-9}
& & & & & shape & direction & length & position\\
\Xhline{1.2pt}
DCB with all components & 0.422 & 0.803 & 4.423 & 1.029 & 0.880 & 0.676 & 0.841 & 0.888\\
w/o history map & 0.419 & 0.800 & 4.397 & 1.042 & 0.882 & 0.672 & 0.844 & 0.887\\
w/o saliency map & 0.419 & 0.795 & 4.403 & 1.029 & 0.880 & 0.675 & 0.840 & 0.887\\
w/o stuff maps & 0.407 & 0.777 & 4.111 & 1.248 & 0.876 & 0.662 & 0.836 & 0.875\\
w/o thing maps & 0.331 & 0.487 & 2.047 & 3.152 & 0.855 & 0.605 & 0.852 & 0.818 \\
w/o target map & 0.338 & 0.519 & 2.274 & 2.926 & 0.866 & 0.613 & 0.837 & 0.820 \\
\end{tabular}
\end{center}
\vskip -0.2in
\caption{{\bf Ablation study of the proposed state representation---dynamic contextual belief}. The full state consists of 1 history map, 1 saliency map, 54 stuff maps, 79 context maps and 1 target map. We mask out one part by setting the map(s) to zeros at each time.}
\label{table:abalation}
\end{table*}
\begin{table*}[t!]
\begin{center}
\begin{tabular}{l|c|c|c|c|cccc}
\multirow{2}{2.5cm}{Scanpath generation policy} & \multirow{2}{1.3cm}{\centering Sequence Score $\bm{\uparrow}$} & \multirow{2}{1.3cm}{\centering Scanpath Ratio $\bm{\uparrow}$} & \multirow{2}{1.2cm}{\centering TFP- AUC $\bm{\uparrow}$} & \multirow{2}{1.7cm}{\centering Probability Mismatch $\bm{\downarrow}$} & \multicolumn{4}{c}{MultiMatch $\bm{\uparrow}$} \\
\cline{6-9}
& & & & & shape & direction & length & position\\
\Xhline{1.2pt}
Based on total reward & {\bf 0.422} & {\bf 0.826} & {\bf 4.509} & {\bf 0.987} & {\bf 0.886} & {\bf 0.695} & 0.866 & {\bf 0.885}\\
Based on immediate reward & 0.375 & 0.704 & 3.893 & 2.143 & 0.886 & 0.648 & {\bf 0.873} & 0.852 \\
\end{tabular}
\end{center}
\vskip -0.2in
\caption{{\bf IRL model predictions using Greedy (immediate reward) and Non-greedy (total reward) policy}.}
\label{table:greedy_cmpr}
\end{table*}
\begin{table*}[t!]
\begin{center}
\begin{tabular}{l|c|c|c|c|cccc}
& \multirow{2}{1.3cm}{\centering Sequence Score $\bm{\uparrow}$} & \multirow{2}{1.3cm}{\centering Scanpath Ratio $\bm{\uparrow}$} & \multirow{2}{1.2cm}{\centering TFP- AUC $\bm{\uparrow}$} & \multirow{2}{1.7cm}{\centering Probability Mismatch $\bm{\downarrow}$} & \multicolumn{4}{c}{MultiMatch $\bm{\uparrow}$} \\
\cline{6-9}
& & & & & shape & direction & length & position\\
\Xhline{1.2pt}
IRL, 20 ipc & 0.415 & 0.808 & 4.324 & 1.140 & 0.875 & 0.672 & 0.832 & 0.879\\
CNN, 20 ipc & 0.408 & 0.723 & 3.906 & 1.325 & 0.884 & 0.664 & 0.849 & 0.878\\
\hline
IRL, 10 ipc & 0.409 & 0.774 & 4.029 & 1.318 & 0.881 & 0.591 & 0.851 & 0.819\\
CNN, 10 ipc & 0.397 & 0.678 & 3.542 & 1.657 & 0.877 & 0.594 & 0.847 & 0.821\\
\hline
IRL, 5 ipc & 0.389 & 0.723 & 3.696 & 1.603 & 0.876 & 0.588 & 0.844 & 0.813\\
CNN, 5 ipc & 0.388 & 0.678 & 3.484 & 1.731 & 0.886 & 0.594 & 0.862 & 0.828\\
\end{tabular}
\end{center}
\vskip -0.2in
\caption{{\bf Data efficiency of IRL and CNN}. ``ipc'' means images per category used for training. For exmaple, IRL 10 ipc means we train the IRL model using 10 images from each category which are randomly selected from the training data. CNN and IRL are trained and tested on the same images for fair comparison.}
\label{table:data_efficiency}
\end{table*}
\subsection{Additional Experiment Results}\label{sec:results}
\myheading{Cumulative distribution of sequence scores}. In the main paper we reported the {\it average} Sequence Score of 0.422 for the scanpaths generated by the IRL model. To put this in perspective, \Fref{fig:seq_score_cdf} plots the cumulative distribution of the sequence scores and shows four qualitative examples that have sequence scores of 0.33, 0.40, 0.50, and 0.75, respectively.
\myheading{Comparing different state representations}. To evaluate the benefits of having DCB as the state representation, we compared its predictive performance with the Cumulative Foveated Image (CFI)~\cite{zelinsky2019benchmarking} under the same IRL framework. CFI is created by extracting CNN feature maps on the retina-transformed images which are progressively more blurred based on the distance away from the currently fixated location. On the other hand, the DCB is created by extracting panoptic segmentations~\cite{kirillov2019panopticfpn} on uniform-blur images which are uniformly blurred except around the fixated region (the level of blurriness applied in DCB is close to the middle level in the blur pyramid of CFI \cite{zelinsky2019benchmarking,huang2015salicon,perry2002gaze}). For a fair comparison, we extract features for CFI using the backbone ResNet-50-FPN network from the Panoptic-FPN \cite{kirillov2019panopticfpn} that was used in DCB. Both DCB and CFI have the same spatial resolution of \htimesw{20}{32}. As shown in \Tref{table:cfi_cmpr}, the IRL model with DCB achieves significantly higher search efficiency and scanpath similarity than when using CFI as state representation. Specifically, DCB reduces the search gap by approximately 45\% and improves the scanpath ratio from 61.9\% to 82.6\%, much closer to the human behavioral ratio of 86.2\%. This result is even more impressive considering the size differences between the policy network used with DCB and CFI: DCB is trained with a smaller policy network, since it is comprised of 134 channels, nearly 8x smaller than CFI of 1024 channels. In our experiment, the policy network with CFI state representation has 29.6M parameters, while the policy network with DCB state representation only has 0.3M parameters. Relatedly, another benefit of having DCB as state representation is that it is memory and operation efficient. Creating DCB requires a smaller computational cost than creating CFI, since there's only a single level of blurriness in DCB and extracted panoptic segmentation maps are smaller by an order of magnitude than the feature maps extracted for CFI. Given that IRL models are particularly difficult to train in high dimensional environments~\cite{tucker2018inverse}, having an efficient representation like DCB can be very helpful.
\myheading{State Ablation}.
DCB is a rich representation that incorporates top-down, bottom-up, and history information. The full representation consists of 136 belief maps, which can be divided into five groups: target object (1 map), ``thing'' (object, 79 maps), ``stuff'' (background classes, 54 maps), saliency (1 map, extracted using DeepGaze2 \cite{kummerer2017understanding}), and history (1 binary map for the locations of previous fixations). To understand the contribution of each factor, we removed the maps of each group one at a time and compared the resulting model's performance. As shown in \Tref{table:abalation}, target object and ``thing'' maps are the most critical for generating human-like scanpaths, followed by ``stuff'' maps, whereas saliency and history do not have strong impact to the model performance.
\myheading{Greedy vs. Non-greedy search behavior}.
How does human search behavior compare to generated scanpaths reflecting either Greedy or Non-greedy reward policies?
Under the greedy policy, the selection of each location to fixate during search reflects a maximization of immediate reward. But the greedy policy is highly short-sighted -- it only seeks reward in the short term. Non-greedy reward seeks to maximize the total reward that would be acquired over the sequence of fixations comprising a scanpath. This policy therefore does not maximize reward in the near term, but rather allows more exploration that leads to higher total reward. As shown in \Tref{table:greedy_cmpr}, we generated greedy and non-greedy policies from our IRL model and compared their predictive performance on human scanpaths. The results show that 1) models using greedy vs. non-greedy policy produce different search behaviors, with the model using non-greedy policy generating more human-like scanpaths by all tested metrics. This is an interesting finding. Despite the high efficiency of human search in our study (1-2 sec), the search process was strategic in that the fixations maximized total reward, even over that short period of time.
\myheading{Data Efficiency}.
Table \ref{table:data_efficiency} shows the full results of IRL and BC-CNN given different numbers of training images across different metrics. Both use DCB as the state representation. The results are consistent with the results presented in the main paper and suggest that IRL is more data-efficient when compared to the CNN -- IRL achieved comparable or better results than the CNN using less training data.
| 896a9bf585b71f70a16b1d548c040d03f89f59ab | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Scoliosis is an abnormal condition characterized by lateral spinal curvature. Early assessment and treatment planning is critical \citep{weinstein2008adolescent}. Conventionally, the assessment of scoliosis is performed manually by clinicians through the identification and localization of vertebral structures in spinal X-ray images. However, large inter-patient anatomical variation and poor image quality challenge clinicians to assess the severity of scoliosis accurately and reliably. Automated measurement promises to enable the reliable quantitative assessment of scoliosis.
Several spinal landmark detection methods are available in the literature: Conventional hand-crafted feature engineering \citep{ebrahimi2019vertebral} is a semi-automatic method involving several sub-tasks. Our approach is automatic convolutional neural network (CNN) models. The CNN model of Wu et al.~\citep{wu2017automatic} requires cropped images and tedious data augmentation. Landmarks can also be detected by segmenting the relevant vertebrae \citep{imran2019end}. Our proposed model is totally end-to-end, requiring no pre-processing, and is fully automatic, eschewing any hand-crafted feature extractions.
\section{Method}
Given an X-ray image, we formulate the landmark detection problem as identifying $n$ landmarks localizing the relevant vertebrae. Each training image $x_i$, for $i=1,\dots, m$, is annotated by an associated $2n$-dimensional landmark vector $y_i$. Through supervised learning, a CNN can be trained to extract landmarks automatically, by minimizing the standard mean squared error (MSE) loss
\begin{equation}
\text{MSE} = \frac{1}{m}\sum_{i=1}^m (y_i - \hat{y}_i)^2,
\end{equation}
where $y_i$ are the ground-truth landmarks and $\hat y_i$ are the predicted landmarks. However, the MSE loss ignores inter-landmark relationships. To guide a CNN in the detection of landmark coordinates while learning spinal shape, we propose a novel distance measure---bipartite distance.
\begin{figure}
\centering
\begin{tabular}{cc}
\includegraphics[width=0.12\linewidth, trim={380 180 470 80}, clip]{bpd}
&
\includegraphics[width=0.79\linewidth, trim={0cm 19cm 3cm 2cm}, clip]{model}\\
(a) & (b)
\end{tabular}
\caption{Convolutional neural networks for landmark detection from spine X-ray images: (a) Illustration of the bipartite distance in a spinal image based on the ground truth (green) and predicted (red) landmarks. (b) Model architecture.}
\label{fig:model}
\end{figure}
Referring to Figure~\ref{fig:model}a, we regard the ground-truth (green) landmarks $A_y$ on the left and $B_y$ on the right of the spine as the two disjoint sets of vertices of a complete bipartite graph whose edges connect every landmark in $A_y$ with all landmarks in $B_y$. The same holds for the predicted (red) landmarks, $A_{\hat{y}}$ and $B_{\hat{y}}$. This leads to a shape-aware loss, which penalizes the CNN model when the pairwise distances between the predicted landmarks deviate from those between the ground truth landmarks. Letting $d^e$ denote the Euclidean distance between ground-truth landmarks connected by edge $e$ of the graph and $\hat d^e$ denote the Euclidean distance between the corresponding predicted landmarks, the bipartite distance (BPD) is
\begin{equation}
\label{eqn:bpd}
\text{BPD} = \frac{1}{m}\sum_{i=1}^{m} \sum_e\left|d_i^e - \hat d_i^e\right|.
\end{equation}
We employ the loss function
\begin{equation}
\label{eqn:loss}
L = \text{MSE} + \alpha\,\text{BPD},
\end{equation}
where $\alpha$ weighs the BPD term against the MSE.
\section{Implementation Details}
Our dataset consists of 100 high-resolution anterior-posterior spinal X-ray images with signs of mild to severe scoliosis. Since the cervical vertebrae are seldom involved in spinal deformity and the identification of the bottom cervical vertebra could be important, we selected 18 vertebrae: C7 (cervical), T1--T12 (thoracic), and L1--L5 (lumbar). Medical experts provided binary segmentation annotation by labeling the $n/4=18$ vertebrae in the X-ray images. The 4 corners of each vertebral region serve as landmarks. They were automatically extracted by applying FAST \citep{rosten2006machine} to the expert-segmented labels. Therefore, associated with each spinal image are 72 landmarks to be estimated.
As shown in Figure~\ref{fig:model}b, our model is a CNN comprising five convolutional layers and three fully-connected (FC) layers. Leaky-ReLU is used as the activation function in each layer. The convolutional layers have feature sizes 16, 32, 64, 128, and 256. In each layer, two $3\times 3$ convolution operations are followed by a $2\times 2$ maxpooling layer. After every convolutional layer, we use a batch-normalization layer and a dropout layer with the rate of 0.25. After two FC layers with 512 neurons, a final FC layer of $2n=144$ neurons is used to produce the image-plane coordinates of the landmarks. The model is implemented in Tensorflow with Python 3 and runs on a Tesla P40 GPU on a 64-bit Intel(R) Xeon(R) 440G CPU.
The dataset was split into training (80 images), testing (15 images), and validation (5 images) sets. All the images were resized to $1024\times 512\times 1$ and normalized to $[0, 1]$ before feeding them to the network. When the model is trained using our MSE-BPD loss, we used $\alpha=0.01$ in (\ref{eqn:loss}). As a baseline, we trained the same architecture using only the MSE loss; i.e., $\alpha=0$. The models were trained with a minibatch size of 4. We used the Adam optimizer with a learning rate of 0.0001 and momentum 0.9.
\section{Experimental Results}
We compared the performance of the proposed model (MSE-BPD) against the baseline (MSE) both qualitatively and quantitatively. Qualitative comparisons (e.g., Figure~\ref{fig:bpd}), show better agreement of our model over the baseline model. The irregular spinal shapes in the baseline model are mitigated by our model. Moreover, the landmark detection performance is also improved in our model, which achieves a correlation score of 0.95 compared to the baseline model's score of 0.92 (Pearson correlation coefficient). Moreover, one-way ANOVA analysis confirms that the landmarks predicted by our model have no significant difference with the ground truth landmarks ($\textit{p-value} <0.05$).
\begin{figure}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|ccc|ccc|}
{\Large Raw Image} & {\Large MSE} & {\Large MSE-BPD} & {\Large Raw Image} & {\Large MSE} & {\Large MSE-BPD}\\[10pt]
\includegraphics[width=0.3\linewidth]{raw_image1}
&
\includegraphics[width=0.2\linewidth, trim={12cm 4cm 11cm 4cm}, clip]{mse1}
&
\includegraphics[width=0.2\linewidth, trim={12cm 4cm 11cm 4cm}, clip]{bpd1}
&
\includegraphics[width=0.3\linewidth]{raw_image2}
&
\includegraphics[width=0.2\linewidth,trim={12cm 4cm 11cm 4cm}, clip]{mse2}
&
\includegraphics[width=0.2\linewidth,trim={12cm 4cm 11cm 4cm}, clip]{bpd2}
\\
\end{tabular} }
\caption{Qualitative comparison shows improved spinal shape and landmark detection performance by our model (MSE-BPD loss) relative the baseline model (MSE loss) in two spinal X-ray images from the test set. Green boxes bound vertebrae based on the ground-truth landmarks; red boxes bound vertebra based on the model-predicted landmarks.}
\label{fig:bpd}
\end{figure}
\section{Conclusions}
The detection of vertebral landmarks is crucial for the accurate measurement of scoliosis in spinal X-ray images. To this end, we proposed a new loss function which guides the training of a CNN vertebral (corner) landmark detection model to perform reliable shape-aware predictions.
\bibliographystyle{agsm}
| 185dabd1d6dce86397b998b11e341e9c6f2a063d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Abstract}
Perceptual illusions based on the spatial correspondence between objects and displayed images have been pursued by artists and scientists since the 15th century, mastering optics to create crucial techniques as the linear perspective and devices as the \textit{Magic Lantern}. Contemporary video projection mapping inherits and further extends this drive to produce perceptual illusions in space by incorporating the required real time capabilities for dynamically superposing the imaginary onto physical objects under fluid real world conditions. A critical milestone has been reached in the creation of the technical possibilities for all encompassing, untethered synthetic reality experiences available to the plain senses, where every surface may act as a screen and the relation to everyday objects is open to perceptual alterations.
\section*{Introduction}
The advancement and popularization of computer technologies over the past three decades has had a significant effect not only on the way images are created, but also on the techniques and technologies which enable their presentation. One of the most popular emerging practices is video projection mapping, also referred as video mapping or 3D mapping, a terminology which stands for a set of imaging techniques in service of the spatially coherent projection of bi dimensional images onto physical three dimensional objects~\cite{yun,naimark1}.
Despite its seemingly novelty in its use in the fields of communication, advertising and visual arts in the recent years, the current practice of video projection mapping is the result of several centuries of development and perfecting of a vast set of techniques derived from optics and the sciences of human perception, as well as mathematical operations for image transformations, and the evolution projection devices. Across all these periods in history, the production of perceptual illusions remains as the fundamental shared goal, of which important milestones are the invention of linear perspective techniques and its application in \textit{Trompe l'oeil} and \textit{Quadratura} installations during the Renassaince, the immersive theatrical qualities of \textit{Panorama} and \textit{Phantasmagoria} in the 18th century, the development of real time 3D computer rendering capabilities and the popularization of video projection mapping during the last part of the 20th century.
The present article analyzes the evolution of video projection mapping in relation to the history of perceptual illusion practices, considering common design methodologies, tools, and implemented techniques, in a transversal selection of scientific and artistic practices, spreading from static compositions on a canvas, to dynamic substitutions in three dimensional space. A commentary on the emerging perceptual and cognitive phenomena as a result of the spatial distribution of synthetic contents is provided, presenting the notion that the current stance of real time dynamic forms of video projection mapping constitutes a turning point in the quest for replacing the real world with synthetic versions that would offer unprecedented means for visual identity and meaning transformations beyond material constraints.
\section*{Spatial illusion in the Renassaince}
Before the 14th Century few attempts were made to create the illusion of three dimensional space and to realistically depict the spatial world in art. The Italian masters Giotto and Duccio during the late thirteenth and early fourteenth centuries, began to explore the idea of depth and volume in their artwork, and may be credited with introducing an early form of three dimensional painting, using shadows and contrasts in colors to create the perceived illusion of depth~\cite{giotto,schet}. Although it was still very far from the accurate and robust three dimensional representations popular in imagery today, this technique known as \textit{Chiaroscuro} helped artists to create the illusion of volume and depth. Popularized during the beginning of the fifteenth century, Chiaroscuro was mastered especially for the processes of modeling and representing the human body and was extensively used during the sixteenth and seventeenth centuries in the Mannerism and Baroque eras as one of the standards for approximating three dimensional representation.
Not satisfied with the state of the art, Florentine painters of the early fifteenth century, wanted to develop painting as a science derived from Euclidean geometry influenced by the Neoplatonic intellectual postulates. In this sense, they did not conform with the intuitive Chiaroscuro technique, but sought scientific methods of representing reality based on mathematical laws, thus giving birth to the knowledge that ultimately led to the devise of the linear perspective system. The linear perspective system projected the illusion of depth onto a two dimensional plane by use of vanishing points to which all lines converged at eye level on the horizon, accurately reproducing the way that objects appear smaller when they are further away, and the way parallel lines appear to meet each other at a point in the distance.
Between 1415 and 1420, the Italian architect Filippo Brunelleschi, considered one of the founding figures of the Renassaince, conducted and documented a series of experiments which helped him develop the correct perspective techniques for accurately representing three dimensional volumes on a flat canvas. One of its results portrayed in accurate spatial proportion and organization, the Florence Baptistery and the Palazzo Vecchio building seen from the front gate of the unfinished cathedral, affording to the spectator the full illusion of being present in the scene. These technical advancements where later systematized by Leon Battista Alberti in his 1436 work \textit{De Pictura}, one of the most important treatises on painting of the Renaissance, enabling artists for the first time to paint imaginary scenes with perfectly accurate three dimensional realism by following the set of rules of perspective studied by Brunelleschi~\cite{albert, alberti2,dunning, brock15}. Gradually, a standardized model of human visual perception was consolidated, from which a set of mathematical rules was derived and applied to govern the disposition of pictorial elements in order to produce a credible illusion of three dimensionality. \\
For the next five centuries, Brunelleschi’s system of perspective and its derived techniques were used to create the illusion of depth on the picture plane, and served as the basis of some of the great art of Western culture~\cite{gablik,gombrich, verstegen}. Among the earliest of those, The Brera Madonna created between 1472 and 1474 by the Italian artist Piero della Francesca. It is a large painting depicting a group of religious figures and icons at a fictitious apse of the church, meticulously rendered using the innovations of linear perspective techniques. But not only did perspective techniques allowed to produce a coherent illusion of space on a flat surface, but perspective was also used to organize and highlight particular elements of the composition, thus allocating meaning in relation to specific spatial positions in the canvas~\cite{banker, shw}. For example, the geometric center of the face of the main character coincides with the vanishing point of the perspective system used in the scene, as well as the egg which hangs from a shell - shaped element below the apse, is perfectly aligned with the scene's vanishing point and marks a perpendicular line to the scene's horizon.
Tommaso di Ser Giovanni di Simone, best known as \textit{Masaccio} and regarded as the first great Italian painter of the \textit{Quattrocento} period of the Italian Renaissance, painted \textit{The Holy Trinity} in 1425 in the dominican church of Santa Maria Novella in Florence. This pioneering mural painting not only accurately presented the illusion of three dimensions using linear perspective techniques, but also engaged in a coherent composition with the pre existing architectural elements that surrounded the mural, mixing three dimensional painted and architectural elements to enhance the illusory effect. This specific type of illusion supported by the coherent merger between painting techniques and architectural elements was known as Quadratura~\cite{masa}.
Artist Andrea Mantegna further experimented developing illusionistic ceiling paintings using perspective techniques to create the perceptual illusions and was responsible in 1465 for one of the most significant examples of Quadratura. Located in the Camera degli Sposi in the Ducal Palace of Mantua, this Trompe l'oeil was applied to ceiling paintings to be seen by spectators from below and upward and presented the illusionistic painting of an oculus that opens to the sky and a number of fictional characters who look down towards the spectators~\cite{wade}. Performed by Andrea Pozzo in 1688, \textit{The Apotheosis of St Ignatius} in Sant Ignazio in Rome is another remarkable example of Quadratura, presenting in a 55 feet wide ceiling a composition combining fictional characters and architectural elements. A new type of site specific art form began maturing, unifying architecture, painting and sculpture, and being conveyed to spectators by means of a full body spatial experience. In a sense, this almost theatrical experience was prescient of the aesthetic ideals of \textit{Gesamtkunstwerk}, postulated years later in mid nineteenth century by the German opera composer Richard Wagner.
\section*{Panorama and Phantasmagoria}
Baroque painters from seventeenth century, notably Caravaggio, Bernini, Rubens, Rembrandt, Velazquez and Vermeer, used the now long established three dimensional representation system in hundreds of varied approaches, as did the painters from the Neoclassicist movement, of whom a notable example is the work of Giovanni Batista Piranesi. \textit{Carceri d'invenzione}, a series prints produced between 1745 and 1761 by Piranesi depicting somber fictional scenes full of mysterious characters, bridges, stairs and mechanisms, trough meticulous utilization of the perspective technique. However, few artists during these decades shared the interest for the experimentation and development of perceptual illusions via spatial experience. In this sense, a major referent is the work of Robert Barker, an Irish painter that developed the idea of erecting a circular plan building to house near real scale paintings featuring panoramic landscapes. In 1792 the Panorama was born, supported on the knowledge on perspective and color contrast practiced during the last two centuries, its paintings covered 360 degrees around the spectator in the horizontal plane and reached 15 meter of height.
Soon, Barker's Panorama grew in size and complexity, and the circular plan installation incorporated topographies and architectural and elements intended to match the painted scene thus increasing its illusory effect in a similar way as Quadratura incorporated architectural elements to its composition. This way, flooring, stairs, railings, water fountains or other urban furniture, not only to regulate the displacements of spectators inside the building, but more importantly, to serve as immersive cues, formal and semantic links between the depicted scenes and the notion of what was accountable as real world experience~\cite{panorama, flieth}. The Panorama continued its popularization in most of Western Europe and North America as an entertainment form, successfully hosting the representation of famous historical battles for the great public. In a time when few people could afford to travel, the Panorama allowed to walk through distant landscapes and contemplate historical events from privileged points of view, pioneering in a primitive form of narrative distributed in space and time, in a sense similar as it is now possible to explore in a virtual reality experience or a video game~\cite{finn,finn2}.
During the precedent decades, the rapid increase of knowledge in optics was not only useful for development and use of perspective related techniques, but also supported a number of relevant inventions regarding optical instruments, among those, light reflection and projection devices. The Jesuit Athanasius Kircher in his book \textit{Ars Magna Lucis et Umbrae} from 1645, described a primitive projection system with a focusing lens and a concave mirror reflecting sunlight that served as canvas to display text or pictures known as the \textit{Steganographic Mirror}. Further elaborating, dutch scientist Christiaan Huygens in 1659 documented the invention of the \textit{Magic Lantern}, an early type of image projector that used semi transparent, hand painted glass plates, one or more lenses, and a light source~\cite{swan}. The Magic Lantern was then consolidated, and increasingly used for education and entertainment purposes during the 18th century~\cite{kember}. In this period, although the basic design of the Magic Lantern suffered various modifications in relation to the light source capabilities and the characteristics of lenses in use, its main fundamental concept remain unchanged as well as its basic operation.
By the end of 18th century, Belgian Physicist and cleric Ettiene Gaspar Robert, and one of the first lanternists of all time Paul Philidor, created a new form of illusion performance in the form of a Magic Lantern show known as Phantasmagoria, a theatrical show that consisted of a mixture of frontal and retro projection of images, from simultaneous magic lanterns, on canvases and solid elements in front of the sitting public~\cite{hecht, barber, phanta}. During the 19th century, the invention was further sophisticated, with more powerful light sources, incorporation of photographic impressions onto glass plates, improvements in optics, and mechanisms that allow the projection of sequences of images, helping consolidate the expressive maturity of a new type visual spectacle still greatly influenced by the use of perspective and contrast in colors to achieve volumetric realism~\cite{clarke}. Following this theatrical illusory tradition, Louis Daguerre in 1822 developed the \textit{Diorama} for a light based type of scene representation device, that featured two immense paintings and various interchangeable moving elements, such as human figures, animals or fictional characters, lit from the front and through the back inside an otherwise pitch black, rotating auditorium. The Diorama combined techniques of opaque and translucent painting and manipulating light in a live spectacle, with color contrast and perspective techniques, to produce a plausible scene with accurate spatial representation. In a large scale version, the Diorama frame could reach 10 meter wide and feature interposing translucent elements of different opacities and colors, actuated by means of ropes that affect the color tone of selective parts of the frame, for example, to simulate intense fog or the shining sun~\cite{Hankins}.
A few decades later, the German theater director and producer Erwin Piscator developed the most consistent scenography experiments using the technological possibilities of the early 20th century to support epic theater works emphasizing the socio political content of drama. Notably, his early production \textit{St\"urmflut} from 1926 the stage was composed by a translucent background screen in a solid black frame with variable aperture serving as background displaying supporting images from four retro projectors for performing artists.
\section*{Performatic spacetime}
Although most of the works of Impressionist painters such as Renoir, Sisley, Monet and Pissarro and Post Impressionists such as Van Gogh, Gauguin and Seurat also depended on these mimetic spatial representation techniques, a significant transformation took place in the broad scientific and artistic domains at the end of the nineteenth century. It challenged the absolute centralized perspective standard, in which the images orbited around the static spectator's eye, increasingly giving way to the development of the multiple viewpoints systems, which aimed to capture time lapses in which the scene should be considered not as static but in motion.
The idea of multiplicity in viewpoints was mainly developed in the arts by the Cubist movement, fundamentally Picasso and Braque in their invention of techniques such as faceting, passage and multiple perspective. This type of representations of multiple viewpoints in space and time, fused in one single image, consolidated as one of the movement's most acclaimed traits. Importantly, these innovations may be interpreted as the movements' critical philosophical stance regarding perspective and the capacity of human beings to faithfully grasp and represent reality. For the Cubists, art would not serve to be a mimesis of reality, as an objective mirror snapshot of \textit{what is there}, but would serve to represent things as experienced in space and time, a subjective snapshot of \textit{what is sensed}. This new method aimed to account for change and flux in a scene, and create a performative situation in which the interpretative role of the spectator becomes central to the evocative power of the Cubist's images~\cite{edgerton}.
Following these lines, Albert Gleizes and Jean Metzinger in the 1912 book \textit{Du "Cubisme"}, the first major text on Cubism, explicitly related the pursue of multiple perspective to an attempt to understand time as a continuum. This may be interpreted as giving symbolic expression to the notion of duration proposed by the philosopher Henri Bergson, according to which, life is subjectively experienced as a continuum, with the past flowing into the present and the present merging into the future~\cite{cubisme,bergson}. Early Futurist paintings by Umberto Boccioni, Giacomo Balla or Gino Severini where also highly influenced by this search for multiple perspective and simultaneity. This technique of representing multiple viewpoints in space and time is advanced in complexity and scale in Gleizes' \textit{Le Dépiquage des Moissons}, displayed at the 1912 \textit{Salon de la Section d'Or}, and Robert Delaunay's City of Paris, exhibited at the \textit{Indépendants} in 1912.
As the invention of linear perspective in fifteenth century Renaissance was a major manifestation of the establishment of the strong anthropocentric interpretation of the world, in which the whole of its Science and Philosophy was founded upon, the multiple perspective paradigm that arrived at the dawn of the 20th century marked the ultimate disenchantment of human as the measure of all things, welcoming novel ideas from Physics regarding the nature of space and time, as well as the atom and atomic dynamics, which ultimately favored many of the ideas that shaped the following decades: The relative over the absolute, multiplicity over centralization, dynamic co creation over pre determination~\cite{lepenies}. Whereas Renaissance artists strove to represent the real by offering viewers an archetypical perceptual illusion, Cubist illusionism subverts this kind of realistic representation by calling attention to its own artifice, to its own perspectival manipulations, and thus to the problematic nature of representation and illusion.
\section*{Expanded cinema}
During the 1950's The Czech artist Josef Svoboda pursued a type of non verbal theatrical format where actors perform in harmony along with moving images projected by a set of Magic Lanterns and elements of the scenography that supported the performative meaning production. This poetic synthesis of traditional methods with technical innovation including architecture, dance, cinema and theater, that in a sense revived Richard Wagner's \textit{Gesamtkunstwerk}, secured Czechoslovakia's best score in the Brussels World Exhibition of 1958. Svoboda not only mastered the classical perspective and color contrast techniques to put adequate spatial illusion in action on the stage, but also multiplied its evocative power by profiting from the emerging interactions between the characteristics of the selected physical elements, the performing actors and the content of the projected images in the composition~\cite{burian}. Moreover, part of this new theatrical format was Svoboda's \textit{Polyekran}, a multi projection system, consisting of screens of various shapes and dimensions suspended with steel cables in arbitrary angulations on a black background. Eight projectors of slides and seven movie projectors, along with synchronized stereo helped completing this instrument in service of the total work of art.\\
The immersive quality of spatially distributed images was also among the interests of the experimental filmmaker Stan VanDerBeek , who between 1963 and 1965 developed his \textit{Movie Drome}, a type of semi spherical theater where people would lie down and experience movies all around them. Notably, influenced by Buckminster Fuller’s dome constructions and Robert Barker's Panorama public installations, VanDerBeek envisioned the Movie Drome as a prototype for a global two way communication system consisting of a network of Movie Dromes linked to orbiting satellites that would store and transmit images between them. Although far from fully accomplishing it vision, Movie Dromes did put in place a new set of challenges for the traditional use of perspective and linear narratives to represent reality. Instead, VanDerBeek presented a multiple perspective design, that achieved multiple gazes over multiple simultaneous events, resulting in a fragmented non linear narrative, a spatially distributed moving image collage~\cite{sutton,ferguson}. Distinct from traditional cinema, this new type of experience sometimes characterized as \textit{Expanded Cinema}~\cite{youngblood} had a clear emphases on breaking the cinematography from its linear and concentrated constraints to give way to a non linear, distributed in space, multi sensory experience, aimed at a type of audience engagement that anticipates contemporary art's interactivity and participatory practices.
Audience engagement produced by carefully controlled perceptual illusions was also researched for popular entertainment purposes. As a proof, on 9 august 1969 the doors to the \textit{Haunted Mansion} spectacle at Disneyland opened to the public, featuring a number of light based perceptual illusions placed inside a three story Victorian house structure built in the New Orleans Square in Disney's Adventure Park in California. The project gathered the works of artists and storytellers Yale Gracey, Marc Davis, Claude Coats, Xavier Atencio and among those, Rolly Crump's innovative \textit{Singing Busts} installation~\cite{disney}. This work consisted of five singing busts, brougth to life by the \textit{Grim Grinning Ghosts} actors, pre recorded in 16mm film singing the theme song of the ride, and then projected onto the plaster busts modeled after each character. By controlling the lighting and contrast of the scene, together with the predetermined path and vantage point of user's, the designers could obtain an unprecedented multi sensory illusion that may be considered as direct precursor of current virtual and augmented reality paradigms.\\
Since the mid 1980s, numerous examples exist of re enactment of cinematographic scenes by the use of video sequences in combination with immersive objects and architectures. Among those the works of Tony Oursler who as one of the key figures in the development of video art experimented early on with the moving image that extended beyond the borders of the TV monitor~\cite{kaye}. Oursler used projections on sculptures and specially crafted scenographic installations. His early 1990s dolls and dummy works or \textit{Troubler} from 1996, rely entirely in the spatial and semantic correspondence between projected film and dolls which are strongly anthropomorphized~\cite{london}. Using centuries old knowledge on optics and perception, controlling light and shadow and the spectator's point of view Oursler mastered a unique type of cinematographic composition based on the combination of found objects, life-like moving faces and archetypal elements coming into life by virtue of the superimposed projected narratives~\cite{malt}.\\
Michael Naimark, is one key multimedia artist and researcher who has also explored spatialized representation. His \textit{Displacements} intallation, developed from 1980 to 1984, further advanced the spatial and semantic interactions between images and surfaces, and is considered among the most important seminal works in this domain. An immersive installation that projected a pre recorded 16mm film on an archetypal living room completely covered in white paint, matching the spatial location of the projected images with selected objects of the living room. The projector located on a rotating turntable at the room's center slowly directed the spectators' gaze towards the unwrapping narrative, with the appearance of the characters and objects as three dimensional.
Artist Jeffrey Shaw has also undertaken since the late 1960s, research on types of Expanded Cinema and the development of various innovative spatial illusion strategies. For example \textit{Corpocinema} from 1967, involved projection into a dome, and the {Diadrama} from 1974 was projected onto a 180 degree screen surrounding the audience.
In its theatrical form, an example of the combination of ancient perspective and contrast techniques with contemporary spatialized narratives to bring three dimensional realism to life is the opera \textit{Le Grande Macabre} by Fura dels Baus. Premiered in Brussels in March 2009, the company made use of the correspondence of projected images and physical objects, profiting from the emerging frictions between the projected images and a large scale anthropomorphic figure that mutates in identity and function within the play. The proposal of Fura dels Baus intends to avoid mimesis, and rather than repeat reality, aims at constructing a type of illusory live experience~\cite{fura}.
Several remarkable examples are found in the works of Robert Lepage, and among them, \textit{Needles and Opium}, which narratively relies on an strong engagement with the space, thus the need for a careful representation of location and context within a semi cubic structure that serves as a group of screens that mutate according to the needs of each scene~\cite{pluta}. The use of these types of neutral spatial cavities that serve as immersive screens has grown in popularity during the last decade. In 2012, as part of a marketing campaign for the Sony Playstation products, the agency Marshmallow Laser Feast presented the video mapping of an entire living room using choreographed live puppetry, controlling every element in the scene and performed in front of a non stop recording camera.
The resulting recorded performances offered a high degree of illusion while consolidating the archetypical centralized and predetermined viewpoint, in this case, the viewpoint of the non stop recording camera.
\subsection*{Interactive virtuality}
The advent and popularization of the interactive computing technologies, a variety of spatialized works of art speculate in the direction of producing full body perceptual illusions no longer tied to the spectator's spatial positions and viewpoints predetermined beforehand, but affording contents that interact with the user, whose three dimensional representations are dynamically adjusted to the view point of the user in real time~\cite{svoboda, szel, majum}.
From this point on, a significant part of research from academic and industry related fields, coincided in the aim for developing techniques and devices for visually compositing virtual objects within real environments, with ever increasing interactive faculties for spectators to become co participants. A significant advance was made in this regard by Carolina Cruz Neira, Daniel J. Sandin, and Thomas A. DeFanti at the University of Illinois, Chicago Electronic Visualization Laboratory in 1992. They presented the notion of \textit{CAVE} acronym for \textit{Cave automatic virtual environment}, a semi cubic installation made up of rear high resolution projection screen walls and floor, and a motion capture system that allowed users to interact by physical movements~\cite{cruz, cruz2,prince}. In 1995 a team of researchers from The University of North Carolina at Chapel Hill, including Ramesh Raskar, Greg Welch, and Henry Fuchs presented The office of the future a type of CAVE application not supported on retro projected surfaces but projection on real office surfaces and objects, moreover incorporated real time computer vision techniques to dynamically extract per pixel depth and reflectance information for the visible surfaces in the office including walls, furniture, objects, and people, and then to either project images on the surfaces, render images of the surfaces, or interpret changes in the surfaces~\cite{raskar, raskar3, wye}.
In 2006 Shaw initiated research into a fully 360 degree 3D projection system at the UNSW iCinema Research Centre, a centre directed by Shaw and co founded with Dennis Del Favero. AVIE \textit{Advanced Visualisation and Interaction Environment} was first installed in the UNSW Scientia building in 2003 and presented a paradigm that allowed the exploration of 360 degree panoramic projected images, within an architectural framework that correlated the design of the virtual landscape with that of the installation itself, thus simultaneously co activating the virtual representations and the real materials of the projection space.
This installation also introduced a fundamental level of user control, allowing the viewer to control the rotation of a projected window. The design of AVIE follows the traditions of the Panorama, as well as the search for embodied impact of the Baroque Tromp l'oeil fresco painting techniques ~\cite{shaw2, Shaw2003}. One remarkable implementation example is \textit{The Pure Land: Inside the Mogao Grottoes}, a project by Sarah Kenderdine and Jeffrey Shaw working in the University City Honk Kong in 2012. Using digitization data from the Mogao Grottoes in China this installation emphasizes the spatial illusion via panoramic immersion and coherent image display techniques~\cite{mogao}. The contents are staged in a 10 meter diameter by 4 meter high cylindrical AVIE theater, while a handheld interface provides interaction with the rendered images, allowing the user to volitionally reveal key parts of the mural paintings on its walls.
These real scale interactive visual display systems allowed the research on the domain of the spatial illusion of reality and helped set the foundations for projector based \textit{Augmented Reality} in the following decades. As defined, Augmented reality aims at building systems that enhances the real world by superimposing computer generated information on top of it, in one or more sensory modalities~\cite{wellner,asselin,billin, caudell}. In its projector based form, Augmented Reality shares the aim to generate perceptual illusion untethered and available to the plain senses, by using synthetic information displayed on real world locations, but distinguishing from mere spatialized motion image in the strong interactive component of the experience~\cite{bimber1, bimber2}.
\section*{Living virtuality}
In the last decades one of the most interesting objects for projector based synthetic augmentation has been the human body. In a collaboration between 1998 and 2002 choreographer Chris Haring and Vienna based director and composer Klaus Obermaier, created two multimedia dance performances, D.A.V.E and Vivisector, that have used the dancer's body as the primary projection surface pre choreographed projections on performers skins that led spectators to believe that the bodies of performers were dramatically reshaped and altered in real time~\cite{ ploe, emph,Dourish2001}. In 2004 together with Ars Electronica Futurelab, Obermaier developed and premiered Apparition, an interactive dance and media performance featuring Desirée Kongerød and Rob Tannion whose movements on stage are captured using an infrared camera and suited with adequate images generated in realtime.\\
At the 8th IEEE International Symposium on Mixed and Augmented Reality ISMAR in 2009, a team of researchers including Greg Welch, Henry Fuchs from The University of North Carolina at Chapel Hill, introduced Shader Lamps Avatars as a new approach for creating a video projected three dimensional avatar using faces of real people. Based on cameras and projectors to capture and map the dynamic motion and appearance of a real person onto a humanoid life sized styrofoam head mounted on an animatronic device, the system delivers a dynamic, real time representation of a person to multiple viewers~\cite{ismar1,ismar2}.
In a similar quest, a collaboration between Japanese media artist Nobumichi Asai, make up artist Hiroto Kuwahara and digital image engineer Paul Lacroix resulted in the Omote project in 2014, using projection mapping techniques to put make up and effects onto a model's face in real time. By using a set of infrared markers painted on the model's face, the team was able to capture and process marker position, estimating face position and orientation, rendering face CG model with animated texture upon, and finally send the image through the projector. This face changing system brings to life a fluid stance of identity in a way similar to the ancient Bian Lian techniques that allow Sichuan Opera's performers to unperceivably change masks in front of spectators.
Further advancing these achievements, The Ichikawa Senoo Laboratory from the University of Tokio aimed in 2015 at overcoming the limited static or quasi static conditions required for video projection mapping. A dynamic projection mapping system was developed, capable of adjusting images in real time onto deforming non rigid surfaces. To this end, the team led by Gaku Narita, Yoshihiro Watanabe, and Masatoshi Ishikawa developed two set of technologies addressing the system's main challenges~\cite{narita, narita2}. First, \textit{Deformable Dot Cluster Marker}, a robust infrared acquisition method for capturing deformations from the surface, even capable of handling large deformation and occlusions. Second, a high speed projector system was developed, \textit{DynaFlash}, capable of displaying 8 bit images in frame rates up to 1,000fps with unprecedented 3 milliseconds latencies. These two innovations mark a breakpoint in the quest for spatially distributed illusion, achieving the perceptual effect as if the projected images were printed onto the target surface. The Ichikawa Senoo Laboratory further demonstrated in 2016 their dynamic projection mapping systems actuating onto a deformed sheet of paper and T shirt on one or multiple targets. Considering the possibility for embedding active or passive markers in clothing and everyday objects~\cite{fibar}, these advancements could represent the novel technical feasibility for most surfaces in the real world to become screens, displaying perceptual illusions coherently and seamlessly~\cite{kott, jasch, borra, landi}.
\section*{Conclusions}
Video projection mapping has been an important step in the evolution of spatially based perceptual illusionism, and on the base of the evolution of its associated techniques and idiosyncrasies a great variety of research lines have been developed, as well as entertainment and artistic applications. Having mastered linear perspective, color contrast, and vantage point control in order to produce static realistic spatial three dimensional representations, in recent years video projection mapping systems paired with computer vision systems have acquired the capability of user interaction and of dynamically adjusting the spatial characteristics of the images in order to achieve almost perfect correspondence between projected contents and physical space. Considering video projection mapping as an intermediate stance in the reality - virtuality continuum~\cite{Milgram1994}, it has been shown to produce at least two types of observable results that occur almost simultaneously~\cite{Pfeiffer2013Multisensory,yun2011}. The first occurs at a sensory level, in which the shape, color and texture properties of the physical surfaces are incorporated by the superimposed images or wane in its benefit. For example the visitors of the Camera degli Sposi are deceived into believing that a two dimensional solid surface such as a brick wall disappears to give way to a different landscape in perfect harmony with the existing architecture. The face in the Omote's models can be suddenly full of make up, and the performers in Obermaier's works may seem to be suddenly in raw flesh.
The second observable phenomena concerns the semantic level of interaction between imaginary contents and physical objects. Depending on the semantic load of each of the interacting elements, the merger between superimposed images and material supports may yield a varying degree of identity transference, as such, physical objects may contrast with projected images or completely assume its identity and its network of semantic associations. In these semantic agreement and opposition relationships, the visitors of the Camera degli Sposi may fall under the impression of being at the residence of a powerful dynasty whose decorated oculus attracts angels from the skies. The models in Omote project can suddenly appear to be skeletons or android humanoids, or it may result unimaginable to think on the bodies in Obermaier's Vivisector as human due to their unprecedented flexibility.
Many steps have been taken in making the real world wane in favor of synthetic superpositions that increasingly become more present, in that every surface acts as a screen and every object contains potential of fluid identity and meaning hosted in the domain of virtual representations. As the production of perceptual illusions appear to be in full effect, new possibilities arise for new entertainment paradigms, spatially distributed learning applications, or massive dataset exploration tools that profit from the embodied cognition that results from spatial interaction with digital contents~\cite{Stahl2006,Stahl13}. Importantly, the relentless idea that fueled the quest for perceptual illusion and pervaded across centuries of artistic and scientific experimentation, seems to gain momentum at the current juncture in history: reality does not satisfy desire.
\section*{Acknowledgments}
Support provided Virtual Sense Systems and Hangar Centre for artistic research and production in Barcelona was greatly appreciated.
| 67cbf9f2750aa22cdb51f6c3674d31b8fb784064 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Newton and Wigner \cite{Newton1949} defined three physical criteria
that must be satisfied by the point-localized state vectors of a massive
or massless particle. The first is that a nonzero spatial translation
of a localized state vector must produce another localized state vector
that is orthogonal to the original. Then a rotation of a state vector
localized at the spatial origin at any time, $t,$ must produce another
state vector localized at the origin. Such a state vector can be labelled
by another quantum number that must carry an irreducible representation
of rotations, with state vectors for different values of this quantum
number being mutually orthogonal. Lastly, the scalar product of a
localized state vector and any boost of this state vector must be
a continuous function of the boost velocity.
It is not possible to satisfy all three criteria for the photon. The
reason is not the masslessness of the photon but because of its limited
helicity spectrum, as we will confirm below. This point was noted
by Wightman \cite{Wightman1962}. Within this limitation, it is possible
to define measures of partial (not point) localization for the photon
\cite{Hoffmann2020a}.
Given the impossibility result, it was surprising to learn that some
authors \cite{Hawton2019,Babaei2017,Hawton2017,Mostafazadeh2006}
claim to have constructed point-localized states for the photon. Perhaps
the result was not derived in suffficient detail, so we remedy that
situation in Section \ref{sec:Impossibility-of-point-localized}.
Then, in Section \ref{sec:Analysis-of-claimed}, we closely examine
some of the claims of point-localized photons and attempt to reproduce
their results. Conclusions follow in Section \ref{sec:Conclusions}.
\section{\label{sec:Impossibility-of-point-localized}Impossibility of point-localized
photon state vectors}
As stated above, the helicity spectrum of a massless particle determines
whether it can be point-localized. A hypothetical particle of zero
mass and zero helicity could be localized at a point, with state vectors
\begin{equation}
|\,x,0\,\rangle=\frac{1}{(2\pi)^{\frac{3}{2}}}\int\frac{d^{3}k}{\sqrt{\omega}}\,|\,k,0\,\rangle\,e^{ik\cdot x},\label{eq:1}
\end{equation}
where $\omega=|\boldsymbol{k}|$ and the covariant normalization of
the momentum-helicity eigenvectors (for general helicity, $\lambda$),
\begin{equation}
\langle\,k_{1},\lambda_{1}\,|\,k_{2},\lambda_{2}\,\rangle=\delta_{\lambda_{1}\lambda_{2}}\omega_{1}\delta^{3}(\boldsymbol{k}_{1}-\boldsymbol{k}_{2}),\label{eq:2}
\end{equation}
will be used throughout this paper. Also, in this paper, we use Heaviside-Lorentz
units, in which $\hbar=c=\epsilon_{0}=\mu_{0}=1.$ Only positive energies
are used in this superposition. These state vectors have the equal-time
scalar product
\begin{equation}
\langle\,(t,\boldsymbol{x}_{1}),0\,|\,(t,\boldsymbol{x}_{2}),0\,\rangle=\delta^{3}(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})\label{eq:3}
\end{equation}
and rotate and translate as expected for localized state vectors.
They boost continuously.
The photon has only two physical helicities, $\lambda=\pm1$. If there
were a $\lambda=0$ photon in addition to these two, we could construct
three state vectors (for $\sigma=-1,0,1$)
\begin{equation}
|\,x,\sigma,3\,\rangle=\frac{1}{(2\pi)^{\frac{3}{2}}}\int\frac{d^{3}k}{\sqrt{\omega}}\sum_{\lambda=-1,0,1}|\,k,\lambda\,\rangle\,\mathcal{R}_{\lambda\sigma}^{(1)-1}[\hat{\boldsymbol{k}}]\,e^{ik\cdot x}.\label{eq:4}
\end{equation}
Here $\mathcal{R}_{\lambda\mu}^{(1)-1}(\hat{\boldsymbol{k}})$ are
$j=1$ matrix elements of the inverse of the standard rotation
\begin{equation}
R_{0}[\hat{\boldsymbol{k}}]=R_{z}(\varphi)R_{y}(\theta)R_{z}(-\varphi)\label{eq:5}
\end{equation}
that takes $\hat{\boldsymbol{z}}$ into the momentum direction $\hat{\boldsymbol{k}}=(\theta,\varphi).$
The basis vectors rotate as
\begin{equation}
U(R)\,|\,k,\lambda\rangle=|\,R\,k,\lambda\,\rangle\,e^{-i\lambda w(R,\hat{\boldsymbol{k}})},\label{eq:6}
\end{equation}
with a Wigner rotation, a rotation about the momentum direction, $\hat{\boldsymbol{k}},$
specified by \cite{Hoffmann2020a}
\begin{equation}
R(w(R,\hat{\boldsymbol{k}})\,\hat{\boldsymbol{k}})=R_{0}^{-1}[R\hat{\boldsymbol{k}}]\,R\,R_{0}[\hat{\boldsymbol{k}}].\label{eq:7}
\end{equation}
Then we have (no sum over $\lambda$)
\begin{equation}
U(R)\,|\,k,\lambda\,\rangle\,\mathcal{R}_{\lambda\sigma}^{(1)-1}(\hat{\boldsymbol{k}})=\sum_{\sigma^{\prime}=-1}^{1}|\,R\,k,\lambda\,\rangle\,\mathcal{R}_{\lambda\sigma^{\prime}}^{(1)-1}[R\hat{\boldsymbol{k}}]\,\mathcal{D}_{\sigma^{\prime}\sigma}^{(1)}(R).\label{eq:8}
\end{equation}
This is an important result, as it shows that each helicity rotates
with the same transformation matrix.
This leads to the required rotation behaviour
\begin{equation}
U(R)\,|\,(t,\boldsymbol{x}),\sigma,3\,\rangle=\sum_{\mu^{\prime}=-1}^{1}|\,(t,R\,\boldsymbol{x}),\sigma^{\prime},3\,\rangle\,\mathcal{D}_{\sigma^{\prime}\sigma}^{(1)}(R),\label{eq:9}
\end{equation}
according to the unitary irreducible rotation representation with
angular momentum quantum number $j=1.$ It is clear that this construction
(Eq. (\ref{eq:4})) is the unique solution (up to phase changes dependent
on helicity) of the requirement in Eq. (\ref{eq:9}).
The unitary transformation
\begin{equation}
\langle\,\sigma\,|\,i\,\rangle=\sqrt{\frac{4\pi}{3}}\,Y_{1\sigma}^{*}(\hat{\boldsymbol{i}})\quad\mathrm{for}\ i=1,2,3,\label{eq:10}
\end{equation}
produces three state vectors for every $x,$
\begin{equation}
|\,x,i,3\,\rangle=\sum_{\sigma=-1}^{1}|\,x,\sigma,3\,\rangle\,\langle\,\sigma\,|\,i\,\rangle\quad\mathrm{for}\ i=1,2,3.\label{eq:11}
\end{equation}
We note that the coefficients in the superposition can be identified
as the components of complex conjugate polarization vectors in a particular
gauge,
\begin{equation}
\epsilon_{i}^{*}(\hat{\boldsymbol{k}},\lambda)=\sum_{\sigma=-1}^{1}\mathcal{R}_{\lambda\sigma}^{(1)-1}[\hat{\boldsymbol{k}}]\,\langle\,\sigma\,|\,i\,\rangle=\sum_{\sigma=-1}^{1}\sum_{j=1}^{3}\sum_{k=1}^{3}\langle\,\lambda\,|\,j\,\rangle\,R_{0jk}^{-1}[\hat{\boldsymbol{k}}]\,\langle\,k\,|\,\sigma\,\rangle\langle\,\sigma\,|\,i\,\rangle=\sum_{j=1}^{3}R_{0ij}[\hat{\boldsymbol{k}}]\,\langle\,\lambda\,|\,j\,\rangle.\label{eq:12}
\end{equation}
They are obtained by a rotation from their $\hat{\boldsymbol{k}}=\hat{\boldsymbol{z}}$
values
\begin{equation}
\boldsymbol{\epsilon}^{*}(\hat{\boldsymbol{z}},+1)=-\frac{1}{\sqrt{2}}\begin{pmatrix}1\\
-i\\
0
\end{pmatrix},\quad\epsilon^{*}(\hat{\boldsymbol{z}},0)=\begin{pmatrix}0\\
0\\
1
\end{pmatrix},\quad\epsilon^{*}(\hat{\boldsymbol{z}},-1)=+\frac{1}{\sqrt{2}}\begin{pmatrix}1\\
+i\\
0
\end{pmatrix}.\label{eq:13}
\end{equation}
So we can write
\begin{equation}
|\,x,i,3\,\rangle=\frac{1}{(2\pi)^{\frac{3}{2}}}\int\frac{d^{3}k}{\sqrt{\omega}}\sum_{\lambda=-1,0,1}|\,k,\lambda\,\rangle\,\epsilon_{i}^{*}(\hat{\boldsymbol{k}},\lambda)\,e^{ik\cdot x}.\label{eq:14}
\end{equation}
The three state vectors in Eq. (\ref{eq:14}) rotate according to
the three-vector representation of rotations
\begin{equation}
U(R)\,|\,(t,\boldsymbol{x}),i,3\,\rangle=\sum_{i^{\prime}=-1}^{1}|\,(t,R\,\boldsymbol{x}),i^{\prime},3\,\rangle\sum_{\sigma^{\prime},\sigma=-1}^{1}\langle\,i^{\prime}\,|\,\sigma^{\prime}\,\rangle\mathcal{D}_{\sigma^{\prime}\sigma}^{(1)}(R)\,\langle\,\sigma\,|\,i\,\rangle=\sum_{i^{\prime}=-1}^{1}|\,(t,R\,\boldsymbol{x}),i^{\prime},3\,\rangle\,R_{i^{\prime}i}.\label{eq:15}
\end{equation}
Both sets of state vectors translate correctly and satisfy the equal-time
orthonormality relations
\begin{equation}
\langle\,(t,\boldsymbol{x}_{1}),\sigma_{1},3\,|\,(t,\boldsymbol{x}_{2}),\sigma_{2},3\,\rangle=\frac{1}{(2\pi)^{3}}\int d^{3}k\,\sum_{\lambda=-1,0,1}\mathcal{R}_{\sigma_{1}\lambda}^{(1)}[\hat{\boldsymbol{k}}]\mathcal{R}_{\lambda\sigma_{2}}^{(1)-1}(\hat{\boldsymbol{k}})\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})}=\delta_{\sigma_{1}\sigma_{2}}\delta^{3}(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})\label{eq:16}
\end{equation}
and
\begin{equation}
\langle\,(t,\boldsymbol{x}_{1}),i_{1},3\,|\,(t,\boldsymbol{x}_{2}),i_{2},3\,\rangle=\delta_{i_{1}i_{2}}\delta^{3}(\boldsymbol{x}_{1}-\boldsymbol{x}_{2}),\label{eq:17}
\end{equation}
as required for sets of three point-localized state vectors.
Now, for the physical photon, we must remove the zero helicity in
the new definition
\begin{equation}
|\,x,\sigma,\gamma\,\rangle=\frac{1}{(2\pi)^{\frac{3}{2}}}\int\frac{d^{3}k}{\sqrt{\omega}}\sum_{\lambda=-1,1}|\,k,\lambda\,\rangle\,\mathcal{R}_{\lambda\sigma}^{(1)-1}(\hat{\boldsymbol{k}})\,e^{ik\cdot x}\quad\mathrm{for}\ \sigma=-1,0,1.\label{eq:18}
\end{equation}
We continue to use the $j=1$ rotation representation, as it is expected
to come closest to the desired result. From the result of Eq. (\ref{eq:8}),
we see that these three still rotate in the desired way. However,
the equal-time overlap becomes
\begin{equation}
\langle\,(t,\boldsymbol{x}_{1}),\sigma_{1},\gamma\,|\,(t,\boldsymbol{x}_{2}),\sigma_{2},\gamma\,\rangle=\frac{1}{(2\pi)^{3}}\int d^{3}k\,\sum_{\lambda=-1,1}\mathcal{R}_{\sigma_{1}\lambda}^{(1)}[\hat{\boldsymbol{k}}]\mathcal{R}_{\lambda\sigma_{2}}^{(1)-1}(\hat{\boldsymbol{k}})\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})}.\label{eq:19}
\end{equation}
The sum over helicities is
\begin{align}
\sum_{\lambda=-1,1}\mathcal{R}_{\sigma_{1}\lambda}^{(1)}[\hat{\boldsymbol{k}}]\mathcal{R}_{\lambda\sigma_{2}}^{(1)-1}(\hat{\boldsymbol{k}}) & =\delta_{\sigma_{1}\sigma_{2}}-\mathcal{R}_{\sigma_{1}0}^{(1)}[\hat{\boldsymbol{k}}]\mathcal{R}_{0\sigma_{2}}^{(1)-1}(\hat{\boldsymbol{k}})\nonumber \\
& =\delta_{\sigma_{1}\sigma_{2}}-\begin{pmatrix}\frac{1}{2}\sin^{2}\theta & -\frac{e^{-i\varphi}}{\sqrt{2}}\sin\theta\cos\theta & -\frac{1}{2}e^{-i2\varphi}\sin^{2}\theta\\
-\frac{e^{+i\varphi}}{\sqrt{2}}\sin\theta\cos\theta & \cos^{2}\theta & +\frac{e^{-i\varphi}}{\sqrt{2}}\sin\theta\cos\theta\\
-\frac{1}{2}e^{+i2\varphi}\sin^{2}\theta & +\frac{e^{+i\varphi}}{\sqrt{2}}\sin\theta\cos\theta & \frac{1}{2}\sin^{2}\theta
\end{pmatrix}_{\sigma_{1}\sigma_{2}}.\label{eq:20}
\end{align}
So
\begin{equation}
\langle\,(t,\boldsymbol{x}_{1}),\sigma_{1},\gamma\,|\,(t,\boldsymbol{x}_{2}),\sigma_{2},\gamma\,\rangle=\delta_{\sigma_{1}\sigma_{2}}\delta^{3}(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})-\frac{1}{(2\pi)^{3}}\int d^{3}k\,M_{\sigma_{1}\sigma_{2}}(\hat{\boldsymbol{k}})\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})},\label{eq:21}
\end{equation}
where $M_{\sigma_{1}\sigma_{2}}(\hat{\boldsymbol{k}})$ is the matrix
appearing in the second term of Eq. (\ref{eq:20}).
For the three-vector states
\begin{align}
\langle\,(t,\boldsymbol{x}_{1}),i_{1},\gamma\,|\,(t,\boldsymbol{x}_{2}),i_{2},\gamma\,\rangle & =\frac{1}{(2\pi)^{3}}\int d^{3}k\,(\delta_{i_{1}i_{2}}-\sum_{\lambda=-1,1}\langle\,i_{1}\,|\,\sigma_{1}\,\rangle\mathcal{R}_{\sigma_{1}0}^{(1)}[\hat{\boldsymbol{k}}]\mathcal{R}_{0\sigma_{2}}^{(1)-1}(\hat{\boldsymbol{k}})\langle\,\sigma_{2}\,|\,i_{2}\,\rangle)\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})}.\label{eq:22}
\end{align}
with summation over $\mu_{1}$ and $\mu_{2}$ implied. We find the
result
\begin{equation}
\langle\,(t,\boldsymbol{x}_{1}),i_{1},\gamma\,|\,(t,\boldsymbol{x}_{2}),i_{2},\gamma\,\rangle=\delta_{i_{1}i_{2}}\delta^{3}(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})-\frac{1}{(2\pi)^{3}}\int d^{3}k\,\hat{\boldsymbol{i}}_{1}\cdot\hat{\boldsymbol{k}}\hat{\boldsymbol{k}}\cdot\hat{\boldsymbol{i}}_{2}\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})},\label{eq:23}
\end{equation}
illustrating the impossibility of point-localized photon state vectors,
at least using the $j=1$ rotation representation.
Clearly using $j=0$ gives the wrong rotation properties for the state
vectors. Using $j\geq2,$ there will be unwanted terms inside the
momentum integral for the scalar products, given by
\begin{equation}
T_{\sigma_{1}\sigma_{2}}^{(j)}(\hat{\boldsymbol{k}})=-\sum_{\{\lambda\}}\mathcal{R}_{\sigma_{1}0}^{(j)}[\hat{\boldsymbol{k}}]\mathcal{R}_{0\sigma_{2}}^{(j)-1}(\hat{\boldsymbol{k}}),\label{eq:24}
\end{equation}
a sum over all helicities other than $\lambda=\pm1.$
\section{\label{sec:Analysis-of-claimed}Analysis of claimed point-localized
states}
Hawton \cite{Hawton2019} (her Eq. (34), summed over $\epsilon$
(the sign of the energy) and $\lambda$) constructs state vectors
(using our normalization conventions)
\begin{equation}
|\,A^{\mu}(x)\,\rangle=\frac{1}{(2\pi)^{\frac{3}{2}}}\,\int\frac{d^{3}k}{\omega}\sum_{\lambda=\pm1}|\,k,\lambda\,\rangle\,\epsilon^{*\mu}(\hat{\boldsymbol{k}},\lambda)\,e^{+ik\cdot x}.\label{eq:25}
\end{equation}
Here $\epsilon^{*\mu}(\hat{\boldsymbol{k}},\lambda)$ is now a four-component
polarization vector, with $\mu=0,1,2,3.$
Negative frequency contributions, $|\,A^{(-)\mu}(x)\,\rangle,$ with
the phase factor $\exp(-ik\cdot x),$ are also included, but they
don't translate correctly:
\begin{equation}
U(T(a))\,|\,A^{(-)\mu}(x)\,\rangle=|\,A^{(-)\mu}(x-a)\,\rangle\label{eq:26}
\end{equation}
instead of
\begin{equation}
U(T(a))\,|\,A^{(-)\mu}(x)\,\rangle=|\,A^{(-)\mu}(x+a)\,\rangle.\label{eq:27}
\end{equation}
It seems that these negative frequencies are not essential to their
result, so we omit them here.
The first issue is gauge variation. The four-component polarization
vectors appear in the electromagnetic field strength operators, $\hat{F}^{\mu\nu}(x),$
and in the four-component gauge field operator, $\hat{A}^{\mu}(x).$
A gauge transformation,
\begin{equation}
\epsilon^{\mu}(\hat{\boldsymbol{k}},\lambda)\rightarrow\epsilon^{\mu}(\hat{\boldsymbol{k}},\lambda)+g(|\boldsymbol{k}|)\,k^{\mu},\label{eq:28}
\end{equation}
for arbitrary $g,$ leaves the observable electromagnetic field strengths
unchanged, but changes the gauge field components. We note that such
a gauge transformation preserves the Lorentz condition,
\begin{equation}
k\cdot\epsilon(\hat{\boldsymbol{k}},\lambda)\equiv0,\label{eq:29}
\end{equation}
which is essential for the gauge fields to satisfy the Maxwell equations.
Unless a particular gauge is chosen, Eq. (\ref{eq:25}) does not give
a unique definition. We choose to consider the radiation gauge, with
polarization vectors $\epsilon_{R}^{\mu}(\hat{\boldsymbol{k}},\lambda),$
in which $\epsilon_{R}^{0}(\hat{\boldsymbol{k}},\lambda)\equiv0$
and the spatial parts are given by Eqs. (\ref{eq:12},\ref{eq:13}).
This is part of the Lorentz family of gauges since $\boldsymbol{k}\cdot\boldsymbol{\epsilon}_{R}(\hat{\boldsymbol{k}},\lambda)=\hat{\boldsymbol{z}}\cdot\boldsymbol{\epsilon}_{R}(\hat{\boldsymbol{z}},\lambda)\equiv0.$
Then we arrive at three state vectors for every $x,$
\begin{equation}
|\,x,i,\mathrm{H}\,\rangle=\frac{1}{(2\pi)^{\frac{3}{2}}}\,\int\frac{d^{3}k}{\omega}\sum_{\lambda=\pm1}|\,k,\lambda\,\rangle\,\epsilon_{R}^{*i}(\hat{\boldsymbol{k}},\lambda)\,e^{+ik\cdot x}\quad\mathrm{for}\ i=1,2,3.\label{eq:30}
\end{equation}
Since a different factor of the rotational invariant, $\omega,$ in
the superposition will not change the rotation properties, we see
that these state vectors rotate according to the irreducible representation.
Their equal-time quantum-mechanical scalar products are
\begin{multline*}
\langle\,(t,\boldsymbol{x}_{1}),i_{1},\mathrm{H}\,|\,(t,\boldsymbol{x}_{2}),i_{2},\mathrm{H}\,\rangle=\frac{1}{(2\pi)^{3}}\int\frac{d^{3}k}{\omega}\sum_{\lambda=\pm1}\epsilon_{R}^{i_{1}}(\hat{\boldsymbol{k}},\lambda)\epsilon_{R}^{*i_{2}}(\hat{\boldsymbol{k}},\lambda)\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})}\\
=\frac{1}{(2\pi)^{3}}\int\frac{d^{3}k}{\omega}(\delta_{i_{1}i_{2}}-\hat{\boldsymbol{i}}_{1}\cdot\hat{\boldsymbol{k}}\hat{\boldsymbol{k}}\cdot\hat{\boldsymbol{i}}_{2})\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})},
\end{multline*}
clearly not giving the required orthogonality result.
It is remarkable that these authors \cite{Hawton2019,Babaei2017,Hawton2017,Mostafazadeh2006}
define an \textit{alternative} scalar product, not equal to the quantum-mechanical
scalar product. If a state vector is defined as a superposition of
basis vectors with known scalar products, then the quantum-mechanical
scalar product of two such state vectors is completely defined and
not subject to arbitrary redefinition. The orthogonality properties
of the basis vectors, in this case, follow from the fact that they
are eigenvectors of Hermitian observables. State vectors can be defined
in different ways, but then their scalar products are fixed.
The alternative scalar product is defined by
\begin{equation}
(\,A_{1}(t,\boldsymbol{x}_{1}),A_{2}(t,\boldsymbol{x}_{2})\,)=\frac{1}{(2\pi)^{3}}\int\frac{d^{3}k}{\omega}\sum_{\lambda=\pm1}\sum_{i=1}^{3}\epsilon_{R}^{i}(\hat{\boldsymbol{k}},\lambda)\epsilon_{R}^{*i}(\hat{\boldsymbol{k}},\lambda)\,e^{i\boldsymbol{k}\cdot(\boldsymbol{x}_{1}-\boldsymbol{x}_{2})}.\label{eq:32}
\end{equation}
To verify the rotation criterion of Newton and Wigner, nine scalar
products for each $t,x_{1},x_{2}$ are needed, one for each pair $i_{1},i_{2}$.
But only one is given here, with the index $i$ summed over. If there
were no sum over $i,$ we would again have
\begin{equation}
\sum_{\lambda=\pm1}\epsilon_{R}^{i_{1}}(\hat{\boldsymbol{k}},\lambda)\epsilon_{R}^{*i_{2}}(\hat{\boldsymbol{k}},\lambda)=\delta_{i_{1}i_{2}}-\hat{\boldsymbol{i}}_{1}\cdot\hat{\boldsymbol{k}}\hat{\boldsymbol{k}}\cdot\hat{\boldsymbol{i}}_{2}.\label{eq:32.1}
\end{equation}
Using
\begin{equation}
\boldsymbol{\epsilon}_{R}(\hat{\boldsymbol{k}},\lambda)\cdot\boldsymbol{\epsilon}_{R}^{*}(\hat{\boldsymbol{k}},\lambda)=\boldsymbol{\epsilon}_{R}(\hat{\boldsymbol{z}},\lambda)\cdot\boldsymbol{\epsilon}_{R}^{*}(\hat{\boldsymbol{z}},\lambda)=1\label{eq:33}
\end{equation}
gives their result
\begin{equation}
(\,A_{1}(t,\boldsymbol{x}_{1}),A_{2}(t,\boldsymbol{x}_{2})\,)=2\,\delta^{3}(\boldsymbol{x}_{1}-\boldsymbol{x}_{2}).\label{eq:34}
\end{equation}
This result does not, as claimed, prove that localized photon state
vectors have been constructed.
\section{\label{sec:Conclusions}Conclusions}
We have confirmed that no point-localized state vectors can be constructed
for the photon that satisfy all three of the criteria of Newton and
Wigner \cite{Newton1949}. The reason is the limited helicity spectrum
of the photon. If there were a zero helicity state in addition to
$\lambda=\pm1,$ such localized states, three for every position and
time, could be constructed.
We considered the claims of other authors \cite{Hawton2019,Babaei2017,Hawton2017,Mostafazadeh2006}
that they could construct point-localized photon states satisfying
the three criteria of Newton and Wigner. It was found that they employ
an alternative definition of scalar product, not equal to the well-defined
quantum-mechanical one. It is only by using this alternative scalar
product that they are able to find a Dirac delta function in position.
But the argument is incomplete. The rotation of their state vectors
involves a subspace of dimension 3. Orthogonality of state vectors
with different values of the index would be required to complete the
proof, but no such result is given and would clearly not be possible.
We note that these authors also claim to have constructed a positive-definite
position probability density for the photon (the zero component of
a locally conserved four-current). If this were the case, it would
have to take the form
\[
\hat{\rho}(x)=\sum_{\sigma}|\,x,\sigma\,\rangle\langle\,x,\sigma\,|.
\]
The basis vectors would have to rotate according to an irreducible
representation and satisfy mutual orthogonality between different
values of the index, $\sigma$. Since we know this is not possible,
the claim is baseless.
\bibliographystyle{vancouver}
| 48ce722d340fff5a81328e45d05b604d6bb9d3c5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\begin{figure}[t]
\centering
\includegraphics*[scale=0.625, viewport=73 83 454 469]{figures/DWT-max_pooling.pdf}
\caption{Comparison of max-pooling, max-unpooling, and wavelet transforms.
Max-pooling is a common down-sampling operation in the deep networks,
which easily removes the data details (the window boundary in area A) and breaks the object structure (the pole in area B).
Based on the pooling indices, max-unpooling recovers the data resolution, but not the details.
DWT keeps the object structure and saves the data details,
while IDWT precisely recovers the original data using the DWT output.
}\label{fig_MaxPooling_DWT_IDWT}
\end{figure}
Due to the advantage of deep network in extracting high-level features,
deep learning has achieved high performances in various tasks, particular in the computer vision.
However, the current deep networks are not good at extracting and processing data details.
While deep networks with more layers are able to fit more functions,
the deeper networks are not always associated with better performances \cite{he2016deep_resnet}.
An important reason is that the details will be lost as the data flow through the layers.
In particular, the lost data details significantly degrade the performacnes of the deep networks for the image segmentation.
Various techniques, such as condition random field,
\`atrous convolution \cite{chen2014semantic_deeplabv1,chen2018encoder_deeplabv3+}, PointRend \cite{kirillov2019pointrend},
are introduced into the deep networks to improve the segmentation performance.
However, these techniques do not explicitly process the data details.
Wavelets \cite{daubechies1992ten,mallat1989a_wavelet}, well known as ``mathematical microscope'',
are effective time-frequency analysis tools,
which could be applied to decompose an image $\textbf{\emph{X}}$ into
the low-frequency component $\textbf{\emph{X}}_{ll}$ containing the image main information
and the high-frequency components $\textbf{\emph{X}}_{lh}, \textbf{\emph{X}}_{hl}, \textbf{\emph{X}}_{hh}$
containing the details (Fig. \ref{fig_MaxPooling_DWT_IDWT}).
In this paper, we rewrite Discrete Wavelet Transform (DWT) and Inverse DWT (IDWT) as the general network layers,
which are applicable to 1D/2D/3D data and various wavelets.
One can flexibly design end-to-end architectures using them,
and directly process the data details in the deep networks.
We design wavelet integrated deep networks for image segmentation, termed WaveSNets,
by replacing the down-sampling with DWT and the up-sampling with IDWT
in the U-Net \cite{ronneberger2015u_net}, SegNet \cite{badrinarayanan2017segnet}, and DeepLabv3+ \cite{chen2018encoder_deeplabv3+}.
When Haar, Cohen, and Daubechies wavelets are used,
we evaluate WaveSNets using dataset
CamVid \cite{brostow2009semantic_CamVid}, Pascal VOC \cite{everingham2015pascal}, and Cityscapes \cite{cordts2016cityscapes}.
The experimental results show that WaveSNets achieve better performances in semantic image segmentation than their vanilla versions,
due to the effective segmentation for the fine and similar objects.
In summary:
\begin{itemize}
\item We rewrite DWT and IDWT as general network layers, which are applicable to various wavelets
and can be applied to design end-to-end deep networks for processing the data details during the network inference;
\item We design WaveSNets using various network architectures,
by replacing the down-sampling with DWT layer and up-sampling with IDWT layer;
\item WaveSNets are evaluated on the dataset of CamVid, Pascal VOC, Cityscapes, and achieve better performance for semantic image segmentation.
\end{itemize}
\section{Related works}
\subsection{Sampling operation}
Down-sampling operations, such as max-pooling, average-pooling, and strided-convolution,
are introduced into the deep networks for local connectivity and weight sharing.
These down-sampling operations usually ignore the chassic sampling theorem \cite{nyquist1928certain},
which result in aliasing among the data components in different frequency intervals.
As a result, the data details presented by the high-frequency components are totally lost,
and random noises showing up in the same components could be sampled into the low resolution data.
In addition, the object basic structures presented by the low-frequency component will be broken.
Fig. \ref{fig_MaxPooling_DWT_IDWT} shows a max-pooling example.
In the signal processing, the low-pass filtering before the down-sampling is the standard method for anti-aliasing.
Anti-aliased CNNs \cite{zhang2019making} integrate the low-pass filtering with the down-sampling in the deep networks,
which achieve increased classification acccuracy and better shift-robustness.
However, the filters used in anti-aliased CNNs are empirically designed based on the row vectors of Pascal's triangle,
which are ad hoc and no theoretical justifications are given.
As no up-sampling operation, i.e., reconstruction, of the low-pass filter is available,
the anti-aliased U-Net \cite{zhang2019making} has to apply the same filtering after normal up-sampling to achieve the anti-aliasing effect.
Up-sampling operations, such as max-unpooling \cite{badrinarayanan2017segnet},
deconvolution \cite{ronneberger2015u_net}, and bilinear interpolation \cite{chen2014semantic_deeplabv1,chen2018encoder_deeplabv3+},
are widely used in the deep networks for image-to-image translation tasks.
These up-sampling operations are usually applied to gradually recover the data resolution,
while the data details can not be recovered from them.
Fig. \ref{fig_MaxPooling_DWT_IDWT} shows a max-unpooling example.
The lost data details would significantly degrade the network performance for the image-to-image tasks, such as the image segmentation.
Various techniques,
including \`atrous convolution \cite{chen2014semantic_deeplabv1,chen2018encoder_deeplabv3+}, PointRend \cite{kirillov2019pointrend}, etc.,
are introduced into the design of deep networks to capture the fine details for performance improvement of image segmentation.
However, these techniques try to recover the data details from the detail-unrelated information.
Their ability in the improvement of segmentation performance is limited.
\subsection{Wavelet}
Wavelets are powerful time-frequency analysis tools,
which have been widely used in signal analysis, image processing, and pattern recognition.
A wavelet is usually associated with scaling function and wavelet functions.
The shifts and expansions of these functions compose stable basis for the signal space,
with which the signal can be decomposed and reconstructed.
The functions are closely related to the low-pass and high-pass filters of the wavelet.
In practice, these filters are applied for the data decomposition and reconstruction.
As Fig. \ref{fig_MaxPooling_DWT_IDWT} shows,
2D Discrete Wavelet Transform (DWT) decompose the image $\textbf{\emph{X}}$ into its low-frequency component $\textbf{\emph{X}}_{ll}$
and three high-frequency components $\textbf{\emph{X}}_{lh}, \textbf{\emph{X}}_{hl}, \textbf{\emph{X}}_{hh}$.
While $\textbf{\emph{X}}_{ll}$ is a low resolution version of the image, keeping its main information,
$\textbf{\emph{X}}_{lh}, \textbf{\emph{X}}_{hl}, \textbf{\emph{X}}_{hh}$ save its horizontal, vertical, and diagonal details, respectively.
2D Inverse DWT (IDWT) could reconstruct the image using the DWT output.
Various wavelets,
including orthogonal wavelets, biorthogonal wavelets, multiwavelets, ridgelet, curvelet, bandelet and contourlet, etc.,
have been designed, studied, and applied in signal processing, numerical analysis, pattern recognition,
computer vision and quantum mechanics, etc.
The \`atrous convolution used in DeepLab \cite{chen2014semantic_deeplabv1,chen2018encoder_deeplabv3+}
is originally developed in the wavelet theory.
In the deep learning, while wavelets are widely applied as data preprocessing or postprocessing tools,
wavelet transforms are also introduced into the design of deep networks by taking them as substitutes of sampling operations.
Multi-level Wavelet CNN (MWCNN) \cite{liu2019multi_MWCNN}
integrates Wavelet Package Transform (WPT) into the deep network for image restoration.
MWCNN concatenates the low-frequency and high-frequency components of the input feature map, and processes them in a unified way.
The details stored in the high-frequency components would be largely wiped out via this processing mode,
because the data amplitude in the components is significantly weaker than that in the low-frequency component.
Convolutional-Wavelet Neural Network (CWNN) \cite{duan2017sar} applies the redundant dual-tree complex wavelet transform (DT-CWT)
to suppress the noise and keep the object structures for extracting robust features from SAR images.
The architecture of CWNN contains only two convolution layers.
While DT-CWT discards the high-frequency components output from DT-CWT,
CWNN takes as its down-sampling output the average value of the multiple low-frequency components.
Wavelet pooling proposed in \cite{williams2018wavelet} is designed using a two-level DWT.
Its back-propagation performs a one-level DWT with a two-level IDWT,
which does not follow the mathematical principle of gradient.
Recently, the application of wavelet transform in image style transfer \cite{yoo2019photorealistic} is studied.
In these works, the authors evaluate their methods using only one or two wavelets,
because of the absence of the general wavelet transform layers;
the data details presented by the high-frequency components are abandoned or processed together with the low-frequency component,
which limits the detail restoration in the image-to-image translation tasks.
\section{Our method}
Our method is to replace the sampling operations in the deep networks with wavelet transforms.
We firstly rewrite Discrete Wavelet Transform (DWT) and Inverse DWT (IDWT) as the general network layers.
Although the following analysis is mainly for orthogonal wavelet and 1D data,
it can be generalized to other wavelets and 2D/3D data with slight changes.
\subsection{Wavelet transform}
For a given 1D data $\textbf{\emph{x}} \in \mathbb{R}^n$,
DWT decomposes it, using two filters, i.e., low-pass filter $\textrm{f}_l$ and high-pass filter $\textrm{f}_h$ of an 1D orthogonal wavelet,
into its low-frequency component $\textbf{\emph{x}}_l$
and high-frequency component $\textbf{\emph{x}}_h$,
where
\begin{equation}\label{eq_DWT_1D}
\textbf{\emph{x}}_c = (\downarrow2)(\textrm{f}_c \ast \textbf{\emph{x}}), \quad c \in \{l,h\},
\end{equation}
and $\ast$, $(\downarrow2)$ denote the convolution and naive down-sampling, respectively.
In theory, the length of every component is $1/2$ of that of $\textbf{\emph{x}}$, i.e.,
\begin{equation}
\textbf{\emph{x}}_c \in \mathbb{R}^{\lfloor \frac{n}{2} \rfloor}, \quad c \in \{l,h\}.
\end{equation}
Therefore, $n$ is usually even number.
IDWT reconstructs the original data $\textbf{\emph{x}}$ based on the two components,
\begin{equation}\label{eq_IDWT_1D}
\textbf{\emph{x}} = \textrm{f}_l \ast (\uparrow2)\,\textbf{\emph{x}}_l + \textrm{f}_h \ast (\uparrow2)\,\textbf{\emph{x}}_h,
\end{equation}
where $(\uparrow2)$ denotes the naive up-sampling operation.
For high-dimensional data, high dimensional DWT could decomposes it into one low-frequency component and multiple high-frequency components,
while the corresponding IDWT could reconstructs the original data from the DWT output.
For example, for a given 2D data $\textbf{\emph{X}} \in \mathbb{R}^{m\times n}$,
2D DWT decomposes it into four components,
\begin{equation}\label{eq_DWT_2D}
\textbf{\emph{X}}_{c_0c_1} = (\downarrow2)(\textrm{f}_{c_0c_1} \ast \textbf{\emph{X}}), \quad c_0, c_1 \in \{l,h\},
\end{equation}
where $\textrm{f}_{ll}$ is the low-pass filter and $\textrm{f}_{lh}, \textrm{f}_{hl}, \textrm{f}_{hh}$ are the high-pass filters of the 2D orthogonal wavelet.
$\textbf{\emph{X}}_{ll}$ is the low-frequency component of the original data which is a low-resolution version containing the data main information;
$\textbf{\emph{X}}_{lh}, \textbf{\emph{X}}_{hl}, \textbf{\emph{X}}_{hh}$ are three high-frequency components which store the vertical, horizontal, and diagonal details of the data.
2D IDWT reconstruct the original data from these components,
\begin{equation}\label{eq_IDWT_2D}
\textbf{\emph{X}} = \sum_{c_0, c_1 \in \{l,h\}}\textrm{f}_{c_0c_1} \ast (\uparrow2)\,\textbf{\emph{X}}_{c_0c_1}.
\end{equation}
Similarly, the size of every component is $1/2$ size of the original 2D data in terms of the two dimensional direction, i.e.,
\begin{equation}
\textbf{\emph{X}}_{c_0c_1} \in \mathbb{R}^{\lfloor\frac{m}{2}\rfloor\times\lfloor\frac{n}{2}\rfloor},\quad c_0,c_1\in\{l,h\}.
\end{equation}
Therefore, $m,n$ are usually even numbers.
Generally, the filters of high-dimensional wavelet are tensor products of the two filters of 1D wavelet.
For 2D wavelet, the four filters could be designed from
\begin{equation}\label{eq_filter_2D}
\textrm{f}_{c_0c_1} = \textrm{f}_{c_1} \otimes \textrm{f}_{c_0}, \quad c_0, c_1 \in \{l,h\},
\end{equation}
where $\otimes$ is the tensor product operation.
For example, the low-pass and high-pass filters of 1D Haar wavelet are
\begin{equation}\label{eq_filter_2D}
\textrm{f}_l^{\textrm{H}} = \frac{1}{\sqrt{2}} (1,1)^T, \quad
\textrm{f}_h^{\textrm{H}} = \frac{1}{\sqrt{2}} (1,-1)^T.
\end{equation}
Then, the filters of the corresponding 2D Haar wavelet are
\begin{align}
\textrm{f}_{ll}^{\textrm{H}} &= \textrm{f}_l^{\textrm{H}} \otimes \textrm{f}_l^{\textrm{H}} =
\frac{1}{2}\left[
\begin{array}{cc}
1&1\\1&1
\end{array}\right]\\
\textrm{f}_{hl}^{\textrm{H}} &= \textrm{f}_l^{\textrm{H}} \otimes \textrm{f}_h^{\textrm{H}} =
\frac{1}{2}\left[
\begin{array}{cc}
1&-1\\1&-1
\end{array}\right]\\
\textrm{f}_{lh}^{\textrm{H}} &= \textrm{f}_h^{\textrm{H}} \otimes \textrm{f}_l^{\textrm{H}} =
\frac{1}{2}\left[
\begin{array}{cc}
1&1\\-1&-1
\end{array}\right]\\
\textrm{f}_{hh}^{\textrm{H}} &= \textrm{f}_h^{\textrm{H}} \otimes \textrm{f}_h^{\textrm{H}} =
\frac{1}{2}\left[
\begin{array}{cc}
1&-1\\-1&1
\end{array}\right]
\end{align}
Eqs. (\ref{eq_DWT_1D}), (\ref{eq_IDWT_1D}), (\ref{eq_DWT_2D}) and (\ref{eq_IDWT_2D}) present the forward propagations for 1D/2D DWT and IDWT.
It is onerous to deduce the gradient for the backward propagations from these equations.
Fortunately, the modern deep learning framework PyTorch \cite{paszke2017automatic} could automatically deduce the gradients for the common tensor operations.
We have rewrote 1D/2D DWT and IDWT as network layers in PyTorch,
which will be publicly available for other researchers.
In the layers, we do DWT and IDWT channel by channel for multi-channel data.
\subsection{WaveSNet}
We design wavelet integrated deep networks for image segmentation (WaveSNets),
by replacing the down-sampling operations with 2D DWT and the up-sampling operations with 2D IDWT.
In this paper, we take U-Net, SegNet, and DeepLabv3+ as the basic architectures.
\begin{figure}[bpt]
\centering
\subfigure[PDDS]{
\label{fig_dual_structure_Pooling_Deconv}
\includegraphics*[scale=0.525, viewport=72 685 285 767]{figures/dual_structures.pdf}
}\hspace{7pt}
\subfigure[PUDS]{
\label{fig_dual_structure_Pooling_Unpooling}
\includegraphics*[scale=0.525, viewport=72 479 285 561]{figures/dual_structures.pdf}
}
\caption{The dual structures used in U-Net and SegNet.}\label{fig_dual_structures}
\end{figure}
\begin{figure}[bpt]
\centering
\includegraphics*[scale=0.7, viewport=72 380 285 462]{figures/dual_structures.pdf}
\caption{The wavelet based dual structure (WADS).}\label{fig_dual_structures_wavelet}
\end{figure}
\textbf{WSegNets}\quad
SegNet and U-Net share a similar symmetrical encoder-decoder architecture, but differ in their sampling operations.
We name the pair of connected down-sampling and up-sampling operations and the associated convolutional blocks as \textbf{dual structure},
where the convolutional blocks process the feature maps with the same size.
Fig. \ref{fig_dual_structure_Pooling_Deconv} and Fig. \ref{fig_dual_structure_Pooling_Unpooling}
show the dual structures used in U-Net and SegNet,
which are named as PDDS (Pooling Deconvolution Dual Structure) and PUDS (Pooling-Unpooling Dual Structure), respectively.
U-Net and SegNet consist of multiple nested dual structures.
While they apply the max-pooling during their down-sampling,
PDDS and PUDS use deconvolution and max-unpooling for the up-sampling, respectively.
As Fig. \ref{fig_dual_structure_Pooling_Deconv} shows, PDDS copys and transmits the feature maps from encoder to decoder,
concatenating them with the up-sampled features and extracting detailed information for the object boundaries restoration.
However, the data tensor injected to the decoder might contain redundant information,
which interferes with the segmentation results and introduces more convolutional paramters.
PUDS transmits the pooling indices via the branch path for the upgrading of the feature map resolution in the decoder.
As Fig. \ref{fig_MaxPooling_DWT_IDWT} shows, the lost data details can not be restored from the pooling indices.
To overcome the weaknesses of PDDS and PUDS, we adopt DWT for down-sampling and IDWT for up-sampling,
and design WADS (WAvelet Dual Structure, Fig. \ref{fig_dual_structures_wavelet}).
During its down-samping, WADS decomposes the feature map into low-frequency and high-requency components.
Then, WADS injects the low-frequency component into the following layers in the deep networks to extract high-level features,
and transmits the high-frequency components to the up-sampling layer for the recovering of the feature map resolution using IDWT.
IDWT could also restore the data details from the high-frequency components during the up-sampling.
We design wavelet integrated encoder-decoder networks using WADS, termed WSegNets, for imag segmentation.
Table \ref{Tab_network_configuration} illustrates the configuration of WSegNets, together with that of U-Net and SegNet.
In this paper, the U-Net consists of eight more convolutional layers than the original one \cite{ronneberger2015u_net}.
\begin{table}[!t]
\scriptsize
\begin{center}
\setlength{\tabcolsep}{1.5mm}{
\begin{tabular}{c||l|r||c|c|c}\hline
\multirow{3}{*}{data size} & \multicolumn{2}{c||}{the number of channels} & \multicolumn{3}{c}{networks} \\ \cline{2-6}
& \multicolumn{1}{c|}{encoder} & \multicolumn{1}{c||}{decoder} &SegNet & U-Net & WSegNet \\ \hline
$512\times512$ & 3, 64 & 64, 64 & PUDS & PDDS & WADS \\
$256\times256$ & 64, 128 & 128, 64 & PUDS & PDDS & WADS \\
$128\times128$ & 128, 256, 256 & 256, 256, 128 & PUDS & PDDS & WADS \\
$64\times64$ & 256, 512, 512 & 512, 512, 256 & PUDS & PDDS & WADS \\
$32\times32$ & 512, 512, 512 & 512, 512, 512 & PUDS & PDDS & WADS \\ \hline
\end{tabular}}
\caption{Configurations of U-Net, SegNet, and WSegNet.} \label{Tab_network_configuration}
\end{center}
\end{table}
In Table \ref{Tab_network_configuration}, the first column shows the input size,
though these networks can process images with arbitrary size.
Every number in the table corresponds to a convolutional layer followed by a Batch Normalization (BN) and Rectified Linear Unit (ReLU).
While the number in the column ``encoder'' is the number of the input channels of the convolution,
the number in the column ``decoder'' is the number of the output channels.
The encoder of U-Net and SegNet consists of 13 convolutional layers
corresponding to the first 13 layers in the VGG16bn \cite{simonyan2014very_VGG}.
A convolutional layer with kernel size of $1\times1$ converts the decoder output into the predicted segmentation result.
\textbf{WDeepLabv3+}\quad
DeepLabv3+ \cite{chen2018encoder_deeplabv3+} employs an unbalanced encoder-decoder architecture.
The encoder applys an \`atrous convolutional version of CNN
to alleviate the detailed information loss due to the common down-sampling operations,
and an \`Atrous Spatial Pyramid Pooling (ASPP) to extract image multiscale representations.
At the begin of the decoder, the encoder feature map is directly upsampled using a bilinear interpolation with factor of 4,
and then concatenated with a low-level feature map transmitted from the encoder.
DeepLabv3+ adopts a dual structure connecting its encoder and decoder,
which differs with PDDS only on the up-sampling and suffers the same drawbacks.
\begin{figure}[bpt]
\centering
\includegraphics*[scale=0.475, viewport=27 473 534 713]{figures/WDeepLabv3+.pdf}
\caption{WDeepLabv3+ structure.}
\label{fig_WDeepLabv3+}
\end{figure}
We design WDeepLabv3+, a wavelet integrated version of DeepLabv3+,
by applying two wavelet version of dual structures.
As Fig. \ref{fig_WDeepLabv3+} shows, the encoder applys a wavelet integrated \`atrous CNN followed by the ASPP,
which output encoder feature map and two sets of high-frequency components.
The encoder feature map is up-sampled by two IDWT, integrated with the detail information contained in the high-frequency components,
while the two IDWT are connected with a convolutional block.
The decoder then apply a few $3\times3$ convolutions to refine the feature map,
followed by a bilinear interpolation to recover the resolution.
\section{Experiment}
We evaluate the WaveSNets (WSegNet, WDeepLabv3+) on the image dataset of
CamVid, Pascal VOC, and Cityscapes, in terms of mean intersection over union (mIoU).
\subsection{WSegNet}
\textbf{CamVid}\quad
CamVid contains 701 road scene images (367, 101 and 233 for the training, validation and test respectively) with size of $360\times480$,
which was used in \cite{badrinarayanan2017segnet} to quantify SegNet performance.
Using the training set, we train SegNet, U-Net, and WSegNet when various wavelets are applied.
The trainings are supervised by cross-entropy loss.
We employ SGD solver,
initial learning rate of $0.007$ with ``poly'' policy and $power = 0.9$, momentum of $0.9$ and weight decay of $0.0005$.
With a mini-batch size of $20$, we train the networks 12K iterations (about 654 epochs).
The input image size is $352 \times 480$.
For every image, we adopt random resizing between $0.75$ and $2$,
random rotation between $-20$ and $20$ degrees, random horizontal flipping and cropping with size of $352 \times 480$.
We apply a pre-trained VGG16bn model for the encoder initialization
and initialize the decoder using the technique proposed in \cite{he2015delving_initilization}.
We do not tune above hyper-parameters for a fair comparison.
\begin{table}[!tbp]
\scriptsize
\centering
\setlength{\tabcolsep}{0.35mm}{
\begin{tabular}{r|c|c|c|cccc|ccccc}\hline
&\multirow{2}{*}{SegNet}&\multirow{2}{*}{U-Net}&\multicolumn{10}{c}{WSegNet}\\\cline{4-13}
&&&haar&ch2.2&ch3.3&ch4.4&ch5.5&db2&db3&db4&db5&db6\\\hline
sky &90.50 &\textbf{91.52} &91.31 &91.38 &91.29 &91.39 &91.24 &91.35 &91.48 &90.99 &90.89 &90.63\\
building &77.80 &\textbf{80.28} &79.90 &79.27 &78.82 &79.37 &78.65 &79.48 &79.60 &78.52 &78.58 &77.84\\
pole &~~9.14 &27.97 &27.99 &\textbf{29.38} &27.90 &28.96 &27.26 &27.91 &28.38 &28.04 &26.66 &25.96\\
road &92.69 &93.71 &93.69 &93.47 &93.47 &93.77 &93.78 &93.72 &\textbf{93.91} &92.57 &92.12 &92.11\\
sidewalk &78.05 &80.05 &80.33 &79.44 &79.34 &79.89 &80.08 &79.67 &\textbf{80.58} &76.95 &75.62 &76.65\\
tree &72.40 &73.51 &73.34 &73.27 &73.21 &73.07 &71.60 &73.61 &\textbf{73.68} &72.90 &72.28 &71.92\\
symbol &37.61 &43.07 &44.44 &42.68 &40.42 &43.57 &42.33 &\textbf{44.72} &44.01 &41.06 &41.72 &39.69\\
fence &19.92 &27.50 &\textbf{32.59} &24.62 &25.59 &28.40 &28.85 &25.52 &29.60 &26.90 &24.15 &29.00\\
car &79.31 &\textbf{85.04} &83.21 &84.43 &82.63 &84.57 &84.14 &84.30 &83.97 &81.92 &81.07 &78.38\\
walker &39.93 &50.35 &49.35 &\textbf{50.98} &50.52 &50.43 &49.09 &50.15 &49.39 &47.69 &48.02 &43.17\\
bicyclist &39.48 &44.45 &50.38 &47.94 &48.69 &47.93 &\textbf{52.64} &51.15 &47.73 &46.08 &43.53 &38.96\\\cdashline{1-13}[3pt/1pt]
mIoU &57.89 &63.40 &\textbf{64.23} &63.35 &62.90 &63.76 &63.61 &63.78 &63.85 &62.15 &61.33 &60.39\\\hline
gl. acc. &91.04 &92.19 &92.08 &92.04 &91.98 &92.10 &91.73 &92.06 &\textbf{92.40} &91.60 &91.30 &91.14\\\hline
para.(e6) &29.52 &37.42&\multicolumn{10}{c}{\textbf{29.52}}\\\hline
\end{tabular}}
\caption{WSegNet results on CamVid test set}\label{tab_mIoU_CamVid_WSegNet}
\end{table}
Table \ref{tab_mIoU_CamVid_WSegNet} shows the mIoU and global accuracy on the CamVid test set,
together with the parameter numbers of U-Net, SegNet, and WSegNet with various wavelets.
In the table, ``dbx'' represents orthogonal Daubechies wavelet with approximation order $x$,
and ``chx.y'' represents biorthogonal Cohen wavelet with approximation orders $(x,y)$.
Their low-pass and high-pass filters can be found in \cite{daubechies1992ten}.
The length of these filters increase as the order increases.
While Haar and Cohen wavelets are symmetric, Daubechies are not.
The mIoU of WSegNet decreases from $63.78\%$ to $60.39\%$ as asymmetric wavelet varies from ``db2'' to ``db6'',
while it varies from $63.35\%$ to $63.61\%$ as symmetric wavelet varies from ``ch2.2'' to ``ch5.5''.
It seems that the performances among different asymmetric wavelets are much diverse than that among various symmetric wavelets.
In the wavelet integrated deep networks, we truncate the DWT output to make them to be $1/2$ size of input data.
As a result, IDWT with asymmetric wavelet can not fully restore an image in the region near the image boundary,
and the region width increases as the wavelet filter length increases.
With symmetric wavelet, however, one can fully restore the image based on symmetric extension of the DWT output.
Consequently, WSegNet with Cohen wavelets performs better than that with Daubechies ones near the image boundary.
Fig. \ref{Fig_CamVid_Comparison_Boundary} shows an example image, its manual annotation,
a region consisting of ``building'', ``road'' and ``sidewalk'' and the segmentation results achieved using different wavelets.
The region is located close to the image boundary and has been enlarged with colored segmentation results for better illustration.
One can observe from the results of ``db5'' and ``db6'' that a long line of ``sidewalk'' pixels,
located near the image boundary, are classified as ``road''.
In comparison, the results of ``ch4.4'' and ``ch5.5'' match well with the ground truth.
\begin{figure}[bpt]
\centering
\includegraphics*[scale=0.55, viewport=41 430 474 570]{figures/detail_comparison.pdf}
\caption{Boundary comparison for images segmented by WSegNet with different wavelets.}
\label{Fig_CamVid_Comparison_Boundary}
\end{figure}
In \cite{badrinarayanan2017segnet}, the authors train the SegNet using an extended CamVid training set containing 3433 images,
which achieved $60.10\%$ mIoU on the CamVid test set.
We train SegNet, U-Net and WSegNet using only 367 CamVid training images.
From Table \ref{tab_mIoU_CamVid_WSegNet}, one can find WSegNet achieves the best mIoU ($64.23\%$) using Haar wavelet,
and the best global accuracy ($92.40\%$) using ``db3'' wavelet.
WSegNet is significantly superior to SegNet in terms of mIoU and the global accuracy,
while they require the same amount of parameters ($29.52\times10^6$).
The segmentation performance of WSegNet is also better than that of U-Net,
while it requires much less parameters than U-Net ($37.42\times10^6$).
Table \ref{tab_mIoU_CamVid_WSegNet} also lists the IoUs for the 11 categories in the CamVid.
Compared with U-Net and WSegNet, SegNet performs very poor on the fine objects, such as ``pole'', ``symbol'', and ``fence'',
as that the max-pooling indices used in SegNet are not helpful for restoring the image details.
While U-Net achieves comparable or even better IoUs on the easily identifiable objects, such as the ``sky'', ``building'', and ``car'',
it does not discriminate well similar objects like``walker'' and ``bicyclist'', or ``building'' and ``fence''.
The feature maps of these two pairs of objects might look similar to the decoder of U-Net, as the data details are not separately provided.
Table \ref{tab_confusion_matrices} shows the confusion matrices on these four categories.
The proportion of ``building'' in the predicted ``fence''
decreases from $34.11\%$ to $30.16\%$, as the network varies from SegNet to WPUNet,
while that of ``bicyclist'' in the predicted ``walker'' decreases from $2.54\%$ to $1.90\%$.
These results suggest that WSegNet is more powerful than SegNet and U-Net in distinguishing similar objects.
\begin{table}
\scriptsize
\centering
\setlength{\tabcolsep}{1.25mm}{
\begin{tabular}{r|cc|cc|cc}
\hline
& \multicolumn{2}{|c|}{SegNet} &\multicolumn{2}{c|}{U-Net} &\multicolumn{2}{c}{WSegNet(haar)}\\\cline{2-7}
& building & fence & building & fence & building & fence\\ \hline
building & 88.01 & ~~1.30 & 89.91 & ~~1.19 & 90.22 & ~~1.08\\
fence & \underline{34.11} & 29.12 & \underline{30.64} & 40.16 & \underline{30.16} & 44.65\\ \hline\hline
& walker& bicyclist& walker& bicyclist& walker& bicyclist\\ \hline
walker & 53.30 & \underline{~~2.54} & 66.83 & \underline{~~1.95}& 68.74 & \underline{~~1.90}\\
bicyclist & 13.69 & 49.66 & 16.73 & 51.45 & 12.52 & 59.80 \\ \hline
\end{tabular}}
\caption{Confusion matrices for the CamVid test set.
\label{tab_confusion_matrices}
\end{table}
\begin{figure}[bpt]
\centering
\includegraphics*[scale=0.675, viewport=40 616 382 776]{figures/detail_comparison.pdf}
\caption{Comparison of SegNet and WSegNet segmentations.}
\label{fig_segmentation_comparison_CamVid}
\end{figure}
Fig. \ref{fig_segmentation_comparison_CamVid} presents a visual example for various networks,
which shows the example image, its manual annotation,
a region consisting of ``building'', ``tree'', ``sky'' and ``pole'', and the segmentation results achieved using SegNet, U-Net and WSegNet.
The region is enlarged with colored segmentation reults for better illustration.
From the figure, one can find in the segmentation result that WSegNet keeps the basic structure of ``tree'',
``pole'', and ``building'' and restores the object details, such as the ``tree'' branches and the ``pole''.
The segmentation result of WSegNet matches the image region much better than that of SegNet and U-Net,
even corrects the annotation noises about ``building'' and “tree” in the ground truth.
\begin{table}[!tbp]
\scriptsize
\centering
\setlength{\tabcolsep}{0.75mm}{
\begin{tabular}{cr|c|c|c|cccc|c}\hline
&&\multirow{2}{*}{SegNet}&\multirow{2}{*}{U-Net}&\multicolumn{6}{c}{WSegNet}\\\cline{5-10}
& & & & haar& ch2.2& ch3.3& ch4.4& ch5.5& db2\\\hline
\multirow{2}{*}{Pascal VOC}&mIoU &61.33 &63.64 &\textbf{63.95} &63.50 &63.46 &63.48 &63.62 &63.48\\\cdashline{2-10}[1pt/1pt]
&gl. acc.&90.14 &\textbf{90.82} &90.79 &90.72 &90.72 &90.73 &90.76 &90.75\\\hline
\multirow{2}{*}{Cityscapes}&mIoU &65.75 &70.05 &70.09 &69.86 &69.73 &70.13 &\textbf{70.63} &70.37 \\\cdashline{2-10}[1pt/1pt]
&gl. acc.&94.65 &\textbf{95.21} &95.20 &95.18 &95.09 &95.17 &95.15 &95.20 \\\hline
\multicolumn{2}{r|}{parameters (e6)}& 29.52&37.42 &\multicolumn{6}{c}{\textbf{29.52}}\\\hline
\end{tabular}}
\caption{WSegNet results on Pascal VOC and Cityscapes \textit{val} set.}\label{tab_mIoU_VOC_Cityscapes_WSegNet}
\end{table}
\textbf{Pascal VOC and Cityscapes}\quad
The original Pascal VOC2012 semantic segmentation dataset contains 1464 and 1449 annotated images
for training and validation, respectively, and contains 20 object categories and one background class.
The images are with various sizes.
We augment the training data to 10582 by extra annotations provided in \cite{hariharan2011semantic}.
We train SegNet, U-Net, and WSegNet with various wavelets on the extended training set 50 epochs with batch size of 16.
During the training, we adopt random horizontal flipping, random Gaussian blurring, and cropping with size of $512\times512$.
The other hyper-parameters are the same with that used in CamVid training.
Table \ref{tab_mIoU_VOC_Cityscapes_WSegNet} presents the results on the validation set for the trained networks.
Cityscapes contains 2975 and 500 high quality annotated images for the training and validation, respectively.
The images are with size of $1024\times2048$.
We train the networks on the training set 80 epochs with batch size of 8 and initial learning rate of $0.01$.
During the training, we adopt random horizontal flipping, random resizing between 0.5 and 2, and random cropping with size of $768\times768$.
Table \ref{tab_mIoU_VOC_Cityscapes_WSegNet} presents the results on the validation set.
From Tabel \ref{tab_mIoU_VOC_Cityscapes_WSegNet}, one can find the segmentation performance of SegNet
is significant inferior to that of U-Net and WSegNet.
While WSegNet achieves better mIoU ($63.95\%$ for Pascal VOC and $70.63\%$ for Cityscapes)
and requires less parameters ($29.52\times10^6$),
the global accuracies of WSegNet on the two dataset are a little lower than that of U-Net.
This result suggest that, compared with U-Net, WSegNet could more precisely classify the pixels at the ``critical'' locations.
\subsection{WDeepLabv3+}
\begin{table}[!tbp]
\scriptsize
\centering
\setlength{\tabcolsep}{1.25mm}{
\begin{tabular}{r|c|c|cccc|c}\hline
&\multirow{2}{*}{DeepLabv3+}&\multicolumn{6}{c}{WDeepLabv3+}\\\cline{3-8}
& & haar& ch2.2& ch3.3& ch4.4& ch5.5& db2\\\hline
background &93.83 &93.82 &93.85 &\textbf{93.94} &93.91 &93.86 &93.87 \\
aeroplane &92.29 &93.14 &92.50 &91.56 &\textbf{93.21} &92.73 &92.41 \\
bike &41.42 &43.08 &42.19 &42.21 &\textbf{43.42} &42.59 &42.84 \\
bird &91.47 &90.47 &\textbf{91.60} &90.34 &91.24 &90.81 &90.73 \\
boat &75.39 &\textbf{75.47} &72.04 &75.19 &74.20 &72.40 &72.90 \\
bottle &82.05 &80.18 &81.14 &82.12 &79.55 &\textbf{82.23} &81.89 \\
bus &\textbf{93.64} &93.25 &93.25 &93.20 &93.07 &93.52 &93.31 \\
car &89.30 &90.36 &90.00 &\textbf{90.67} &88.79 &86.31 &87.11 \\
cat &93.69 &92.80 &93.31 &92.62 &93.56 &93.84 &\textbf{93.97} \\
\underline{chair} &38.28 &40.80 &40.75 &39.32 &39.79 &\textbf{43.27} &41.60 \\
\textbf{\emph{cow}} &86.60 &89.72 &89.04 &90.49 &88.42 &\textbf{92.04} &88.17 \\
\underline{table} &61.37 &63.24 &62.67 &65.49 &64.93 &\textbf{67.31} &65.58 \\
dog &\textbf{91.16} &90.24 &91.04 &89.54 &89.97 &90.38 &90.65 \\
\textbf{\emph{horse}} &86.60 &88.86 &88.88 &90.23 &89.00 &\textbf{91.02} &89.19 \\
motorbike &88.47 &87.94 &87.89 &\textbf{88.61} &88.36 &87.30 &87.58 \\
person &86.71 &86.88 &86.61 &86.35 &86.59 &86.32 &\textbf{86.96} \\
\underline{plant} &64.48 &64.33 &64.69 &\textbf{68.45} &65.50 &68.01 &65.41 \\
\textbf{\emph{sheep}} &83.04 &87.45 &86.50 &87.43 &85.70 &\textbf{88.14} &85.95 \\
sofa &49.24 &47.43 &48.06 &49.85 &46.74 &\textbf{51.94} &50.19 \\
train &85.49 &84.18 &83.88 &85.47 &83.88 &85.16 &\textbf{86.69} \\
monitor &77.80 &76.55 &78.31 &79.04 &78.32 &78.78 &\textbf{79.11} \\\cdashline{1-8}[3pt/1pt]
mIoU &78.68 &79.06 &78.96 &79.62 &78.96 &\textbf{79.90} &79.34 \\\hline
gl. acc. &94.64 &94.69 &94.68 &94.75 &94.68 &\textbf{94.77} &94.75 \\\hline
para. (e6) &\textbf{59.34} &\multicolumn{6}{c}{60.22} \\\hline
\end{tabular}}
\caption{WDeepLabv3+ results on Pascal VOC \textit{val} set.}\label{tab_mIoU_VOC_WDeepLabv3+}
\end{table}
Taking ResNet101 as the backbone, we build DeepLabv3+ and WDeepLabv3+ with $output\ stride = 16$,
and train them on the Pascal VOC using the same training policy with that used in the training of WSegNet.
Table \ref{tab_mIoU_VOC_WDeepLabv3+} shows the segmentation results on the validation set.
We achieve $78.68\%$ mIoU using DeepLabv3+ on the Pascal VOC validation set,
which is comparable to that ($78.85\%$) obtained by the inventors of this network \cite{chen2018encoder_deeplabv3+}.
With a few increase of parameter ($0.88 \times 10^6,~1.48\%$),
the segmentation performance of WDeepLabv3+ with various wavelet is always better than that of DeepLabv3+,
in terms of mIoU and global accuracy.
Using ``ch5.5'' wavelet, WDeepLabv3+ achieves the best performance, $79.90\%$ mIoU and $94.77\%$ global accuracy.
From Table \ref{tab_mIoU_VOC_WDeepLabv3+}, one can find that the better performance of WDeepLabv3+
is mainly resulted from its better segmentation
for the fine objects (``chair'', ``table'', and ``plant'') and similar objects (``cow'', ``horse'', and ``sheep'').
The above results justify the high efficiency of DWT/IDWT layers in processing data details.
\begin{figure}[bpt]
\centering
\includegraphics*[scale=0.415, viewport=22 260 581 810]{figures/segmentation_comparison.pdf}
\caption{Comparison of DeepLabv3+ and WDeepLabv3+ results.}
\label{fig_comparison_WDeepLabv3+_VOC}
\end{figure}
Fig. \ref{fig_comparison_WDeepLabv3+_VOC} shows four visual examples of segmentation results for DeepLabv3+ and WDeepLabv3+.
The first and second columns present the original images and the segmentation ground truth,
while the third and fourth columns show the segmentation results of DeepLabv3+ and WDeepLabv3+ with ``ch5.5'' wavelet, respectively.
We show the segmentation results with colored pictures for better illustration.
For the example image in Fig. \ref{fig_comparison_WDeepLabv3+_VOC}(a),
DeepLabv3+ falsely classifies the pixels in some detail regions on the cow and her calfs as ``background'' or ``horse'',
i.e., the regions for the cow's left horn, the hind leg of the brown calf, and some fine regions on the two calfs.
While WDeepLabv3+ correctly classifies the horse and the sheep
in the Fig. \ref{fig_comparison_WDeepLabv3+_VOC}(b) and Fig. \ref{fig_comparison_WDeepLabv3+_VOC}(c),
DeepLabv3+ classifies them as ``cow'' because of the similar object structures.
In Fig. \ref{fig_comparison_WDeepLabv3+_VOC}(d), the ``table'' segmented by WDeepLabv3+ is more complete than that segmented by DeepLabv3+.
These results illustrate that WDeepLabv3+ performs better on the detail regions and the similar objects.
\section{Conclusion}
Our proposed general DWT and IDWT layers are applicable to various wavelets,
which can be used to extract and process the data details during the network inference.
We design WaveSNets (WSegNet and WDeepLabv3+) by replacing the down-sampling with DWT
and replacing the up-sampling with IDWT, in U-Net, SegNet, and DeepLabv3+.
Experimental results on the CamVid, Pascal VOC, and Cityscapes show that
WaveSNets could well recover the image details and perform better for segmenting similar objects than their vanilla versions.
\bibliographystyle{named}
| 245bce748aa00023b90df25be7736348be751d74 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Appendix}
\subsection{Bounding the Error of Monte Carlo Estimation} \label{apx:MC}
As shown in Proposition~\ref{pro:cAcB},
the smoothed model $f^\rs$ is certified robust at an input $\X$ in the sense of \eqref{equ:certified}
if
\begin{align*}
\Delta_\X
& = {\ell^\rs}(\X, y)
-{\ell^\rs}(\X, y_B)
-2 q_\X \\
& = {\ell^\rs}(\X, y)
- \max_{c\ne y
{\ell^\rs}(\X, c)
-2q_\X > 0,
\end{align*}
where $y$ is the true label of $\X$,
and
$$
{\ell^\rs}(\X, c) := \mathbb{P}_{\Z\sim\Pi_{\X}}\left(f(\Z)=c\right)
=
\mathbb{E}_{\Z\sim\Pi_{\X}}\left[\mathbb I\{ f(\Z)=c\}\right].
$$
Assume $\{\Z^{(i)}\}_{i=1}^n$ is an i.i.d. sample
from $\Pi_{\X}$.
By Monte Carlo approximation, we can estimate ${\ell^\rs}(\X, c)$ for all $c\in\mathcal{Y}$ jointly, via
$$
{\hat{\ell}^\rs}(\X, c ):=\frac{1}{n}\sum_{i=1}^{n}\mathbb{I}\left\{ f(\Z^{(i)})=c\right\},$$
and estimate $\Delta_\X$ via
$$
\hat \Delta_\X :=
\frac{1}{n}\sum_{i=1}^{n}\mathbb{I}\left\{ f(\Z^{(i)})=y\right\} -
\max_{c\ne y}\frac{1}{n}\sum_{i=1}^{n}\mathbb{I}\left\{ f(\Z^{(i)})=c\right\}
- 2 q_\X.
$$
To develop a rigorous procedure for testing $\Delta_\X >0$,
we need to bound the non-asymptotic error of the Monte
Carlo estimation, which can be done with a simple application of Hoeffding's concentration inequality and union bound.
\begin{proposition}\label{pro:concentration}
Assume $\{\Z^{(i)}\}$ is i.i.d. drawn from $\Pi_\X$.
For any $\delta\in (0,1)$, with probability at least $1-\delta$,
we have
\begin{align*}
\Delta_\X \ge &
\hat\Delta_\X - 2\sqrt{\frac{\log\frac{1}{\delta}+\log\left|\mathcal{Y}\right|}{2n}}.
\end{align*}
\end{proposition}
We can now frame the robustness certification problem into a hypothesis test problem.
Consider the null hypothesis $\text{H}_0$ and alternatively hypothesis $\text{H}_a$:
\begin{align*}
\text{H}_{0}: & \Delta_\X \leq 0~~~ (f^{\rs}\text{\ is\ not\ certified\ robust\ to\ }\X)\\
\text{H}_{a}: & \Delta_\X > 0~~~ (f^{\rs}\text{\ is\ certified\ robust\ to\ }\X).
\end{align*}
Then according to Proposition~\ref{pro:concentration}, we can reject the null hypothesis $\text{H}_0$ with a significance level $\delta$ if
\[
\hat\Delta_\X - 2\sqrt{\frac{\log\frac{1}{\delta}+\log\left|\mathcal{Y}\right|}{2n}} > 0.
\]
In all the experiments, we set $\delta = 0.01$ and $n=5000$.
\subsection{Proof of the Main Theorems} \label{apx:pf}
In this section, we give the proofs of the theorems in the main text.
\newcommand{{\X'\in S_\X}}{{\X'\in S_\X}}
\renewcommand{\H}{\mathcal{H}_{[0,1]}}
\subsubsection{Proof of Proposition \ref{pro:cAcB}}
According to the definition of $f^\rs$, it is certified robust at $\X$, that is, $y = f^\rs(\X')$ for $\forall \X' \in S_\X$, if
\begin{align} \label{equ:grddfdi}
{\ell^\rs}(\X', ~ y) \geq \max_{c\neq y}{\ell^\rs}(\X', c),~~~~~ \X' \in S_\X.
\end{align}
Obviously
\begin{align*}
{\ell^\rs}(\X', ~ y) - \max_{c\neq y}{\ell^\rs}(\X', c)
& \geq \min_{\X'\in S_\X}{\ell^\rs}(\X', ~ y) - \max_{c\neq y} \max_{\X' \in S_\X}{\ell^\rs}(\X', c) \\
& \geq \left ({\ell^\rs}(\X, ~ y) - q_\X \right) - \max_{c\neq y} \left ({\ell^\rs}(\X, c) + q_\X \right)
~~~\text{//by Theorem~\ref{thm:lb}.} \\
& = \Delta_\X.
\end{align*}
Therefore, $\Delta_\X > 0$ must imply \eqref{equ:grddfdi} and hence certified robustness.
\subsubsection{Proof of Theorem \ref{thm:lb}}
Our goal is to calculate the upper and lower bounds
$\max_{\X'\sim \Pi_\X}{\ell^\rs}(\X', c)$ and
$\min_{\X'\sim \Pi_\X}{\ell^\rs}(\X', c)$.
Our key idea is to frame the computation of the upper and lower bounds into a variational optimization.
\begin{lemma}\label{lem:first}
Define $\H$ to be the set of all bounded functions mapping from $\mathcal{X}$ to $[0,1]$, For any $h \in \H$, define
\[
\Pi_{\X}[h]=\mathbb{E}_{\Z\sim\Pi_{\X}}[h(\Z)].
\]
Then we have for any $\X$ and $c \in \mathcal Y$,
\begin{align*}
& \min_{\X'\sim \Pi_\X}{\ell^\rs}(\X', c) \geq \min_{h \in \H }\min_{\X'\sim \Pi_\X} \left \{ \Pi_{\X'}[h] ~~~s.t.~~~ \Pi_{\X}[h] = {\ell^\rs}(\X, c) \right\} : = g^{\rs}_{low}(\X, c),
\\
& \max_{\X'\sim \Pi_\X}{\ell^\rs}(\X', c) \leq \max_{h \in \H }\max_{\X'\sim \Pi_\X} \left \{ \Pi_{\X'}[h] ~~~s.t.~~~ \Pi_{\X}[h] = {\ell^\rs}(\X, c) \right\} : = g^{\rs}_{up}(\X, c).
\end{align*}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lem:first}]
The proof is straightforward. Define $h_0(\X)= \mathbb I\{ f(\X)=c\}$.
Recall that
$$
{\ell^\rs}(\X, c) = \mathbb{P}_{\Z\sim\Pi_{\X}}\left(f(\Z)=c\right) = \Pi_{\X}[h_0].
$$
Therefore, $h_0$ satisfies the constraints in the optimization, which makes it obvious that
$$
{\ell^\rs}(\X', c) = \Pi_{\X'}[h_0] \geq
\min_{h \in \H }\left \{ \Pi_{\X'}[h] ~~~s.t.~~~ \Pi_{\X}[h] = {\ell^\rs}(\X, c) \right\}.
$$
Taking $\min_{\X' \in S_\X}$ on both sides yields the lower bound. The upper bound follows the same derivation.
\end{proof}
Therefore, the problem reduces to deriving bounds for the optimization problems.
\begin{theorem} \label{thm:constrain}
Under the assumptions of Theorem~\ref{thm:lb},
for the optimization problems in Lemma~\ref{lem:first},
we have
\begin{align*}
g^{\rs}_{low}(\X, c) \geq \max ({\ell^\rs}(\X,c) - q_\X,~ 0),
&&
g^{\rs}_{up}(\X, c) \leq \min({\ell^\rs}(\X, c) + q_\X, ~ 1).
\end{align*}
where $q_\X$ is the quantity defined in Theorem \ref{thm:lb} in the main text.
\end{theorem}
Now we proceed to prove Theorem \ref{thm:constrain}.
\begin{proof}[Proof of Theorem~\ref{thm:constrain}]
We only consider the minimization problem because the maximization follows the same proof. For notation, we denote $p = {\ell^\rs}(\X, c)$.
Applying the Lagrange multiplier to the constraint optimization problem
and exchanging the min and max, we have
\begin{align*}
g^{\rs}_{low}(\X, c) = & \min_{{\X'\in S_\X}}\min_{h\in\H}\mathrm{\max_{\lambda\in \mathbb{R}}}\ \Pi_{\X'}[h]-\lambda\Pi_{\X}[h]+\lambda p\\
\ge & \mathrm{\max_{\lambda\in \mathbb{R}}}\min_{{\X'\in S_\X}}\min_{h\in\H}\ \Pi_{\X'}[h]-\lambda\Pi_{\X}[h]+\lambda p\\
= & \mathrm{\max_{\lambda\in \mathbb{R}}}\min_{{\X'\in S_\X}}\min_{h\in\H}\ \int h(\Z)\left(d\Pi_{\X'}(\Z)-\lambda d\Pi_{\X}(\Z)\right)+\lambda p\\
= & -\mathrm{\max_{\lambda\in \mathbb{R}}}\max_{{\X'\in S_\X}}\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}+\lambda p
\\
= & -\mathrm{\max_{\lambda\ge0}}\max_{{\X'\in S_\X}}\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}+\lambda p.
\end{align*}
Here $d\Pi_{\X}^{0}(\Z)$ and $d\Pi_{\X'}^{0}(\Z)$ is the counting
measure and $(s)_{+}=\max(s,0)$. Now we calculate $\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}$.
\begin{lemma} \label{lem:analytic}
Given $\x,\x'$, define $n_{\x}=\left|P_{\x}\right|$, $n_{\x'}=\left|P_{\x'}\right|$
and $n_{\x,\x'}=\left|P_{\x}\cap P_{\x'}\right|$. We have the following identity
\begin{align*}
& \int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}
\\
= & \lambda\left[1-\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}\right]+\left[\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}\right]\left[\lambda-\prod_{j\in[L],\x_{j}\ne\x'_{j}}\frac{n_{\x_{j}}}{n_{\x'_{j}}}\right]_{+}.
\end{align*}
As a result,
under the assumption that $n_x=\left|P_{\x}\right|=\left|P_{\x'}\right|=n_{x'}$ for every word $x$ and
its synonym $\x'\in S_{\x}$, we have
\begin{align*}
\int \left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}
= &
\lambda\left[1-\prod_{j\in[L],\x_{j}\neq\x'_{j}}
\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}
\right]+\left[\prod_{j\in[L],\x_{j}\neq\x'_{j}}
\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}
\right]\left(\lambda-1\right)_{+}.
\end{align*}
\end{lemma}
We now need to solve the optimization of $\max_{{\X'\in S_\X}}\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}$.
\newcommand{\tilde{\x}^*}{\tilde{\x}^*}
\begin{lemma} \label{lem:optperterb}
For any word $\x$, define $\tilde{\x}^*=\underset{\x'\in S_{\x}}{\arg\min}\ n_{\x,\x'}/n_{\x}$.
For a given sentence $\X = x_1,\ldots, x_L$, we define an ordering of the words
$x_{\ell_1}, \ldots, x_{\ell_L}$ such that
$n_{\x_{\ell_{i}},\tilde{\x}^*_{\ell_{i}}}/n_{\x_{\ell_{i}}}\le n_{\x_{\ell_{j}},\tilde{\x}^*_{\ell_{j}}}/n_{\x_{\ell_{j}}}$
for any $i\le j$.
For a given $\X$ and $R$, we define an adversarial perturbed sentence
$
\X^* = x^*_1,\ldots, x^*_L, $
where
$$
x_i^* = \begin{cases}
\tilde{\x}_{i}^{*} & \text{if $i \in [\ell_1, \ldots, \ell_R]$} \\
{\x}_{i} & \text{if $i \notin [\ell_1, \ldots, \ell_R]$}. \\
\end{cases}
$$
Then for any $\lambda\ge0$, we have that $\X^*$ is the optimal solution of $\max_{{\X'\in S_\X}}\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}$, that is,
\[
\max_{{\X'\in S_\X}}\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}=\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X^*}(\Z)\right)_{+}.
\]
\end{lemma}
Now by Lemma \ref{lem:optperterb}, the lower bound becomes
\begin{align}
g^{\rs}_{low}(\X, c)
= ~& -\mathrm{\max_{\lambda\ge0}}\max_{{\X'\in S_\X}}\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}+\lambda p \notag \\
=~ & - \mathrm{\max_{\lambda\ge0}}\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X^*}(\Z)\right)_{+}+\lambda p\notag \\
= ~& \mathrm{\max_{\lambda\ge0}}\mathrm{\ }(p-q_\X)\lambda-(1-q_\X)(\lambda-1)_{+} \label{equ:dfdfdjijiojioj}\\
= ~ & \max( p - q_{\X}, ~ 0 ), \notag
\end{align}
where $q_\X$ is consistent with the definition in Theorem~\ref{thm:lb}: \[
q_\X=1-\prod_{j\in[L],\x_{j}\neq\tilde{\x}_{j}^{*}}\frac{n_{\x_{j},\tilde{\x}_{j}^{*}}}{n_{\x_{j}}} =
1- \prod_{j=1}^R q_{x_{\ell_j}}.
\]
Here equation (\ref{equ:dfdfdjijiojioj}) is by calculation using the assumption of Theorem \ref{thm:lb}.
The optimization of $\max_{\lambda\geq 0}$ in \eqref{equ:dfdfdjijiojioj} is an elementary step:
if $p\le q$, we have $\lambda^{*}=0$ with
solution $0$; if $p\ge q$, we have $\lambda^{*}=1$ with solution $(p-q_\X)$.
This finishes the proof of the lower bound. The proof the upper bound follows similarly.
\end{proof}
\paragraph{Proof of Lemma \ref{lem:analytic}}
Notice that we have
\begin{align*}
\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}
= & \sum_{\Z\in S_{\X'}\cap S_{\X}}\left(\lambda\left|S_{\X}\right|^{-1}-\left|S_{\X'}\right|^{-1}\right)_{+}
+ \lambda\sum_{\Z\in S_{\X}-S_{\X'}}\left|S_{\X}\right|^{-1}\\
= & \left|S_{\X'}\cap S_{\X}\right|\left(\lambda\left|S_{\X}\right|^{-1}-\left|S_{\X'}\right|^{-1}\right)_{+}
+ \lambda\left|S_{\X}-S_{\X'}\right|\left|S_{\X}\right|^{-1}.
\end{align*}
Also notice that $\left|S_{\X}\right| =\prod_{j=1}^{L}n_{\x_{j}}$; $\left|S_{\X'}\right| =\prod_{j=1}^{L}n_{\x_{j}'}$; $\left|S_{\X'}\cap S_{\X}\right| = \prod_{j=1}^{L}n_{\x_{j},\x_{j}'}$ and $\left|S_{\X}-S_{\X'}\right| =\prod_{j=1}^{L}n_{\x_{j}}-\prod_{j=1}^{L}n_{\x_{j},\x_{j}'}$.
Plugging in the above value, we have
\begin{align*}
\left|S_{\X}-S_{\X'}\right|\left|S_{\X}\right|^{-1}
= & \frac{\prod_{j=1}^{L}n_{\x_{j}}-\prod_{j=1}^{L}n_{\x_{j},\x_{j}'}}{\prod_{j=1}^{L}n_{\x_{j}}}\\
= & 1-\prod_{j=1}^{L}\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}\\
= & 1-\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}.
\end{align*}
And also,
\begin{align*}
\left(\lambda\left|S_{\X}\right|^{-1}-\left|S_{\X'}\right|^{-1}\right)_{+}
= & \left(\lambda\prod_{j=1}^{L}n_{\x_{j}}^{-1}-\prod_{j=1}^{L}n_{\x_{j}'}^{-1}\right)_{+}\\
= & \left(\lambda\prod_{j\in[L],\x_{j}=\x'_{j}}n_{\x_{j}}^{-1}\prod_{j\in[L],\x_{j}\ne\x'_{j}}n_{\x_{j}}^{-1}-\prod_{j\in[L],\x_{j}=\x'_{j}}n_{\x_{j}}^{-1}\prod_{j\in[L],\x_{j}\ne\x'_{j}}n_{\x'_{j}}^{-1}\right)_{+}\\
= & \prod_{j\in[L],\x_{j}=\x'_{j}}n_{\x_{j}}^{-1}\left(\lambda\prod_{j\in[L],\x_{j}\ne\x'_{j}}n_{\x_{j}}^{-1}-\prod_{j\in[L],\x_{j}\ne\x'_{j}}n_{\x'_{j}}^{-1}\right)_{+}.
\end{align*}
Plugging in the above value, we have
\begin{align*}
\left|S_{\X'}\cap S_{\X}\right|\left(\lambda\left|S_{\X}\right|^{-1}-\left|S_{\X'}\right|^{-1}\right)_{+}
= & \prod_{j=1}^{L}n_{\x_{j},\x_{j}'}\left(\lambda\left|S_{\X}\right|^{-1}-\left|S_{\X'}\right|^{-1}\right)_{+}\\
= & \prod_{j\in[L],\x_{j}=\x'_{j}}n_{\x_{j}}\prod_{j\in[L],\x_{j}\neq\x'_{j}}n_{\x_{j},\x_{j}'}\left(\lambda\left|S_{\X}\right|^{-1}-\left|S_{\X'}\right|^{-1}\right)_{+}\\
= & \prod_{j\in[L],\x_{j}\neq\x'_{j}}n_{\x_{j},\x_{j}'}\left(\lambda\prod_{j\in[L],\x_{j}\ne\x'_{j}}n_{\x_{j}}^{-1}-\prod_{j\in[L],\x_{j}\ne\x'_{j}}n_{\x'_{j}}^{-1}\right)_{+}\\
= & \prod_{j\in[L],\x_{j}\neq\x'_{j}}n_{\x_{j},\x_{j}'}\prod_{j\in[L],\x_{j}\ne\x'_{j}}n_{\x_{j}}^{-1}\left(\lambda-\prod_{j\in[L],\x_{j}\ne\x'_{j}}\frac{n_{\x_{j}}}{n_{\x'_{j}}}\right)_{+}\\
= & \left(\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}\right)\left(\lambda-\prod_{j\in[L],\x_{j}\ne\x'_{j}}\frac{n_{\x_{j}}}{n_{\x'_{j}}}\right)_{+}.
\end{align*}
Combining all the calculation, we get
\begin{align*}
& \int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}
\\
= & \lambda\left[1-\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}\right]+\left[\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}'}}{n_{\x_{j}}}\right]\left[\lambda-\prod_{j\in[L],\x_{j}\ne\x'_{j}}\frac{n_{\x_{j}}}{n_{\x'_{j}}}\right]_{+}.
\end{align*}
\paragraph{Proof of Lemma \ref{lem:optperterb}}
It is sufficient to proof that, for any $\X'\neq\X^*$, we have
\[
\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X^*}(\Z)\right)_{+}\ge\int\left(\lambda d\Pi_{\X}(\Z)-d\Pi_{\X'}(\Z)\right)_{+}.
\]
Notice that for any $\lambda\ge0$, define
\[
Q(\X,\X'')=\lambda\left[1-\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}''}}{n_{\x_{j}}}\right]+\left[\prod_{j\in[L],\x_{j}\neq\x'_{j}}\frac{n_{\x_{j},\x_{j}''}}{n_{\x_{j}}}\right]\left(\lambda-1\right)_{+}.
\]
Given any \X, we can view $Q(\X,\X'')$ as the function of $n_{\x_{i},\x_{i}''}/n_{\x_{i}}$, $i\in[L]$. And $Q(\X,\X'')$ is a decreasing function of $n_{\x_{i},\x_{i}''}/n_{\x_{i}}$
for any $i\in[L]$ when fixing $\frac{n_{\x_{j},\x_{j}''}}{n_{\x_{j}}}$ for all other $j \neq i$. Suppose $\tilde{r}_k$ is the $k$-th smallest quantities of $n_{\x_{i},\tilde{\x}_{i}^{*}}/n_{\x_{i}}$, $i\in[L]$ and $r'_k$ is the $k$-th smallest quantities of $n_{\x_{j},\tilde{\x}_{j}^{*}}/n_{\x_{i}}$, $i\in[L]$. By the construction of $\X^*$, we have $\tilde{r}_k \le r'_k$ for any $k\in[L]$. This implies that
\[
Q(\X,\X^*)\ge Q(\X,\X').
\]
\subsubsection{Proof of Theorem \ref{thm:tight}}
We denote ${\ell^\rs}(\X, y) = p_A$, ${\ell^\rs}(\X, y_B) = p_B$ and $q = q_\X$ in this proof for simplicity.
The $\X^*$ below is the one defined in the proof of Lemme \ref{lem:optperterb}.
Our proof is based on constructing a randomized smoothing classifier that
satisfies the desired property we want to prove.
\paragraph{Case 1 $p_{A}\ge q$ and $p_{B}+q\le1$}
Note that in this case $\left|S_{\X}\cap S_{\X^{*}}\right|/\left|S_{\X}\right|=1-q\ge(p_{A}-q)+p_{B}$,
where the inequality is due to $p_{A}+p_{B}\le1$. Therefore, we can
choose set $U_{1}$ and $U_{2}$ such that $U_{1}\subseteq S_{\X}\cap S_{\X^{*}}$;
$U_{2}\subseteq S_{\X}\cap S_{\X^{*}}$; $U_{1}\cap U_{2}=\emptyset$;
$\left|U_{1}\right|/\left|S_{\X}\right|=p_{A}-q$ and $\left|U_{2}\right|/\left|S_{\X}\right|=p_{B}$. We
define the classifier:
\[
f^{*}(\Z)=\begin{cases}
y & \text{if}\ \Z\in\left(S_{\X}-S_{\X^{*}}\right)\cap U_{1}\\
y_B & \text{if}\ \Z\in \left(S_{\X^{*}}-S_{\X}\right) \cup U_{2} \\
\text{other class ($c \neq y$ or $y_B$)} & \text{if}\ \Z\in S_{\X}\cap S_{\X^{*}}-\left(U_{1}\cup U_{2}\right)\\
\text{any\ class ($c\in \mathcal Y$)} & \text{otherwise}
\end{cases}
\]
This classifier is well defined for binary classification because
$S_{\X}\cap S_{\X^{*}}-\left(U_{1}\cup U_{2}\right)=\emptyset$.
\paragraph{Case 2 $p_{A}<q$ and $p_{B}+q\le1$}
In this case, we can choose set $U_{1}$ and $U_{2}$ such that $U_{1}\subseteq S_{\X}-S_{\X^{*}}$;
$U_{2}\subseteq S_{\X}\cap S_{\X^{*}}$; $\left|U_{1}\right|/\left|S_{\X}\right|=p_{A}$
and $\left|U_{2}\right|/\left|S_{\X}\right|=p_{B}$. We define the classifier:
\[
f^{*}(\Z)=\begin{cases}
y & \text{if}\ \Z\in U_{1}\\
y_B & \text{if}\ \Z\in U_{2}\cup\left(S_{\X^{*}}-S_{\X}\right)\\
\text{other class ($c \neq y$ or $y_B$)} & \text{if}\ \Z\in S_{\X}-\left(U_{1}\cup U_{2}\right)\\
\text{any\ class ($c\in \mathcal Y$)} & \text{otherwise}
\end{cases}
\]
This classifier is well defined for binary classification because
$S_{\X}-\left(U_{1}\cup U_{2}\right)=\emptyset$.
\paragraph{Case 3 $p_{A}\ge q$ and $p_{B}+q>1$}
This case does not exist since we would have $p_{A}+p_{B}>1$.
\paragraph{Case 4 $p_{A}<q$ and $p_{B}+q>1$}
We choose set $U_{1}$ and $U_{2}$ such that $U_{1}\subseteq S_{\X}-S_{\X^{*}}$;
$U_{2}\in S_{\X}-S_{\X^{*}}$; $U_{1}\cap U_{2}=\emptyset$; $\left|U_{1}\right|/\left|S_{\X}\right|=p_{A}$
and $\left|U_{2}\right|/\left|S_{\X}\right|=p_{B}-(1-q)$. Notice that the intersect of
$U_{1}$ and $U_{2}$ can be empty as $\left|U_{1}\right|/\left|S_{\X}\right|+\left|U_{2}\right|/\left|S_{\X}\right|=p_{A}+p_{B}-(1-q)\le1-(1-q)=q=\left|S_{\X}-S_{\X^{*}}\right|/\left|S_{\X}\right|$.
We define the classifier:
\[
f^{*}(\Z)=\begin{cases}
y & \text{if}\ \Z\in U_{1}\\
y_B & \text{if}\ \Z\in U_{2}\cup S_{\X^{*}}\\
\text{other class ($c \neq y$ or $y_B$)} & \text{if}\ \Z\in\left(S_{\X}-S_{\X^{*}}\right)-\left(U_{1}\cup U_{2}\right)\\
\text{any\ class ($c\in \mathcal Y$)} & \text{otherwise}
\end{cases}
\]
This classifier is well defined for binary classification because
$S_{\X}-S_{\X^{*}}-\left(U_{1}\cup U_{2}\right)=\emptyset$.
It can be easily verified that for each case, the defined classifier
satisfies all the conditions in Theorem \ref{thm:tight}.
\section{Additional Experiment Details}
We set $R = L$ in adversarial attacking, that is,
all words in the sentence can be perturbed simultaneously by the attacker.
We use 5,000 random draws in the Monte Carlo estimation of $\Delta_\X$, and use the same method in \citet{jia2019certified} to tune the hyper-parameters when training the base models {e.g. learning rate, batch size and the schedule of loss function}.
For the IMDB dataset, we train the IBP models and ours for 60 and 10 epochs, respectively. For the Amazon dataset, we train the IBP models and ours for 100 and 20 epochs, respectively.
We test our algorithm on two different datasets, IMDB and Amazon.
The IMDB movie review dataset~\citep{maas2011imdb} is a sentiment classification dataset.
It consists of 50,000 movie review comments with binary sentiment labels.
The Amazon review dataset~\citep{mcauley2013hidden} is an extremely large dataset
that contains 34,686,770 reviews with 5 different types of labels. Similar to \citet{cohen2019certified}, we test the models on randomly selected subsets of the test set with 1,250 and 6,500 examples for IMDB and Amazon dataset, respectively.
\section{Conclusion}
We proposed a robustness certification method, which provably guarantees that all the possible perturbations cannot break down the system. Compared with previous work such as \citet{jia2019certified, huang2019achieving}, our method is structure-free and thus can be easily applied to any pre-trained models (such as BERT) and character-level models (such as Char-CNN).
The construction of the perturbation set is of critical importance to our method. In this paper, we used a heuristic way based on the synonym network to construct the perturbation set, which may not be optimal. In further work, we will explore more efficient ways for constructing the perturbation set. We also plan to generalize our approach to achieve certified robustness against other types of adversarial attacks in NLP, such as the out-of-list attack. An na\"ive way is to add the ``OOV" token into the synonyms set of every word, but potentially better procedures can be further explored.
\section*{Acknowledgement}
This work is supported in part by NSF CRII 1830161 and NSF CAREER 1846421.
\section{Experiments} \label{sec:exp}
We test our method on both IMDB~\citep{maas2011imdb} and Amazon~\citep{mcauley2013hidden} text classification tasks, with various types of models, including text CNN~\citep{kim2014textcnn}, Char-CNN~\citep{zhang2015charnn} and BERT~\citep{devlin2018bert}.
We compare with the recent IBP-based methods \citep{jia2019certified, huang2019achieving} as baselines.
Text CNN ~\citep{kim2014textcnn} was used in \citet{jia2019certified} and achieves the best result therein.
All the baseline models are trained and tuned using the schedules recommended in the corresponding papers.
We consider the case when $R=L$ during attacking, which means all words in the sentence can be perturbed simultaneously by the attacker.
Code for reproducing our results can be found in \url{https://github.com/lushleaf/Structure-free-certified-NLP}.
\paragraph{Synonym Sets}
Similar to \citet{jia2019certified, alzantot2018generating}, we construct the synonym set $S_x$ of word $x$ to be the set of words with $\geq 0.8$ cosine similarity in the GLOVE vector space.
The word vector space is constructed by post-processing the pre-trained GLOVE vectors~\citep{pennington2014glove} using the counter-fitted method \citep{mrkvsic2016counter} and the ``all-but-the-top'' method \citep{mu2018allbutthetop} to ensure that synonyms are near to each other while antonyms are far apart.
\paragraph{Perturbation Sets}
We say that two words $x$ and $ x'$ are connected synonymously if there exists a path of words $x = x_1, x_2, \ldots, x_\ell = x'$, such that all the successive pairs are synonymous.
Let $B_x$ to be the set of words connected to $x$ synonymously.
Then we define the perturbation set $P_x$ to consist of the top $K$ words in $B_x$ with the largest GLOVE cosine similarity if $|B_x| \geq K$, and set $P_x = B_x$ if $|B_x| < K$.
Here $K$ is a hyper-parameter that controls the size of $P_x$ and hence
trades off the smoothness and accuracy of $f^\rs$.
We use $K=100$ by default and investigate its effect in
Section \ref{apx:trades}.
\paragraph{Evaluation Metric}
We evaluate the certified robustness of a model $f^\rs$
on a dataset with the \emph{certified accuracy} \citep{cohen2019certified},
which equals the percentage of data points on which
$f^\rs$ is certified robust, which, for our method, holds when $\Delta_\X > 0$ can be verified.
\cutforspace{
For the IBP-based methods \citep{jia2019certified, huang2019achieving},
the certified accuracy is calculated similarly, with certified robustness holds
when the lower bound of the logits of the true labels are larger than the upper bounds of all the other classes based on IBP.
}
\subsection{Main Results}
We first demonstrate that adversarial word substitution is able to give strong attack in our experimental setting. Using IMDB dataset, we attack the vanilla BERT \citep{devlin2018bert} with the adversarial attacking method of \citet{jin2019bert}. The vanilla BERT achieves a $91\%$ clean accuracy (the testing accuracy on clean data without attacking), but only a $20.1\%$ adversarial accuracy (the testing accuracy under the particular attacking method by \citet{jin2019bert}). We will show later that our method is able to achieve $87.35\%$ certified accuracy and thus the corresponding adversarial accuracy must be higher or equal to $87.35\%$.
We compare our method with IBP ~\citep{jia2019certified, huang2019achieving}.
in Table~\ref{tab:textcnn}.
We can see that our method clearly outperforms the baselines.
In particular,
our approach significantly outperforms IBP on Amazon by improving the $14.00\%$ baseline to $24.92\%$.
Thanks to its structure-free property,
our algorithm can be easily applied to any pre-trained models and character-level models,
which is not easily achievable with \citet{jia2019certified} and \citet{huang2019achieving}.
Table ~\ref{tab:model} shows that
our method can further improve the result using Char-CNN (a character-level model) and
BERT~\citep{devlin2018bert},
achieving an $87.35\%$ certified accuracy on IMDB.
In comparison, the IBP baseline only achieves a $79.74\%$ certified accuracy under the same setting.
\begin{table}[btp]
\begin{center}\scalebox{.9}{
\begin{tabular}{l|cc}
\toprule
\bf Method & \bf IMDB & \bf Amazon \\
\hline
\citet{jia2019certified} & 79.74 & 14.00 \\
\citet{huang2019achieving} & 78.74 & 12.36 \\
Ours & \bf 81.16 & \bf 24.92 \\
\bottomrule
\end{tabular}}
\end{center}
\caption{\label{tab:textcnn} The certified accuracy of our method and the baselines on the IMDB and Amazon dataset.}
\end{table}
\begin{table}[btp]
\begin{center}\scalebox{.9}{
\begin{tabular}{l|l|c}
\toprule
\bf Method & \bf Model & \bf Accuracy \\
\hline
\citet{jia2019certified} & CNN & 79.74 \\
\citet{huang2019achieving} & CNN & 78.74 \\
\hline
\multirow{3}{*}{Ours} & CNN & 81.16 \\
& Char-CNN & 82.03 \\
& BERT & \bf 87.35 \\
\bottomrule
\end{tabular}}
\end{center}
\caption{\label{tab:model} The certified accuracy of different models and methods on the IMDB dataset.}
\end{table}
\subsection{Trade-Off between Clean Accuracy and Certified Accuracy}
\label{apx:trades}
We
investigate the trade-off between smoothness and accuracy while tuning $K$ in Table \ref{tab:trades}.
We can see that the clean accuracy decreases when $K$ increases, while
the gap between the clean accuracy and certified accuracy,
which measures the smoothness, decreases when $K$ increases. The best certified accuracy is achieved when $K=100$.
\begin{table}[hbtp]
\begin{centering}
\begin{tabular}{l|ccccc}
\toprule
$K$ & 20 & 50 & 100 & 250 & 1000 \\
\hline
\small{Clean (\%)} & \small{88.47} & \small{88.48} & \small{88.09} & \small{84.83} & \small{67.54} \\
\small{Certified (\%)} & \small{65.58} & \small{77.32} & \small{81.16} & \small{79.98} & \small{65.13} \\
\bottomrule
\end{tabular}\caption{\label{tab:trades}Results of the smoothed model $f^\rs$ with different $K$ on IMDB using text CNN. ``Clean" represents the accuracy on the clean data without adversarial attacking and ``Certified" the certified accuracy.
}
\par\end{centering}
\end{table}
\section{Introduction}
Deep neural networks have achieved state-of-the-art results in many NLP tasks,
but also have been shown to be brittle to carefully crafted adversarial perturbations, such as
replacing words with similar words ~\citep{alzantot2018generating},
adding extra text ~\citep{wallace2019universal},
and replacing sentences with semantically similar sentences ~\citep{ribeiro2018semantically}.
These adversarial perturbations are
imperceptible to humans,
but can fool deep neural networks and break their performance.
Efficient methods for defending these attacks
are of critical importance for deploying modern deep NLP models to practical automatic AI systems.
In this paper, we focus on defending the synonymous
word substitution attacking ~\citep{alzantot2018generating},
in which an attacker attempts to alter the output of the model by replacing words in the input sentence with their synonyms according to
a synonym table, while keeping the meaning of this sentence unchanged.
A model is said to be \emph{certified robust} if such an attack is guaranteed to fail, no matter how the attacker manipulates the input sentences.
Achieving and verifying certified robustness is highly challenging even if the synonym table used by the attacker is known during training
\citep[see][]{jia2019certified},
because it requires
to check every possible
synonymous word substitution, whose number is exponentially large.
Various defense methods against
synonymous word substitution attacks have been developed
~\citep[e.g.,][]{wallace2019universal, ebrahimi2017hotflip}, most of which, however,
are
not {certified robust} in that they may eventually be broken by stronger attackers.
Recently,
\citet{jia2019certified, huang2019achieving} proposed
the first certified robust methods against word substitution attacking. Their methods are based on the interval bound propagation (IBP) method
\citep{dvijotham2018ibp} which computes the range of the model output by propagating the interval constraints of the inputs layer by layer.
However, the IBP-based methods of \citet{jia2019certified, huang2019achieving}
are limited in several ways.
First,
because IBP only works for certifying neural networks with continuous inputs, the inputs in \citet{jia2019certified} and \citet{huang2019achieving} are taken to be the word embedding vectors of the input sentences, instead of the discrete sentences.
This makes
it inapplicable to character-level~\citep{zhang2015charnn} and subword-level~\citep{bojanowski2017subword} model, which are more widely used in practice \citep{wu2016google}.
\cutforspace{
Secondly,
IBP requires to access the structures of the models,
and needs to be derived and implemented at a model-by-model basis,
which makes it difficult to use complex models, especially large-scale pre-trained models such as
BERT ~\citep{devlin2018bert}.
Finally,
the embedding vectors in NLP models can be non-local and hence make the IBP bounds relatively loose,
because semantically similar words can have very different embedding vectors~\citep{mu2018allbutthetop}.
In comparison, the adversarial perturbations in image classifications are constrained in small balls (e.g., $\ell_2$ or $\ell_1$), on which IBP can provide tighter bounds.
}
In this paper,
we
propose a \emph{structure-free}
certified defense method
that applies to arbitrary models that can be queried in a black-box fashion,
without any requirement on the model structures.
Our method is based on the idea of randomized smoothing,
which smooths
the model with random word substitutions
build on the synonymous network,
and leverage the statistical properties of the randomized ensembles to construct provably certification bounds.
Similar ideas of provably certification
using
randomized smoothing
have been developed recently in deep learning \citep[e.g.,][]{cohen2019certified, salman2019provably, zhang2020black, lee2019tight},
but mainly
for computer vision tasks whose inputs (images) are in a continuous space \citep{cohen2019certified}.
Our method admits a substantial extension
of the randomized smoothing technique to discrete and structured input spaces for NLP.
We test our method on various types of NLP models, including text CNN \citep{kim2014textcnn}, Char-CNN \citep{zhang2015charnn}, and BERT \citep{devlin2018bert}.
Our method significantly outperforms the recent IBP-based methods \citep{jia2019certified, huang2019achieving}
on both IMDB and Amazon text classification.
In particular, we achieve an 87.35\% certified accuracy on IMDB by applying our method on the state-of-the-art BERT,
on which previous certified robust methods are not applicable.
\section{Adversarial Word Substitution} \label{sec:background}
In a text classification task,
a model $f(\X)$ maps an input sentence $\X\in\mathcal{X}$ to a label $c$ in a set $\mathcal{Y}$ of discrete categories, where
$\X = \x_{1},\ldots, \x_{L}$ is a sentence consisting of $L$ words.
In this paper, we focus on adversarial word substitution in which
an attacker
arbitrarily replaces the words in the sentence by their synonyms according to a synonym table to alert the prediction of the model.
Specifically, for any word $\x$, we consider a pre-defined synonym set $S_{\x}$ that contains the synonyms of $\x$ (including $\x$ itself).
We assume the synonymous relation is symmetric, that is,
$\x$ is in the synonym set of all its synonyms.
The synonym set $S_{\x}$ can be built based on GLOVE~\citep{pennington2014glove}.
With a given input sentence $\X = \x_{1}$,\ldots, $\x_{L}$,
the attacker may construct an adversarial sentence ${\X'}=\x'_{1},\ldots,\x'_{L}$ by perturbing at most
$R\le L$ words $x_i$ in $\X$ to any of their synonyms $ x'_i \in S_{x_i}$
\begin{align*
\begin{split}
&
S_{\X} := \left \{
{\X'}\colon
\norm{\X' - \X}_{0} \leq R, ~~\x'_{i}\in S_{\x_{i}}, \forall i
\right\},
\end{split}
\end{align*}
where $S_{\X}$ denotes the candidate set of adversarial sentences available to the attacker.
Here $\norm{ \X' - \X}_{0} := \sum_{i=1}^{L}\mathbb{I}\left\{ \x'_{i}\ne\x_{i}\right\}$ is the Hamming distance, with $\mathbb{I}\{\cdot\}$ the indicator function.
It is expected that all $ \X' \in S_{\X}$ have the same semantic meaning as $\X$ for human readers, but they may have different outputs from the model.
The goal of the attacker is to find
$ \X' \in S_{\X}$ such that
$f(\X) \neq f( \X')$.
\paragraph{Certified Robustness}
Formally,
a model $f$
is said to be \emph{certified robust} against word substitution attacking on an input $\X$ if it is able to give consistently correct predictions for all the possible word substitution perturbations, i.e,
\begin{align}\label{equ:certified}
\text{$y = f(\X) = f( \X')$,~~~~ for all $ \X' \in S_\X,$}
\end{align}
where $y$ denotes the true label of sentence $\X$.
Deciding if $f$ is certified robust can be highly challenging,
because,
unless additional structural information is available,
it requires to exam all the candidate sentences in $S_\X$,
whose size grows exponentially with $R$. In this work, we mainly consider the case when $R = L$, which is the most challenging case.
\section{Certifying Smoothed Classifiers}
Our idea is to replace $f$ with a more smoothed model that is easier to verify
by averaging the outputs of a set of randomly perturbed inputs based on random word substitutions.
The smoothed classifier $f^{\rs}$ is constructed by introducing random perturbations on the input space,
\begin{align*}
& f^{\rs}(\X)=\underset{c\in\mathcal{Y}}{\arg\max}\
\mathbb{P}_{\Z\sim\Pi_{\X}}\left(f(\Z)=c\right),
\end{align*}
where $\Pi_{\X}$ is a probability distribution on the input space that prescribes a random perturbation around $\X$.
For notation, we define
$$
{\ell^\rs}(\X, c) :=\mathbb{P}_{\Z\sim\Pi_{\X}}\left(f(\Z)=c\right) ,
$$
which is the ``soft score'' of class $c$ under $f^\rs$.
The perturbation distribution ${\Pi_\X}$ should be chosen properly so that $f^\rs$ forms a close approximation to the original model $f$ (i.e.,
$f^\rs(\X) \approx f(\X)$), and is also sufficiently random to ensure that $f^\rs$ is smooth enough to allow certified robustness (in the sense of Theorem~\ref{thm:lb} below).
\cutforspace{
For vision tasks, Gaussian distribution is a natural choice for $\Pi_{\X}$ as the input is in a continuous Euclidean space \cite{cohen2019certified}.
However,
Gaussian perturbation can not be directly applied on
discrete sentence inputs.
}
In our work,
we define $\Pi_\X$ to be the uniform distribution
on a set of random word substitutions.
Specifically, let $P_{\x}$ be
a \emph{perturbation set} for word $\x$ in the vocabulary,
which is different from the \emph{synonym set} $S_\x$.
In this work, we construct $P_{\x}$ based on the top $K$ nearest neighbors
under the cosine similarity of GLOVE vectors,
where $K$ is a hyperparameter that controls the size of the perturbation set; see Section \ref{sec:exp} for more discussion on $P_x$.
For a sentence $\X=\x_{1},\ldots,\x_{L}$,
the sentence-level perturbation distribution $\Pi_\X$ is
defined by randomly and independently perturbing each word $x_i$
to a word in its perturbation set $P_{x_i}$
with equal probability, that is,
\vspace{-10pt}
\[
\Pi_{\X}(\Z)=\prod_{i=1}^{L}\frac{\mathbb{I}\left\{ z_{i}\in P_{\x_{i}}\right\} }{\left|P_{\x_{i}}\right|},
\]
\vspace{-10pt}
\noindent where $\Z = z_{1},\ldots, z_{L}$ is the perturbed sentence and $|P_{x_i}|$ denotes the size of $P_{x_i}$.
Note that the random perturbation $\Z$ and the adversarial candidate $\X' \in S_\X$ are different.
\begin{figure*}
\vspace{-15pt}
\centering
\includegraphics[scale = 0.55]{img/figure_new3.pdf
\vspace{-5pt}
\caption{A pipeline of the proposed robustness certification approach.}
\vspace{-15pt}
\label{fig:framework}
\end{figure*}
\subsection{Certified Robustness}
We now discuss how to certify the robustness of the smoothed model $f^{\rs}$.
Recall that $f^\rs$ is certified robust if $y = f^\rs(\X')$ for any $\X' \in S_{\X}$, where $y$ is the true label.
A sufficient condition for this is
$$
\min_{\X'\in S_{\X}} {\ell^\rs}(\X', y) \geq
\max_{\X' \in S_{\X}}{\ell^\rs}({\X'}, c) ~~~~~\forall c \neq y,
$$
where the lower bound of ${\ell^\rs}(\X', y)$ on $\X' \in S_\X$
is larger than the upper bound of ${\ell^\rs}(\X', c)$ on $\X' \in S_\X$ for every $c\neq y$.
The key step is hence to calculate the upper and low bounds of $ {\ell^\rs}(\X', c) $ for $\forall c\in \mathcal Y$ and $\X'\in S_\X$, which we address in Theorem~\ref{thm:lb} below.
All proofs are in Appendix \ref{apx:pf}.
\begin{theorem} \textbf{(Certified Lower/Upper Bounds)} \label{thm:lb}
Assume the perturbation set $P_{\x}$ is constructed such that
$\left|P_{\x}\right|=\left|P_{{\x'}}\right|$ for every word $x$ and its synonym ${\x'}\in S_{\x}$. Define
\[
q_{\x}=\min_{{\x'}\in S_{\x}}{\left|P_{\x}\cap P_{{\x'}}\right|}/{\left|P_{\x}\right|},
\]
where $q_{\x}$ indicates the overlap between the two different perturbation sets.
For a given sentence $\X=\x_{1},\ldots,\x_{L}$,
we sort the words according to $q_{\x}$, such that
$q_{\x_{i_{1}}}\le q_{\x_{i_{2}}}\le\cdots\le q_{\x_{i_{L}}}$.
Then
\begin{align*}
\min_{\X'\in S_{\X}}
{\ell^\rs}(\X', c)
& \ge\max({\ell^\rs}(\X, c) -q_\X,0)\\
\max_{\X'\in S_{\X}}
{\ell^\rs}(\X', c)
& \le\min({\ell^\rs}(\X, c) + q_\X,1).
\end{align*}
where $q_\X:=1-\prod_{j=1}^{R}q_{\x_{i_{j}}}$.
Equivalently, this says
\begin{align*}
\left | {\ell^\rs}(\X', c) - {\ell^\rs}(\X, c) \right | \leq q_\X,~~~~
\text{any label $c \in \mathcal Y$}.
\end{align*}
\end{theorem}
The idea is that, with the randomized smoothing,
the difference between ${\ell^\rs}(\X', c)$ and ${\ell^\rs}(\X, c)$ is at most $q_\X$ for any adversarial candidate $\X' \in S_\X$. Therefore, we can give adversarial upper and lower bounds of ${\ell^\rs}(\X', c)$ by ${\ell^\rs}(\X, c)\pm q_\X$, which,
importantly,
avoids the difficult adversarial optimization of ${\ell^\rs}(\X', c)$ on ${\X' \in S_\X}$, and instead just needs to evaluate ${\ell^\rs}(\X, c)$ at the original input $\X$.
We are ready to describe a practical criterion for checking the certified robustness.
\begin{proposition} \label{pro:cAcB}
For a sentence $\X$ and its label $y$, we define
\begin{align*}
& y_B = \arg\max_{c\in \mathcal Y, c \neq y} {\ell^\rs}(\X, c).
\end{align*}
Then under the condition of Theorem~\ref{thm:lb}, we can certify that $f(\X') = f(\X) = y$ for any $\X' \in S_{\X}$ if
\begin{align} \label{equ:DeltaX}
\Delta_\X \overset{def}{=} {\ell^\rs}(\X, y) - {\ell^\rs}(\X, y_B) - 2q_{\X} > 0.
\end{align}
\end{proposition}
Therefore, certifying whether the model gives consistently correct prediction reduces to checking if $\Delta_{
\X}$ is positive, which can be easily achieved with Monte Carlo estimation as we show in the sequel.
\paragraph{Estimating ${\ell^\rs}(\X, c)$ and $\Delta_\X$}
Recall that ${\ell^\rs}(\X, c)= \mathbb{P}_{\Z\sim\Pi_{\X}}(f(\Z)=c)$.
We can estimate ${\ell^\rs}(\X, c)$ with a Monte Carlo estimator $\sum_{i=1}^{n}\mathbb{I} \{ f(\Z^{(i)})=c\}/n$,
where $\Z^{(i)}$ are i.i.d. samples from $\Pi_{\X}$.
And $\Delta_\X$ can be approximated accordingly.
Using concentration inequality, we can quantify the non-asymptotic approximation error.
This allows us to construct rigorous statistical procedures to
reject the null hypothesis that $f^\rs$ is not certified robust at $\X$ (i.e., $\Delta_{\X} \leq 0$)
with a given significance level (e.g., $1\%$).
See Appendix \ref{apx:MC} for the algorithmic details of the testing procedure.
We can see that our procedure is
\emph{structure-free} in that it only requires the black-box assessment of the output $f(\Z^{(i)})$ of the random inputs, and does not require any other structural information of $f$ and $f^\rs$, which makes our method widely applicable to various types of complex models.
\paragraph{Tightness}
A key question is if our bounds are sufficiently tight.
The next theorem shows that the lower/upper bounds in Theorem \ref{thm:lb} are tight and can not be further improved unless further information of the model $f$ or $f^\rs$ is acquired.
\begin{theorem}
\textbf{(Tightness)} \label{thm:tight}
Assume the conditions of Theorem \ref{thm:lb} hold.
For a model $f$ that satisfies $f^\rs(\X) = y$ and $y_B$ as defined in Proposition~\ref{pro:cAcB},
there exists a model $f_*$ such that its related smoothed classifier ${{\ell}^\rs_*}$
satisfies
${{\ell}^\rs_*}(\X, c)={\ell^\rs}(\X, c)$ for $c = y$ and $c = y_B$, and
\begin{align*}
\min_{\X'\in S_{\X}}
{{\ell}^\rs_*}(\X', y)
& =\max({{\ell}^\rs_*}(\X, y)-q_\X,0)\\
\max_{\X'\in S_{\X}}
{{\ell}^\rs_*}(\X',y_B)
& =\min({{\ell}^\rs_*}(\X, y_B) +q_\X,1),
\end{align*}
where $q_\X$ is defined in Theorem~\ref{thm:lb}.
\end{theorem}
In other words,
if we access ${\ell^\rs}$ only through the evaluation of ${\ell^\rs}(\X, y)$ and ${\ell^\rs}(\X, y_B)$,
then the bounds in Theorem~\ref{thm:lb} are the tightest possible that we can achieve, because we can not distinguish between ${\ell^\rs}$ and the ${{\ell}^\rs_*}$ in Theorem~\ref{thm:tight} with the information available.
\subsection{Practical Algorithm} \label{sec:prac}
Figure \ref{fig:framework} visualizes the pipeline of the
proposed approach. Given the synonym sets $S_\X$, we generate the perturbation sets $P_\X$ from it.
When an input sentence $\X$ arrives,
we draw perturbed sentences $\{\Z^{(i)}\}$ from $\Pi_\X$ and average their outputs to estimate $\Delta_\X$, which is used to decide if the
model is certified robust for $\X$.
\paragraph{Training the Base Classifier $f$}
Our method needs to
start with a base classifier $f$.
Although it is possible to train $f$ using standard learning techniques, the result can be improved by considering that the method uses the smoothed $f^\rs$, instead of $f$.
To improve the accuracy of $f^\rs$,
we introduce a data augmentation induced by the perturbation set.
Specifically, at each training iteration,
we first sample a mini-batch of data points (sentences) and randomly perturbing the sentences using the perturbation distribution $\Pi_\X$.
We then apply gradient descent on the model based on the perturbed mini-batch.
Similar training procedures
were also used for Gaussian-based random smoothing on continuous inputs \citep[see e.g.,][]{cohen2019certified}.
Our method can easily leverage powerful pre-trained models such as BERT. In this case, BERT is used to construct feature maps and only the top layer weights are finetuned using the data augmentation method.
| 0e766ec1747bf96e3079d9883f300728af9edb9f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The aim of this note is to discuss a new variational time-discretization scheme
adapted to the structure of General Equations for Non-Equilibrium
Reversible-Irreversible Coupling (GENERIC). Introduced by Grmela \& \"Ottinger,
this formulation provides a unified frame for describing the time evolution of physical systems out of equilibrium in
presence of reversible and irreversible dynamics \cite{Oettinger}.
Let $y$ denote the {\it state} of a closed, nonequilibrium physical system
and let $E(y)$ and $S(y)$ be
the corresponding total {\it energy} and total {\it entropy}, respectively. The GENERIC formulation of the time
evolution of the system reads
\begin{equation}
\label{gen0}
y' = L(y) \,\mathrm{D} E(y) +K(y) \, \mathrm{D} S (y).
\end{equation}
Here, $y'$ denotes the time derivative, $L$ is the antisymmetric {\it
Poisson} operator, and $K$ is the symmetric and positive definite
{\it Onsager} operator (details in Section \ref{sec:gen}).
The gist of the GENERIC system \eqref{gen0} is that conservative and
dissipative dynamics are clearly separated. Under a specific compatibility condition, see
\eqref{nic} below, this entails that a solution $t\mapsto y(t) $ to
\eqref{gen0} conserves energy and accumulates entropy in a
quantifiable way, namely
$$ \frac{\d}{\d t} E(y) = 0 \quad \text{and} \quad \frac{\d}{\d t}
S(y) =\langle \mathrm{D} S(y) , K(y) \mathrm{D} S\rangle\geq 0.$$
The conservation of energy and the quantified dissipative
character are the distinguishing traits of the GENERIC
system \eqref{gen0}.
To replicate these properties at the discrete level leads to
so-called {\it structure-preserving} approximations. These have
drawn interest in the last years, giving rise to different
numerical solutions adapted to different applicative
contexts.
In the last years, GENERIC
has attracted increasing
attention and has been applied to a number of situations ranging from
complex fluids \cite{Grmela}, to dissipative quantum mechanics \cite{Mielke13}, to
thermomechanics \cite{Betsch,Huetter,Huetter2,Huetter3,Mielke11}, to
electromagnetism \cite{Jelic}, to shape-memory alloys \cite{generic_souza},
to the Vlasov-Fokker-Planck \cite{Duong,Hoyuelos,Peletier} and the
Boltzmann equation \cite{Montefusco}, and to large-deviation limits of reversible stochastic processes \cite{Kraaij,Kraaij2}.
Numerical schemes conserving energy
can be found for instance in
\cite{Gonzales,QuMc08,Romero,Romero3,Romero2,SMSF15}; see
\cite{CGMMOOQ12} for a review and \cite{Suzuki} for a contribution
explicitly focusing on GENERIC formulations. These schemes are often {\it discrete-gradient}
methods, where gradients of functionals are specifically modified in
order to fulfill a discrete chain rule and exactly replicate
conservation \cite{Gonzales,Gonzales2,HLW06,MQR99}. In the thermomechanical
context, structure-preserving discretizations either in terms of the absolute
temperature \cite{Conde,Conde2,Garcia,Portillo} or of the internal energy
or the entropy \cite{Betsch,Betsch2} have been obtained. The reader is
referred to
\cite{Krueger} for an approach to open systems.
GENERIC integrators
able to conserve energy and accurately describe entropy accumulation
have been proposed in
\cite{Oettinger2}, where however some limitations are also mentioned. In particular, energy and Onsager
operators have to be suitably modified in a time-step dependent manner in order energy conservation to
hold. Integrators are actually constructed in case of single dissipation
mechanism only (were $K$ be a matrix, it would have to
have rank one) and no convergence theory is provided. The discretization of the
damped harmonic oscillator is addressed in \cite{Oettinger2}, where
nonetheless the temperature is given by a prescribed heat bath. In this case,
explicit solutions have to be used in order to specify the GENERIC integrator.
To the best of our knowledge, all available
numerical schemes directly target GENERIC systems in their
differential form~\eqref{gen0}. Our focus is here on a time-discretization of variational nature instead,
fitting into the general scheme of so-called {\it minimizing
movements} \cite{Ambrosio95,Ambrosio08}. In particular, at all discrete steps we
aim at solving a specific minimization problem. We start by defining
the entropy-production potential
$\Psi(y,\xi) = \langle\xi, K(y)\xi \rangle/2$ and denoting by
$\Psi^*(y,\cdot)$ its conjugate in the second variable. Then,
the GENERIC relation \eqref{gen0} in $[0,T]$ can be equivalently rewritten in
terms of the scalar equation
\begin{equation}-S(y(t)) + \int_0^t \Psi^*(y,y'{-}L(y) \, \mathrm{D} E(y)) + \int_0^t
\Psi(y,\mathrm{D} S(y)) +S(y(0))=0\label{gen01}
\end{equation}
for all $t\in [0,T]$. The
reformulation of dissipative systems in terms of scalar equations
as \eqref{gen01} is usually referred to as {\it De Giorgi's Energy-Dissipation
Principle} \cite{Dondl,Liero,Mielke20}. This principle has already
been applied in a variety of different contexts, including generalized
gradient flows \cite{Bacho,Rossi08}, curves
of maximal slope in metric spaces \cite{Ambrosio08,DeGiorgi}, rate-independent systems
\cite{Mielke12,Roche}, and optimal control \cite{portinale}. For
GENERIC systems, the reformulation of \eqref{gen0} in terms of such
variational principle has been already presented in \cite{Peletier}.
Our new minimizing-movements approach to \eqref{gen0} consists in tackling a
discrete version of relation \eqref{gen01}.
Assume to be
given a time partition $0=t_0<t_1 < \dots<t_N=T$ with steps $\tau_i =
t_i - t_{i-1}$, and an initial state $y^0$. We define the time-discrete trajectory $\{ y_i \}_{i=0}^N$ by letting $y_0=y^0$ and
subsequently perform the minimization
$$\min_y\left\{ -S(y) +\tau_i \Psi^*\left(y, \frac{y- y_{i-1}}{\tau_i} -
L(y) \, \mathrm{D} E(y)\right) + \tau_i \Psi(y,\mathrm{D} S(y)) +
S( y_{i-1})\right\}$$
for $ i=1,\dots, N$. The latter is nothing but a localized and
discretized version of the De Giorgi's Energy-Dissipation Principle
\eqref{gen01}.
The minimizing-movements scheme above has been
introduced in \cite{generic_euler}. The
theory in \cite{generic_euler} is however restricted to the case of state-independent
operators $L$ and $K$, which severely limits the applicability to real GENERIC systems. In addition, the convergence analysis in
\cite{generic_euler} relies on a suitable set of a priori assumptions,
leaving open the discussion whether these can be met in practice.
The aim
of this note is then threefold. First, we extend the reach of the
numerical method to include the case of state dependent operators $L(y)$ and
$K(y)$, hence covering the full extent of the GENERIC theory (Section
\ref{sec:gen}). Our main result is the conditional convergence of
Theorem \ref{thm:cond}. Second, we
provide the detailed analysis of a specific case, the damped
harmonic oscillator, in which the above-mentioned convergence assumptions can
actually be proved to hold (Section \ref{sec:oscillator}). Eventually, we present numerical
experiments assessing the performance of the minimizing-movements scheme and
compare it to the classical
implicit Euler scheme (Subsection \ref{sec:num}).
\section{The minimizing-movements scheme for GENERIC
systems} \label{sec:gen}
In this section, we recall the structure of a GENERIC system \cite{Oettinger} by
specifying the assumptions on functionals and operators that will be
used throughout. Moreover, we formulate our minimizing-movements scheme
and present a conditional convergence result, namely Theorem \ref{thm:cond}.
The GENERIC system
\begin{equation}
\label{gen}
y' = L(y) \,\mathrm{D} E(y) +K(y) \, \partial S (y) \quad \text{a.e. in} \ [0,T]
\end{equation}
is defined by specifying the quintuple
$(Y,E,S,L,K)$. In the following, the state space $Y$ is assumed to be a reflexive Banach
space. The functionals $E$ and $S$ represent the total energy and the total entropy of the
system, respectively. We assume $E$ to be Fr\'echet differentiable, with
strongly-weakly continuous differential $\mathrm{D} E$, and $-S:Y \to
(-\infty,\infty]$ to be proper and lower semicontinuous
with single-valued and strongly-weakly continuous Fr\'echet subdifferential $\partial(-S)$. Recall
that one has $\xi \in \partial(-S)(y)$ iff $ S(y)>-\infty $ and
$$ \liminf_{x\to y}\frac{S(y)-S(x) - \langle \xi , x-y \rangle}{\| x-y
\|_Y}\geq 0.$$
In the following, we also make use of the obvious notation $-\partial S
= \partial (-S)$.
The operators $L$ and $K $ define a {\it
Poisson} and an {\it Onsager} geometric structure on $Y$,
respectively. In particular, for all
states $y\in Y$ we assume that
$L(y)$ and $K(y)$ are linear and continuous from $Y^*$ to
$Y$. Moreover, $L(y)$ is required to be antisymmetric, $L^*(y)=-L(y)$, and to fulfill the Jacobi
identity $\{ \{g_1,g_2\},g_3 \}+\{ \{g_2,g_3\},g_1 \}+\{
\{g_3,g_1\},g_2 \}=0$. Here, $g_i$ denotes any differentiable function on $Y$
and the Poisson bracket is defined as $\{g,\tilde g\}(y)=\langle \mathrm{D} g(y),
L(y)\mathrm{D} \tilde g(y)\rangle $, where $\langle \cdot,\cdot \rangle$ denotes the
duality pairing
between $Y^*$ (dual) and $Y$. Moreover, we assume the strong-weak
continuity
\begin{equation}
y_n
\stackrel{Y}{\to} y \ \ \text{and} \ \ \xi_n \stackrel{Y^*}{\rightharpoonup}
\xi \ \ \Rightarrow \ \ L(y_n)\xi_n \stackrel{Y}{\rightharpoonup} L(y)\xi.\label{eq:contL}
\end{equation}
The mapping $K(y)$ is asked to be symmetric
and positive definite, namely $ K(y)=K^*(y)\geq 0$. We
associate to $K$ the so-called {\it entropy-production potential} $\Psi: Y \times Y^* \to [0,\infty)$ given
by $\Psi(y,\xi) = \langle \xi, K(y)\xi\rangle/2$ and let $\Psi^*$
be its conjugate in the second variable, namely $\Psi^*(y,\eta) = \sup_{\xi}(\langle \xi,
\eta \rangle - \Psi(y,\xi))$. We also assume the lower semicontinuity of the sum of
the entropy-production potential and its dual, that is
\begin{align}&
y_n
\stackrel{Y}{\to} y, \ \ \eta_n \stackrel{Y}{\rightharpoonup}
\eta, \ \ \text{and} \ \ \xi_n \stackrel{Y^*}{\rightharpoonup}
\xi \nonumber \\
&\quad \Rightarrow \ \ \liminf_{n \to\infty} \left(\Psi^*(y_n,\eta_n)+
\Psi(y_n,\xi_n)\right) \geq \Psi^*(y,\eta) + \Psi(y,\xi). \label{eq:contK}
\end{align}
In addition, functionals and operators are asked to satisfy the crucial
{\it noninteraction} condition
\begin{equation}
\label{nic}
L^*(y)\, \partial S(y) = K^*(y)\, \mathrm{D} E(y) = 0 \quad \forall y \in Y.
\end{equation}
This condition ensures that the system of dissipative
actions $K(y)\,\partial S(y)$ does not spoil energy conservation and the system of conservative actions $L(y)\,\mathrm{D} E(y)$
does not contribute to dissipation. Indeed, by
assuming sufficient smoothness one checks that
\begin{align}
\frac{\d}{\d t} E(y) &= \langle \mathrm{D} E(y), y' \rangle \stackrel{\eqref{gen}}{
=}\langle \mathrm{D} E(y), L(y) \mathrm{D} E(y) \rangle + \langle \mathrm{D} E(y), K(y) \partial S(y) \rangle
\nonumber \\
&= 0 + \langle \partial S(y), K^*(y)\mathrm{D} E(y) \rangle\stackrel{\eqref{nic}}{
=} 0, \nonumber \\
\frac{\d}{\d t} S(y) &= \langle \partial S(y), y' \rangle \stackrel{\eqref{gen}}{
=}\langle \partial S(y), L(y) \mathrm{D} E(y) \rangle + \langle \partial S(y), K(y) \partial S(y) \rangle \nonumber
\\
&= \langle \mathrm{D} E(y) , L^*(y)\partial S(y)\rangle + \langle \partial S(y), K(y) \partial S(y) \rangle \nonumber\\
&\stackrel{\eqref{nic}}{
= } \langle \partial S(y), K(y) \partial S(y) \rangle \geq 0.\label{S}
\end{align}
The noninteraction condition \eqref{nic} hence implies that trajectories $y$ solving
\eqref{gen} have constant total energy and entropy rate $\langle \partial
S(y), K(y) \partial S(y) \rangle $. In particular, the entropy is
nondecreasing and entropy production results solely from irreversible processes.
In computing \eqref{S}
we have used the chain rule $(\d/\d t)S(y) = \langle \partial
S(y),y'\rangle$ almost everywhere. This is classical in case $-S$ is (a
regular perturbation of) a convex functional \cite[Prop. 3.3,
p. 73]{Brezis73}. The reader is referred to \cite{Rossi-Savare06} for
a general discussion out of the convex case.
Before moving on, let us remark that the structure of GENERIC is
geometric in nature. Indeed, it is invariant by coordinate changes. Let $y = \phi(\tilde y)$ for
$\tilde y \in \tilde Y$ and define
$\tilde E(\tilde y) = E(\phi( \tilde y))$, $-\tilde S(\tilde y) =
-S(\phi(\tilde y))$, $\tilde L(\tilde y) = \mathrm{D} \phi(\tilde y)^{-1}
L(\phi(\tilde y)) \mathrm{D} \phi(\tilde y)^{-*}$, and $\tilde K(\tilde y) = \mathrm{D} \phi(\tilde y)^{-1}
K(\phi(\tilde y)) \mathrm{D} \phi(\tilde y)^{-*}$, where $\mathrm{D} \phi(\tilde y) ^{-*}:
\tilde Y^* \to Y^*$ is the adjoint of
the inverse of $\mathrm{D} \phi(\tilde y) : \tilde Y \to
Y$. Then, the quintuple $(\tilde Y, \tilde E, \tilde
S, \tilde L,\tilde K)$ satisfies the above structural assumptions and
the GENERIC structure \eqref{gen} can be rewritten as $\tilde y' = \tilde L( \tilde y)\, \mathrm{D} \tilde
E(\tilde y) + \tilde K (\tilde y)\, \partial \tilde S(\tilde y)$.
We now reconsider the discussion leading to \eqref{gen01} and specify it
further by remarking that relation \eqref{gen} is actually
equivalent to the {\it inequality}
\begin{align}&-S(y(t)) + \int_0^t \Psi^*\big(y,y'{-}L(y) \, \mathrm{D}
E(y)\big) + \int_0^t
\Psi\big(y,\partial S(y)\big)+S(y(0))\leq 0 \label{gen1}
\end{align}
for all $t \in [0,T]$.
The equivalence between \eqref{gen} and \eqref{gen1} follows from
Fenchel's relations. Applied to the entropy-production potential
$\Psi$, these relations read
\begin{align}
\Psi^*(y,\eta) + \Psi(y,\xi) \geq \langle \xi,\eta \rangle \quad \forall
y,\,\eta \in Y, \ \xi \in Y^*,
\label{eq:fenchel}\\
\Psi^*(y,\eta) + \Psi(y,\xi) = \langle \xi,\eta \rangle\quad \Leftrightarrow \quad
\xi\in \partial \Psi(y,\eta),
\label{eq:fenchel2}
\end{align}
where the subdifferential is taken in the second variable only.
By noting that $\partial \Psi(y,\partial S(y)) = K(y)\, \partial S(y)$ and using \eqref{eq:fenchel}-\eqref{eq:fenchel2}, one can prove the
equivalences
\begin{align*}
&\eqref{gen} \quad \Leftrightarrow \ \ y' - L(y)\, \mathrm{D} E(y)
= \partial \Psi(y,\partial S(y)) \ \ \text{a.e.}\quad
\nonumber\\
&\quad\stackrel{\eqref{eq:fenchel2}}{\Leftrightarrow}
\Psi^*(y,y'{-}L(y)\, \mathrm{D} E(y))+ \Psi(y,\partial S(y)) - \langle y'{-}L(y)\, \mathrm{D}
E(y), \partial S(y)\rangle \leq 0 \ \ \text{a.e.}\quad
\nonumber\\
&\quad\stackrel{\eqref{nic}}{\Leftrightarrow} \Psi^*(y,y'{-}L(y)\, \mathrm{D} E(y))+ \Psi(y,\partial
S(y)) - \frac{\d}{\d t} S(y) \leq 0 \ \ \text{a.e.} \nonumber\\
&\quad \stackrel{\eqref{eq:fenchel}}{\Leftrightarrow} \quad \eqref{gen1}.
\end{align*}
In particular, the last left-to-right implication follows by
integration while the right-to-left counterpart from the nonnegativity
of the integrand, given \eqref{nic} and \eqref{eq:fenchel}.
The minimizing-movements scheme corresponds to a discretization of
inequality \eqref{gen1}.
To each time partition $0=t_0<t_1<\dots< t_{N}=T$, we associate the
time steps $\tau_i = t_i - t_{i-1}$ and the diameter $\tau = \max
\tau_i$. Given the vector $\{y_i\}_{i=0}^N\in Y^{N+1}$, we introduce the
backward piecewise
constant and piecewise linear interpolations $\overline y:[0,T] \to Y$ and
$\widehat y:[0,T] \to Y$,
\begin{align*}
&\overline y(0)= \widehat y(0), \ \overline y(t)=y_i, \ \widehat y(t) =
\frac{t-t_{i-t}}{\tau_i}y_{i-1}+ \frac{t_i - t}{\tau_i}y_i,\\
&\qquad
\forall t \in (t_{i-1},t_i], \ i =1, \dots, N.
\end{align*}
We define the {\it incremental} functional by
$G : (0,\infty) \times Y\times Y \to (-\infty,\infty]$ as
$$G(\tau,\eta; y) = -S(y) +\tau \Psi^*\left(y, \frac{y-\eta}{\tau} -
L(y) \, \mathrm{D} E(y)\right) + \tau \Psi(y,\partial S(y)) +
S(\eta).$$
By letting $y_0= y^0$ we find the discrete solution
$\{y_i\}_{i=0}^{N}$ by subsequently solving the minimization
problem
\begin{equation}\min_y G(\tau_i,y_{i-1};y ) \quad \text{for} \ i =1, \dots,
N.\label{eq:min}
\end{equation}
Note that, for all $\tau>0$ and $y_{i-1}\in Y$ with $S(y_{i-1})>-\infty$,
the map $y \mapsto G(\tau_i,y_{i-1};y)$ is strongly lower
semicontinuous because of the lower semicontinuity of $-S$, the lower
semicontinuity \eqref{eq:contK} of $\Psi^*+\Psi$, the weak-strong continuity of $\mathrm{D} E $ and $\partial
S$, and the continuity \eqref{eq:contL} of $L$. In order to solve
problem \eqref{eq:min} one has hence to check that $y \mapsto G(\tau_i,y_{i-1};y)$ is strongly
coercive.
The main result of this section is the following {\it conditional} convergence result.
\begin{theorem}[Conditional convergence]\label{thm:cond} Under the above assumptions, let a sequence of partitions $0=t_0^n<\dots<t_{N^n}^n=T$ be
given with $ \tau^n\to 0$ as $n \to \infty$, and let $\{y_i^n\}_{i=1}^{N^n}$ be such that $y_0^n = y^0$. Assume that
\begin{align}
&\sum_{i=1}^{N^n}\big(G(\tau_i^n , y_{i-1}^n; y_i^n)\big)^+ \to 0 \ \
\text{as} \ n \to \infty, \label{eq:G}\\
& \{\widehat y^n\} \ \text{is bounded in} \ H^1(0,T;Y) \ \text{and
takes values in} \ K \subset\subset Y,\label{eq:G2}\\
&\{L(\overline y^n)\,\mathrm{D} E(\overline y^n)\} \ \text{is bounded in} \ L^2(0,T;Y), \label{eq:G3}\\
&\{\partial S(\overline y^n)\} \ \text{is bounded in} \ L^2(0,T;Y^*). \label{eq:G4}
\end{align}
Then, up to a subsequence, $\widehat y^n \rightharpoonup y$ in $H^1(0,T;H)$, where
$y$ is a solution of the \emph{GENERIC} system \eqref{gen} in the sense of inequality \eqref{gen1}.
\end{theorem}
\begin{proof}[Proof of Theorem \ref{thm:cond}]
We infer from the Aubin-Lions lemma \cite[Cor. 7]{Simon87}, upon extracting not relabeled subsequences, that
\begin{align}
&\widehat y^n \rightharpoonup y \ \ \text{in} \ \ H^1(0,T;Y), \nonumber \\
& \overline y^n, \, \widehat y^n \to y \ \ \text{in} \ \ C([0,T];Y), \nonumber \\
& L(\overline y^n)\,\mathrm{D} E(\overline y^n) \rightharpoonup \ell \ \ \text{in} \ \ L^2(0,T;Y), \label{conv1}\\
&\partial S(\overline y^n) \rightharpoonup s \ \ \text{in} \ \ L^2(0,T;Y^*). \label{conv2}
\end{align}
As $\partial S$ is strongly-weakly closed, we have that $s=\partial
S(y)$ almost everywhere. On the other hand, the strong-weak continuity
of $\mathrm{D} E$ and the continuity \eqref{eq:contL} of $L$
imply that $L(\overline y^n)\,\mathrm{D} E(\overline y^n) \to L(y)\, \mathrm{D} E(y)$ pointwise
in time,
so that $\ell = L(y)\, \mathrm{D} E(y)$ almost everywhere.
Fix any $t\in (0,T]$ and let $t^n_m$ be such that $t \in (t^n_{m-1},t^n_m]$.
The discrete sequence of solutions $\{y_i^n\}_{i=0}^{N^n}$ fulfills
\begin{align}
& -S(\overline y^n(t)) + \int_0^{t^n_m} \Psi^*\big(\overline y^n, (\widehat y^n)'- L(\overline
y^n)\, \mathrm{D} E(\overline y^n)\big) + \int_0^{t^n_m} \Psi\big(\overline y^n, \partial
S(\overline y^n)\big) + S(y^0) \nonumber\\
&\quad = \sum_{i=1}^m G(\tau_i^n , y_{i-1}^n; y_i^n) .\label{to_pass}
\end{align}
As $\Psi(y,\cdot)\geq \Psi(y,0)=0$, we conclude that $\Psi^*\geq 0$ as
well. By restricting integrals to the interval $[0,t] \subset
[0,t^n_m]$ equation \eqref{to_pass} implies that
\begin{align}
& -S(\overline y^n(t)) + \int_0^{t} \Psi^*(\overline y^n, (\widehat y^n)'- L(\overline
y^n)\, \mathrm{D} E(\overline y^n)) + \int_0^{t} \Psi(\overline y^n, \partial
S(\overline y^n)) + S(y^0) \nonumber\\
&\quad \leq \sum_{i=1}^{N^n} \big(G(\tau_i^n , y_{i-1}^n; y_i^n)\big)^+. \label{to_pass2}
\end{align}
We now pass to the limit inferior as $n\to \infty$ in relation
\eqref{to_pass2}. By using convergence \eqref{eq:G}, the lower
semicontinuity of $-S$, convergences \eqref{conv1}-\eqref{conv2}, and
the lower semicontinuity \eqref{eq:contK}, we readily obtain that the
limit $y$ fulfills
inequality~\eqref{gen1}.
\end{proof}
The conditional convergence result of Theorem \ref{thm:cond} relies
on the possibility of solving the inequality $G(\tau_i^n , y_{i-1}^n; y_i^n)\leq
0$ up to a small, controllable error, and establishing some a priori
bounds on the discrete solution. The validity of these conditions
has to be checked on the specific problem at hand. In
the coming Section~\ref{sec:oscillator} we give an example of a
situation where \eqref{eq:G}-\eqref{eq:G4} actually hold.
In case $-S$ is convex, an example of $\{y_i^n\}_{i=0}^{N^n}$
fulfilling \eqref{eq:G} are the solutions of the implicit Euler scheme
\begin{equation}\frac{y_i^n-y_i^n}{\tau_i^n} = L(y_i^n) \, \mathrm{D} E(y_i^n) + K(y_i^n) \, \partial S(y_i^n)
\label{eq:euler}
\end{equation}
whenever available. Indeed, such $\{y_i^n\}_{i=0}^{N^n}$ fulfills
\begin{align}
G(\tau_i^n , y_{i-1}^n; y_i^n) & \leq \tau_i^n \Psi^*\left(y_i^n,
\frac{y_i^n-y_i^n}{\tau_i^n} - L(y_i^n) \, \mathrm{D} E(y_i^n) \right) +\tau_i^n
\Psi(y_i^n,\partial S(y_i^n))\nonumber\\
&\quad -\langle \partial S(y_i^n), y_i^n - y_{i-1}^n\rangle =0,\label{eq:euler2}
\end{align}
where the last equality follows from \eqref{eq:fenchel2} and \eqref{eq:euler}.
Note that the incremental minimization problem \eqref{eq:min} may have
no solutions, even if the Euler scheme is solvable. In the purely dissipative case $L=0$, the
existence of a solution to \eqref{eq:min} is ensured as soon as $\tau^*>0$ exists, so that, for all $\tau\in
(0,\tau^*)$ and $y_{i-1}\in Y$ with $S(y_{i-1})>-\infty$, the function $y \mapsto -S(y) +
\tau\Psi(y,(y-y_{i-1})/\tau)$ is strongly coercive. The latter follows whenever
$-S$ has strongly compact sublevels and would imply in
particular that the Euler scheme is also solvable. We anticipate however that the example discussed in
Section~\ref{sec:oscillator} is not purely dissipative and
$-S$, albeit convex, does not have strongly compact sublevels.
Let us mention that, in specific applications, the bounds \eqref{eq:G2}-\eqref{eq:G4}
may follow from \eqref{eq:G}. For instance, this would be the case if
the coercivity
\begin{equation}\Psi^*(y,\eta)+\Psi(y,\xi)\geq c\| \eta \|_Y^2 + c\|
\xi\|^{2}_{Y^*} -\frac1c\label{eq:coer}
\end{equation}
holds for some $c>0$ and all $y,\, \eta\in Y$ and $\xi \in Y^*$, namely in case $K(y)$ is
positive definite and bounded, uniformly with respect to $y$. This
however does not apply to
the example of the damped harmonic oscillator from
Section~\ref{sec:oscillator}, for $K$ is
singular there.
The quadratic nature of the entropy-production potential
could be generalized to the case of $p$-growth with $p>1$ without any
specific intricacy. In particular, one could consider the polynomial
case $\partial \Psi(y,\xi) = K(y) \| \xi\|^{p-2}_{Y^*}\xi$ and
coercivity \eqref{eq:coer} would then read
$$\Psi^*(y,\eta)+\Psi(y,\xi)\geq c\| \eta \|_Y^{p'}+ c\|
\xi\|^{p}_{Y^*} -\frac1c$$
for $1/p+1/p'=1$.
This setting would correspond to the case of doubly-nonlinear GENERIC
dynamics, namely
$$ y' = L(y)\, \mathrm{D} E(y) + K(y) \| \partial
S(y)\|^{p-2}_{Y^*}\partial S(y) .$$
Again, under the noninteraction condition \eqref{nic}, suitably regular
solutions conserve
energy, since $\langle \mathrm{D} E(y), \partial \Psi(y,\partial S(y))\rangle
= \| \partial S(y)\|_{Y^*}^{p-2}\langle \partial S(y), K^*(y)\, \mathrm{D}
E(y)\rangle=0$, and have entropy rate $(\d/\d t)(S(y)) = \| \partial S(y)\|_{Y^*}^{p-2}\langle \partial S(y), K(y)\, \mathrm{D}
S(y)\rangle \geq 0$.
Let us further remark that the conditional convergence result of Theorem \ref{thm:cond}
can serve as an a-posteriori tool to check the convergence
of time-discrete approximations $\{y^n_i\}$, regardless of the method used to
generate them. In particular, relation \eqref{eq:G} can be seen as a
sort of a-posteriori error estimator.
We conclude this section by pointing out that the statement of
Theorem \ref{thm:cond} would hold also for other minimizing-movements
schemes, where nonlinearities are possibly evaluated explicitly. In
particular, instead of $G$ one could consider the functional $\tilde G:
(0,\infty)\times Y \times Y \to (-\infty,\infty]$ defined by
$$
\tilde G(\tau,\eta;y) = -S(y) +\tau \Psi^*\left(y^1, \frac{y-\eta}{\tau} -
L(y^2) \, \mathrm{D} E(y^3)\right) \nonumber
+ \tau \Psi(y^1,\partial S(y)) + S(\eta),
$$
where $y^j$, $j=1,2,3$, are independently chosen to be either $y_i$ or $y_{i-1}$
(or else, for instance $(y_i {+} y_{i-1})/2$). Under conditions
\eqref{eq:G}-\eqref{eq:G4}, convergence would again follow as in
Theorem \ref{thm:cond}.
\section{The minimizing-movements scheme for the damped harmonic oscillator}\label{sec:oscillator}
In this section, we analyze the simplest GENERIC system fulfilling conditions
\eqref{eq:G}-\eqref{eq:G4}. We consider the case of the damped harmonic
oscillator, namely
\begin{align}
&mq''+\nu q' +\kappa q +\lambda \theta =0,\label{eq:d1}\\
&c\theta' = \nu (q')^2 +\lambda q'\theta. \label{eq:d2}
\end{align}
Here, $q\in {\mathbb R}$ represents the position of the harmonic oscillator and $\theta>0$
is its absolute temperature. Note that the case of $\theta$ being the
given, constant
temperature of a surrounding heat bath is considered in
\cite{Oettinger2} instead.
The positive constants $m$, $\nu$, $\kappa$, $\lambda$ and $c$ are
the mass of the oscillator, the viscosity
of the medium, the elastic modulus, the thermal-exchange coefficient, and
the heat capacity, respectively.
By letting $p\in {\mathbb R}$ be the momentum of the harmonic oscillator, we
rewrite \eqref{eq:d1}-\eqref{eq:d2} as the first-order system
\begin{align}
&q' = \frac{p}{m} , \label{eq:o1}\\
&p'=-\frac{\nu p}{m} - \kappa q -\lambda \theta , \label{eq:o2}\\
&\theta' = \frac{\nu p^2}{m^2c} +\frac{\lambda p\theta}{mc} .\label{eq:o3}
\end{align}
System \eqref{eq:o1}-\eqref{eq:o3} can be written in the GENERIC
form \eqref{gen} by letting $y=(q,p,\theta)\in Y = {\mathbb R}^2\times(0,\infty)$ represent the state of the harmonic oscillator and by defining the total
energy and the total entropy as
$$E(q,p,\theta) = \frac{p^2}{2m} + c\theta+\frac{\kappa q^2}{2} \ \
\text{and} \ \
S(q,p,\theta) = -\lambda q + c +c\ln\theta.$$
These definitions comply with the classical Helmholtz relations under
the choice for the free energy
$$\Phi(q,p,\theta) = \frac{\kappa q^2}{2}+\lambda q\theta - c \theta \ln
\theta.$$
In particular, $S =-\partial_\theta \Phi$ and $E = p^2/(2m) + \Phi +
\theta S$.
Note that both $E$ and $S$ are smooth, $-S$ is convex, and
$S(y)>-\infty$ iff $\theta>0$.
The operators $L, \, K : {\mathbb R}^3 \to {\mathbb R}^{3\times 3}$
are given by
\begin{equation*}
L(x)=
\begin{pmatrix}
0 & 1 & 0 \\
-1 & 0 & -\lambda\theta/c \\
0 & \lambda\theta/c & 0
\end{pmatrix}
\ \ \text{and} \ \ \ K(x)=\nu\theta
\begin{pmatrix}
0 & 0 & 0 \\
0 & 1 & -p/(mc) \\
0 & -p/(mc) & p^2/(mc)^2
\end{pmatrix}.
\end{equation*}
One readily checks that $L$ is antisymmetric and a tedious
computation shows that the Jacobi identity holds. On the other hand, $K$ is
symmetric and positive semidefinite, while not being invertible. Note
that $K(y)$ has rank two for all $p\not =0 $, so that the construction
in \cite{Oettinger2} does not apply here.
By computing the gradients
\[
\mathrm{D} E(y)=(\kappa q, p/m,c), \qquad \mathrm{D} S(y)=(-\lambda,0,c/\theta),
\]
one can easily check that the noninteraction condition \eqref{nic}
holds.
Letting $\xi=(\xi_q,\xi_p,\xi_\theta)\in {\mathbb R}^3$ and
$\eta=(\eta_q,\eta_p,\eta_\theta) \in {\mathbb R}^3$, the entropy-production
potential and its dual read
\begin{align*}
&\Psi(y,\xi) = \frac{1}{2}\xi \cdot
K(y)\xi=\frac{\nu\theta}{2}\Bigl(\xi_p-\frac{p}{mc}\xi_\theta\Bigr)^2,\\
&\Psi^*(y,\eta) =
\left\{
\begin{array}{ll}
\displaystyle
\frac{1}{2}\frac{\eta_p^2}{\nu\theta} &\quad\text{if} \
\eta_q =0
\ \text{and} \
p \eta_p + mc
\eta_\theta =0, \\[2mm]
\displaystyle
\infty &\quad\text{otherwise.}
\end{array}
\right.
\end{align*}
With the definitions above, the incremental functional
$G:(0,\infty)\times {\mathbb R}^3 \times {\mathbb R}^3 \to (-\infty,\infty]$ for
$y=(q,p,\theta)$ and $\eta=(\eta_q,\eta_p,\eta_\theta)$ takes the
form
\begin{align*}
&G(\tau,\eta;y)=\lambda q -c\ln \theta
+\frac{\tau}{2}\frac{1}{\nu\theta}\Bigl( \frac{p -
\eta_p}{\tau}+\kappa q+\lambda\theta \Bigr)^2
+\frac{\tau}{2}\nu\theta\Bigr(\frac{p}{m\theta}\Bigl)^2-\lambda
\eta_q
+c\ln\eta_\theta\\[2mm]
& \quad \text{if} \ \ \theta>0,\ \ \displaystyle
\frac{q-\eta_q}{\tau}=\frac{p}{m}, \ \ \text{and} \ \ \displaystyle c \frac{\theta-\eta_\theta}{\tau} = -\frac{p}{m}\frac{p-\eta_p}{\tau}- \frac{\kappa
pq}{m},\\[2mm]
&G(\tau,\eta;y)=\infty \quad \text{otherwise}.
\end{align*}
Assume to be given the time partition
$0=t_0<\dots<t_N=T$ with $\tau_i=t_i - t_{i-1}$ for $i=1,\dots,N$. By letting
$\eta=(q_{i-1},p_{i-1},\theta_{i-1})$ and defining $(q_0,p_0,\theta_0)
= (q^0,p^0,\theta^0)$ for some given initial state $ (q^0,p^0,\theta^0)\in {\mathbb R}^2 \times (0,\infty)$, the incremental minimization
problem \eqref{eq:min} becomes
\begin{align}
\nonumber
&\min_{(q_i,p_i,\theta_i)}\Bigl\{\lambda q_i -c\ln \theta_i
+\frac{\tau_i}{2 \nu\theta_i}\Bigl( \frac{p_i - p_{i-1}}{\tau_i}+\kappa q_i+\lambda\theta_i \Bigr)^2
+\frac{\tau_i\nu p_i^2}{2m\theta_i}\\
&\qquad \qquad -\lambda q_{i-1}+c\ln\theta_{i-1}\Bigr\}\quad \label{eq:min2}\\[2mm]
&\text{under the constraints}\nonumber\\[2mm]
&\qquad \theta_i >0, \label{eq:c0}\\[1mm]
&
\qquad \frac{q_i-q_{i-1}}{\tau_i}=\frac{p_i}{m},\label{eq:c1}\\
&\qquad c\frac{\theta_i-\theta_{i-1}}{\tau_i}=-\frac{\kappa q_i p_i}{m}
-\frac{p_i}{m}\frac{p_i-p_{i-1}}{\tau_i}, \label{eq:c2}\\[1mm]
&\text{for} \ i=1,\dots,N.\nonumber
\end{align}
We devote the remainder of this section to prove that the convergence
result of Theorem \ref{thm:cond} applies to the scheme \eqref{eq:min2}-\eqref{eq:c2},
namely that conditions \eqref{eq:G}-\eqref{eq:G4} are fulfilled. In
particular, we check that the
minimization problem admits a solution $\{y_i\}_{i=0}^N =
\{(q_i,p_i,\theta_i)\}_{i=0}^N $ (Subsection \ref{sec:mini}), that
such solution is unique if time steps are small enough (Subsection \ref{sec:uni}),
that each such solutions fulfills condition
\eqref{eq:G} in the stronger sense $G(\tau_i,y_{i-1};y_i) \leq 0$
(Subsection \ref{sec:eu}), and that the bounds \eqref{eq:G2}-\eqref{eq:G4}
can be established, independently of the partition (Subsection
\ref{sec:bounds}). Eventually, numerical simulations are presented in Subsection \ref{sec:num}.
\subsection{Existence of minimizers}\label{sec:mini}
Assume to be given $\eta=(q_{i-1},p_{i-1},\theta_{i-1})\in {\mathbb R}^2\times(0,\infty).$
As $y \mapsto G(\tau_i,\eta;y)$ is smooth on its domain, in order to
prove that the scheme \eqref{eq:min2}-\eqref{eq:c2} admits a solution,
we just need to check coercivity, namely that sublevels of $y \mapsto
G(\tau_i,\eta;y)$ are bounded.
By using the constraints \eqref{eq:c1}-\eqref{eq:c2}, we can express
$q_i$ and $\theta_i$ in terms of $p_i$ and the values
$(q_{i-1},p_{i-1}, \theta_{i-1})$
in the form
\begin{equation}
q_i = \tau_i \frac{p_i}{m} + q_{i-1} \ \ \text{and} \ \ \theta_i=f(p_i),\label{eq:redefine}
\end{equation}
where the function $f$ is defined by
\[
f(p_i) = \theta_{i-1}-\frac{\tau_i^2\kappa}{m^2c}p_i^2-\frac{\tau_i\kappa}{mc}p_i q_{i-1}-\frac{1}{mc}p_i^2+\frac{1}{mc}p_i p_{i-1}.
\]
By substituting the two expressions in the minimum problem
\eqref{eq:min2}-\eqref{eq:c2}, we can reduce it to a minimization in the variable $p_i$
only. Indeed, problem \eqref{eq:min2}-\eqref{eq:c2} is equivalent to
\begin{equation}
\label{eq:min3}
\min_{p_i} F(p_i) \ \ \text{under the constraint} \ \ f(p_i)>0,
\end{equation}
where we have defined
\begin{align}
\nonumber
F(p_i) &:= \frac{\tau_i\lambda p_i}{m
-c\ln f(p_i) +\frac{\tau_i}{2\nu f(p_i)}\Bigl(
\frac{p_i-p_{i-1}}{\tau_i}+\frac{\tau_i\kappa}{m} p_i+\kappa q_{i-1}
+\lambda f(p_i)\Bigr)^2\nonumber\\
&\phantom{:xx} + \frac{\tau_i\nu p_i^2}{2m f(p_i)}
+c\ln\theta_{i-1}. \nonumber
\end{align}
Again, as $F$ is smooth on its domain, in order to solve the
minimization problem \eqref{eq:min3} we just need to prove that $F$ has bounded sublevels. This
follows from the fact that $F(p_i)\to \infty$ if $f(p_i)\to 0^+$ and
that $f(p_i)>0$ iff $p_i$ belongs to a bounded interval, depending on
$(q_{i-1},p_{i-1},\theta_{i-1})$. In fact, as $f$ is quadratic and
$\theta_{i-1}>0$, it follows that $f(p_i)>0$ if and only if
$|p_i - \widehat p|<\widehat r$ with
\begin{align*}
&\widehat p = \frac{p_{i-1}-\tau_i \kappa q_{i-1}}{2+2\tau_i^2 \kappa/m
} \ \ \text{and} \ \
\widehat r=\frac{\left({(p_{i-1}{-}\tau_i\kappa
q_{i-1})^2} +
4(1{+}\tau_i^2\kappa/m)mc\theta_{i-1}\right)^{1/2} }{2+2\tau_i^2\kappa/m}
.
\end{align*}
Hence, for all given $(q_{i-1},p_{i-1},\theta_{i-1})\in{\mathbb R}^2\times
(0,\infty)$, we can find a solution $p_i$ of \eqref{eq:min3}.
We conclude from \eqref{eq:redefine} that
$(q_i,p_i,\theta_i)$ solves \eqref{eq:min2}-\eqref{eq:c2}. In particular, we have
that $\theta_i>0$.
\subsection{Uniqueness of minimizers}\label{sec:uni}
Let us remark that the minimizers
$(q_i,p_i,\theta_i)$ of \eqref{eq:min2}-\eqref{eq:c2} need not be unique in general. Still, the
function $F$ is strictly convex for $\tau_i$
small enough, depending on the values $(q_{i-1},p_{i-1},\theta_{i-1})$
and on material parameters.
\newpage
Let us start by rewriting $F$ as
\begin{align}
F(p_i)&= -c \ln f(p_i) + \frac{(1+\tau_i^2
\kappa/m)^2 }{2\tau_i \nu }\frac{ (p_i+h_i)^2}{ f(p_i)}
\nonumber\\
&\quad + \frac{\tau_i \nu}{2m}\frac{p_i^2}{f(p_i)}
+ \frac{\tau_i\lambda^2f(p_i)}{2\nu}+ r_i(p_i), \label{todiff}
\end{align}
where $h_i =(-p_{i-1}+\tau_i \kappa q_{i-1})/(1+ \tau_i^2 \kappa/m)$ and $r_i$
is the affine function
$$r_i(p_i)= \frac{\tau_i \lambda p_i}{m}
+ \frac{\lambda (1+ \tau_i^2 \kappa/m)}{\nu }(p_i + h_i)
+ c \ln \theta_{i-1}.$$
Using the following facts
\begin{align*}
\left(\frac{(p_i+h_i)^2}{f(p_i)}\right)'' &= \frac{\big(\big(2
f^3(p_i)\big)^{1/2} -\big(2
f(p_i)\big)^{1/2}(p_i+h_i) f'(p_i)\big)^2 }{ f^4(p_i)}\\
& \,\quad - \frac{ (p_i+h_i)^2
f^2(p_i)f''(p_i)}{ f^4(p_i)}> 0,\\
\left(\frac{p_i^2}{f(p_i)}\right)'' &= \frac{\big(\big(2
f^3(p_i)\big)^{1/2} -\big(2
f(p_i)\big)^{1/2} p_i f'(p_i)\big)^2 }{ f^4(p_i)}\\
& \,\quad - \frac{ p_i ^2
f^2(p_i)f''(p_i)}{ f^4(p_i)}> 0,
\end{align*}
and differentiating $F$ in \eqref{todiff} twice, we obtain
\begin{align*}
F''(p_i) &> -c (\ln f(p_i))'' +\frac{\tau_i
\lambda^2}{2\nu}f''(p_i) = \frac{c(f'(p_i))^2}{f^2(p_i)} +
f''(p_i)\left(- \frac{c}{f(p_i)} + \frac{\tau_i
\lambda^2}{2\nu}\right)\\
&\geq f''(p_i)\left(- \frac{c}{f(p_i)} + \frac{\tau_i
\lambda^2}{2\nu}\right)=:g(p_i).
\end{align*}
As $f''(p_i) =
-2(1+\tau_i^2 \kappa/m)/(mc)<0$, the function $g$ is nonnegative as long
as $f(p_i)\leq 2\nu c/(\tau_i \lambda^2)$. It hence suffices to choose
$\tau_i$ so small that
\begin{equation}
\label{eq:small}
\max f \leq \frac{2\nu c}{\tau_i \lambda^2}.
\end{equation}
This condition depends on the size of the time-step
and on material parameters and
implies that $F$ is strictly convex, thus admitting a unique minimizer
$p_i$. Consequently, problem \eqref{eq:min2}-\eqref{eq:c2} has a unique
minimizer $(q_i,p_i,\theta_i)$.
We can make the time-step dependent bound \eqref{eq:small} on $\tau_i$
more explicit by directly computing the
maximum of $f$:
\begin{align*}
\max f &= \frac{\left(\displaystyle \frac{-\tau_i\kappa q_{i-1}}{mc} +
\frac{p_{i-1}}{mc}\right)^2+4\left( \displaystyle \frac{\tau_i^2\kappa}{m^2c} +
\frac{1}{mc}\right) \theta_{i-1}}{4\left(\displaystyle
\frac{\tau_i^2\kappa}{m^2c} + \frac{1}{mc}\right)}\\
&=
\frac{(-\tau_i \kappa q_{i-1} + p_{i-1})^2}{4(\tau_i^2 \kappa c + mc)} + \theta_{i-1}
\leq
\frac{p_{i-1}^2}{2mc} + \frac{\tau_i^2 \kappa^2 q_{i-1}^2}{2mc} +
\theta_{i-1}.
\end{align*}
In Subsection \ref{sec:bounds} we will check that above right-hand
side can be
bounded by a positive constant $c_0$ in terms of the initial data and of material parameters only, uniformly with
respect to $i$, see \eqref{eq:bound1}.
Condition
\eqref{eq:small} can hence be strengthened by asking the partition to
be such that
\begin{equation} \label{eq:small2}
c_0 \leq \min_i\frac{2\nu c}{\tau_i
\lambda^2}.
\end{equation}
This stronger condition ensures the
unique solvability of the minimization problem \eqref{eq:min2}-\eqref{eq:c2} for all $i=1,\dots,N$.
\subsection{Minimizers fulfill condition \eqref{eq:G}}\label{sec:eu}
Following the discussion of Section \ref{sec:gen}, the convexity of
$-S$ ensures that
condition \eqref{eq:G} holds as soon as one can prove that the
implicit Euler scheme \eqref{eq:euler} is solvable. In the current
setting, the implicit Euler scheme reads
\begin{align}
&\displaystyle
\frac{q_i-q_{i-1}}{\tau_i}=\frac{p_i}{m}, \label{eq:Eimp1}\\[1mm]
& \displaystyle
\frac{p_i-p_{i-1}}{\tau_i}=-\frac{\nu p_i}{m}-\kappa q_i
-\lambda \theta_i, \label{eq:Eimp2}\\[1mm]
& \displaystyle
\frac{\theta_i-\theta_{i-1}}{\tau_i}=\frac{\nu
p_i^2}{m^2c}+\frac{\lambda p_i\theta_i}{mc} \label{eq:Eimp3}
\end{align}
for $i=1,\dots,N$.
Let the initial data $(q^0,p^0,\theta^0)\in {\mathbb R}^2 \times (0,\infty)$
be given and assume to have solved \eqref{eq:Eimp1}-\eqref{eq:Eimp3} up to
level $i-1$. In particular, let
$(q_{i-1},p_{i-1},\theta_{i-1}) \in {\mathbb R}^2 \times (0,\infty)$.
Relation \eqref{eq:Eimp1} can be written as
\[
q_i=\frac{\tau_i p_i}{m}+q_{i-1}
\]
and we can substitute it into \eqref{eq:Eimp2} obtaining
\begin{equation}
p_i = \frac{-\tau_im\lambda}{m+\tau_i\nu+\tau^2_i\kappa} \theta_i +
\frac{-\tau_im\kappa
q_{i-1}+mp_{i-1}}{m+\tau_i\nu+\tau^2_i\kappa}
=:\alpha_i\theta_i+\beta_i \label{eq:pi}
\end{equation}
where $\alpha_i$ and $\beta_i$ depend just on material parameters, $\tau_i$, and on $q_{i-1}$
and $p_{i-1}$.
We may hence rewrite equation \eqref{eq:Eimp3} as
\begin{align*}
\theta_i &=\theta_{i-1}+\frac{\tau_i\nu p_i^2}{m^2c} +\frac{\tau_i \lambda p_i\theta_i}{mc}\\
&=\theta_{i-1}+\frac{\tau_i\nu}{m^2c}\Bigl(\alpha^2_i\theta_i^2 +2\alpha_i\beta_i\theta_i +\beta^2_i\Bigr) +\frac{\tau_i\lambda}{mc}\Bigl(\alpha_i\theta_i^2 +\beta_i\theta_i\Bigr),
\end{align*}
which gives the expression
\begin{align}
0&=\Bigl(\frac{\tau_i \nu \alpha_i^2}{m^2c} + \frac{\tau_i \lambda
\alpha_i}{mc} \Bigr)\theta_i^2 + \Bigl(\frac{2\tau_i \nu
\alpha_i\beta_i}{m^2c}+ \frac{\tau_i \lambda
\beta_i}{mc}-1\Bigr)\theta_i + \Bigl( \theta_{i-1}+\frac{\tau_i \nu
\beta_i^2}{m^2c}\Bigl)\nonumber\\
&=:\gamma_i\theta_i^2 + \delta_i\theta_i +\epsilon_i, \label{eq:zero}
\end{align}
with $\gamma_i$, $\delta_i$, and $\epsilon_i$ depending on $\alpha_i$,
$\beta_i$, $\theta_{i-1}$, and data. Note that
$$
\gamma_i =
\frac{\tau_i}{mc}\bigg(\frac{\nu\alpha_i}{m}+\lambda\bigg)\alpha_i =
- \frac{\tau_i^2\lambda^2}{c} \frac{m+\tau_i^2\kappa}{(m+\tau_i\nu+\tau_i^2\kappa)^2}<0.
$$
We infer from $\theta_{i-1}>0$ that $\epsilon_i>0$. Hence, the
second-order equation \eqref{eq:zero} has a unique solution
$\theta_i>0$. This uniquely defines $p_i$ and $q_i$ via \eqref{eq:pi}
and \eqref{eq:Eimp1}, respectively. In particular, the implicit Euler
scheme \eqref{eq:Eimp1}-\eqref{eq:Eimp3} has a unique solution $y^e_i
=(q_i,p_i,\theta_i)$.
Let now $y_i $ be a solution to the incremental minimization
problem \eqref{eq:min2}. Owing to \eqref{eq:euler2}, we conclude that
$$
G(\tau_i,y_{i-1};y_i) = \min_y G(\tau_i,y_{i-1};y) \leq
G(\tau_i,y_{i-1};y_i^e) \stackrel{\eqref{eq:euler2}}{=}0,
$$
where we have also used the fact that $y^e_i $ fulfills the
constraints \eqref{eq:c0}-\eqref{eq:c2}. In particular, condition \eqref{eq:G}
holds in the even stronger form
\begin{equation} G(\tau_i,y_{i-1};y_i) \leq 0 \quad \text{for} \ \
i=1,\dots,N. \label{eq:stronger}
\end{equation}
\subsection{A priori bounds}\label{sec:bounds}
We now prove that condition \eqref{eq:G2}-\eqref{eq:G4} of Theorem
\ref{thm:cond} hold. This will follow by checking a priori bounds on minimizers $\{y_i\}_{i=0}^N$ of
\eqref{eq:min2}, independently from the time partition.
Let us start by remarking that
the energy is nonincreasing. By \eqref{eq:c1}, one can equivalently rewrite the constraint \eqref{eq:c2}
as
\begin{align*}
c\theta_i-c\theta_{i-1} =-\frac{\kappa q_i^2}{2}+\frac{\kappa
q_{i-1}^2}{2} -\frac{\kappa}{2}|q_i-q_{i-1}|^2 -\frac{p_i^2}{2m} +\frac{p^2_{i-1}}{2m}-\frac{1}{2m}|p_i-p_{i-1}|^2.
\end{align*}
A rearrangement of the terms leads to
$$E(y_i) + \frac{\kappa}{2}|q_i-q_{i-1}|^2+\frac{1}{2m}|p_i-p_{i-1}|^2 =
E(y_{i-1}).$$
Hence, $ E(y_i)$ is nonincreasing and $E(y_i) =E(y_{i-1})$ iff
$q_i=q_{i-1}$ and $p_i=p_{i-1}$. By taking the sum for $i=1,\dots,j$ for
$j\leq N$, we find that
\begin{equation}
\label{eq:energy}
E(y_j) + D^q_j +D^p_j=E(y^0) \quad \forall j =1,\dots,N,
\end{equation}
where we have set
$$D^q_j
=\sum_{i=1}^j\frac{\kappa}{2}|q_i-q_{i-1}|^2 \ \ \text{and} \ \ D^p_j
=\sum_{i=1}^j\frac{1}{2m}|p_i-p_{i-1}|^2.$$
The nonnegative terms
$D^q_j$ and $D^p_j$ exactly quantify the energy dissipativity of the
scheme. Owing to \eqref{eq:energy}, we obtain the uniform bound
\begin{equation}
|p_i| + |\theta_i| + |q_i|\leq
C \quad \forall i=1,\dots,N,\label{eq:bound1}
\end{equation}
where, here and in the following, the symbol $C$ denotes a generic
positive constant depending
on the initial data $y^0$ and on material
parameters, but not on the time partition. Bound \eqref{eq:bound1} and constraint
\eqref{eq:c1} imply that
\begin{equation}
\left|\frac{q_i-q_{i-1}}{\tau_i}\right|\leq
C \quad \forall i=1,\dots,N.\label{eq:bound2}
\end{equation}
Moreover, we readily check that
\begin{equation}
|L(y_i)\, \mathrm{D} E(y_i)|= |(p_i/m,-kq_i-\lambda \theta_i,\lambda
p_i\theta_i/m)| \stackrel{\eqref{eq:bound1}}{\leq} C \quad \forall i=1,\dots,N.\label{eq:bound20}
\end{equation}
Let us take the sum for $i=1,\dots,j$ for $j \leq N$ in \eqref{eq:stronger} getting
$$ - S(y_j) + \psi^*_j+ \psi_j \leq -S(y^0),$$
where
$$\psi^*_j = \sum_{i=1}^j \frac{\tau_i}{2 \nu\theta_i}\Bigl( \frac{p_i - p_{i-1}}{\tau_i}+\kappa q_i+\lambda\theta_i \Bigr)^2
\ \ \text{and} \ \ \psi_j=\sum_{i=1}^j \frac{\tau_i\nu
p_i^2}{2m\theta_i}.$$
As $\psi^*_j$ and $\psi_j$ are nonnegative, we infer that $ S(y_i)$
is nondecreasing. In particular,
$$ - c\ln\theta_j \leq -S(y^0) - \lambda q_j +c
\stackrel{\eqref{eq:bound1}}{\leq }C \quad \forall j=1,\dots,N. $$
and consequently,
\begin{equation}
\label{eq:bound3}
\theta_i \geq \theta_{\rm min}>0 \quad \forall i=1,\dots,N,
\end{equation}
for some given $\theta_{\rm min}$ depending just on $y^0$ and the material
parameters. This ensures that
\begin{equation}
\label{eq:bound30}
|\mathrm{D} S(y_i)| = |(-\lambda,0,c/\theta_i)| \leq C \quad \forall i=1,\dots,N.
\end{equation}
It follows from the bound \eqref{eq:bound1} on $q_i$ and
$\theta_i$ that $S(y_i) =
-\lambda q_i + c + c \ln \theta_i \leq
C$ for all $i=1,\dots,N$. We conclude that $\psi_N \leq
-S(y^0)+S(y_N)\leq C$ and
\begin{align}
&\sum_{i=1}^N \frac{\tau_i}{2\nu
\theta_i} \left|\frac{p_i - p_{i-1}}{\tau_i} \right|^2 \nonumber\\
&\quad= \psi_N^* - \sum_{i=1}^N \frac{\tau_i}{2\nu
\theta_i} \left(\kappa q_i +\lambda\theta_i\right)^2 - 2 \sum_{i=1}^N \frac{\tau_i}{2\nu
\theta_i} \left(\frac{p_i - p_{i-1}}{\tau_i} \right) \left(\kappa
q_i +\lambda\theta_i\right)\nonumber\\
&\quad\leq \psi_N^* + \sum_{i=1}^N \frac{\tau_i}{2\nu
\theta_i} \left(\kappa q_i +\lambda\theta_i\right)^2
+\sum_{i=1}^N\frac{\tau_i}{4\nu \theta_i} \left|\frac{p_i - p_{i-1}}{\tau_i}
\right|^2 \nonumber\\
&\quad
\stackrel{\eqref{eq:bound1}}{\leq} C +\sum_{i=1}^N \frac{\tau_i}{4\nu
\theta_i} \left|\frac{p_i - p_{i-1}}{\tau_i} \right|^2.\label{eq:pp}
\end{align}
Using the fact that $\theta_i$ is
uniformly bounded by \eqref{eq:bound1}, we can hence bound
\begin{align}
\sum_{i=1}^N \tau_i \left|\frac{p_i - p_{i-1}}{\tau_i} \right|^2
\leq 4\nu \,(\max_i\theta_i) \sum_{i=1}^N \frac{\tau_i}{4\nu
\theta_i} \left|\frac{p_i - p_{i-1}}{\tau_i} \right|^2
\stackrel{\eqref{eq:bound1}+\eqref{eq:pp}}{\leq} C.\label{eq:bound4}
\end{align}
Eventually, we infer from constraint \eqref{eq:c2} that
\begin{equation}
\label{eq:bound5}
\sum_{i=1}^N \tau_i \left|\frac{\theta_i - \theta_{i-1}}{\tau_i}
\right|^2 \leq 2\sum_{i=1}^N \tau_i \left(\frac{\kappa^2 q_i^2
p_i^2}{m^2c^2} + \frac{p_i^2}{m^2c^2}\left| \frac{p_i-p_{i-1}}{\tau_i}\right|^2 \right)\leq C.
\end{equation}
Bounds \eqref{eq:bound1}-\eqref{eq:bound2} and
\eqref{eq:bound4}-\eqref{eq:bound5} imply condition
\eqref{eq:G2} from Theorem \ref{thm:cond}. On the other hand, bounds
\eqref{eq:bound20} and \eqref{eq:bound30} imply conditions
\eqref{eq:G3} and \eqref{eq:G4}, respectively. Hence, the convergence
statement of Theorem
\ref{thm:cond} holds. In particular, given initial values
$(q^0,p^0,\theta^0)\in {\mathbb R}^3$ with $\theta^0>0$, a
sequence of partitions $0=t_0^n<\dots<t^n_{N^n}=T$ with $\tau^n=\max_i( t^n_i
- t^n_{i-1})\to 0$, and corresponding minimizers $\{(q_i^n,p_i^n,
\theta_i^n)\}_{i=0}^N$ of \eqref{eq:min2}, it follows that
\begin{equation}
(\widehat q^n,\widehat p^n,\widehat \theta^n) \rightharpoonup (q,p,\theta) \ \
\text{in} \ \ H^1(0,T;{\mathbb R}^3),\label{eq:conf}
\end{equation}
where $ (q,p,\theta)$ solves the damped harmonic oscillator system
\eqref{eq:o1}-\eqref{eq:o3} with initial value $
(q(0),p(0),\theta(0))= (q^0,p^0,\theta^0)$. Note that solutions of
\eqref{eq:o1}-\eqref{eq:o3} are
unique. Hence, convergence \eqref{eq:conf}
holds for the whole sequence of
discrete solutions, not just for a subsequence.
As $E$ and $S$ are smooth, $\overline y^n$ is bounded, and $\theta_i \geq \theta_{\rm min}>0$, we find that
$E(\overline y(\cdot)) \to E(y(\cdot))\equiv E(y^0)$ and $ S(\overline y(\cdot))
\to S(y(\cdot))$ uniformly, together with their time derivatives of all orders.
More precisely, upon remarking that
\begin{align*}
&D_N^q \leq \tau^n\sum_{i=1}^N\frac{\tau_i\kappa}{2} \left|\frac{q_i -
q_{i-1}}{\tau_i} \right|^2\stackrel{\eqref{eq:bound2}}{\leq}
C\tau^n, \\
&D_N^p \leq \tau^n\sum_{i=1}^N\frac{\tau_i}{2m} \left|\frac{p_i -
p_{i-1}}{\tau_i} \right|^2\stackrel{\eqref{eq:bound4}}{\leq}
C\tau^n,
\end{align*}
we may use \eqref{eq:energy} and check that
$$ -C\tau^n + E(\overline y^n(t)) \leq E^0 \quad \forall t \in [0,T].$$
In particular, we control the energy dissipation as follows:
\begin{equation}
\max_{t\in [0,T]}\| E(\overline y^n(t)) - E(y^0)\|_{L^\infty(0,T)} \leq C\tau^n.
\label{errore}
\end{equation}
\subsection{Numerical tests}\label{sec:num}
We record here some numerical evidence on the performance of the
minimizing-movements scheme \eqref{eq:min2}-\eqref{eq:c2}. All computations are
performed in Matlab. In the following, we choose
\begin{equation}
m=\nu=\kappa=\lambda=c=1, \quad y^0 = (q^0,p^0,\theta^0)= (1,1,1),
\quad T=15.\label{data}
\end{equation}
Given a
uniform time-partition with time step $\tau=T/N$, we find a
solution $\{y_i\}_{i=0}^N$ of the minimizing-movements scheme
\eqref{eq:min2}-\eqref{eq:c2} via Newton's method. Note that, by
choosing \eqref{data} and recalling the bound \eqref{eq:energy}, one
has that the constant $c_0$ in \eqref{eq:small2} can be bounded from above
by $E(y^0)=2$, for all $\tau\leq 1$. In particular, condition
\eqref{eq:small2} is fulfilled by all $\tau\leq 1$ and the solution of
the minimizing-movements scheme is unique.
The unique solution $\{y^e_i\}_{i=0}^N$ of the
implicit Euler scheme \eqref{eq:Eimp1}-\eqref{eq:Eimp3} is obtained
directly from \eqref{eq:pi}-\eqref{eq:zero}. Eventually, the numerical
reference
solution $t\mapsto y(t)$ is calculated by means of the Matlab solver
ode45 choosing $10^{-4}$ for the maximal time step and $10^{-8}$ for the absolute tolerance.
Figures~\ref{q1}-\ref{entropy1} illustrate the
numerical
reference solution and the time-discrete solutions for
$\tau=1/4$. Both the minimizing-movements and the Euler scheme
dissipate energy, see Figure~\ref{entropy1} left. This is of
course an undesired effect, which is however attenuated as $\tau \to
0$, see \eqref{errore}. Energy dissipation seems to be more pronounced
for the Euler scheme. On the other hand, entropy is nondecreasing for
the minimizing-movements scheme whereas the Euler scheme shows a nonmonotone
entropy, which is nonphysical, see Figure~\ref{entropy1} right.
\begin{figure}[ht]
\centering
\includegraphics[width=60mm]{figures/q1.eps}
\includegraphics[width=60mm]{figures/p1.eps}
\caption{Position $q$ (left) and momentum $p$ (right) with respect to time for the numerical reference solution (red, dotted),
the minimizing-movements scheme (solid), and the Euler scheme (blue, dash-dotted).}
\label{q1}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=60mm]{figures/theta1.eps}
\caption{Temperature with respect to time for the numerical reference solution (red, dotted),
the minimizing-movements scheme (solid), and the Euler scheme (blue, dash-dotted).}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=60mm]{figures/energy1.eps}
\includegraphics[width=60mm]{figures/entropy1.eps}
\caption{Energy (left) and entropy (right) as function of time for the numerical reference solution (red dotted),
the minimizing-movements scheme (solid), and the Euler scheme (blue, dash-dotted).}
\label{entropy1}
\end{figure}
A second set of experiments is illustrated in Figure \ref{error}. For
the same choices in \eqref{data} and different time steps
\begin{equation}\label{times}
\tau^n = 2^{-n} \quad \text{for} \ \ n =-1,0,\dots,11,
\end{equation}
we
compute the uniform errors of the temperature component of the solution and of the energy with
respect to the numerical reference solution.
\begin{figure}[ht]
\centering
\includegraphics[width=60mm]{figures/error_theta.eps}
\includegraphics[width=60mm]{figures/error_energy.eps}
\caption{ Errors with respect to
$1/\tau^n$ from \eqref{times}, logarithmic scales for the variable
$\theta$ (on the left) and the energy (on the right). The (green) dotted lines indicate the order of convergence $1$.
Left: $\max_{t\in [0,T]}|\theta(t)-\widehat \theta (t)|$
(solid), $\max_{t\in [0,T]}|\theta (t)-\widehat \theta^{\,e}(t)|$ (blue, dash-dotted); right: $\max_{t\in [0,T]}| E(\widehat
y(t)) - E(y^0)|$ (solid), and $\max_{t\in [0,T]}| E(\widehat
y^{\,e}(t)) - E(y^0)|$ (blue, dash-dotted). }
\label{error}
\end{figure}
As $\tau$ converges to $0$, our computations confirm that both the minimizing-move\-ments and the Euler
scheme are of order $\tau$, see also \eqref{errore}. Let us mention
that a proof of first-order convergence for the minimizing-movements scheme in the nondissipative
regime ($L=0$, $K$ independent of the state) is given in \cite[Prop.\ 4.3]{generic_euler}. Both schemes do not conserve
energy. Still, the minimizing-movements scheme is more accurate than
the Euler scheme when the time steps are large.
\section*{Acknowledgements}
This research is supported by Austrian Science Fund (FWF) project
F\,65 and W\,1245.
AJ is supported by the FWF projects P\,30000 and P\,33010.
US is supported by the FWF projects I\,4354 and P\,32788 and by the
Vienna Science and Technology Fund (WWTF) through Project
MA14-009.
| 3791de1e3a5acfaa88794e8b658aade99244787f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\setcounter{theorem}{0}
We saw in Chapter~1 that underlying every inverse semigroup is a groupoid,
but this groupoid is not enough on its own to recover the original inverse semigroup multiplication.
This raises the question of what else is needed, and we answer this question in Section~2.
The underlying groupoid of an inverse semigroup combined with the natural partial order lead to a structure called an inductive groupoid
and inverse semigroups and inductive groupoids are two ways of looking at the same thing.
In fact, Ehresmann worked with inductive groupoids rather than inverse semigroups.
Our second categorical description takes its cue from how substructures are represented in a category.
This leads to left (or right) cancellative categories and their actions on principal groupoids
as a way of constructing arbitrary inverse semigroups.
This is described in Section~4.
Section~3 forms a bridge between Sections~2 and 4.
In it, we describe the extent to which ordered groupoids are related to left cancellative categories.
In the case of inverse monoids, this leads to a complete description in terms of categories.
\section{Ordered groupoids}\setcounter{theorem}{0}
The motivation for this construction is described in Section~2.1 and the main theorem,
the Ehresmann-Schein-Nambooripad theorem, is proved in Section~2.3.
\subsection{Motivation}
The following result shows how the usual product in an inverse semigroup can be reconstructed from the restricted product and the natural partial order.
\begin{lemma} Let $S$ be an inverse semigroup.
\begin{enumerate}
\item Let $s \in S$ and $e$ an idempotent
such that $e \leq s^{-1}s$.
Then $a = se$ is the unique element in $S$ such that
$a \leq s$ and $a^{-1}a = e$.
\item Let $s \in S$ and $e$ an idempotent
such that $e \leq ss^{-1}$.
Then $a = es$ is the unique element in $S$ such that $a \leq s$
and $aa^{-1} = e$.
\item Let $s,t \in S$.
Then $st = s' \cdot t'$ where $s' = se$, $t' = et$
and $e = s^{-1}stt^{-1}$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) From the definition of the natural partial order
we have that $a \leq s$.
Also, $a^{-1}a = (se)^{-1}se = es^{-1}se = e$.
Now let $b \leq s$ be such that $b^{-1}b = e$.
Then $b = sb^{-1}b$, so that $b = se = a$.
(2) Similar to (1).
(3) Put $s' = se$ and $t' = et$ where $e = s^{-1}stt^{-1}$.
Then $s' \leq s$ and $t' \leq t$.
It is easy to check that
${\bf d}(s') = e$ and ${\bf r}(t') = e$.
Thus $s' \cdot t'$ exists.
But $s'\cdot t' = set = st$.
\end{proof}
A function $\theta \colon \: S \rightarrow T$ between inverse
semigroups is said to be a
{\em prehomomorphism}
if $\theta (st) \leq \theta (s)\theta (t)$
for all $s,t \in S$.\footnote{The function is called a
{\em dual prehomomorphism}
if
$\theta (s)\theta (t) \leq \theta (st)$.}
Inverse semigroups and prehomomorphisms form a category
that contains the usual category of inverse semigroups and homomorphisms.
We can easily construct examples of prehomomorphisms
which are not homomorphisms.
Let $L$ and $M$ be meet semilattices and
let $\theta \colon \: L \rightarrow M$ be an order-preserving function.
Let $e,f \in L$.
Then $e \wedge f \leq e,f$ and so
$\theta (e \wedge f) \leq \theta (e),\theta (f)$ since
$\theta$ is order-preserving.
Thus
$$\theta (e \wedge f) \leq \theta (e) \wedge \theta (f)$$
since $M$ is a meet semilattice.
It follows that $\theta$ is a prehomomorphism
from the inverse semigroup $(L,\wedge)$ to the
inverse semigroup $(M,\wedge)$,
but not in general a homomorphism.
\begin{lemma} Let $\theta \colon \: S \rightarrow T$ be a function between inverse semigroups.
\begin{enumerate}
\item $\theta$ is a prehomomorphism if, and only if, it preserves the restricted product and the natural partial order.
\item $\theta$ is a homomorphism if, and only if, it is a prehomomorphism which satisfies $\theta (ef) = \theta (e)\theta (f)$
for all idempotents $e,f \in S$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) Let $\theta \colon \: S \rightarrow T$ be a prehomomorphism.
We first prove that
$\theta (s^{-1}) = \theta (s)^{-1}$ for each $s \in S$.
By definition
$$\theta (s)
= \theta (s(s^{-1}s)) \leq \theta (s)\theta (s^{-1}s)
\leq \theta (s)\theta (s^{-1})\theta (s).$$
Similarly,
$$\theta (s^{-1}) \leq \theta (s^{-1})\theta (s)\theta (s^{-1}).$$
Put $a = \theta (s)$ and $b = \theta (s^{-1})$.
Then $a \leq aba$ and $b \leq bab$.
Now $ab \leq abab$, so that
$$ab = (ab)^{2}(ab)^{-1}(ab) = (ab)^{2}.$$
Similarly, $ba = (ba)^{2}$.
Thus $a(ba) \leq a \leq aba$, and so $a = aba$.
Similarly, $b = bab$.
Hence $\theta (s^{-1}) = \theta (s)^{-1}$.
Next we show that if $e$ is an idempotent then $\theta (e)$ is an idempotent.
Let $e$ be an idempotent.
Then $\theta (e) = \theta (e^{-1}) = \theta (e)^{-1}$ by the result above.
Thus
$$\theta (e) = \theta (ee) \leq \theta (e)\theta (e)
= \theta (e)\theta (e^{-1}e) \leq \theta (e)\theta (e)^{-1}\theta (e)
= \theta (e),$$
and so $\theta (e) = \theta (e)\theta (e)$.
We can now prove that $\theta$ is order-preserving.
Let $s \leq t$.
Then $s = te$ for some idempotent $e$.
Thus
$$\theta (s) = \theta (te) \leq \theta (t)\theta (e) \leq \theta (t)$$
since $\theta (e)$ is an idempotent.
A key ingredient in proving that $\theta$ preserves restricted
products is the following:
$$\theta (ss^{-1}) = \theta (s) \theta (s)^{-1}
\mbox{ and }
\theta (s^{-1}s) = \theta (s)^{-1} \theta (s)$$
for every $s \in S$.
We show that $\theta (ss^{-1}) = \theta (s)\theta (s)^{-1}$;
the proof of the other case is similar.
Clearly, $\theta (ss^{-1}) \leq \theta (s)\theta (s)^{-1}$,
so that
$$\theta (ss^{-1}) \theta (s) \leq \theta (s)\theta (s)^{-1}\theta (s)
= \theta (s).$$
But then
$$\theta (s) = \theta ((ss^{-1})s) \leq \theta (ss^{-1})\theta (s)
\leq \theta (s).$$
Thus $\theta (s) = \theta (ss^{-1})\theta (s)$, and so
$$\theta (s)\theta (s)^{-1} = \theta (ss^{-1})(\theta (s)\theta (s)^{-1}).$$
Hence $\theta (s)\theta (s)^{-1} \leq \theta (ss^{-1})$.
But $\theta (ss^{-1}) \leq \theta (s)\theta (s)^{-1}$.
It follows that $\theta (s)\theta (s)^{-1} = \theta (ss^{-1})$.
We can now prove that $\theta$ preserves restricted products.
Suppose that $s \cdot t$ is defined.
Then by the result above so too is $\theta (s) \cdot \theta (t)$.
It remains to show that $\theta (s \cdot t) = \theta (s) \cdot \theta (t)$.
Clearly $\theta (s \cdot t) \leq \theta (s) \cdot \theta (t)$.
Now
$$\theta (s \cdot t)^{-1} \theta (s \cdot t)
= \theta ((s \cdot t)^{-1}(s \cdot t)) = \theta (t^{-1}s^{-1}st)
= \theta (t^{-1}t),$$
and
$$[\theta (s)\cdot \theta (t)]^{-1}[\theta (s) \cdot \theta (t)]
= \theta (t)^{-1}\theta (t) = \theta (t^{-1}t).$$
Hence $\theta (s \cdot t) = \theta (s) \cdot \theta (t)$ as required.
To prove the converse, suppose that $\theta$ preserves the
restricted product and the natural partial order.
We show that it is a prehomomorphism.
Let $st$ be a full product in $S$.
Then $st = (se)\cdot (et)$ where $e = s^{-1}stt^{-1}$ by Lemma~2.1.
Thus, by assumption, $\theta (st) = \theta (se)\cdot \theta (et)$.
But $se \leq s$ and $et \leq t$ so that
$\theta (se) \leq \theta (s)$ and
$\theta (et) \leq \theta (t)$.
Hence $\theta (st) \leq \theta (s)\theta (t)$ as required.
(2) We now prove that
if $\theta \colon \: S \rightarrow T$
is a prehomomorphism satisfying
$\theta (ef) = \theta (e)\theta (f)$
for all idempotents $e,f \in S$, then $\theta $ is a homomorphism.
Let $st$ be a full product in $S$.
Then $st = (se)\cdot (et)$ where $e = s^{-1}stt^{-1}$.
Thus $\theta (st) = \theta (se)\cdot \theta (et)$.
We show that $\theta (se) = \theta (s)\theta (e)$.
Clearly, $\theta (se) \leq \theta (s)\theta (e)$.
Now
$$\theta (se)^{-1}\theta (se) = \theta ((se)^{-1}(se))
= \theta (es^{-1}se) = \theta (e),$$
and $$[\theta (s) \theta (e)]^{-1} \theta (s) \theta (e)
= \theta (e)^{-1} \theta (s)^{-1} \theta (s) \theta (e)
= \theta (e)^{-1} \theta (s^{-1}s)\theta (e)
= \theta (e)$$
since $\theta (e) = \theta (s^{-1}s) \theta (tt^{-1})$ by assumption.
Thus $\theta (se) = \theta (s)\theta (e)$.
Similarly, $\theta (et) = \theta (e)\theta (t)$.
It now follows that
$$\theta (st) = \theta (s)\theta(e)\theta(e)\theta(t)
= \theta (s)\theta (s^{-1}s)\theta (tt^{-1})\theta (t)
= \theta (s)\theta(t).$$
\end{proof}
Lemma~2.2 implies that every prehomomorphism between inverse semigroups induces a functor between their associated groupoids.
\subsection{Inductive groupoids}
Let $(G,\cdot)$ be a groupoid, and let $\leq$ be a partial order defined on $G$.
Then $(G,\cdot,\leq )$ is an {\em ordered groupoid} if the following axioms hold:
\begin{description}
\item[{\rm (OG1)}] $x \leq y$ implies $x^{-1} \leq y^{-1}$
for all $x,y \in G$.
\item[{\rm (OG2)}] For all $x,y,u,v \in G$,
if $x \leq y, u \leq v, \exists xu$ and $\exists yv$ then
$xu \leq yv$.
\item[{\rm (OG3)}] Let $x \in G$ and let $e$ be an identity
such that $e \leq {\bf d}(x)$.
Then there exists a unique element $(x\, | \,e)$,
called the {\it restriction of}\index{restriction in an ordered groupoid}
$x$ {\it to} $e$, such that $(x\, | \,e) \leq x$ and ${\bf d}(x\,|\,e) = e$.
\item[{\rm (OG3)$^{\ast}$}] Let $x \in G$ and let $e$ be an identity
such that $e \leq {\bf r}(x)$.
Then there exists a unique element $(e \, | \, x)$,
called the {\it corestriction of}\index{corestriction in an order\-ed group\-oid}
$x$ {\it to} $e$,
such that $(e \, | \, x) \leq x$ and ${\bf r}(e\, | \,x) = e$.
\end{description}
An ordered groupoid is said to be {\em inductive} if the partially ordered set of identities forms a meet-semilattice.
This term was used by Ehresmann to refer to a more restricted class of ordered groupoids than we have defined, but the terminology is now well-established.
As we shall see, if the groupoid is actually a group the order degenerates to equality and so ordered groupoids do not generalize ordered groups.
A functor between two ordered groupoids is said to
be {\em ordered}
if it is order-preserving.
An ordered functor between two inductive groupoids is said to be
{\em inductive}\
if it preserves the meet operation on the set of identities.
An {\em isomorphism} of ordered groupoids is a bijective ordered
functor whose inverse is an ordered functor.
\begin{lemma} Let $\theta \colon G \rightarrow H$ be an ordered
functor between ordered groupoids.
\begin{enumerate}
\item If $(x \, | \, e)$ is defined in $G$
then $(\theta (x) \, | \, \theta (e))$ is defined in $H$ and
$\theta (x \, | \, e) = (\theta (x) \, | \, \theta (e))$.
\item If $(e \, | \, x)$ is defined in $G$ then
$(\theta (e) \, | \, \theta (x))$ is defined in $H$ and
$\theta (e \, | \, x) = (\theta (e) \, | \, \theta (x))$.
\end{enumerate}
\end{lemma}
\begin{proof}
We shall prove (1); the proof of (2) is similar.
By definition $(x \, | \, e) \leq x$ and so
$\theta (x \, | \, e) \leq \theta (x)$
since $\theta$ is an ordered functor.
But
$${\bf d}(\theta (x \, | \, e))
= \theta ({\bf d}(x \, | \, e)) = \theta (e)$$
since $\theta$ is a functor.
But by axiom (OG3), $(\theta (x) \, | \, \theta (e))$ is the unique element
less than $\theta (x)$ and with domain $\theta (e)$.
Thus $\theta (x \, | \, e) = (\theta (x) \, | \, \theta (e))$.
\end{proof}
We now establish some of the basic properties of ordered groupoids.
\begin{lemma} Let $(G,\cdot,\leq)$ be an ordered groupoid.
\begin{enumerate}
\item[{\rm (1)}] If $x \leq y$ then ${\bf d}(x) \leq {\bf d}(y)$
and ${\bf r}(x) \leq {\bf r}(y)$.
\item[{\rm (2)}] The order $\leq $ restricted to {\rm hom-sets} is trivial.
\item[{\rm (3)}] If $\exists xy$ and $e$ is an identity such that
$e \leq {\bf d}(xy)$ then
$$(xy \, | \, e) = (x \, | \, {\bf r} ( y \, | \, e))(y \, | \, e).$$
\item[{\rm (4)}] If $\exists xy$ and $e$ is an identity such that
$e \leq {\bf r}(xy)$ then
$$(e \, | \, xy) = (e \, | \, x)({\bf d}(e \, | \, x)\, | \, y).$$
\item[{\rm (5)}] If $z \leq xy$ then there exist elements
$x'$ and $y'$ such that $\exists x'y'$, $x' \leq x$, $y' \leq y$
and $z = x'y'$.
\item[{\rm (6)}] Axiom {\em (OG3)}$^*$ is a consequence of the axioms
{\em (OG1)} and {\em (OG3)}.
\item[{\rm (7)}] The set of identities $G_{o}$ is an order ideal of $G$.
\item[{\rm (8)}] If $f \leq e \leq {\bf d}(x)$ then
$(x \, | \, f) \leq (x \, | \, e) \leq x$.
\item[{\rm (9)}] If $f \leq e \leq {\bf r}(x)$ then
$(f \, | \, x) \leq (e \, | \, x) \leq x$.
\item[{\rm (10)}] Let $x,y,e,f \in G$ such that $x \leq y$, $f \leq e$,
$f \leq {\bf d}(x)$ and $e \leq {\bf d}(y)$.
Then $(x \, | \, f) \leq (y \, | \, e)$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) This is immediate from axioms (OG1) and (OG2).
(2) Suppose that
$${\bf d}(x) = {\bf d}(y), \, {\bf r}(x) = {\bf r}(y)
\mbox{ and } x \leq y.$$
In particular, ${\bf d}(x) \leq {\bf d}(y)$
and so by axiom (OG3) there is a unique element $(y\, | \, {\bf d}(x))$
such that
$$(y\, | \, {\bf d}(x)) \leq y
\mbox{ and }
{\bf d}(y\, | \, {\bf d}(x)) = {\bf d}(x).$$
But the element $x$ also has the property that
$x \leq y$ and ${\bf d}(x) = {\bf d}(x)$.
Thus by uniqueness $(y\, | \, {\bf d}(x)) = x$.
However, $(y \, | \, {\bf d}(y)) = x$
since ${\bf d}(x) = {\bf d}(y)$.
But $(y\, | \, {\bf d}(y)) = y$.
Hence $x = y$.
(3) Since $e \leq {\bf d}(xy) = {\bf d}(y)$
the restriction $(y\, | \, e)$ is defined.
Since $(y\, | \, e) \leq y$ we have that
${\bf r}(y\, | \, e) \leq {\bf r}(y) = {\bf d}(x)$.
Thus $(x \, | \, {\bf r} (y\, | \, e))$ exists and the product
$(x\, | \, {\bf r}(y\, | \, e))(y \, | \, e)$ exists.
Clearly, $(x\, | \, {\bf r}(y\, | \, e))(y \, | \, e) \leq xy$.
But
$${\bf d}((x\, | \, {\bf r}(y\, | \, e))(y \, | \, e)) = e,$$
\noindent
and so $(x\, | \, {\bf r}(y \, | \, e)) (y \, | \, e) = (xy\, | \, e)$.
(4) Similar to the proof of (3).
(5) Let $z \leq xy$.
Then ${\bf d}(z) \leq {\bf d}(xy)$ by (1).
Thus $(xy\, | \, {\bf d}(z))$ exists.
Now ${\bf d}(xy\, | \, {\bf d}(z)) = {\bf d}(z)$
and $(xy\, | \, {\bf d}(z)) \leq xy$, so that $z = (xy\, | \, {\bf d}(z))$.
By (3),
$$(xy\, | \, {\bf d}(z))
= (x\, | \, {\bf r}(y \, | \, {\bf d}(z)))(y \, | \, {\bf d}(z)).$$
Put $x' = (x\, | \, {\bf r}(y \, | \, {\bf d}(z)))$
and $y' = (y\, | \, {\bf d}(z))$, and we have the result.
(6) Suppose that the axioms (OG1) and (OG3) hold.
We show that axiom (OG3)$^*$ holds.
Let $e \leq {\bf r}(x)$.
Then $e \leq {\bf d}(x^{-1})$,
so that $(x^{-1}\, | \, e)$ exists by axiom (OG3).
Define $(e\, | \, x) = (x^{-1}\, | \, e)^{-1}$.
Then $(e \,| \,x) \leq x$ by axiom (OG1),
and ${\bf r}(e \,|\, x) = {\bf d}(x^{-1}\, | \, e) = e$.
Now for uniqueness.
Suppose that $y \leq x$ and ${\bf r}(y) = e$.
Then $y^{-1} \leq x^{-1}$ by axiom (OG1)
and ${\bf d}(y^{-1}) = e$.
Thus $y^{-1} = (x^{-1}\, | \, e)$ by (OG3),
and so $y = (e\, | \, x)$ by (OG1).
(7) Let $x \leq e$ where $e$ is an identity.
Then ${\bf d}(x) \leq e$.
But $x,{\bf d}(x) \leq e$ and ${\bf d}(x) = {\bf d}({\bf d}(x))$.
Thus $x = {\bf d}(x)$.
(8) Let $f \leq e \leq {\bf d}(x)$.
Both $(x\, | \, e)$ and $(x\, | \, f)$ exist.
Now $f \leq {\bf d}(x\, | \, e)$ and so the element
$((x\, | \, e) \, |\, f)$ exists.
But ${\bf d}(x \, | \, f) = f$ and $(x \, | \, f) \leq x$.
Thus $((x \, | \, e)\, | \, f) = (x\, | \, f)$,
and so $(x\, | \, f) \leq (x \, | \, e)$.
(9) Similar to (8).
(10) By (8) we have that $(y \, | \, f) \leq (y \, | \, e)$.
However, $(x \, | \, f),(y \, | \, f) \leq y$
and ${\bf d}(x \, | \, f) = {\bf d}(y \, | \, f)$.
Thus by axiom (OG3), we have that $(x \, | \, f) = (y \, | \, f)$.
Hence $(x \, | \, f) \leq (y \, | \, e)$.
\end{proof}
Let $G$ be an ordered groupoid and $H$ a subset of $G$.
Then we say that $H$
is an {\em ordered subgroupoid}
if it is a subgroupoid of $G$
and an ordered groupoid with respect to the induced order.
This is equivalent to the condition that $H$ be a subgroupoid of $G$
and that if $x \in H$ and $e \in H_{o}$ and $e \leq {\bf d}(x)$
then $(x \, | \, e) \in H$.
Let $\theta \colon \: G \rightarrow K$
be an injective ordered functor.
The image of $\theta$ is a subgroupoid of $K$,
because if $\theta (x)\theta(y)$ is defined in $K$
then $\theta ({\bf d}(x)) = \theta ({\bf r}(y))$
and so $xy$ is defined in $G$;
this gives $\theta (x)\theta (y) = \theta (xy)$.
However, the image of $\theta$ need not be an ordered subgroupoid of $K$.
A stronger notion than an injective ordered functor is what we
term an {\em ordered embedding};
this is an ordered functor $\theta \colon \:G \rightarrow K$
such that for all $g,h \in G$
$$g \leq h \Leftrightarrow \theta (g) \leq \theta (h).$$
The image of $\theta$ is an ordered subgroupoid of $K$
which is isomorphic to $G$.
In verifying that a structure is an ordered groupoid, it is
sometimes more convenient to use the following characterization.
We shall need the following two axioms.
Let $G$ be a groupoid and $\leq$ a partial order defined on $G$.
The axioms (OI) and (OG4) are defined as follows:
\begin{description}
\item[{\rm (OI)}] $G_{o}$ is an order ideal of $G$.
\item[{\rm (OG4)}] For all $x \in G$ and $e \in G_{o}$,
if $e \leq {\bf d}(x)$ then there exists $y \in G$
such that $y \leq x$ and ${\bf d}(y) = e$.
\end{description}
\begin{lemma} Let $(G,\cdot)$ be a groupoid and $\leq$
a partial order defined on $G$.
Then $(G,\cdot,\leq)$ is an ordered groupoid if, and only if,
the axioms {\em (OG1)}, {\em (OG2)}, {\em (OI)} and {\em (OG4)} hold.
\end{lemma}
\begin{proof}
If $G$ is an ordered groupoid
then axioms (OG1) and (OG2) hold by definition,
axiom (OI) holds by Lemma~2.4(7) and axiom (OG3)
implies that axiom (OG4) holds.
To prove the converse, it is enough to show that axiom (OG3) holds
because axiom (OG3)$^{\ast}$ follows from the other axioms for an ordered
groupoid by Lemma~2.4(6).
Let $u,v \leq x$ be such that ${\bf d}(u) = {\bf d}(v) = e$.
We shall show that $u = v$ which, together with axiom (OG4), will imply
that axiom (OG3) holds.
Clearly ${\bf d}(u) = {\bf r}(v^{-1}) = e$.
Thus $uv^{-1}$ is defined.
By axiom (OG1), we have that $v^{-1} \leq x^{-1}$.
Thus by axiom (OG2), we have that $uv^{-1} \leq xx^{-1}$.
Now $xx^{-1}$ is an identity and so by axiom (OI), the element $uv^{-1}$
is an identity.
Thus $u = v$, as required.
\end{proof}
\subsection{The Ehresmann-Schein-Nambooripad theorem}
From our results in Chapter~1 and by Lemma~2.1, we have the following.
\begin{proposition} Let $S$ be an inverse semigroup.
Then $(S,\cdot,\leq )$ is an inductive groupoid.
\end{proposition}
The inductive groupoid associated with $S$ is denoted by ${\bf G}(S)$.
We now show how to construct an inverse semigroup from an inductive groupoid.
Let $G$ be an ordered groupoid and let
$x,y \in G$ be such that
$e = {\bf d}(x) \wedge {\bf r}(y)$ exists.
Put
$$x \otimes y = (x\, | \, e)(e \, | \, y),$$
and call $x \otimes y$ the
{\it pseudoproduct}
of $x$ and $y$.
It is immediate from the definition that the pseudoproduct is
everywhere defined in an inductive groupoid.
The next result provides a neat, order-theoretic way of viewing the
pseudoproduct.
\begin{lemma} Let $G$ be an ordered groupoid.
For each pair $x,y \in G$ put
$$\langle x,y \rangle\,
= \{(x',y') \in G \times G \colon
{\bf d}(x' ) = {\bf r}(y') \mbox{ and } x' \leq x
\mbox{ and } y' \leq y\},$$
regarded as a subset of the ordered set $G \times G$.
Then $x \otimes y$ exists if, and only if, there is a maximum element
$(x',y')$ of $\langle x,y \rangle$.
In which case, $x \otimes y = x'y'$.
\end{lemma}
\begin{proof}
Suppose that $x \otimes y$ exists.
Then $e = {\bf d}(x) \wedge {\bf r}(y)$ exists,
and $$((x \,|\, e),(e \,|\, y)) \in \langle x,y \rangle.$$
Let \mbox{$(u,v) \in \,\langle x,y \rangle$.}
Then
$$u \leq x,\, v \leq y \mbox{ and } {\bf d}(u) = {\bf r}(v) = f,$$
say.
Thus by axioms (OG3) and (OG3)$^{\ast}$ we have that
$u = (x \, | \, f)$ and $v = (f \, | \, y)$.
By Lemma~2.4(1),
${\bf d}(u) \leq {\bf d}(x)$ and ${\bf r}(v) \leq {\bf r}(y)$.
Thus $f \leq {\bf d}(x),{\bf r}(y)$,
and so, by assumption, $f \leq e$.
By Lemma~2.4(8),(9), it follows that
$$u = (x \, | \, f) \leq (x \, | \, e)
\mbox{ and }
v = (f \, | \, y) \leq (e \, | \, y).$$
Thus $((x \, | \, e),(e \, | \, y))$ is the maximum
element of $\langle x,y \rangle$.
Conversely, suppose that the maximum element of $\langle x,y \rangle$
exists and equals $(x',y')$.
Put $e = {\bf d}(x') = {\bf r}(y')$.
Clearly, $e \leq {\bf d}(x),{\bf r}(y)$.
Now let $f$ be any identity such that $f \leq {\bf d}(x),{\bf r}(y)$.
Then
$$(x\, | \, f) \leq x, \, (f\, | \, y) \leq y
\mbox{ and }
{\bf d}(x \, | \, f) = f = {\bf r}(f \, | \, y).$$
Thus $((x \, | \, f),(f \, | \, y)) \in \, \langle x,y \rangle$
and so $((x \, | \, f),(f \, | \, y)) \leq (x',y')$.
Hence $f \leq e$, which implies that $e = {\bf d}(x) \wedge {\bf r}(y)$.
Thus $x \otimes y$ exists.
The proof of the last assertion is now immediate.
\end{proof}
It will be an immediate consequence of the following result that
the pseudoproduct on an inductive groupoid is associative.
\begin{lemma}
Let $G$ be an ordered groupoid.
Then for all $x,y,z \in G$
if $x \otimes (y \otimes z)$ and
$(x \otimes y) \otimes z$ both exist
then they are equal.
\end{lemma}
\begin{proof}
Let $(x \otimes y) \otimes z = az'$
where $(a,z')$ is the maximum element of
$\langle x \otimes y,z \rangle$.
Let $x \otimes y = x'y'$ where $(x',y')$
is the maximum element of $\langle x,y \rangle$.
Then
$$a \leq x \otimes y, \, z' \leq z, \, x' \leq x
\mbox{ and } y' \leq y.$$
By Lemma~2.4(5), $a \leq x'y'$ implies that there
are elements
$x'' \leq x'$ and $y'' \leq y'$
such that $a = x''y''$.
Thus
$$(x \otimes y) \otimes z
= (x''y'')z' = x''(y''z').$$
Now, $y'' \leq y' \leq y$ and $z' \leq z$,
so that $(y'',z') \in ~ \langle y,z \rangle$.
Thus $y''z' \leq y \otimes z$.
Similarly, $(x'',y''z') \in ~\langle x,y \otimes z \rangle$ and so
$x''(y''z') \leq x \otimes (y \otimes z)$.
Hence
$(x \otimes y) \otimes z \leq x \otimes (y \otimes z)$.
The reverse inequality follows by symmetry.
\end{proof}
If $(G,\cdot,\leq)$ is an inductive groupoid, then $(G,\otimes)$ will be denoted by ${\bf S}(G)$.
We can now show how to construct an inverse semigroup from an
inductive groupoid.
\begin{proposition} Let $(G,\cdot ,\leq )$ be an inductive groupoid.
\begin{enumerate}
\item $(G,\otimes )$ is an inverse semigroup.
\item ${\bf G}({\bf S}(G,\cdot ,\leq )) = (G,\cdot ,\leq )$.
\item For any inverse semigroup $S$ we have that ${\bf S}({\bf G}(S)) = S$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) By Lemma~2.8, $(G,\otimes )$ is a semigroup.
If $x,y \in G$ and $\exists x\cdot y$ in the groupoid $G$
then $x\cdot y = x \otimes y$.
But for each element $x \in G$ we have that
$x = x\cdot x^{-1}\cdot x$ and $x^{-1} = x^{-1}\cdot x\cdot x^{-1}$.
Thus $(G,\otimes )$ is a regular semigroup.
It is easy to check that the idempotents of $(G,\otimes )$
are precisely the identities of $(G,\cdot )$.
Let $e$ and $f$ be two idempotents of $(G,\otimes )$.
Then
$$e \otimes f = (e \,|\, e \wedge f)(e \wedge f \,| \,f )
= e \wedge f
= (f \,|\, e \wedge f) (e \wedge f \,|\, e) = f \otimes e$$
so that the idempotents commute.
It follows that $(G,\otimes )$ is an inverse semigroup.
(2) We show first that the natural partial order on $(G,\otimes )$
is just $\leq$.
Suppose that $x = e \otimes y$ in $(G,\otimes )$ for some idempotent $e$.
Then $x = (e \wedge {\bf r}(y)\, | \, y)$ and so
$x \leq y$ in $(G,\cdot ,\leq)$.
Conversely, suppose that $x \leq y$ in $(G,\cdot ,\leq)$.
Then $x = ({\bf r}(x)\, | \, y)$.
But $({\bf r}(x)\, | \, y) = {\bf r}(x) \otimes y$.
Thus $x \leq y$ in $(G,\otimes )$.
Now we turn to the restricted product.
The restricted product of $x$ and $y$ is defined in $(G,\otimes )$
precisely when $x^{-1} \otimes x = y \otimes y^{-1}$.
But from the properties of the pseudoproduct,
we have that $x^{-1} \otimes x = x^{-1}\cdot x$
and $y \otimes y^{-1} = y\cdot y^{-1}$ in $(G,\cdot)$.
Thus the restricted product of $x$ and $y$ exists in $(G,\otimes)$
precisely when the product $x\cdot y$ exists in $(G,\cdot )$.
Thus ${\bf G}({\bf S}(G,\cdot ,\leq )) = (G,\cdot,\leq )$.
(3) The pseudoproduct in ${\bf G}(S)$ is given by
$$s \otimes t = (s\, | \, e)\cdot (e\, | \, t),$$
\noindent
where $e = {\bf d}(s) \wedge {\bf r}(t)$ and the product on the
right is the restricted product in $S$.
But $(s\, | \, e) = se$ and $(e\, | \, t) = et$ and $e = s^{-1}stt^{-1}$
by Lemma~2.1.
Thus $s \otimes t = st$.
Hence ${\bf S}({\bf G}(S)) = S$.
\end{proof}
\begin{theorem}[Ehresmann-Schein-Nambooripad] The category of inverse semigroups and pre\-homo\-morph\-isms
(respectively, homo\-morph\-isms) is isomorphic to the category of
inductive groupoids and ordered functors (respectively, inductive functors).
\end{theorem}
\begin{proof}
Define a function ${\bf G}$ from the category
of inverse semigroups and prehomomorphisms
to the category of inductive groupoids and ordered functors
as follows:
for each inverse semigroup $S$ we define
${\bf G}(S) = (S,\cdot,\leq )$, an inductive groupoid by Proposition~2.6,
and if $\theta \colon \: S \rightarrow T$ is a prehomomorphism
then ${\bf G}(\theta ) \colon {\bf G}(S) \rightarrow {\bf G}(T)$
is defined to be the same function on the underlying sets;
this is an ordered functor by Lemma~2.2.
It is easy to check that ${\bf G}$ defines a functor.
Define a function ${\bf S}$ from the category of inductive
groupoids and ordered functors to the category of inverse
semigroups and prehomomorphisms as follows:
for each inductive groupoid $G$ we define
${\bf S}(G) = (G,\otimes )$, an inverse semigroup by Proposition~2.9,
and if $\phi \colon \:G \rightarrow H$
is an ordered functor between inductive groupoids
then ${\bf S}(\phi ) \colon {\bf S}(G) \rightarrow {\bf S}(H)$
is defined to be the same function on the underlying sets;
this is a prehomomorphism by Lemma~2.2
since it preserves the restricted product
of ${\bf S}(G)$ and the natural partial order.
It is easy to check that ${\bf S}$ is a functor.
By Proposition~2.9, we have that
${\bf G}({\bf S}(G,\cdot,\leq ))
= (G,\cdot,\leq )$ and ${\bf S}({\bf G}(S)) = S$.
It is now immediate that the category of inverse semigroups and
prehomomorphisms is isomorphic to the category of inductive
groupoids and ordered functors.
By Lemma~2.2 this isomorphism restricts to an
isomorphism between the category of inverse semigroups and
semigroup homomorphisms and the category of inductive groupoids
and inductive functors.
\end{proof}
The above theorem can be viewed as a wide-ranging generalization of the result that a commutative idempotent semigroup can also be regarded as a meet semilattice.
One of its uses is in proving that a structure suspected of being an inverse semigroup actually is.
As usual in semigroup theory we have the slight annoyance of having to deal with the case of semigroups with and without zero separately.
An ordered groupoid is said to be {\em $\ast$-inductive}
if the following condition holds for each pair of identities:
if they have a lower bound, they have a greatest lower bound.
A $\ast$-inductive groupoid gives rise to an inverse semigroup with zero $(G^{0},\otimes)$:
adjoin a zero to the set $G$, and extend the pseudoproduct on $G$ to $G^{0}$ in such a way
that if $s,t \in G$ and $s \otimes t$ is not defined then put $s \otimes t = 0$,
and define all products with $0$ to be $0$.
Every inverse semigroup with zero arises in this way.
\subsection{Applications}
Ordered groupoids can be viewed as wide-ranging generalizations of inverse semigroups.
`Wide-ranging' because both groupoids and partially ordered sets are examples of ordered groupoids.
The category of ordered groupoids provides much more space for working with inverse semigroups.
In particular, constructions that lead out of the category of inverse semigroups may actually be possible in the larger category of ordered groupoids.
In this section, I shall illustrate this idea.
Let $G$ be an ordered subgroupoid of the ordered groupoid $H$.
We say that $H$ is an {\em enlargement}
of $G$ if the following
three axioms hold:
\begin{description}
\item[{\rm (GE1)}] $G_{o}$ is an order ideal of $H_{o}$.
\item[{\rm (GE2)}] If $x \in H$ and ${\bf d}(x),{\bf r}(x) \in G$
then $x \in G$.
\item[{\rm (GE3)}] If $e \in H_{o}$ then there exists $x \in H$
such that ${\bf r}(x) = e$ and ${\bf d}(x) \in G$.
\end{description}
The following is proved in Section~8.3 of \cite{Lawson98}.
\begin{theorem}[The maximum enlargement theorem] \mbox{}
\begin{enumerate}
\item Let $p \colon \: H \rightarrow K$ be an ordered,
star injective functor between ordered groupoids.
Then there is an ordered groupoid $G$,
an ordered embedding $i \colon \: H \rightarrow G$,
and an ordered covering functor $p'\colon \: G \rightarrow K$
such that $p'i = p$ where $G$ is an enlargement of $i(H)$.
\item Let $j \colon \: H \rightarrow G'$ be any ordered embedding
and let $p'' \colon \: G' \rightarrow K$ be an ordered covering
functor such that $p''j = p$.
Then there is a unique ordered functor $\theta \colon \: G \rightarrow G'$
such that $\theta i = j$ and $p'' \theta = p'$.
\end{enumerate}
\end{theorem}
The key point of the above theorem is that every star injective ordered functor can be factorized as an enlargement followed by a star bijective ordered functor,
and that this can be done in essentially one way.
Star bijective or covering functors have pleasant properties,
and if an ordered groupoid is an enlargement of another ordered groupoid then it is similar to it in structure.
We shall apply the above theorem to determine the structure of the $E$-unitary inverse semigroups introduced in Chapter~1.
It hinges on two observations.
First, in Theorem~2.26, we proved that an inverse semigroup is $E$-unitary if and only if the natural homomorphism
to its maximum group image is star injective;
second, in Theorem~3.6 an inverse semigroup is a semidirect product of a semilattice by a group if and only if that selfsame natural map is star bijective.
We begin by defining semidirect products of partially ordered sets by groups.
Let $G$ be a group and $X$ a partially ordered set.
We suppose that $G$ acts on $X$ by order automorphisms on the left.
Define a partial multiplication on $X \times G$ by
$(x,g)(y,h) = (x,gh)$ if $x = g \cdot y$ and undefined otherwise
and define a partial order on $X \times G$ by
$$(x,g) \leq (y,h) \Leftrightarrow x \leq y \mbox{ and } g = h.$$
The set $X \times G$ equipped with this partial multiplication
and partial order is denoted by $P(G,X)$.
The following is proved as Theorem~1 of Section~8.1 of \cite{Lawson98}.
\begin{lemma} $P(G,X)$ is an ordered groupoid,
and the function $\pi_{2} \colon \: \: P(G,X) \rightarrow G$
defined by $\pi_{2}(x,g) = g$ is a surjective, ordered covering functor.
\end{lemma}
We say that the ordered groupoid
$P(G,X)$ is a {\em semidirect product of a partially ordered set by a group}.
Such semidirect product ordered groupoids can be characterized abstractly.
The proof of the following is Theorem~3 of Section~8.1 of \cite{Lawson98}.
\begin{proposition}
Let $\pi \colon \: \Pi \rightarrow G$
be an ordered covering functor from the ordered groupoid $\Pi$
onto a group $G$.
Then $G$ acts on the poset $X = \Pi_{o}$
by order automorphisms, and there is an isomorphism of ordered groupoids
$\theta \colon \: \Pi \rightarrow P(G,X)$
such that $\pi_{2}\theta = \pi$.
\end{proposition}
The key point of Lemma~2.12 and Proposition~2.13 is that we can recognize when an ordered groupoid
is a semidirect product of a partially ordererd set by a group by checking to see if it admits
an ordered covering functor onto as group.
This is the ordered groupoid version of Theorem~3.6 of Chapter~1.
Now let $S$ be an $E$-unitary inverse semigroup.
Then the natural map from $S$ to its maximum group image is star injective.
This can be factorized using the maximum enlargment theorem into an enlargement followed by a covering map to a group.
It follows that every $E$-unitary inverse semigroup has an enlargement which is a semidirect product of a partially ordered set and a group.
For the full proof of the following see Theorem~4 of Section~8.1 of \cite{Lawson98}.
\begin{theorem} Let $S$ be an inverse semigroup with associated
inductive groupoid ${\bf G}(S)$.
Then $S$ is $E$-unitary if, and only if,
there is an ordered embedding
$\iota \colon \: {\bf G}(S) \rightarrow P(G,X)$
into some semidirect product of a poset by a group
such that $P(G,X)$ is an enlargement of $\iota ({\bf G}(S))$
in such a way that the function
$\pi_{2} \colon \: P(G,X) \rightarrow G$ restricted to $\iota({\bf G}(S))$
is surjective.
\end{theorem}
The above result characterizes $E$-unitary inverse semigroups in terms of semidirect products,
but not of semilattices by groups as might be expected but of partially ordered sets by groups.
The disadvantage of the above result is that the structure of $E$-unitary inverse semigroups is described using ordered groupoids.
However, the information contained in the theorem can be couched in purely semigroup-theoretic language.
To do this we need the following classical definition.
Let $G$ be a group and $X$ a partially ordered set.
We shall suppose that $G$ acts on $X$ on the left by order automorphisms.
We denote the action of $g \in G$ on $x \in X$ by $g \cdot x$.
Let $Y$ be a subset of $X$ partially ordered by the induced ordering.
We say that $(G,X,Y)$ is a {\em McAlister triple} if the following three axioms hold:
\begin{description}
\item[{\rm (MT1)}] $Y$ is an order ideal of $X$ and a meet
semilattice under the induced ordering.
\item[{\rm (MT2)}] $G \cdot Y = X$.
\item[{\rm (MT3)}] $g \cdot Y \cap Y \neq \emptyset$ for every $g \in G$.
\end{description}
Let $(G,X,Y)$ be a McAlister triple.
Put
$$P(G,X,Y) = \{(y,g) \in Y \times G \colon \: g^{-1} \cdot y \in Y \}.$$
\begin{lemma} Let $(e,g),(f,h) \in P(G,X,Y)$.
Then $e \wedge g \cdot f$ exists in the partially
ordered set $X$ and
$(e \wedge g \cdot f,gh) \in P(G,X,Y)$.
\end{lemma}
\begin{proof}
By assumption, $g^{-1} \cdot e \in Y$.
Thus $g^{-1} \cdot e \wedge f$ exists since $Y$ is a semilattice.
Put $i = g^{-1} \cdot e \wedge f$.
Then $i \leq g^{-1} \cdot e$ and $i \leq f$.
Thus $g \cdot i \leq e$ and $g \cdot i \leq g \cdot f$.
Now let $j \leq e, g \cdot f$.
Then $g^{-1} \cdot j \leq g^{-1} \cdot e$ and $g^{-1} \cdot j \leq f$.
Hence $g^{-1} \cdot j \leq i$,
and so $j \leq g \cdot i$.
We have therefore shown that the meet $e \wedge g \cdot f$ exists.
To show that
$(e \wedge g \cdot f,gh) \in P(G,X,Y)$,
we have to show that the element
$(gh)^{-1} \cdot (e \wedge g \cdot f)$
belongs to $Y$.
Now
$$(gh)^{-1} \cdot (e \wedge g \cdot f)
=
h^{-1} \cdot (g^{-1} \cdot (e \wedge g \cdot f))$$
and
$$h^{-1} \cdot (g^{-1} \cdot (e \wedge g \cdot f))
\leq h^{-1} \cdot (g^{-1} \cdot (g \cdot f))
\leq h^{-1} \cdot f \in Y.$$
But $Y$ is an order ideal of $X$ and so
$(gh)^{-1} \cdot (e \wedge g \cdot f) \in Y$.
\end{proof}
Define a product on $P(G,X,Y)$ by
$$(e,g)(f,h) = (e \wedge g \cdot f,gh).$$
It is well-defined by the above lemma.
A proof of the following is Theorem~9 of Section~7.2 of \cite{Lawson98}.
\begin{proposition} $P(G,X,Y)$ is an $E$-unitary inverse semigroup,
with semilattice of idempotents isomorphic to $Y$ and maximum
group homomorphic image isomorphic to $G$.
\end{proposition}
If we combine Theorem~2.4 with Proposition~2.16, we get the following first proved by Don McAlister.
\begin{theorem}[The $P$-theorem]
An inverse semigroup is $E$-unitary if and only if it is isomorphic to a $P$-semigroup.
\end{theorem}
\section{Ordered groupoids and left/right cancellative categories}
This section is a bridge between the inductive groupoid approach to studying inverse semigroups described in Section~2 and the
category action approach that we describe in Section~4.
Our goal is to show the extent to which ordered groupoids are related to left (respectively, right) cancellative categories.
The motivation for this section comes from the category-theoretic definition of a subobject.
\subsection{From left cancellative categories to ordered groupoids }
A {\em left cancellative} category is a category in which $xy = xz$ implies $y = z$.
It is therefore precisely a category of monomorphisms.
We define a {\em right cancellative} category dually, and a {\em cancellative} category is one which is both left and right cancellative.
A left (respectively, right) cancellative category with one identity is a left (respectively, right) cancellative monoid.
If $C$ is a subcategory of $D$, we say that it is {\em isomorphism-dense}
in $D$ if for each identity $e \in D_{o}$ there exists an identity
$f \in C_{o}$ and an isomorphism $x \in D$ such that
$e \stackrel{x}{\longrightarrow} f$.
If $C$ is a subcategory of $D$, we say it is {\em full}
if $x \in D$ such that $\mathbf{d} (x), \mathbf{r} (x) \in C$ then $x \in C$.
A functor $F \colon C \rightarrow D$ is an {\em equivalence} if it is full, faithful and essentially surjective;
the first two conditions mean that the restriction $F \colon \mbox{hom}(e,f) \rightarrow \mbox{hom}(F(e),F(f))$
is surjective and injective respectively for all identities $e,f \in C_{o}$,
and the last condition means that each identity in $D$ is isomorphic to the image of an identity under $F$.
We now define two categories of structures, the relationship between them being the subject of this section.
\begin{description}
\item[$\mathcal{LC}$] The category of left cancellative categories
and their functors.
\item[$\mathcal{OG}$] The category of ordered groupoids and ordered functors.
\end{description}
We shall construct functors
$${\bf G} \colon \: \mathcal{LC} \rightarrow \mathcal{OG}
\text{ and }
{\bf L} \colon \: \mathcal{OG} \rightarrow \mathcal{LC},$$
and describe their composites ${\bf G} {\bf L}$ and ${\bf L} {\bf G}$.
Let $C$ be a left cancellative category.
We shall construct an ordered groupoid ${\bf G} (C)$ from $C$.
Put
$$U = \{(a,b) \in C \times C \colon \: \mathbf{d} (a) = \mathbf{d} (b)\}.$$
Define a relation $\sim$ on $U$ as follows:
$$(a,b) \sim (a',b') \Leftrightarrow (a,b) = (a',b')u \mbox{ for some isomorphism } u$$
where $(a',b')u = (a'u,b'u)$.
Then $\sim$ is an equivalence relation on $U$.
Denote the equivalence class containing $(a,b)$ by $[a,b]$,
and the set of equivalence classes by ${\bf G}(C)$.
Define
$$\mathbf{d} [a,b] = [b,b], \quad \mathbf{r} [a,b] = [a,a]
\mbox{ and }
[a,b]^{-1} = [b,a].$$
These are all well-defined.
Suppose $\mathbf{d} [a,b] = \mathbf{r} [c,d]$.
Then there exists an isomorphism $u$ in $C$ such that $b = cu$.
In this case, define the partial product
$$[a,b] \cdot [c,d] = [a,du].$$
The picture below illustrates why:
$$\diagram
&
&
&
\\
&
&
&\uto^{d} \lto_{c}
\\
&
&\lto_{a} \uto^{b} \urto_{u}
&
\enddiagram$$
We shall usually denote the partial product by concatenation.
Define a relation $\leq$ on ${\bf G}(C)$
by
$$[a,b] \leq [c,d] \Leftrightarrow (a,b) = (c,d)p \mbox{ for some }p \in C.$$
This is well-defined and is a partial order.
If $\theta \colon \: C \rightarrow D$ is a functor
between two left cancellative categories,
define the function
${\bf G}(\theta) \colon \: {\bf G}(C) \rightarrow {\bf G}(D)$
by ${\bf G}(\theta)([a,b]) = [\theta (a),\theta (b)]$.
\begin{proposition} Let $C$ be a left cancellative category.
Then ${\bf G}(C)$ is an ordered groupoid,
and ${\bf G}$ is a functor from
$\mathcal{LC}$ to $\mathcal{OG}$.
\end{proposition}
\begin{proof}
The set $({\bf G} (C),\cdot)$ equipped with the partial binary operation defined above
is a groupoid in which the set of identities is
$${\bf G} (C)_{o} = \{[a,a] \colon \: a \in C \}.$$
To show that $({\bf G} (C), \cdot,\leq)$ is an ordered groupoid,
we verify that the axioms (OG1)--(OG3) hold.
It is immediate that (OG1) holds,
and the proof of (OG2) is easy by left cancellativity.
We now prove that both (OG3) and (OG3)$^{\ast}$ hold.
Let $[a,a] \leq \mathbf{d} [x,y]$.
Then $a = yp$ for some $p \in C$.
Define
$$([x,y]\,|\,[a,a]) = [xp,a].$$
It is easy to check that this is a well-defined restriction
whose uniqueness is a consequence of left cancellativity.
Let $[b,b] \leq \mathbf{r} [x,y]$.
Then $b = xq$ for some $q \in C$.
Define
$$([b,b]\,|\,[x,y]) = [b,yq].$$
It is easy to check that this is a well-defined corestriction,
which is likewise unique.
It follows that ${\bf G} (C)$ is an ordered groupoid.
If $\theta \colon \: C \rightarrow D$ is a functor
between left cancellative categories,
then
${\bf G} (\theta)$ is an ordered functor from
${\bf G} (C)$ to ${\bf G} (D)$.
The proof that ${\bf G}$ is a functor is also straightforward.
\end{proof}
It is useful to know when the pseudoproduct of two
elements of ${\bf G}(C)$ exists.
\begin{lemma} Let $C$ be a left cancellative category, and ${\bf G}(C)$ its associated ordered groupoid.
\begin{enumerate}
\item Let $[x,y],[w,z] \in {\bf G}(C)$.
Then $[x,y] \otimes [w,z]$ is defined
if and only if
$y$ and $w$ have a pullback in $C$.
\item Let $C$ have the following additional property:
if $a$ and $b$ are any elements of $C$ such that $\mathbf{r} (a) = \mathbf{r} (b)$
and that can be completed to a commutative square $aa' = bb'$ for some elements
$a'$ and $b'$, then $a$ and $b$ have a pullback.
Then ${\bf G} (C)$ is $\ast$-inductive.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) Consider the identities $[y,y]$ and $[w,w]$.
They have a lower bound
iff elements $p,q \in C$ can be found such that $yp = wq$;
and they have a greatest lower bound iff $y$ and $w$ have
a pullback in $C$;
left cancellativity is once again used crucially.
Let $a = yp = wq$ be a pullback of $y$ and $w$.
Then $[y,y] \wedge [w,w] = [a,a]$.
A simple calculation using the forms of the restriction
and corestriction yields
$$[x,y] \otimes [w,z]
=
[xp,zq].$$
In diagrammatic terms this is just:
$$\diagram
&
&
&
\\
&
&
&\uto^{z} \lto_{w}
\\
&
&\lto_{x} \uto^{y}
&\lto_{p} \uto_{q}
\enddiagram$$
(2) This follows immediately from the proof of (1).
\end{proof}
Ordered groupoids of the form ${\bf G} (C)$ have extra properties.
Let $G$ be an ordered groupoid.
We say that $G$ has {\em maximal identities} if there is
a function $G_{o} \rightarrow G_{o}$, denoted by $e \mapsto e^{\circ}$,
that has the following two properties:
\begin{description}
\item[{\rm (MI1)}] $e \leq e^{\circ}$.
\item[{\rm (MI2)}] If $e \leq i^{\circ},j^{\circ}$
then $i^{\circ} = j^{\circ}$.
\end{description}
Observe that $e^{\circ}$ really is a maximal identity,
for if $e^{\circ} \leq f$, where $f$ is an identity,
then $e^{\circ} \leq f \leq f^{\circ}$ by (MI1).
But then by (MI2), we have that $e^{\circ} = f^{\circ}$,
and so $e^{\circ} = f$.
We define the relation $\mathcal{D}$ on an ordered groupoid $G$ by
$g \,\mathcal{D}\, h$ iff they are in the same connected component
of the groupoid $G$.
It is immediate that $\mathcal{D}$ is an equivalence relation on $G$.
\begin{proposition} Let $C$ be a left cancellative category.
Then ${\bf G} (C)$ is an ordered groupoid with maximal identities.
In addition, each $\mathcal{D}$-class contains a maximal identity.
\end{proposition}
\begin{proof}
For each identity $[a,a] \in {\bf G} (C)$ define
$$[a,a]^{\circ} = [\mathbf{r} (a),\mathbf{r} (a)].$$
This is evidently a function from ${\bf G} (C)_{o}$ to ${\bf G} (C)_{o}$.
Because $(a,a) = (\mathbf{r} (a),\mathbf{r} (a))a$ we have that
$[a,a] \leq [a,a]^{\circ}$.
Thus (MI1) holds.
Suppose that
$$[a,a] \leq [\mathbf{r} (b), \mathbf{r} (b)],[\mathbf{r} (c),\mathbf{r} (c)].$$
Then $a = \mathbf{r} (b)p = \mathbf{r} (c)q$ for some $p$ and $q$.
But then $a = p = q$ and so $\mathbf{r} (b) = \mathbf{r} (p) = \mathbf{r} (a)$,
and $\mathbf{r} (c) = \mathbf{r} (p)$.
Thus $[\mathbf{r} (b), \mathbf{r} (b)] = [\mathbf{r} (c),\mathbf{r} (c)]$,
and so (MI2) holds.
It follows that ${\bf G} (C)$ has maximal identities.
Observe that an identity is of the form $[a,a]^{\circ}$
iff it is of the form $[e,e]$ for some identity $e \in C_{o}$.
Let $[a,a]$ be an arbitrary identity.
Consider the maximal identity
$[\mathbf{d} (a),\mathbf{d} (a)]$.
Then $[a, \mathbf{d} (a)] \in {\bf G} (C)$ is
such that
$\mathbf{d} [a, \mathbf{d} (a)] = [\mathbf{d} (a), \mathbf{d} (a)]$
and
$\mathbf{r} [a, \mathbf{d} (a)] = [a,a]$.
Hence $[a,a] \, \mathcal{D} \, [\mathbf{d} (a), \mathbf{d} (a)]$.
It follows that each $\mathcal{D}$-class contains a maximal identity.
\end{proof}
We have shown how to construct an ordered groupoid ${\bf G} (C)$ from a left cancellative category.
It is likewise possible to construct an ordered groupoid ${\bf G}' (C)$ from a {\em right} cancellative category $C$.
Just as the construction of ${\bf G} (C)$ is a generalization of the construction of subobjects,
so the construction of ${\bf G}' (C)$ is a generalization of quotient objects.
The set $U'$ is defined to consist of those pairs of elements $(a,b)$
of $C$ such that ${\bf r}(a) = {\bf r}(b)$.
The equivalence relation $\sim$ is defined by
$$(a,b) \sim (a',b') \Leftrightarrow (a,b) = u(a',b')$$
for some isomorphism $u$.
We define ${\bf G}' (C)$ to consist of $\sim$-equivalence classes.
The following is immediate
\begin{proposition} Let $C$ be a right cancellative category.
Then ${\bf G}' (C)$ is an ordered groupoid with maximal identities.
In addition, each $\mathcal{D}$-class contains a maximal identity.
\end{proposition}
\subsection{From ordered groupoids to left cancellative categories}
Let $G$ be an ordered groupoid.
Define
$${\bf L} (G) = \{(e,x) \in G_{o} \times G \colon \: {\bf r}(x) \leq e\}$$
and
$$\mathbf{d} (e,x) = (\mathbf{d} (x),\mathbf{d} (x))
\mbox{ and }
\mathbf{r} (e,x) = (e,e)$$
and define a partial product on ${\bf L} (G)$ as follows:
if $\mathbf{d} (e,x) = \mathbf{r} (f,y)$ then
$$(e,x)(f,y) = (e,x \otimes y),$$
otherwise it is undefined.
If $\theta \colon \: G \rightarrow H$ is an ordered functor
between two ordered groupoids, define
${\bf L}(\theta) \colon \: {\bf L}(G) \rightarrow {\bf L}(H)$
by ${\bf L}(\theta)(e,x) = (\theta (e),\theta (x))$.
Since $\mathbf{r} (x) \leq e$ and $\theta$ is an order-preserving functor,
we have that $\theta (\mathbf{r} (x)) = \mathbf{r} ( \theta (x))$
and $\mathbf{r} (\theta (x)) \leq \theta (e)$.
Thus ${\bf L} (\theta)(e,x)$ is an element of ${\bf L} (G)$,
and so ${\bf L} (\theta)$ is a well-defined function.
\begin{proposition} Let $G$ be an ordered groupoid.
Then ${\bf L}(G)$ is a left cancellative category,
and
${\bf L}$ defines a functor from $\mathcal{OG}$ to $\mathcal{LC}$.
The identities of ${\bf L}(G)$ are those elements of the form
$(e,e)$ where $e$ is an identity of $G$.
The invertible elements of ${\bf L} (G)$ are those of the form $(\mathbf{r} (x),x)$
and constitute a groupoid isomorphic to $G$.
\end{proposition}
\begin{proof}
We regard the element $(e,x)$ in diagrammatic terms
as follows
$$\diagram
&e \ddotted
&\mathbf{d} (x)\dlto^{x}
\\
&\mathbf{r} (x)
&
\enddiagram$$
The dotted line indicates the natural partial order.
The fact that $\mathbf{d} (e,x) = \mathbf{r} (f,y)$
means, in diagrammatic terms, that we have
$$\diagram
&e \ddotted
&f \dlto^{x} \ddotted
&\mathbf{d} (y) \dlto^{y}
\\
&\mathbf{r} (x)
&\mathbf{r} (y)
&
\enddiagram$$
Thus using the restriction we can construct the following
$$\diagram
&e \ddotted
&f \dlto^{x} \ddotted
&\mathbf{d} (y) \dlto^{y}
\\
&\ddotted
&\mathbf{r} (y) \dlto^{(x\,|\,\mathbf{r} (y))}
&
\\
&\mathbf{r} (x \,|\, \mathbf{r} (y))
&
&
\enddiagram$$
But $(x \,|\, \mathbf{r} (y))y = x \otimes y$.
It is now evident that ${\bf L} (G)$ is a category
whose set of identities is
$${\bf L} (G)_{o} = \{(e,e) \colon \: e \in C_{o} \}.$$
We prove that it is left cancellative.
Suppose that
$$(e,x)(f,y) = (e,x)(f,z).$$
Then $(x \,|\, \mathbf{r} (y))y = (x \,|\, \mathbf{r} (z))z$.
But $(x \,|\, \mathbf{r} (y)),(x \,|\, \mathbf{r} (z)) \leq x$
and $\mathbf{r} (x \,|\, \mathbf{r} (y)) = \mathbf{r} (x \,|\, \mathbf{r} (z))$.
Hence $(x \,|\, \mathbf{r} (y)) = (x \,|\, \mathbf{r} (z))$ and so $y = z$.
We have therefore proved that $(f,y) = (f,z)$.
The proof that ${\bf L} (\theta)$ is a functor
follows from the fact that if
$\theta \colon \: G \rightarrow H$ is an ordered functor and $e \leq \mathbf{d} (x)$,
then $\theta (e) \leq \mathbf{d} (\theta (x))$
and so $\theta (x \,|\, e) = (\theta (x) \,|\, \theta (e))$.
Consequently ${\bf L}$ is a functor.
It is easy to check that the inverse of $(\mathbf{r} (x),x)$ is $(\mathbf{r} (x^{-1}), x^{-1})$,
and that only elements of the form $(\mathbf{r} (x),x)$ are invertible.
The function from $G$ to ${\bf L} (G)$ defined by $x \mapsto (\mathbf{r} (x),x)$
induces an isomorphism between $G$ and the groupoid of invertible elements of ${\bf L} (G)$.
\end{proof}
We may also construct a {\em right} cancellative category
from an ordered groupoid.
Let
$${\bf R} (G) = \{(x,e) \in G \times G_{o} \colon \: {\bf d}(x) \leq e\}.$$
Define
$${\bf d}(x,e) = (e,e)
\mbox{ and }
{\bf r}(x,e) = ({\bf r}(x),{\bf r}(x))$$
and define a partial product on ${\bf R} (G)$ as follows:
if ${\bf d}(x,e) = {\bf r}(y,f)$ then
$(x,e)(y,f) = (x \otimes y,f)$,
else it is undefined.
\begin{proposition} Let $G$ be an ordered groupoid.
Then ${\bf R} (G)$ is a right cancellative category,
and ${\bf L} (G)^{\scriptsize op}$ is isomorphic
to ${\bf R} (G)$.
\end{proposition}
\begin{proof}
It is easy to check that ${\bf R} (G)$ is a category.
We now prove that ${\bf L} (G)^{\scriptsize op}$ is isomorphic
to ${\bf R} (G)$.
Define a function
$\theta \colon {\bf L}(G)^{\scriptsize op} \rightarrow {\bf R}(G)$
by $\theta (e,x)^{\scriptsize op} = (x^{-1},e)$.
It is immediate that this function is bijective
and maps identities to identities.
Suppose that
$(f,y)^{\scriptsize op}(e,x)^{\scriptsize op}$
is defined in ${\bf L}(G)^{\scriptsize op}$.
By definition
$$(f,y)^{\scriptsize op}(e,x)^{\scriptsize op}
= ((e,x)(f,y))^{\scriptsize op}$$
and so
${\bf d}(x) = f$.
By definition,
$$\theta (f,y)^{\scriptsize op} = (y^{-1},f)
\mbox{ and }
\theta (e,x)^{\scriptsize op} = (x^{-1},e).$$
It follows that
$\theta (f,y)^{\scriptsize op}\theta (e,x)^{\scriptsize op}$
is defined in ${\bf R}(G)$.
We have that
$$\theta (f,y)^{\scriptsize op}\theta (e,x)^{\scriptsize op}
=
(y^{-1},f)(x^{-1},e)
=
(y^{-1} \otimes x^{-1},e)
=
\theta (e,x \otimes y).$$
Thus $\theta$ is an isomorphism from ${\bf L}(G)^{\scriptsize op}$
to ${\bf R}(G)$.
\end{proof}
\subsection{Forward and back}
We shall now describe the relationship between ${\bf L} {\bf G} (C)$ and $C$, and between ${\bf G} {\bf L} (G)$ and $G$.
Our first theorem tells us that up to equivalence {\em every} left cancellative category can be constructed from an ordered groupoid.
\begin{theorem} Let $C$ be a left cancellative category.
Then $C' = {\bf L}{\bf G}(C)$ is a left cancellative category.
Define
$$\iota \colon \: C \rightarrow C'
\mbox{ by }
\iota (a) = ([\mathbf{r} (a),\mathbf{r} (a)],[a, \mathbf{d}(a)]).$$
Then $\iota$ is an injective functor which embeds
$C$ in $C'$ as a full, isomorphism-dense subcategory.
In particular, $C$ and ${\bf L}{\bf G}(C)$
are equivalent categories.
\end{theorem}
\begin{proof}
Identities of $C'$ have the form
$([z,z],[z,z])$ where $z \in C$.
Thus $\iota$ maps identities to identities.
Suppose that $ab$ is defined in $C$.
Then $\iota (a)\iota (b)$ is defined in $C'$.
From the fact that
$$[a,\mathbf{d} (a)] \otimes [b,\mathbf{d} (b)] = [ab,\mathbf{d} (b)]$$
we quickly deduce that
$\iota (ab) = \iota (a)\iota (b)$.
Thus $\iota$ is a functor.
Suppose that $\iota (a) = \iota (b)$.
Then there are isomorphisms $u,v \in C$ such that
$$\mathbf{r} (a) = \mathbf{r} (b)u, \quad a = bv \mbox{ and } \mathbf{d} (a) = \mathbf{d} (b)v.$$
Thus $v = \mathbf{d} (a)$ and so $a = b$.
Hence $\iota$ is injective.
Let $([c,c],[a,b])$ be an arbitrary element of $C'$ such that
$$\mathbf{d} ([c,c],[a,b]), \mathbf{r} ([c,c],[a,b]) \in \iota (C).$$
From $\mathbf{r} [a,b] \leq [c,c]$,
we get that $a = cp$ for some $p \in C$.
By our assumption, we have that $[b,b] = [e,e]$ and $[c,c] = [f,f]$ for some
$e,f \in C_{o}$.
Thus there are isomorphisms $u,v$ such that
$b = eu$ and $c = fv$.
Now
$$[c,c] = [\mathbf{r} (v), \mathbf{r} (v)]
\mbox{ and }
[a,b] = [vpu^{-1},\mathbf{r} (u)].$$
But $\iota (vpu^{-1}) = ([\mathbf{r} (v),\ f (v)],[vpu^{-1},\mathbf{r} (p)])$.
Thus $\iota$ is a full functor.
Finally, let $([a,a],[a,a])$ be an arbitrary identity in $C'$.
Consider the element $([a,a],[a,\mathbf{d} (a)])$ of $C'$.
It is an isomorphism such that
$$\mathbf{r} ([a,a],[a,\mathbf{d} (a)]) = ([a,a],[a,a])
\mbox{ and }
\mathbf{d} ([a,a],[a,\mathbf{d} (a)]) = \iota (\mathbf{d} (a)).$$
\end{proof}
We shall now describe the relationship between $G$ and ${\bf G} {\bf L} (G)$.
This is not as satisfactory and explains why we cannot put ordered groupoids and left cancellative categories on a par.
The ordered groupoid ${\bf G} {\bf L} (G)$ is a somewhat complicated object, so to help us see what is going on, we construct an isomorphic copy that is easier to understand.
Let $G$ be an ordered groupoid.
Define $\overline{G}$ as follows:
$$\overline{G} = \{\langle e,x,f \rangle \in G_{o} \times G \times G_{o} \colon \: \mathbf{d} (x) \leq f, \mathbf{r} (x) \leq e \}.$$
Define
$$\mathbf{d} (\langle e,x,f \rangle) = \, \langle f, \mathbf{d} (x), f \rangle
\mbox{ and }
\mathbf{r} (\langle e,x,f \rangle) = \, \langle e, \mathbf{r} (x), e \rangle,$$
and
$$\langle e,x,f \rangle^{-1} = \, \langle f,x^{-1},e \rangle.$$
These are all elements of $\overline{G}$.
If $\mathbf{d} \langle e,x,f \rangle = \mathbf{r} \langle i,y,j \rangle$,
define
$$\langle e,x,f \rangle \langle i,y,j \rangle = \langle e,xy,j \rangle.$$
Then $\overline{G}$ is easily seen to be a groupoid in which the identities are the elements of the form $\langle e,f,e \rangle$.
Define
$$\langle e,x,f \rangle \, \leq \, \langle e',x',f' \rangle$$
iff $e = e'$, $f = f'$ and $x \leq x'$.
Finally, define
$$\langle e,f,e \rangle^{\circ} = \langle e,e,e \rangle.$$
Observe that if $f \leq e$ where $e,f \in G_{o}$,
then $\mathbf{d} (\langle e,f,f \rangle) = \langle f,f,f \rangle$
and $\mathbf{r} (\langle e,f,f \rangle) = \langle e,f,e \rangle$.
\begin{lemma} Let $G$ be an ordered groupoid.
Then $\overline{G}$ is an ordered groupoid with maximal identities in which each $\mathcal{D}$-class contains a maximal identity.
The map $\kappa \colon \: \overline{G} \rightarrow G$ is an ordered functor that is order-reflecting.
\end{lemma}
The connection with ${\bf G} {\bf L} (G)$ is described in the following result.
\begin{proposition} Let $G$ be an ordered groupoid.
Then $G' = {\bf G} {\bf L} (G)$ is isomorphic to $\overline{G}$.
\end{proposition}
\begin{proof}
We show first that every element of $G'$
can be written uniquely in the form $[(e,w),(f,\mathbf{d} (w))]$ for some $w \in G$.
Let $[(e,x),(f,y)] \in G'$.
Then $\mathbf{d} (x) = \mathbf{d} (y)$ and $\mathbf{r} (x) \leq e$
and $\mathbf{r} (y) \leq f$.
The pair $(\mathbf{d} (x),y^{-1})$ is an isomorphism in $\L (G)$,
and
$$(e,x)(\mathbf{d} (x),y^{-1}) = (e,xy^{-1})
\mbox{ and }
(f,y)(\mathbf{d} (x),y^{-1}) = (f,\mathbf{r} (y)).$$
Put $w = xy^{-1}$.
We have proved that
$$[(e,x),(f,y)] = [(e,w),(f, \mathbf{d} (w))].$$
Suppose now that
$$[(e,w),(f,\mathbf{d} (w))] = [(e',w'),(f',\mathbf{d} (w'))].$$
Then there is an invertible element
$(\mathbf{d} (x),x)$ in ${\bf L} (G)$
such that
$$(f ,\mathbf{d} (w)) = (f',\mathbf{d}(w'))(\mathbf{d} (x),x)$$
where $\mathbf{d} (w') = \mathbf{d} (x)$.
It follows that $f = f'$ and $x = \mathbf{d} (w)$.
Thus $(\mathbf{d} (x),x)$ is an identity and so
$(e,w) = (e',w')$ and $(f,\mathbf{d} (w)) = (f',\mathbf{d} (w'))$.
It follows that $e = e'$, $f = f'$ and $w = w'$.
Define
$$\alpha \colon \: \overline{G} \rightarrow G'
\mbox{ by }
\alpha (\langle e,x,f \rangle) = [(e,x),(f ,\mathbf{d} (x))].$$
Then the two results above show that $\alpha$ is a bijection.
We shall prove that it is an isomorphism of ordered groupoids.
Suppose that $\langle e,x,f \rangle \, \leq \, \langle e,x',f \rangle$.
Then $x \leq x'$.
The pair $(\mathbf{d} (x'), \mathbf{d} (x)) \in {\bf L} (G)$
and
$$(e,x')(\mathbf{d} (x'),\mathbf{d} (x)) = (e,x)
\mbox{ and }
(f,\mathbf{d} (x))(\mathbf{d} (x'),\mathbf{d} (x)) = (f, \mathbf{d} (x)).$$
Thus $\alpha (\langle e,x,f \rangle) \, \leq \, \alpha (\langle e,x'f \rangle)$, and so
$\alpha$ is order-preserving.
Suppose that $\alpha (\langle e,x,f \rangle) \, \leq \, \alpha (\langle e',x',f' \rangle)$.
Then there is an element $(i,a) \in {\bf L} (G)$ such that
$$(e,x) = (e',x')(i,a)
\mbox{ and }
(f, \mathbf{d} (x)) = (f',\mathbf{d} (x'))(i,a).$$
It readily follows from this that
$e = e'$, $f = f'$ and $x \leq x'$.
Hence $\langle e,x,f \rangle \, \leq \, \langle e',x',f' \rangle$.
We have therefore proved that $\alpha$ is an order isomorphism.
It remains to show that $\alpha$ is a functor.
It is easy to check that $\alpha$ maps identities to identities.
Suppose that $\langle f,x,e \rangle \langle e,y,i \rangle$ is defined in $\overline{G}$.
By definition
$$\alpha (\langle f,x,e \rangle ) = [(f,x),(e,\mathbf{d}(x))]
\mbox{ and }
\alpha (\langle e,y,i \rangle) = [(e,y),(i,\mathbf{d} (y))].$$
Observe that $(e,\mathbf{d} (x)) = (e,y)(\mathbf{d}(y),y^{-1})$
where $(\mathbf{d} (y),y^{-1})$ is invertible.
It follows that
$\alpha (\langle f,x,e \rangle)\alpha (\langle e,y,i \rangle)$
is defined in $G'$
and is equal to
$[(f,x),(i,y^{-1})]$.
But
$$[(f,x),(i,y^{-1})] = [(f,xy),(i,\mathbf{d} (xy))] = \alpha (\langle f,xy,i \rangle)$$
making use of the isomorphism $(\mathbf{d} (x),y)$.
Thus $\alpha$ is a functor.
\end{proof}
To get a sharper connection betweeen $G$ and $\overline{G}$ we need to assume more about $G$.
\begin{proposition} Let $G$ be an ordered groupoid with maximal identities in which each $\mathcal{D}$-class contains
a maximal identity.
Then $\overline{G}$ is an enlargement of $G$.
\end{proposition}
\begin{proof} Define $\pi \colon G \rightarrow \overline{G}$ by $\pi (g) = (\mathbf{r} (g)^{\circ}, g, \mathbf{d} (g)^{\circ})$.
Then $\pi$ is an ordered embedding and $\overline{G}$ is an enlargement of the image of $\pi$.
\end{proof}
\subsection{Rooted categories}
The results we have obtained so far on ordered groupoids and left cancellative categories show that we almost have a correspondence between them.
But to get sharper results, we need to restrict the class of left cancellative categories we consider and correspondingly the class of ordered groupoids.
In both cases, we need an `anchor'.
A {\em weak terminal identity} in a category is an identity $1$ with the property that for each identity $e$ there is an arrow from $e$ to $1$.
\begin{lemma} Let $G$ be an ordered groupoid with maximum identity $1$.
Then ${\bf L} (G)$ has a weak terminal identity $(1,1)$.
\end{lemma}
\begin{proof}
Let $(e,e) \in {\bf L} (G)_{o}$.
Then $e \leq 1$ and so $(1,e) \in {\bf L} (G)_{o}$.
\end{proof}
Let $C$ be a left cancellative category.
The set $[a,a] \otimes {\bf G} (C) \otimes [a,a]$
is the set of all products $[a,a] \otimes [x,y] \otimes [a,a]$
where they are defined and forms an ordered subgroupoid of ${\bf G} (C)$.
\begin{lemma}
Let $C$ be a left cancellative category with weak terminal identity $1$.
Then $[a,b] \in [1,1] \otimes {\bf G} (C) \otimes [1,1]$
if and only if $\mathbf{r} (a) = 1 = \mathbf{r} (b)$.
\end{lemma}
\begin{proof}
Let $[a,b] \in [1,1] \otimes {\bf G}(C) \otimes [1,1]$.
Then $[a,a],[b,b] \leq [1,1]$.
It follows that there are $p,q \in C$
such that $a = 1p$ and $b = 1q$.
Thus, in particular, ${\bf r}(a) = {\bf r}(b) = 1$.
Conversely, suppose that
${\bf r}(a) = {\bf r}(b) = 1$.
Then $a = 1a$ and $b = 1b$ and so
$[a,a],[b,b] \leq [1,1]$.
Hence $[a,b] \in [1,1] \otimes {\bf G}(C) \otimes [1,1]$.
\end{proof}
Let $C$ be a left cancellative category with weak terminal identity $1$.
Put ${\bf G}_{1}(C) = [1,1] \otimes {\bf G}(C) \otimes [1,1]$.
\begin{proposition} Let $G$ be an ordered groupoid with maximum identity $1$.
Then $G$ is isomorphic to ${\bf G}_{1}({\bf L} (G))$.
\end{proposition}
\begin{proof}
From the proof of Proposition~3.10,
the function $\pi \colon \: G \rightarrow \bar{G}$
defined by
$$\pi (x) = \langle 1,x,1 \rangle$$
is a well-defined, injective ordered functor
since here $e^{\circ} = 1$ for each identity $e$ in $G$.
In addition, $G$ is isomorphic to $\pi (G)$.
Under the isomorphism $\alpha \colon \: \overline{G} \rightarrow {\bf G} {\bf L} (G)$
of Proposition~3.9,
we have that $\alpha \pi$ is contained in ${\bf G}_{1}{\bf L} (G)$.
Let $[(1,x),(1,y)]$ be an arbitrary element of ${\bf G}_{1}{\bf L} (G)$.
Then $(\mathbf{d} (y),y^{-1})$ is an isomorphism
and $(1,x)(\mathbf{d} (y),y^{-1}) = (1,xy^{-1})$
and $(1,y))(\mathbf{d} (y),y^{-1}) = (1,yy^{-1})$.
Thus
$$[(1,x),(1,y)] = [(1,xy^{-1}),(1,yy^{-1})]
=
\alpha \pi (xy^{-1}).$$
It follows that $G$ is isomorphic to ${\bf G}_{1}({\bf L} (G))$.
\end{proof}
We now cast the above results into a more usable form.
A left cancellative category $C$ with a weak terminal identity 1 is called a {\em left rooted category}.
Put
$$U = \{(a,b) \in C \times C \colon \: \mathbf{d} (a) = \mathbf{d} (b), \mathbf{r} (a) = 1 = \mathbf{r} (b)\}.$$
Observe that both $a$ and $b$ have codomain 1.
Define a relation $\sim$ on $U$ by
$$(a,b) \sim (a',b') \Leftrightarrow (a,b) = (a',b')u \mbox{ for some isomorphism } u$$
where $(a',b')u = (a'u,b'u)$.
Then $\sim$ is an equivalence relation on $U$.
Denote the equivalence class containing $(a,b)$ by $[a,b]$,
and the set of equivalence classes by ${\bf G}^{l}(C)$.
We may think of $[a,b]$ as modelling a partial bijection with domain of definition described by $[b,b]$ and with range described by $[a,a]$
both of these being subobjects of 1.
Define
$$\mathbf{d} [a,b] = [b,b], \quad \mathbf{r} [a,b] = [a,a]
\mbox{ and }
[a,b]^{-1} = [b,a].$$
If $\mathbf{d} [a,b] = \mathbf{r} [c,d]$.
Then there exists an isomorphism $u$ in $C$ such that $b = cu$.
In this case, define the partial product
$$[a,b] \cdot [c,d] = [a,du].$$
Define a relation $\leq$ on ${\bf G}^{l}(C)$
by
$$[a,b] \leq [c,d] \Leftrightarrow (a,b) = (c,d)p \mbox{ for some }p \in C.$$
This is well-defined and is a partial order.
In this way, ${\bf G}^{l}(C)$ is an ordered groupoid with a maximum identity.
Let $C$ and $D$ be two left rooted categories with weak terminal identities $1_{C}$ and $1_{D}$ respectively.
Let $\theta \colon \: C \rightarrow D$ be an equivalence of categories such that $F(1_{C}) = 1_{D}$.
Then the function
${\bf G}^{l}(\theta) \colon \: {\bf G}^{l}(C) \rightarrow {\bf G}^{l}(D)$
defined by ${\bf G}^{l}(\theta)([a,b]) = [\theta (a),\theta (b)]$
is an isomorphism of ordered groupoids with maximum identities.
We shall say that a left rooted category has {\em all allowable pullbacks}
if whenever $a$ and $b$ are elements of $C$ such that $\mathbf{r} (a) = \mathbf{r} (b)$
and that can be completed to a commutative square $aa' = bb'$ for some elements $a'$ and $b'$,
then $a$ and $b$ have a pullback.
It is worth noting that it is enough to assume this condition for those pairs $a$ and $b$ where in addition $\mathbf{r} (a) = \mathbf{r} (b)$.
In this case, ${\bf G}^{l} (C)$ is $\ast$-inductive and so when we adjoin a zero we get an inverse monoid with zero.
The ordered groupoid ${\bf G}^{l} (C)$ is inductive when $C$ has {\em all pullbacks}.
If $S$ is an inverse monoid with zero,
we denote by $\mathcal{L}(S)$ the left rooted category of the ordered groupoid of non-zero elements of $S$,
its right rooted category, denoted by $\mathcal{R}(S)$, is defined similarly.
We therefore have the following theorem with which is associated an evident dual theorem referring to right Leech categories.
\begin{theorem} \mbox{}
\begin{enumerate}
\item Let $G$ be an ordered groupoid with a maximum identity.
Then ${\bf L} (G)$ is a left rooted category and ${\bf G}^{l}({\bf L} (G))$ is isomorphic to $G$.
\item Let $C$ be a left rooted category.
Then ${\bf G}^{l}(C)$ is an ordered groupoid with a maximum identity
and ${\bf L} ({\bf G}^{l}(C))$ is a left rooted category equivalent to $C$.
\item Each inverse monoid with zero is determined by a left cancellative category with a weak terminal identity
that has all allowable pullbacks.
\item Each inverse monoid is determined by a left cancellative category with a weak terminal identity that has all pullbacks.
\end{enumerate}
\end{theorem}
\begin{proof} We prove (2).
For each identity $e$ in $C$ choose an arrow $c_{e} \colon e \rightarrow 1$.
We choose $c_{1} = 1$.
Let $f \stackrel{a}{\longleftarrow} e$ be an element of $C$.
Then $a$ and $c_{e}$ have the same domain.
Thus $c_{f}a$ and $c_{e}$ have both the same domain and range equal to 1.
Hence $[c_{f}a,c_{e}]$ is an element of ${\bf G}^{l}(C)$.
Also $[c_{f},c_{f}]$ is an identity of ${\bf G}^{l}(C)$.
By construction $\mathbf{r} [c_{f}a,c_{e}] \leq [c_{f},c_{f}]$.
Thus $([c_{f},c_{f}], [c_{f}a,c_{e}])$ is an element of ${\bf L} ({\bf G}^{l}(C))$.
Define $\theta \colon C \rightarrow {\bf L} ({\bf G}^{l}(C))$
by
$$\theta (a) = ([c_{f},c_{f}], [c_{f}a,c_{e}]).$$
The function $\theta$ is a functor that maps the weak terminal identity $1$ to the weak terminal identity $([1,1],[1,1])$
and is full, faithful and essentially surjective.\end{proof}
There is now an obvious question which has a familiar answer.
\begin{proposition}
Let $S$ be an inverse monoid.
Then its associated left rooted category is cancellative if and only if $S$ is $E$-unitary.
\end{proposition}
\begin{proof}
Suppose that $\mathcal{L}(S)$ is cancellative.
Let $e \leq s$.
Then
$$(1,s)(s^{-1}s,e) = (1,s^{-1}s)(s^{-1}s,e)$$
and so by right cancellation $s = s^{-1}s$, an idempotent as required.
Conversely, suppose that $S$ is $E$-unitary.
We prove that $\mathcal{L}(S)$ is right cancellative.
Let $(e,s)(f,u) = (e,t)(f,u)$.
Then $su = tu$.
It follows that $s^{-1}suu^{-1} = s^{-1}tuu^{-1}$ and so $s^{-1}t$ is an idempotent
using the fact that $S$ is $E$-unitary.
Similarly $st^{-1}$ is an idempotent.
Hence $s$ and $t$ are compatible but also $s^{-1}s = t^{-1}t$.
Thus by Lemma~2.24, we have that $s = t$ as required.
\end{proof}
It is interesting to consider what happens in the case when the inverse monoid is $0$-bisimple because we can then replace categories by monoids.
Let $S$ be $0$-bisimple.
Put $L_{1} = \{s \in S \colon s^{-1}s = 1 \}$.
Then $L_{1}$ is a left cancellative monoid.
It is isomorphic to the local monoid at the identity $(1,1)$ in the left rooted category $\mathcal{L}(S)$
and under our assumption on $S$ is actually equivalent to it.
We therefore have the following.
\begin{theorem}
Each $0$-bisimple inverse monoid with zero is determined by a left cancellative monoid
with the property that any two principal right ideals are either disjoint or their intersection is again a principal right ideal.
\end{theorem}
\begin{remark}
{\em What we have proved in this section for inverse monoids can also be generalized to inverse categories}
\end{remark}
\section{Affine systems}\setcounter{theorem}{0}
Inverse monoids and, more generally, ordered groupoids with maximum identities can be described by means of suitable left cancellative categories
equipped with weak terminal identities or by means of suitable right cancellative categories equipped with weak initial identities.
The problem now is describing arbitrary inverse semigroups or, more generally, arbitrary ordered groupoids.
The solution is similar to what happens when we want to replace vector spaces, which have a distinguished origin, with affine spaces which don't:
we have to work with actions.
In our case, and choosing sides, we work with (right cancellative) categories acting on the left on principal groupoids,
the groupoids arising from equivalence relations.
We shall see that our earlier description of inverse monoids or ordered groupoids with maximum identities is a special case.
In outline, we do the following
\begin{itemize}
\item A category $C$ acts on a principal groupoid $H$ on the left.
\item This action induces a preorder $\preceq$ on $H$ whose associated equivalence relation is $\equiv$.
\item The quotient structure $H/\equiv$ is a groupoid on which the preorder induces an order.
\item The groupoid $H/\equiv$ is ordered and every ordered groupoid is isomorphic to one constructed in this way.
\end{itemize}
\subsection{From category actions to ordered groupoids}
Before I give the formal definition, I want to motivate it by reconsidering how we defined an ordered groupoid
with identity from a {\em right} cancellative category $C$ with a weak {\em initial} identity 1.
Incidently, the shift from left to right rooted categories is simply a consequence of the fact that I shall work with {\em left} actions below.
Form the set of ordered pairs
$$U = \{(a,b) \in C \times C \colon \mathbf{r} (a) = \mathbf{r} (b), \mathbf{d} (a) = 1 = \mathbf{d} (b) \}.$$
We may define a partial action of $C$ on $U$ by defining $c \cdot (a,b) = (ca,cb)$ only when $\mathbf{d} (c) = \mathbf{r} (a)$.
If we define $\pi (a,b) = \mathbf{r} (a)$ then we can rewrite the condition for the action of $c$ on $(a,b)$ to be defined by $\mathbf{d} (a) = \pi (a,b)$.
What is $U$?
It is the groupoid corresponding to an equivalence relation defined on the set $C1$:
namely $a$ and $b$ are related if and only if $\mathbf{r} (a) = \mathbf{r} (b)$.
Suppose that $(a,b) = c(a',b')$ and $(a',b') = d(a,b)$.
Then because $C$ is right cancellative, the elements $c$ and $d$ are mutually invertible.
It follows that $(a,b) \sim (a',b')$.
Thus the action of $C$ on $U$ can be used to construct the elements of the ordered groupoid ${\bf G}^{r}(C)$.
In order to define $U$ in this case we needed a weak initial identity,
but the same construction would go through if we started with the pair $(G,U)$.
By axiomatizing the properties of such pairs leads to the construction of ordered groupoids
from categories acting on principal groupoids which we now describe.
Let $C$ be a category and $G$ a groupoid.
Let $\pi \colon \: G \rightarrow C_{o}$ be a function to the set of identities of $C$.
Define
$$C \ast G = \{(a,x) \in C \times G \colon \: \mathbf{d} (a) = \pi (x) \}.$$
We say that $C$ {\em acts on} $G$ if there is a function from
$C \ast G$ to $G$, denoted by $(a,x) \mapsto a \cdot x$, which satisfies the axioms (A1)--(A6) below.
Note that I write $\exists a \cdot x$ to mean that $(a,x) \in C \ast G$.
I shall also use $\exists$ to denote the existence of products in the categories
$C$ and $G$.
\begin{description}
\item[{\rm (A1)}] $\exists \pi (x) \cdot x$ and $\pi (x) \cdot x = x$.
\item[{\rm (A2)}] $\exists a \cdot x$ implies that $\pi (a \cdot x) = \mathbf{r} (a)$.
\item[{\rm (A3)}] $\exists a \cdot (b \cdot x)$ iff $\exists (ab) \cdot x$,
and if they exist they are equal.
\item[{\rm (A4)}] $\exists a \cdot x$ iff $\exists a \cdot \mathbf{d} (x)$, and if they exist then
$\mathbf{d} (a \cdot x) = a \cdot \mathbf{d} (x)$;
$\exists a \cdot x$ iff $\exists a \cdot \mathbf{r} (x)$,
and if they exist then $\mathbf{r} (a \cdot x) = a \cdot \mathbf{r} (x)$.
\item[{\rm (A5)}] If $\pi (x) = \pi (y)$ and $\exists xy$ then $\pi (xy) = \pi (x)$.
\item[{\rm (A6)}] If $\exists a \cdot (xy)$ then $\exists (a \cdot x)(a \cdot y)$
and $a \cdot (xy) = (a \cdot x)(a \cdot y)$.
\end{description}
We write $(C,G)$ to indicate the fact that $C$ acts on $G$.
If $C$ acts on $G$ and $x \in G$ define
$$C \cdot x = \{a \cdot x \colon \: \exists a \cdot x \}.$$
Define $x \preceq y$ in $G$ iff there exists $a \in C$ such that $x = a \cdot y$.
The relation $\preceq$ is a preorder on $G$.
Let $\equiv$ be the associated equivalence:
$x \equiv y$ iff $x \preceq y$ and $y \preceq x$.
Observe that $x \preceq y$ iff $C \cdot x \subseteq C \cdot y$.
Thus $x \equiv y$ iff $C \cdot x = C \cdot y$.
Denote the $\equiv$-equivalence class containing $x$ by $[x]$,
and denote the set of $\equiv$-equivalence classes by $J(C,G)$.
The set $J(C,G)$ is ordered by $[x] \leq [y]$ iff $x \preceq y$.
\begin{remarks}{\em \mbox{}
\begin{enumerate}
\item Axioms (A1), (A2) and (A3) are the usual axioms for the action of a category on a set.
\item If $x \equiv y$ then ${\bf d}(x) \equiv {\bf d}(y)$ and ${\bf r}(x) \equiv {\bf r}(y)$ by axiom (A4).
\item For each $e \in C_{o}$ put $G_{e} = \pi^{-1}(e)$.
Let $x \in G_{e}$.
Then $\pi (x) = e$.
Thus $\exists e \cdot x$ and so by axiom (A4), we also have that $\exists e \cdot \mathbf{d} (x)$ and $\exists e \cdot \mathbf{r} (x)$.
Thus $\mathbf{d} (x), \mathbf{r} (x) \in G_{e}$.
Also $\exists \mathbf{d} (x)$ means that $e \cdot (x^{-1} x)$ and so by axiom (A6),
we have that $\exists e \cdot x^{-1}$ and so $x^{-1} \in G_{e}$.
By axiom (A5), if $x,y \in G_{e}$ and $xy$ is defined then $xy \in G_{e}$.
It follows that $G_{e}$ is a subgroupoid of $G$,
and by axiom (A4) it must be a union of connected components of $G$.
If $x$ is an identity in $G$ and $\exists a \cdot x$ then $a \cdot x$ is an identity in $G$.
This follows by (A4), since ${\bf d}(a \cdot x) = a \cdot {\bf d}(x) = a \cdot x$.
Combining this with axiom (A6), we see that if $f \stackrel{a}{\longleftarrow} e$ in $C$, then the function
$x \mapsto a \cdot x$ from $G_{e}$ to $G_{f}$ is a functor.
\end{enumerate}
}
\end{remarks}
We shall be interested in actions of categories $C$ on groupoids $G$ that satisfy two further conditions:
\begin{description}
\item[{\rm (A7)}] $G$ is principal.
\item[{\rm (A8)}] $\mathbf{d} (a \cdot x) = \mathbf{d} (b \cdot x)$ iff $\mathbf{r} (a \cdot x) = \mathbf{r} (b \cdot x)$.
\end{description}
Condition (A7) is to be expected;
condition (A8) will make everything work, as will soon become clear.
The axioms (A7) and (A8) together imply that if $\mathbf{d} (a \cdot x) = \mathbf{d} (b \cdot x)$ then $a \cdot x = b \cdot x$.
\begin{theorem} Let $C$ be a category acting on the groupoid $G$,
and suppose in addition that both (A7) and (A8) hold.
Then
\begin{enumerate}
\item $J(C,G)$ is an ordered groupoid.
\item $J(C,G)$ is $\ast$-inductive iff for all identities $e,f \in G$ we have that
$C \cdot e \cap C \cdot f$ non-empty implies there exists an identity $i$
such that $C \cdot e \cap C \cdot f = C \cdot i$.
\end{enumerate}
\end{theorem}
\begin{proof}
(1) Define
$$\mathbf{d} [x] = [\mathbf{d} (x)]
\mbox{ and }
\mathbf{r} [x] = [\mathbf{r} (x)].\footnote{Strictly speaking, I should write
$\mathbf{d} ([x])$ but I shall omit the outer pair of brackets.}$$
These are well-defined by Remarks~4.1(2).
We claim that $\mathbf{d} [x] = \mathbf{r} [y]$ iff
there exists $x' \in [x]$ and $y' \in [y]$ such that $\exists x'y'$.
To prove this, suppose first that $\mathbf{d} [x] = \mathbf{r} [y]$.
Then $\mathbf{d} (x) \equiv \mathbf{r} (y)$.
There exist elements $a,b \in C$ such that
$\mathbf{d} (x) = a \cdot \mathbf{r} (y)$ and $\mathbf{r} (y) = b \cdot \mathbf{d} (x)$.
Thus by (A3) and (A4), we have that
$$\mathbf{r} (b \cdot (a \cdot y)) = \mathbf{r} (y).$$
By (A8), this implies that
$$\mathbf{d} (b \cdot (a \cdot y)) = \mathbf{d} (y).$$
By (A7), this means that $y = b \cdot (a \cdot y)$.
Hence $y \equiv a \cdot y$ and $\exists x (a \cdot y)$, as required.
The converse follows by Remarks~4.1(2).
We define a partial product on $J(C,G)$ as follows:
if $\mathbf{d} [x] = \mathbf{r} [y]$ then
\begin{center}
$[x][y] = [x'y']$ where $x' \in [x]$, $y' \in [y]$ and $\exists x'y'$,
\end{center}
otherwise the partial product is not defined.
To show that it is well-defined we shall use (A7) and (A8).
Let $x'' \in [x]$ and $y'' \in [y]$ be such that $\exists x''y''$.
We show that $x'y' \equiv x''y''$.
By definition there exist $a,b,c,d \in C$ such that
$$
x' = a \cdot x,
\quad
x = b \cdot x',
\quad
x'' = c \cdot x,
\quad
x = d \cdot x''
$$
and there exist $s,t,u,v \in C$ such that
$$
y' = s \cdot y,
\quad
y = t \cdot y',
\quad
y'' = u \cdot y,
\quad
y = v \cdot y''.$$
Now
$x = b \cdot x'$ and $x'' = c \cdot x$.
Thus $x'' = (cb) \cdot x'$ by (A3).
Now $\exists x'y'$ and so
$\pi (x'y') = \pi (x')$ by (A5).
Thus $\exists (cb) \cdot (x'y')$.
Hence
$(cb) \cdot (x'y') = [(cb) \cdot x'][(cb) \cdot y']$ by (A6) which is $x''[(cb) \cdot y']$.
We shall show that $(cb) \cdot y' = y''$, which proves that $x''y'' \preceq x'y'$;
the fact that $x'y' \preceq x''y''$ holds by a similar argument so that $x'y' \equiv x''y''$ as required.
It therefore only remains to prove that $(cb) \cdot y' = y''$.
We have that
$y'' = (ut) \cdot y'$ and $\mathbf{d} (x'') = \mathbf{r} (y'')$.
Thus
$\mathbf{d} (x'') = \mathbf{r} (y'') = (ut) \cdot \mathbf{r} (y')$ by (A4).
But $\mathbf{d} (x'') = (cb) \cdot \mathbf{r} (y')$.
Thus $(ut) \cdot \mathbf{r} (y') = (cb) \cdot \mathbf{r} (y')$.
Hence
$$\mathbf{r} ((ut) \cdot y') = \mathbf{r} ((cb) \cdot y')$$
by (A4).
By axioms (A7) and (A8), it follows that
$(cb) \cdot y' = (ut) \cdot y' = y''$
and so the partial product is well-defined.
Thus $J(C,G)$ is a groupoid in which $[x]^{-1} = [x^{-1}]$, and the identities are the elements of the form $[x]$ where $x \in G_{o}$.
The order on $J(C,G)$ is defined by $[x] \leq [y]$ iff $x = a \cdot y$ for some $a \in C$.
It remains to show that $J(C,G)$ is an ordered groupoid with respect to this order.
(OG1) holds: let $[x] \leq [y]$.
Then $x = a \cdot y$.
By axiom (A4), we have that $\mathbf{d} (x) = a \cdot \mathbf{d} (y)$.
Thus $x^{-1}x = a \cdot (y^{-1}y)$.
By axiom (A6), $x^{-1}x = (a \cdot y^{-1})(a \cdot y)$.
Similarly $xx^{-1} = (a \cdot y)(a \cdot y^{-1})$.
But $G$ is a principal groupoid and so $x^{-1 } = (a \cdot y^{-1})$ and so $[x^{-1}] \leq [y^{-1}]$, as required.
(OG2) holds: let $[x] \leq [y]$ and $[u] \leq [v]$ and suppose that the partial products $[x][u]$ and $[y][v]$ exist.
Then there exist $x' \in [x]$, $u' \in [u]$, $y' \in [y]$ and $v' \in [v]$ such that
$[x][u] = [x'u']$ and $[y][v] = [y'v']$.
By assumption, $[x'] \leq [y']$ and $[u'] \leq [v']$ so that there exist $a,b \in C$ such that
$x' = a \cdot y'$ and $u' = b \cdot v'$.
We need to show that $x'u' \preceq y'v'$.
Now $\mathbf{d} (x') = \mathbf{r} (u')$ and so $a \cdot \mathbf{d} (y') = b \cdot \mathbf{r} (v')$.
But $\mathbf{d} (y') = \mathbf{r} (v')$.
Thus
$a \cdot \mathbf{d} (y') = b \cdot \mathbf{d} (y')$.
Hence
$$\mathbf{d} (a \cdot y') = \mathbf{d} (b \cdot y').$$
By (A8), we therefore have that
$$\mathbf{r} (a \cdot y') = \mathbf{r} (b \cdot y'),$$
and so $a \cdot y' = b \cdot y'$ by (A7).
Thus
$x'u' = (a \cdot y')(b \cdot v') = (b \cdot y')(b \cdot v')$.
Now $\exists y'v'$ and so by (A5) and (A6) we have that
$(b \cdot y')(b \cdot v') = b \cdot (y'v')$.
Thus $x'u' = b \cdot (y'v')$ and so $x'u' \preceq y'v'$, as required.
(OG3) holds: let $[e] \leq \mathbf{d} [x]$ where $e \in G_{o}$.
Then $e \preceq \mathbf{d} (x)$ and so $e = a \cdot \mathbf{d} (x)$ for some $a \in C$.
Now $\exists a \cdot x$ by (A4).
Define
$$([x]\,|\,[e]) = [a \cdot x].$$
Clearly $[a \cdot x] \leq [a]$,
and $\mathbf{d} [a \cdot x] = [a \cdot \mathbf{d} (x)] = [e]$.
It is also unique with these properties as we now show.
Let $[y] \leq [x]$ such that $\mathbf{d} [y] = [e]$.
Then $y = b \cdot x$ for some $b \in C$ and $\mathbf{d} (y) \equiv e$.
Because of the latter, there exists $c \in C$ such that $e = c \cdot \mathbf{d} (y)$.
Thus $e = (cb) \cdot \mathbf{d} (x)$.
But $e = a \cdot \mathbf{d} (x)$ and so $(cb) \cdot \mathbf{d} (x) = a \cdot \mathbf{d} (x)$.
By (A7) and (A8), we therefore have that $c \cdot y = a \cdot x$.
It follows that we have shown that $a \cdot x \preceq y$.
From $\mathbf{d} (y) \equiv e$, there exists $d \in C$
such that $\mathbf{d} (y) = d \cdot e$.
Using (A7) and (A8), we can show that $y = d \cdot (a \cdot x)$,
and so $y \preceq a \cdot x$.
We have therefore proved that $y \equiv a \cdot x$.
Hence $[y] = [a \cdot x]$, as required.
(OG3)$^{\ast}$ holds: although this axiom follows from the others,
we shall need an explicit description of the corestriction.
Let $[e] \leq \mathbf{r} [x]$ where $e \in G_{o}$.
Then $e \preceq \mathbf{r} (x)$ and so $e = b \cdot \mathbf{r} (x)$ for some $b \in C$.
Now $\exists b \cdot x$ by (A4).
Define
$$([e]\,|\,[x]) = [b \cdot x].$$
The proof that this has the required properties is similar to the one above.
(2) We now turn to the properties of the pseudoproduct in $J(C,G)$.
Let $[e],[f]$ be a pair of identities in $J(C,G)$.
It is immediate from the definition of the partial order that
$[e]$ and $[f]$ have a lower bound iff $C \cdot e \cap C \cdot f \neq \emptyset$.
Next, a simple calculation shows that
$[i] \leq [e],[f]$ iff $C \cdot i \subseteq C \cdot e \cap C \cdot f$.
It is now easy to deduce that
$[i] = [e] \wedge [f]$ iff $C \cdot i = C \cdot e \cap C \cdot f$.
It will be useful to have a description of the pseudoproduct itself.
If $C \cdot i = C \cdot e \cap C \cdot f$
then denote by
\begin{center}
$e \ast f$ and $f \ast e$
\end{center}
elements of $C$, not necessarily unique,
such that
$$i = (e \ast f) \cdot f = (f \ast e) \cdot e.$$
Suppose that $[x],[y]$ are such that the pseudoproduct $[x]\otimes [y]$ exists.
Then by definition $[\mathbf{d} (x)] \wedge [\mathbf{r} (y)]$ exists.
Thus $C \cdot \mathbf{d} (x) \cap C \cdot \mathbf{r} (y) = C \cdot e$ for some $e \in G_{o}$.
It follows that
$$[x] \otimes [y] = ([x]\,|\,[e])([e]\,|\,[y]).$$
Now
$$([x]\,|\,[e]) = [(\mathbf{r} (y) \ast \mathbf{d} (x)) \cdot x]$$
and
$$([e]\,|\,[y]) = [(\mathbf{d} (x) \ast \mathbf{r} (y))\cdot y].$$
Hence
$$[x] \otimes [y]
=
[((\mathbf{r} (y) \ast \mathbf{d} (x)) \cdot x)
((\mathbf{d} (x) \ast \mathbf{r} (y))\cdot y)
].
$$
\end{proof}
The condition that if $C \cdot e \cap C \cdot f$ is non-empty,
where $e$ and $f$ are identities,
then there exists an identity $i$ such that $C \cdot e \cap C \cdot f = C \cdot i$
will be called the {\em orbit condition} for the pair $(C,G)$.
Part (2) of Theorem~4.1 can therefore be stated thus:
$J(C,G)$ is $\ast$-inductive iff $(C,G)$ satisfies the orbit condition.
\subsection{Universality of the construction}
In this section, I shall show that every ordered groupoid is isomorphic to
one of the form $J(C,H)$ for some action of a category $C$ on a principal groupoid $H$.
Let $G$ be an ordered groupoid.
There are three ingredients needed to construct $J(C,H)$:
a category, which I shall denote by $\mathbf{R}(G)$, a principal groupoid, which I shall denote by $R(G)$,
and a suitable action of the former on the latter.
We define these as follows:
\begin{itemize}
\item We define the category ${\bf R} (G)$ as in Section~3 as follows:
an element of ${\bf R} (G)$ is an ordered pair $(x,e)$
where $(x,e) \in G \times G_{o}$ and $\mathbf{d} (x) \leq e$.
We define a partial product on ${\bf R} (G)$ as follows:
if $(x,e),(y,f) \in {\bf R} (G)$ and $e = \mathbf{r} (y)$ then
$(x,e)(y,f) = (x \otimes y, f)$.
Thus ${\bf R} (G)$ is a right cancellative category with identities $(e,e) \in G_{o} \times G_{o}$.
\item We define the groupoid $R(G)$ as follows:
its elements are pairs $(x,y)$ where $\mathbf{r} (x) = \mathbf{r} (y)$.
Define $\mathbf{d} (x,y) = (y,y)$ and $\mathbf{r} (x,y) = (x,x)$.
The partial product is defined by $(x,y)(y,z) = (x,z)$.
Evidently, $R(G)$ is the groupoid associated with the equivalence relation
that relates $x$ and $y$ iff $\mathbf{r} (x) = \mathbf{r} (y)$.
\item We shall now define what will turn out to be an action of ${\bf R} (G)$ on $R(G)$.
Define $\pi \colon \: R(G) \rightarrow {\bf R} (G)_{o}$ by $\pi (x,y) = (\mathbf{r} (x),\mathbf{r} (y))$,
a well-defined function.
Define $(g,e) \cdot (x,y) = (g \otimes x, g \otimes y)$ iff
$e = \mathbf{r} (x) = \mathbf{r} (y)$.
This is a well-defined function from ${\bf R} (G) \ast R(G)$ to $R(G)$.
\end{itemize}
\begin{proposition} Let $G$ be an ordered groupoid.
With the above definition, the pair $({\bf R} (G),R(G))$ satisfies axioms (A1)--(A8).
\end{proposition}
\begin{proof}
The verification of axioms (A1)--(A7) is routine.
We show explicitly that (A8) holds.
Suppose that
$$\mathbf{r} [(s,e) \cdot (x,y)] = \mathbf{r} [(t,e) \cdot (x,y)].$$
Then $s \otimes x = t \otimes x$.
The groupoid product $x^{-1}y$ is defined,
and the two ways of calculating the pseudoproduct of the triple $(s,x,x^{-1}y)$
are defined, and the two ways of calculating the pseudoproduct of the triple
$(t,x,x^{-1}y)$ are defined.
It follows that $s \otimes y = t \otimes y$;
that is,
$$\mathbf{d} [(s,e) \cdot (x,y)] = \mathbf{d} [(t,e) \cdot (x,y)].$$
The converse is proved similarly.
\end{proof}
The next theorem establishes what we would hope to be true is true.
\begin{theorem} Let $G$ be an ordered groupoid.
Then $J({\bf R} (G),R(G))$ is isomorphic to $G$.
\end{theorem}
\begin{proof}
Define $\alpha \colon \: G \rightarrow J({\bf R} (G),R(G))$ by $\alpha (g) = [(\mathbf{r} (g),g)]$.
We show first that $\alpha$ is a bijection.
Suppose that $\alpha (g) = \alpha (h)$.
Then $(\mathbf{r} (g),g) \equiv (\mathbf{r} (h),h)$.
Thus
$(a,\mathbf{r} (g)) \cdot (\mathbf{r} (g),g) = (\mathbf{r} (h),h)$
and
$(b,\mathbf{r} (h)) \cdot (\mathbf{r} (h),h)= (\mathbf{r} (g),g)$
for some category elements $(a,\mathbf{r} (g))$ and $(b,\mathbf{r} (h))$.
Hence
$$a \otimes \mathbf{r} (g) = \mathbf{r} (h),\quad
b \otimes \mathbf{r} (h) = \mathbf{r} (g), \quad
a \otimes g = h,
\mbox{ and }
b \otimes h = g.$$
It follows that $a$ and $b$ are identities
and so $h \leq g$ and $g \leq h$, which gives $g = h$.
Thus $\alpha$ is injective.
To prove that $\alpha$ is surjective,
observe that if $[(x,y)]$ is an arbitrary element of $J({\bf R} (G),R(G))$,
then $(x,y) \equiv (\mathbf{d} (x), x^{-1}y)$ because
$$(x^{-1},\mathbf{r} (x))\cdot (x,y) = (\mathbf{d} (x),x^{-1}y)
\mbox{ and }
(x,\mathbf{d} (x)) \cdot (\mathbf{d} (x),x^{-1}y) = (x,y).$$
Next we show that $\alpha$ is a functor.
It is clear that identities map to identities.
Suppose that $gh$ is defined in $G$.
Now $\alpha (g) = [(\mathbf{r} (g),g)]$ and $\alpha (h) = [(\mathbf{r} (h),h)]$.
We have that
$\mathbf{d} [(\mathbf{r} (g),g)] = [(g,g)]$ and $\mathbf{r} [(\mathbf{r} (h),h)] = [(\mathbf{r} (h), \mathbf{r} (h))]$.
Now $(g,g) \equiv (\mathbf{d} (g), \mathbf{d} (g))$ because
$$(g^{-1},\mathbf{d} (g)) \cdot (g,g) = (\mathbf{d} (g),\mathbf{d} (g))$$
and
$$(g,\mathbf{d} (g)) \cdot (\mathbf{d} (g),\mathbf{d} (g)) = (g,g).$$
Thus $\alpha (g)\alpha (h)$ is also defined.
Now $(\mathbf{r} (h),h) \equiv (g,gh)$ because
$$(g,\mathbf{r} (h)) \cdot (\mathbf{r} (h),h) = (g,gh)$$
and
$$(g^{-1},\mathbf{r} (g)) \cdot (g,gh) = (\mathbf{r} (h),h).$$
Thus $\alpha (g) \alpha (h) = [(\mathbf{r} (g),gh)] = \alpha (gh)$.
It follows that $\alpha$ is a functor.
Finally, we prove that $\alpha$ is an order isomorphism.
Suppose first that $g \leq h$ in $G$.
Then $g^{-1} \leq h^{-1}$
and
$(\mathbf{d} (g)|h^{-1}) \leq h^{-1}$
and
$\mathbf{r} (\mathbf{d} (g)|h^{-1}) = \mathbf{d} (g) = \mathbf{r} (g^{-1})$.
Thus $(\mathbf{d} (g)|h^{-1}) = g^{-1}$.
It is now easy to check that
$(\mathbf{r} (g),g) = (g \otimes h^{-1},\mathbf{r} (h)) \cdot (\mathbf{r} (h),h)$.
Thus $\alpha (g) \leq \alpha (h)$.
Now suppose that $\alpha (g) \leq \alpha (h)$.
Then $(\mathbf{r} (g),g) = (a,\mathbf{r} (h)) \cdot (\mathbf{r} (h),h)$.
It follows that $a$ is an identity and that $g = a \otimes h$ and so $g \leq h$.
We have proved that $\alpha$ is an order isomorphism.
Hence $\alpha$ is an isomorphism of ordered groupoids.
\end{proof}
Theorem~4.4 tells us that ordered groupoids can be constructed from pairs $(C,G)$ satisfying some additional conditions.
First, we may assume that $C$ is right cancellative.
Second, $\pi \colon G \rightarrow C_{o}$ is a surjective map.
Third if $a \cdot x = b \cdot x$ then $a = b$;
we call this the {\em right cancellation condition}.
A pair satisfying these three additional condition is called an {\em affine system} and are the big sisters to the RP-systems of classical semigroup theory \cite{McA,Reilly}.
In an affine system, the equivalence relation $\equiv$ is determined by the isomorphisms in $C$;
for suppose that $y = a \cdot x$ and $x = b \cdot y$ then
$y = (ab) \cdot y$ and $x = (ba) \cdot x$ and so by the right cancellation condition
$a$ is invertible with inverse $b$.
\subsection{Morphisms between affine systems}
Let $(C,G)$ and $(D,H)$ be affine systems.
A {\em morphism} from $(C,G)$ to $(D,H)$ is a pair $\alpha = (\alpha_{1},\alpha_{2})$
where $\alpha_{1} \colon C \rightarrow D$ and $\alpha_{2} \colon G \rightarrow H$ are functors
such that if $a \cdot x$ is defined then $\alpha_{1} (a) \cdot \alpha_{2} (x)$ is defined and
$\alpha_{2} (a \cdot x) = \alpha_{1} (a) \cdot \alpha_{2} (x)$.
\begin{lemma} Let $\alpha \colon (C,G) \rightarrow (D,H)$ be a morphism of affine systems.
Then $\bar{\alpha} \colon J(C,G) \rightarrow J(D,H)$,
defined by $\bar{\alpha}[x] = [\alpha_{2}(x)]$,
is an ordered functor.
\end{lemma}
\begin{proof} Observe that if $x \preceq y$ then $\alpha_{2} (x) \preceq \alpha_{2} (y)$.
Thus the function $\bar{\alpha}$ is well-defined.
Since $\alpha_{2}$ is a functor $\bar{\alpha}$ maps identities to identities.
Suppose that $[x][y]$ is defined in $J(C,G)$.
Then there exist $x' \equiv x$ and $y' \equiv y$ such that $x'y'$ is defined in $G$.
Thus $\alpha_{2}(x)\alpha_{2}(y)$ is defined in $H$.
We have that
$$\bar{\alpha}([x][y]) = \bar{\alpha}[x'y'] = [\alpha_{2}(x'y')] = [\alpha_{2}(x')\alpha_{2}(y')] = [\alpha_{2}(x')][\alpha_{2}(y')]$$
which is just
$\bar{\alpha}[x] \bar{\alpha}[y]$.
Thus $\bar{\alpha}$ is a functor and it is an ordered by our first observation.
\end{proof}
A {\em morphism} $\alpha = (\alpha_{1},\alpha_{2})$ from $(C,G)$ to $(D,H)$ is said to be an {\em equivalence} if
the following three conditions hold:
\begin{description}
\item[{\rm (E1)}] $\alpha_{1}$ is an equivalence of categories.
\item[{\rm (E2)}] $\alpha_{2} (x) \preceq \alpha_{2}(y)$ implies $x \preceq y$.
\item[{\rm (E3)}] For each $y \in H$ there exists $x \in G$ such that $y \equiv \alpha_{2} (x)$.
\end{description}
\begin{lemma} Let $\alpha \colon (C,G) \rightarrow (D,H)$ be an equivalence of affine systems.
Then $\bar{\alpha} \colon J(C,G) \rightarrow J(D,H)$ is an isomorphism of ordered groupoids.
\end{lemma}
\begin{proof}
(E2) implies that $\bar{\alpha}$ is injective and (E3) that $\bar{\alpha}$ is surjective.
The condition (E2) also implies that $\bar{\alpha}$ is an order isomorphism.
\end{proof}
\subsection{Special cases and examples}
We sketch out a few special cases of the above theory and touch on some interesting examples.
We show first that the theory of ordered groupoids with a maximum identity, described in Section~3.4, is a special case of this new theory.
The proof of the following is immediate from the definitions.
It is not surprising given the motivation described at the beginning of Section~4.1.
\begin{proposition} Let $C$ be a right cancellative category with weak initial identity 1.
Put $X = C1$ and $X \ast X = \{(a,b) \in X \times X \colon \mathbf{r} (a) = \mathbf{r} (b) \}$.
Define an action of $C$ on $X \ast X$ by $a \cdot (x,y) = (ax, ay)$ if $\mathbf{d} (a) = \mathbf{r} (x)$.
Then $(C,X \ast X)$ is an affine system and $\mathbf{G}^{r}(C)$, the ordered groupoid constructed from
the right rooted category $C$ is isomorphic to $J(C,X \ast X)$.
\end{proposition}
An affine system $(C,G)$ is said to be {\em cyclic} if the following two conditions hold:
\begin{description}
\item[{\rm (C1)}] There exists $x_{0} \in G_{o}$ such that $G_{o} = C \cdot x_{0}$.
\item[{\rm (C2)}] If $a,b \in C$ such that $\mathbf{r} (a) = \mathbf{r} (b)$ and $\mathbf{d} (a) = \mathbf{d} (b) = \pi (x_{0})$ then
there exists $g \in G$ such that $\mathbf{r} (g) = a \cdot x_{0}$ and $\mathbf{d} (g) = b \cdot x_{0}$.
\end{description}
Ordered groupoids with maximum identity correspond to cyclic affine systems.
Our second special case deals with the situation where our ordered groupoid is connected.
We consider an affine system $(C,G)$ where $C$ is a right cancellative monoid and $G = X \times X$ is a universal principal groupoid.
It follows that in fact we have a left monoid action of $C$ on $X$ which satisfies the right cancellation condition.
We call $(C,X)$ an {\em affine monoid system}.
\begin{proposition}
Affine monoid systems describe connected ordered groupoids.
\end{proposition}
Affine systems lead to a natural description of arbitrary inverse semigroups with zero.
Let $S$ be an inverse semigroup with zero and let $\mathbf{R} = \mathbf{R}(S^{\ast})$,
the right cancellative category associated with the non-zero elements of $S$.
The principal groupoid $R = R(S^{\ast})$ consists of those pairs $(s,t)$ such that
$s$ and $t$ are both non-zero and $ss^{-1} = tt^{-1}$.
By Theorem~4.4, the ordered groupoid $S^{\ast}$ is isomorphic to $J(\mathbf{R},R)$.
Thus the inverse semigroup $S$ is isomorphic to $J(\mathbf{R},R)^{0}$ equipped with the pseudoproduct.
We may summarize these results as follows.
\begin{theorem} Every inverse semigroup with zero $S$ is determined upto isomorphism by three ingredients:
the right cancellative category $\mathbf{R}(S^{\ast})$, Green's $\mathcal{R}$-relation on the non-zero elements,
and the action of the category on the groupoid determined by Green's $\mathcal{R}$-relation.
\end{theorem}
One obvious question is how categories acting on principal groupoids arise.
We now describe one example.
Let $(C,X)$ be a pair consisting of a category $C$ acting on a set $X$
where we denote by $\pi \colon \: X \rightarrow C_{o}$ the function used in
defining the action.
Define the relation $\mathcal{R}^{\ast}$ on the set $X$ as follows:
$x \,\mathcal{R}^{\ast}\, y$ iff $\pi (x) = \pi (y)$ and for all $a,b \in C$
we have that, when defined,
$$a \cdot x = b \cdot x \Leftrightarrow a \cdot y = b \cdot y.$$
Observe that $\mathcal{R}^{\ast}$ is an equivalence relation on the set $X$.
In addition, $x \,\mathcal{R}^{\ast}\, y$ implies that
$c \cdot x \,\mathcal{R}^{\ast}\, c \cdot y$
for all $c \in C$ where $c \cdot x$ and $c \cdot y$ are defined.
Consequently, we get a principal groupoid
$$G(C,X) = \{(x,y) \colon \: x \,\mathcal{R}^{\ast}\, y \}.$$
Define $\pi' \colon \: G(C,X) \rightarrow C_{o}$ by
$\pi' (x,y) = \pi (x)$,
and define an action of $C$ on $G(C,X)$ by
$a \cdot (x,y) = (a \cdot x, a \cdot y)$
when ${\bf d}(a) = \pi' (x,y)$.
It is easy to check that axioms (A1)--(A8) hold.
We may therefore construct an ordered groupoid from the pair
$(C,G(C,X))$.
This is identical to the ordered groupoid constructed in
\cite{Lawson99} directly from the pair $(C,X)$.
Constructing inverse semigroups from rooted categories and affine systems has practical applications.
One can try to relate the categorical properties of the rooted category to the algebraic properties of the inverse monoid \cite{Leech98}.
It also leads to the perspective that inverse semigroup theory can be viewed not just as the abstract theory
of partial bijections but also the abstract theory of bijections between quotients \cite{F,FL}.
The theory restricted to a class of $0$-bisimple inverse monoids allows self-similar group actions \cite{Nek} to be
described in terms of a class of inverse semigroups \cite{Lawson08}.
Graph inverse semigroups can be constructed from free categories \cite{Lawson99}.
Such inverse semigroups are important in the theory of $C^{\ast}$-algebras \cite{Raeburn}.
Finally, general affine systems \cite{Lawson06} can be used to better understand the nature of the inverse semigroups that Dehornoy first constructed \cite{Dehornoy93}.
\section{Notes on Chapter 2}
The theory described in Section~2 originates in Ehresmann's work \cite{Ehresmann} and was first developed
within inverse semigroup by Boris Schein \cite{Schein} and then successfully generalized
to arbitrary regular semigroups by Nambooripad \cite{Nam}.
The maximum enlargement theorem comes under scrutiny in \cite{Steinberg} and \cite{Lawson02a}.
I wrote extensively about the ordered groupoid approach to inverse semigroup theory in my book \cite{Lawson98} so I will say no more here.
The origins of Section~3 go right back to the beginnings of inverse semigroup theory.
Clifford \cite{Clifford} showed how to describe inverse monoids in terms of left cancellative monoids.
The realization that this could be generalized to arbitrary inverse monoids by replacing left cancellative monoids by
left cancellative categories is due to Leech \cite{Leech} who gives duw credit to Logananthan's trail-blazing paper \cite{Log}
on how category theory can be applied to semigroup theory;
Loganathan showed that the cohomology of an inverse semigroup,
introduced by Lausch \cite{Lausch}, was the same as the cohomology of the associated left rooted category.
The extension of Clifford's original construction to $0$-bisimple inverse monoids is due to Reilly \cite{Reilly} and McAlister \cite{McA}.
The routine extension of Leech's construction to inverse monoids with zero may be found in \cite{Lawson99a}.
Note that the fact that this paper is labelled `I' and that the references refer to two further papers by the same author labelled `II' and `III' can safely be ignored.
The more general results on ordered groupoids were proved in \cite{Lawson04} and \cite{Lawson04b}.
In Section~2.4, I described how ordered groupoids could be used to find a structural description of $E$-unitary inverse semigroups.
In \cite{James}, the authors show how the left rooted category associated with an $E$-unitary
inverse monoid has a groupoid of fractions from which a proof of the $P$-theorem can also be deduced.
There are almost as many proofs of the $P$-theorem as there are semigroup theorists and a survey of some of them can be found in \cite{Lawson07}.
Chapter~X of Petrich \cite{Petrich} contains much more on the theory of bisimple inverse monoids.
Section~4 arises as unfinished business from Section~3.
We know how to construct inverse monoids with zero from left rooted categories but the question remains of how to deal with the semigroup case.
McAlister \cite{McA} and Reilly \cite{Reilly} discovered how to deal with bsimple inverse smeigroups and $0$-bisimple inverse semigroups
by using what were called {\em RP-systems} and {\em generalized RP-systems}.
As a result of reading a paper by Girard on linear logic \cite{Girard},
I was led to the construction described in \cite{Lawson99},
which shows how inverse semigroups can be constructed from
categories acting on {\em sets}.
Claas R\"over pointed out to me that the paper by Dehornoy \cite{Dehornoy93}.
Dehornoy constructs an inverse semigroup from any variety,
in the sense of universal algebra,
that is described by equations which are balanced,
meaning that the same variables occur on either side of the equation.
This construction was clearly related to the construction in \cite{Lawson99}
and an analysis of the connections between the two that led to \cite{Lawson05}
and the {\em affine systems} described in this section.
This raises the obvious question: what have inverse semigroups got to do with linear logic?
| a8a3b01f17e2efce8c0c8f941dd31a0532402c9c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\label{sec:NewNuclearData}New Nuclear Data}
\section{\label{sec:Introduction}Introduction}
The ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg is the main neutron source in evolved massive stars (M$>$10M$_\odot$) for the weak $s$-process (see Ref. \cite{pignatari:10}), producing most of the s-elements between iron and strontium (60$<$A$<$90). In low-mass AGB stars between 1M$_\odot$ and 4M$_\odot$ \cite{herwig:05} the reaction is activated during He-flash events, leaving its fingerprint in the final abundance of specific isotopes like ${^{87}}$Rb and ${^{96}}$Zr, which are directly observable through spectroscopy or laboratory measurements of meteoritic material \cite{herwig:05,battino:16,PALMERINI201821}. In both astrophysical scenarios, the ratio of the reaction rates of the two competing reactions $^{22}$Ne($\alpha,n$)$^{25}$Mg and $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg has the largest impact on the nucleosynthesis, determining the total neutron budget.
Since the last evaluation of the $^{22}$Ne$+\alpha$ reaction rates in 2012 by Longland and collaborators \cite{PhysRevC.85.065809}, there have been a number of experimental investigations of nuclear states in $^{26}$Mg with associated re-evaluations of the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rates. The various reaction rates resulting from these studies vary by up to a factor of 500 in the astrophysically relevant region \cite{Massimi20171,PhysRevC.93.055803}. An evaluation of the nuclear data culminating in newly calculated rates is required in order to resolve or identify the main sources of these discrepancies in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rates.
In performing the present evaluation we have chosen to use the \textsc{RatesMC} Monte-Carlo calculation code \cite{0067-0049-207-1-18} used previously by Longland {\it et al.} \cite{PhysRevC.85.065809}. In addition we adopt as a starting point the nuclear-data evaluation of Longland {\it et al.} \cite{PhysRevC.85.065809}. This ensures that any change in the reaction rates between the current evaluation and that of Longland {\it et al.} is the result of changes in the input nuclear data rather than in the method used to evaluate the reaction rates. This is the same approach used by Talwar {\it et al.} \cite{PhysRevC.93.055803} who also adopted the \textsc{RatesMC} code to evaluate the reaction rates.
In Sec~\ref{sec:NewNuclearData} the various new sources of nuclear data are introduced along with properties about $^{26}$Mg states that may be extracted from each experimental method. The available information about $^{26}$Mg are then discussed in Section \ref{sec:SynthesisOfTheData}. Particular focus was given to spin and parity assignments of observed $^{26}$Mg levels. Since both $^{22}$Ne and $\alpha$ particle have spin-parity $J^\pi=0^+$, the $^{22}$Ne$+\alpha$ reaction can populate only natural-parity states, {\it i.~e.} $0^+$, $1^-$, $2^+...$, and so only a subset of observed nuclear levels in $^{26}$Mg can contribute to the $^{22}$Ne($\alpha$,$n$)$^{25}$Mg and $^{22}$Ne($\alpha$,$\gamma$)$^{26}$Mg reaction rates. The resulting level assignments and nuclear data on levels in $^{26}$Mg are summarised in Section \ref{sec:LevelAssignments}.
The present characterisation of level structure in $^{26}$Mg was used for the evaluation of the reaction rates, discussed in Section~\ref{ReactionRateEvaluationSTARLIB}. Comparisons with previous reaction-rate evaluations together with the study of the contribution of individual $^{26}$Mg nuclear levels to the reaction rates are presented in Section \ref{RateComparisons}. Section \ref{sec:SuggestedFutureExperiments} presents priorities for future measurements, noting whence the major uncertainties in the reaction rates arise, and suggesting mitigating studies. In Section \ref{sec:impact} the suggested reaction rate is used in stellar models to see the effect on AGB stars.
\section{\label{sec:NewNuclearData}New Nuclear Data}
In the following section we briefly report recent measurements performed to study the $^{26}$Mg levels. In particular, we discuss recent experiments using neutron-induced reactions on $^{25}$Mg; proton, deuteron and $\alpha$-particle inelastic-scattering reactions; the $\alpha$-particle transfer reaction $^{22}$Ne($^6$Li,$d$)$^{26}$Mg and $\gamma$-ray spectroscopy following fusion-evaporation reactions. We present and compare the different experimental work, and we discuss the nuclear data that may be extracted from each technique.
Updated resonance parameters used in the present study are listed in Tables \ref{tab:KnownResonances} and \ref{tab:ULResonances} including the resonance strength, defined by:
\begin{equation}\label{eq:strength_alpha_gamma}
\omega\gamma_{(\alpha,\gamma)}=\omega\frac{\Gamma_\alpha\Gamma_\gamma}{\Gamma}
\end{equation}
and
\begin{equation}\label{eq:strength_alpha_n}
\omega\gamma_{(\alpha,n)}=\omega\frac{\Gamma_\alpha\Gamma_n}{\Gamma},
\end{equation}
where $\Gamma$ is the total width, $\Gamma_{\alpha,n,\gamma}$ are the $\alpha$-particle, neutron and $\gamma$-ray partial widths, respectively, and $\omega=(2J+1)/[(2I+1)(2i+1)]$ the statistical spin factor, where $J$, $I$ and $i$ are the spins of the resonant level, target and projectile, respectively. Since $I=i=0$, in this case $\omega = 2J+1$.
\subsection{$^{22}$Ne($\alpha,\gamma$)$^{26}$Mg direct measurement}
Recently, the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg resonance strength for the $E_r = 706$-keV resonance has been studied in direct kinematics at TUNL \cite{PhysRevC.99.045804}. This experiment used blister-resistant $^{22}$Ne-implanted targets, reducing the problems encountered when using an extended gas target \cite{Wolke1989}. This experiment confirmed the resonance in $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and provided a new measurement of the resonance energy ($E_{\alpha,\mathrm{lab}} = 835.2(30)$ keV) and the resonance strength ($\omega\gamma = 0.046(12)$ meV). No corresponding $^{22}$Ne($\alpha,n$)$^{25}$Mg data have been reported.
\subsection{$^{25}$Mg$+n$ transmission and capture reactions}
The $^{25}$Mg($n,\gamma$)$^{26}$Mg reaction cross section was studied at the neutron time-of-flight (TOF) facility n\_TOF at CERN, whilst the neutron total cross section was measured at the GELINA TOF facility at JRC-GEEL \cite{Massimi20171}. These data provide useful information on nuclear levels above the neutron threshold. These recent measurements aimed at solving some inconsistencies related to neutron data in the literature~\cite{PhysRevC.66.055805, PhysRevC.85.044615}. No information about levels of $^{26}$Mg below the neutron threshold are available from these experiments.
The n\_TOF facility at CERN generates neutrons using an high-energy proton beam incident upon a lead target. Neutrons pass down a 185-m flight path. At the end of the flight path, a detector system consisting of two C$_6$D$_6$ scintillation detectors placed on either side of the neutron beam is present. The scintillator detectors identify $\gamma$ rays resulting from $^{25}$Mg($n,\gamma$)$^{26}$Mg capture reactions. The time between the proton pulse hitting the lead target and the $\gamma$-ray detection in the scintillator detectors provides a measurement of the time of flight of the neutrons and thus the neutron energy.
The total $n+^{25}$Mg cross section was measured on the 50-m station at the GELINA facility. At GELINA, electrons from a linear accelerator are directed onto an uranium target producing Bremsstrahlung. Photoneutrons are generated from reactions of the Bremsstrahlung $\gamma$ rays with the uranium. Neutrons are detected in a $^6$Li-glass detector. The observable in this case is the proportion of neutrons transmitted through the target to the neutron detector as a function of neutron energy.
The energy-dependent cross sections obtained with the TOF technique have sub-keV energy resolution.
Resonant neutron scattering from $^{25}$Mg can be used to simultaneously obtain the excitation energies, spin-parities, and neutron and $\gamma$-ray partial widths of $^{26}$Mg levels above the neutron threshold. This is done using an R-matrix fit to the data. For the experiments described in Ref. \cite{Massimi20171}, there is a combination of neutron-transmission data (data from GELINA) and neutron-capture data (data from n\_TOF); the R-matrix analysis is performed simultaneously on both data.
\subsection{$^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg}
Two studies of the $^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg reaction have been carried out, one by Talwar {\it et al.} using the Grand Raiden magnetic spectrometer at RCNP Osaka \cite{PhysRevC.93.055803} and the other with the K600 magnetic spectrometer at iThemba LABS by Adsley {\it et al.} \cite{PhysRevC.96.055802}. Alpha-particle inelastic scattering from even-even targets is selective to isoscalar states with natural parity. Therefore, it is well suited to probe levels in $^{26}$Mg that contribute to the $^{22}$Ne$+\alpha$ reactions, which must also be isoscalar and of natural parity. The differential cross sections of these reactions allow the spin and parity of the populated states to be determined.
In both experiments, dispersion-matched beams of $E_\alpha = 200$ MeV were incident upon metallic enriched $^{26}$Mg targets. The scattered $\alpha$ particles were momentum analysed in the spectrometers. The focal-plane detection systems of both spectrometers consisted of drift chambers which provide information on the horizontal and vertical positions and trajectories of the particles detected at the focal plane. Paddle detectors made from plastic scintillator allow the time-of-flight through the spectrometer and the total residual energy to be measured; from these quantities the particles detected at the focal plane may be identified. In both experiments, the energy resolution was around 65 keV.
Talwar {\it et al.} \cite{PhysRevC.93.055803} measured up to scattering angles of approximately 12 degrees but the number of points in the differential cross sections is only 4, making it difficult to distinguish between different spin-parities. Adsley {\it et al.} \cite{PhysRevC.96.055802} only measured up to a scattering angle of 6 degrees, and assignments are therefore limited to $J=0$ and $J=1$.
Some discrepancies were observed between the results the two $^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg experiments \cite{PhysRevC.93.055803,PhysRevC.96.055802}. These will be discussed in Section \ref{sec:SynthesisOfTheData}.
\subsection{$^{22}$Ne($^6$Li,$d$)$^{26}$Mg}
In addition to studying the $^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg reaction, Talwar {\it et al.} also measured the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg reaction with the Grand Raiden spectrometer at RCNP Osaka \cite{PhysRevC.93.055803}. This reaction was performed with $E_{^6\mathrm{Li}} = 82.3$ MeV. The resolution of this experiment was around 100 keV due to the use of a gas-cell target.
The $^{22}$Ne($^6$Li,$d$)$^{26}$Mg differential cross sections are only available at one or two angles for many of the states in the region of interest. The difficulties in these types of measurements are due to the interplay between the poor energy resolution and the high level density in the region of interest. This, in turn, makes clear identification of the resonant levels difficult, limiting the possibility of linking them to other measurements available in the literature. In addition, as discussed in Section \ref{sec:SynthesisOfTheData} there are inconsistencies between this measurement and a previous $^{22}$Ne($^6$Li,$d$)$^{26}$Mg measurement \cite{GIESEN199395} which introduce considerable uncertainty into the reaction rates.
Two additional $^{22}$Ne($^6$Li,$d$)$^{26}$Mg datasets are also available. One of these experiments used sub-Coulomb barrier $\alpha$-particle transfer with a 1 MeV/u $^{22}$Ne beam \cite{JAYATISSA2020135267}. This method allows for less model-dependence in the extraction of the $\alpha$-particle partial width compared to traditional, higher-energy transfer reactions.
The other experimental study used the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg transfer reaction in inverse kinematics at a beam energy of 7 MeV/u \cite{OTA2020135256}. In this case the deuteron ejectiles were detected in an array of silicon detectors and the $^{25}$Mg and $^{26}$Mg recoils were detected at the focal plane of the MDM magnetic spectrograph. This allowed the decay branching of the populated states in $^{26}$Mg to be determined by comparing the number of deuterons associated with $^{25}$Mg and $^{26}$Mg recoils.
\subsection{$^{26}$Mg($p,p^\prime$)$^{26}$Mg and $^{26}$Mg($d,d^\prime$)$^{26}$Mg}
A study of the $^{26}$Mg($p,p^\prime$)$^{26}$Mg and $^{26}$Mg($d,d^\prime$)$^{26}$Mg reactions using 18-MeV proton and deuteron beams impinging on an enriched $^{26}$MgO target was performed using the Munich Q3D magnetic spectrograph \cite{PhysRevC.97.045807}. The proton inelastic-scattering reaction at these energies is highly unselective to the structure of the states of interest \cite{PhysRevC.89.065805,413,Moss1976429}, whilst the deuteron inelastic-scattering reaction is selective to isoscalar transitions, i.e., transitions to states with the same isospin as the ground state \cite{PhysRevC.97.045807,KAWABATA20076}. Due to the low incident energies these inelastic scattering processes are unable to give any information on the spin and parity of the resonance.
New states were observed in this experiment both above and below the neutron threshold, the existence of the states and their origin from $^{26}$Mg could be confirmed by ensuring that the kinematic shifts between angles were consistent with those expected for the $^{26}$Mg($p,p^\prime$)$^{26}$Mg reaction.
\subsection{$^{11}$B($^{16}$O,$p$)$^{26}$Mg}
Gamma-ray spectroscopy data using the Gammasphere array of high-purity germanium detectors located at Argonne National Laboratory following the $^{11}$B($^{16}$O,$p$)$^{26}$Mg reaction are available \cite{Lotay2019}. Spin and parity assignments of levels are made using the branching ratios and angular correlations of the observed $\gamma$ rays. Due to the relatively high angular momentum imparted to the system in this fusion-evaporation reaction no $J=0$ or $J=1$ states are observed.
\section{Combining the experimental data}
\label{sec:SynthesisOfTheData}
Having briefly introduced the updated nuclear data collected since the last evaluation of the rates, it is now necessary to build as consistent as possible a picture of all the nuclear data. In this way we are able to constrain some of the properties of possible resonances by combining data from multiple experiments.
In making new assignments updated with respect to those from Longland {\it et al.} \cite{PhysRevC.85.065809}, we adopt the following rules:
\begin{enumerate}
\item The excitation energies of $^{26}$Mg are preferentially taken from the resonant neutron-scattering measurements of Massimi {\it et al.} \cite{Massimi20171} and the photon-scattering measurements of Longland {\it et al.} \cite{PhysRevC.80.055803} as these data have better energy resolution.
\item For levels below the neutron threshold or for levels which were not observed in the resonant neutron-scattering data, the energies are preferentially taken from the high-resolution proton inelastic-scattering measurement of Adsley {\it et al.} \cite{PhysRevC.97.045807} or the $\gamma$-ray spectroscopy data \cite{Lotay2019}.
\item Due to the poor energy resolutions of the experiments using $\alpha$-particle inelastic scattering \cite{PhysRevC.96.055802,PhysRevC.93.055803} and $\alpha$-particle transfer \cite{PhysRevC.93.055803}, we do not use the excitation energies resulting from these experiments.
\item Spin-parity assignments above the neutron threshold are preferentially taken from Massimi {\it et al.} \cite{Massimi20171}.
\item Other spin-parity assignments are taken from the $^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg data of Adsley {\it et al.} \cite{PhysRevC.96.055802}, Talwar {\it et al.} \cite{PhysRevC.93.055803} and the Gammasphere data of Ref. \cite{Lotay2019}.
\item Due to the limited number of angles for the $\alpha$-particle transfer reaction of Talwar {\it et al.} \cite{PhysRevC.93.055803}, we approach the spin parities and spectroscopic factors derived from those data with some caution and, were possible, seek confirmation from another source.
\item Levels which have not been observed in the neutron resonant-scattering data of Massimi {\it et al.} \cite{Massimi20171} but have been observed in the proton and deuteron inelastic-scattering measurements of Adsley {\it et al.} \cite{PhysRevC.97.045807} have been assigned upper limits on neutron widths based on the sensitivity of the neutron-scattering experiments.
\item Resonance strengths for higher-lying resonances are taken from Longland {\it et al.} except for the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg resonance strength for the $E_{r}^{CM} = 706$-keV resonance, which has been recalculated using the results of a recently published direct measurement \cite{PhysRevC.99.045804}.
\item For lower-lying resonances in $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and the recalculation of the $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance strength for the $E_r = 706$-keV resonance, the results of the experiments at Texas A\&M by Jayatissa {\it et al.} \cite{JAYATISSA2020135267} and Ota {\it et al.} \cite{OTA2020135256} have been used for the calculations. When partial widths or resonance strengths are adopted from these studies is noted in the text.
\end{enumerate}
The states in $^{26}$Mg for which some re-determination must be made or some uncertainty exists are discussed below. We do not discuss levels for which no assignment is made or changed. Table \ref{tab:LevelInformationForDiscussion} gives detailed information on the sources of excitation energies, and spins and parities for some of the levels being discussed for ease of reference.
\begin{table*}[tphb]
\caption{Excitation energies, and spins and parities for some of the levels discussed in Section \ref{sec:SynthesisOfTheData}. When there are multiple experiments of the same type (e.g. $^{25}$Mg$+n$ data from Refs. \cite{Massimi20171} and \cite{PhysRevC.85.044615}) the most up-to-date values are given. The information in the table is linked to the index for the subsection in Section \ref{sec:SynthesisOfTheData}. The adopted excitation energy, and spin and parity (if available) are given along with the excitation energies, and spins and parities available from different sources.}
\label{tab:LevelInformationForDiscussion}
\begin{ruledtabular}
\begin{tabular}{c c c c c c}
Section & $E_x$ [MeV] & $J^\pi$ & Source 1 & Source 2 & Source 3 \\ \hline \hline \\
\ref{statesec1} & $10.6507(4)$ & $7^-$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802} \\ $E_x = 10.650(1)$ MeV,\\ no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019}\\$E_x = 10.6507(4)$ MeV,\\$J^\pi = 7^-$} & \\
\ref{statesec2} & $10.719(1)$ & $3^+$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802} \\ $E_x = 10.719(1)$ MeV\\ no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019} \\ $E_x = 10.7227(22)$ MeV\\ $J^\pi = (3^+)$} \\
\ref{statesec3} & $10.746(1)$ & $3^+$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802} \\ $E_x = 10.746(3)$ MeV\\ no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019} \\ $E_x = 10741.7(30)$ MeV\\ $J^\pi = 3^+$} \\
\ref{statesec4} & $10.771(1)$ & $(3-7)$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802} \\ $E_x = 10.771(1)$ MeV\\ no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019} \\ $E_x = 10765.1(30)$\\ J$^\pi = (3-7)$} \\
\ref{statesec5} & $10.8057(7)$ & $1^-$ & \makecell{$^{26}$Mg($\gamma,\gamma^\prime$)$^{26}$Mg \cite{PhysRevC.80.055803}\\$E_x = 10.8057(7)$ MeV\\$J^\pi = 1^-$} & \makecell{$^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802}\\$E_x = 10.806(10)$ MeV\\$J^\pi = 1^-$} & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802}\\
$E_x = 10.806(1)$ MeV\\no $J^\pi$}\\
\ref{statesec5} & $10.818(1)$ & $0^+$ & \makecell{$^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg \cite{PhysRevC.93.055803}/\cite{PhysRevC.96.055802} \\$E_x = 10.822(10)/10.824(10)$,\\$J^\pi = 0^+$} & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802}\\$E_x = 10.818(1)$ MeV,\\no $J^\pi$} \\
\ref{statesec5} & $10.826(1)$ & $(2^+)$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802}\\$E_x = 10.826(1)$ MeV\\ no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019}\\$E_x = 10.8226(30)$ MeV\\$J^\pi = 2^+$} \\
\ref{statesec8} & $10.915(1)$ & $2^+$ or $6^+$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802}\\
$E_x = 10.915(1)$ MeV\\
no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019}\\$E_x = 10.9128(30)$ MeV,\\ $J^\pi = 2^+$ or $6^+$} \\
\ref{statesec9} & $11.017(1)$ & $2^+, 3, 4, 5^+$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802} \\ $E_x = 11.017(1)$ MeV\\ no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019} \\ $E_x = 11.0173(36)$ MeV\\ $J^\pi = 2^+, 3, 4, 5^+$} \\
\ref{statesec10} & $11.084(1)$ & $2^+$ & \makecell{$^{26}$Mg($p,p^\prime$)$^{26}$Mg \cite{PhysRevC.96.055802} \\ $E_x = 11.084(1)$ MeV\\ no $J^\pi$} & \makecell{Gammasphere \cite{Lotay2019}\\$E_x = 11.0809(40)$ MeV,\\$J^\pi = 2^+$} \\
\ref{statesec11} & $11.16310(4)$ & $2^+$ & \makecell{$^{25}$Mg$+n$ \cite{Massimi20171}\\$E_x = 11.16310(4)$ MeV\\$J^\pi=2^+$} & & \\
\ref{statesec11} & $11.16926(4)$ & $3^-$ & \makecell{$^{25}$Mg$+n$ \cite{Massimi20171}\\$E_x = 11.16926(4)$ MeV\\$J^\pi = 3^-$} & \\
\ref{statesec11} & $11.1707(4)$ & $(2^+)$ & \makecell{Gammasphere \cite{Lotay2019}\\$E_x = 11.1717(30)$\\$J^\pi = (2^+)$} & \makecell{$^{25}$Mg$+n$ \cite{Massimi20171}\\$E_x = 11.1707(4)$ MeV\\no $J^\pi$} & \\
\ref{statesec12} & $11.321(1)$ & $0^+$ or $1^-$ & See text & \\
\end{tabular}
\end{ruledtabular}
\end{table*}
\subsubsection{$E_x = 10.6507$ MeV; $E_r = 36$ keV}
\label{statesec1}
The level at $E_x = 10.6507(4)$ MeV was assigned $J^\pi = 7^-$ from the Gammasphere data \cite{Lotay2019}. Due to the high spin of the state it has negligible astrophysical impact. Note that we do not assume that this state is the same as the $E_x = 10.650(1)$-MeV, $J^\pi = 1^+$ state assigned by Adsley {\it et al.} \cite{PhysRevC.97.045807} based on the $^{26}$Mg($p,p^\prime$)$^{26}$Mg data of Crawley {\it et al.} \cite{PhysRevC.39.311} as a $J=1$ state would not have been observed in the Gammasphere data \cite{Lotay2019}.
\subsubsection{$E_x = 10.719$ MeV; $E_r = 104$ keV}
\label{statesec2}
The level at $E_x = 10.719(2)$ MeV observed in the $^{26}$Mg($p,p^\prime$)$^{26}$Mg inelastic-scattering reaction at low energies \cite{PhysRevC.97.045807} was assigned as $J^\pi = 3^+$ under the assumption that it is the $E_x = 10.7227(22)$-MeV state observed in the Gammasphere experiment \cite{Lotay2019}.
\subsubsection{$E_x = 10.746$ MeV; $E_r = 131$ keV}
\label{statesec3}
The level at $E_x = 10.746(1)$ MeV observed in the $^{26}$Mg($p,p^\prime$)$^{26}$Mg inelastic-scattering reaction at low energies \cite{PhysRevC.97.045807} was assigned to $J^\pi = 3^+$ under the assumption that it is the $E_x = 10.7417(30)$-MeV state observed in the Gammasphere experiment \cite{Lotay2019}.
\subsubsection{$E_x = 10.771$ MeV; $E_r = 149$ keV}
\label{statesec4}
The level at $E_x = 10.7651(30)$~MeV was assigned to $J^\pi = (3-7)$ from the Gammasphere data \cite{Lotay2019}. This state is assumed to be the {$E_x = 10.771(1)$-MeV} state in the $^{26}$Mg($p,p^\prime$)$^{26}$Mg inelastic-scattering reaction at low energies \cite{PhysRevC.97.045807}.
\subsubsection{$E_x = 10.806$, $E_x = 10.818$ and $E_x = 10.826$ MeV; $E_r = 191$, $E_r = 203$ and $E_r = 211$ keV}
\label{statesec5}
The $\alpha$-particle inelastic-scattering measurement of Talwar {\it et al.} \cite{PhysRevC.93.055803} gave an assignment of $J^\pi = 1^-$ to the state observed at $E_x = 10.82$ MeV assuming it to be the $J^\pi = 1^-$ level observed by Longland {\it et al.} in a $\gamma$-ray inelastic-scattering measurement \cite{PhysRevC.80.055803}. However, Adsley {\it et al.} \cite{PhysRevC.96.055802}, also using $\alpha$-particle inelastic scattering, showed that there is a $J^\pi = 0^+$ level at around $E_x = 10.82$ MeV. Multiple levels were confirmed by the high-resolution proton and deuteron inelastic-scattering data of Adsley {\it et al.} \cite{PhysRevC.97.045807}. Adsley {\it et al.} argue that the third state at around this excitation energy observed in the high-resolution proton scattering experiment at Munich \cite{PhysRevC.97.045807} corresponds to a $J^\pi = 1^+$ level at $E_x = 10.81$ MeV observed in high-energy $^{26}$Mg($p,p^\prime$)$^{26}$Mg scattering data \cite{PhysRevC.39.311}.
Lotay {\it et al.} \cite{Lotay2019} give a tentative $J^\pi = 2^+$ assignment for the state at $E_x = 10.8226(30)$ MeV combining a $J = 2-6$ assignment from the Gammasphere data with the $^{26}$Mg($e,e^\prime$)$^{26}$Mg reaction study of Lees {\it et al.}, which finds only a $J^\pi = 2^+$ state in this excitation-energy region though with poor energy resolution.
The summary of the experimental studies is this: there are three levels at $E_x = 10.8057(7)$, $10.818(1)$ and $10.826(1)$ MeV based on the $^{26}$Mg($p,p^\prime$)$^{26}$Mg data \cite{PhysRevC.97.045807}. The level at $E_x = 10.8057(7)$ MeV has $J^\pi = 1^-$ \cite{PhysRevC.80.055803}. The level at $E_x = 10.826(1)$ MeV must have $J>1$ as it was observed in the Gammasphere study \cite{Lotay2019}. The result is therefore that the previous assignment of $J^\pi = 0^+$ to the $E_x = 10.826(1)$-MeV level was incorrect, and the $J^\pi = 0^+$ assignment should instead be associated with the $E_x = 10.818(1)$ MeV state.
Finally, for computing the reaction rates, the $\alpha$-particle partial widths must be estimated. The previous evaluation of Longland {\it et al.} used the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg data of Ugalde \cite{PhysRevC.76.025802} to calculate the $\alpha$-particle width for the $J^\pi = 1^-$ state at $E_x = 10.806$ MeV. Adsley {\it et al.} \cite{PhysRevC.96.055802} relaxed the constraints on the $\alpha$-particle partial widths since the states at these excitation energies are not resolved and it is not clear to which $^{26}$Mg state or states the transfer reaction proceeds. The sub-Coulomb transfer data of Jayatissa {\it et al.} \cite{JAYATISSA2020135267} show that a state is populated at around $E_x = 10.8$ MeV but it is not clear which of the three states in $^{26}$Mg is the one populated. We adopt the widths from Jayatissa {\it et al.} \cite{JAYATISSA2020135267} for the computation of the recommended reaction rate.
\subsubsection{$E_x = 10.915$ MeV; $E_r = 300$ keV}
\label{statesec8}
The level observed at $E_x = 10.9128(30)$ MeV in the Gammasphere data \cite{Lotay2019} was constrained as $J^\pi = 2^+$ or $6^+$. This state is assumed to be the level at $E_x = 10.915(1)$~MeV observed in the low-energy $^{26}$Mg($p,p^\prime$)$^{26}$Mg measurement of Adsley {\it et al.} \cite{PhysRevC.97.045807}.
As this state does not have a confirmed spin-parity assignment it has not been included in the current evaluation. However, this state does have a natural parity and could potentially contribute to the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate.
\subsubsection{$E_x = 11.017$ MeV; $E_r = 402$ keV}
\label{statesec9}
The level observed at $E_x = 11.0173(36)$ MeV in the Gammasphere data \cite{Lotay2019} was constrained to $J^\pi = 2^+, 3, 4, 5^+$. This state is assumed to be the level observed at $E_x = 11.017(1)$~MeV in the low-energy $^{26}$Mg($p,p^\prime$)$^{26}$Mg measurement of Adsley {\it et al.} \cite{PhysRevC.97.045807}.
This state does not have a confirmed spin-parity assignment and has therefore not been included in the current evaluation. It is not clear if this state has a natural parity and so it is unclear if it may contribute to the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate.
\subsubsection{$E_x = 11.084$ MeV; $E_r = 469$ keV}
\label{statesec10}
This state was observed in $^{26}$Mg($p,p^\prime$)$^{26}$Mg ($E_x = 11.084(1)$ MeV) reactions \cite{PhysRevC.97.045807,Moss1976429} and in the Gammasphere data \cite{Lotay2019} ($E_x = 11.809(40)$ MeV) in which an assignment of $J^\pi = 2^+$ is made. The $\alpha$-particle partial width was determined for this resonance to be: $\Gamma_\alpha = 5.7(15) \times 10^{-11}$ eV from sub-Coulomb $\alpha$-particle transfer \cite{JAYATISSA2020135267}.
\subsubsection{$E_x = 11.102$ MeV; $E_r = 487$ keV}
A state at $E_x = 11.102(1)$ MeV was observed in the $^{26}$Mg($p,p^\prime$)$^{26}$Mg data of Adsley {\it et al.} \cite{PhysRevC.97.045807}. This state lies $8$ keV above the neutron threshold but was not observed in the $^{25}$Mg$+n$ experiments of Massimi {\it et al.} \cite{Massimi20171}. Based on this non-observation it is possible to estimate the maximum neutron width for this state. Depending on the $\gamma$-ray partial width used in the calculation, the neutron partial width, $\Gamma_n$, must be below $2-4$ eV. No spin-parity assignment is available for this state.
\subsubsection{$E_x = 11.119$ MeV; $E_r = 504$ keV}
This state was observed in the $^{26}$Mg($p,p^\prime$)$^{26}$Mg data of Adsley {\it et al.} \cite{PhysRevC.97.045807}. This state lies $26$ keV above the neutron threshold but was not observed in the $^{25}$Mg$+n$ experiments of Massimi {\it et al.} \cite{Massimi20171}. Using the same methodology as for the $E_x = 11.102$-MeV state, the upper limit of the neutron partial width is found to be $\Gamma_n = 6-8$ eV. No spin-parity assignment is available for this state.
\subsubsection{$E_x = 11.163$, $E_x = 11.169$ and $E_x = 11.171$ MeV; $E_r = 548$, $E_r = 554$, $E_r = 556$ keV}
\label{statesec11}
\label{sec:11167keVState}
The region from $E_x = 11.163$ to $11.171$ MeV is extremely important, due to the possible observation of an $\alpha$-particle cluster state in the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg study of Talwar {\it et al.} \cite{PhysRevC.93.055803}. However, as we discuss below, there are outstanding problems in the nuclear data in this region and further confirmation of the proposed state of Talwar {\it et al.} is required.
First let us consider the assignments in this region which are simple to make. Two states have been observed in the $^{25}$Mg$+n$ experiments of Massimi {\it et al.} at $E_x = 11.163$ and $11.167$ MeV with $J^\pi = 2^+$ and $3^-$, respectively \cite{Massimi20171}.
A third state, at $E_x = 11.171$ MeV was also observed by Massimi {\it et al.} \cite{Massimi20171}, but no $J^\pi$ could be assigned. This state has comparable $\gamma$-ray and neutron partial widths, and is thus likely to have an observable $\gamma$-ray decay. This is therefore likely the state observed by Lotay {\it et al.} at $E_x = 11.1717(30)$ MeV and given a tentative $J^\pi = 2^+$ assignment \cite{Lotay2019}.
This third state has been associated with the strong $\alpha$-cluster candidate at $E_x = 11.17$ MeV observed by Talwar {\it et al.} in $^{22}$Ne($^6$Li,$d$)$^{26}$Mg reactions \cite{PhysRevC.93.055803}. This association was made because no corresponding strong $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance was observed in the direct measurements of Jaeger {\it et al.} \cite{PhysRevLett.87.202501}, meaning that the state cannot have a large $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance strength. This, in turn, means that the $\gamma$-ray partial width is larger than the neutron partial width. Talwar {\it et al.} also reported observing the $E_x = 11.31$-MeV state corresponding to the $E_r = 706$-keV resonance of Jaeger \cite{PhysRevLett.87.202501}, Wolke \cite{Wolke1989} and Giesen \cite{GIESEN199395}. In summary, Talwar {\it et al.} report two $\alpha$-cluster states at $E_x = 11.17$ and $11.31$ MeV, separated by around 140 keV.
There are older $^{22}$Ne($^6$Li,$d$)$^{26}$Mg data, those of Giesen {\it et al.} \cite{GIESEN199395} in which no $E_x = 11.17$-MeV $\alpha$-cluster state is observed. Rather two states are observed at $E_x = 11.31$ and $11.45$ MeV, separated by around 140 keV. This $E_x = 11.45$ MeV $\alpha$-cluster state matches a resonance observed in the direct measurement of Jaeger {\it et al.} at $E_x = 11.441$ MeV \cite{PhysRevLett.87.202501}. Unfortunately, Talwar {\it et al.} \cite{PhysRevC.93.055803} report that the $E_x = 11.44$-MeV state does not fall on the focal plane in their measurement, though it should not be entirely resolved from the $E_x = 11.31$-MeV state and so some indication of its population may be observable in their data.
In addition to the Giesen and Talwar $^{22}$Ne($^6$Li,$d$)$^{26}$Mg data, the more recent $^{22}$Ne($^6$Li,$d$)$^{26}$Mg results from Texas A\&M at 7 MeV/u (corresponding to a $^6$Li beam energy of 42 MeV) \cite{OTA2020135256} and the sub-Coulomb barrier transfer results at 1 MeV/u \cite{JAYATISSA2020135267} both fail to observe a strongly populated state at around $E_x = 11.17$ MeV. In the case of the sub-Coulomb transfer measurement, the non-observation of the state could be due to the state of interest having a high spin \cite{JAYATISSA2020135267}.
The higher-energy measurement from Texas A\&M (at 7 MeV/u) is at an energy between the Giesen and Talwar measurements, and a state at $E_x = 11.17$ MeV is observed. The change of beam energy between the various $^{22}$Ne($^6$Li,$d$)$^{26}$Mg measurements may be the cause of why the $E_x = 11.17$-MeV state is not seen in the experiments of Giesen {\it et al.} \cite{GIESEN199395} and Jayatissa {\it et al.} \cite{JAYATISSA2020135267}, though there is no clear evidence of the change in the population ratio of e.g. the $E_x = 11.31$-MeV and the $E_x = 11.45$-MeV states between Ota measurement \cite{OTA2020135256} and that of Giesen {\it et al.} \cite{OTA2020135256} which may be expected if the beam-energy changes were causing marked changes in the population of levels with different spins.
It may be coincidental that the two $^{22}$Ne($^6$Li,$d$)$^{26}$Mg measurements using magnetic spectrographs (the data of Talwar {\it et al.} and Giesen {\it et al.}) observe two $\alpha$-cluster states separated by approximately the same energy. However, it could be that one of the measurements has a problem in the excitation-energy calibration which may cause this discrepancy.
Given the discrepancies between the Talwar {\it et al.} \cite{PhysRevC.93.055803} result and the results of two above-barrier $^{22}$Ne($^6$Li,$d$)$^{26}$Mg measurements \cite{GIESEN199395,OTA2020135256}, one sub-barrier $^{22}$Ne($^6$Li,$d$)$^{26}$Mg measurement \cite{JAYATISSA2020135267}, and the direct $^{22}$Ne($\alpha,n$)$^{25}$Mg measurement of Jaeger \cite{PhysRevLett.87.202501}, we have decided not to use the spectroscopic factors derived from the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg measurement of Talwar {\it et al.} in our computation of the reaction rates. We have instead used the $J^\pi = 2^+$ assignment for the $E_x = 11.171$-MeV state and have used the experimental upper limits on the $\alpha$-particle partial widths from the sub-Coulomb barrier transfer measurement of Jayatissa {\it et al.} \cite{JAYATISSA2020135267} and the transfer-reaction of Ota {\it et al.} \cite{OTA2020135256} to constrain the resonance strengths for the states at $E_x = 11.163$, $11.169$ and $11.171$ MeV.
\subsubsection{$E_x = 11.27963(4)$ MeV; $E_r = 665$ keV}
The state at $E_x = 11.27963(4)$ MeV observed in resonant neutron scattering at $E_n = 194.01(2)$ keV has been reassigned, on the basis of the neutron-scattering data, from $J^\pi = 4^{(-)}$ to $J^\pi = 3^-$, a change from unnatural to natural parity. This state now potentially contributes to the reaction rates \cite{Massimi20171}.
\subsubsection{$E_x = 11.321$ MeV; $E_r = 706$ keV}
\label{statesec12}
This is the location of the lowest-energy directly measured resonances. Wolke {\it et al.} \cite{Wolke1989} measured a resonance in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction at $E_{\alpha,\mathrm{lab}} = 828(5)$ keV ($E_x = 11.315(4)$ MeV) with $\omega\gamma_{(\alpha,\gamma)} = 0.036(4)$ meV.
Hunt {\it et al.} measured a resonance in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction at $E_{\alpha,\mathrm{lab}} = 835.2(30)$ keV ($E_r = 705(3)$ keV, $E_x = 11.3195(25)$ MeV) with $\omega\gamma_{(\alpha,\gamma)}=0.046(12)$ meV. Based on the dipole-or-E2 rule of Endt \cite{ENDT19901}, a spin-parity assignment of $J^\pi = 0^+$, $1^-$, $2^+$ or $3^-$ is given.
Using the results of the measurements above, the weighted average of the resonance strength is $\omega\gamma = 0.037(4)$ meV. As the results of the two $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg measurements \cite{Wolke1989,PhysRevC.99.045804} agree within their uncertainties, we do not perform the procedure described in Ref. \cite{PhysRevC.85.065809} to account for unknown systematic biases within the direct measurements.
Jaeger {\it et al.} \cite{PhysRevLett.87.202501} measured a resonance in the $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction at $E_{\alpha,\mathrm{lab}} = 832(2)$ keV ($E_x = 11.319(2) $ MeV) with $\omega\gamma_{(\alpha,n)} = 0.118(11)$ meV with a total width of $\Gamma = \Gamma_n+\Gamma_\gamma+\Gamma_\alpha =0.25(17)$ keV.
The $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction has also been measured by Giesen {\it et al.} \cite{GIESEN199395}, Drotleff {\it et al.} \cite{1993ApJ...414..735D} and Harms {\it et al.} \cite{PhysRevC.43.2849}. Giesen {\it et al.} determined a resonance energy $E_r = 701(5)$ keV and a resonance strength $\omega\gamma = 0.234(77)$ meV. Drotleff {\it et al.} found $E_r = 703(3)$ keV and $\omega\gamma = 0.18(3)$ meV. Harms {\it et al.} measured $E_r = 702(3)$ keV and $\omega\gamma = 0.083(24)$ meV. Longland {\it et al.} \cite{PhysRevC.85.065809} used a method of estimating parameters and uncertainties including potential systematic effects, which resulted in $\omega\gamma = 0.14(3)$ meV.
From the weighted averages of the direct measurements described above, the $\alpha$-particle partial width may be computed resulting in $\Gamma_\alpha = 180(30) \mu$eV ($60(10) \mu$eV) assuming a $J^\pi = 0^+$ ($J^\pi = 1^-$) assignment.
The $\alpha$-particle partial width for this resonance has been measured with sub-Coulonb barrier $\alpha$-particle transfer at Texas A\&M \cite{JAYATISSA2020135267}, and found to be smaller than expected based on the weighted averages of the direct $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg measurements. This, coupled with a smaller $\Gamma_n/\Gamma_\gamma$ ratio determined by Ota {\it et al.}, detecting $^{25}$Mg and $^{26}$Mg recoils using the MDM magnetic spectrograph at Texas A\&M following the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg reaction in inverse kinematics leads to a reduced $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance strength. Jayatissa {\it et al.} \cite{JAYATISSA2020135267} give a resonance strength of $\omega\gamma_{(\alpha,n)} = 42(11)$ $\mu$eV combining the weighted average of the direct $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg measurements of the resonance strength with the neutron/gamma branching ratio measured by Ota {\it et al.} \cite{OTA2020135256}. The resonance strength for the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg channel remains unchanged.
The measurements of Jayatissa {\it et al.} \cite{JAYATISSA2020135267} and Ota {\it et al.} \cite{OTA2020135256} both favour $J^\pi = 0^+$ or $J^\pi = 1^-$ assignments, and we adopt $J^\pi = 0^+$ for our calculations; as the resonance is narrow the choice of the $J^\pi$ assignment does not change the calculated reaction rate.
A weighed average of the $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance strengths from the Jaeger \cite{PhysRevLett.87.202501}, Giesen \cite{GIESEN199395}, Drotleff \cite{1993ApJ...414..735D}, Jayatissa \cite{JAYATISSA2020135267} and Ota \cite{OTA2020135256} measurements gives a resonance strength of $\omega\gamma = 71(22)$ $\mu$eV, falling between the result for the TAMU experiments and the previous weighted average with a rate that falls between the rates calculated using the TAMU and Longland resonance strengths. However, as the two TAMU results are consistent with one another and differ from the Longland evaluation of the resonance strength by $3.1\sigma$ we instead evaluate the rate separately, using the results from the Texas A\&M experiments for one evaluation and the Longland weighted average for the other rather than using the weighted average of the global data.
There remains some dispute as to whether there is one or two resonances at this energy. Koehler \cite{PhysRevC.66.055805} argued that the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg resonances cannot be the same since, using the total width determined by Jaeger and the measured resonance strengths, the $\gamma$-ray partial width would be $\Gamma_\gamma = 56$ eV, much larger than a typical $\gamma$-ray partial width. However, Longland {\it et al.} \cite{PhysRevC.85.065809} noted that, due to the high uncertainty on the total width from Jaeger {\it et al.}, the uncertainty on the calculated $\Gamma_\gamma$ is extremely high, meaning that this argument cannot be used to conclude that the resonances are not the same.
Talwar {\it et al.} \cite{PhysRevC.93.055803} argued that the resonances observed in $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg are the same. The basis of their argument is that the experimental energy resolution of the study of Jaeger {\it et al.} leads to an upper limit of 0.42 keV for the total width. In addition, they cite the unpublished $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg data in the PhD thesis of Jaeger \cite{JaegerThesis} which uses the same resonance parameters as those of the $^{22}$Ne($\alpha,n$)$^{25}$Mg study \cite{PhysRevLett.87.202501}.
We note two additional pieces of evidence that the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg resonances originate from the same state. First, in the high-resolution $^{26}$Mg($p,p^\prime$)$^{26}$Mg data of Adsley {\it et al.} \cite{PhysRevC.97.045807} a state is observed at $E_x = 11.321(1)$ MeV, matching the energies of $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg resonances within uncertainties. No other state is observed at around this excitation energy.
Second, the non-observation of a level at around this energy in the $^{25}$Mg$+n$ scattering study by Massimi {\it et al.} \cite{Massimi20171} implies that the neutron partial width of the resonance is below the experimental sensitivity of the $^{25}$Mg$+n$ scattering study, which is of the order of 20 eV at $E_n = 234$ keV. As a consequence, the neutron partial width is much smaller than might be expected from the total width taken from Jaeger {\it et al.} and the $\gamma$-ray partial width need not to be abnormally high to reproduce the relative resonance strengths. This is consistent with the argument of Talwar {\it et al.} \cite{PhysRevC.93.055803} that the total width determined by Jaeger {\it et al.} \cite{PhysRevLett.87.202501} should instead be interpreted as an upper limit, and the real total width could be much smaller.
In conclusion, the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg resonances correspond to the same state in $^{26}$Mg and the reaction rate calculations are performed under this assumption in the present paper. The resonance energy is taken from the proton inelastic scattering data of Adsley {\it et al.} \cite{PhysRevC.97.045807}. The $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg resonance strength is taken from the weighted average of the direct measurements, and the $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance strength is taken from the weighted average of the direct measurements and the indirect Texas A\&M $^{22}$Ne($^6$Li,$d$)$^{26}$Mg measurements given by Jayatissa {\it et al.} \cite{JAYATISSA2020135267}.
\section{Tables of level assignments \label{sec:LevelAssignments}}
The available nuclear data and the properties of states in $^{26}$Mg are summarised in Tables \ref{tab:KnownResonances}-\ref{tab:UnknownResonances}.
Table \ref{tab:KnownResonances} contains the properties of resonances which are known to be natural parity and for which estimates or measurements of the $\alpha$-particle partial width exist. These include states which have been directly measured in $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions, and states which have been observed in $\alpha$-particle transfer reactions. In contrast to the evaluation of Longland {\it et al.} \cite{PhysRevC.85.065809}, when resonances have been observed in both $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions at the same energy we use the weighted average of the resonance energy, though for all resonances below the resonance at $E_{r,\mathrm{lab}} = 1433.7(12)$ keV, the resonance energies for the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions are identical.
\begin{table*}
\caption{Relevant data for resonances in $^{22}$Ne$+\alpha$ with known partial widths or resonance strengths. Resonance strengths and partial widths are taken from the evaluation of Longland {\it et al.} \cite{PhysRevC.85.065809} except for the $E_{r}^{CM} = 706$-keV resonance which has been recalculated using the results of a recently published study \cite{PhysRevC.99.045804} (see Section \ref{statesec12}). Excitation and resonance energies ($E_{r}^{CM}$ is the resonance energy in the $\alpha+^{22}$Ne centre of mass system) have been re-calculated where appropriate (see Section \ref{sec:SynthesisOfTheData}). The final column indicates whether the contribution of the resonance is numerically integrated to account for the energy dependence of the partial widths.\label{tab:KnownResonances}}
\begin{ruledtabular}
\begin{tabular}{ c c c c c c c c c }
\makecell{$E_x$\\{[MeV]}} & \makecell{$E_{r}^{CM}$\\{[keV]}} & $J^\pi$ & \makecell{$\omega\gamma_{(\alpha,\gamma)}$\\{[eV]}} & \makecell{$\omega\gamma_{(\alpha,n)}$\\{[eV]}} & \makecell{$\Gamma_\alpha$\\{[eV]}} & \makecell{$\Gamma_\gamma$\\{[eV]}} & \makecell{$\Gamma_n$\\{[eV]}} & \makecell{Integrate\\resonance?} \\ \\
\hline \\
$10.6963(4)$ & $81.6(4)$ & $4^+$ & & & $3.5(18) \times 10^{-46}$ & $3.0(15)$ & $0$ & No \\
$11.084(1)$ & $469(1)$ & $2^+$ & & & $5.7(1.5) \times 10^{-11}$ & $3.0(15)$& 0 & No \\
$11.321(1)$& $706(1)$\footnote{Resonance energy is taken using the state observed in Ref. \cite{PhysRevC.97.045807} and assuming a single resonance.} & $0^+/1^-$& $3.7(4) \times 10^{-5}$ & $4.2(11) \times 10^{-5}$ & & & & No \\
$11.44120(4)$& $826.46(5)$ & $3^-$ & &$3.9(10)\times10^{-5}$ & $5.50(14) \times 10^{-6}$ & $3.0(15)$ & $1.47(8) \times 10^{3}$ & Yes \\
$11.46574(6)$ & $851.00(6)$& $3^-$ & & $5.5(17)\times10^{-5}$ & $7.9(2.4) \times 10^{-6}$ & $3.0(15)$ & $6.55(9) \times 10^3$ & Yes\\
$11.5080(9)$ & $893.3(9)$ & $1^-$ & & $3.5(6)\times10^{-4}$& $1.2(4)\times10^{-4}$&$3.0(15)$ & $1.27(25)\times10^3$& Yes\\
$11.5260(15)$ & $911.3(15)$&$1^-$ & & $1.3(4)\times10^{-3}$ &$4.3(11)\times10^{-4}$ &$3.0(15)$ &$1.80(25)\times10^3$ & Yes\\
$11.630(1)$ & $1015.3(14)$ & $1^-$ & &$7.1(15)\times10^{-3}$ &$2.4(5)\times10^{-3}$ &$3.0(15)$ & $13.5(17)\times10^3$&Yes \\
$11.749(5)$ & $1133(6)$ & $1^-$ & & $5.9(8)\times10^{-2}$ &$2.0(3)\times 10^{-2}$ & $3.0(15)$& $64(9)\times10^3$&Yes \\
$11.787(3)$ & $1172(3)$ &$1^-$ & &$2.5(9)\times10^{-2}$ &$8(3)\times10^{-3}$ &$3.0(15)$ &$24.5(24)\times10^3$ &Yes \\
$11.828(1)$ & $1213(1)$ &$2^+$ & &$2.5(3)\times10^{-4}$ & $1.8(1)\times10^{-1}$ &$3.0(15)$ &$1.10(25)\times10^3$ &Yes \\
$11.863(3)$ & $1248(3)$ & $1^-$ & & & $1.5(10) \times 10^{-2}$ & $3.0(15)$ & $2.45(34) \times 10^4$ & Yes \\
$11.880(3)$ & $1265(3)$ & $1^-$ & & $1.9(19) \times 10^{-1}$ & $6.30(63) \times 10^{-2}$ & $3.0(15)$ & $3.0(15) \times 10^3$ & No \\
$11.895(4)$ & $1280(4)$ &$1^-$ & $2.0(2)\times10^{-3}$& $4.1(4)\times10^{-1}$& & & & No\\
$11.911(1)$ & $1297(3)$ &$1^-$ & $3.4(4)\times10^{-3}$&$1.4(1)$ &$1.9(9.8)$ &$3.0(15)$ &$5(2)\times10^3$ & Yes\\
$11.953(3)$ & $1338(3)$ &$2^+$ & $3.4(4)\times10^{-3}$& $1.60(13)$ & $3.2(1.7)\times 10^{-1}$ & $3.0(15)$ & $2(1) \times 10^3$ & Yes\\
$12.050(1)$ & $1436(3)$ &$2^+$ & $6.0(7)\times10^{-3}$&$4.7(3)$ &$1.1(3)\times10^{-1}$ &$3.0(15)$ &$4(1)\times10^3$ & Yes\\
$12.141(1)$ & $1526(3)$ &$1^-$ & $1.0(2)\times10^{-3}$&$2.4(2)$ & $1.7(5)$& $3.0(15)$& $1.5(2)\times10^{4}$& Yes\\
$12.184(5)$ & $1569(7)$ &$0^+$ & $1.1(2)\times10{-3}$& $1.21(29)\times10^1$& $0.90(11)$&$3.0(15)$ & $3.3(5)\times10^4$& Yes\\
$12.270(5)$ & $1658(7)$ &$0^+$ &$8.9(1)\times10^{-3}$ &$2.1(2)\times10^1$ &$2.2(4)\times10^2$ &$3.0(15)$ & $7.3(9)\times10^{4}$& Yes\\
$12.344(2)$ & $1728(4)$ &$0^+$ &$5.4(7)\times10^{-2}$ &$1.57(10)\times10^{2}$&$6.30(12)\times10^{2}$ &$3.0(15)$ & $3.5(5)\times10^{4}$ & Yes \\
\end{tabular}
\end{ruledtabular}
\end{table*}
Tables \ref{tab:ULResonances} contains properties of resonances which are known to have natural parity and for which the spin of the state is known. Therefore, these states can in principle contribute to the reaction rates but no estimate of the $\alpha$-particle partial width is available. The upper limits for the $\alpha$-particle partial widths for these states come from the Wigner limits or, in some cases, from $\alpha$-particle transfer-reaction data.
\begin{table*}
\caption{Nuclear data inputs for unobserved resonances in the $^{22}$Ne$+\alpha$ system with known natural parity. Alpha-particle partial widths are computed from the Wigner limits unless a lower estimate is available. Excitation and resonance energies ($E_{r}^{CM}$ is the resonance energy in the $\alpha-^{22}$Ne centre of mass system) have been re-calculated where required.\label{tab:ULResonances}}
\begin{ruledtabular}
\begin{tabular}{ c c c c c c c }
\makecell{$E_x$\\{[MeV]}} & \makecell{$E_r^{CM}$\\{[keV]}} & $J^\pi$ & \makecell{$\Gamma_{\alpha,UL}$\\{[eV]}} & \makecell{$\Gamma_\gamma$\\{[eV]}} & \makecell{$\Gamma_n$\\{[eV]}} & \makecell{Integrate\\resonance?} \\ \\
\hline \\
$10.6507(4)$ & $36.0(4)$ & $7^-$ & $1.60 \times 10^{-76}$ & $3.0(15)$ & 0 & No\\
$10.8057(7)$ & $191.0(7)$ & $1^-$ & $3.2 \times 10^{-23}$ & $0.72(18)$ & 0 & No \\
$10.818(1)$ & $203(1)$ & $0^+$ & $1.29\times 10^{-20}$ & $3.0(15)$ & 0 & No \\
$10.826(1)$ & $211(1)$ & $(2^+)$ & $6.65\times10^{-21}$ & $3.0(15)$ & 0 & No \\
$10.8976(47)$ & $278(1)$ & $(4^+)$ & $1.41\times10^{-18}$ & $3.0(15)$ & 0 & No \\
$10.9491(1)$ & $334.4(8)$ & $1^-$ & $2.90 \times 10^{-15}$ &$1.9(3)$ & 0 & No \\
$11.11223(4)$ & $497.49(5)$ & $2^+$ & $4.3 \times 10^{-10}$ & $1.37(6)\times10^{-2}$ &$2.095(5)\times10^3$ & Yes \\
$11.16310(4)$ & $548.36(5)$ & $2^+$ & $5.2\times10^{-9}$ & $2.8(2)$ & $5.31(5)\times10^3$ & Yes \\
$11.16926(4)$ & $554.52(5)$ &$3^-$ & $4.4\times10^{-10}$& $3.3(2)$ &$1.94(2)\times10^{3}$ & Yes \\
$11.17107(4)$ & $556.33(5)$ &$2^+$ & $1.3\times10^{-11}$& $3(2)$ & $0.8(7)$ & No \\
$11.27380(4)$ & $659.06(5)$ & $2^+$& $1.00\times10^{-6}$&$2.2(2)$ &$4.1(1)\times10^2$& Yes \\
$11.27963(4)$ & $664.89(5)$ & $3^-$&$9.20\times10^{-8}$ &$3(1)\times10^{-1}$ &$1.81(2)\times10^{3}$ & Yes \\
$11.30100(9)$ & $686.26(9)$ & $(2^+)$&$1.53\times10^{-5}$ &$<3$ & $<2.0\times10^1$& No \\
$11.32768(4)$ & $712.94(5)$ & $(1^-)$&$1.80\times10^{-6}$ &$2.2(3)$ &$1.71(6)\times10^2$ & Yes \\
$11.33696(4)$ & $722.22(5)$ &$(1^-)$ &$1.74\times10^{-4}$ & $<3$& $<2.0\times10^1$& No \\
$11.34389(9)$ & $729.15(9)$ & $(2^+)$& $1.10\times10^{-6}$&$1.0(2)$ &$<1.95\times10^3$ & Yes \\
$11.50022(4)$ & $885.48(5)$ & $1^-$&$1.95\times10^{-1}$ &$3.0(15)$ & $3.0(15)\times10^3$& Yes \\
\end{tabular}
\end{ruledtabular}
\end{table*}
States with unknown or uncertain spin-parities which could potentially contribute to the reaction rates are given in Table \ref{tab:UnknownResonances} along with potentially useful nuclear data such as constraints
on the partial widths for the state. Where available, known limits on the spin-parity of the states are also included.
\begin{table*}
\caption{States in $^{26}$Mg without confirmed spin-parity (i.e. unknown or tentative). Any known limit on the spin and parity of a state is included in the table. Note that this list is limited to levels between the $\alpha$-particle threshold and the lowest directly measured resonance at $E_{r}^{CM} =~706$-keV ($E_x = 11.321$ MeV). We assume that all important resonances at higher energy would have been observed in Refs. \cite{PhysRevLett.87.202501,Wolke1989}.\label{tab:UnknownResonances}}
\begin{ruledtabular}
\begin{tabular}{ c c c c }
\makecell{$E_x$\\{[MeV]}} & \makecell{$E_{r}^{CM}$\\{[keV]}} & \makecell{Possible\\$J^\pi$} & Comments\\ \\
\hline \\
$10.684(1)$ & $69(3)$& & \\
$10.7037(22)$ & $89(2)$ & $2^+,4$ & \\
$10.7089(22)$ & $94(2)$ & $3-6$ & \\
$10.730(1)$ & $115(1)$ & $3^+$ & Tentative assignment from the Gammasphere data \cite{Lotay2019} \\
$10.882(1)$ &$267(1)$ & & \\
$10.915(1)$ &$300(1)$ & $2^+/6^+$ & \\
$10.928(1)$ &$313(1)$ & & \\
$10.978(1)$ &$363(1)$ & & \\
$11.017(1)$ &$402(1)$ & $2-5$ & \\
$11.0386(20)$ & $424(3)$ & $2-6$ & \\
$11.047(1)$ &$432(1)$ & & \\
$11.074(1)$ &$459(1)$ & & \\
$11.102(1)$ &$487(1)$ & & $E_n = 9$ keV, $\Gamma_n < 3(1)$ eV \\
$11.1818(20)$ & $567(2)$ & $2-6$ & \\
$11.119(1)$ &$504(1)$ & & $E_n = 27$ keV, $\Gamma_n < 7(1)$ eV \\
\end{tabular}
\end{ruledtabular}
\end{table*}
\section{Reaction-rate evaluation using RatesMC}
\label{ReactionRateEvaluationSTARLIB}
The re-evaluation of the reaction rates was performed using the STARLIB Monte-Carlo code \textsc{RatesMC} \cite{0067-0049-207-1-18,PhysRevC.85.065809}, accessible online \cite{STARLIBWebsite}. This was done to ensure that the only variations observed in the reaction rates were due to changes in the nuclear data while the calculation methodology remained identical to that of Longland {\it et al.} \cite{PhysRevC.85.065809}. The resulting reaction rates are given in Tables \ref{tab:Ne22_alphagamma_rate} and \ref{tab:Ne22_alphan_rate}. The rates were evaluated using the resonances given in Table \ref{tab:KnownResonances} as measured resonance values and the resonances listed in Table \ref{tab:ULResonances} as upper limits for resonances.
To show the effect of the changes in the nuclear data for the $E_r = 706$-keV resonance, the reaction rates for $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg without the inclusion of the results from the Texas A\&M experiments were also computed. These reaction rate are given in Tables \ref{tab:Ne22_alphagamma_rate_no_TAMU} and \ref{tab:Ne22_alphan_rate_no_TAMU}. For details on how \textsc{RatesMC} estimates the reaction rates, refer to Refs. \cite{0067-0049-207-1-18,PhysRevC.85.065809}.
For the avoidance of doubt: the rates in the present paper which are recommended for use are those in Tables \ref{tab:Ne22_alphagamma_rate} and \ref{tab:Ne22_alphan_rate} which incorporate the results from the Texas A\&M experiments. We recommend the rates with the inclusion of data from the Texas A\&M measurements as these two experiments show consistent results, suggesting a reduction of the ${22}$Ne($\alpha,n$)$^{25}$Mg resonance strength for the $E_r = 706$-keV resonance.
\begin{table*}
\caption{Recommended $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg rate from the present evaluation incorporating the TAMU results. The low, median and high rates correspond to the 16\%, 50\% and 84\% values of the cumulative probability distribution. The $\mu$ and $\sigma$ parameters are resulting from fitting the distribution of rates at that temperature to a log-normal distribution. The Anderson-Darling statistic is a measure of how the data are well described by a log-normal distribution - see Refs. \cite{0067-0049-207-1-18,PhysRevC.85.065809} for details. All rates and the $\mu$ and $\sigma$ parameters are given in cm$^3$ mol$^{-1}$ s$^{-1}$. At temperatures above 1.33 GK, the rates should be taken from Hauser-Feshbach models - see Ref. \cite{PhysRevC.85.065809}.}
\begin{ruledtabular}
\begin{tabular}{c c c c c c c }
T [GK] & Low rate & Median rate & High rate & Log-normal $\mu$ & Log-normal $\sigma$ & Anderson-Darling statistic \\
\hline \\
0.010 & 3.38$\times$10$^{-79}$ & 5.52$\times$10$^{-79}$ &
9.04$\times$10$^{-79}$ & -1.802$\times$10$^{+02}$ &
5.01$\times$10$^{-01}$ & 2.73$\times$10$^{-01}$ \\
0.011 & 1.59$\times$10$^{-75}$ & 2.58$\times$10$^{-75}$ &
4.21$\times$10$^{-75}$ & -1.717$\times$10$^{+02}$ &
4.95$\times$10$^{-01}$ & 3.22$\times$10$^{-01}$ \\
0.012 & 1.82$\times$10$^{-72}$ & 2.93$\times$10$^{-72}$ &
4.75$\times$10$^{-72}$ & -1.647$\times$10$^{+02}$ &
4.89$\times$10$^{-01}$ & 3.22$\times$10$^{-01}$ \\
0.013 & 6.96$\times$10$^{-70}$ & 1.12$\times$10$^{-69}$ &
1.81$\times$10$^{-69}$ & -1.588$\times$10$^{+02}$ &
4.85$\times$10$^{-01}$ & 3.26$\times$10$^{-01}$ \\
0.014 & 1.13$\times$10$^{-67}$ & 1.81$\times$10$^{-67}$ &
2.93$\times$10$^{-67}$ & -1.537$\times$10$^{+02}$ &
4.83$\times$10$^{-01}$ & 3.25$\times$10$^{-01}$ \\
0.015 & 9.26$\times$10$^{-66}$ & 1.48$\times$10$^{-65}$ &
2.40$\times$10$^{-65}$ & -1.493$\times$10$^{+02}$ &
4.83$\times$10$^{-01}$ & 3.23$\times$10$^{-01}$ \\
0.016 & 4.34$\times$10$^{-64}$ & 6.93$\times$10$^{-64}$ &
1.13$\times$10$^{-63}$ & -1.454$\times$10$^{+02}$ &
4.83$\times$10$^{-01}$ & 3.10$\times$10$^{-01}$ \\
0.018 & 2.60$\times$10$^{-61}$ & 4.16$\times$10$^{-61}$ &
6.75$\times$10$^{-61}$ & -1.390$\times$10$^{+02}$ &
4.85$\times$10$^{-01}$ & 2.84$\times$10$^{-01}$ \\
0.020 & 4.27$\times$10$^{-59}$ & 6.83$\times$10$^{-59}$ &
1.12$\times$10$^{-58}$ & -1.339$\times$10$^{+02}$ &
4.87$\times$10$^{-01}$ & 2.55$\times$10$^{-01}$ \\
0.025 & 5.79$\times$10$^{-55}$ & 9.65$\times$10$^{-55}$ &
1.59$\times$10$^{-54}$ & -1.244$\times$10$^{+02}$ &
5.03$\times$10$^{-01}$ & 9.62$\times$10$^{-01}$ \\
0.030 & 8.22$\times$10$^{-50}$ & 3.94$\times$10$^{-49}$ &
1.40$\times$10$^{-48}$ & -1.116$\times$10$^{+02}$ &
1.37$\times$10$^{+00}$ & 2.50$\times$10$^{+01}$ \\
0.040 & 1.18$\times$10$^{-41}$ & 4.00$\times$10$^{-41}$ &
1.10$\times$10$^{-40}$ & -9.314$\times$10$^{+01}$ &
1.13$\times$10$^{+00}$ & 2.34$\times$10$^{+01}$ \\
0.050 & 9.31$\times$10$^{-37}$ & 2.76$\times$10$^{-36}$ &
6.55$\times$10$^{-36}$ & -8.200$\times$10$^{+01}$ &
1.01$\times$10$^{+00}$ & 3.40$\times$10$^{+01}$ \\
0.060 & 1.70$\times$10$^{-33}$ & 4.78$\times$10$^{-33}$ &
1.07$\times$10$^{-32}$ & -7.455$\times$10$^{+01}$ &
9.68$\times$10$^{-01}$ & 4.07$\times$10$^{+01}$ \\
0.070 & 3.65$\times$10$^{-31}$ & 9.84$\times$10$^{-31}$ &
2.19$\times$10$^{-30}$ & -6.921$\times$10$^{+01}$ &
9.44$\times$10$^{-01}$ & 3.71$\times$10$^{+01}$ \\
0.080 & 2.20$\times$10$^{-29}$ & 5.64$\times$10$^{-29}$ &
1.22$\times$10$^{-28}$ & -6.513$\times$10$^{+01}$ &
8.60$\times$10$^{-01}$ & 1.85$\times$10$^{+01}$ \\
0.090 & 8.06$\times$10$^{-28}$ & 1.72$\times$10$^{-27}$ &
3.24$\times$10$^{-27}$ & -6.169$\times$10$^{+01}$ &
7.24$\times$10$^{-01}$ & 2.00$\times$10$^{+01}$ \\
0.100 & 1.68$\times$10$^{-26}$ & 4.26$\times$10$^{-26}$ &
8.66$\times$10$^{-26}$ & -5.852$\times$10$^{+01}$ &
7.83$\times$10$^{-01}$ & 4.51$\times$10$^{+01}$ \\
0.110 & 2.73$\times$10$^{-25}$ & 8.38$\times$10$^{-25}$ &
2.12$\times$10$^{-24}$ & -5.553$\times$10$^{+01}$ &
9.06$\times$10$^{-01}$ & 5.08$\times$10$^{+01}$ \\
0.120 & 5.43$\times$10$^{-24}$ & 1.44$\times$10$^{-23}$ &
3.57$\times$10$^{-23}$ & -5.263$\times$10$^{+01}$ &
8.12$\times$10$^{-01}$ & 6.23$\times$10$^{+01}$ \\
0.130 & 1.11$\times$10$^{-22}$ & 2.11$\times$10$^{-22}$ &
4.36$\times$10$^{-22}$ & -4.989$\times$10$^{+01}$ &
5.97$\times$10$^{-01}$ & 4.85$\times$10$^{+01}$ \\
0.140 & 1.71$\times$10$^{-21}$ & 2.62$\times$10$^{-21}$ &
4.30$\times$10$^{-21}$ & -4.738$\times$10$^{+01}$ &
4.15$\times$10$^{-01}$ & 2.27$\times$10$^{+01}$ \\
0.150 & 1.90$\times$10$^{-20}$ & 2.63$\times$10$^{-20}$ &
3.62$\times$10$^{-20}$ & -4.509$\times$10$^{+01}$ &
3.09$\times$10$^{-01}$ & 6.46$\times$10$^{+00}$ \\
0.160 & 1.58$\times$10$^{-19}$ & 2.07$\times$10$^{-19}$ &
2.68$\times$10$^{-19}$ & -4.303$\times$10$^{+01}$ &
2.63$\times$10$^{-01}$ & 1.91$\times$10$^{+00}$ \\
0.180 & 5.56$\times$10$^{-18}$ & 7.02$\times$10$^{-18}$ &
8.96$\times$10$^{-18}$ & -3.950$\times$10$^{+01}$ &
2.41$\times$10$^{-01}$ & 1.76$\times$10$^{-01}$ \\
0.200 & 1.08$\times$10$^{-16}$ & 1.33$\times$10$^{-16}$ &
1.67$\times$10$^{-16}$ & -3.655$\times$10$^{+01}$ &
2.19$\times$10$^{-01}$ & 9.84$\times$10$^{-01}$ \\
0.250 & 5.68$\times$10$^{-14}$ & 6.42$\times$10$^{-14}$ &
7.25$\times$10$^{-14}$ & -3.038$\times$10$^{+01}$ &
1.22$\times$10$^{-01}$ & 3.76$\times$10$^{-01}$ \\
0.300 & 7.63$\times$10$^{-12}$ & 8.60$\times$10$^{-12}$ &
9.78$\times$10$^{-12}$ & -2.547$\times$10$^{+01}$ &
1.28$\times$10$^{-01}$ & 4.29$\times$10$^{+00}$ \\
0.350 & 2.85$\times$10$^{-10}$ & 3.23$\times$10$^{-10}$ &
3.71$\times$10$^{-10}$ & -2.185$\times$10$^{+01}$ &
1.36$\times$10$^{-01}$ & 6.78$\times$10$^{+00}$ \\
0.400 & 4.32$\times$10$^{-09}$ & 4.89$\times$10$^{-09}$ &
5.64$\times$10$^{-09}$ & -1.913$\times$10$^{+01}$ &
1.40$\times$10$^{-01}$ & 9.12$\times$10$^{+00}$ \\
0.450 & 3.52$\times$10$^{-08}$ & 3.99$\times$10$^{-08}$ &
4.61$\times$10$^{-08}$ & -1.703$\times$10$^{+01}$ &
1.42$\times$10$^{-01}$ & 1.13$\times$10$^{+01}$ \\
0.500 & 1.86$\times$10$^{-07}$ & 2.11$\times$10$^{-07}$ &
2.44$\times$10$^{-07}$ & -1.536$\times$10$^{+01}$ &
1.44$\times$10$^{-01}$ & 1.33$\times$10$^{+01}$ \\
0.600 & 2.19$\times$10$^{-06}$ & 2.48$\times$10$^{-06}$ &
2.89$\times$10$^{-06}$ & -1.289$\times$10$^{+01}$ &
1.46$\times$10$^{-01}$ & 1.68$\times$10$^{+01}$ \\
0.700 & 1.25$\times$10$^{-05}$ & 1.42$\times$10$^{-05}$ &
1.65$\times$10$^{-05}$ & -1.115$\times$10$^{+01}$ &
1.46$\times$10$^{-01}$ & 1.94$\times$10$^{+01}$ \\
0.800 & 4.70$\times$10$^{-05}$ & 5.31$\times$10$^{-05}$ &
6.15$\times$10$^{-05}$ & -9.829$\times$10$^{+00}$ &
1.44$\times$10$^{-01}$ & 1.95$\times$10$^{+01}$ \\
0.900 & 1.36$\times$10$^{-04}$ & 1.54$\times$10$^{-04}$ &
1.80$\times$10$^{-04}$ & -8.762$\times$10$^{+00}$ &
1.46$\times$10$^{-01}$ & 1.71$\times$10$^{+01}$ \\
1.000 & 3.36$\times$10$^{-04}$ & 3.85$\times$10$^{-04}$ &
4.56$\times$10$^{-04}$ & -7.846$\times$10$^{+00}$ &
1.60$\times$10$^{-01}$ & 1.89$\times$10$^{+01}$ \\
1.250 & 2.11$\times$10$^{-03}$ & 2.55$\times$10$^{-03}$ &
3.23$\times$10$^{-03}$ & -5.946$\times$10$^{+00}$ &
2.23$\times$10$^{-01}$ & 2.58$\times$10$^{+01}$ \\
\end{tabular}
\end{ruledtabular}
\label{tab:Ne22_alphagamma_rate}
\end{table*}
\begin{table*}
\caption{The $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate evaluated without the incorporation of the TAMU results. The low, median and high rates correspond to the 16\%, 50\% and 84\% values of the cumulative probability distribution. The $\mu$ and $\sigma$ parameters are resulting from fitting the distribution of rates at that temperature to a log-normal distribution. The Anderson-Darling statistic is a measure of how the data are well described by a log-normal distribution - see Refs. \cite{0067-0049-207-1-18,PhysRevC.85.065809} for details. All rates and the $\mu$ and $\sigma$ parameters are given in cm$^3$ mol$^{-1}$ s$^{-1}$. At temperatures above 1.33 GK, the rates should be taken from Hauser-Feshbach models - see Ref. \cite{PhysRevC.85.065809}.}
\begin{ruledtabular}
\begin{tabular}{c c c c c c c }
T [GK] & Low rate & Median rate & High rate & Log-normal $\mu$ & Log-normal $\sigma$ & Anderson-Darling statistic \\
\hline \\
0.010 & 3.38$\times$10$^{-79}$ & 5.55$\times$10$^{-79}$ &
9.05$\times$10$^{-79}$ & -1.802$\times$10$^{+02}$ &
4.96$\times$10$^{-01}$ & 1.42$\times$10$^{-01}$ \\
0.011 & 1.59$\times$10$^{-75}$ & 2.58$\times$10$^{-75}$ &
4.20$\times$10$^{-75}$ & -1.717$\times$10$^{+02}$ &
4.91$\times$10$^{-01}$ & 1.88$\times$10$^{-01}$ \\
0.012 & 1.82$\times$10$^{-72}$ & 2.94$\times$10$^{-72}$ &
4.79$\times$10$^{-72}$ & -1.647$\times$10$^{+02}$ &
4.86$\times$10$^{-01}$ & 2.32$\times$10$^{-01}$ \\
0.013 & 6.96$\times$10$^{-70}$ & 1.13$\times$10$^{-69}$ &
1.82$\times$10$^{-69}$ & -1.588$\times$10$^{+02}$ &
4.83$\times$10$^{-01}$ & 2.77$\times$10$^{-01}$ \\
0.014 & 1.13$\times$10$^{-67}$ & 1.83$\times$10$^{-67}$ &
2.95$\times$10$^{-67}$ & -1.537$\times$10$^{+02}$ &
4.81$\times$10$^{-01}$ & 3.15$\times$10$^{-01}$ \\
0.015 & 9.23$\times$10$^{-66}$ & 1.49$\times$10$^{-65}$ &
2.42$\times$10$^{-65}$ & -1.493$\times$10$^{+02}$ &
4.81$\times$10$^{-01}$ & 3.50$\times$10$^{-01}$ \\
0.016 & 4.30$\times$10$^{-64}$ & 6.99$\times$10$^{-64}$ &
1.14$\times$10$^{-63}$ & -1.454$\times$10$^{+02}$ &
4.82$\times$10$^{-01}$ & 3.65$\times$10$^{-01}$ \\
0.018 & 2.57$\times$10$^{-61}$ & 4.21$\times$10$^{-61}$ &
6.84$\times$10$^{-61}$ & -1.390$\times$10$^{+02}$ &
4.84$\times$10$^{-01}$ & 2.76$\times$10$^{-01}$ \\
0.020 & 4.22$\times$10$^{-59}$ & 6.93$\times$10$^{-59}$ &
1.13$\times$10$^{-58}$ & -1.339$\times$10$^{+02}$ &
4.87$\times$10$^{-01}$ & 2.27$\times$10$^{-01}$ \\
0.025 & 5.83$\times$10$^{-55}$ & 9.83$\times$10$^{-55}$ &
1.62$\times$10$^{-54}$ & -1.244$\times$10$^{+02}$ &
5.05$\times$10$^{-01}$ & 1.44$\times$10$^{+00}$ \\
0.030 & 8.41$\times$10$^{-50}$ & 4.14$\times$10$^{-49}$ &
1.48$\times$10$^{-48}$ & -1.116$\times$10$^{+02}$ &
1.41$\times$10$^{+00}$ & 3.25$\times$10$^{+01}$ \\
0.040 & 1.19$\times$10$^{-41}$ & 4.25$\times$10$^{-41}$ &
1.18$\times$10$^{-40}$ & -9.310$\times$10$^{+01}$ &
1.16$\times$10$^{+00}$ & 3.31$\times$10$^{+01}$ \\
0.050 & 9.66$\times$10$^{-37}$ & 2.96$\times$10$^{-36}$ &
7.00$\times$10$^{-36}$ & -8.195$\times$10$^{+01}$ &
1.03$\times$10$^{+00}$ & 4.45$\times$10$^{+01}$ \\
0.060 & 1.78$\times$10$^{-33}$ & 5.14$\times$10$^{-33}$ &
1.11$\times$10$^{-32}$ & -7.450$\times$10$^{+01}$ &
9.83$\times$10$^{-01}$ & 5.10$\times$10$^{+01}$ \\
0.070 & 3.77$\times$10$^{-31}$ & 1.06$\times$10$^{-30}$ &
2.21$\times$10$^{-30}$ & -6.916$\times$10$^{+01}$ &
9.54$\times$10$^{-01}$ & 4.60$\times$10$^{+01}$ \\
0.080 & 2.28$\times$10$^{-29}$ & 5.95$\times$10$^{-29}$ &
1.23$\times$10$^{-28}$ & -6.509$\times$10$^{+01}$ &
8.67$\times$10$^{-01}$ & 2.45$\times$10$^{+01}$ \\
0.090 & 8.31$\times$10$^{-28}$ & 1.76$\times$10$^{-27}$ &
3.28$\times$10$^{-27}$ & -6.167$\times$10$^{+01}$ &
7.28$\times$10$^{-01}$ & 2.19$\times$10$^{+01}$ \\
0.100 & 1.75$\times$10$^{-26}$ & 4.25$\times$10$^{-26}$ &
8.54$\times$10$^{-26}$ & -5.852$\times$10$^{+01}$ &
7.87$\times$10$^{-01}$ & 4.46$\times$10$^{+01}$ \\
0.110 & 2.42$\times$10$^{-25}$ & 7.97$\times$10$^{-25}$ &
2.02$\times$10$^{-24}$ & -5.561$\times$10$^{+01}$ &
9.86$\times$10$^{-01}$ & 5.22$\times$10$^{+01}$ \\
0.120 & 3.02$\times$10$^{-24}$ & 1.22$\times$10$^{-23}$ &
3.24$\times$10$^{-23}$ & -5.294$\times$10$^{+01}$ &
1.10$\times$10$^{+00}$ & 6.52$\times$10$^{+01}$ \\
0.130 & 3.78$\times$10$^{-23}$ & 1.47$\times$10$^{-22}$ &
3.61$\times$10$^{-22}$ & -5.048$\times$10$^{+01}$ &
1.09$\times$10$^{+00}$ & 7.52$\times$10$^{+01}$ \\
0.140 & 3.97$\times$10$^{-22}$ & 1.40$\times$10$^{-21}$ &
3.09$\times$10$^{-21}$ & -4.823$\times$10$^{+01}$ &
1.04$\times$10$^{+00}$ & 8.16$\times$10$^{+01}$ \\
0.150 & 3.47$\times$10$^{-21}$ & 1.09$\times$10$^{-20}$ &
2.19$\times$10$^{-20}$ & -4.616$\times$10$^{+01}$ &
9.84$\times$10$^{-01}$ & 6.05$\times$10$^{+01}$ \\
0.160 & 2.45$\times$10$^{-20}$ & 6.86$\times$10$^{-20}$ &
1.47$\times$10$^{-19}$ & -4.424$\times$10$^{+01}$ &
9.44$\times$10$^{-01}$ & 2.59$\times$10$^{+01}$ \\
0.180 & 8.92$\times$10$^{-19}$ & 1.99$\times$10$^{-18}$ &
4.96$\times$10$^{-18}$ & -4.073$\times$10$^{+01}$ &
8.47$\times$10$^{-01}$ & 1.92$\times$10$^{+00}$ \\
0.200 & 2.77$\times$10$^{-17}$ & 5.00$\times$10$^{-17}$ &
1.06$\times$10$^{-16}$ & -3.747$\times$10$^{+01}$ &
6.40$\times$10$^{-01}$ & 2.56$\times$10$^{+01}$ \\
0.250 & 4.52$\times$10$^{-14}$ & 5.44$\times$10$^{-14}$ &
6.84$\times$10$^{-14}$ & -3.052$\times$10$^{+01}$ &
2.13$\times$10$^{-01}$ & 1.87$\times$10$^{+01}$ \\
0.300 & 7.45$\times$10$^{-12}$ & 8.54$\times$10$^{-12}$ &
9.85$\times$10$^{-12}$ & -2.548$\times$10$^{+01}$ &
1.42$\times$10$^{-01}$ & 2.44$\times$10$^{+00}$ \\
0.350 & 2.86$\times$10$^{-10}$ & 3.25$\times$10$^{-10}$ &
3.73$\times$10$^{-10}$ & -2.184$\times$10$^{+01}$ &
1.39$\times$10$^{-01}$ & 5.49$\times$10$^{+00}$ \\
0.400 & 4.33$\times$10$^{-09}$ & 4.94$\times$10$^{-09}$ &
5.66$\times$10$^{-09}$ & -1.912$\times$10$^{+01}$ &
1.42$\times$10$^{-01}$ & 8.52$\times$10$^{+00}$ \\
0.450 & 3.52$\times$10$^{-08}$ & 4.02$\times$10$^{-08}$ &
4.63$\times$10$^{-08}$ & -1.702$\times$10$^{+01}$ &
1.44$\times$10$^{-01}$ & 1.10$\times$10$^{+01}$ \\
0.500 & 1.86$\times$10$^{-07}$ & 2.12$\times$10$^{-07}$ &
2.45$\times$10$^{-07}$ & -1.536$\times$10$^{+01}$ &
1.46$\times$10$^{-01}$ & 1.32$\times$10$^{+01}$ \\
0.600 & 2.19$\times$10$^{-06}$ & 2.50$\times$10$^{-06}$ &
2.89$\times$10$^{-06}$ & -1.289$\times$10$^{+01}$ &
1.49$\times$10$^{-01}$ & 1.69$\times$10$^{+01}$ \\
0.700 & 1.25$\times$10$^{-05}$ & 1.43$\times$10$^{-05}$ &
1.66$\times$10$^{-05}$ & -1.115$\times$10$^{+01}$ &
1.49$\times$10$^{-01}$ & 1.99$\times$10$^{+01}$ \\
0.800 & 4.71$\times$10$^{-05}$ & 5.33$\times$10$^{-05}$ &
6.19$\times$10$^{-05}$ & -9.826$\times$10$^{+00}$ &
1.47$\times$10$^{-01}$ & 2.06$\times$10$^{+01}$ \\
0.900 & 1.36$\times$10$^{-04}$ & 1.55$\times$10$^{-04}$ &
1.81$\times$10$^{-04}$ & -8.758$\times$10$^{+00}$ &
1.50$\times$10$^{-01}$ & 1.86$\times$10$^{+01}$ \\
1.000 & 3.36$\times$10$^{-04}$ & 3.87$\times$10$^{-04}$ &
4.58$\times$10$^{-04}$ & -7.843$\times$10$^{+00}$ &
1.65$\times$10$^{-01}$ & 1.83$\times$10$^{+01}$ \\
1.250 & 2.10$\times$10$^{-03}$ & 2.56$\times$10$^{-03}$ &
3.25$\times$10$^{-03}$ & -5.944$\times$10$^{+00}$ &
2.29$\times$10$^{-01}$ & 2.58$\times$10$^{+01}$ \\
\end{tabular}
\end{ruledtabular}
\label{tab:Ne22_alphagamma_rate_no_TAMU}
\end{table*}
\begin{table*}
\caption{Recommended $^{22}$Ne($\alpha,n$)$^{25}$Mg rate from the present evaluation incorporating the TAMU results. For a description of the table content, refer to the caption of Table \ref{tab:Ne22_alphagamma_rate}. At temperatures below 0.03 GK the rate is $<10^{-99}$ and is omitted from the table. At temperatures above 1.33 GK, the rates should be taken from Hauser-Feshbach models - see Ref. \cite{PhysRevC.85.065809}.}
\begin{ruledtabular}
\begin{tabular}{c c c c c c c }
T [GK] & Low rate & Median rate & High rate & Log-normal $\mu$ & Log-normal $\sigma$ & Anderson-Darling statistic \\
\hline \\
0.030 & 6.98$\times$10$^{-88}$ & 4.57$\times$10$^{-87}$ &
1.82$\times$10$^{-86}$ & -1.990$\times$10$^{+02}$ &
1.60$\times$10$^{+00}$ & 4.18$\times$10$^{+01}$ \\
0.040 & 1.67$\times$10$^{-67}$ & 1.38$\times$10$^{-66}$ &
5.65$\times$10$^{-66}$ & -1.519$\times$10$^{+02}$ &
1.72$\times$10$^{+00}$ & 5.50$\times$10$^{+01}$ \\
0.050 & 3.11$\times$10$^{-55}$ & 2.89$\times$10$^{-54}$ &
1.19$\times$10$^{-53}$ & -1.236$\times$10$^{+02}$ &
1.77$\times$10$^{+00}$ & 6.28$\times$10$^{+01}$ \\
0.060 & 4.86$\times$10$^{-47}$ & 4.69$\times$10$^{-46}$ &
1.94$\times$10$^{-45}$ & -1.047$\times$10$^{+02}$ &
1.79$\times$10$^{+00}$ & 6.49$\times$10$^{+01}$ \\
0.070 & 3.73$\times$10$^{-41}$ & 3.38$\times$10$^{-40}$ &
1.39$\times$10$^{-39}$ & -9.121$\times$10$^{+01}$ &
1.76$\times$10$^{+00}$ & 6.13$\times$10$^{+01}$ \\
0.080 & 1.04$\times$10$^{-36}$ & 8.27$\times$10$^{-36}$ &
3.38$\times$10$^{-35}$ & -8.107$\times$10$^{+01}$ &
1.70$\times$10$^{+00}$ & 5.31$\times$10$^{+01}$ \\
0.090 & 3.13$\times$10$^{-33}$ & 2.13$\times$10$^{-32}$ &
8.56$\times$10$^{-32}$ & -7.317$\times$10$^{+01}$ &
1.62$\times$10$^{+00}$ & 4.35$\times$10$^{+01}$ \\
0.100 & 1.96$\times$10$^{-30}$ & 1.16$\times$10$^{-29}$ &
4.50$\times$10$^{-29}$ & -6.683$\times$10$^{+01}$ &
1.54$\times$10$^{+00}$ & 3.52$\times$10$^{+01}$ \\
0.110 & 3.97$\times$10$^{-28}$ & 2.01$\times$10$^{-27}$ &
7.48$\times$10$^{-27}$ & -6.164$\times$10$^{+01}$ &
1.45$\times$10$^{+00}$ & 2.92$\times$10$^{+01}$ \\
0.120 & 3.35$\times$10$^{-26}$ & 1.51$\times$10$^{-25}$ &
5.32$\times$10$^{-25}$ & -5.730$\times$10$^{+01}$ &
1.37$\times$10$^{+00}$ & 2.56$\times$10$^{+01}$ \\
0.130 & 1.43$\times$10$^{-24}$ & 5.97$\times$10$^{-24}$ &
1.95$\times$10$^{-23}$ & -5.361$\times$10$^{+01}$ &
1.30$\times$10$^{+00}$ & 2.37$\times$10$^{+01}$ \\
0.140 & 3.65$\times$10$^{-23}$ & 1.42$\times$10$^{-22}$ &
4.24$\times$10$^{-22}$ & -5.044$\times$10$^{+01}$ &
1.22$\times$10$^{+00}$ & 2.19$\times$10$^{+01}$ \\
0.150 & 6.24$\times$10$^{-22}$ & 2.24$\times$10$^{-21}$ &
6.19$\times$10$^{-21}$ & -4.767$\times$10$^{+01}$ &
1.13$\times$10$^{+00}$ & 1.91$\times$10$^{+01}$ \\
0.160 & 7.87$\times$10$^{-21}$ & 2.54$\times$10$^{-20}$ &
6.58$\times$10$^{-20}$ & -4.521$\times$10$^{+01}$ &
1.02$\times$10$^{+00}$ & 1.51$\times$10$^{+01}$ \\
0.180 & 6.91$\times$10$^{-19}$ & 1.63$\times$10$^{-18}$ &
3.61$\times$10$^{-18}$ & -4.098$\times$10$^{+01}$ &
7.82$\times$10$^{-01}$ & 8.65$\times$10$^{+00}$ \\
0.200 & 3.27$\times$10$^{-17}$ & 5.61$\times$10$^{-17}$ &
1.03$\times$10$^{-16}$ & -3.740$\times$10$^{+01}$ &
5.43$\times$10$^{-01}$ & 9.32$\times$10$^{+00}$ \\
0.250 & 5.27$\times$10$^{-14}$ & 6.80$\times$10$^{-14}$ &
8.70$\times$10$^{-14}$ & -3.032$\times$10$^{+01}$ &
2.52$\times$10$^{-01}$ & 2.42$\times$10$^{-01}$ \\
0.300 & 8.33$\times$10$^{-12}$ & 1.05$\times$10$^{-11}$ &
1.30$\times$10$^{-11}$ & -2.528$\times$10$^{+01}$ &
2.26$\times$10$^{-01}$ & 4.96$\times$10$^{-01}$ \\
0.350 & 3.31$\times$10$^{-10}$ & 4.12$\times$10$^{-10}$ &
5.09$\times$10$^{-10}$ & -2.161$\times$10$^{+01}$ &
2.18$\times$10$^{-01}$ & 4.74$\times$10$^{-01}$ \\
0.400 & 5.60$\times$10$^{-09}$ & 6.83$\times$10$^{-09}$ &
8.28$\times$10$^{-09}$ & -1.880$\times$10$^{+01}$ &
2.00$\times$10$^{-01}$ & 7.07$\times$10$^{-01}$ \\
0.450 & 5.63$\times$10$^{-08}$ & 6.64$\times$10$^{-08}$ &
7.86$\times$10$^{-08}$ & -1.652$\times$10$^{+01}$ &
1.70$\times$10$^{-01}$ & 1.48$\times$10$^{+00}$ \\
0.500 & 4.21$\times$10$^{-07}$ & 4.81$\times$10$^{-07}$ &
5.52$\times$10$^{-07}$ & -1.454$\times$10$^{+01}$ &
1.37$\times$10$^{-01}$ & 1.76$\times$10$^{+00}$ \\
0.600 & 1.39$\times$10$^{-05}$ & 1.59$\times$10$^{-05}$ &
1.87$\times$10$^{-05}$ & -1.104$\times$10$^{+01}$ &
1.50$\times$10$^{-01}$ & 9.31$\times$10$^{+00}$ \\
0.700 & 2.60$\times$10$^{-04}$ & 3.09$\times$10$^{-04}$ &
3.76$\times$10$^{-04}$ & -8.071$\times$10$^{+00}$ &
1.86$\times$10$^{-01}$ & 9.47$\times$10$^{+00}$ \\
0.800 & 2.92$\times$10$^{-03}$ & 3.48$\times$10$^{-03}$ &
4.25$\times$10$^{-03}$ & -5.649$\times$10$^{+00}$ &
1.86$\times$10$^{-01}$ & 9.15$\times$10$^{+00}$ \\
0.900 & 2.16$\times$10$^{-02}$ & 2.54$\times$10$^{-02}$ &
3.05$\times$10$^{-02}$ & -3.664$\times$10$^{+00}$ &
1.70$\times$10$^{-01}$ & 9.47$\times$10$^{+00}$ \\
1.000 & 1.15$\times$10$^{-01}$ & 1.33$\times$10$^{-01}$ &
1.57$\times$10$^{-01}$ & -2.008$\times$10$^{+00}$ &
1.50$\times$10$^{-01}$ & 7.96$\times$10$^{+00}$ \\
1.250 & 2.72$\times$10$^{+00}$ & 3.06$\times$10$^{+00}$ &
3.46$\times$10$^{+00}$ & 1.122$\times$10$^{+00}$ &
1.20$\times$10$^{-01}$ & 1.19$\times$10$^{+00}$ \\
\end{tabular}
\end{ruledtabular}
\label{tab:Ne22_alphan_rate}
\end{table*}
\begin{table*}
\caption{The $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rate evaluated without the incorporation of the TAMU results. For a description of the table contents, refer to the caption of Table \ref{tab:Ne22_alphagamma_rate}. At temperatures below 0.03 GK the rate is $<10^{-99}$ and is omitted from the table. At temperatures above 1.33 GK, the rates should be taken from Hauser-Feshbach models - see Ref. \cite{PhysRevC.85.065809}.}
\begin{ruledtabular}
\begin{tabular}{c c c c c c c }
T [GK] & Low rate & Median rate & High rate & Log-normal $\mu$ & Log-normal $\sigma$ & Anderson-Darling statistic \\
\hline \\
0.030 & 7.71$\times$10$^{-88}$ & 4.87$\times$10$^{-87}$ &
2.01$\times$10$^{-86}$ & -1.989$\times$10$^{+02}$ &
1.62$\times$10$^{+00}$ & 2.84$\times$10$^{+01}$ \\
0.040 & 1.90$\times$10$^{-67}$ & 1.47$\times$10$^{-66}$ &
6.28$\times$10$^{-66}$ & -1.518$\times$10$^{+02}$ &
1.74$\times$10$^{+00}$ & 4.00$\times$10$^{+01}$ \\
0.050 & 3.62$\times$10$^{-55}$ & 3.05$\times$10$^{-54}$ &
1.32$\times$10$^{-53}$ & -1.235$\times$10$^{+02}$ &
1.79$\times$10$^{+00}$ & 4.71$\times$10$^{+01}$ \\
0.060 & 5.76$\times$10$^{-47}$ & 4.96$\times$10$^{-46}$ &
2.15$\times$10$^{-45}$ & -1.046$\times$10$^{+02}$ &
1.81$\times$10$^{+00}$ & 4.88$\times$10$^{+01}$ \\
0.070 & 4.29$\times$10$^{-41}$ & 3.58$\times$10$^{-40}$ &
1.55$\times$10$^{-39}$ & -9.110$\times$10$^{+01}$ &
1.77$\times$10$^{+00}$ & 4.44$\times$10$^{+01}$ \\
0.080 & 1.16$\times$10$^{-36}$ & 8.80$\times$10$^{-36}$ &
3.76$\times$10$^{-35}$ & -8.095$\times$10$^{+01}$ &
1.69$\times$10$^{+00}$ & 3.56$\times$10$^{+01}$ \\
0.090 & 3.54$\times$10$^{-33}$ & 2.29$\times$10$^{-32}$ &
9.52$\times$10$^{-32}$ & -7.304$\times$10$^{+01}$ &
1.60$\times$10$^{+00}$ & 2.59$\times$10$^{+01}$ \\
0.100 & 2.34$\times$10$^{-30}$ & 1.25$\times$10$^{-29}$ &
4.98$\times$10$^{-29}$ & -6.669$\times$10$^{+01}$ &
1.50$\times$10$^{+00}$ & 1.79$\times$10$^{+01}$ \\
0.110 & 4.78$\times$10$^{-28}$ & 2.21$\times$10$^{-27}$ &
8.33$\times$10$^{-27}$ & -6.148$\times$10$^{+01}$ &
1.40$\times$10$^{+00}$ & 1.23$\times$10$^{+01}$ \\
0.120 & 4.22$\times$10$^{-26}$ & 1.68$\times$10$^{-25}$ &
5.92$\times$10$^{-25}$ & -5.713$\times$10$^{+01}$ &
1.31$\times$10$^{+00}$ & 8.87$\times$10$^{+00}$ \\
0.130 & 1.89$\times$10$^{-24}$ & 6.73$\times$10$^{-24}$ &
2.21$\times$10$^{-23}$ & -5.342$\times$10$^{+01}$ &
1.22$\times$10$^{+00}$ & 6.64$\times$10$^{+00}$ \\
0.140 & 5.06$\times$10$^{-23}$ & 1.62$\times$10$^{-22}$ &
4.89$\times$10$^{-22}$ & -5.022$\times$10$^{+01}$ &
1.12$\times$10$^{+00}$ & 4.77$\times$10$^{+00}$ \\
0.150 & 9.10$\times$10$^{-22}$ & 2.62$\times$10$^{-21}$ &
7.27$\times$10$^{-21}$ & -4.741$\times$10$^{+01}$ &
1.01$\times$10$^{+00}$ & 3.09$\times$10$^{+00}$ \\
0.160 & 1.25$\times$10$^{-20}$ & 3.13$\times$10$^{-20}$ &
7.94$\times$10$^{-20}$ & -4.490$\times$10$^{+01}$ &
8.88$\times$10$^{-01}$ & 3.23$\times$10$^{+00}$ \\
0.180 & 1.37$\times$10$^{-18}$ & 2.42$\times$10$^{-18}$ &
4.83$\times$10$^{-18}$ & -4.051$\times$10$^{+01}$ &
6.09$\times$10$^{-01}$ & 1.47$\times$10$^{+01}$ \\
0.200 & 7.79$\times$10$^{-17}$ & 1.08$\times$10$^{-16}$ &
1.65$\times$10$^{-16}$ & -3.672$\times$10$^{+01}$ &
3.80$\times$10$^{-01}$ & 1.81$\times$10$^{+01}$ \\
0.250 & 1.49$\times$10$^{-13}$ & 1.81$\times$10$^{-13}$ &
2.24$\times$10$^{-13}$ & -2.934$\times$10$^{+01}$ &
2.08$\times$10$^{-01}$ & 3.87$\times$10$^{-01}$ \\
0.300 & 2.48$\times$10$^{-11}$ & 3.03$\times$10$^{-11}$ &
3.72$\times$10$^{-11}$ & -2.422$\times$10$^{+01}$ &
2.04$\times$10$^{-01}$ & 7.57$\times$10$^{-01}$ \\
0.350 & 9.76$\times$10$^{-10}$ & 1.19$\times$10$^{-09}$ &
1.46$\times$10$^{-09}$ & -2.055$\times$10$^{+01}$ &
2.02$\times$10$^{-01}$ & 7.98$\times$10$^{-01}$ \\
0.400 & 1.54$\times$10$^{-08}$ & 1.87$\times$10$^{-08}$ &
2.27$\times$10$^{-08}$ & -1.780$\times$10$^{+01}$ &
1.96$\times$10$^{-01}$ & 9.26$\times$10$^{-01}$ \\
0.450 & 1.37$\times$10$^{-07}$ & 1.63$\times$10$^{-07}$ &
1.96$\times$10$^{-07}$ & -1.563$\times$10$^{+01}$ &
1.83$\times$10$^{-01}$ & 1.27$\times$10$^{+00}$ \\
0.500 & 8.52$\times$10$^{-07}$ & 9.89$\times$10$^{-07}$ &
1.17$\times$10$^{-06}$ & -1.382$\times$10$^{+01}$ &
1.61$\times$10$^{-01}$ & 1.77$\times$10$^{+00}$ \\
0.600 & 1.94$\times$10$^{-05}$ & 2.21$\times$10$^{-05}$ &
2.53$\times$10$^{-05}$ & -1.072$\times$10$^{+01}$ &
1.35$\times$10$^{-01}$ & 2.20$\times$10$^{+00}$ \\
0.700 & 2.94$\times$10$^{-04}$ & 3.43$\times$10$^{-04}$ &
4.11$\times$10$^{-04}$ & -7.965$\times$10$^{+00}$ &
1.71$\times$10$^{-01}$ & 8.99$\times$10$^{+00}$ \\
0.800 & 3.04$\times$10$^{-03}$ & 3.59$\times$10$^{-03}$ &
4.37$\times$10$^{-03}$ & -5.614$\times$10$^{+00}$ &
1.82$\times$10$^{-01}$ & 9.35$\times$10$^{+00}$ \\
0.900 & 2.19$\times$10$^{-02}$ & 2.56$\times$10$^{-02}$ &
3.07$\times$10$^{-02}$ & -3.651$\times$10$^{+00}$ &
1.69$\times$10$^{-01}$ & 9.09$\times$10$^{+00}$ \\
1.000 & 1.16$\times$10$^{-01}$ & 1.34$\times$10$^{-01}$ &
1.57$\times$10$^{-01}$ & -2.003$\times$10$^{+00}$ &
1.51$\times$10$^{-01}$ & 7.67$\times$10$^{+00}$ \\
1.250 & 2.72$\times$10$^{+00}$ & 3.05$\times$10$^{+00}$ &
3.45$\times$10$^{+00}$ & 1.121$\times$10$^{+00}$ &
1.20$\times$10$^{-01}$ & 1.84$\times$10$^{+00}$ \\
\end{tabular}
\end{ruledtabular}
\label{tab:Ne22_alphan_rate_no_TAMU}
\end{table*}
Figure \ref{fig:FractionalRateContributions} shows the fractional contributions of individual resonances to the recommended reaction rates for the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions. Above 0.2 GK, both reaction rates are dominated by the $E_r = 706$-keV resonance and other higher-lying directly measured resonances. In this region, we expect little deviations from the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg rate of Longland {\it et al.} \cite{PhysRevC.85.065809} as both evaluations are based on the similar nuclear data, with only the updated (and consistent) measurement of Hunt {\it et al.} \cite{PhysRevC.99.045804}. The updated, reduced, $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance strength for this state is expected to cause a concomitant reduction in the reaction rate above 0.2 GK.
\begin{figure}[htbp]
\centering
\subfigure{\includegraphics[angle=0,width=0.95\columnwidth]{FractionalContributions_alphagamma.pdf}}
\subfigure{\includegraphics[angle=0,width=0.95\columnwidth]{FractionalContributions_alphaneutron.pdf}}
\caption{Fractional contributions of selected resonances to the (top) $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and (bottom) $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rates. These fractional contributions are for the recommended reaction rates, which incorporate the Texas A\&M results. The shaded region gives the 68\% coverage limit for the contribution of each resonance. Note that only the most significant resonances are included in the figure; the sum of the contributions may not reach 100\% due to contributions from omitted resonances.}
\label{fig:FractionalRateContributions}
\end{figure}
Below 0.2 GK, lower-energy resonances in $^{26}$Mg are predicted to dominate the reaction rate. Some of these resonances have been observed in the $^{25}$Mg$+n$ reactions of Massimi {\it et al.} \cite{Massimi20171,PhysRevC.85.044615} and neutron and $\gamma$-ray partial widths have been determined. No $\alpha$-particle partial widths have been directly measured for the resonances below $E_{cm} = 706$ keV but an estimate of the partial width for the $E_r = 469$-keV resonance is available from the transfer data of Jayatissa {\it et al.} \cite{JAYATISSA2020135267}. Evaluations of the reaction rates are based on the potential upper limits for the other resonances resulting in larger uncertainties for the rates.
The states below the $E_r = 706$-keV resonance which have been identified as potentially strongly contributing to the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate are those at $E_r = 191$, $203$, $211$, $278$, $334$, and $469$ keV ($E_x = 10.806$, $10.818$, $10.826$, $10.893$, $10.949$, and $11.084$ MeV. The states at around $E_x = 11.17$ MeV are predicted to have negligible impact due to the weak cross section observed in the sub-barrier transfer reaction of Jayatissa {\it et al.} \cite{JAYATISSA2020135267}.
The states below the $E_r = 706$-keV resonance which have been identified as potentially strongly contributing to the $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rate are those at $E_r = 497$ and $548$ keV ($E_x = 11.112$ and $11.163$ MeV).
\section{Comparison to previous reaction-rate evaluations}
\label{RateComparisons}
We now compare the presently computed reaction rates with those of Longland {\it et al.} \cite{PhysRevC.85.065809}, Talwar {\it et al.} \cite{PhysRevC.93.055803} and Massimi {\it et al.} \cite{Massimi20171}. The Longland \cite{PhysRevC.85.065809} and Talwar \cite{PhysRevC.93.055803} evaluations both use the \textsc{RatesMC} code whereas the Massimi evaluation was differently performed and reported upper limits rather than median reaction rates.
The ratios of the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rates calculated in the present study and the reaction rates from Longland {\it et al.} \cite{PhysRevC.85.065809} and Talwar {\it et al.} \cite{PhysRevC.93.055803} are shown in Figure \ref{fig:RatioToHistorics}.
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[angle=0, width=0.95\columnwidth]{Ne22agRateRatios.pdf}}
\subfigure{
\includegraphics[angle=0, width=0.95\columnwidth]{Ne22anRateRatios.pdf}}
\caption{Ratio of the (top) $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and (bottom) $^{22}$Ne($\alpha,n$)$^{25}$Mg recommended reaction rates of the present calculation to the rates from (black, circles, solid line) Longland {\it et al.} \cite{PhysRevC.85.065809}, (red, triangles, broken line) Talwar {\it et al.} \cite{PhysRevC.93.055803} and (blue, crosses, broken line) the calculations performed for this measurement without using the Texas A\&M data. The solid points are the calculated ratios of median reaction rates; lines are drawn to guide the eye.}
\label{fig:RatioToHistorics}
\end{figure}
Above 0.2 GK, the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rates are dominated by resonances for which direct measurements exist, in particular by the $706$-keV resonance discussed in Section \ref{statesec12}. The consistency in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate (upper panel of Figure \ref{fig:RatioToHistorics}) above 0.2 GK is due to the fact that the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg resonance strength for the $E_r = 706$-keV resonance is unchanged between the different evaluations. There is a reduction in the recommended $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rate relative to the Longland and Talwar rates due to the results from the two transfer reactions \cite{OTA2020135256,JAYATISSA2020135267} carried out at Texas A\&M.. A small reduction in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate is observed due to the conclusion that the resonances observed in $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions are the same and the subsequent shift in the resonance energy. This reduction is within the uncertainty on the reaction rate computed by Longland {\it et al.} \cite{PhysRevC.85.065809}.
There is a small increase in the recommended $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate relative to the Longland rate below 0.1 GK. This is due to the additional contributions from the $J^\pi = 0^+$ state at $E_x = 10.818$ MeV from the $^{26}$Mg($\alpha,\alpha^\prime$)$^{26}$Mg data of Adsley {\it et al.} \cite{PhysRevC.96.055802}.
The increase in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate between 0.1 and 0.2 GK seen when comparing the recommended rate to the rate computed without the Texas A\& M results and the Longland evaluation is due to there now being an estimated resonance strength rather than an upper limit for the the $J^\pi = 2^+$ state at $E_x = 11.084$ MeV ($E_r = 469$ keV) \cite{OTA2020135256,JAYATISSA2020135267}.
The present $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate is much smaller than that of Talwar {\it et al.} \cite{PhysRevC.93.055803} between 0.1 and 0.4 GK due to the increased contribution from the $E_x = 11.171$-MeV state in the Talwar evaluation. As discussed in Section \ref{sec:11167keVState}, there is some uncertainty about the existence and properties of the states around $E_x = 11.17$ MeV; for this reason it has not been included in the present evaluation.
At the lowest temperatures, the significant increase in the reaction rate with respect to that of Talwar {\it et al.} \cite{PhysRevC.93.055803} is due to the treatment of the low-lying resonances at $E_r = 191$ and $E_r = 211$ keV for which the limits on the spectroscopic factors have been relaxed. However, these resonances are at low energy and are unlikely to have any astrophysical impact.
At temperatures above 0.2 GK, Massimi {\it et al.} \cite{Massimi20171} calculated an upper limit for the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate which is much smaller than the rate of Longland {\it et al.} \cite{PhysRevC.85.065809} and the current rate. This is because the methodology used to calculate the reaction rates is based on the ratio of the rates which are inferred from the ratios of the neutron and $\gamma$-ray partial widths for the observed resonances. As a consequence, the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate is underestimated by Massimi {\it et al.} \cite{Massimi20171}.
\subsection{Reaction-rate ratio}
The neutron yield depends on the $^{22}$Ne abundance and on the competition between the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions at various temperatures. Therefore, it is necessary to evaluate both the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and the $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rates, and their ratio.
The ratio of the two rates as a function of temperature is shown in Figure \ref{fig:RRRs}. At lower temperatures the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate is expected to dominate as the $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction is endothermic. As the temperature increases, the $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rate becomes stronger relative to the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg until it eventually exceeds the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate at around 0.2 GK. The dominant states in the temperature region around 0.2 GK for $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction are those at $E_x = 10.949$ and $11.084$ MeV, while the dominant states for the $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction are those at $E_x = 11.112$ and $11.163$ MeV. It is likely that the strengths of these resonances determine at which temperature the $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction becomes stronger than the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg one.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\columnwidth]{ReactionRateRatioContour.pdf}
\caption{Ratio of $^{22}$Ne($\alpha,n$)$^{25}$Mg and $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rates for the recommended rate (black solid line) and the rate computed without the Texas A\& M results (red solid line). Values above unity imply that neutron production is more likely than radiative capture. The broken (dotted) lines are the 68\% (95\%) confidence limits for the ratio of the recommended rates, the colours represent the same evaluations.}
\label{fig:RRRs}
\end{figure}
The Monte-Carlo calculations performed in this evaluation and in the evaluations of e.g. Refs. \cite{PhysRevC.85.065809,PhysRevC.93.055803} are independent calculations of the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rates. However, the properties of resonances in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions are correlated which can lead to overestimation of the uncertainties in the ratio of the reaction rates. This is not a problem for the reaction rate calculated without the inclusion of the Texas A\&M results as all of these measurements are independent.
Combining the $\Gamma_n/\Gamma_\gamma$ ratio determined from the measurement of Ota {\it et al.} \cite{OTA2020135256} with the weighted average of the the direct $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg measurements means that the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg resonance strengths are no longer independent.
However, in the present case the uncertainty is dominated by the uncertainty in the ratio of $\Gamma_n$ and $\Gamma_\gamma$ (23\%) not the uncertainty in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg resonance strength (11\%). This means that the overestimation in the uncertainty from the correlations arising from the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg resonance strength are a small contribution to the total uncertainty and, therefore, that the uncertainty in the ratio is not significantly overestimated.
Bearing in mind the above arguments, the 68\% and 95\% confidence limits on the reaction-rate ratio have been computed from the Monte-Carlo samples for each rate at each temperature. The probability distribution function of the ratio of the reaction rates has been constructed from these samples and the confidence limits extracted.
\section{Priorities for future measurements\label{sec:SuggestedFutureExperiments}}
As discussed in the previous section, the comparison to previous reaction-rate estimations shows sizeable uncertainties at low temperatures and significant disagreement with the rates of Talwar {\it et al.} \cite{PhysRevC.93.055803}. The fair agreement between our results and those of Longland {\it et al.} \cite{PhysRevC.85.065809} for the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate may also be attributable to the common method (i.e. the \textsc{RatesMC} Monte-Carlo code \cite{0067-0049-207-1-18}) adopted for the evaluation of the $\alpha+^{22}$Ne reaction rates. Although a full R-Matrix evaluation of the reaction rate would be preferable, the lack of experimental information on several $^{26}$Mg states as well as on the interference patterns between distant levels prevented this kind of approach.
As a consequence, future measurements should aim at determining the ($\alpha$,$n$), ($\alpha,\gamma$) and elastic scattering cross sections as well as characterising $^{26}$Mg levels in terms of spin and parity, eventually leading to an estimation of the reaction rate based on a full R-matrix analysis of available experimental data, including the effect of interference also due to sub-threshold resonances. Hereafter we report a high-priority request list based on the impact on the reaction rate and astrophysical implications.
\subsubsection{$^{22}$Ne($\alpha,\gamma$)$^{26}$Mg}
The $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate at temperature of interest to the $s$-process is linked to a few resonances. These levels could be studied in inverse kinematics using a recoil separator, as well as with direct measurements. In particular:
\begin{itemize}
\item The resonance at $E_x=11.321$~MeV ($E_r=706$~keV) is crucial at temperatures relevant to massive stars (see Figure~\ref{fig:FractionalRateContributions}). Two consistent direct measurements of the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg resonance strength are reported in the literature~\cite{Wolke1989,PhysRevC.99.045804}. Further direct measurements of this resonance with high beam intensities and improved resonance-energy resolution are required in order to determine any possible interference effects. The blister-proof neon-implanted targets of Hunt {\it et al.} \cite{PhysRevC.99.045804} would appear to be a productive approach.
\item The tentative $E_r = 557$-keV resonance reported by Talwar {\it et al.} \cite{PhysRevC.93.055803} is a source of great uncertainty in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate, increasing it by up to a factor of 100 relative to our evaluation. Based on the estimate of Talwar {\it et al.} \cite{PhysRevC.93.055803}, should this resonance exist (which we consider an open question) then it should be within reach of direct measurements. Confirmation or refutation of this state with additional measurements of the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg and $^{22}$Ne($^7$Li,$t$)$^{26}$Mg reactions would help to reduce the uncertainties in the rate considerably. The results of the Texas A\&M experiments appear to suggest that this state does not exist but higher-resolution measurements would be preferred in order to render a definitive verdict.
\item An associated important piece of experimental data would be the confirmation of the $E_x = 11.44$-MeV state observed by Giesen {\it et al.} \cite{GIESEN199395} in $^{22}$Ne($^6$Li,$d$)$^{26}$Mg and Jaeger {\it et al.} \cite{PhysRevLett.87.202501} in $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions. This state falls at the edge of the focal-plane acceptance of the experiment of Talwar {\it et al.} \cite{PhysRevC.93.055803}, being around 140 keV above the $E_x = 11.31$-MeV state.
\item A re-examination of the focal-plane spectra of Refs. \cite{PhysRevC.93.055803,TalwarThesis} and re-measurements of the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg reaction or a measurement of the $^{22}$Ne($^7$Li,$t$)$^{26}$Mg reaction may help to clarify which and how many $\alpha$-cluster states are populated in this region. As explained in Section \ref{sec:SynthesisOfTheData}, both Giesen {\it et al.} \cite{GIESEN199395} and Talwar {\it et al.} \cite{PhysRevC.93.055803} observed two $\alpha$-particle cluster states separated by $\sim150$ keV using the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg reaction. This may be coincidental but one possible cause is that the energy calibration in one of the experiments is shifted. A valuable step which may take place without requiring any further experiments is the re-analysis of the data of Talwar {\it et al.} \cite{PhysRevC.93.055803}, confirming that the two states observed in Talwar {\it et al.} \cite{PhysRevC.93.055803} at $E_x = 11.17$ and $11.31$ MeV are not, in fact, the $E_x = 11.31$- and $11.44$-MeV levels observed by Giesen {\it et al.} \cite{GIESEN199395} and Jaeger {\it et al.} \cite{PhysRevLett.87.202501}. \color{black} Re-analyses of the data of Giesen {\it et al.} \cite{GIESEN199395} and Jaeger {\it et al.} \cite{PhysRevLett.87.202501} would also be valuable but those data may no longer be available. This reconsideration of the results of the transfer reactions is supported by the non-observation of the $E_x = 11.17$-MeV state of Talwar {\it et al.} in the results from the Texas A\&M measurements of Jayatissa {\it et al.} \cite{JAYATISSA2020135267}. We note that the low-energy transfer reaction of Jayatissa {\it et al.} \cite{JAYATISSA2020135267} cannot populate high-spin states which may support the argument of Ota {\it et al.} \cite{OTA2020135256} that this state has high spin, as it was observed in that experiment. Additionally, the relative population of states of different spins in the various transfer measurements (Refs. \cite{GIESEN199395,PhysRevC.93.055803,JAYATISSA2020135267,OTA2020135256}) could also be considered as this would provide a practical method of testing the high-spin hypothesis of Ota {\it et al.} \cite{OTA2020135256}.
\item Resonances at $E_x=10.949$ MeV ($E_r=334$~keV) and $E_x=11.084$ MeV ($E_r=469$~keV) dominate the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate at temperature between 0.1 and 0.2~GK. No direct measurement of the $E_x = 10.949$-MeV state are available yet though some progress has been made \cite{10.1007/978-3-030-13876-9_50} but improved estimates with transfer reactions would also help to reduce uncertainties.
\end{itemize}
\subsubsection{$^{22}$Ne($\alpha$,$n$)$^{25}$Mg}
For an accurate determination of the $^{22}$Ne($\alpha$,$n$)$^{26}$Mg reaction rate and its uncertainty, several new experimental data are required. In particular:
\begin{itemize}
\item The partial widths or resonance strengths of the $E_x=11.112$ MeV ($E_r=497$ keV), $E_x=11.163$ MeV ($E_r=548$ keV), $E_x=11.321$ MeV ($E_r=706$ keV) and $E_x=11.911$ MeV ($E_r=1297$ keV) levels in $^{26}$Mg. Estimates of some of these quantities are available from the decay branching measurements of Ota {\it et al.} \cite{OTA2020135256} but higher-resolution data obtained with a less selective reaction may help in quantifying the branching of these states.
\item As for the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction, a firm spin and parity assignment of the potential cluster state observed in the vicinity of $E_x = 11.167$ MeV and a clear connection between this state and the states observed in the $n+^{25}$Mg experiments \cite{PhysRevC.85.044615,Massimi20171} are required. Confirmation of the $E_x = 11.44$-MeV state would also be useful in clarifying the properties of the levels in this region.
\item The interference pattern between distant levels, including sub-threshold states. The implanted targets of Hunt {\it et al.} \cite{PhysRevC.99.045804} provide an excellent opportunity to constrain the interference pattern.
\end{itemize}
\section{Impact for $s$-process calculations}\label{sec:impact}
\subsection{Main s-process from low-mass AGB star}
We re-computed the $s$-process nucleosynthesis for stellar models of 2 and 3M$_\odot$
adopting our new evaluations for ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg and ${^{22}}$Ne($\alpha$,$\gamma$)${^{26}}$Mg reaction rates using two different codes: the stellar models computed with the MESA code \cite{paxton:11} presented by Battino {\it et al.} \cite{battino:19} and the NEWTON code \cite{PALMERINI201821} presented by Trippella {\it et al.} \cite{trippella}.
In both cases, we include the results obtained adopting the reaction rates of Longland {\it et al.} \cite{PhysRevC.85.065809} and the rates evaluated in the present work with and without the Texas A\&M results.
In Figures \ref{fig:rb_cstars} and \ref{fig:Rb_cstars_sara} we compare our calculated Rb abundances as a function of the total $s$-process abundances observed through the spectroscopy of carbon stars \citep{abia:02,zamora:09} of initial metal content similar to our models (Z = 0.02 and Z=0.03). The M3.z3m2-hCBM model computed by Battino {\it et al.} \cite{battino:19} are reported in Figure \ref{fig:rb_cstars} while Figure \ref{fig:Rb_cstars_sara} reports the comparison of the NEWTON code output with observations. The variations in the predictions from the models using the Longland rates and the rates in the current paper without the Texas A\&M results included is small. However, the variations when including the Texas A\&M results is much more significant due to the reduced $^{22}$Ne($\alpha,n$)$^{25}$Mg reaction rate.
The ${^{22}}$Ne($\alpha,n$)$^{25}$Mg reaction does not dominate the bulk of $s$-process production in low-mass AGB stars but it still leaves its fingerprint on isotopic ratios involving branching-points and neutron-magic nuclei. Figures \ref{fig:Ba_isoratios} shows the MESA prediction of barium isotopic ratios compared to laboratory measurement of Silicon-Carbide (SiC) grains \citep{liu:14,liu:15}, which condensed in the ejected atmosphere of AGB stars, polluting the pristine solar nebula \citep{lugaro:03a,zinner:14}. The impact of the rate evaluated without the Texas A\&M results is visible but is smaller than the observational uncertainties. In the rate using the Texas A\&M results a small yet visible impact is present shifting the theoretical tracks towards the observational data, favouring a better agreement. In Figure \ref{fig:Ba_isoratios_sara} the NEWTON theoretical predictions reach similar values for the ${^{138}}$Ba/${^{136}}$Ba ratio even adopting a lower metallicity (Z=0.02) because the ${^{13}}$C-pocket of the NEWTON code, which is assumed to form because of the stellar magnetic field, is larger (4.8$\times 10^{-3}$M$_{\odot}$) and poor in ${^{14}}$N (see Ref. \cite{trippella} for more details).
The main differences between the results shown in Figures \ref{fig:rb_cstars}/\ref{fig:Rb_cstars_sara} and \ref{fig:Ba_isoratios}/\ref{fig:Ba_isoratios_sara} are due to the stellar nucleosynthesis models employed for calculation and to the different initial metal content. The theoretical tracks in Figure \ref{fig:Ba_isoratios} cover a good fraction of the values in ${^{138}}$Ba/${^{136}}$Ba because of the high metallicity adopted (Z=0.03), which favour the production of first-peak elements (Sr,Y and Zr) over second peak ones (Ba and La) due to the higher neutrons over Fe seeds ratio \citep{lugaro:18}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\columnwidth]{Rb.pdf}
\caption{Rb abundances as a function of $s$-process abundances - results obtained using the MESA models of Battino {\it et al.} \cite{battino:19} compared to spectroscopy.}
\label{fig:rb_cstars}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=0.95\columnwidth]{fig5a-color.png}}
\subfigure{
\includegraphics[width=0.95\columnwidth]{fig5b-color.png}
}
\caption{As in Figure \ref{fig:rb_cstars}, but showing the theoretical results obtained with the NEWTON code for a 3$M_{\odot}$ model (upper panel) and 2$M_{\odot}$ model (lower panel).}
\label{fig:Rb_cstars_sara}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=\columnwidth]{Ba138.pdf}
\caption{Comparison of measured Ba isotopic rations from presolar SiC grains with the results of MESA stellar models using the reaction rates from the present work.}
\label{fig:Ba_isoratios}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=0.95\columnwidth]{fig7a-color.png}
}
\subfigure{
\includegraphics[width=0.95\columnwidth]{fig7b-color.png}
}
\caption{As in Figure \ref{fig:Ba_isoratios}, but showing the theoretical results obtained with the NEWTON code. Lower panel: 2$M_{\odot}$ model. Upper panel: 3$M_{\odot}$ model.}
\label{fig:Ba_isoratios_sara}
\end{figure}
Figure \ref{fig:Zr_isoratios} shows the impact on the zirconium isotopic ratios of our newly evaluated rates for ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg and ${^{22}}$Ne($\alpha$,$\gamma$)${^{26}}$Mg with MESA code. Including the Texas A\&M results, the difference between the theoretical predictions is much larger, mainly due to a factor three reduction in the ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg rate at $T=0.3$ GK when our rate is compared to that of Longland {\it et al.} \cite{PhysRevC.85.065809}, in particular bringing the theoretical predictions in agreement with measured barium isotopic ratio from presolar grains. The lower ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg rate directly affects the $s$-process branching at ${^{95}}$Zr, impacting the production of ${^{96}}$Zr and lowering the predicted ${^{96}}$Zr/${^{94}}$Zr ratio; the comparison with measured ratios is greatly improved.
In Figure \ref{fig:Zr_isoratios_sara} the Zr isotopic mixture of the sample of grains in Figure \ref{fig:Zr_isoratios} is compared with the predictions of the $s$-process using the NEWTON code for a 3$M_{\odot}$ star, this model shows a reduction on the ${^{96}}$Zr$/{^{94}}$Zr ratio too, but the sensitivity to the $^{22}$Ne$+\alpha$ reaction rates is less pronounced, as the ${^{96}}$Zr/${^{94}}$Zr ratio is more efficiently reduced because of the extended $^{13}$C pocket.
The results of the MESA and NEWTON calculations are consistent in showing that the ${^{96}}$Zr/${^{94}}$Zr ratio is a factor of two lower when computed by using the recent nuclear data inputs of the Texas A\&M measurements \cite{OTA2020135256,JAYATISSA2020135267}. The shift between the model curves is less visible in Figure \ref{fig:Zr_isoratios_sara} because the values of the ${^{96}}$Zr/${^{94}}$Zr ratio are smaller than in Figure \ref{fig:Zr_isoratios} and the $\delta$ notation reduces the visible size of this effect: the $\delta_i = 1000(\frac{(^iX/^jX)_\mathrm{grain}}{(^iX/^jX)_\mathrm{standard}}-1)$. An expanded plot of the $\delta(^{96}$Zr/$^{94}$Zr$)$ values is shown in the lower panel of Figure \ref{fig:Zr_isoratios_sara}.
The differing predictions of the zirconium isotopic ratios from MESA and NEWTON are a consequence of considering two different mixing processes at the convective boundaries, resulting in greater $^{94}$Zr production in the NEWTON models compared to the MESA models, and a concomitant lower $^{96}$Zr/$^{94}$Zr ratio. The nuclear reaction rates can be a source of uncertainty comparable to the model uncertainties (see Figs. \ref{fig:Zr_isoratios} and \ref{fig:Zr_isoratios_sara}), and minimising the nuclear-physics uncertainties is a necessary step before the impact of the convective boundary mixing schemes on $s$-process nucleosynthesis in evolved stars may be understood.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\columnwidth]{Zr90.pdf}
\caption{Comparison of MESA stellar models results with measured Zr isotopic ratios from presolar SiC grains. The theoretical tracks show the results obtained with the ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg and ${^{22}}$Ne($\alpha$,$\gamma$)${^{26}}$Mg evaluated with and without the recent nuclear data inputs of the Texas A\&M measurements \cite{OTA2020135256,JAYATISSA2020135267} and the rates of Longland {\it et al.} \cite{PhysRevC.85.065809}.}
\label{fig:Zr_isoratios}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=0.95\columnwidth]{fig9b-color.png}}
\subfigure{
\includegraphics[width=0.95\columnwidth]{fig9b-zoom.pdf}}
\caption{As in Figure \ref{fig:Zr_isoratios} but showing the theoretical results obtained with the NEWTON code. The 3 theoretical tracks show the results obtained with ${^{22}}$Ne$+\alpha$ reaction rates of Longland {\it et al.} \cite{PhysRevC.85.065809} (in green) and the one suggested in this paper evaluated using the Texas A\&M measurements \cite{OTA2020135256,JAYATISSA2020135267} (in red) and without this last input (in blue). The lower panel is an enhanced region of the upper panel.}
\label{fig:Zr_isoratios_sara}
\end{figure}
\subsection{Weak $s$-process from massive star}
The weak $s$-process component (producing solar system $s$-process abundances between iron and zirconium; e.g. \cite{travaglio:04}) is produced during convective core helium burning and convective shell carbon-burning stages in massive stars (M$>$10M$_\odot$).
The ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg reaction is the principal neutron source for the weak $s$-process component. An important role is also played by the ${^{22}}$Ne($\alpha$,$\gamma$)${^{26}}$Mg reaction, which competes with the ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg reaction in typical weak $s$-process conditions, lowering the number of neutrons available for the $s$-process. In the following section, we first discuss the $s$-process production without the inclusion of the Texas A\&M results, and then with those results.
In Figure \ref{fig:MS_pf} we present the impact on the weak $s$-process nucleosynthesis of our new rates computed using the MESA code without the inclusion of the Texas A\&M results. We post-processed the stellar structure model of a 25 M$_{\odot}$ star with initial metal content Z=0.006 from Ref. \cite{ritter:18}. The Top and bottom panels of Figure \ref{fig:MS_pf} show the isotopic distribution at the end of core helium burning and shell carbon-burning respectively, extracted at Lagrangian mass and time coordinate specified in Figure \ref{fig:MS_kippe}. Overall, the impact is not large. Elements between the first-peak elements (Sr, Y and Zr) and A=130 are all overproduced by two order of magnitudes compared to their initial abundances, and they are all reduced when our rates are adopted instead of the rates from Ref. \cite{PhysRevC.85.065809}, but always by less than a factor of two.
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=0.95\columnwidth]{iso_distrib_Hecore.png}}
\subfigure{
\includegraphics[width=0.95\columnwidth]{iso_distrib_Cshell.png}}
\caption{Production factors of heavy isotopes between the iron and lead peak from the 25 M${_\odot}$, Z=0.006 model of Ref. \cite{ritter:18}. Top and bottom panels show the isotopic distribution at the end of core helium burning and shell carbon burning respectively. Lagrangian mass and time coordinate selected to extract the predictions here presented, are specified in figure \ref{fig:MS_kippe}. For each isotope, two different values are presented, corresponding to the adoption of the ${^{22}}$Ne+$\alpha$ rates without the inclusion of the Texas A\&M results which are presented in this work and from Longland {\it et al.} \cite{PhysRevC.85.065809}.}
\label{fig:MS_pf}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=\columnwidth]{kippenhahn_m25z6m3.png}}
\caption{Kippenhahn diagrams of our adopted 25 M${_\odot}$, Z=0.006 model. The convective zones are marked in gray. Lagrangian mass-coordinate is shown on the Y-axis, while the logarithm of the time remaining before the end of the star's life is given on the X-axis. Vertical red dashed lines mark the time selected to extract the abundance distributions presented in our plot, at the mass-coordinate specified by the red horizontal dotted line.}
\label{fig:MS_kippe}
\end{figure}
In Figure \ref{fig:MS_pf2}, we repeat the same test in a massive rotating star model (25 M$_{\odot}$ star, Z=0.0001) from \cite{hirschi:04}.
Rotation has indeed an important role in the weak $s$-process in massive star, as it allows the primary production of ${^{14}}$N, which is later converted into ${^{22}}$Ne by $\alpha$-particle captures. As a consequence, a larger quantity of neutrons are released by ${^{22}}$Ne($\alpha,n$)${^{25}}$Mg reactions, globally increasing the $s$-process efficiency. In this case, we again confirm the low impact of the new rates without the Texas A\&M results compared to Longland {\it et al.} \cite{PhysRevC.85.065809}, as the variations we obtained are all smaller than a factor of two.
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=0.95\columnwidth]{pf_ws_rot.pdf}}
\subfigure{
\includegraphics[width=0.95\columnwidth]{pf-ws-rot-1.pdf}}
\caption{Same test as presented in Figure \ref{fig:MS_pf}, but performed on the rotating massive star model with 25 M${_\odot}$ star, Z=0.00001 from Ref. \cite{hirschi:04}}
\label{fig:MS_pf2}
\end{figure}
In the same manner, in Figure \ref{fig:MS_pf_tamu} we show the impact of our new recommended reaction rates for ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg and ${^{22}}$Ne($\alpha$,$\gamma$)${^{26}}$Mg using the input nuclear data from the Texas A\&M experiments \cite{OTA2020135256,JAYATISSA2020135267}. As already seen for the main $s$-process from AGB stars, the difference from the theoretical predictions adopting the Longland {\it et al.} \cite{PhysRevC.85.065809} reaction rates is large. In particular, the production of Sr, Y and Zr drops by about a factor of five, and by one order of magnitude for A$>$100, or even two if the lower limit of the ${^{22}}$Ne($\alpha,n$)${^{25}}$Mg reaction rate is adopted.
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=0.95\columnwidth]{iso_distrib_Hecore_TAMU.png}}
\subfigure{
\includegraphics[width=0.95\columnwidth]{iso_distrib_Cshell_TAMU.png}}
\caption{Same as in Figure \ref{fig:MS_pf}, but showing our theoretical results obtained adopting the recommended ${^{22}}$Ne($\alpha$,$n$)${^{25}}$Mg and ${^{22}}$Ne($\alpha$,$\gamma$)${^{26}}$Mg reaction rates evaluated using the recent input nuclear data from Texas A\&M \cite{OTA2020135256,JAYATISSA2020135267}.}
\label{fig:MS_pf_tamu}
\end{figure}
\section{Conclusions}
The $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg and $^{22}$Ne($\alpha,n$)$^{25}$Mg reactions control the production of neutrons for the weak $s$-process in evolved massive stars and in the helium-flash in low-mass AGB stars. In this paper we have critically re-evaluated
the available nuclear data on the states in $^{26}$Mg which govern these reaction rates and have re-analysed these rates. We find that the rates are approximately unchanged from the evaluation of Longland {\it et al.} \cite{PhysRevC.85.065809}, in contrast to the evaluation of Talwar {\it et al.} \cite{PhysRevC.93.055803} which found a greatly increased $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate, and the evaluation of Massimi {\it et al.} \cite{Massimi20171} which found a greatly reduced $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate. Our disagreement with the results of Massimi {\it et al.} \cite{Massimi20171} may be attributed to the different methodologies for calculating the reaction rates as the inclusion of their nuclear data has only a small impact on the current evaluation.
The primary source of uncertainty in the $^{22}$Ne($\alpha,\gamma$)$^{26}$Mg reaction rate is whether a strong $\alpha$-cluster state exists at around $E_x = 11.17$ MeV. Evidence from two different measurements of the $^{22}$Ne($^6$Li,$d$)$^{26}$Mg $\alpha$-particle transfer reaction is contradictory, with experiments which have approximately the same excitation-energy resolution both observing two strong $\alpha$-cluster resonances separated by a similar energy gap but with inconsistent excitation energies. Resolution of this disagreement is an high priority for future experimental studies. Additional information, in particular estimates of resonance strengths and/or $\alpha$-particle partial widths of the states at $E_x = 10.949$, $11.112$, $11.163$ and $11.171$ MeV which are expected to control the behaviour of the reaction rates below the $E_r = 706$-keV resonance, and on the interference pattern between higher-energy resonances are also required.
In this work we have evaluated the reaction rates twice, once without the inclusion of the new results from two experiments performed at Texas A\&M and once with the inclusion of those results. This latter evaluation, with the inclusion of the Texas A\&M results, is our recommended rate.
The evaluations of the rates without the inclusion of the new Texas A\&M results produces no substantial change in the predictions of $s$-process nucleosynthesis. The inclusion of the new resonance strength for the $E_r = 706$-keV resonance derived from the Texas A\&M measurements results in predicted barium and zirconium isotopic ratios which better match the measured ratios from presolar SiC grains.
\section*{Acknowledgements}
This paper is based upon work from the \textquoteleft ChETEC\textquoteright\ COST Action (CA16117), supported by COST (European Cooperation in Science and Technology). PA thanks the Claude Leon Foundation for a fellowship under which part of this work was performed. UB and AB acknowledge support from the European Research Council (Grant agreements ERC-2015-STG No. 677497 and ERC-2019-STG No. 852016, respectively). RH acknowledges support from the World Premier International Research Centre Initiative (WPI Initiative).
| 8e9bc3119a64de271ac740331aa94ded61fb0d38 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In this work we bring forward the study of the normalized Laplacian that has been established for \textsl{chemical hypergraphs}: hypergraphs with the additional structure that each vertex in a hyperedge is either an input, an output or both (in which case we say that it is a catalyst for that hyperedge). Chemical hypergraphs have been introduced in \cite{Hypergraphs} with the idea of modelling chemical reaction networks and related ones, such as metabolic networks. In this model, each vertex represents a chemical element and each hyperedge represents a chemical reaction. Furthermore, in \cite{Master-Stability}, chemical hypergraphs have been used for modelling dynamical systems with high order interactions. In this model, the vertices represent oscillators while the hyperedges represent the interactions on which the dynamics depends.\newline
The spectrum of the normalized Laplacian $L$ reflects many structural properties of the network and several theoretical results on the eigenvalues have been established in \cite{Hypergraphs,Sharp,MulasZhang}. Furthermore, as shown in \cite{Sharp}, by defining the vertex degree in a way that it does not take catalysts into account, studying the spectrum of $L$ for chemical hypergraphs is equivalent to studying the spectrum of the \textsl{oriented hypergraphs} introduced in \cite{ReffRusnak} by Reff and Rusnak, in which catalysts are not included. Therefore, without loss of generality we can work on oriented hypergraphs. Here, in particular, we focus on the \textsl{bipartite} case and we show that the spectrum of the normalized Laplacian for bipartite chemical hypergraphs coincides with the spectrum of the \textsl{signless normalized Laplacian} that we introduce for classical hypergraphs. Furthermore, we establish the spectra of the signless normalized Laplacian for special families of such classical hypergraphs.\newline
Classical hypergraphs are widely used in various disciplines. For instance, they offer a valid model for transport networks \cite{andreotti2020eigenvalues}, neural networks (in whose context they are often called \textsl{neural codes}) \cite{neuro1,neuro2,neuro3,neuro5,neuro6,neuro7,neuro9,neuro10}, social networks \cite{ZhangLiu} and epidemiology networks \cite{BodoKatonaSimon}, just to mention some examples. It is worth noting that a simplicial complex $\mathcal{S}$ is a particular case of hypergraph with the additional constraint that, if a hyperedge belongs to $\mathcal{S}$, then also all its subsets belong to $\mathcal{S}$. Simplicial complexes are also widely present in applications. On the one hand, their more precise structure allows for a deeper theoretical study, compared to general hypergraphs. On the other hand, the constraints of simplicial complexes can be translated as constraints on the model, and this is not always convenient. Consider, for instance, a collaboration network that represents coauthoring of research papers: in this case, the fact that authors $A$, $B$ and $C$ have written a paper all together does not imply that $A$, $B$ and $C$ have all written single author papers, nor that $A$ and $B$ have written a paper together without $C$. In this case, a hypergraph would give a better model than a simplicial complex. \newline
\textbf{Structure of the paper.} In Section \ref{section:oriented} we introduce the basic definitions which are needed throughout the paper, while in Section \ref{section:twin} we introduce and discuss twin vertices. In Section \ref{section:bipartite} we prove new properties of bipartite oriented hypergraphs and we show that, from the spectral point of view, these are equivalent to classical hypergraphs with no input/output structure. Finally, in Section \ref{section:families} we investigate the spectra of new hypergraph structures that we introduce with the idea of generalizing well known graph structures.
\section{Basic definitions}\label{section:oriented}
\begin{definition}[\cite{ReffRusnak,Hypergraphs}]
An \textbf{oriented hypergraph} is a pair $\Gamma=(\mathcal{V},\mathcal{H})$ such that $\mathcal{V}$ is a finite set of vertices and $\mathcal{H}$ is a set such that every element $h$ in $\mathcal{H}$ is a pair of disjoint elements $(h_{in},h_{out})$ (input and output) in $\mathcal{P}(\mathcal{V})\setminus\{\emptyset\}$. The elements of $\mathcal{H}$ are called the \textbf{oriented hyperedges}. Changing the orientation of a hyperedge $h$ means exchanging its input and output, leading to the pair $(h_{out},h_{in})$.
\end{definition}
\begin{definition}
Given $h\in \mathcal{H}$, we say that two vertices $i$ and $j$ are \textbf{co-oriented} in $h$ if they belong to the same orientation sets of $h$; we say that they are \textbf{anti-oriented} in $h$ if they belong to different orientation sets of $h$.
\end{definition}
From now on, we fix a chemical hypergraph $\Gamma=(\mathcal{V},\mathcal{H})$ on $N$ vertices $v_1,\ldots,v_N$ and $M$ hyperedges $h_1,\ldots, h_M$. For simplicity, we assume that $\Gamma$ has no isolated vertices.
\begin{remark}\label{remark:graphs1}
Simple graphs can be seen as oriented hypergraphs such that $\# h_{in}=\# h_{out}=1$ for each $h\in \mathcal{H}$, that is, each edge has exactly one input and one output.
\end{remark}
\begin{definition}
The \textbf{underlying hypergraph} of $\Gamma$ is $\Gamma':=(\mathcal{V},\mathcal{H}')$ where
\begin{equation*}
\mathcal{H}':=\{(h_{in}\cup h_{out},\emptyset):h=(h_{in},h_{out})\in \mathcal{H}\}.
\end{equation*}
\end{definition}
\begin{definition}[\cite{Sharp}]
The \textbf{degree} of a vertex $v$ is
\begin{equation*}
\deg(v):=\#\text{ hyperedges containing $v$.}
\end{equation*}
Similarly, the \textbf{cardinality} of a hyperedge $h$ is
\begin{equation*}
\# h:=\#\{h_{in}\cup h_{out}\}.
\end{equation*}
\end{definition}
\begin{definition}[\cite{Hypergraphs,MulasZhang}]
The \textbf{normalized Laplace operator} associated to $\Gamma$ is the $N\times N$ matrix
\begin{equation*}
L:=\id -D^{-1}A,
\end{equation*}where $\id$ is the $N\times N$ identity matrix, $D$ is the \textbf{diagonal degree matrix} and $A$ is the \textbf{adjacency matrix} defined by $A_{ii}:=0$ for each $i=1,\ldots,n$ and
\begin{align*}
A_{ij}:=& \# \{\text{hyperedges in which }v_i \text{ and }v_j\text{ are anti-oriented}\}+\\
&-\# \{\text{hyperedges in which }v_i \text{ and }v_j\text{ are co-oriented}\}
\end{align*}for $i\neq j$.
\end{definition}
We define the \textbf{spectrum of $\Gamma$} as the spectrum of $L$. As shown in \cite{Hypergraphs,MulasZhang}, this spectrum is given by $N$ real, nonnegative eigenvalues whose sum is $N$. We denote them by
\begin{equation*}
\lambda_1\leq\ldots\leq\lambda_N.
\end{equation*}
\begin{definition}
We say that two vertices $v_i$ and $v_j$ are \textbf{adjacent}, denoted $v_i\sim v_j$, if they are contained at least in one common hyperedge.
\end{definition}
\begin{remark}\label{remark:graphs2}
Consider a graph $\Gamma$ and let $\Gamma'$ be its underlying hypergraph. Then, the adjacency matrix $A$ of $\Gamma$ and the adjacency matrix $A'$ of $\Gamma'$ are such that $A'=-A$, while the degree matrices of $\Gamma$ and $\Gamma'$ coincide. Therefore, the normalized Laplacians of $\Gamma$ and $\Gamma'$ are
\begin{equation*}
L=\id -D^{-1}A \qquad \text{ and }\qquad L'=\id +D^{-1}A=2\cdot \id -L,
\end{equation*}respectively. Hence, $\lambda$ is an eigenvalue for $L$ if and only if $2-\lambda$ is an eigenvalue for $L'$.
\end{remark}
\begin{definition}
Let $\Gamma$ be an oriented hypergraph and let $\Gamma'$ be its underlying hypergraph. The \textbf{signless normalized Laplacian} of $\Gamma$ is the normalized Laplacian of $\Gamma'$.
\end{definition}
\section{Twin vertices}\label{section:twin}
\begin{definition}[\cite{MulasZhang}]
Two vertices $v_i$ and $v_j$ are \textbf{duplicate} if $A_{ik}=A_{jk}$ for all $k$. In particular, $A_{ij}=A_{ji}=A_{ii}=0$.
\end{definition}
In \cite{MulasZhang} it is shown that $\hat{n}$ \textsl{duplicate vertices} produce the eigenvalue $1$ with multiplicity at least $\hat{n}-1$. Similarly, in this section we discuss \textsl{twin vertices}.
\begin{definition}\label{def:twins}
We say that two vertices $v_i$ and $v_j$ are \textbf{twins} if they belong exactly to the same hyperedges, with the same orientations. In particular, $A_{ij}=-\deg (v_i)=-\deg (v_j)$ and $A_{ik}=A_{jk}$ for all $k\neq i,j$.
\end{definition}
\begin{remark}
While duplicate vertices are known also for graphs, twin vertices cannot exist for graphs, since in this case one assumes that each edge has one input and one output.
\end{remark}
We now generalize the notions of duplicate vertices and twin vertices by defining \textsl{duplicate families of twin vertices}.
\begin{definition}
Let $\Gamma=(\mathcal{V},\mathcal{H})$ be an oriented hypergraph. We say that a family of vertices $\mathcal{V}_1\sqcup\ldots\sqcup \mathcal{V}_l\subset \mathcal{V}$ is a \textbf{$l$-duplicate family of $t$-twin vertices} if
\begin{itemize}
\item For each $i\in \{1,\ldots,l\}$, $\#\mathcal{V}_i=t$ and the $t$ vertices in $\mathcal{V}_i$ are twins to each other;
\item For each $i,j\in \{1,\ldots,l\}$ with $i\neq j$, for each $v_i\in \mathcal{V}_i$ and for each $v_j\in \mathcal{V}_j$, we have that $A_{ij}=0$ and $A_{ik}=A_{jk}$ for all vertices $v_k$ that are not in the $l$-family, i.e. $v_k\in \mathcal{V}\setminus \mathcal{V}_1\sqcup\ldots\sqcup \mathcal{V}_l$.
\end{itemize}
\end{definition}
\begin{proposition}\label{prop:lt}
If $\Gamma$ contains a $l$-duplicate family of $t$ twins, then:
\begin{itemize}
\item $t$ is eigenvalue with multiplicity at least $l-1$;
\item $0$ is an eigenvalue with multiplicity at least $l(t-1)$.
\end{itemize}
\end{proposition}
\begin{proof}In order to show that $t$ is eigenvalue with multiplicity at least $l-1$, consider the following $l-1$ functions. For $i=2,\ldots,l$, let $f_i:\mathcal{V}\rightarrow \mathbb{R}$ such that $f_i:=1$ on $\mathcal{V}_1$, $f_i:=-1$ on $\mathcal{V}_i$ and $f_i:=0$ otherwise. Then,
\begin{itemize}
\item For each $v_1\in \mathcal{V}_1$,
\begin{equation*}
Lf(v_1)=1-\frac{1}{\deg v_1}\sum_{v_1\neq v_j\in \mathcal{V}_1}-\deg v_1=1+t-1=t\cdot f(v_1);
\end{equation*}
\item For each $v_i\in \mathcal{V}_i$,
\begin{equation*}
Lf(v_i)=-1-\frac{1}{\deg v_i}\sum_{v_i\neq v_j\in \mathcal{V}_i}\deg v_i=-1-(t-1)=t\cdot f(v_i);
\end{equation*}
\item For each $v_k\in \mathcal{V}\setminus \mathcal{V}_1\sqcup\ldots\sqcup \mathcal{V}_l$,
\begin{equation*}
Lf(v_k)=-\frac{1}{\deg v_k}\left(\sum_{v_1\in \mathcal{V}_1}A_{1k}-\sum_{v_i\in \mathcal{V}_i}A_{ik} \right)=0=t\cdot f(v_k).
\end{equation*}Therefore, $f_i$ is an eigenfunction for $t$. Furthermore, the functions $f_2,\ldots, f_l$ are linearly independent. Therefore, $t$ is an eigenvalue with multiplicity at least $l-1$.
\end{itemize}
Similarly, in order to prove that $0$ is eigenvalue with multiplicity at least $l(t-1)$, let $\mathcal{V}_i=\{v^i_1,\ldots,v^i_t\}$ and consider the $l(t-1)$ functions $g_j^i:\mathcal{V}\rightarrow\mathbb{R}$ defined as follows, for $i=1,\ldots,l$ and $j=2,\ldots,t$. Let $g_j^i(v_1^i):=1$, $g_j^i(v_j^i):=-1$ and $g_j^i:=0$ otherwise. Then, by \cite[Equation (5)]{Hypergraphs}, it is clear that each $g_j^i$ is an eigenfunction for $0$. Since, furthermore, these are $l(t-1)$ linearly independent functions, $0$ has multiplicity at least $l(t-1)$.
\end{proof}
\begin{proposition}\label{prop:twins}
If $\Gamma$ has $\hat{n}$ vertices that are twins to each other, $0$ is an eigenvalue with multiplicity at least $\hat{n}-1$. Furthermore, if $v_i$ and $v_j$ are twin vertices and $f$ is an eigenfunction for $L$ with eigenvalue $\lambda\neq 0$, then $f(v_i)=f(v_j)$.
\end{proposition}
\begin{proof}
The first claim follows from Proposition \ref{prop:lt}, by taking $t=1$.\newline
Now, assume that $v_i$ and $v_j$ are twin vertices and let $f$ be an eigenfunction for $L$ with eigenvalue $\lambda\neq 0$. Then,
\begin{equation*}
\lambda f(v_i)= Lf(v_i)=f(v_i)+f(v_j)-\frac{1}{\deg v_i}\left(\sum_{k\neq i,j}A_{ik}f(v_k)\right)=Lf(v_j)=\lambda f(v_j).
\end{equation*}Since $\lambda\neq 0$, this implies that $f(v_i)=f(v_j)$.
\end{proof}
\section{Bipartite hypergraphs}\label{section:bipartite}
\begin{definition}[\cite{Hypergraphs}\label{def:bipartite}]
We say that a hypergraph $\Gamma$ is \textsl{bipartite} if one can decompose the vertex set as a disjoint union $\mathcal{V}=\mathcal{V}_1\sqcup \mathcal{V}_2$ such that, for every hyperedge $h$ of $\Gamma$, either $h$ has all its inputs in $\mathcal{V}_1$ and all its outputs in $\mathcal{V}_2$, or vice versa (Figure \ref{fig:bipartiteh}).
\end{definition}
\begin{figure}[t!]
\begin{center}
\begin{tikzpicture}
\node (v3) at (1,0) {};
\node (v2) at (1,1) {};
\node (v1) at (1,2) {};
\node (v6) at (5,0) {};
\node (v5) at (5,1) {};
\node (v4) at (5,2) {};
\begin{scope}[fill opacity=0.5]
\filldraw[fill=gray!70] ($(v1)+(0,0.5)$)
to[out=180,in=180] ($(v2) + (0,-0.5)$)
to[out=0,in=180] ($(v5) + (0,-0.5)$)
to[out=0,in=0] ($(v4) + (0,0.5)$)
to[out=180,in=0] ($(v1)+(0,0.5)$);
\filldraw[fill=white!70] ($(v2)+(0,0.5)$)
to[out=180,in=180] ($(v3) + (0,-0.5)$)
to[out=0,in=180] ($(v6) + (0,-0.5)$)
to[out=0,in=0] ($(v5) + (0,0.5)$)
to[out=180,in=0] ($(v2)+(0,0.5)$);
\end{scope}
\fill (v1) circle (0.05) node [right] {$v_1$} node [above] {\color{black}$+$};
\fill (v2) circle (0.05) node [right] {$v_2$} node [above] {\color{black}$+$} node [below] {\color{gray}$-$};
\fill (v3) circle (0.05) node [right] {$v_3$} node [below] {\color{gray}$-$};
\fill (v4) circle (0.05) node [left] {$v_4$} node [above] {\color{black}$-$};
\fill (v5) circle (0.05) node [left] {$v_5$} node [above] {\color{black}$-$}node [below] {\color{gray}$+$};
\fill (v6) circle (0.05) node [left] {$v_6$} node [below] {\color{gray}$+$};
\node at (0,2) {\color{black}$h_1$};
\node at (0,0) {\color{gray}$h_2$};
\end{tikzpicture}
\end{center}
\caption{A bipartite hypergraph with $\mathcal{V}_1=\{v_1,v_2,v_3\}$ and $\mathcal{V}_2=\{v_4,v_5,v_6\}$.}\label{fig:bipartiteh}
\end{figure}
We now give the definition of \textsl{vertex-bipartite} hypergraph that, as we shall see in Lemma \ref{lemma:vertexbipartite} below, coincides with the definition of bipartite hypergraph.
\begin{definition}
We say that a hypergraph $\Gamma$ is \textbf{vertex-bipartite} if one can decompose the hyperedge set as a disjoint union $\mathcal{H}=\mathcal{H}_1\sqcup \mathcal{H}_2$ such that, for every vertex $v$ of $\Gamma$, either $v$ is an input only for hyperedges in $\mathcal{H}_1$ and it is an output only for hyperedges in $\mathcal{H}_2$, or vice versa. \end{definition}
\begin{lemma}\label{lemma:vertexbipartite}
Up to changing the orientation of some hyperedges, a hypergraph is bipartite if and only if it is vertex-bipartite.
\end{lemma}
\begin{proof}Assume that $\Gamma$ is bipartite. Up to changing the orientation of some hyperedges, we can assume that the vertex set has a decomposition $\mathcal{V}=\mathcal{V}_1\sqcup \mathcal{V}_2$ such that each hyperedge $h$ has all its inputs in $\mathcal{V}_1$ and all its outputs in $\mathcal{V}_2$. Therefore, every vertex in $\mathcal{V}_1$ is an input only for hyperedges in $\mathcal{H}$, and every vertex in $\mathcal{V}_2$ is only an output for hyperedges in $\mathcal{H}$. It follows that the decomposition of the hyperedge set as $\mathcal{H}=\mathcal{H}\sqcup \emptyset$ gives a vertex-bipartition.\newline
Now, assume that $\Gamma$ is vertex-bipartite, with $\mathcal{H}=\mathcal{H}_1\sqcup \mathcal{H}_2$. Assume, by contradiction, that $\Gamma$ is not bipartite. Then, up to changing the orientation of some hyperedges, there exist two vertices $v,w\in \mathcal{V}$ and two hyperedges $h_1,h_2\in \mathcal{H}$ such that:
\begin{enumerate}
\item\label{item1} $h_1$ has both $v$ and $w$ as inputs;
\item\label{item2} $h_2$ has $v$ as input and $w$ as output.
\end{enumerate}The fact that $v$ is an input in both $h_1$ and $h_2$ implies that $h_1$ and $h_2$ are in the same $\mathcal{H}_i$. On the other hand, the fact that $w$ is an input for $h_1$ and an output for $h_2$ implies that $h_1$ and $h_2$ do not belong to the same $\mathcal{H}_i$. This brings to a contradiction. Therefore, $\Gamma$ is bipartite.
\end{proof}
\begin{proposition}\label{prop_bipartite}
If $\Gamma$ is bipartite, it is isospectral to its underlying hypergraph, therefore, in particular, also to every other bipartite hypergraph that has the same underlying hypergraph as $\Gamma$.
\end{proposition}
\begin{proof}Since $\Gamma$ is bipartite, up to switching (without loss of generality) the orientations of some hyperedges we can assume that all the inputs are in $\mathcal{V}_1$ and all the outputs are in $\mathcal{V}_2$, with $\mathcal{V}=\mathcal{V}_1\sqcup \mathcal{V}_2$. Furthermore, by Lemma 49 in \cite{Hypergraphs}, we can move a vertex from $\mathcal{V}_1$ to $\mathcal{V}_2$ or vice versa, by letting it be always an output or always an input, without affecting the spectrum. In particular, if we move all vertices to $\mathcal{V}_1$, we obtain the underlying hypergraph of $\Gamma$.
\end{proof}
\begin{remark}
As a consequence of Proposition \ref{prop_bipartite}, without loss of generality we can always assume that a bipartite hypergraph $\Gamma$ has only inputs, when studying the spectrum of the normalized Laplacian. In this case,
\begin{itemize}
\item $A_{ij}=-\# \{h\in \mathcal{H}:v_i,v_j\in \mathcal{H}\}$ for each $i\neq j$;
\item $\sum_jA_{ij}=-\sum_{h\ni v_i}(\# h-1)$, for each $v_i\in \mathcal{V}$.
\end{itemize}
\end{remark}
From here on we work on a hypergraph $\Gamma=(\mathcal{V},\mathcal{H})$ that has only inputs. Therefore, we focus on the signless normalized Laplacian of classical hypergraphs.
\section{Families of hypergraphs}\label{section:families}
\subsection{Hyperflowers}\label{section:flowers}
We now introduce and study \textsl{hyperflowers}: hypergraphs in which there is a set of nodes, the \textsl{core,} that is well connected to the other vertices, and a set of \textsl{peripheral nodes} such that each of them is contained in exactly one hyperedge. Hyperflowers are therefore a generalization of star graphs \cite{ANDREOTTI2018206}.
\begin{definition}\label{def:hyperflowers}
A \textbf{$(l,r)$-hyperflower with t twins} (Figure \ref{fig:hyperflowertwins}) is an hypergraph $\Gamma=(\mathcal{V},\mathcal{H})$ whose vertex set can be written as $ \mathcal{V}=U\sqcup \mathcal{W}$, where:
\begin{itemize}
\item $U$ is a set of $t\cdot l$ nodes $v_{11},\ldots,v_{1l},\ldots,v_{t1},\ldots,v_{tl}$ which are called \textbf{peripheral};
\item There exist $r$ disjoint sets of vertices $h_1,\ldots,h_r \in \mathcal{P}(\mathcal{W})\setminus\{\emptyset\}$ such that
$$\mathcal{H}=\{h|h=h_i\cup\bigcup_{z=1}^{t} v_{zj} \mbox{ for }i=1,\ldots,r \mbox{ and } j=1,\ldots,l \}.$$
\end{itemize}If $t=1$, we simply say that $\Gamma$ is a \textbf{$(l,r)$-hyperflower}.\newline
If $r=1$, we simply say that $\Gamma$ is a \textbf{$l$-hyperflower with t twins}.
\end{definition}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=6cm]{Hyperflower.jpg}
\caption{A $5$-hyperflower with $3$ twins.} \label{fig:hyperflowertwins}
\end{center}
\end{figure}
\begin{remark}
The $(l,r)$-hyperflowers in Definition \ref{def:hyperflowers} are a particular case of the hyperstars in \cite{andreotti2020eigenvalues}, that also include weights and non-disjoint sets $h_1,\ldots,h_r$. Here we choose to study the particular structure of $(l,r)$-hyperflowers (and their generalizations with twins) because the strong symmetries of these structures allows for a deeper study of the spectrum.
\end{remark}
\begin{proposition}\label{prop:l2hyperflower}
The spectrum of the $(l,2)$-hyperflower on $N$ nodes is given by:
\begin{itemize}
\item $0$, with multiplicity $N-l-1$;
\item $1$, with multiplicity $\geq l-1$;
\item $\lambda_N>1$;
\item $\lambda_{N-1}=N-\lambda_N-l+1\geq 1$.
\end{itemize}In the particular case in which $\#h$ is constant for each $h\in \mathcal{H}$, $\lambda_N=\frac{N-l}{2}+1$ and $\lambda_{N-1}=\frac{N-l}{2}$.
\end{proposition}
\begin{proof}
By \cite[Corollary 3.5]{MulasZhang}, $1$ is an eigenvalue with multiplicity at least $l-1$. Now, the $N-l$ vertices $v_{l+1},\ldots,v_N$ form two classes of twin vertices that generate the eigenvalue $0$ with multiplicity at least $N-l-2$. In particular, there exist $N-l-2$ linearly independent corresponding eigenfunctions $f_i:\mathcal{V}\rightarrow\mathbb{R}$ such that $f_i(v)=1$ for some $v\notin\{v_1,\ldots,v_l\}$, $f_i(w)=-1$ for a given $w$ twin of $v$, and $f_i=0$ otherwise. If we let $g(v_j):=1$ for each $j=1,\ldots,l$, $g(v'_1):=-1$ for exactly one $v'_1\in h_1$ and $g(v'_2):=-1$ for exactly one $v'_2\in h_2$, it's easy to see that $g$ is also an eigenfunction of $0$. Furthermore, the $f_i$'s and $g$ are all linearly independent, which implies that $0$ has multiplicity at least $N-l-1$.\newline
Now, by \cite[Theorem 3.1]{Sharp}, $\lambda_N\geq \frac{\sum_{h\in \mathcal{H}}\# h}{|\mathcal{H}|}>1$. We have therefore listed already $N-1$ eigenvalues and there is only one eigenvalue $\lambda$ missing. Since $\sum_{i=1}^N\lambda_i=N$, we have that $\lambda=N-\lambda_N-l+1$. In particular, since by \cite[Theorem 3.1]{Sharp} $\lambda_N\leq\max_{h\in \mathcal{H}}\#h$ with equality if and only if $\#h$ is constant, and $\max_{h\in \mathcal{H}}\#h\leq N-l$, we have that
\begin{equation*}
\lambda=N-\lambda_N-l+1\geq 1,
\end{equation*}with equality if and only if $\#h$ is constant and equal to $N-l$, that is, if and only if $\#h_1=\#h_2=1$. Hence, $\lambda=\lambda_{N-1}$ and we have that $\lambda_{N-1}=1$ if and only if $\#h_1=\#h_2=1$.\newline
In general, if $\#h$ is constant for each $h\in \mathcal{H}$, then by \cite[Theorem 3.1]{Sharp} $\lambda_N=\#h=\frac{N-l}{2}+1$ and therefore $\lambda_{N-1}=\frac{N-l}{2}$.
\end{proof}
\begin{proposition}\label{prop:r1}
Let $\Gamma$ be an $(l,r)$--hyperflower with peripheral vertices $v_1,\ldots,v_l$. Let $\hat{\Gamma}:=(\hat{\mathcal{V}},\hat{\mathcal{H}})$ be the $(1,r)$--hyperflower defined by
\begin{equation*}
\hat{\mathcal{V}}:=\mathcal{V}\setminus\{v_2,\ldots,v_l\} \qquad\text{and}\qquad \hat{\mathcal{H}}:=\{h\in \mathcal{H}: v_2,\ldots,v_l\notin h\}.
\end{equation*}Then, the spectrum of $\Gamma$ is given by:
\begin{itemize}
\item The $N-l+1$ eigenvalues of $\hat{\Gamma}$, with multiplicity;
\item $1$, with multiplicity at least $l-1$.
\end{itemize}
\end{proposition}
\begin{proof}
By \cite[Corollary 3.5]{MulasZhang}, adding $v_2,\ldots,v_l$ to $\hat{\Gamma}$ produces the eigenvalue $1$ with multiplicity $l-1$. Therefore, it is left to show that, if $\lambda$ is an eigenvalue of $\hat{\Gamma}$, then $\lambda$ is also an eigenvalue of $\Gamma$. Let $L$ and $A$ be the Laplacian and the adjacency matrix on $\Gamma$, respectively, and let $\hat{L}$ and $\hat{A}$ be the Laplacian and the adjacency matrix on $\hat{\Gamma}$, respectively. Let also $\hat{f}$ be an eigenfunction for $\hat{\Gamma}$ corresponding to the eigenvalue $\lambda$. Then,
\begin{equation*}
\hat{L}\hat{f}(v_k)=\hat{f}(v_k)-\frac{1}{\deg_{\hat{\Gamma}}v_k}\sum_{v_i\in \hat{\mathcal{V}}\setminus \{v_k\}}\hat{A}_{ik}\hat{f}(v_i)=\lambda\cdot \hat{f}(v_k), \qquad\text{for all }v_k\in \hat{\mathcal{V}}.
\end{equation*}Now, let $f:\mathcal{V}\rightarrow \mathbb{R}$ be such that $f:=\hat{f}$ on $\hat{\mathcal{V}}$ and $f(v_2):=\ldots:=f(v_l):=\hat{f}(v_1)$.
Then,
\begin{align*}
Lf(v_1)&=f(v_1)-\frac{1}{\deg v_1}\sum_{v_i\in \hat{\mathcal{V}}\setminus \{v_1\}}A_{i1}f(v_i)=\hat{f}(v_1)-\frac{1}{\deg_{\hat{\Gamma}} v_1}\sum_{v_i\in \hat{\mathcal{V}}\setminus \{v_1\}}\hat{A}_{i1}\hat{f}(v_i)\\&=\hat{L}\hat{f}(v_k) =\lambda\cdot \hat{f}(v_1) =\lambda\cdot f(v_1).
\end{align*}
Similarly, for $j\in 2,\ldots,l$,
\begin{equation*}
Lf(v_j)=f(v_j)-\frac{1}{\deg v_j}\sum_{v_i\in \hat{\mathcal{V}}\setminus \{v_1\}}A_{ij}f(v_i)=\hat{f}(v_1)-\frac{1}{\deg_{\hat{\Gamma}} v_1}\sum_{v_i\in \hat{\mathcal{V}}\setminus \{v_1\}}\hat{A}_{i1}\hat{f}(v_i)=\lambda\cdot \hat{f}(v_1)=\lambda\cdot f(v_j).
\end{equation*}Furthermore, for each $v_k\in \mathcal{V}\setminus\{v_1,\ldots,v_l\}$, we have that
\begin{itemize}
\item $\deg_{\hat{\Gamma}}(v_k)=1$ while $\deg(v_k)=l$;
\item For each $v_{k'}\in \mathcal{V}\setminus\{v_1,\ldots,v_l,v_k\}$ such that $\hat{A}_{kk'}\neq 0$, $\hat{A}_{kk'}=-1$ while $A_{kk'}=-l$;
\item $\hat{A}_{k1}=A_{k1}=-1$, and $A_{kj}=-1$ for each $j\in 2,\ldots, l$.
\end{itemize}Therefore, for for each $v_k\in \mathcal{V}\setminus\{v_1,\ldots,v_l\}$,
\begin{align*}
Lf(v_k)&=f(v_k)-\frac{1}{\deg v_k}\left(\sum_{k'}A_{kk'}f(v_{k'})+\sum_{j=1}^lA_{kj}f(v_j)\right)
\\&=\hat{f}(v_k)-\frac{1}{l}\left(\sum_{k'}(-l)\hat{f}(v_{k'})+(-1)\sum_{j=1}^l\hat{f}(v_1)\right)
\\&= \hat{f}(v_k)+\sum_{k'}\hat{f}(v_{k'})+\hat{f}(v_1)
\\&=\hat{L}\hat{f}(v_k)=\lambda\cdot \hat{f}(v_k)=\lambda\cdot f(v_k).
\end{align*}This proves that $\lambda$ is an eigenvalue for $L$, and $f$ is a corresponding eigenfunction.
\end{proof}
\begin{remark}
Proposition \ref{prop:r1} tells us that, in order to know the spectrum of a $(l,r)$--hyperflower, we can study the spectrum of the $(1,r)$--hyperflower obtained by deleting $l-1$ peripheral vertices and the hyperedges containing them, and then add $l-1$ $1$'s to the spectrum.
\end{remark}
\begin{proposition}The spectrum of the $l$-hyperflower with $t$ twins is given by:
\begin{itemize}
\item $0$, with multiplicity $N-l$;
\item $t$, with multiplicity $l-1$;
\item $\lambda_N=N-tl+t$.\end{itemize}\end{proposition}
\begin{proof}Since all hyperedges have cardinality $N-tl+t$, by \cite[Theorem 3.1]{Sharp} we have that $\lambda_N=N-tl+t$. Furthermore, by Proposition \ref{prop:lt}, $t$ is an eigenvalue with multiplicity at least $l-1$. Since, clearly, $N-tl+t>t$, we have listed $l$ eigenvalues whose sum is $N$. It follows that $0$ has multiplicity $N-l$.
\end{proof}
\subsection{Complete hypergraphs}
\begin{definition}[\cite{MulasZhang}]
We say that $\Gamma=(\mathcal{V},\mathcal{H})$ is the \textbf{$c$-complete hypergraph}, for some $c\geq 2$, if $\mathcal{V}$ has cardinality $N$ and $\mathcal{H}$ is given by all possible ${N \choose c}$ hyperedges of cardinality $c$.
\end{definition}
\begin{proposition}\label{prop:complete}
The spectrum of the $c$-complete hypergraph is given by:
\begin{itemize}
\item $\frac{N-c}{N-1}$, with multiplicity $N-1$;
\item $c$, with multiplicity $1$.
\end{itemize}
\end{proposition}
\begin{proof}By \cite[Theorem 3.1]{Sharp}, $\lambda_N=c$. Now, observe that each vertex $v$ has degree $d:={N-1 \choose c-1}$, while $a:=A_{ij}=-{N-2 \choose c-2}$ is constant for all $i\neq j$. Therefore, $\frac{a}{d}=-\frac{c-1}{N-1}$ and
\begin{equation*}
Lf(v)=f(v)-\frac{a}{d}\left(\sum_{w\neq v}f(w)\right)=f(v)+\frac{c-1}{N-1}\left(\sum_{w\neq v}f(w)\right), \qquad \forall v\in \mathcal{V}.
\end{equation*}Now, for each $i=2,\ldots,N$, let $f(v_1):=1$, $f(v_i):=-1$ and $f:=0$ otherwise. Then,
\begin{itemize}
\item $Lf(v_1)=1-\frac{c-1}{N-1}=\frac{N-c}{N-1}\cdot f(v_1)$,
\item $Lf(v_i)=-1+\frac{c-1}{N-1}=\frac{N-c}{N-1}\cdot f(v_i)$, and
\item $Lf(v_j)=0=\frac{N-c}{N-1}\cdot f(v_j)$ for all $j\neq 1,i$.
\end{itemize}Therefore, the $f_i$'s are $N-1$ linearly independent eigenfunctions for $\frac{N-c}{N-1}$. This proves the claim.
\end{proof}
\begin{ex}
Proposition \ref{prop:complete} tells us that the signless spectrum of the complete graph on $N$ nodes is given by $\frac{N-2}{N-1}$, with multiplicity $N-1$, and $2$ with multiplicity $1$. By Remark \ref{remark:graphs2}, this is equivalent to saying that the spectrum of the complete graph is given by $\frac{N}{N-1}$, with multiplicity $N-1$, and $0$ with multiplicity $1$. This is a well known result (see \cite{Chung}) and Proposition \ref{prop:complete} generalizes it.
\end{ex}
\subsection{Lattice Hypergraphs}
\textsl{Lattice graphs}, also called \textsl{grid graphs}, are well known both in graph theory and in applications \cite{Asllani2015, Andreotti2015ModelingTF,Lattice1,Lattice2,Lattice3,Lattice4,Lattice5,Lattice6}. For instance, they model topologies used in transportation networks, such as the Manhattan street network, and crystal structures used in crystallography. These structures and their spectra are also widely used in statistical mechanics, in the study of ASEP, TASEP and SSEP models \cite{Mallick_2011, Speer1994, Schtz2017FluctuationsIS}, which have applications in the Ising model, (lattice) gas and which also describe the movement of ribosomes along the mRNA \cite{Gritsenko2015UnbiasedQM}. In this section we generalize the notion of lattice graph to the case of hypergraphs.
\begin{definition}
Given $l\in \mathbb{N}_{\geq 2}$, we define the \textbf{$l$-lattice} as the hypergraph $\Gamma=(\mathcal{V},\mathcal{H})$ on $l^2$ nodes and $2l$ hyperedges that can be drawn so that:
\begin{itemize}
\item The vertices form a $l\times l$ grid, and
\item The hyperedges are exactly the rows and the columns of the grid (Figure \ref{fig:lattice}).
\end{itemize}
\end{definition}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=4cm]{Reticolo.png}
\caption{A $3$-lattice.} \label{fig:lattice}
\end{center}
\end{figure}
\begin{proposition}
The spectrum of the $l$-lattice is given by:
\begin{itemize}
\item $0$, with multiplicity $l^2-2l+1$;
\item $\frac{l}{2}$, with multiplicity $2(l-1)$;
\item $l$, with multiplicity $1$.
\end{itemize}
\end{proposition}
\begin{proof}By \cite[Theorem 3.1]{Sharp}, $\lambda_{l^2}=l$. Furthermore, by \cite[Corollay 33]{Hypergraphs}, since the maximum number of \textsl{linearly independent hyperedges} is $2l-1$, this implies that $0$ is an eigenvalue with multiplicity $l^2-2l+1$.\newline
Now, observe that $\deg v=2$ for each $v$ and
\begin{equation*}
A_{ij}=\begin{cases}-1 &\text{if } v_i\sim v_j\\ 0 &\text{otherwise,}\end{cases}
\end{equation*}for all $i\neq j$. Therefore,
\begin{equation}\label{eq:Llattice}
Lf(v)=f(v)+\frac{1}{2}\left(\sum_{w\sim v}f(w)\right), \qquad \text{for all }v\in \mathcal{V}.
\end{equation}Fix a row of the $l$-lattice given by the vertices $w_1,\ldots,w_l$. For $i=1,\ldots,l-1$, let $f_i:\mathcal{V}\rightarrow\mathbb{R}$ be $1$ on the neighbors of $w_i$ with respect to the row, $-1$ on the neighbors of $w_i$ with respect to its column, and $0$ otherwise. Then, by \eqref{eq:Llattice}, it is easy to check that $f_i$ is an eigenfunction for $\frac{l}{2}$. Since the $f_i$'s are linearly independent, this proves the claim.
\end{proof}
\subsection{Hypercycles}
\begin{definition}
Fix $N$ and $l\in\{2,\ldots,\frac{N}{2}\}$. We say that $\Gamma=(\mathcal{V},\mathcal{H})$ is the \textbf{$l$-hypercycle} on $N$ nodes (Figure \ref{fig:cycle}) if $\mathcal{V}=\{v_1,\ldots,v_N\}$, $\mathcal{H}=\{h_1,\ldots,h_N\}$ and
\begin{equation*}
h_i=\{v_i,\ldots,v_{i+l-1}\},
\end{equation*}where we let $v_{N+i}:=v_{i}$ for each $i=1,\ldots, N$.
\end{definition}
\begin{theorem}
The eigenvalues of the $l$-hypercycle are
\begin{equation*}
\lambda_i=1+\frac{\sum_{r=1}^Nm(r)\cdot \cos \left(\frac{2\pi ir}{N}\right)}{l},\qquad \text{for }i=1,\ldots,N,
\end{equation*}where $m:\{0,\ldots,N\}\rightarrow\mathbb{Z}$ is such that:
\begin{itemize}
\item $m(r):=l-r$ for all $r\in \{1,\ldots,l-1\}$
\item $m(N-k):=m(k)=l-k$ for all $k\in \{1,\ldots,l-1\}$
\item $m:=0$ otherwise.
\end{itemize}
\end{theorem}
\begin{proof}By construction, all vertices have degree $l$. Therefore, by \cite[Remark 2.17]{MulasZhang}, proving the claim is equivalent to proving that the eigenvalues of the adjacency matrix are
\begin{equation*}
\mu_i=-\sum_{r=1}^Nm(r)\cdot \cos \left(\frac{2\pi ir}{N}\right),\qquad \text{for }i=1,\ldots,N.
\end{equation*}Observe that the adjacency matrix can be written as
\begin{equation*}A=-\begin{bmatrix}
0 & l-1 & l-2 &\ldots & 1 & 0 & \ldots & 0 & 1 & \ldots & l-2 &l-1\\
l-1 & 0 & l-1 & l-2 & \ldots & 1 & 0 & \ldots & 0 & 1 & \ldots & l-2\\
l-2 & l-1 & 0 & \ddots & \ddots & \ddots & \ddots & \ddots & & \ddots & \ddots & \vdots\\
\vdots & \ddots& \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & & \ddots & 1\\
1 & & & & & & & & & & & 0\\
0 & 1 & & & & & & & & & \ddots & \vdots\\
\vdots & 0 & \ddots & & & & & & & & \ddots & 0\\
0 & & \ddots& & & & & & & & & 1\\
1 & 0 & & & & & & & & \ddots & \ddots& \vdots\\
\vdots & \ddots & \ddots& & & & & & \ddots & \ddots & l-1 & l-2\\
l-2 & & \ddots & \ddots & & \ddots & \ddots & & & l-1 & 0 & l-1\\
l-1 & l-2 & \ldots & \ldots & 0 & \ldots & 0 & 1 & \ldots & \ldots & l-1 & 0 \\
\end{bmatrix}\end{equation*}
Therefore,
\begin{equation*}A=-\begin{bmatrix}
m(0) & m(N-1) & m(N-2) &\ldots & m(1)\\
m(1) & m(0) & m(N-1) &\ldots & m(2)\\
m(2) & m(1) & m(0) &\ldots & m(3)\\
\vdots & \vdots & \vdots & \ddots & \vdots\\
m(N-1) & m(N-2) & m(N-3) & \ldots & m(0)
\end{bmatrix}\end{equation*} where \begin{itemize}
\item $m(r):=l-r$ for all $r\in \{1,\ldots,l-1\}$
\item $m(N-k):=m(k)=l-k$ for all $k\in \{1,\ldots,l-1\}$
\item $m:=0$ otherwise.
\end{itemize}
Hence, $A$ is a (symmetric) circulant matrix. By \cite{circulant}, the eigenvalues of $A$ are
\begin{equation*}
\mu_i=-\sum_{r=1}^Nm(r)\cdot \cos \left(\frac{2\pi ir}{N}\right),\qquad \text{for }i=1,\ldots,N.
\end{equation*}This proves the claim.
\end{proof}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=5cm]{Cycle2.jpg}
\caption{The $3$-hypercycle on $6$ nodes.} \label{fig:cycle}
\end{center}
\end{figure}
| ce625891ed6e8de575250ae24c3316eb2f7bd499 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Grids}
\begin{table}[ht]
\centering
\begin{tabular}[t]{cccccc}
\hline
\hline
$E_{\text{trans}}$ [eV]&$v$&$j$&$E_{\text{trans}}'$ [eV]&$v'$&$j'$\\
\hline
0.0&0&0&0.1&0&0\\
0.2&2&15&0.3&2&20\\
0.4&4&30&0.5&4&40\\
0.6&6&45&1.0&6&60\\
1.0&9&60&1.5&9&80\\
1.5&12&90&2.0&12&100\\
2.0&15&120&2.5&15&125\\
2.5&18&150&3.0&18&150\\
3.0&21&180&3.5&21&175\\
3.5&24&210&4.5&24&200\\
4.5&27&240&5.5&27&220\\
5.5&30&-&6.5&30&240\\
6.5&33&-&7.5&33&-\\
7.5&36&-&8.5&36&-\\
8.5&-&-&9.5&42&-\\
9.5&-&-&10.5&47&-\\
10.5&-&-&-&-&-\\
11.5&-&-&-&-&-\\
\hline
\hline
\end{tabular}
\caption{Grid points used in this work for sampling the
reactant ($P(E_{\text{trans}}), P(v), P(j)$) and product state
distributions ($ P(E_{\text{trans}}'), P(v'), P(j')$).}
\label{sifig:grid_table}
\end{table}%
\newpage
\section{Statistical Evaluation}
\label{si:stat_eval}
For statistical evaluation, NNs in the F- and G-based models were
trained on 10 independent random splits of $N_{\rm tot}$ into $N_{\rm
train}$, $N_{\rm valid}$ and $N_{\rm test}$ for Set1 as well as Set1
and Set3, respectively. For each of the 10 resulting F- and G-based
models, ${\rm RMSD}_{\text{NN}}$, $R^2_{\text{NN}}$ and ${\rm
RMSD}_{\text{QCT}}$, $R^2_{\text{QCT}}$ values were evaluated over
the test set and subsequently the corresponding mean and standard
deviation were calculated. These results are displayed in Table
\ref{sifig:statistical_eval}. Taking the reported standard deviations
as a reference, we can be confident that any performance difference
between two models larger than $\sim 0.0001$ is not solely of
statistical nature. In particular, we expect this to apply to
approaches and data sets other than the ones reported in Table
\ref{sifig:statistical_eval} and will therefore refer to this estimate
throughout this work.\\
\begin{table}[ht]
\centering
\begin{tabular}[t]{l||c|c|c}
\hline
\hline
DTD model&F-NN&G-NN&G-NN\\
\hline
Training \& test&Set1&Set1&Set3\\
${\rm RMSD}_{\rm NN}$&$0.00072\pm0.00005$&$0.00089\pm0.00002$&$0.00092\pm0.00003$\\
${R^2}_{\rm NN}$&$0.99948\pm0.00011$&$0.99930\pm0.00005$&$0.99905\pm0.00010$\\
${\rm RMSD}_{\rm QCT}$&$0.00142\pm0.00004$&$0.00107\pm0.00003$&$0.00106\pm0.00004$\\
${R^2}_{\rm QCT}$&$0.99816\pm0.00014$&$0.99901\pm0.00005$&$0.99882\pm0.00011$\\
\hline
\hline
\end{tabular}
\caption{Performance measures (${\rm RMSD}_{\rm NN}$, ${R^2}_{\rm
NN}$, ${\rm RMSD}_{\rm QCT}$ and ${R^2}_{\rm QCT}$) for
statistical evaluation of F- (F-NN) and G-based models (G-NN)
trained and tested on Set 1/3.}
\label{sifig:statistical_eval}
\end{table}%
\section{Statistical Measures}
The RMSD$_{\rm QCT}$ and $R^2_{\rm QCT}$ values of all G-based models
and variants, applied to the different data sets are summarized in
Table \ref{sitab:perfomance_measures}.\\
\begin{table}[h!]
\centering
\begin{tabular}[t]{l|cccc}
\hline
\hline
DTD model&Training&Test&${\rm RMSD}_{\rm QCT}$&${R^2}_{\rm QCT}$\\
\hline
G-NN&Set1&Set1&0.0010&0.9991\\
\hline
G-NN&Set1&Set2&0.0016&0.9974\\
G-NN&Set2&Set2&0.0011&0.9990\\
G$'_{1}$-NN&Set2&Set2&0.0022&0.9878\\
G$'_{2}$-NN&Set2&Set2&0.0011&0.9990\\
G($\bm{T}$)-NN&Set2&Set2&0.0011&0.9990\\
G($\bm{\mu}$)-NN&Set2&Set2&0.0011&0.9990\\
\hline
G-NN&Set3&Set3&0.0011&0.9989\\
G$'_{2}$-NN&Set3&Set3&0.0013&0.9984\\
G$'_{3}$-NN&Set3&Set3&0.0011&0.9988\\
G($\bm{w},\bm{T}$)-NN&Set3&Set3&0.0011&0.9988\\
G($\bm{\mu}$)-NN&Set3&Set3&0.0020&0.9959\\
G($\bm{\mu},\bm{\sigma}$)-NN&Set3&Set3&0.0011&0.9987\\
\hline
G-NN&Set3&Set2&0.0012&0.9988\\
G$'_{3}$-NN&Set3&Set2&0.0012&0.9987\\
G($\bm{w},\bm{T}$)-NN&Set3&Set2&0.0125&0.8907\\
G($\bm{\mu},\bm{\sigma}$)-NN&Set3&Set2&0.0012&0.9988\\
\hline
G-NN&Set3&Set3A&0.0009&0.9991\\
G$'_{3}$-NN&Set3&Set3A&0.0009&0.9990\\
G($\bm{\mu},\bm{\sigma}$)-NN&Set3&Set3A&0.0011&0.9988\\
\hline
\hline
\end{tabular}
\caption{Performance measures (${\rm RMSD}_{\rm QCT}$ and ${R^2}_{\rm
QCT}$) of all G-based models and variants considered in this
work. Here, G$'_{i}$-NN denoted the G$'$-based model with $i$ grid
points per reactant state distribution. The column labelled training
denotes the data set on which the model was trained on, whereas the
test column specifies the data set whose test set was used to
calculate these performance measures. The number of significant
digits being reported is based on the findings of Section
\ref{si:stat_eval} in the SI.}
\label{sitab:perfomance_measures}
\end{table}%
\newpage
\section{G$'$-based models for Equilibrium Reactant State Distributions}
\label{si:limits_reduced_G_based_models}
G$'$-based models are variants of G-based models which use a
significantly reduced number of grid points per reactant state
distribution. Consider the case of a G$'$-based model, developed for
equilibrium reactant state distributions such as Set2.\\
\noindent
As these are equilibrium distributions they are characterized by a
corresponding temperature once the system-specific parameters are
fixed (here these are the energies of the rovibrational state of the
diatom). Thus, providing the value (``amplitude'') of the distribution
at a single, "suitably chosen" grid point uniquely determines the
corresponding distribution in the absence of noise and it is expected
that this suffices as input to accurately predict the corresponding
product state distributions. This is supported by the finding that
G($\bm{\mu}$)- and G($\bm{T}$)-based models yield accurate predictions
once ($\bm{\mu}$ and $\bm{T}$) are specified. However, it is important
that the grid point chosen does not correspond to a crossing point
between two equilibrium distributions at two different temperatures.\\
\noindent
In practice, however, the reactant state distributions considered in
this work suffer from noise due to finite sample statistics. In the
presence of noise the grid points should be placed at locations where
the difference between equilibrium distributions at different
temperatures is largest. Grid points located where this difference is
small, such as at the tail of these distributions, further raises the
difficulty of distinguishing between equilibrium distributions at
different temperatures when taking noise into account. Consequently,
the presence of noise serves as an explanation on why G$'$-based
models developed for equilibrium reactant state distributions of Set2
suffer from a significant drop in the prediction accuracy when the
number of grid points is reduced to a single point per distribution.\\
\newpage
\section{Additional Figures}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.99\textwidth]{NN_comparison_accurate_R2.png}
\caption{Product state distributions obtained by explicit QCT
simulations (QCT) as well as the corresponding references (-R) and
predictions (-NN) obtained in the (a-c) F-based (F-R, F-NN), (d-f)
K-based (K-R, K-NN) and (g-i) G-based approaches (G-R,
G-NN). Furthermore, the amplitudes to construct the reference
RKHS-based representations in the K- and G-based approaches are
displayed (circles). The data sets considered here are from the test
set of Set1 and result in predictions with the largest $R^2_{\rm
NN}$ value in the test set: (a-c) $\bm{T}=($17750 K, 12500 K,
12500 K), RMSD$_{\text{NN}}=0.0002$, $R^2_{\text{NN}}=0.99997$,
(d-f) $\bm{T}=($16750 K, 8000 K, 8000 K), RMSD$_{\text{NN}}=0.0006$,
$R^2_{\text{NN}}=0.9997$, (g-i) $\bm{T}=($9750 K, 5000 K, 5000 K),
RMSD$_{\text{NN}}=0.0005$, $R^2_{\text{NN}}=0.9998$.}
\label{sifig:NN_comparison_accurate}
\end{center}
\end{figure}
\newpage
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.99\textwidth]{NN_comparison_inaccurate_R2.png}
\caption{Product state distributions obtained by explicit QCT
simulations (QCT) as well as the corresponding references (-R) and
predictions (-NN) obtained in the (a-c) F-based (F-R, F-NN), (d-f)
K-based (K-R, K-NN) and (g-i) G-based approaches (G-R,
G-NN). Furthermore, the amplitudes to construct the reference
RKHS-based representations in the K- and G-based approaches are
displayed (circles). The data sets considered here are from the test
set of Set1 and result in predictions with the smallest $R^2_{\rm
NN}$ value in the test set: (a-c) $\bm{T}=($5000 K, 8000 K, 8000
K), RMSD$_{\text{NN}}=0.0020$, $R^2_{\text{NN}}=0.9973$, (d-f)
$\bm{T}=($6750 K, 18750 K, 18750 K), RMSD$_{\text{NN}}=0.0036$,
$R^2_{\text{NN}}=0.9887$, (g-i) $\bm{T}=($19000 K, 18750 K, 18750
K), RMSD$_{\text{NN}}=0.0015$, $R^2_{\text{NN}}=0.9984$.}
\label{sifig:NN_comparison_inaccurate}
\end{center}
\end{figure}
\newpage
\begin{figure}[h]
\includegraphics[width=0.65\textwidth]{Schwarzentruber.png}
\caption{$P(v')$ and $P(j')$ obtained by explicit QCT simulations
(QCT) as well as the corresponding fits to the parametric surprisal
model\cite{schwartz:2018} (Model). The data sets considered here are
from Set1: (a-b) $\bm{T}$ = (20000 K, 5000 K, 5000 K), (c-d)
$\bm{T}$ = (5500 K, 20000 K, 20000 K). While the model closely
matches the QCT data for (a-b), it is insufficiently described by
the model for (c-d), in particular $P(v')$.}
\label{sifig:Schwarzentruber}
\end{figure}
\newpage
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\textwidth]{alternative_function_based_model.png}
\caption{$P(v')$ obtained by explicit QCT simulations (QCT) as well as
the corresponding references obtained in the F-based approach using
Eq.~6 (F-R 1) and Eq.~12
(F-R 2). The data sets considered here are from Set1: (a) $\bm{T}$ =
(12500 K, 5750 K, 5750 K), (b) $\bm{T}$ = (9500 K, 16000 K, 16000
K), (c) $\bm{T}$ = (5750 K, 19250 K, 19250 K). These results
illustrate that Eq.~12 leads to a better
quality of fit compared to Eq.~6.}
\label{sifig:alternative_function_based_model}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\textwidth]{pv_kernel_coefficients.png}
\caption{(a) $P(v')$ obtained by explicit QCT simulations for
($T_{\text{trans}}=12000$ K, $T_{\text{rovib}}=5250$ K; black) and
($T_{\text{trans}}=12000$ K, $T_{\text{rovib}}=5500$ K; red) with
similar shape. (b-c) Corresponding featurizations using the K-based
approach. Here, the displayed features (kernel coefficients) are (b)
non-standardized and (c) standardized. These results illustrate that
positive and negative kernel coefficients can cancel. Hence,
different combinations of kernel coefficients are able to model
similarly shaped distributions, here $P(v')$.}
\label{sifig:pv_kernelcoeff}
\end{center}
\end{figure}
\newpage
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.99\textwidth]{results_set2_differentinputs.png}
\caption{Product state distributions obtained from explicit QCT
simulations (QCT), as well as the corresponding predictions of the
G- (G-NN), G$'_{2}$- (G$'_{2}$-NN), G($\bm{T}$)- (G($\bm{T}$)-NN)
and G($\bm{\mu}$)-based models (G($\bm{\mu}$)-NN). G$'_{2}$-NN uses
2 grid points per reactant state distribution (see main text). The
data sets considered here are from the test set of Set2: (a-c)
$\bm{T}$ = (20000 K, 7000 K, 5000 K), (d-f) $\bm{T}$ = (10000 K,
14000 K, 9000 K), (g-i) $\bm{T}$ = (5000 K, 20000 K, 8000 K).}
\label{sifig:set2_differentinputs}
\end{center}
\end{figure}
\newpage
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.6\textwidth]{LearningCurve.png}
\caption{Learning curve for the G-based model trained and evaluated on
Set3 with a variable size of the training and validation sets. Here,
both RMSD$_{\rm NN}$ and $1-R^2_{\rm NN}$ are measures for the NN
prediction accuracy. The NN prediction accuracy did not
significantly increase when $N_{\text{train}}+N_{\text{valid}}$ was
increased from 5000 to 30000 in increments of 5000.}
\label{sifig:probmodel_learningcurve}
\end{center}
\end{figure}
\begin{table}[h]
\centering
\begin{tabular}[t]{l|cc}
\hline
\hline
Panels&$w_{n}/w_{\rm tot}$&$\bm{T}$ [K]\\
\hline
(a-c)&2/5&$(11750, 6250, 6250)$\\
&2/5&$(7750,8250,8250)$\\
&1/5&$(16750, 19000, 19000)$\\
\hline
(d-f)&1/2&$(10000, 14750, 14750)$\\
&1/2&$(20000, 18000, 13000)$\\
\hline
(g-i)&2/3&$(19500, 17750,17750)$\\
&1/3&$(5000, 7000, 19000)$\\
\hline
\hline
\end{tabular}
\caption{Normalized weights $w_{n}/w_{\rm tot}$ and sets of
temperatures $\bm{T}=(T_{\rm trans}, T_{\rm vib}, T_{\rm rot})$
characterizing the data sets displayed in Figure~8.}
\label{sifig:probmodel_differentinputs_R2_table}
\end{table}%
\newpage
\begin{table}[h]
\centering
\begin{tabular}[t]{l|cc}
\hline
\hline
Panels&$w_{n}/w_{\rm tot}$&$\bm{T}$ [K]\\
\hline
(a-c)&39/77&$(15000,19750, 19750)$\\
&6/77&$(14250, 9500 ,9500)$\\
&32/77&$(10250, 10250, 10250)$\\
\hline
(d-f)&71/611&$(15000, 19750 ,19750)$\\
&82/611&$(15250 ,12250 ,12250)$\\
&84/611&$(14250 ,18500 ,18500)$\\
&16/611&$(9500,19750 ,19750)$\\
&24/611&$(10000, 14000 ,15000)$\\
&23/611&$(20000, 11000 ,16000)$\\
&95/611&$(9750 ,20000,20000)$\\
&89/611&$(19500 ,17750 ,17750)$\\
&51/611&$(14500 ,19500,19500)$\\
&76/611&$(10250, 10250,10250)$\\
\hline
(g-i)&18/77&$(17500, 7500, 7500)$\\
&32/77&$(7750 ,16000,16000)$\\
&5/77&$(10000, 18000, 8000)$\\
&22/77&$(18750 ,5000 ,5000)$\\
\hline
\hline
\end{tabular}
\caption{Normalized weights $w_{n}/w_{\rm tot}$ and sets of
temperatures $\bm{T}=(T_{\rm trans}, T_{\rm vib}, T_{\rm rot})$
characterizing the data sets displayed in Figure~9.}
\label{sifig:probmodel_diverse_differentinputs_R2_table}
\end{table}%
\clearpage
\section{Introduction}
The realistic description of chemical and reactive systems with a
large number of available states, such as explosions, hypersonic gas
flow around space vehicles upon re-entry into the atmosphere, or
meteorites penetrating deep into the lower parts of Earth's or a
planet's atmosphere, requires an understanding of the relevant
processes at a molecular
level.\cite{park:1993,cummings:2003,MM.hypersonics:2020} Correctly
describing the population of the available state space under such
non-equilibrium conditions (e.g. high temperatures in atmospheric
re-entry with $T > 10000$ K) from ground-based experiments is
extremely challenging. Such high gas temperatures make the gathering
of experimental data exceedingly difficult but is essential for
simulations of hypersonic flight.\cite{boyd:2015} On the other hand, a
comprehensive modeling of gas phase chemical reactions through
explicit molecular-level simulations remains computationally
challenging due to the large number of accessible states and
transitions between them.\cite{grover:2019,MM.nncs:2019} There are
also other, similar situations in physical chemistry such as the
spectroscopy in hot environments (e.g. on the sun) for which small
polyatomic molecules can populate a large number of rovibrational
states\cite{HITRAN} between which transitions can take
place. Exhaustively probing and enumerating all allowed transitions or
creating high-dimensional analytical representations for them is
usually not possible. Nevertheless, it is essential to have complete
line lists available because, if specific states that are involved in
important transitions are omitted, modeling of the spectroscopic bands
becomes difficult or even impossible.\cite{tennyson:2014} This points
towards an important requirement for such models, namely that they
contain the majority of the important information while remaining
sufficiently fast to evaluate.\\
\noindent
In such situations, machine learning approaches can provide an
alternative to address the problem of characterizing product
distributions from given reactant state distributions. In previous
work\cite{MM.nncs:2019}, a model for state-to-state (STS) cross
sections of an atom-diatom collision system using a neural network
(NN) has been proposed. Motivated by the success of such an approach,
the present work attempts to develop a NN-based
distribution-to-distribution (DTD) model for the relative
translational energy $E_{\rm trans}$, the vibrational $v$ and
rotational $j$ states of a reactive atom-diatom collision system. In
other words, given the reactant state distributions
($P(E_{\text{trans}}), P(v), P(j)$) such a model predicts the three
corresponding product state distributions ($P(E_{\text{trans}}'),
P(v'), P(j')$). Here, $P(v)$ and $P(j)$ are marginal distributions,
i.e. $P(v)=\sum_{j} P(v,j)$ and $P(j)=\sum_{v} P(v,j)$, where $(v,j)$
labels the rovibrational state of the diatom.\cite{schwartz:2018}
Hence, instead of considering all possible combinations of
($E_{\text{trans}}, v, j$) on the reactant (input) and product
(output) side explicitly, one is rather interested in a description of
these microscopic quantities by means of their underlying probability
distributions.\\
\noindent
While the state-to-state specificity is lost, such a probabilistic
approach considerably reduces the computational complexity of the
problem at hand. While a STS-approach is still feasible for an
atom-diatom collision system with $\sim 10^{7}$ STS cross
sections\cite{MM.nncs:2019}, it becomes
intractable\cite{schwartzentruber:2018} even for a diatom-diatom type
collision system due to the dramatic increase in the number of STS
cross sections to $\sim 10^{15}$. Moreover, such DTD models still
allow for the prediction of quantities relevant to hypersonic flow,
such as the reaction rates or the average vibrational and rotational
energies.\cite{singh2:2019}\\
\noindent
Here, the N + O$_{2}\rightarrow$ NO + O reaction is considered, which
is relevant in the hypersonic flight regime and for which accurate,
fully dimensional potential energy surfaces (PESs) are
available.\cite{MM.no2:2020} The necessary reference data to train the
NN-based models was obtained by running explicit quasi-classical
trajectories (QCT) for reactive N + O$_2$ collisions. In particular,
from a diverse set of equilibrium reactant state distributions
($P(E_{\text{trans}}), P(v), P(j)$) for N + O$_{2}$, the corresponding
product distributions for NO + O are obtained by means of QCT
simulations. In this work, three different approaches for learning and
characterizing these distributions are pursued, including function-,
kernel-, and grid-based models (F-, K-, and G-based models in the
following). The microscopic description provided by such DTD models
can, e.g., be used as an input or to develop models for more
coarse-grained approaches, including Direct Simulation Monte
Carlo\cite{dsmc} (DSMC) or computational fluid dynamics (CFD)
simulations. Furthermore, the core findings of this work also carry
over to applications in other areas where a DTD model is of interest,
such as in demographics\cite{flaxman:2016} or
economics\cite{perotti:1996}.\\
\noindent
This work is structured as follows. First, the methods including three
different approaches to construct NN-based DTD models are
described. Then, the performance of the models is assessed for various
data sets and improvements in particular related to the input features
are explored. Finally, implications for modeling hypersonic gas flow,
based on DSMC and CFD simulations, are discussed and conclusions are
drawn.\\
\section{Methods}
\subsection{Quasi-Classical Trajectory Calculations}
\label{QCT}
\noindent
Explicit QCT simulations for the N + O$_{2}$ collision system were
carried out on the $^4$A$'$ PES of NO$_2$ following previous
work.\cite{tru79,hen11,kon16:4731,MM.cno:2018,MM.no2:2020}
Specifically, the reactive channel for the N + O$_2$ $\rightarrow$ NO
+ O collision was considered. The $^4$A$'$ PES is chosen here, because
this state contributes most to the equilibrium rate.\cite{MM.no2:2020}
Briefly, Hamilton's equations of motion are solved in reactant Jacobi
coordinates using a fourth-order Runge-Kutta method with a time step
of $\Delta t = 0.05$~fs, which guarantees conservation of the total
energy and angular momentum. The initial conditions for each
trajectory were randomly chosen using standard Monte Carlo sampling
methods.\cite{tru79,hen11} The initial relative translational energies
$E_{\rm trans}$ were sampled from Maxwell-Boltzmann distributions
($E_{\rm trans,min} = 0.0$ eV; $E_{\rm trans,max} = 19.8$ eV;$\Delta
E_{\rm trans} = 0.1$ eV) and reactant vibrational $v$ and rotational
$j$ states were sampled from Boltzmann distributions ($v_{\rm min}=0,
v_{\rm max}=38, \Delta v = 1$; $j_{\rm min}=0, j_{\rm max}=242, \Delta
j = 1$), characterized by $T_{\text{trans}}$, $T_{\text{vib}}$ and
$T_{\text{rot}}$, respectively.\cite{tru79,bender:2015} The impact
parameter $b$ was sampled from 0 to $b_{\rm max} = 10$ \AA\/ using
stratified sampling\cite{tru79,bender:2015} with 6 equidistant strata.
The rovibrational reactant (O$_{2}$; $(v,j)$) and product diatom (NO;
$(v',j')$) states are calculated following semiclassical theory of
bound states.\cite{kar65:3259} The states of the product diatom are
assigned using histogram binning.\\
\noindent
First, models were constructed for the case
$T_{\text{rovib}}=T_{\text{rot}}=T_{\text{vib}}$, for which QCT
simulations were performed at $T_{\text{trans}}$ and
$T_{\text{rovib}}$ ranging from 5000 K to 20000 K in increments of 250
K. This yielded 3698 sets of reactant state distributions and
corresponding product state distributions which will be referred to as
``Set1''. Next, for the more general case $T_{\text{rot}} \neq
T_{\text{vib}}$, further QCT simulations were performed for
$T_{\text{trans}} = 5000, 10000, 15000, 20000$ K with $T_{\text{vib}}$
and $T_{\text{rot}}$ each ranging from 5000 K to 20000 K in increments
of 1000 K. This gives an additional 960 data sets and a total of 4658
data sets that include both cases, $T_{\text{rot}} = T_{\text{vib}}$
and $T_{\text{rot}} \neq T_{\text{vib}}$, collectively referred to as
``Set2''.\\
\noindent
The reactant and product state distributions of Set1 and Set2
constitute representative reference data to train and validate
NN-based models. For both sets the temperatures $\bm{T}=
(T_{\text{trans}}$, $T_{\text{vib}}$, $T_{\text{rot}}$) completely
specify the reactant and product state distributions as they are
related through the explicit QCT simulations. Hence, for brevity a
specific set of reactant and product state distributions is referred
to as $\bm{T}$.\\
\subsection{Generating Nonequilibrium Data Sets}
\label{Generating_noneq_Datasets}
In hypersonic applications it is known that quantities such as
$P(E_{\text{trans}})$, $P(v)$, or $P(j)$ are typically nonequilibrium
probability distributions.\cite{boyd:2015} This has also been
confirmed in explicit simulation studies, starting from equilibrium
energy and state distributions as is commonly done in QCT
simulations.\cite{alp17:2392,MM.cno:2018} Therefore, a general DTD model
should be able to correctly predict (nonequilibrium) product state
distributions starting from nonequilibrium reactant state
distributions. With this in mind, and with Set2 at hand, new reactant
and product state distributions were generated by means of a weighted
sum of the existing distributions according to
\begin{equation}
P(i)=\frac{1}{w_{\text{tot}}}\sum_{n=1}^N w_n \cdot P_{n}(i).
\label{eq:multit}
\end{equation}
Here, $i \in [E_{\text{trans}},v,j,E_{\text{trans}}',v',j']$ labels
the degree of freedom, $n$ labels the data set, $N \in [2,3]$ is the
total number of distributions drawn from Set2, the corresponding
distributions $P_{n}(i)$ used for and obtained from QCT simulations
and the random weights $w_{n} \in [1,2]$ determine how much these
contribute to the total sum. The resulting distributions are scaled by
$w_{\text{tot}}=\sum_{n=1}^N w_n$ to conserve probability. Such
distributions then constitute Set3. It is assumed that any
nonequilibrium state distribution can be represented as a
decomposition in terms of a linear combination of equilibrium
distributions given by Eq.~\ref{eq:multit}. For instance, general
nonequilibrium distributions for nitrogen and oxygen relaxation at
high temperature ($> 8000$ K) conditions, obtained via direct
molecular simulations (DMS), which is equivalent to solving the full
master equation, have been successfully modelled as a weighted sum of
two Boltzmann distributions.\cite{singh3:2019} Consequently, DTD
models that are successfully trained and validated on Set3 are also
expected to generalize well to most nonequilibrium situations
encountered in practice.\\
\noindent
In the following, a single data set for Set3 is generated by randomly
specifying the number of distributions $N \in [2,3]$ to be combined
although larger values for $N$ are possible and will be explored
later. The final set of reactant and product state distributions is
characterized by $N$ sets of temperatures $\bm{T}$ and corresponding
normalized weights $\bm{w}=(w_{1}/w_{\rm tot},...,w_{N}/w_{\rm
tot})$. The product state distributions obtained by this procedure
are akin to explicit QCT simulations using Monte Carlo sampling of the
reactant state distributions by sampling each of the equilibrium
distributions in the corresponding weighted sum. In the following,
three different possibilities for characterizing reactant and product
state distributions are described.\\
\subsection{Function (F)-Based Approach}
In the F-based approach, each set of relative translational energy,
vibrational and rotational state distributions of reactant and product
are fitted to parametrized model functions, see Figure \ref{fig:fig1}
for an example. The corresponding fitting parameters in
Eqs. \ref{eq:func_E_r} to \ref{eq:func_j_p} constitute the input and
output of a NN, respectively (see Section \ref{Neural Network} for
details on the NN). Together with the parametrized model functions
(Eqs. \ref{eq:func_E_r} to \ref{eq:func_j_p}) this serves as a map
between reactant and product state distributions, i.e., a DTD
model. In this work the F-based approach was only applied to Set1.\\
\noindent
The set of model functions used here was
\begin{equation}
\tilde{P}(E_{\text{trans}}) = a_{1} E_{\text{trans}}\cdot
\exp(-E_{\text{trans}}/a_{2}),
\label{eq:func_E_r}
\end{equation}
\begin{equation}
\tilde{P}(v) = b_{1}\exp(-v/b_{2}),
\label{eq:func_v_r}
\end{equation}
\begin{equation}
\tilde{P}(j) = c_{1}\exp(-j/c_{2}) + c_{3}\exp(-(\ln(2c_{4}(j-c_{5})/c_{6}+1)/c_{4})^2),
\label{eq:func_j_r}
\end{equation}
\begin{equation}
\tilde{P}(E_{\text{trans}}') =
d_{1}\exp(-(\ln(2d_{2}(E_{\text{trans}}'-d_{3})/d_{4}+1)/d_{2})^2),
\label{eq:func_E_p}
\end{equation}
\begin{equation}
\tilde{P}(v') = e_{1}v'^4+e_{2}v'^3+e_{3}v'^2+e_{4}v'+e_{5},
\label{eq:func_v_p}
\end{equation}
\begin{equation}
\tilde{P}(j') = f_{1}\exp(-j'/f_{2}) + f_{3}\exp(-(\ln(2f_{4}(j'-f_{5})/f_{6}+1)/f_{4})^2),
\label{eq:func_j_p}
\end{equation}
where $\bm{a}=(a_{1},a_{2})$ through $\bm{f}=(f_{1},...,f_{6})$ are
the fitting parameters of the model functions. In total, this results
in 10 and 15 fitting parameters for one set of reactant or product state
distribution, respectively. For the reactant and product rotational
state distributions, $P(j)$ and $P(j')$, the same model function was
used. Such a parametric approach has its foundation in surprisal
analysis\cite{levine:1978} which was recently used in models for
hypersonics.\cite{schwartz:2018,singh1:2019,singh2:2019}\\
\noindent
The reactant state distributions in Set1 are equilibrium distributions
and the model functions (Eqs. \ref{eq:func_E_r} to \ref{eq:func_j_r})
were chosen accordingly. For the product state distributions, which
are typically nonequilibrium distributions, modified parametrizations
were used after inspection of the results from the QCT simulations.
Here, it is worth mentioning that using alternative parametrizations
for model functions are possible, too, which will be briefly explored
later.\\
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\textwidth]{methods_function_based.png}
\caption{Reactant state distributions for the F-based approach: $P(E_{\rm
trans})$ ($T_{\text{trans}} = 12500$ K, panel (a), $P(v)$ (panel
(b)) and $P(j)$ (panel (c)) distributions ($T_{\text{rovib}} = 5750$
K) for explicit QCT simulations (black) and corresponding fits (red)
obtained using Eqs. \ref{eq:func_E_r} to \ref{eq:func_j_r}.}
\label{fig:fig1}
\end{center}
\end{figure}
\subsection{Kernel (K)-Based Approach}
\label{K-based method}
The representer theorem\cite{scholkopf2001generalized} states that,
given $N$ grid points $x_i$, the function $f(x)$ can always be
approximated as a linear combination of suitable functions
\begin{equation}
f(x) \approx \widetilde{f}(x) = \sum_{i = 1}^{N} c_i K(x,x_i)
\label{RKHS}
\end{equation}
where $c_i$ are coefficients and $K(x,x_i)$ is a kernel function. The
reproducing property asserts that $f(x') = \langle f(x),K(x,x')
\rangle$ where $\langle \cdot \rangle$ is the scalar product,
$K(x,x')$ is the kernel\cite{aronszajn:1950} and the coefficients
$c_i$ are determined through matrix inversion. This leads to a
reproducing kernel Hilbert space (RKHS) representation that exactly
reproduces the function at the grid points
$x_i$.\cite{aronszajn:1950,ho96:2584,MM.RKHS:2017} In the present work
the amplitudes of the distributions at the chosen grid points are used
for inter- and extrapolation based on a RKHS-based representation and
the the coefficients $c_{i}$ serve as input and output of the
NN. Hence, given the kernel coefficients for the reactant state
distributions, the NN is trained to predict the coefficients of the
corresponding product state distributions. Together with the
associated grids, one obtains a continuous (K+NN-based) prediction of
the product state distributions, i.e., a DTD model. The K-based
approach was also only applied to Set1.\\
\noindent
In this work, a Gaussian kernel
\begin{equation}\label{gaussian_ker}
K(x,x')=\exp(-|x-x'|^2/2\sigma^2),
\end{equation}
with hyperparameter $\sigma$ was found to perform well as a
reproducing kernel. Furthermore, a variable amount of regularization
as specified by the regularization rate $\lambda$ in the
Tikhonov-scheme is used. The hyperparameters $\sigma$ and $\lambda$
remain to be optimized systematically. However, the present choices
yielded sufficiently accurate representations. Assigning $\sigma$ to
the average spacing between neighbouring grid points was found to be
advantageous. Alternatively, it is also possible to choose $\sigma$ at
each grid point to be equal to the larger of the two neighbouring grid
spacings. In this work, the first approach is used for $P(v)$ and
$P(v')$, whereas the second approach is applied for all other
distributions. Moreover, such choices for $\sigma$ significantly
reduced the number of grid points required for accurate RKHS
approximations and the resulting kernel coefficients lead to accurate
NN predictions. While the same level of accuracy for the RKHS
approximations can be achieved with larger values for $\sigma$, the
accuracy of the resulting NN predictions was found to
deteriorate. Consequently, only the regularization rate $\lambda$
needed to be tuned and accurate RKHS approximations were obtained
after a few iterations. \\
\begin{figure}[b!]
\begin{center}
\includegraphics[width=0.99\textwidth]{methods_kernel_based.png}
\caption{Reactant state distributions for K- and G-based approaches: $P(E_{\rm
trans})$ ($T_{\text{trans}} = 9500$ K, panel (a), $P(v)$ (panel
(b)) and $P(j)$ (panel (c)) distributions ($T_{\text{rovib}} =
16000$ K) for explicit QCT simulations (black), their RKHS
representations (red lines) and the locally averaged values at the
corresponding grid points (red circles) used for the G-based approach.}
\label{fig:fig2}
\end{center}
\end{figure}
\noindent
The location and number of grid points for the reactant and product
state distributions is largely arbitrary but should be governed by the
overall shape of the distributions, see Figure \ref{fig:fig2} for an
example. The grids used here are reported in Table~S1. The number of grid points for reactant and
product state distributions differs because they are equilibrium and
nonequilibrium distributions, respectively. Also, depending on the
shape of the distributions to be represented, additional points may be
required to avoid unphysical undulations in the RKHS
approximations. For the system considered here, this is mainly
observed for $P(v')$ which requires a denser grid than the
corresponding reactant state distributions $P(v)$.\\
\noindent
Instead of directly evaluating the distributions at the grid points,
local averaging over neighboring data points according to
\begin{equation}
\bar{P}(x_{i}) = \frac{1}{2n+1}\sum\limits_{j=i-n}^{i+n} P(x_{j}),
\label{eq:local_averaging}
\end{equation}
was performed to obtain $\bar{P}(x_{i})$. Here $n = {\rm min}(n_{\rm
max},n_{\rm nb})$, with $n_{\rm nb}$ the maximum number of
neighbouring data points to the left or the right. If the first and
last data points are chosen as grid points they were assigned the
unaveraged distribution values. The value of $n_{\text{max}}$ can
differ for each of the
($E_{\text{trans}},v,j,E_{\text{trans}}',v',j'$) distributions. For
the K-based approach these values were $n_{\text{max},E_{\rm trans}} =
2$, $n_{\text{max},v} = 1$, $n_{\text{max},j} = 12$ for the reactant
and $n_{\text{max},E_{\rm trans}'} = 3$, $n_{\text{max},v'} = 2$, and
$n_{\text{max},j'} = 13$ for the product state distributions. Local
averaging can be seen as an implicit regularization as it reduces the
noise partially arising due to finite sample statistics in the QCT
simulations.\\
\subsection{Grid (G)-Based Approach}
For the G-based approach, the same grids as in the K-based approach
are considered (see Table~S1). Furthermore,
similar to the K-based approach local averaging is performed with
$n_{\text{max},E_{\rm trans}} = 2$, $n_{\text{max,}v} = 1$,
$n_{\text{max,}j} = 9$ for the reactant and $n_{\text{max},E_{\rm
trans}'} = 3$, $n_{\text{max,}v'} = 2$, and $n_{\text{max,}j'} =
10$ for the product state distributions. These values were adjusted
such as to obtain accurate discrete representations of the
corresponding distributions. In the G-based approach the locally
averaged values of reactant state distributions at the grid points
(referred to as ``amplitudes'') directly serve as the input of a NN,
see the red circles in Figure \ref{fig:fig2}. The NN then predicts the
product state distributions on the corresponding grids, where the
amplitudes serve as the reference. The resulting discrete product
distributions are finally represented as a continuous RKHS,
establishing a DTD model. Similarly to the K-based approach, a
Gaussian kernel (Eq. \ref{gaussian_ker}) was used for the RKHS of the
product state distributions, as this choice still yielded accurate
approximations. Furthermore, the corresponding hyperparameters
$\sigma$ and $\lambda$ are were chosen as in the K-based approach.\\
\subsection{Neural Network}
\label{Neural Network}
The NN architecture for training the three models is a multilayer
perceptron with two hidden layers, see Figure \ref{fig:fig4}. The
input and output layers consist of 10/43/43 input and 15/44/44 output
nodes in the F-, K-, and G-based approaches. The input/output are the
fitting parameters (F-based approach), kernel coefficients (K-based
approach), and amplitudes (G-based approach) characterizing reactant
and product state distributions, respectively. When training a NN
using Set1 to Set3 the two hidden layers contain 6, 12, and 9 nodes
each, respectively.\\
\begin{figure}[h!]
\includegraphics[width=0.35\textwidth]{NN_architecture.png}
\caption{Schematic diagram of the NN architecture. The activation
vector of each hidden layer is \textbf{a$_{i}$}. The input and output
vectors are $\mathbf{x}$ and $\mathbf{y}$ and the weight matrix and
bias vector for each layer are $\mathbf{W}_{i}$ and
$\mathbf{b}_{i}$, respectively. The activation function of the
hidden layers is $\sigma(\cdot)$ and corresponds to a shifted
softplus\cite{dugas:2001} function $\sigma(\cdot) =
\ln(1+e^{(\cdot)})-\ln(2)$. Here, activation functions act
element-wise on input vectors $(\cdot)$. The constant $\ln(2)$
centers the mean activation of the hidden nodes at zero, thereby
decreasing the bias shift effect which speeds up the
learning.\cite{clevert:2015, lecun:2012} The activation function of
the output layer is $C \cdot \tanh(\cdot)$\cite{glorot:2011}, where
$C$ is an overall scaling with $C=8/12/8$ in the F-, K- and G-based
approaches.}
\label{fig:fig4}
\end{figure}
\noindent
For training, the input and output of the NN are standardized
according to $x_{i}' = (x_{i} - \bar{x}_{i})/\sigma_{i}$, where
$x_{i}$, $\bar{x}_{i}$ and $\sigma_{i}$ are the $i$-th input/output, and
the mean and standard deviation of their distribution over the entire set of training data. Scaling of the
input, here by means of standardization, is common practice in the
data pre-processing step for machine learning tasks relying on
gradient descent for optimization, as it generally yields a faster
convergence rate.\cite{lecun:2012} The additional standardization of
the output allows one to use a root-mean-square deviation (RMSD) loss
function, as the non-standardized values can differ drastically in
magnitude and spread. Thus, in particular low and high product state
probabilities can be predicted with similar accuracy. It would also be
possible to simply normalize the output but the additional offset
gives the flexibility to use a scaled hyperbolic tangent as
an activation function for the output layer which increases the NN
prediction accuracy compared to other/no activation functions. The
RMSD loss function $\mathcal{L}$ used here is
\begin{equation}
\label{loss_func}
\mathcal{L} = \sqrt{\frac{1}{N}\sum\limits_{i=1}^{N} \left( y_{i} -
y'_{i}\right)^2},
\end{equation}
with $y_{i}$ and $y'_{i}$ the predicted and reference output values.\\
\noindent
The weights and biases of the NN are initialized according to the
Glorot scheme\cite{glorot2010understanding} and optimized using
Adam\cite{kingma2014adam} with an exponentially decaying learning
rate. The NN is trained using TensorFlow \cite{tf:2016} and the set of
weights and biases resulting in the lowest loss as evaluated on the
validation set are used for predictions. When training a NN using Set1
with $N_{\text{tot}}=3698$ data sets, $N_{\text{train}}=3000$ were
randomly selected for training, $N_{\text{valid}}=600$ for validation
and $N_{\text{test}}=98$ as a test set, whereas for Set2
$N_{\text{tot}}=4658$, $N_{\text{train}}=3600$,
$N_{\text{valid}}=900$, and $N_{\text{test}}=158$.\\
\noindent
To train models on reactant state distributions that can {\it not} be
characterized as a single set of temperatures $\bm{T}$, Set3 was
constructed by means of a weighted sum of the distributions in Set2
(see Section \ref{Generating_noneq_Datasets}). For this, 158 data sets
are randomly selected from $N_{\text{tot}}=4658$ data sets of
Set2. They constitute the subset from which the final test set of Set3
is generated. Here, $N_{\text{test}}=125$, see Section
\ref{Generating_noneq_Datasets}. The remaining 4500 data sets make up
the subset from which the data sets for training and validation are
constructed. In particular, $N_{\text{train}}+N_{\text{valid}}= 5000,
10000,15000, 20000, 25000, 30000$ data sets are generated by means of
the same procedure, making up the final training and validation sets
of Set3 with $N_{\text{train}} = 0.8 \times
(N_{\text{train}}+N_{\text{valid}})$ and $N_{\text{valid}} = 0.2
\times (N_{\text{train}}+N_{\text{valid}})$. All NNs in this work were
trained on a 1.8 GHz Intel Core i7-10510U CPU with training times
shorter than 10 minutes.\\
\section{Results}
The results section first presents DTD models for the F-, K-, and
G-based approaches for $T_{\text{vib}} = T_{\text{rot}}$. This is
followed by discussing the influence of featurization, computational
cost and generalizability of the approaches considered. As the G-based
approach is found to perform best from a number of different
perspectives, models are then trained for $T_{\text{vib}} \neq
T_{\text{rot}}$ and for nonequilibrium reactant state distributions. Also,
variations of the G-based approach requiring fewer input data are
explored. Then, the findings are discussed in a broader context and
conclusions are drawn.\\
\subsection{Distribution-to-Distribution Models for $T_{\text{vib}} = T_{\text{rot}}$}
\label{Dist_to_Dist}
First, an overall assessment of the three different approaches for
describing the reactant state distributions for Set1
(i.e. $T_{\text{vib}} = T_{\text{rot}}$) is provided. As they are
generated according to the typical sampling procedures followed by QCT
simulations and not from direct function evaluations of equilibrium
distributions they contain noise. This is done because the entire work
is concerned with a situation typically encountered in QCT simulations
of reactive processes.\cite{kar65:3259} Also, it is noted that for the
reactant state distributions one should find - as demonstrated here -
that they are characterized by one parameter only, namely
temperature. This, however, is an open point and not guaranteed for
the product state distributions, which are nonequilibrium
distributions in general.\\
\noindent
For the F- and K-based approaches the reactant state distributions are
first represented either as a parametrized model function or as a
RKHS, respectively, and the agreement is found to be excellent (see
Figures \ref{fig:fig1} and \ref{fig:fig2}). For the G-based approach
this step is not required, as the input are the amplitudes of the
reactant state distributions themselves (see Figure \ref{fig:fig2}).\\
\noindent
Having established that all three (F-, K-, and G-based) approaches are
suitable to describe reactant state distributions, a NN for each
of the three models was trained on Set1 with $N_{\text{train}}=3000$
and $N_{\text{valid}}=600$. The quality of the final model for
predicting product state distributions depends on two aspects: 1. The
ability of the NN to {\it learn} and predict the product state
distributions obtained from the QCT simulations 2. The ability of the
(F-, K- or G-based) approaches to {\it describe} these
distributions.\\
\noindent
{\it 1. Quality of the NN Prediction:} For the first aspect, RMSD and
coefficient of determination ($R^2$) values, referred to as RMSD$_{\rm
NN}$ and $R^2_{\rm NN}$, are considered as performance measures. For
a single data set from the test set these are calculated by comparing
the normalized reference representations of each of the models and the
corresponding normalized NN predictions on the grid for which QCT data
is available for each of the three product state distributions
separately and averaging over the resulting values. The normalization
factors were calculated by numerical integration of the distributions
obtained from the QCT simulations. The final RMSD and $R^2$ values are
then obtained by averaging over the entire test set with
$N_{\text{test}}=98$, see Table \ref{tab:Set1_perfomance_measures}.\\
\begin{table}[ht]
\centering
\begin{tabular}[t]{l|cc|cc}
\hline
\hline
DTD model&${\rm RMSD}_{\rm NN}$&${R^2}_{\rm NN}$&${\rm RMSD}_{\rm QCT}$&${R^2}_{\rm QCT}$\\
\hline
F-NN&0.0007&0.9996&0.0014&0.9982\\
K-NN&0.0013&0.9984&0.0014&0.9981\\
G-NN&0.0009&0.9994&0.0010&0.9991\\
\hline
\hline
\end{tabular}
\caption{Performance measures of the F-, K- and G-based models (F-NN,
K-NN and G-NN) trained and evaluated on Set1. The number of
significant digits being reported is based on the findings of Table~S2.}
\label{tab:Set1_perfomance_measures}
\end{table}%
\noindent
For three different data sets from the test set, the results from
explicit QCT simulations, the NN predictions, and the reference
representation of the corresponding approach are shown in Figure
\ref{fig:fig5}. These results are representative of NN predictions for
data from the test set for each of the three approaches as they are
characterized by an $R^2_{\rm NN}$ value closest to the mean $R^2_{\rm
NN}$ value as evaluated over the test set. Figures
S1 and
S2 show the predictions that are
characterized by the highest (``accurate'' prediction) and lowest
(``inaccurate'' prediction) $R^2_{\rm NN}$ value in the test set,
respectively.\\
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\textwidth]{NN_comparison_representative_R2.png}
\caption{Product state distributions obtained by explicit QCT
simulations (QCT) as well as the corresponding references (-R) and
predictions (-NN) from the (a-c) F-based (F-R, F-NN), (d-f) K-based
(K-R, K-NN) and (g-i) G-based approaches (G-R, G-NN). Also, the
amplitudes to construct the reference RKHS-based representations in
the K- and G-based approaches are displayed (circles). The data sets
considered here are from the test set of Set1 and result in
predictions that are characterized by an $R^2_{\rm NN}$ value
closest to the mean $R^2_{\rm NN}$ value as evaluated over the test
set: (a-c) $\bm{T}=($9500 K, 16000 K, 16000 K),
RMSD$_{\text{NN}}=0.0005$, $R^2_{\text{NN}}=0.9996$, (d-f)
$\bm{T}=($10250 K, 19250 K, 19250 K), RMSD$_{\text{NN}}=0.0013$,
$R^2_{\text{NN}}=0.9984$, (g-i) $\bm{T}=($12000 K, 9750 K, 9750 K),
RMSD$_{\text{NN}}=0.0009$, $R^2_{\text{NN}}=0.9994$.}
\label{fig:fig5}
\end{center}
\end{figure}
\noindent
In general, all three approaches are very accurate,
as their predictions closely match the corresponding reference
representations. Closer inspection of Figure \ref{fig:fig5} reveals
that for these particular examples the K-based approach appears to
yield slightly less accurate predictions than the other two
approaches (see, for example Figure \ref{fig:fig5}f). The RMSD$_{\rm
NN}$ values are 0.0007, 0.0013 and 0.0009 for the F-, K-, and
G-based models respectively, and the corresponding $R^2_{\rm NN}$
values are 0.9996, 0.9984 and 0.9994. These performance measures
indicate that the F-based approach yields the most accurate
predictions, followed by the G- and K-based approaches.\\
\noindent
{\it 2. Quality of the F-, K-, and G-Based Model Predictions:} For
the product state distributions the prediction accuracy depends on the
accuracy of the NN and the accuracy with which the representations
approximate them. Figure \ref{fig:fig6} compares the final model
predictions from the three approaches. The examples illustrate the
variety of product state distributions in Set1. It is found that
despite the appreciable variation in shapes (in particular for
$P(v')$) all three models correctly describe the product state
distributions. Distributions $P(v')$ are not well represented as a
single equilibrium distribution which is typical for vibrational
states at high temperatures.\cite{boyd:2015,schwartz:2018}\\
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\textwidth]{final_model_comparison.png}
\caption{Product state distributions obtained by explicit QCT
simulations (QCT) as well as the corresponding model predictions
obtained in the (a-c) F-based (F-NN), (d-f) K-based (K-NN) and (g-i)
G-based approaches (G-NN). The data sets considered here are from
the test set of Set1: (a-c) $\bm{T}=$ (12500 K, 5750 K, 5750 K) ,
(d-f) $\bm{T}=$ (9500 K, 16000 K, 16000 K), (g-i) $\bm{T}=$ (5750 K,
19250~K, 19250 K).}
\label{fig:fig6}
\end{center}
\end{figure}
\noindent
A quantitative measure for the performance of the three models are the
RMSD and $R^2$ values calculated by comparing the locally averaged
($n_{\text{max},E_{\rm trans}'} = 3$, $n_{\text{max},v'} = 2$,
$n_{\text{max},j'} = 10$) QCT data and the model predictions following
a similar procedure as for RMSD$_{\rm NN}$ and $R^2_{\rm NN}$ which
will be called RMSD$_{\rm QCT}$ and $R^2_{\rm QCT}$, respectively. For
Set1 these performances are reported in Table
\ref{tab:Set1_perfomance_measures}. Again, all models are of high
quality with the G-based approach performing best. The somewhat lower
quality of the F-based approach when compared to the G-based approach
can largely be attributed to the fits of the product state
distributions. The representation of the F-based approach leads to
differences, in particular for $P(v')$ (e.g., deviations for small and
high $v'$ or extra undulations in Figure \ref{fig:fig6}b). However,
the deviations in the F-based approach observed for high $v'$ are only
partially relevant, as the accessible vibrational and rotational state
space is finite in practice, here $v'_{\rm max}=47$, $j'_{\rm
max}=240$. Since state space is limited, extrapolation is not always
required. Considering the K- and G-based approaches, the reference
representations describing product distributions are nearly identical
and reproduce the QCT data very closely. Hence, the lower accuracy of
the final model in the K-based approach when compared to the G-based
approach can largely be attributed to its lower NN prediction
accuracy. \\
\noindent
For the F-based approach, finding an optimal set of model functions
(Eqs. \ref{eq:func_E_r} to \ref{eq:func_j_p}) specific to the system
at hand is expected to be a difficult task. Such parametric models for
nonequilibrium conditions are still a current topic of
research.\cite{schwartz:2018, singh1:2019, singh2:2019} To highlight
the performance of different models, a parametric model for transient
vibrational and rotational state distributions based on surprisal
analysis\cite{schwartz:2018} was applied to Set1. $P(v')$ and $P(j')$
distributions for two different sets of temperatures
(($T_{\text{trans}} = 20000$ K, $T_{\text{rovib}} = 5000$ K) and
($T_{\text{trans}} = 5500$ K, $T_{\text{rovib}} = 20000$ K)) from QCT
simulation are modelled following the parametrization of
Ref. \citen{schwartz:2018} (see Figure~S3). While for the first set of temperatures
(translationally hot and rovibrationally moderately hot) the QCT
results for $P(v')$ and $P(j')$ are closely matched by the model, for
the second set (translationally moderately hot and rovibrationally
hot) both distributions are insufficiently described by the model, in
particular for $P(v')$ which is consistent with
Ref. \citen{schwartz:2018}. The fact that the shape of the $P(v')$
appears to vary more widely for different $\bm{T}$ compared to
$P(E_{\rm trans}')$ and $P(j')$ could be a partial explanation of why
developing a universally valid parametric model for $P(v')$ is more
challenging. It should be emphasised that comparison between
predictions based on the model and explicit QCT simulations is
mandatory to validate the model function used.\\
\subsection{Sensitivity of Performance to Feature Selection}
As in all machine learning tasks, feature selection for representing
the raw data is crucial for the complexity and prediction accuracy of
the resulting NN-based model.\cite{goodfellow:2016, bengio:2012} Here,
the main difference between the three approaches are the features that
represent reactant and product state distributions and which serve as
input/output of the NNs. Hence, any difference in the NN prediction
accuracy is due to the features used (i.e. the
``featurization''). Here, the features are fitting parameters (F-based),
kernel coefficients (K-based), and amplitudes (G-based) and together they
constitute a feature vector. Hence, a good featurization allowing for
an accurate NN to be trained is characterized by the fact that
similarly shaped distributions are described by similar feature
vectors.\cite{faber:2015} Here, ``similarity'' is measured by an
appropriate metric, such as an Euclidean norm.\\
\noindent
For the F-based approach, the choice of model functions (see
Eqs. \ref{eq:func_E_r} to \ref{eq:func_j_p}) turned out to
yield a satisfactory featurization. Conversely, for the K-based
approach it was necessary to increase the regularization rate
$\lambda$ and averaging over more neighbouring data points which in
essence smooths out sharp variations in the kernel coefficients
between neighboring data sets in temperature space. In the G-based
approach, accurate NN predictions were obtained through local
averaging because the amplitudes are the features.\\
\noindent
For the F-based approach the dependence of NN performance on feature
selection was explicitly explored by choosing an alternative
parametrization for
\begin{equation}
\tilde{P}(v') = g_{1}\exp(-(\ln(2g_{2}(v'-g_{3})/g_{4}+1)/g_{2})^2) +
g_{5}\exp(-(\ln(2g_{6}(v'-g_{7})/g_{8}+1)/g_{6})^2),
\label{eq:func_v_p_alternative}
\end{equation}
where $\bm{g}=(g_{1},...,g_{8})$ are the corresponding fitting
parameters. The resulting fit (see Figure~S4) to the QCT data
demonstrates that Eq. \ref{eq:func_v_p_alternative} yields a better
fit than Eq. \ref{eq:func_v_p}. However, training the corresponding NN
turned out to be difficult and the resulting NN predictions were
highly inaccurate (see below).\\
\noindent
Figure \ref{fig:pv_features} illustrates these points for $P(v')$ for
three combinations of simulation temperatures with 1)
$T_{\text{trans}} \sim T_{\text{rovib}}$ ($T_{\text{trans}} = 5000$ K,
$T_{\text{rovib}} = 5000,5250,5500,5750$ K; black) 2)
$T_{\text{trans}} < T_{\text{rovib}}$ ($T_{\text{trans}} = 5000$ K,
$T_{\text{rovib}} = 10000,10250,10500,10750$ K; red), and 3)
$T_{\text{trans}} > T_{\text{rovib}}$ ($T_{\text{trans}} = 12000$ K,
$T_{\text{rovib}} = 5000,5250,5500,5750$ K; green). As Figure
\ref{fig:pv_features}a demonstrates, the shapes of all $P(v')$ are
comparable and for each color there are four largely overlapping
distributions which can not be separated because the differences in
$T_{\text{rovib}}$ are too small. Using the F-based approach with
Eq. \ref{eq:func_v_p} for $P(v')$ yields parameter values that are
clustered (Figure \ref{fig:pv_features}b) for the black, red, and
green $P(v')$, respectively. For such input a robust NN can be
trained. Conversely, using Eq. \ref{eq:func_v_p_alternative}, even the
fitting parameters for one set of $P(v')$ spread considerably
and mix with those from $P(v')$ of other temperature combinations, see
Figure \ref{fig:pv_features}c. Thus, similarity in the shape of
$P(v')$ does not translate into similarity of the fitting parameters
used for the featurization. This is an unfavourable situation for
training a NN which compromises the prediction ability of such an
F-based model. The F-based model trained with Eq. \ref{eq:func_v_p}
yields an accurate prediction (${\rm RMSD}_{\rm NN}=0.0007, R^2_{\rm
NN} = 0.9996$) whereas the one trained with
Eq. \ref{eq:func_v_p_alternative} fails to predict $P(v')$ (RMSD$_{\rm
NN}=0.0348$, $R^2_{\rm NN}=-10.6133$; i.e., this model is worse than
a baseline model with $R^2 =0$). Similarly, a K-based approach can lead
to considerable spread of the kernel coefficients (Figure
\ref{fig:pv_features}d) which is not observed for the amplitudes in
the G-based approach (Figure \ref{fig:pv_features}e). Such differences
in the featurization leads to differences in the NN prediction
accuracies.\\
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\textwidth]{pv_representations.png}
\caption{Comparison of the featurization for three different groups of
similarly shaped $P(v')$ (black, red and green). Panel a: the
distributions as obtained from QCT simulations (fluctuating
lines). Panels b and c: Fitting Parameters for the F-based approach
with model functions from Eqs. \ref{eq:func_v_p} and
\ref{eq:func_v_p_alternative}; panel d: kernel coefficients for the
K-based approach and panel e: amplitudes for G-based approach. The
temperatures are ($T_{\text{trans}} = 5000$ K, $T_{\text{rovib}} =
5000,5250,5500,5750$ K; black), ($T_{\text{trans}} = 5000$ K,
$T_{\text{rovib}} = 10000,10250,10500,10750$ K; red), and
($T_{\text{trans}} = 12000$ K, $T_{\text{rovib}} =
5000,5250,5500,5750$ K; green). The quality of all fits in panel a
is as good as in Figure \ref{fig:fig6} and all features are
standardized.}
\label{fig:pv_features}
\end{center}
\end{figure}
\noindent
Another difference between the NNs in the three approaches is the fact
that prediction errors in the features translate into errors in the
corresponding predicted product state distributions in different
ways. For the G-based approach, an error in the predicted features
directly translates into an error in the predicted product state
distributions. This is not the case for the F- or K-based
approaches. As an example, the model functions for the product state
distributions in an F-based approach depend nonlinearly on the fitting
parameters (features). Hence, small errors in the NN predictions can
lead to large errors in the predicted distributions. This is
problematic, as the NNs are trained on a loss function that measures
the errors in feature space, whereas one is rather interested in the
quality of the predicted product state distributions. By using a loss
function that depends on errors in the predicted product distributions
this problem can be avoided. In the K-based approach this is partially
resolved by the choice of a Gaussian kernel where the hyperparameter
$\sigma$ is assigned according to the procedure described in Section
\ref{K-based method}. This results in a local kernel with kernel
coefficients largely determined by the amplitude at the corresponding
grid points and its nearest neighbours.\cite{bengio:2006}
Consequently, errors in the predicted kernel coefficients are also
restricted to impact the predicted model function locally, similar to
the G-based approach.\\
\subsection{Computational Cost and Generalizability}
To compare the computational cost of the final models in the F-, K-
and G-based approaches, the evaluation times of the final models for
1000 randomly selected data sets (from Set1) are considered. Here, a
single evaluation is defined as a prediction of the product
($E_{\text{trans}}',v',j'$) distributions at 201, 48 and 241 evenly
spaced points in the interval between $E_{\text{trans}}' = 0-20$ eV,
$v' = 0-47$ and $j' = 0-240$, respectively, given the reactant state
distributions. The evaluation times on a 1.8 GHz Intel Core i7-10510U
CPU are $(9.0 \pm 0.1)$ s, $(29.0 \pm 0.3)$ s, and $(28.9 \pm 0.3)$ s
for the F-, K- and G-based models, respectively. For the F-based
approach the evaluation time is 3 times faster compared to the two
other methods and is dominated by fitting the reactant state
distributions to Eqs. \ref{eq:func_E_r} to \ref{eq:func_j_r} whereas
for the K- and G-based approaches the evaluation time is dominated by
the evaluation of the RKHS-based representations of the product
distributions given the predicted kernel coefficients or amplitudes,
respectively. This may be further improved for the K- and G-based
approaches by using a computationally efficient kernel
toolkit.\cite{MM.RKHS:2017}\\
\noindent
In terms of generality and transferability, an F-based model can {\it
not} be easily generalized to distributions with widely different
shapes emanating from the QCT simulations. New optimal model
functions, also suitable for training a NN would need to be found for
every single system. Conversely, with the K- and G-based approaches
all desired features of the distributions can be captured by an
appropriate choice of reproducing kernel and grid, requiring
fine-tuning of the corresponding hyperparameters. Compared to the
K-based approach, a G-based model only requires tuning of the
corresponding hyperparameters for the product state distributions. In
addition, for a G-based model it is also possible to use a linear
interpolation instead of a RKHS if one is not concerned with
extrapolation. Then, the grid for the product state distributions
needs to be chosen sufficiently dense suitable for linear
interpolation at the cost of an increased number of NN
parameters.\\
\subsection{Grid-Based Models for $T_{\text{vib}} \neq T_{\text{rot}}$}
As vibrational relaxation is often slow in hypersonic flow, assuming
$T_{\text{vib}} = T_{\text{rot}}$ is often not a good
approximation.\cite{boyd:2015,bender:2015} Therefore, the G-based
approach is extended to and tested for the case of $T_{\rm vib} \neq
T_{\rm rot}$ using Set2. Restricting this to the G-based approach is
motivated by the fact that it performed best so far, both in terms of
final model accuracy and practicability.\\
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.99\textwidth]{results_set2.png}
\caption{Performance of the G-based approach on Set2 from training on
Set1 (green) or Set2 (red). Product state distributions from
explicit QCT simulations (QCT) compared with model predictions from
the G-based approach by training on Set1 (G-NN (Set1)) and Set2
(G-NN (Set2)) for: (a-c) $\bm{T}$ = (20000 K, 7000 K, 5000 K) ,
(d-f) $\bm{T}$ = (10000 K, 14000 K, 9000 K), (g-i) $\bm{T}$ = (5000
K, 20000 K, 8000 K).}
\label{fig:tunequal}
\end{center}
\end{figure}
\noindent
First, predictions for Set2 were made based on the G-based model
(G-NN) trained on Set1 ($T_{\text{vib}} = T_{\text{rot}}$) and
compared with QCT data (QCT), see Figure \ref{fig:tunequal}. The
accuracy of this G-based model deteriorates (see Table
S3 for all performance measures) as the
difference between $T_{\text{vib}}$ and $T_{\text{rot}}$ increases
(green lines in Figure \ref{fig:tunequal}). Consequently, a new
G-based model was trained and evaluated on Set2. The resulting model
predictions (red lines in Figure \ref{fig:tunequal}) are very
accurate, close to the level of accuracy of the G-based model trained
and evaluated on Set1. Thus, the G-based approach performs equally
well for $T_{\text{vib}} = T_{\text{rot}}$ and $T_{\text{vib}} \neq
T_{\text{rot}}$.\\
\noindent
In an attempt to further improve the G-based approach, three
alternative types of input were considered. They are all based on
reducing the number of input which not only decreases computational
cost, but also removes redundant features which can improve prediction
accuracy.\cite{chandra:2014} Again, continuous distributions were
obtained from an RKHS of the discrete predictions. The first
(G$'$-based) model used values of the reactant state distributions at
a fixed but reduced number of grid points compared with the G-based
model used so far (see Table S1). Next, a model
using only the three temperatures characterizing the reactant state
distributions $\bm{T}$ as input (G($\bm{T}$)-based model) is
considered. This is meaningful because the value of $\bm{T}$ entirely
specifies the equilibrium reactant state distributions. A third model
used averages $\bm{\mu}=(\mu_{E_{\rm trans}},\mu_{v},\mu_{j})$ of the
reactant state distributions as input (G($\bm{\mu}$)-based model). For
generality, these models will be trained and evaluated on Set2 and all
performance measures are summarized in Table
S3.\\
\noindent
A G$'$-based approach using two grid points per reactant state
distribution ($E_{\text{trans}}=0.3,3.5$ eV; $v=2,12$; $j=30,150$)
still allows for accurate predictions of the product state
distributions. The location of these grid points is largely arbitrary,
but they should be sufficiently spaced to provide information about
the distribution at different locations. However, reducing to a single
grid point per reactant state distribution ($E_{\text{trans}}=0.6$ eV,
$v=6$, $j=60$) leads to a significant drop in the prediction
accuracy. The fact that 2 grid points per reactant state distribution
are required for accurate predictions can mainly be attributed to the
presence of noise in the distributions arising from finite sample
statistics (see Section IV in SI
for further clarification). The resulting predictions for the above
mentioned models for selected data sets from the test set are
displayed in Figure~S6 in the SI.\\
\noindent
For the G($\bm{T}$)-based model the performance is close to the
original G-based model trained and evaluated on Set2. This is
expected, as $\bm{T}$ entirely specifies the equilibrium reactant
state distributions and allows a NN to predict corresponding product
state distributions. Finally, providing the mean $\mu$ of each of the
reactant state distributions as input in a G($\bm{\mu}$)-based model
also leads to highly accurate predictions. This can be explained by
the fact, that the mean values $\bm{\mu}$ of the reactant equilibrium
distributions are uniquely linked to the corresponding set of
temperatures $\bm{T}$.\\
\noindent
The results of this Section highlight that all three {\it variants} of
the G-based model yield similarly high levels of accuracy as the
G-based approach, which makes them preferable as they are
computationally less expensive. These results may be specific to
reactant state distributions that can be uniquely specified by a
single parameter, such as a temperature or its mean value $\mu$. To
explore this and to further demonstrate the generality of the G-based
approach and its variants, a more diverse dataset for nonequilibrium
conditions (Set3) was finally considered, for which the reactant state
distributions are characterized by multiple sets of temperatures
$\bm{T}$.\\
\subsection{Grid-based Models for Nonequilibrium Product State Distributions}
As a final application, nonequilibrium DTD models are constructed for
Set3 which was generated by means of a weighted sum (see
Eq. \ref{eq:multit}) using Set2 (see Section
\ref{Generating_noneq_Datasets}) with $N \in [2,3]$, and $w_{n} \in
[1,2]$. Training and validation sets of variable sizes were
considered, whereas $N_{\text{test}}=125$ throughout. First, a
G-based model was trained on Set3 with
$N_{\text{train}}+N_{\text{valid}}=5000$. Again, all performance
measures are summarized in Table
S3.\\
\noindent
The predictions of this G-based model for three different data sets
from the test set are shown in Figure
\ref{fig:probmodel_differentinputs_R2}. In particular, the predictions
for these three data sets are characterized by a $R^2_{\rm QCT}$ value
closest to the mean $R^2_{\rm QCT}$ value as evaluated over the entire
test set, as well as the highest and lowest $R^2_{\rm QCT}$ value in
the test set, respectively. Once again, the G-based approach gives a
very accurate DTD model, close to the G-based model trained and
evaluated on Set2. However, the predictions of the G-based model for
Set3 have a larger variance compared to the G-based model for Set2. To
assess the influence of the training and validation set size on the
prediction accuracy, a learning curve was computed (Figure~S7). The NN prediction accuracy does
not significantly increase when $N_{\text{train}}+N_{\text{valid}}$
was increased from 5000 to 30000 in increments of 5000. Hence, this
justifies training and validating the variants of the G-based model
only on $N_{\text{train}}+N_{\text{valid}} = 5000$.\\
\noindent
In an attempt to further improve and reduce this model for Set3, the
dependence on different amount of input information (as was done for
Set2) was tested again. Accurate predictions are still possible with
amplitudes of $P(E_{\text{trans}}), P(v), P(j)$ at three different
grid points ($E_{\text{trans}}=0.3,1.5,3.5$ eV; $v=2, 6, 12$; $j = 30,
60, 150$), but the prediction accuracy decreases when reducing this to
two grid points ($E_{\text{trans}} = 0.3, 3.5$ eV; $v = 2,12$; $j =
30, 150$). Again it is found that with a G$'$-based approach the
number of grid points characterizing the reactant state distributions
can be significantly reduced compared to the grids in Table
S1. This suggests that the possibility to reduce
the number of input features (amplitudes) in such G-based models is a
generic property which can be systematically explored.\\
\begin{figure}[th!]
\begin{center}
\includegraphics[width=0.99\textwidth]{probmodel_differentinputs_R2.png}
\caption{Performance of G-based models and variants trained and evaluated on
Set3. Product state distributions obtained from explicit QCT
simulations (QCT), together with the predictions from the G- (G-NN),
G$'_{3}$- (G$'_{3}$-NN), G($\bm{\mu},\bm{\sigma}$)-
(G($\bm{\mu},\bm{\sigma}$)-NN) and G($\bm{w},\bm{T}$)-based
(G($\bm{w},\bm{T}$)-NN) models trained on Set3. G$'_{3}$-NN uses 3
grid points per reactant state distribution (see text). The data
sets considered here are from the test set of Set3. In particular,
the predictions for these three data sets are characterized by (a-c)
a $R^2_{\rm QCT}$ value closest to the mean $R^2_{\rm QCT}$ value as
evaluated over the entire test set, as well as (d-f) the largest and
(g-i) smallest $R^2_{\rm QCT}$ value in the test set,
respectively. The normalized weights $w_{n}/w_{\rm tot}$ and sets of
temperatures $\bm{T}$ characterizing the data sets displayed here
are given in Table S4.}
\label{fig:probmodel_differentinputs_R2}
\end{center}
\end{figure}
\noindent
Providing the mean $\mu$ and standard deviation $\sigma$ for each
reactant state distribution ($E_{\text{trans}}$,$v$,$j$) in a
G($\bm{\mu},\bm{\sigma}$)-based approach also yields a good model
whereas omitting the standard deviations $\bm{\sigma}$ as input
information results in a G($\bm{\mu}$)-based model with a
significantly lower prediction accuracy. This should be compared with
the G($\bm{\mu}$)-based model for Set2 which yielded accurate
predictions. The aforementioned differences of the G$'$- and
G($\bm{\mu}$)-based models for Set2 and Set3 can be attributed to the
fact that the reactant state distributions in Set3 show more diverse
shapes compared to Set2, which makes it necessary to provide
additional information to maintain a high prediction accuracy. In
particular, reactant state distributions in Set3 are nonequilibrium
distributions and consequently can {\it not} be uniquely specified by
a single parameter, such as a temperature or its mean value $\mu$, as
was the case in Set2. Rather, the G($\bm{\mu},\bm{\sigma}$)-based
approach for Set3 showed that the set of reactant state distributions
is characterized by specifying ($\bm{\mu},\bm{\sigma}$).\\
\noindent
Keeping this in mind, extending the G($\bm{T}$)-based approach to Set3
can be achieved by providing the sets of temperatures $\bm{T}$ from
which the particular reactant state distributions of Set3 were
generated, together with the set of weights $\bm{w}$ with which these
contributed (see Section \ref{Generating_noneq_Datasets}). This
results in a G($\bm{w},\bm{T}$)-based model. To always guarantee the
same number of NN input being specified, as expected by the NN used in
this work (see Section \ref{Neural Network}), zero padding was
used. Such a G($\bm{w},\bm{T}$)-based model for Set3 leads to accurate
predictions. The predictions of the G$'$-, G($\bm{\mu},\bm{\sigma}$)-
and G($\bm{w},\bm{T}$)-based model for selected data sets from the
test set are also reported in Figure
\ref{fig:probmodel_differentinputs_R2}.\\
\noindent
Interestingly, the G-, G$'$- and G($\bm{\mu},\bm{\sigma}$)-based
models trained on Set3 can accurately predict the product state
distributions when given the reactant state distributions from
Set2. The performance measures (see Table
S3) were calculated by considering the
subset of 158 data sets of Set2 from which the test set of Set3 was
generated. Even though these models were trained on reactant state
distributions given as a linear combination of two to three
equilibrium distributions, they can accurately generalize to
equilibrium reactant state distributions. This is not the case for the
G($\bm{w},\bm{T}$)-based model, which yields unreliable predictions
when applied to the reactant state distributions of Set2. This can be
attributed to the zero padding. In particular, such a model can not
generalize {\it at all} to reactant state distributions being composed
of more than three equilibrium distributions, as this requires more NN
input to be specified than there are input nodes. This point can be
addressed in the future by considering a different NN architecture,
allowing for a variable input size.\\
\noindent
Conversely, generalizing to reactant state distributions being
composed of more than three equilibrium distributions is possible for
the G-, G$'$- and G($\bm{\mu},\bm{\sigma}$)-based models trained on
Set3. Specifically, a set of 125 reactant and product state
distributions given as a linear combination of $N \in [1-10]$
distributions from Set2 (i.e., from the subset of 158 data sets of
Set2 from which the test set of Set3 was generated) with integer
weights $w_{n} \in [1-100]$ (see Section
\ref{Generating_noneq_Datasets}) was generated, referred to as
Set3A. When applied to the reactant state distributions of Set3A,
these models still predict the corresponding product state
distributions with high accuracy, see Table
S3. The final DTD model predictions using
the G-based approach as well as its variants for such data sets is
shown in Figure
\ref{fig:probmodel_diverse_differentinputs_R2}. Consequently, as
discussed in Section \ref{Generating_noneq_Datasets}, these models are
also expected to generalize well to most nonequilibrium distributions
encountered in practice.\\
\begin{figure}[th!]
\begin{center}
\includegraphics[width=0.99\textwidth]{probmodel_diverse_differentinputs_R2.png}
\caption{Performance of G-based models and variants trained on Set3
and evaluated on Set3A. Product state distributions obtained from
explicit QCT simulations (QCT), together with the predictions from
the G- (G-NN), G$'_{3}$- (G$'_{3}$-NN), G($\bm{\mu},\bm{\sigma}$)-
(G($\bm{\mu},\bm{\sigma}$)-NN) and G($\bm{w},\bm{T}$)-based
(G($\bm{w},\bm{T}$)-NN) models trained on Set3. G$'_{3}$-NN uses 3
grid points per reactant state distribution (see text). The data
sets considered here are from Set3A. In particular, the predictions
for these three data sets are characterized by (a-c) a $R^2_{\rm
QCT}$ value closest to the mean $R^2_{\rm QCT}$ value as evaluated
over the entire Set3A, as well as (d-f) the largest and (g-i)
smallest $R^2_{\rm QCT}$ value in Set3A, respectively. The
normalized weights $w_{n}/w_{\rm tot}$ and sets of temperatures
$\bm{T}$ characterizing the data sets displayed here are given in
Table S5.}
\label{fig:probmodel_diverse_differentinputs_R2}
\end{center}
\end{figure}
\section{Discussion and Conclusions}
The present work demonstrates that machine learning of product state
distributions from the corresponding reactant state distributions for
reactive atom + diatom collisions based on a NN (DTD model)
constitutes a promising alternative to a full but computationally very
demanding (or even unfeasible, e.g., for diatom + diatom type
collisions) treatment by means of explicit QCT simulations. For such
DTD models, only a subset of the state space of the reactant needs to
be sampled which drastically reduces the computational complexity of
the problem at hand.\cite{MM.nncs:2019} In particular, DTD models for
the N + O$_2$ $\rightarrow$ NO + O reaction were constructed following
three distinct (F-, K- and G-based) approaches for $T_{\rm vib} =
T_{\rm rot}$. Although all three approaches yield accurate predictions
for the product state distributions as judged from $R^2$ and RMSD
measures, the G-based approach performs best in terms of prediction
accuracy, generality and practical implementation. On the other hand
the F-based approach is computationally more efficient by a factor of
3 compared with the K- and G-based approaches. For the K- and G-based
approaches it is found that RMSD$_{\rm NN}$ and $R^2_{\rm NN}$ are
close to RMSD$_{\rm QCT}$ and $R^2_{\rm QCT}$, respectively. This is
different for the F-based approach, where RMSD$_{\rm NN}$ is smaller
than RMSD$_{\rm QCT}$ by a factor of 2 (similarly for $R^2$), see
Table \ref{tab:Set1_perfomance_measures}. This indicates that the
parametrizations used for the F-based model can still be improved. In
general, an F-based approach is feasible if a universally valid and
accurate parametrization for the distributions can be found, which
also allows for an accurate NN to be trained. However, finding such a
parametrization may not always be possible. Consequently, the G-based
approach is generally preferred.\\
\noindent
The G-based approach and its input-reduced variants (G$'$-,
G($\bm{\mu},\bm{\sigma}$)) were found to perform well, too, for
$T_{\rm vib} \neq T_{\rm rot}$ (Set2) and nonequilibrium reactant
state distributions (Set3 and Set3A). Consequently, the G$'$- and
G($\bm{\mu},\bm{\sigma}$)-based models are generally preferred over
the standard G-based model, as the reduced number of input lowers
their computational cost. Moreover, G-, G$'$- and
G($\bm{\mu},\bm{\sigma}$)-based models trained on Set3 are also
expected to generalize well to most realistic nonequilibrium
distributions. This is of particular relevance for applications in
hypersonics for which nonequilibrium effects are of importance.\\
\noindent
Therefore, it is also of interest to discuss the present findings in
the context of the methods traditionally employed in DSMC\cite{dsmc}
and CFD simulations for hypersonics. Continuum-level reaction rates
are required in a multi-temperature framework usually employed in CFD
solvers\cite{gnoffo1990upwind,wright1998data,nompelis2011implementation}. The
expression for the exchange reaction rates
\begin{equation}
k_{\rm exc} (\bm{T}) = \left(\cfrac{8 k_B T_{\rm trans}}{\pi \mu}
\right)^{1/2} \pi b_{max}^2 P_{\rm r},
\end{equation}
where $k_B$ is the Boltzmann constant, $\mu$ is the reduced mass of
the reactants. Here, $P_{\rm r}$ is the reaction probability, which
can be obtained in a computationally inexpensive way by integrating
one of the predicted product state distributions. While rates derived
in this manner are based on equilibrium distributions characterized by
$\bm{T}$, the vibrational population is nonequilibrium at high
temperatures ($T \geq 8000$ K).\cite{schwartz:2018} Nonequilibrium
effects are particularly relevant for diatomic dissociation because
high vibrational states have significantly increased probability for
dissociation. For instance, for the dissociation of N$_2$ (in N$_2$ +
N$_2$) studied in Ref. \citen{bender:2015}, at $T_{\rm trans} =T_{\rm
rot} = 10000$ K, the dissociation probability to form N$_2$ + N + N
increases by a factor of 500, when $T_{\rm vib}$ is increased from
8000 K to 20000 K. Conversely, for the exchange reaction considered in
the present work, at $T_{\rm trans} = T_{\rm rot} = 10000$ K,
increasing $T_{\rm vib}$ from 5000 K to 18000 K results in an increase
in the reaction rate by only 40 \% (see Figure
\ref{fig:fig10}). Therefore due to the weaker dependence of the
exchange reaction probability on vibrational energy, a Boltzmann
distribution at $T_{\rm vib}$ may be sufficient for modeling exchange
rates. However, if necessary, the simple model for non-Boltzmann
distribution developed in Ref. \citen{singh1:2019} can be approximated
by a linear combination of Boltzmann distributions to include
non-Boltzmann effects in the reaction rates for exchange reactions as
well. The reactant state distributions can then expressed as a linear
combination of equilibrium distributions, as was done here for Set3,
from which $P_{\rm r}$ can be calculated. Furthermore, the average
vibrational energy change due to decomposition reactions, another key
input required in CFD, can also be obtained by taking an appropriate
moment of the product state distributions.\cite{singh2:2019}\\
\noindent
As an alternative to CFD for hypersonic flow, coarse grained Master
equations (ME) are being used for modeling chemical
kinetics.\cite{panesi2013rovibrational,magin2012coarse,andrienko2017state}
Here, several rovibrational states are lumped together in groups and
only the transition rates between these groups are required which
considerably speeds up such simulations. The accuracy of such an
approach directly depends on the criterion with which the groups are
generated,
though.\cite{macdonald2018_QCT,macdonald2018construction_DMS} A DTD
model as developed here constitutes a new framework for tracking the
time evolution of the population in each rovibrational state in a
computationally feasible manner. In the context of the present work
the DTD model can be repeatedly used for drawing reactant state
distributions at each time step for propagating the ME. This is
similar to sequential QCT proposed in
Refs. \cite{bruehl1988theoretical1,bruehl1988theoretical2} and DMS
method \cite{schwartzentruber:2018}, but computationally more
efficient because it \textit{avoids} explicit trajectory
calculations.\\
\begin{figure}
\centering \includegraphics[width=3.5in]{Fig10.png}
\caption{Exchange reaction rates $k_{\rm exc} (\bm{T})$
for N + O$_2$ $\rightarrow$ NO + O as a function of $T_{\rm vib}$, where
$T_{\rm trans} = T_{\rm rot}$.}
\label{fig:fig10}
\end{figure}
\noindent
The product state distributions predicted from the DTD models can also
be used for developing simple function-based, state-specific models
for exchange reactions in
DSMC.\cite{bird1976molecular,boyd2017nonequilibrium,singh1:2019,gimelshein2017modeling}
Such a model can be used within DSMC to estimate state-specific
exchange (forward and backward) reaction probabilities instead of the
total collision energy (TCE) \cite{bird1981monte} model. Furthermore,
DTD models also provide a QCT-, physics-based alternative to the
phenomenological Borgnakke-Larsen model\cite{borgnakke1975statistical}
which is currently employed to sample internal energy and
translational energy of products formed in exchange reactions.\\
\noindent
There is scope to further extend and improve the present methods. One
of them concerns the application of the G-based models to predict
product state distributions which can subsequently be used as reactant
state distributions for QCT simulations or DTD models. This way,
starting from a set of reactant state distributions transient
distributions can be obtained after a certain number of cycles. This
will be of particular relevance for applications in
hypersonics. Moreover, data construction schemes, such as constructing
nonequilibrium distributions as a linear combination of equilibrium
distributions, may prove useful for training DTD models from a small
set of reactant and product state distributions obtained from explicit
QCT simulations that generalize far beyond the conditions used for the
reactant state distributions.\\
\noindent
Overall, the present work establishes that NN-based models for
distribution-to-distribution learning can be developed based on
explicit trajectory-based data. This will apply to both, data
generated from QCT and quantum simulations if sufficiently converged
and complete data can be generated. More generally, the approach
presented in this work will also be applicable to situations in which
initial distributions are mapped on final distributions by means of a
deterministic algorithm such as molecular dynamics simulations. In the
future it may also be of interest to consider a fourth approach to DTD
learning based on a distribution regression network\cite{kou:2019}
promising a higher prediction accuracy with fewer NN parameters
compared to the approaches investigated in this work. Moreover, it may
also be interesting to explore the possibility for constructing a
``state-to-distribution'' model which would be intermediate between
the DTD model and the earlier STS model\cite{MM.nncs:2019}.\\
\section*{Data and Code Availability}
All data required to train the NNs has been made available on zenodo
\url{https://doi.org/QQQ/zenodo/QQQ} and the code for training the DTD
models is available at \url{https://github.com/MMunibas/DTD}.
\section{Acknowledgments}
This work was supported by the Swiss National Science Foundation
through grants 200021-117810, 200020-188724, the NCCR MUST, and the
University of Basel.
| bb4883c5837c7768f590fe278598ea4d0b4a1749 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:Introduction}
Diffractive hadron interactions at high scattering energies are characterised by final states consisting of two systems well separated in rapidity, which carry the quantum numbers of the initial state hadrons. Most diffractive phenomena are governed by \text{soft}, large distance processes. Despite being dominated by the strong interaction, they remain largely inaccessible to the description by Quantum Chromodynamics (QCD) in terms of quark and gluon interactions. In many cases, perturbative QCD calculations are not applicable because the typical scales involved are too low. Instead, other models have to be employed, such as Regge theory~\cite{Collins:1977jy}, in which interactions are described in terms of the coherent exchange of reggeons and the pomeron.
\begin{figure}[!htb]\centering
\includegraphics[width=0.33\textwidth,align=t]{d20-080f
}
\caption{Diffractive vector meson electroproduction.}
\label{fig:intro_VMProd}
\end{figure}
Exclusive vector meson (VM) \text{electroproduction} $e+p \rightarrow e + \ensuremath{{\mathrm{VM}}}\xspace + Y$ is particularly suited to study diffractive phenomena. This process is illustrated in \figref{fig:intro_VMProd}. In leading order QED, the interaction of the electron with the proton is the exchange of a virtual photon which couples to the proton in a diffractive manner to produce a VM (${\ensuremath{{\rho^0}}}\xspace$, $\omega$, $\phi$, $J/\psi$, \dots) in the final state. The proton is scattered into a system $Y$, which can be a proton (\textit{elastic} scattering) or a diffractively excited system (\textit{diffractive proton dissociation}). Scales for the process are provided by the vector meson mass squared $m^2_{\ensuremath{{\mathrm{VM}}}\xspace}$, the photon virtuality $Q^2 = -q^2$, and the squared $4$-momentum transfer at the proton vertex $t$, the dependence on each of which can be studied independently.
In this paper, elastic and proton-dissociative \textit{photoproduction} ($Q^2=0$) of ${\ensuremath{{\rho^0}}}\xspace$ mesons is studied using electroproduction data at small $Q^2 < 2.5~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$. Electroproduction of {\ensuremath{{\rho^0}}}\xspace mesons has been studied previously at HERA in both the photoproduction regime and for large $Q^2\gg \Lambda_\mathrm{QCD}^2$ (the perturbative cut-off in QCD), as well as for elastic and proton-dissociative scattering~\cite{Aaron:2009wg, Aaron:2009xp, Aktas:2006qs, Adloff:2002tb, Adloff:1999kg, Adloff:1997jd, Aid:1996ee, Aid:1996bs, Chekanov:2007zr, Chekanov:2002rm, Breitweg:1999fm, Breitweg:1999jy, Breitweg:1998nh, Breitweg:1997ed, Derrick:1996vw, Derrick:1995vq, Derrick:1995yd,Abramowicz:2011pk}. Measurements at lower photon-proton centre-of-mass energy {\ensuremath{{W_{\gamma p}}}}\xspace have been published in fixed-target interactions~\cite{Ballam:1971wq,Park:1971ts,Ballam:1972eq,Struczinski:1975ik,Egloff:1979mg,Aston:1982hr}. Most recently, a measurement of exclusive {\ensuremath{{\rho^0}}}\xspace photoproduction has been performed at the CERN LHC in ultra-peripheral lead-proton collisions~\cite{Sirunyan:2019nog}.
The present measurement is based on a data set collected during the 2006/2007 HERA running period by the H1 experiment. Since {\ensuremath{{\rho^0}}}\xspace mesons decay almost exclusively into a pair of charged pions, the analysis is based on a sample of {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction events. Compared with previous HERA results, the size of the sample makes possible a much more precise measurement with a statistical precision at the percent level. It is then possible to extract up to three-dimensional kinematic distributions as a function of the invariant mass of the {\ensuremath{{\pi^+\pi^-}}}\xspace system ${\ensuremath{{m_{\pi\pi}}}}\xspace$, of ${\ensuremath{{W_{\gamma p}}}}\xspace$, and of $t$. However, the size of the dataset is such that the systematic uncertainties of the modelling of the H1 experiment are important.
The structure of the paper is as follows: Theoretical details of {\ensuremath{{\rho^0}}}\xspace meson photoproduction are discussed in \secref{sec:theory} with a focus on the description of {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction in terms of Regge theory (\secref{sec:theo_pipiRegge}), cross section definitions (\secref{sec:theo_xSecDef}), and Monte Carlo modelling of relevant processes (\secref{sec:theo_MCModel}). In \secref{sec:experimental}, the experimental method is detailed, including a description of the H1 detector (\secref{sec:exp_detector}), the dataset underlying the analysis (\secref{sec:exp_dataSample}), the unfolding procedure applied to correct detector level distributions (\secref{sec:exp_unfolding}), and systematic uncertainties of the measurement (\secref{sec:systematics}). Results are presented in \secref{sec:results}. They encompass a measurement of the integrated {\ensuremath{{\pi^+\pi^-}}}\xspace production cross section in the fiducial phase space (\secref{sec:res_sigmaPiPiFid}), a study of the invariant {\ensuremath{{m_{\pi\pi}}}}\xspace distributions (\secref{sec:res_sigmaRhoFid}), measurements of the scattering energy (\secref{sec:res_sigmaRhoOfW}) and $t$ dependencies of the {\ensuremath{{\rho^0}}}\xspace meson production cross sections (\secref{sec:res_dSigmaRhoOft}), as well as the extraction of the leading Regge trajectory from the two-dimensional $t$ and {\ensuremath{{W_{\gamma p}}}}\xspace dependencies (\secref{sec:res_dSigmaRhoOfWt}).
\section{Theory}
\label{sec:theory}
\subsection{{\ensuremath{{\pi^+\pi^-}}}\xspace and {\ensuremath{{\rho^0}}}\xspace meson photoproduction}
\label{sec:theo_pipiRegge}
In electron\footnote{In the following, the term ``electron'' is used indistinctly to refer to both positrons and electrons.}-proton collisions, {\ensuremath{{\pi^+\pi^-}}}\xspace and {\ensuremath{{\rho^0}}}\xspace meson photoproduction is studied in the scattering process
\begin{equation}
e(e) + p(p) \rightarrow e(e^\prime) + \pi^+(k_1) + \pi^-(k_2) + Y(p^\prime)\, \mathrm,
\end{equation}
where the 4-momenta of the participating particles are given in parentheses. The relevant kinematic variables are the electron-proton centre-of-mass energy squared
\begin{equation}
s = (e+p)^2 \mathrm,
\end{equation}
the photon virtuality, i.e.,\ the negative squared $4$-momentum transfer at the electron vertex
\begin{equation}
Q^2 = -q^2 = (e-e^\prime)^2 \mathrm,
\end{equation}
the inelasticity
\begin{equation}
y= (q \cdot p)/(e\cdot p) \,\mathrm,
\end{equation}
the $\gamma p$ centre-of-mass energy squared
\begin{equation}
W_{\gamma p}^2 = (q+p)^2 \mathrm,
\label{eqn:theo_Wgp}
\end{equation}
the invariant mass of the {\ensuremath{{\pi^+\pi^-}}}\xspace system squared
\begin{equation}
{\ensuremath{{m_{\pi\pi}^2}}}\xspace = (k_1+k_2)^2\mathrm,
\end{equation}
the squared $4$-momentum transfer at the proton vertex
\begin{equation}
t = (p-p^\prime)^2\mathrm,
\label{eqn:theo_t}
\end{equation}
and the invariant mass squared of the (possibly dissociated) final state proton system
\begin{equation}
{\ensuremath{{m_Y^2}}}\xspace = (p^\prime)^2\mathrm.
\end{equation}
In general, diffractive photoproduction of (light) vector mesons shares the characteristics of soft hadron-hadron scattering: In the high energy limit, the total and elastic cross sections are observed to rise slowly with increasing centre-of-mass energy. Differential cross sections $\ensuremath{{\mathrm{d}}} \sigma/\ensuremath{{\mathrm{d}}} t$ are \textit{peripheral}, favouring low $|t|$, \textit{forward} scattering. With increasing scattering energy, the typical $|t|$ of elastic cross sections becomes smaller, i.e.~forward peaks appear to shrink.
In vector dominance models (VDM)~\cite{VDM}, the photon is modelled as a superposition of (light) vector mesons which can interact strongly with the proton to subsequently form a bound VM state. Like hadron-hadron interactions in general, VM production is dominated by colour singlet exchange in the $t$-channel. The lack of a sufficiently hard scale makes these exchanges inaccessible to perturbative QCD in a large portion of the phase space. Empirical and phenomenological models are used instead to describe the data. At low $|t| \ll 1~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$, differential cross sections are observed to follow exponential dependencies $\ensuremath{{\mathrm{d}}} \sigma/ \ensuremath{{\mathrm{d}}} t \propto \exp(bt)$. In the optical interpretation, the exponential slope $b$ is related to the transverse size of the scattered objects. Towards larger $|t|$, cross section dependencies change into a less steep power-law dependence $\ensuremath{{\mathrm{d}}} \sigma / \ensuremath{{\mathrm{d}}} t \propto |t|^a$. In Regge theory~\cite{Collins:1977jy}, the dependence of hadronic cross sections on the scattering energy ${\ensuremath{{W_{\gamma p}}}}\xspace$ and the shrinkage of the forward peak are characterised by \textit{Regge trajectories} $\alpha(t)$. The contribution of a single \textit{Regge pole} to the differential elastic cross section is $\ensuremath{{\mathrm{d}}}\sigma_{\ensuremath{{\mathrm{el}}}\xspace} / \ensuremath{{\mathrm{d}}} t({\ensuremath{{W_{\gamma p}}}}\xspace) \propto W_{\gamma p}^{4 (\alpha(t)-1)}$~\cite{Collins:1977jy}. At low energies ${\ensuremath{{W_{\gamma p}}}}\xspace \lesssim 10~{\ensuremath{{\mathrm{GeV}}}}\xspace$, \textit{reggeon trajectories} $\alpha_{\ensuremath{{I\!\!R}}\xspace}(t)$ dominate which are characterised by intercepts $\alpha_{\ensuremath{{I\!\!R}}\xspace}(0) < 1$, i.e.,\ they result in cross sections that fall off with increasing energy~\cite{Donnachie:1994zb}. In the high energy limit, only the contribution of what is known as the \textit{pomeron} Regge pole remains because its trajectory $\alpha_{\ensuremath{{I\!\!P}}\xspace}(t)$ has a large enough intercept $\alpha_{\ensuremath{{I\!\!P}}\xspace}(0) \gtrsim 1$ for it not to have decreased to a negligible level. The shrinkage of the elastic forward peak with increasing energy is the result of a positive slope $\alpha_{\ensuremath{{I\!\!P}}\xspace}^\prime > 0$ of the trajectory $\alpha(t)$ at $t=0$.
\begin{figure}[!htb]\centering
\includegraphics[width=0.33\textwidth,align=c]{d20-080f2a}\hspace*{.1\textwidth}
\includegraphics[width=0.33\textwidth,align=c]{d20-080f2b}
\caption{Diagram of {\ensuremath{{\rho^0}}}\xspace meson production and decay in elastic (left) and proton-dissociative (right) $ep$ scattering in the VDM and Regge picture, where the interaction is governed by soft pomeron exchange in the high energy limit.}
\label{fig:intro_eptorhopDiag}
\end{figure}
Feynman-like diagrams which illustrate the interpretation of {\ensuremath{{\rho^0}}}\xspace meson photoproduction in the VDM/Regge picture are given in \figref{fig:intro_eptorhopDiag}. In the diagrams, the {\ensuremath{{\rho^0}}}\xspace meson is shown to decay into a pair of charged pions. This is the dominant decay channel with a branching ratio ${\ensuremath{{\mathcal{BR}}}}\xspace({\ensuremath{{\rho^0}}}\xspace \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace ) \simeq 99\%$~\cite{PDG}.
The decay structure of the {\ensuremath{{\rho^0}}}\xspace meson into {\ensuremath{{\pi^+\pi^-}}}\xspace is described by two decay angles~\cite{Schilling:1973ag}. These also give insight into the production mechanism of the {\ensuremath{{\rho^0}}}\xspace meson. Contributions with $s$-channel helicity conservation are expected to dominate, such that the {\ensuremath{{\rho^0}}}\xspace meson retains the helicity of the photon, i.e.,\ in photoproduction the {\ensuremath{{\rho^0}}}\xspace meson is transversely polarised~\cite{Breitweg:1997ed}.
While vector mesons dominate photon-proton interactions, the VDM approach does not provide a complete picture. This is particularly evident for {\ensuremath{{\rho^0}}}\xspace meson production where in the region of the {\ensuremath{{\rho^0}}}\xspace meson resonance peak also non-resonant {\ensuremath{{\pi^+\pi^-}}}\xspace production plays an important role. The non-resonant {\ensuremath{{\pi^+\pi^-}}}\xspace amplitude interferes with the resonant {\ensuremath{{\rho^0}}}\xspace meson production amplitude to produce a skewing of the Breit-Wigner resonance profile in the {\ensuremath{{\pi^+\pi^-}}}\xspace mass spectrum~\cite{Soding:1965nh}. Newer models thus aim to take a more general approach, e.g.,\ a recently developed model for {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction based on tensor pomeron exchange~\cite{Nachtmann:pipi}. That model seems to be in fair agreement with H1 data when certain model parameters are adjusted \cite{Phd:Abolz}. A more detailed investigation is beyond the scope of this paper.
\subsection{Cross section definition}
\label{sec:theo_xSecDef}
\subsubsection{Photon flux normalisation}
In this paper, photoproduction cross sections $\sigma_{{\ensuremath{{\gamma p}}}\xspace}$ are derived from the measured electron-proton cross sections $\sigma_{ep}$. At low $Q^2$, these can be expressed as a product of a flux of virtual photons $f^{T/L}_{\gamma/e}$ and a virtual photon-proton cross section $\sigma^{T/L}_{\gamma^* p}$:
\begin{equation}
\dfrac{\ensuremath{{\mathrm{d}}}^2 \sigma_{ep}}{\ensuremath{{\mathrm{d}}} y \ensuremath{{\mathrm{d}}} Q^2} =f^T_{\gamma/e}\left(y, Q^2\right) \, \sigma^T_{\gamma^* p}\left({\ensuremath{{W_{\gamma p}}}}\xspace(y,Q^2), Q^2\right) + f^L_{\gamma/e}\left(y, Q^2\right) \, \sigma^L_{\gamma^* p}\left({\ensuremath{{W_{\gamma p}}}}\xspace(y,Q^2), Q^2\right).
\label{eqn:sigmaEpFactoriz}
\end{equation}
A distinction is made between transversely and longitudinally polarised photons, as indicated by the superscripts $T$ and $L$, respectively. In the low $Q^2$ regime studied here, the transverse component is dominant. The transverse and longitudinal photon fluxes are given in the Weizs\"acker-Williams approximation~\cite{equivPhoton} by
\begin{align}
f^T_{\gamma/e}(y,Q^2) &= \dfrac{\alpha_\mathrm{em}}{2\pi} \dfrac{1}{yQ^2}\left( 1 + (1-y)^2 - 2(1-y) \dfrac{Q^2_\mathrm{min}}{Q^2} \right)
\intertext{ and}
f^L_{\gamma/e}(y,Q^2) &= \dfrac{\alpha_\mathrm{em}}{\pi} \dfrac{1}{yQ^2}\left(1-y\right) \text,
\end{align}
respectively, where $\alpha_\mathrm{em}$ is the fine structure constant, $m_e$ the mass of the electron, and $Q^2_\mathrm{min} = m_e^2 y^2/(1-y)$ is the smallest kinematically accessible $Q^2$ value.
In vector meson production in the VDM approach, the cross section for virtual photon interactions ($Q^2>0$) can be related to the corresponding photoproduction cross section $\sigma_{\gamma p}$ at $Q^2=0$:
\begin{align}
\sigma^T_{\gamma^* p}\left({\ensuremath{{W_{\gamma p}}}}\xspace, Q^2\right) &= \sigma_{\gamma p}\left({\ensuremath{{W_{\gamma p}}}}\xspace\right) \, \left( 1 + \frac{Q^2}{m^2_{\ensuremath{{\mathrm{VM}}}\xspace}} \right)^{-2} \mathrm, \\
\sigma^L_{\gamma^* p}\left({\ensuremath{{W_{\gamma p}}}}\xspace, Q^2\right) &= \sigma^T_{\gamma^* p}\left({\ensuremath{{W_{\gamma p}}}}\xspace, Q^2\right) \, \dfrac{Q^2}{m^2_{\ensuremath{{\mathrm{VM}}}\xspace}} \xi^2\mathrm.
\end{align}
In the equations, $m_{\ensuremath{{\mathrm{VM}}}\xspace}$ denotes the mass of the considered vector meson and $\xi^2$ is a proportionality constant, that in the following is set to unity. The real photoproduction cross section can then be factored out of \eqnref{eqn:sigmaEpFactoriz} according to
\begin{equation}
\dfrac{\ensuremath{{\mathrm{d}}}^2 \sigma_{ep}}{\ensuremath{{\mathrm{d}}} y \ensuremath{{\mathrm{d}}} Q^2} = \sigma_{\gamma p}\left( {\ensuremath{{W_{\gamma p}}}}\xspace(y) \right)\, \varphi_\mathrm{eff}\left(y, Q^2\right) \mathrm,
\end{equation}
were the so-called \textit{effective photon flux} is given by
\begin{equation}
\varphi_\mathrm{eff}(y, Q^2) = \dfrac{\alpha_\mathrm{em}}{2\pi} \dfrac{1}{yQ^2} \left[ 1 + (1-y)^2 - 2(1-y) \left( \dfrac{Q^2_\mathrm{min}}{Q^2}-\dfrac{Q^2}{m^2_{\ensuremath{{\mathrm{VM}}}\xspace}} \right) \right] \left( 1 + \frac{Q^2}{m^2_{\ensuremath{{\mathrm{VM}}}\xspace}} \right)^{-2}\mathrm.
\label{eqn:theo_effFlux}
\end{equation}
In practice, measurements of $\sigma_{ep}$ are evaluated as integrals over finite regions in $Q^2$ and ${\ensuremath{{W_{\gamma p}}}}\xspace$. In order to extract corresponding photoproduction cross sections at $Q^2=0$ and for an appropriately chosen average energy $\langle {\ensuremath{{W_{\gamma p}}}}\xspace \rangle$, the measured values are normalised by the effective photon flux integrated over the corresponding $Q^2$ and ${\ensuremath{{W_{\gamma p}}}}\xspace$ ranges:
\begin{equation}
\sigma_{\gamma p}\left(\langle {\ensuremath{{W_{\gamma p}}}}\xspace \rangle\right) = \dfrac{ \sigma_{ep} }{\ensuremath{{\Phi_{\gamma/e}}}\xspace} \mathrm,
\end{equation}
with
\begin{equation}
\ensuremath{{\Phi_{\gamma/e}}}\xspace = \displaystyle \int_{\scriptsize\begin{array}{c}Q^2_{\min}<Q^2<Q^2_{\max} \\
W_{\min}<{\ensuremath{{W_{\gamma p}}}}\xspace<W_{\max}\end{array}} \varphi_\mathrm{eff}(y, Q^2) \, \ensuremath{{\mathrm{d}}} y\ensuremath{{\mathrm{d}}} Q^2 \mathrm.
\label{eqn:theo_intFlux}
\end{equation}
\subsubsection{{\ensuremath{{\rho^0}}}\xspace meson photoproduction cross section}
\label{sec:theo_xSecRho}
With the H1 detector, {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction is measured. In the kinematic region studied, there are significant contributions from the {\ensuremath{{\rho^0}}}\xspace meson resonance, non-resonant {\ensuremath{{\pi^+\pi^-}}}\xspace production, the $\omega$ meson resonance and excited $\rho$ meson states~\cite{PDG,Nachtmann:pipi}. Photoproduction of {\ensuremath{{\rho^0}}}\xspace mesons has to be disentangled from these processes by analysis of the {\ensuremath{{\pi^+\pi^-}}}\xspace mass spectrum.
Here, a model is used which parametrises the spectrum in terms of {\ensuremath{{\rho^0}}}\xspace and $\omega$ meson, as well as non-resonant amplitudes~\cite{Phd:Abolz}, similar to the original proposal made by S\"oding~\cite{Soding:1965nh}.
The contributions are added at the amplitude level, and interference effects are taken into account. The model is defined as
\begin{equation}
\dfrac{\ensuremath{{\mathrm{d}}} \sigma_{\pi\pi}}{\ensuremath{{\mathrm{d}}}{\ensuremath{{m_{\pi\pi}}}}\xspace }({\ensuremath{{m_{\pi\pi}}}}\xspace) = A\ \dfrac{q^3({\ensuremath{{m_{\pi\pi}}}}\xspace)}{q^3(m_\rho)} \ \bigg \vert A_{\rho,\omega}({\ensuremath{{m_{\pi\pi}}}}\xspace) + A_{\ensuremath{{\mathrm{nr}}}\xspace}({\ensuremath{{m_{\pi\pi}}}}\xspace) \bigg \vert^2\text,
\label{eqn:theo_rhoSoedMass}
\end{equation}
where $A$ is a global normalisation factor and $q^3({\ensuremath{{m_{\pi\pi}}}}\xspace)$ describes the phase space, with $q({\ensuremath{{m_{\pi\pi}}}}\xspace) = \frac{1}{2} \sqrt{{\ensuremath{{m_{\pi\pi}^2}}}\xspace - 4m_\pi^2}$ being the momentum of one of the pions in the {\ensuremath{{\pi^+\pi^-}}}\xspace centre-of-mass frame~\cite{Jackson:1964zd}. It is normalised to the value $q^3(m_\rho)$ at the {\ensuremath{{\rho^0}}}\xspace mass. The amplitude $A_{\rho,\omega}$ takes into account the ${\ensuremath{{\rho^0}}}\xspace$ and $\omega$ meson resonance contributions, whereas the non-resonant component is modelled by $A_{\ensuremath{{\mathrm{nr}}}\xspace}({\ensuremath{{m_{\pi\pi}}}}\xspace)$. The components are considered to be fully coherent. Since additional resonances with masses above 1~{\ensuremath{{\mathrm{GeV}}}}\xspace are neglected, the model can only be applied to the region near the {\ensuremath{{\rho^0}}}\xspace meson mass peak. Models of this form have been widely used in similar past analyses because they preserve the physical amplitude structure while being parametric and thus easily applicable.
More sophisticated models are designed to preserve unitarity \cite{Basdevant:1977ya} or are regularised by barrier factors at higher masses \cite{VonHippel:1972fg}.
The combined ${\ensuremath{{\rho^0\text{-}\omega}}}\xspace$ amplitude is modelled following a parametrisation given by \cite{Akhmetshin:2001ig}
\begin{equation}
A_{\rho,\omega}({\ensuremath{{m_{\pi\pi}}}}\xspace) = \mathcal{BW}_{\rho}({\ensuremath{{m_{\pi\pi}}}}\xspace) \ \left( 1 + f_\omega \ensuremath{{\mathrm{e}}}^{\ensuremath{{\mathrm{i}}} \phi_{\omega} } \dfrac{ {\ensuremath{{m_{\pi\pi}^2}}}\xspace}{ m_\omega^2 } \mathcal{BW}_{\omega}({\ensuremath{{m_{\pi\pi}}}}\xspace) \right),
\end{equation}
where $f_\omega$ and $\phi_\omega$ are a normalisation factor and a mixing phase for the $\omega$ contribution, respectively. The $\omega$ meson is not expected to decay into a pair of charged pions directly because of the conservation of G-parity by the strong interaction. However, electromagnetic $\omega\rightarrow{\ensuremath{{\rho^0}}}\xspace$-mixing with a subsequent ${\ensuremath{{\rho^0}}}\xspace \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace$ decay is possible~\cite{OConnell:1995nse}. Both resonances are modelled by a relativistic Breit-Wigner function~\cite{Breit:1936zzb}:
\begin{equation}
\mathcal{BW}_{\ensuremath{{\mathrm{VM}}}\xspace}({\ensuremath{{m_{\pi\pi}}}}\xspace) = \dfrac{ m_{\ensuremath{{\mathrm{VM}}}\xspace} \Gamma_{{\ensuremath{{\mathrm{VM}}}\xspace}} }{ m^2_{\pi\pi} - m_{\ensuremath{{\mathrm{VM}}}\xspace}^2 + \ensuremath{{\mathrm{i}}} \, m_{\ensuremath{{\mathrm{VM}}}\xspace} \Gamma({\ensuremath{{m_{\pi\pi}}}}\xspace) }.
\label{eqn:effingBW}
\end{equation}
The parameters $m_{\ensuremath{{\mathrm{VM}}}\xspace}$ and $\Gamma_{{\ensuremath{{\mathrm{VM}}}\xspace}}$ are the respective vector meson's Breit-Wigner mass and width. The Breit-Wigner function is normalised to $|\mathcal{BW}_{\ensuremath{{\mathrm{VM}}}\xspace}(m_{\ensuremath{{\mathrm{VM}}}\xspace})|=1$. For the {\ensuremath{{\rho^0}}}\xspace resonance a p-wave mass-dependent width~\cite{Jackson:1964zd} is used:
\begin{equation}
\Gamma({\ensuremath{{m_{\pi\pi}}}}\xspace) = \Gamma_{{\ensuremath{{\mathrm{VM}}}\xspace}} \dfrac{q^3({\ensuremath{{m_{\pi\pi}}}}\xspace)}{q^3(m_{\ensuremath{{\mathrm{VM}}}\xspace})} \dfrac{m_{\ensuremath{{\mathrm{VM}}}\xspace}}{{\ensuremath{{m_{\pi\pi}}}}\xspace}\mathrm{,}
\label{eqn:gamma_rho}
\end{equation}
whereas a constant width is assumed for the very narrow $\omega$ resonance.
The unknown non-resonant amplitude is parametrised by the function
\begin{equation}
A_{\ensuremath{{\mathrm{nr}}}\xspace}({\ensuremath{{m_{\pi\pi}}}}\xspace) = \dfrac{f_{\ensuremath{{\mathrm{nr}}}\xspace}}{ \left( {\ensuremath{{m_{\pi\pi}^2}}}\xspace - 4m_\pi^2 + \Lambda^2_{\ensuremath{{\mathrm{nr}}}\xspace} \right)^{\delta_{\ensuremath{{\mathrm{nr}}}\xspace}}} \mathrm{,}
\label{eqn:theo_Anr}
\end{equation}
where the relative normalisation is given by $f_{\ensuremath{{\mathrm{nr}}}\xspace}$, while $\Lambda_{\ensuremath{{\mathrm{nr}}}\xspace}$ and $\delta_{\ensuremath{{\mathrm{nr}}}\xspace}$ are free model parameters. They can shape the amplitude for the modelling of a possible internal structure of the non-resonant $\gamma\pi\pi$-coupling. In similar past analyses, typically a purely real non-resonant amplitude has been assumed. Following that assumption, $f_{\ensuremath{{\mathrm{nr}}}\xspace}$ is set to be real. For $\delta_{\ensuremath{{\mathrm{nr}}}\xspace} > \frac{3}{4}$, the non-resonant contribution to the cross section (cf. \eqnref{eqn:theo_rhoSoedMass}) has a local maximum at
\begin{equation}
m_{\ensuremath{{\mathrm{nr}}}\xspace}^{\text{max}} = \dfrac{ \sqrt{\Lambda^2_{\ensuremath{{\mathrm{nr}}}\xspace} + (\frac{4}{3}\delta_{\ensuremath{{\mathrm{nr}}}\xspace}-1)\, 4m_\pi^2} }{\sqrt{\frac{4}{3}\delta_{\ensuremath{{\mathrm{nr}}}\xspace}-1}}\mathrm,
\end{equation}
and falls proportionally to $\left( 1/{\ensuremath{{m_{\pi\pi}^2}}}\xspace \right)^{2\delta_{\ensuremath{{\mathrm{nr}}}\xspace}-3}$ in the high mass region.
In order to extract the {\ensuremath{{\rho^0}}}\xspace meson contribution to the {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross section, the measured {\ensuremath{{\pi^+\pi^-}}}\xspace mass distributions are fitted using \eqnref{eqn:theo_rhoSoedMass}. The {\ensuremath{{\rho^0}}}\xspace meson Breit-Wigner contribution is then conventionally defined by the integral
\begin{equation}
\sigma(\gamma p \rightarrow {\ensuremath{{\rho^0}}}\xspace Y) = \dfrac{A}{q^3(m_\rho)} \int_{2m_\pi}^{m_\rho + 5 \Gamma_{\rho} } \big\vert \mathcal{BW}_{\rho}(m) \big\vert^2 q^3(m) \ensuremath{{\mathrm{d}}} m.
\label{eqn:theo_intRho}
\end{equation}
As the {\ensuremath{{\rho^0}}}\xspace meson resonance decays almost exclusively into two charged pions, this is taken to be equal to the total {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross section without correcting for the ${\ensuremath{{\rho^0}}}\xspace \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace$ branching fraction.
Kinematic dependencies of the {\ensuremath{{\rho^0}}}\xspace meson production cross section on the variables {\ensuremath{{W_{\gamma p}}}}\xspace, $t$, and {\ensuremath{{m_Y}}}\xspace are measured by fitting the mass distributions in bins of the respective variables, such that all model parameters may have kinematic dependencies.
Physical considerations and statistical and technical limitations affect the assumed dependencies. Physical parameters, i.e.,\ $m_{\ensuremath{{\mathrm{VM}}}\xspace}$, $\Gamma_{{\ensuremath{{\mathrm{VM}}}\xspace}}$, and $\delta_{\ensuremath{{\mathrm{nr}}}\xspace}$ are assumed to be constants. The small width of the $\omega$ meson cannot be constrained by the present data and the PDG value $\Gamma_{\omega}=8.5~{\ensuremath{{\mathrm{MeV}}}}\xspace$~\cite{PDG} is assumed and kept fixed in all fits. Dependencies of $f_\omega$, $\phi_\omega$, and $f_{\ensuremath{{\mathrm{nr}}}\xspace}$ on $t$ or {\ensuremath{{W_{\gamma p}}}}\xspace cannot be constrained with the present data. However, these parameters are allowed to depend on {\ensuremath{{m_Y}}}\xspace, i.e.,\ to be different for elastic and proton-dissociative distributions. The non-resonant background is observed to change with $t$. This is modelled by a $t$ dependence of the parameter $\Lambda_{\ensuremath{{\mathrm{nr}}}\xspace}$, which also can be different for elastic and proton-dissociative distributions. The normalisation $A$ is a free parameter in each kinematic bin. All fit set-ups with the corresponding parameter assumptions are summarised in \tabref{tab:theo_SodingPars}.
\begin{table}\centering
\small
\begin{tabular}{@{}l@{\hspace{0.5em}} l@{\hspace{1.0em}} c c c c@{}}
\toprule
& & \multicolumn{4}{c}{Number of free fit parameters} \\
\cmidrule{3-6}
Parameter & Dependencies & $\frac{\ensuremath{{\mathrm{d}}} \sigma}{\ensuremath{{\mathrm{d}}} m}(m;\, {\ensuremath{{m_Y}}}\xspace)$ & $\frac{\ensuremath{{\mathrm{d}}} \sigma}{\ensuremath{{\mathrm{d}}} m}(m;\, {\ensuremath{{m_Y}}}\xspace,W)$ & $\frac{\ensuremath{{\mathrm{d}}}^2 \sigma}{\ensuremath{{\mathrm{d}}} m \ensuremath{{\mathrm{d}}} t}(m;\, {\ensuremath{{m_Y}}}\xspace, t)$ & $\frac{\ensuremath{{\mathrm{d}}}^2 \sigma}{\ensuremath{{\mathrm{d}}} m \ensuremath{{\mathrm{d}}} t}(m;\, {\ensuremath{{m_Y}}}\xspace, W, t)$\\
\midrule
$A$ & ${\ensuremath{{m_Y}}}\xspace,\ W,\ t$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $9_{W}^{\ensuremath{{\mathrm{el}}}\xspace\vphantom{p}} + 6_{W}^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $12_{t}^{\ensuremath{{\mathrm{el}}}\xspace\vphantom{p}} + 9_{t}^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $4_{W}^{\ensuremath{{\mathrm{el}}}\xspace\vphantom{p}}\cdot7_{t}^{\ensuremath{{\mathrm{el}}}\xspace\vphantom{p}} + 4_{W}^{\ensuremath{{\mathrm{pd}}}\xspace}\cdot5_{t}^{\ensuremath{{\mathrm{pd}}}\xspace}$ \\
$m_\rho$ & - & $1$ & $1$ & $1$ & $1$ \\
$\Gamma_\rho$ & - & $1$ & $1$ & $1$ & $1$ \\
$f_\omega$ & ${\ensuremath{{m_Y}}}\xspace$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & fixed & fixed & fixed \\
$\phi_\omega$ & ${\ensuremath{{m_Y}}}\xspace$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & fixed & fixed & fixed \\
$m_\omega$ & - & $1$ & fixed & fixed & fixed \\
$\Gamma_\omega$ & - & PDG & PDG & PDG & PDG \\
$f_{\ensuremath{{\mathrm{nr}}}\xspace}$ & ${\ensuremath{{m_Y}}}\xspace$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ \\
$\delta_{\ensuremath{{\mathrm{nr}}}\xspace}$ & - & $1$ & $1$ & $1$ & $1$ \\
$\Lambda_{\ensuremath{{\mathrm{nr}}}\xspace}$ & ${\ensuremath{{m_Y}}}\xspace,\ t$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $1^{\ensuremath{{\mathrm{el}}}\xspace}+1^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $12_{t}^{\ensuremath{{\mathrm{el}}}\xspace\vphantom{p}} + 9_{t}^{\ensuremath{{\mathrm{pd}}}\xspace}$ & $7_{t}^{\ensuremath{{\mathrm{el}}}\xspace\vphantom{p}} + 5_{t}^{\ensuremath{{\mathrm{pd}}}\xspace}$\\
\midrule
\multicolumn{2}{@{}l}{Total} & $14$ & $22$ & $47$ & $65$ \\
\bottomrule
\end{tabular}
\caption{%
Parameter assumptions and resulting number of parameters used to fit \eqnref{eqn:theo_rhoSoedMass} to the invariant {\ensuremath{{\pi^+\pi^-}}}\xspace mass distributions.
The number of parameters depends on the number of bins in the extracted cross sections. The number of bins in which a parameter is fitted freely is given and the corresponding distribution indicated by sub- and superscripts. For the fits of the {\ensuremath{{m_{\pi\pi}}}}\xspace distributions in multiple {\ensuremath{{W_{\gamma p}}}}\xspace or $t$ bins, the $\omega$ meson model parameters are fixed to the values obtained from the fit to the one-dimensional distributions. The $\omega$ meson width is always fixed to the PDG value. }
\label{tab:theo_SodingPars}
\end{table}
\subsection{Monte Carlo modelling}
\label{sec:theo_MCModel}
For the purpose of quantifying detector effects, the data are modelled using Monte Carlo (MC) simulations of elastic and proton-dissociative electroproduction of ${\ensuremath{{\rho^0}}}\xspace$, $\omega(782)$, $\phi(1020)$, $\rho(1450)$, and $\rho(1700)$ vector mesons, as well as for non-resonant diffractive photon dissociation. The samples are all generated using the DIFFVM event generator\cite{List:1999} that models VM production on the principles of equivalent photon approximation\cite{equivPhoton}, VDM\cite{VDM}, and pomeron exchange \cite{ReggeTraj}. Proton dissociation is modelled by DIFFVM assuming the following dependence of the cross section on the mass of the dissociated system:
\begin{equation}
\dfrac{\ensuremath{{\mathrm{d}}} \sigma_{\ensuremath{{\gamma p}}}\xspace}{\ensuremath{{\mathrm{d}}} {\ensuremath{{m_Y^2}}}\xspace} = \dfrac{f({\ensuremath{{m_Y}}}\xspace)}{ ({\ensuremath{{m_Y^2}}}\xspace)^{1+\epsilon_Y} }\mathrm{.}
\end{equation}
Here, $\epsilon_Y = 0.0808$ and $f({\ensuremath{{m_Y}}}\xspace)$ is a phenomenological function that is fitted to the experimental data~\cite{ProtonDissoc} to parametrise the low-mass resonance structure in the region $m_p < {\ensuremath{{m_Y}}}\xspace < 1.9~{\ensuremath{{\mathrm{GeV}}}}\xspace$. For ${\ensuremath{{m_Y}}}\xspace>1.9~{\ensuremath{{\mathrm{GeV}}}}\xspace$, the function $f({\ensuremath{{m_Y}}}\xspace)=1$ becomes constant. In the low mass region the dissociative system is treated as an $N^*$ resonance and decays are modelled according to measured branching fractions\cite{PDG}. For higher masses the decay is simulated using the Lund fragmentation model as implemented in JETSET~\cite{Sjostrand:1986hx}. Non-resonant photon dissociation is modelled analogously by assuming a dissociative mass $m_X$ spectrum $\ensuremath{{\mathrm{d}}} \sigma_{{\ensuremath{{\gamma p}}}\xspace}/\ensuremath{{\mathrm{d}}} m_X^2 = 1/(m_X^2)^{1+\epsilon_X}$ with $\epsilon_X = \epsilon_Y$, and simulating the decay using the Lund model.
\begin{table}[htb] \centering
\small
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{ @{}l@{\hspace{2.5em}} l @{\hspace{2em}}r@{\hspace{2em}} c c@{}}
\toprule
& & & \multicolumn{2}{c}{Number of events} \\
\cmidrule{4-5}
Process & Decay modes & ${\ensuremath{{\mathcal{BR}}}}\xspace$ $[\%]$ & elastic & $p$-dissociative \\
\midrule
\multirow{2}{*}{${\ensuremath{{\rho^0}}}\xspace(770)$} & $\pi^+\pi^-$ & $ 99.0$ & \multirow{2}{*}{$10^7$} & \multirow{2}{*}{$10^7$} \\
& ${\ensuremath{{\pi^+\pi^-}}}\xspace\gamma$ & $ 1.0$ \\
\multicolumn{3}{l}{$\quad \hookrightarrow$ reweighted to describe all {\ensuremath{{\pi^+\pi^-}}}\xspace final states} &\\
\cmidrule(r){1-3} \cmidrule{4-5}
\multirow{3}{*}{$\omega(782)$} & $\pi^+\pi^-\pi^0$ & $ 89.2$ & \multirow{3}{*}{$10^6$} & \multirow{3}{*}{$10^6$} \\
& $\pi^0\gamma$ & $ 8.6$ \\
& $\pi^+\pi^-$ (removed, included in signal) & $ 2.2$ \\
\cmidrule(r){1-3} \cmidrule{4-5}
\multirow{5}{*}{$\phi(1020)$} & $K^+ K^-$ & $ 49.0$ & \multirow{5}{*}{$10^6$} & \multirow{5}{*}{$10^6$}\\
& $K_L K_S$ & $ 34.4$ \\
& $\pi^+\rho^-,\ \pi^-\rho^+,\ \pi^0\rho^0$ & $4.3$, $4.3$, $4.3$ \\
& $\pi^+\pi^-\pi^0$ & $ 2.4 $\\
& $\eta \gamma$ & $ 1.3 $\\
\cmidrule(r){1-3} \cmidrule{4-5}
\multirow{2}{*}{$\rho(1450)$} & $\rho^0\pi^+\pi^-,\ \rho^+\pi^-\pi^0,\ \rho^-\pi^+\pi^0$ & $25.0$, $25.0$, $ 25.0 $ & \multirow{2}{*}{$10^6$} & \multirow{2}{*}{$10^6$} \\
\multirow{2}{*}{\& } & $\pi^+\pi^-\pi^+\pi^-$ & $ 15.0$ \\
\multirow{2}{*}{$\rho(1700)$} & $\pi^+\pi^-\pi^0\pi^0$ & $ 8.0$ & \multirow{2}{*}{$10^6$} & \multirow{2}{*}{$10^6$}\\
& $\pi^+\pi^-$ (removed, included in signal) & $ 2.0$ \\
\multicolumn{3}{l}{$\quad \hookrightarrow$ merged $\rho(1450):\rho(1700) = 1:1$ } & \\
\cmidrule(r){1-3} \cmidrule{4-5}
\multirow{2}{*}{$\gamma$-dissoc.} & \multicolumn{2}{l}{Lund fragmentation model } & \multirow{2}{*}{$10^7$}& \multirow{2}{*}{$10^7$} \\
& \multicolumn{2}{l}{ (exclusive ${\ensuremath{{\pi^+\pi^-}}}\xspace$ removed, included in signal) } \\
\bottomrule
\end{tabular}
\caption{DIFFVM MC samples used to model the {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction dataset. All decay modes with a branching fraction $\gtrsim1\%$ are simulated~\cite{List:1999}. For the $\rho(1450)$ and $\rho(1700)$ samples a ratio of 1:1 is assumed. The {\ensuremath{{\pi^+\pi^-}}}\xspace final states are removed from background samples and included in the signal definition.}
\label{tab:theo_diffVMSamples}
\end{table}
In \tabref{tab:theo_diffVMSamples}, details on the samples and in particular on the simulated decay modes are listed\footnote{For the simulation of the $\rho(1450)$ and $\rho(1700)$ mesons, DIFFVM was modified to account for the finite width of intermediate $\rho(770)$ resonances, and decay modes involving the final state ${\ensuremath{{\pi^+\pi^-}}}\xspace \pi^0\pi^0$ were added.}. Several of the considered processes result in an exclusive {\ensuremath{{\pi^+\pi^-}}}\xspace final state. They are simulated by DIFFVM independently, so that interference effects are not considered. However, these can be significant. For example, the interference between the {\ensuremath{{\rho^0}}}\xspace meson resonance and non-resonant {\ensuremath{{\pi^+\pi^-}}}\xspace production causes a strong skewing of the resonance lineshape. To consider these interference effects, the ${\ensuremath{{\rho^0}}}\xspace$ meson samples are reweighted to describe exclusive {\ensuremath{{\pi^+\pi^-}}}\xspace production including contributions from {\ensuremath{{\rho^0}}}\xspace, $\omega$, and a single {\ensuremath{{\rho^\prime}}}\xspace meson resonance and non-resonant production, that are all added at the amplitude level.
For the reweighting, a $t$ and {\ensuremath{{m_{\pi\pi}}}}\xspace dependent lineshape is used, which is similar to the model introduced in \secref{sec:theo_xSecRho} but is extended by an additional {\ensuremath{{\rho^\prime}}}\xspace Breit-Wigner amplitude~\cite{Phd:Abolz}.
All generated events are passed through the full GEANT-3\cite{Brun:1994aa} based simulation of the H1 apparatus and are reconstructed using the same program chain as used for the data. Trigger scaling factors are applied to correct differences in the trigger performance between data and simulation. They are obtained from a {\ensuremath{{\pi^+\pi^-}}}\xspace electroproduction sample, that is triggered independently of the tracking devices~\cite{Phd:Abolz}.
A template is constructed from all MC samples to describe the data. For a better description of the {\ensuremath{{W_{\gamma p}}}}\xspace and $t$ distributions, all samples are tuned to data~\cite{Phd:Abolz}. An additional background contribution from beam-gas events is estimated in a data driven method. The MC normalisations are obtained from data control regions that are enriched with events from a given process through modified event selection requirements as described below. The $\rho(1450)$ and $\rho(1700)$ samples cannot be well distinguished experimentally in this analysis. They are thus combined at a $1{:}1$ ratio and treated as a single MC sample. In order to obtain normalisations for the elastic and proton-dissociative samples independently, information from the forward detector components is used as described below.
Neither initial and final state radiation of real photons from the electron, nor vacuum polarisation effects are simulated. Consequently, these effects are not corrected for in the present measurement. In a comparable phase space, their effect on the overall cross section has been estimated to be smaller than 2\%~\cite{Kurek:1996ez}.
\section{Experimental Method and Data Analysis}
\label{sec:experimental}
\subsection{H1 detector}
\label{sec:exp_detector}
A detailed description of the H1 detector is given elsewhere~\cite{Abt:1996xvhi,Appuhn:1996na}. The components that are relevant for the present analysis are briefly described in the following. A right-handed Cartesian coordinate system is used with the origin at the nominal $ep$ interaction point. The proton beam direction defines the positive $z$-axis (\textit{forward direction}). Transverse momenta are measured in the $x$-$y$ plane. Polar ($\theta$) and azimuthal ($\phi$) angles are measured with respect to this frame of reference.
The interaction point is surrounded in the central region ($15^\circ < \theta < 165^\circ$) by the central tracking detector. Two large coaxial cylindrical jet chambers (CJC1 and CJC2) for precise track reconstruction form its core. They are supported by the central inner proportional chamber (CIP) used for the reconstruction of the primary vertex position on the trigger level, a $z$-drift chamber for an improved reconstruction of $z$ coordinates, and a silicon vertex detector for the reconstruction of secondary decay vertices~\cite{Pitzl:2000wz}. In the forward direction (${7^\circ < \theta < 30^\circ}$), additional coverage is provided by the forward tracking detector, a set of planar drift chambers. The tracking detectors are operated in a solenoidal magnetic field of 1.16~T. For offline track reconstruction, track helix parameters are fitted to the inner detector hits in a general broken lines fit~\cite{Blobel:2006yi}. The procedure considers multiple scattering and corrects for energy loss by ionisation in the detector material. The primary vertex position is calculated from all tracks and optimised as part of the fitting procedure. Transverse track momenta are measured with a resolution of ${\sigma(p_T)/p_T \simeq 0.002~p_T/{\ensuremath{{\mathrm{GeV}}}}\xspace \oplus 0.015}$. The CJCs also provide a measurement of the specific energy loss of charged particles by ionisation ${\ensuremath{{\dd E/\dd x}}}\xspace$ with a relative resolution of 6.5\% for long tracks.
The tracking detectors are surrounded by the liquid argon (LAr) sampling calorimeter~\cite{Andrieu:1993kh}. It provides coverage in the region ${4^\circ < \theta < 154^\circ}$ and over the full azimuthal angle. The inner electromagnetic section of the LAr is interlaced with lead, the outer hadronic section with steel absorbers. With the LAr, the energies of electromagnetic and hadronic showers are measured with a precision of ${\sigma(E)/E \simeq 12\%/\sqrt{E/{\ensuremath{{\mathrm{GeV}}}}\xspace} \oplus 1\%}$ and ${\sigma(E)/E \simeq 50\%/\sqrt{E/{\ensuremath{{\mathrm{GeV}}}}\xspace}\oplus 2\%}$, respectively~\cite{Andrieu:1994yn}. In the backward region \linebreak (${153^\circ < \theta < 178^\circ}$), energies are measured with a spaghetti calorimeter (SpaCal) of lead absorbers interlaced with scintillating fibres~\cite{Appuhn:1996na}.
Detector components positioned in the very forward direction are used in this analysis to identify proton dissociation events. These are the forward muon detectors (FMD), the PLUG calorimeter and the forward tagging system (FTS). The lead-scintillator plug calorimeter is positioned around the beampipe at $z=4.9~{\ensuremath{{\mathrm{m}}}}\xspace$ to measure the energies of particles in the pseudorapidity region $3.5 < \eta < 5.5$. The FMD is a system of six drift chambers positioned outside of the LAr and covering the range $1.9 < \eta < 3.7$. Particles at larger pseudorapidity up to $\eta \lesssim 6.5$ can still induce spurious signals via secondary particles produced in interactions with the beam transport system and detector support structures~\cite{Phd:Dirkmann}. The very forward region, $6.0 < \eta < 7.5$, is covered by an FTS station of scintillation detectors positioned around the beampipe at $z=28~{\ensuremath{{\mathrm{m}}}}\xspace$.
The H1 trigger is operated in four stages. The first trigger level (L1) is implemented in dedicated hardware reading out fast signals of selected sub-detector components. Those signals are combined and refined at the second level (L2). A third, software-based level (L3) combines L1 and L2 information for partial event reconstruction. After full detector read-out and full event reconstruction, events are subject to a final software-based filtering (L4). The data used for the present analysis are recorded using mainly information from the fast track trigger (FTT)~\cite{Baird:2001xc}. The FTT makes it possible to measure transverse track parameters at the first trigger level and complete three-dimensional tracks at L2. This is achieved through applying pattern recognition and associative memory technology to identify predefined tracks in the hit-patterns produced by charged particles in a subset of the CJC signal wires.
The instantaneous luminosity is measured by H1 with a dedicated photon detector located close to the beampipe at $z=-103~{\ensuremath{{\mathrm{m}}}}\xspace$. With it, the rate of the Bethe-Heitler process $ep \rightarrow ep\gamma$ is monitored. The integrated luminosity is measured more precisely with the main H1 detector using the elastic QED Compton process. In this process, the electron and photon in the $ep\gamma$ final state have large transverse momenta and can be reconstructed in a back-to-back topology in the SpaCal. The integrated luminosity has been measured with a total uncertainty of $2.7\%$~\cite{Aaron:2012kn} that is dominated by systematic effects.
\subsection{Data sample}
\label{sec:exp_dataSample}
The present analysis is based on data collected by the H1 experiment during the 2006/2007 HERA running period. In that period, the accelerator was operated with positrons having an energy of $E_e = 27.6~{\ensuremath{{\mathrm{GeV}}}}\xspace$ and protons with an energy of $E_p = 920~{\ensuremath{{\mathrm{GeV}}}}\xspace$. Due to bandwidth limitations, only a subset of the H1 dataset is available for the trigger conditions relevant for this analysis, corresponding to an effective integrated luminosity of $\mathcal{L}_\mathrm{int} = 1.3~{\ensuremath{{\mathrm{pb}^{-1}}}}\xspace$. In the kinematic range considered in this analysis, the pions from ${\ensuremath{{\rho^0}}}\xspace \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace$ photoproduction are produced within the acceptance of the CJC and with low transverse momenta $p_T \lesssim 0.5~{\ensuremath{{\mathrm{GeV}}}}\xspace$. In the diffractive photoproduction regime, both the outgoing proton and electron avoid detection by escaping through the beampipe\footnote{In the studied energy region, the elastically scattered protons are mostly outside of the acceptance region of the H1 forward proton spectrometer (FPS) and the very forward proton spectrometer (VFPS)~\cite{FPSVFPS}.}.
\subsubsection{Trigger}
A dedicated, track-based {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction trigger condition was used for online event selection. Track information within the \mbox{2.3~$\mu$s} decision time of the L1 trigger was available through the FTT. For a positive trigger decision, at least two FTT tracks above a transverse momentum threshold of 160~{\ensuremath{{\mathrm{MeV}}}}\xspace and at most three tracks above a threshold of 100~{\ensuremath{{\mathrm{MeV}}}}\xspace were required. The sum of the charges of these tracks was restricted to $\pm 1$ elementary electric charge. In addition, trigger information from the CIP was used to ensure a low multiplicity interaction within the nominal interaction region along the \mbox{$z$-axis}. Vetoes on the inner forward part of the LAr calorimeter and on a scintillator wall in the forward direction were applied to suppress non-diffractive inelastic interactions. Further SpaCal and timing vetoes rejected events from beam-gas and beam-machine interactions. To keep under control the expected rate from the large {\ensuremath{{\rho^0}}}\xspace meson production cross section, the trigger was scaled down by an average factor of $\sim100$.
\subsubsection{Event reconstruction and selection}
In order to select a sample of {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction events, a set of offline selection cuts is applied on top of the trigger requirements:
\begin{itemize}
\item The {\ensuremath{{\pi^+\pi^-}}}\xspace topology is ensured by requiring exactly two primary-vertex fitted, central tracks to be reconstructed. They need to satisfy some additional quality requirements, have opposite charge, and be within the acceptance region\footnote{The polar acceptance is reduced with respect to the CJC geometry to improve the performance of the {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction trigger and its MC simulation.} defined as $25^\circ < \theta < 155^\circ$ and \mbox{$p_T>0.16~{\ensuremath{{\mathrm{GeV}}}}\xspace$}. Low-momentum kaons, protons, and deuterons are suppressed using the difference between the measured energy loss ${\ensuremath{{\dd E/\dd x}}}\xspace$ of the tracks in the CJC and the expected loss for the respective particle hypothesis in a likelihood-based approach. The two tracks are then taken to be the pion candidates, and their 4-momentum vectors are calculated with the corresponding mass hypothesis.
\item The photoproduction kinematic regime is ensured by vetoing events with a scattered electron candidate in the SpaCal or LAr. The SpaCal acceptance then limits the photon virtuality to $Q^2\lesssim 2.5~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$.
\item The diffractive topology is ensured by requiring a large rapidity gap between the central tracks and any forward detector activity. Events with LAr clusters above a noise level of 0.6~{\ensuremath{{\mathrm{GeV}}}}\xspace in the forward region $\theta < 20^\circ$ are rejected. Information from the FTD is used to reconstruct forward tracks, and events with more than one forward track that cannot be matched to one of the central tracks are also rejected. The presence of a single unmatched track is permitted to reduce the sensitivity on the modelling of the forward energy flow in the forward detectors. This rapidity gap selection in particular also limits the mass of the proton-dissociative system to approximately ${\ensuremath{{m_Y}}}\xspace \lesssim 10~{\ensuremath{{\mathrm{GeV}}}}\xspace$.
\item Background processes with additional neutral particles or charged particles outside of the central tracker acceptance are suppressed by cuts on the LAr and SpaCal energy. LAr and SpaCal clusters above respective noise levels of 0.6~{\ensuremath{{\mathrm{GeV}}}}\xspace and 0.3~{\ensuremath{{\mathrm{GeV}}}}\xspace are geometrically matched to the two central tracks: A cluster is associated to a track if it is within a cylinder of a 60~{\ensuremath{{\mathrm{cm}}}}\xspace radius in the direction of the track upon calorimeter entry. The energies from clusters not associated to either track are summed up. Events are rejected if the total unassociated LAr or SpaCal energies exceed thresholds of 0.8~{\ensuremath{{\mathrm{GeV}}}}\xspace or 0.4~{\ensuremath{{\mathrm{GeV}}}}\xspace, respectively. This allows for a small amount of unassociated energy to account for residual noise or secondary particles produced in interactions of the pion candidates with the detector material. A further suppression of background events with additional final state particles is achieved by requiring a transverse opening angle between the two pion tracks $\Delta \phi > 50^\circ$.
\item For a reliable trigger performance and MC modelling thereof, the difference in the FTT track angles\footnote{The FTT $\phi$ angle is determined at a radial distance of $r=22~{\ensuremath{{\mathrm{cm}}}}\xspace$ from the $z$ axis.} must exceed $\Delta \phi_\mathrm{FTT} > 20^\circ$.
\item The background is further reduced by rejecting out-of-time events via cuts on the LAr and CJC event timing information. Background events from beam-gas and beam-wall interactions are suppressed by restricting the $z$ coordinate of the primary vertex to be within 25~{\ensuremath{{\mathrm{cm}}}}\xspace of the nominal interaction point.
\end{itemize}
The reaction $ep \rightarrow e {\ensuremath{{\pi^+\pi^-}}}\xspace Y$ is kinematically underconstrained since only the two pions in the final state are reconstructed. The mass of the {\ensuremath{{\pi^+\pi^-}}}\xspace system {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace is reconstructed from the 4-momenta of the two tracks under pion hypothesis. The momentum transfer at the proton vertex $t$ and the scattering energy ${\ensuremath{{W_{\gamma p}}}}\xspace$ are reconstructed from the two pion 4-momenta:
\begin{align}
\trec &= -{\ensuremath{{\left(p_{T,\pi\pi}^{\mathrm{rec}}\right)^2}}}\xspace \label{eqn:exp_trec}\\
\intertext{and}
{\ensuremath{{W_{\gamma p}^{\mathrm{rec}}}}}\xspace &= \sqrt{2E_p \left( {\ensuremath{{E_{\pi\pi}^{\mathrm{rec}}}}}\xspace - {\ensuremath{{p_{z,\pi\pi}^{\mathrm{rec}}}}}\xspace \right)}\,\text. \label{eqn:exp_wrec}
\end{align}
Here, $E_p$ denotes the proton-beam energy and ${\ensuremath{{E_{\pi\pi}^{\mathrm{rec}}}}}\xspace$, ${\ensuremath{{p_{T,\pi\pi}^{\mathrm{rec}}}}}\xspace$, and ${\ensuremath{{p_{z,\pi\pi}^{\mathrm{rec}}}}}\xspace$ are the measured energy, transverse, and longitudinal 4-momentum components of the {\ensuremath{{\pi^+\pi^-}}}\xspace system. These two equations are approximations to \eqnref{eqn:theo_Wgp} and \eqnref{eqn:theo_t}, respectively. In some regions of the probed phase space, $Q^2$ may be similar in size to $t$, or ${\ensuremath{{m_Y}}}\xspace$ may be similar in size to {\ensuremath{{W_{\gamma p}}}}\xspace, such that these approximations are poor. Such effects are corrected for in the unfolding procedure discussed later in the text (cf. \secref{sec:exp_unfolding}).
The analysis phase space probed by this measurement is explicitly defined by detector-level cuts $15 < {\ensuremath{{W_{\gamma p}^{\mathrm{rec}}}}}\xspace <90~{\ensuremath{{\mathrm{GeV}}}}\xspace$, $\trec < 3~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$, and $0.3 <{\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace < 2.3~{\ensuremath{{\mathrm{GeV}}}}\xspace$. The exclusivity requirements, which veto events with detector activity not related to the {\ensuremath{{\pi^+\pi^-}}}\xspace pair, further restrict the phase space to $Q^2 \lesssim 2.5~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$ and ${\ensuremath{{m_Y}}}\xspace\lesssim 10~{\ensuremath{{\mathrm{GeV}}}}\xspace$. The mean and median $Q^2$ in that phase space are approximately $0.02~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$ and $8\cdot10^{-6}~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$, respectively, as evaluated in the MC simulation.
A total of $943\,962$ {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction event candidates pass all selection requirements. In \figref{fig:exp_ctrl}, the selected number of events is shown as a function of {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace, {\ensuremath{{W_{\gamma p}^{\mathrm{rec}}}}}\xspace, and \trec. The distributions are compared to the MC model introduced in \secref{sec:theo_MCModel}. The {\ensuremath{{\rho^0}}}\xspace meson resonance at a mass of $\sim 770~{\ensuremath{{\mathrm{MeV}}}}\xspace$ clearly dominates the sample. Background contamination amounts to about 11\% and is investigated in the next section.
\subsubsection{Background processes}
The {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction sample includes various background contributions, even after the full event selection. The dominant background processes are the decays of diffractively produced $\omega \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace \pi^0$, $\phi\rightarrow K^+K^-$, or $\rho^\prime \rightarrow 4\pi$, as well as diffractive photon dissociation. Another source of background originates from interactions of the electron and proton beams with residual gas. Such \textit{reducible} background events are wrongly selected when charged kaons or protons are misidentified as pions or additional charged or neutral particles escape detection, e.g.,\ by being outside the central tracker acceptance or having energies below the calorimeter noise threshold. In addition to the {\ensuremath{{\rho^0}}}\xspace meson, some other vector mesons also decay to an exclusive {\ensuremath{{\pi^+\pi^-}}}\xspace state (cf. \tabref{tab:theo_diffVMSamples}). Rather than being treated as \textit{irreducible} background, these are included in the signal for the analysis of the {\ensuremath{{\pi^+\pi^-}}}\xspace production cross section.
To study the reducible background contributions in more detail, multiple dedicated control regions are introduced:
\begin{itemize}
\item $\omega(782)$ mesons predominantly decay into the ${\ensuremath{{\pi^+\pi^-}}}\xspace \pi^0$ final state. The $\pi^0$ meson can be identified via energy deposits in the calorimeters that are not associated with either of the two pion tracks. An $\omega$ control region is defined by replacing the empty calorimeter selection by a cut $E_\mathrm{LAr}^\mathrm{!assoc} > 0.8~{\ensuremath{{\mathrm{GeV}}}}\xspace$ on the unassociated energy deposited in the LAr. Events with an $\omega$ meson are distinguished from those with a {\ensuremath{{\rho^\prime}}}\xspace meson by cuts on ${\ensuremath{{m_{\pi\pi}}}}\xspace < 0.55~{\ensuremath{{\mathrm{GeV}}}}\xspace$ and on the invariant mass of both tracks (assumed to be pions) and all unassociated clusters $m_\mathrm{evt} < 1.2~{\ensuremath{{\mathrm{GeV}}}}\xspace$. The $\omega$ meson purity achieved in this region is roughly 54\%.
\item $\phi(1020)$ mesons predominantly decay into pairs of charged kaons. A $\phi$ control region is defined by replacing the $\ensuremath{{\mathrm{d}}} E/\ensuremath{{\mathrm{d}}} x$ pion identification selection cuts by a kaon selection and requiring the invariant $K^+K^-$ mass to be within 15~{\ensuremath{{\mathrm{MeV}}}}\xspace of the $\phi$ meson mass. Also the cut on the opening angle between the two tracks at the vertex is removed. The $\phi$ meson purity achieved in this region is roughly 89\%.
\item The excited ${\ensuremath{{\rho^\prime}}}\xspace$ meson dominantly decay into 4 pions in various charge configurations. Due to the track veto in the trigger, additional tracks cannot be used to identify ${\ensuremath{{\rho^\prime}}}\xspace \rightarrow 4\pi$ events. Instead, unassociated energy deposits in the LAr $E_\mathrm{LAr}^\mathrm{!assoc} > 0.8~{\ensuremath{{\mathrm{GeV}}}}\xspace$ are required in place of the empty LAr signal selection. Events with a ${\ensuremath{{\rho^\prime}}}\xspace$ meson are distinguished from those with an $\omega$ meson by requiring $m_\mathrm{evt}>1.2~{\ensuremath{{\mathrm{GeV}}}}\xspace$. The ${\ensuremath{{\rho^\prime}}}\xspace$ meson purity achieved in this region is roughly 48\%.
\item Particles from photon dissociation emerge primarily in the backwards direction. A photon dissociation control region is thus defined by replacing the empty SpaCal signal requirement by a cut $ 4 < E_\mathrm{SpaCal}^\mathrm{!assoc} < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace$ on the unassociated energy deposit in the SpaCal. The lower cut removes $\omega$ and ${\ensuremath{{\rho^\prime}}}\xspace$ events, the upper cut is retained as a veto against the scattered electron. The photon dissociation purity achieved in this region is roughly 78\%.
\end{itemize}
For the {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross section measurement, reducible background processes are subtracted in the unfolding procedure, where the templates for the respective diffractive background processes are taken from the DIFFVM MC samples, as described in \secref{sec:theo_MCModel}. The respective normalisation factors are determined by making use of the control regions in the unfolding process. The residual beam-gas induced background is modelled in a data driven approach, using events from electron and proton {\em pilot} bunches. For electron (proton) pilot bunches, there is a corresponding gap in the proton (electron) beam bunch structure, such that only interactions with rest-gas atoms may occur. The beam gas background shape predictions estimated from pilot bunch events are scaled to match the integrated beam current of the colliding bunches. The beam-gas induced background amounts to about 2\%.
\subsubsection{Proton dissociation tagging}
\label{sec:forwardTagging}
\noindent
In proton-dissociative events, the proton remnants are produced in the very forward direction where the H1 detector is only sparsely instrumented. Consequently, the remnants cannot be fully reconstructed, and elastic and proton-dissociative scattering cannot be uniquely identified on an event-by-event basis. However, in many cases some of the remnants do induce signals in the forward instruments, either directly or via secondary particles that are produced in interactions with the detector or machine infrastructure. These signals are used to tag proton-dissociative events (\textit{tagging fraction}). At a much lower level, such signals can also be present in elastic scattering events (\textit{mistagging fraction}), e.g.,\ in the presence of detector noise or when the elastically scattered proton produces secondaries in interactions with the beam transport system. By simulating the respective tagging and mistagging fractions of proton-dissociative and elastic scattering, the respective proton-dissociative and elastic contributions to the cross section can be determined.
The forward detectors used for tagging in this analysis are the FMD, the PLUG, and the FTS. An event is considered to be tagged by the FMD if there is at least one hit in any of the first three FMD layers, by the PLUG if there is more than one cluster above a noise level of 1.2~{\ensuremath{{\mathrm{GeV}}}}\xspace, or by the FTS if it produces at least one hit. A small contribution to the FTS signal, induced by secondary particles produced by the elastically scattered proton hitting a beam collimator, is discarded by applying acceptance cuts depending on $\trec$ and the location of hits in the FTS~\cite{Phd:Abolz}.
The tagging information from the three detectors is combined by summing the total number of tags in an event: $ 0 \leq N_\mathrm{tags} \leq 3$. In turn, three tagging categories are defined: a zero-tag ($N_\mathrm{tag} = 0$), single-tag ($N_\mathrm{tag} = 1$), and multi-tag ($N_\mathrm{tag} \geq 2$) category. The respective tagging fractions for events passing the standard selection cuts are shown in \figref{fig:exp_ctrl_tagEff_ptsq} as a function of \trec. The tagging categories are used to split the dataset into 3 tagging control regions. The zero-tag region is dominated by $\sim 90\%$ elastic events, in the single-tag region their fraction is reduced to $\sim 64\%$, and in the multi-tag region proton dissociation dominates at $\sim 91\%$.
\subsection{Unfolding}
\label{sec:exp_unfolding}
An unfolding procedure is applied to correct binned reconstructed detector-level distributions for various detector effects. The unfolding corrects the data for reducible background contributions, the finite resolution of reconstructed variables, and efficiency and acceptance losses. Furthermore, it is set up to separate elastic and proton-dissociative scattering events. This makes possible the determination of elastic and proton-dissociative particle level distributions from which the corresponding differential {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections are derived. The cross sections are measured in a fiducial phase space that is slightly smaller than the analysis phase space defined by the event selection. This makes it possible to account for contributions migrating into and out of the fiducial phase space. The fiducial and analysis phase space cuts are summarised in \tabref{tab:exp_fiducialPS}.
\begin{table}\centering
\small
\renewcommand{\arraystretch}{1.2}
\setlength{\tabcolsep}{3pt}
\begin{tabular}{@{}c@{\hspace{5em}} c@{}}
\toprule
\multicolumn{1}{@{}l}{Analysis phase space} & \multicolumn{1}{@{}l}{Fiducial measurement phase space} \\
\midrule
\begin{tabular}[t]{ @{}r c c c r l@{} }
15.0 &$<$ & ${\ensuremath{{W_{\gamma p}}}}\xspace$ & $<$ & 90.0 & {\ensuremath{{\mathrm{GeV}}}}\xspace \\
& & $|t|$ & $<$ & 3.0 & {\ensuremath{{\mathrm{GeV}^2}}}\xspace \\
0.3 &$<$ & ${\ensuremath{{m_{\pi\pi}}}}\xspace$ & $<$ & 2.3 & {\ensuremath{{\mathrm{GeV}}}}\xspace\\
& & $Q^2$ & $<$ & 2.5 & {\ensuremath{{\mathrm{GeV}^2}}}\xspace \\
& & \multirow{2}{*}{${\ensuremath{{m_Y}}}\xspace$} & \multirow{2}{*}{$<$} & \multirow{2}{*}{10.0} & \multirow{2}{*}{{\ensuremath{{\mathrm{GeV}}}}\xspace}\\
\end{tabular}
&
\begin{tabular}[t]{ @{}l r c c c r l@{}}
&20.0 &$<$ & ${\ensuremath{{W_{\gamma p}}}}\xspace$ & $<$ & 80.0 & {\ensuremath{{\mathrm{GeV}}}}\xspace \\
& & & $|t|$ & $<$ & 1.5 & {\ensuremath{{\mathrm{GeV}^2}}}\xspace \\
&0.5 &$<$ & ${\ensuremath{{m_{\pi\pi}}}}\xspace$ & $<$ & 2.2 & {\ensuremath{{\mathrm{GeV}}}}\xspace\\
& & & $Q^2$ & $<$ & 2.5 & {\ensuremath{{\mathrm{GeV}^2}}}\xspace\\
\multicolumn{1}{@{}l}{elastic:} & & & ${\ensuremath{{m_Y}}}\xspace$ & $=$ & $m_p$ \\
\multicolumn{1}{@{}l}{$p$-dissociative:} & $m_p$ & $<$ & ${\ensuremath{{m_Y}}}\xspace$ & $<$ & 10.0 & {\ensuremath{{\mathrm{GeV}}}}\xspace \\
\end{tabular}\\
\bottomrule
\end{tabular}
\caption{Analysis and fiducial measurement phase space. At detector level, the respective cuts are applied to {\ensuremath{{W_{\gamma p}^{\mathrm{rec}}}}}\xspace, \trec, and {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace, the $Q^2$ cut is replaced by the veto on the reconstruction of the scattered electron, and the ${\ensuremath{{m_Y}}}\xspace$ cut is replaced by the rapidity gap requirement as detailed in the text. }
\label{tab:exp_fiducialPS}
\end{table}
\subsubsection{Regularised template fit}
Differential cross section measurements are performed for various distributions of the variables ${\ensuremath{{m_{\pi\pi}}}}\xspace$, {\ensuremath{{W_{\gamma p}}}}\xspace, and $t$, and combinations thereof. The beam-gas background template is subtracted from the considered data distribution and the result is used as input to the unfolding. The unfolding is performed by means of a regularised template fit within the TUnfold framework~\cite{Schmitt:2012kp}. A response matrix $\bm A$ is introduced, with elements $A_{ij}$ describing the probability that an event generated in bin $j$ of a truth-level distribution $\vec{x}$ is reconstructed in bin $i$ of a detector-level distribution $\vec{y}$.
In the definition of the response matrix, elastic and proton-dissociative signal and background MC processes are included as dedicated sub-matrices. This makes it possible to separate the elastic and proton-dissociative signal components in the unfolding. Also, background subtraction is implicitly performed during the unfolding where the normalisation of the backgrounds is determined in the fit. At detector level, the response matrix is split into signal and background control regions as defined above. The signal region is further split into three and the background regions into two orthogonal forward tagging categories. This constrains the respective MC contributions in the template fit. Migrations into and out of the fiducial phase space are considered by including side bins in each sub-matrix both at detector and at truth level~\cite{Phd:Abolz}. These contain events passing the analysis phase space cuts but failing the fiducial cuts (cf. \tabref{tab:exp_fiducialPS}).
As an illustration, the response matrix used for unfolding the one-dimensional {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace distributions is given in \figref{fig:exp_respMatrix_mpipi}. For all response matrices, most truth bins are found to have good constraints by at least one reconstruction level bin with a purity above 50\%. As a result, the statistical correlations between bins of the unfolded one-dimensional distributions are small. The somewhat larger statistical correlations in the unfolding of multi-dimensional distributions are reduced by introducing a Tikhonov regularisation term on the curvature into the template fit~\cite{Schmitt:2012kp,tikonov}. The truth-level MC distributions, that are adjusted to data as explained in \secref{sec:theo_MCModel}, are used as a bias in the regularisation. However, regularisation is not used in the unfolding of the one-dimensional {\ensuremath{{m_{\pi\pi}}}}\xspace distributions. This makes possible the measurement of the $\omega$ meson mass and production cross section, which were found to be very sensitive to any sort of regularisation.
\subsubsection{{\ensuremath{{\pi^+\pi^-}}}\xspace cross section definition}
\label{sec:exp_xSecPiPi}
Differential {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections are calculated from the number of unfolded {\ensuremath{{\pi^+\pi^-}}}\xspace events. The double-differential cross section in {\ensuremath{{m_{\pi\pi}}}}\xspace and $t$ is defined as
\begin{equation}
\left[\dfrac{\ensuremath{{\mathrm{d}}}^{2} \sigma(\gamma p \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace Y) ({\ensuremath{{m_{\pi\pi}}}}\xspace, t; {\ensuremath{{W_{\gamma p}}}}\xspace)}{ \ensuremath{{\mathrm{d}}} t \, \ensuremath{{\mathrm{d}}} {\ensuremath{{m_{\pi\pi}}}}\xspace} \right]_j =
\dfrac{ \hat{x}_j }{ \Delta t \, \Delta {\ensuremath{{m_{\pi\pi}}}}\xspace} \dfrac{1}{ \mathcal{L}\, \ensuremath{{\Phi_{\gamma/e}}}\xspace\left({\ensuremath{{W_{\gamma p}}}}\xspace\right) }
\label{eqn:exp_xSecDef}
\end{equation}
in bin $j$ of an unfolded distribution and with $\hat x_j$ being the number of unfolded events. Differentiation is approximated by dividing the event yields by the {\ensuremath{{m_{\pi\pi}}}}\xspace and $t$ bin widths, $\Delta t$ and $\Delta {\ensuremath{{m_{\pi\pi}}}}\xspace$, respectively. No bin centre correction is performed for the variables {\ensuremath{{m_{\pi\pi}}}}\xspace and $t$. Instead, predictions or fit functions are integrated over the respective bin width when comparing to data. The event yields are normalised by the integrated $ep$ luminosity $\mathcal{L}$. Photoproduction cross sections are calculated by normalising the $ep$ cross sections by means of the integrated effective photon flux $\ensuremath{{\Phi_{\gamma/e}}}\xspace$, that is calculated for the relevant {\ensuremath{{W_{\gamma p}}}}\xspace ranges according to \eqnref{eqn:theo_intFlux}.
\subsection{Cross section uncertainties}
\label{sec:systematics}
The measured cross sections are subject to statistical and systematic uncertainties. Statistical uncertainties originate from the limited size of the available data and MC samples. Systematic uncertainties originate from the MC modelling of the signal and background processes and their respective kinematic distributions, as well as from the simulation of the detector response. Detector uncertainties are estimated by either varying the simulated detector response to MC events or by modifying the event selection procedure simultaneously in both the MC samples and data. Model uncertainties are estimated by varying the generation and reweighting parameters of the MC samples.
\subsubsection{Statistical uncertainties}
Two sources of statistical uncertainties are considered: uncertainties of the data distributions and uncertainties of the unfolding factors obtained from the MC samples. The data uncertainties include contributions from the subtraction of the beam-gas background. Statistical uncertainties are propagated through the unfolding as described in the TUnfold documentation~\cite{Schmitt:2012kp}. The finite detector resolution imposes correlations on the statistical uncertainties after the unfolding.
\subsubsection{Experimental systematic uncertainties}
In order to estimate uncertainties related to the knowledge of the H1 detector response, the unfolding is repeated with systematically varied response matrices. Two types of variations are considered for different sources of uncertainty. Variations of the MC samples on detector level are introduced and are propagated to the cross sections by repeating the unfolding with varied response matrices. Other uncertainties are estimated by modifying the selection procedure. Such variations affect both the migration matrices and the input data distributions simultaneously. For two-sided variations, the full absolute shifts between the nominal and varied unfolded distributions are used to define up and down uncertainties. For one-sided variations, the full shift is taken as a two-sided uncertainty~\cite{Phd:Abolz}.
Concerning the trigger, the following uncertainties are considered:
\begin{itemize}
\item The trigger scale factors for the CIP and FTT correction are varied to account for statistical and shape uncertainties. A further uncertainty is estimated to cover the propagation of the correction factors from the electroproduction regime, where they are derived, to photoproduction~\cite{Phd:Abolz}.
\item The trigger scale factors for the correction of the forward vetoes are varied to account for statistical and shape uncertainties~\cite{Phd:Abolz}.
\item Pions in the very backward direction may potentially interfere with the SpaCal trigger veto via secondary particles produced in {\em nuclear interactions}. Nuclear interactions of the pions with the beam transport system and detector material are not modelled perfectly by the simulation. To estimate a potential impact, the cut on the upper track polar angle acceptance is varied between $150^\circ$ and $160^\circ$.
\end{itemize}
Concerning the track reconstruction and simulation, the following uncertainties are considered:
\begin{itemize}
\item The uncertainty of the modelling of the geometric CJC acceptance in $p_T$ is estimated by varying the acceptance cut to $p_T >0.18~{\ensuremath{{\mathrm{GeV}}}}\xspace$.
\item The uncertainty for potential mismodelling of the forward tracking detectors is estimated by varying the veto on the number of forward tracks to allow either none or up to two tracks in the selection.
\item The uncertainty of the MC $z$-vertex distribution is estimated by varying the MC distribution to cover small discrepancies in the mean and tails relative to the observed $z$-vertex distribution in data.
\item In the particle identification, the cuts on the kaon, proton, and deuteron {\ensuremath{{\dd E/\dd x}}}\xspace rejection likelihoods are varied up and down while simultaneously varying the pion {\ensuremath{{\dd E/\dd x}}}\xspace selection likelihood down and up.
\item Inhomogeneities in the magnetic field may not be fully modelled in the simulation. Nor may the modelling of nuclear interactions of the pions with the detector material be completely accurate. These effects are estimated to result in an uncertainty on the modelled track $p_T$ resolution of 20\%. Their impact is evaluated by smearing the reconstructed track $p_{T,\mathrm{rec}} \rightarrow p_{T,\mathrm{rec}} \pm 0.2\, \left(p_{T,\mathrm{rec}}-p_{T,\mathrm{gen}} \right)$ with respect to the generated true $p_{T,\mathrm{gen}}$ in the signal {\ensuremath{{\pi^+\pi^-}}}\xspace MC samples. The result is propagated to all kinematic variables reconstructed from the two pion 4-momenta.
\item Similarly, a 20\% uncertainty is assumed on the resolution of the track $\theta$ measurement to cover potential mismodellings in the simulation.
\end{itemize}
Concerning the track momentum scale, the following uncertainties are considered:
\begin{itemize}
\item The field strength of the H1 solenoid is known to a level of 0.3\%~\cite{Abt:1996xvhi} or better. To estimate a potential impact on the absolute track $p_T$ scale, the reconstructed MC $p_T$ values are varied up and down by $\pm 0.3\%$. The variation is performed simultaneously for both tracks.
\item The energy loss correction depends on a precise knowledge of the material budget, which is known at a level of 7\%. Through the energy loss correction in the track reconstruction, this uncertainty results in an average track $p_T$ uncertainty of $0.4~{\ensuremath{{\mathrm{MeV}}}}\xspace$.
\end{itemize}
Concerning the calorimeters, the following uncertainties are considered:
\begin{itemize}
\item The noise cuts on LAr and SpaCal energy clusters are independently varied up and down by $\pm 0.2~{\ensuremath{{\mathrm{GeV}}}}\xspace$ and $\pm 0.1~{\ensuremath{{\mathrm{GeV}}}}\xspace$, respectively.
\item The energy scales of the LAr and SpaCal clusters are independently varied by $\pm 10\%$. These variations are applied prior to the respective cluster noise cut.
\item A mismodelling of nuclear interactions may also affect the calorimeter response. A potential impact is estimated by varying the cut applied for associating calorimeter clusters to tracks by $\pm 10~{\ensuremath{{\mathrm{cm}}}}\xspace$.
\end{itemize}
Concerning the forward detectors, the following tagging uncertainties are considered:
\begin{itemize}
\item The tagging and mistagging fractions of the forward detectors are varied independently. The FTS tagging fraction is varied by $\pm 5\%$ in the proton-dissociative MC samples and the mistagging fraction by $\pm 50\%$ in the elastic samples. Similarly, these fractions in the FMD are varied by $\pm 5\%$ and $\pm 5\%$ and in the PLUG by $\pm 5\%$ and $\pm 100\%$, respectively. The sizes of the variations are estimated to cover observed mismodellings in the tagging fractions of the respective detectors.
\end{itemize}
\subsubsection{Model uncertainties}
In order to estimate model uncertainties, the MC samples are modified on generator-level. Model uncertainties are propagated to the cross sections by repeating the unfolding with the varied migration matrices and MC bias distributions. Cross section uncertainties are then calculated from the shifted unfolded spectra in the same manner as for the experimental uncertainties. Generally, the unfolding approach is rather insensitive to many aspects of the MC modelling. The following model variations are considered:
\begin{itemize}
\item Uncertainties on the $Q^2$ and ${\ensuremath{{m_Y}}}\xspace$ dependencies of the MC samples are estimated~\cite{Alexa:2013xxa}. The $Q^2$ dependence of the MC samples is varied by applying an event weight ${(1+Q^2/m_{\ensuremath{{\mathrm{VM}}}\xspace}^2)^{\pm 0.07}}$ in agreement with experimental uncertainties~\cite{Aaron:2009xp}. The ${\ensuremath{{m_Y}}}\xspace$ dependence of the proton-dissociative sample is varied by applying a weight ${(1/{\ensuremath{{m_Y^2}}}\xspace)^{\pm 0.15}}$.
\item The tuning parameters of the MC samples derived to describe the kinematic data distributions (\secref{sec:theo_MCModel}) are independently varied up and down~\cite{Phd:Abolz}.
\item For the ${\ensuremath{{\rho^\prime}}}\xspace$ background MC samples, an uncertainty on the relative ratio of $\rho(1450)$ and $\rho(1700)$ is estimated by varying it from $1{:}1$ up and down to $2{:}1$ and $1{:}2$, \linebreak respectively. An uncertainty on the ${\ensuremath{{\rho^\prime}}}\xspace$ decay modes is estimated by varying \linebreak ${{\ensuremath{{\mathcal{BR}}}}\xspace({\ensuremath{{\rho^\prime}}}\xspace \rightarrow \rho^\pm \pi^\mp \pi^0) = (50\pm25)\%}$ while simultaneously scaling all other decay modes proportionally.
\item A shape uncertainty on the mass distribution of the photon-dissociative mass is estimated by reweighting the distribution by $(1/m_{X}^2)^{\pm 0.15}$.
\end{itemize}
\subsubsection{Normalisation uncertainties}
Normalisation uncertainties are directly applied to the unfolded cross sections. The following uncertainties are considered:
\begin{itemize}
\item The uncertainty of the track reconstruction efficiency due to modelling of the detector material is 1\% per track, leading to a 2\% normalisation uncertainty.
\item The integrated luminosity of the used dataset is known at a precision of 2.7\%.
\item Higher order QED effects have been estimated to be smaller than 2\%~\cite{Kurek:1996ez} which is taken as a normalisation uncertainty.
\end{itemize}
This results in a combined 3.9\% normalisation uncertainty on the measured cross sections.
\subsection{Cross section fits}
\label{sec:exp_fits}
For the analysis, various parametric models are fitted to the measured cross section distributions. Fits are performed by varying model parameters to minimise a $\chi^2$ function. Parametrisations of differential distributions are integrated over the respective kinematic bins and divided by their bin widths in order to match the differential cross section definition (cf. \eqnref{eqn:exp_xSecDef}). The $\chi^2$ function takes into account only the statistical uncertainties, represented by the covariance matrix. Nominal fit parameters are obtained by fitting the nominal measured cross section distributions. The corresponding $\chi^2$ value relative to the number of degrees of freedom $\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace$ is used as a measure of the quality of a fit. Systematic uncertainties are then propagated through a fit via an \textit{offset method}, where each systematic cross section variation is fitted independently. The resulting shifts in the fit parameters relative to the nominal parameters are used to define the systematic parameter uncertainties.
\section{Results}
\label{sec:results}
\noindent
Elastic and proton-dissociative {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections are measured. The measurement is presented integrated over the fiducial phase space and as one-, two-, and three-dimensional cross section distributions as functions of {\ensuremath{{m_{\pi\pi}}}}\xspace, {\ensuremath{{W_{\gamma p}}}}\xspace, and $t$.
Subsequently, {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections as functions of {\ensuremath{{W_{\gamma p}}}}\xspace and $t$ are extracted via fits to the invariant {\ensuremath{{\pi^+\pi^-}}}\xspace mass distributions (cf.~\eqnref{eqn:theo_rhoSoedMass}).
The procedure is exemplified with the differential {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections {\ensuremath{{\dd \sigma(\gamma p \rightarrow \pipi Y)/\dd \mpipi}}}\xspace as functions of {\ensuremath{{m_{\pi\pi}}}}\xspace. One- and two-dimensional {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross section distributions are then parametrised and interpreted using fits. In particular, a Regge fit of the elastic {\ensuremath{{\rho^0}}}\xspace meson production cross section as a function of {\ensuremath{{W_{\gamma p}}}}\xspace and $t$ makes possible the extraction of the effective leading Regge trajectory $\alpha(t)$.
\subsection{Fiducial {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections}
\label{sec:res_sigmaPiPiFid}
\noindent
The fiducial {\ensuremath{{\pi^+\pi^-}}}\xspace electroproduction cross section is measured in the phase space defined in \tabref{tab:exp_fiducialPS} by unfolding one-dimensional {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace distributions, integrating the event yield over all bins, and normalising the result by the integrated luminosity. It is subsequently turned into a photoproduction cross section by normalisation to the integrated effective photon flux. In the phase space considered, the integrated effective flux calculated with \eqnref{eqn:theo_intFlux} is \ensuremath{{\Phi_{\gamma/e}}}\xspace = 0.1368\text.
This yields a {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross section of
\begin{equation}
\ensuremath{{\sigma(\gamma p \rightarrow \pipi Y)}}\xspace = 16.20 \pm 0.05\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+1.11}_{-1.15}\ (\ensuremath{{\mathrm{syst.}}}\xspace)\ \ensuremath{{\mu\mathrm{b}}}\xspace \text, \ \text{ for }\ m_p \leq {\ensuremath{{m_Y}}}\xspace < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace \text.
\end{equation}
The cross section is measured at an average energy $\langle {\ensuremath{{W_{\gamma p}}}}\xspace \rangle \simeq 43~{\ensuremath{{\mathrm{GeV}}}}\xspace$ as estimated from the MC simulation. The elastic and proton-dissociative components of the cross section are separated through the unfolding which yields
\begin{alignat}{3}
\ensuremath{{\sigma(\gamma p \rightarrow \pipi p)}}\xspace =&&\ 11.&52 &&\pm 0.06\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.76}_{-0.78}\ (\ensuremath{{\mathrm{syst.}}}\xspace)\ \ensuremath{{\mu\mathrm{b}}}\xspace \text{ and} \\
\ensuremath{{\sigma(\gamma p \rightarrow \pipi Y)}}\xspace =&&\ 4.&68 &&\pm 0.06\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.62}_{-0.64}\ (\ensuremath{{\mathrm{syst.}}}\xspace)\ \ensuremath{{\mu\mathrm{b}}}\xspace \text, \ \text{ for }\ m_p < {\ensuremath{{m_Y}}}\xspace < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace \text,
\end{alignat}
respectively\footnote{The fiducial cross sections can be obtained by alternatively unfolding the $\trec$, {\ensuremath{{W_{\gamma p}^{\mathrm{rec}}}}}\xspace, or multi-dimensional distributions. The variations of the results are similar in size to the statistical uncertainties. Comparable statistical and systematic uncertainties are observed.}. The uncertainties of the two components are correlated with a statistical and symmetrised total Pearson correlation coefficient of $\rho_\mathrm{stat} = -0.59$, and $\rho_\mathrm{tot} = +0.30$, respectively. The measurements are statistically very precise at a level of 0.5\% (1.2\%) but have large systematic uncertainties of 6.3\% (13.2\%) for the elastic (proton-dissociative) component. The uncertainty of the elastic component is dominated by the trigger and the normalisation uncertainties of 4.1\% and 3.9\%, respectively, whereas the uncertainties associated with the tagging and the calorimeter dominate the proton-dissociative component at 8.4\% and 7.3\%, respectively. A more detailed breakdown of the cross section uncertainties is given in \tabref{tab:res_fidUncGroups}.
\begin{table}[!ttt]\centering
\small
\begin{tabular}{@{}l@{\hspace{4em}} c@{\hspace{5em}} c@{}} \\
\toprule
& \multicolumn{2}{c}{Relative cross section uncertainty [\%]} \\
\cmidrule{2-3}
Source of uncertainty & ${{{\ensuremath{{m_Y}}}\xspace} = m_p}$ & {${m_p < {{\ensuremath{{m_Y}}}\xspace} < 10~{{\ensuremath{{\mathrm{GeV}}}}\xspace}}$}\\
\midrule
Statistical & 0.5 & 1.2 \\[6pt]
Trigger & 4.1 & 5.3 \\
Tracking & 1.4 & 1.3 \\
Momentum scale & 0.1 & 0.1 \\
Calorimeter & 1.5 & 7.3 \\
Tagging & 2.0 & 8.4 \\[6pt]
Normalisation & 3.9 & 3.9 \\[6pt]
MC model (${\ensuremath{{m_Y}}}\xspace,Q^2,$ bgr.) & 2.0 & 2.7 \\
MC model (${\ensuremath{{m_{\pi\pi}}}}\xspace,{\ensuremath{{W_{\gamma p}}}}\xspace,t$) & 0.1 & 0.4 \\
\midrule
Total & 6.6 & 13.3 $\ $ \\
\bottomrule
\end{tabular}
\caption{Summary of the combined impact of systematic uncertainties on the fiducial {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections. The numbers are obtained by unfolding the one-dimensional {\ensuremath{{m_{\pi\pi}}}}\xspace distribution and with symmetrised systematic uncertainties.}
\label{tab:res_fidUncGroups}
\end{table}
Photoproduction of {\ensuremath{{\pi^+\pi^-}}}\xspace mesons has a fairly large cross section and forms a sizeable contribution to the total $\gamma p$ cross section, which is known for similar {\ensuremath{{W_{\gamma p}}}}\xspace from cosmic ray experiments~\cite{Vereshkov:2003cp}. Within the restrictions of the fiducial phase space in {\ensuremath{{m_{\pi\pi}}}}\xspace, $t$, and {\ensuremath{{m_Y}}}\xspace, the elastic and proton-dissociative processes contribute to the total cross section by about 9\% and 4\%, respectively. The elastic {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross section has been measured before in $ep$ collisions at slightly higher {\ensuremath{{W_{\gamma p}}}}\xspace \cite{Aid:1996bs,Breitweg:1997ed}. Considering the differences in {\ensuremath{{W_{\gamma p}}}}\xspace and the energy dependence of the elastic cross section (see below), the measurements are found to be consistent. There has not been a previous measurement of the proton-dissociative {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross section in a comparable phase space.
\subsection{Differential {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections}
\label{sec:res_sigmaRhoFid}
\noindent
The elastic and proton-dissociative differential {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections \linebreak {\ensuremath{{\dd \sigma(\gamma p \rightarrow \pipi Y)/\dd \mpipi}}}\xspace are measured as functions of {\ensuremath{{m_{\pi\pi}}}}\xspace by unfolding one-dimensional {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace distributions. Numerical results are given in \apxref{apx:sec_pipiXSecTables} and the cross section distributions are displayed in \figref{fig:results_dSigma_dm}. The photoproduction of {\ensuremath{{\pi^+\pi^-}}}\xspace mesons is dominated by the {\ensuremath{{\rho^0}}}\xspace meson resonance peaking at a mass of around 770~{\ensuremath{{\mathrm{MeV}}}}\xspace. The differential cross sections fall off steeply towards higher masses with a second broad excited $\rho^\prime$ meson resonance appearing at around 1600~{\ensuremath{{\mathrm{MeV}}}}\xspace. At the {\ensuremath{{\rho^0}}}\xspace meson resonance peak, the proton-dissociative cross section is around 40\% of the elastic cross section with the difference becoming smaller towards higher masses. The uncertainties of the measurement vary slowly with {\ensuremath{{m_{\pi\pi}}}}\xspace. At the {\ensuremath{{\rho^0}}}\xspace meson peak the statistical and systematic uncertainty of the elastic differential cross section are roughly 1\% and 7\%, respectively. At the position of the $\rho^\prime$ peak they have grown to 7\% and 13\%, respectively. The corresponding uncertainties of the differential proton-dissociative cross section are about twice as large.
\subsubsection{Fit of the {\ensuremath{{m_{\pi\pi}}}}\xspace dependence}
In order to extract the {\ensuremath{{\rho^0}}}\xspace meson contributions to the {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections, the {\ensuremath{{m_{\pi\pi}}}}\xspace lineshape defined in \eqnref{eqn:theo_rhoSoedMass} is fitted to the data.
As this paper focusses on the analysis of {\ensuremath{{\rho^0}}}\xspace meson production, the fit is only performed in the analysis range $0.6~{\ensuremath{{\mathrm{GeV}}}}\xspace \leq {\ensuremath{{m_{\pi\pi}}}}\xspace \leq 1.0~{\ensuremath{{\mathrm{GeV}}}}\xspace$ where contributions from excited $\rho^\prime$ meson states can be neglected. The parametrisation is fitted simultaneously to the elastic and proton-dissociative distributions with the parameter settings described in \tabref{tab:theo_SodingPars}. The model describes the data well with the fit yielding $\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace= 24.6/24$.
The resulting model parameters are presented in \tabref{tab:results_dSigma_dm_pars}.
The fitted curves are shown in \figref{fig:results_dSigma_dm_fit}, where they are compared to the measured data.
\begin{table}[!thb]\centering
\small
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{@{}l@{\hspace{4em}} d{3.4} d{1.4} l c@{\hspace{4em}} d{3.4} d{1.4} l@{}}
\toprule
Shared parameter & \multicolumn{1}{l}{Value} & \multicolumn{1}{l}{$\Delta_{\ensuremath{{\mathrm{stat.}}}\xspace}$} & \multicolumn{1}{l}{$\Delta_{\ensuremath{{\mathrm{syst.}}}\xspace}$} &\\
\midrule
$ m_{\rho}~[{\ensuremath{{\mathrm{MeV}}}}\xspace]$ & 770.8 & 1.3 &$ ^{+2.3}_{-2.4}$ \\
$ \Gamma_{\rho}~[{\ensuremath{{\mathrm{MeV}}}}\xspace]$ & 151.3 & 2.2 &$ ^{+1.6}_{-2.8}$ \\
$ m_{\omega}~[{\ensuremath{{\mathrm{MeV}}}}\xspace]$ & 777.9 & 2.2 &$ ^{+4.3}_{-2.2}$ \\
$ \Gamma_{\omega}~[{\ensuremath{{\mathrm{MeV}}}}\xspace]$ & 8.5 & \multicolumn{2}{l}{PDG fixed} \\
$ \delta_{nr}$ & 0.76 & 0.35 &$ ^{+0.14}_{-0.08}$ \\
\midrule
& \multicolumn{3}{c}{${\ensuremath{{m_Y}}}\xspace = m_p$} && \multicolumn{3}{c}{$m_p < {\ensuremath{{m_Y}}}\xspace < 10~{{\ensuremath{{\mathrm{GeV}}}}\xspace}$} \\
\cmidrule{2-4} \cmidrule{6-8}
Independent parameter & \multicolumn{1}{l}{Value} & \multicolumn{1}{l}{$\Delta_{\ensuremath{{\mathrm{stat.}}}\xspace}$} & \multicolumn{1}{l}{$\Delta_{\ensuremath{{\mathrm{syst.}}}\xspace}$} && \multicolumn{1}{l}{Value} & \multicolumn{1}{l}{$\Delta_{\ensuremath{{\mathrm{stat.}}}\xspace}$} & \multicolumn{1}{l}{$\Delta_{\ensuremath{{\mathrm{syst.}}}\xspace}$}\\
\midrule
$ A~[\ensuremath{{\mu\mathrm{b}}}\xspace/{\ensuremath{{\mathrm{GeV}}}}\xspace^{2}]$ & 48.4 & 0.8 & $^{+3.3}_{-3.2}$ && 20.8 & 0.4 & $^{+2.7}_{-2.6}$\\
$ f_{\omega}$ & 0.166 & 0.017 & $^{+0.008}_{-0.023}$ && 0.135 & 0.042 & $^{+0.042}_{-0.036}$\\
$ \phi_{\omega}$ & -0.53 & 0.22 & $^{+0.21}_{-0.17}$ && -0.02 & 0.34 & $^{+0.31}_{-0.19}$\\
$ f_{nr}$ & 0.189 & 0.026 & $^{+0.025}_{-0.016}$ && 0.145 & 0.025 & $^{+0.029}_{-0.014}$\\
$ \Lambda_{nr}~[{\ensuremath{{\mathrm{GeV}}}}\xspace]$ & 0.18 & 0.59 & $^{+0.20}_{-0.10}$ && 0.1 & 1.3 & $^{+0.4}_{-0.3}$\\
\bottomrule
\end{tabular}\\
\caption{%
Free parameters for the fit of the single-differential elastic and proton-dissociative cross section {\ensuremath{{\dd \sigma(\gamma p \rightarrow \pipi Y)/\dd \mpipi}}}\xspace as a function of {\ensuremath{{m_{\pi\pi}}}}\xspace.
The nominal fit values and statistical and systematic uncertainties are given. A full uncertainty breakdown and statistical correlations are provided~\cite{H1_data}. The corresponding fit is shown in \figref{fig:results_dSigma_dm_fit}.}
\label{tab:results_dSigma_dm_pars}
\end{table}
Within the fit model, the differential {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections obtain significant contributions from the {\ensuremath{{\rho^0}}}\xspace meson resonance and non-resonant {\ensuremath{{\pi^+\pi^-}}}\xspace production.
The two corresponding amplitudes give rise to large interference terms that result in a skewing of the resonance lineshapes. At the measured {\ensuremath{{\rho^0}}}\xspace meson mass, the relative contributions of the {\ensuremath{{\rho^0}}}\xspace resonance and non-resonant {\ensuremath{{\pi^+\pi^-}}}\xspace production to the elastic differential {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross section are $[79.1 \pm 3.2~(\ensuremath{{\mathrm{stat.}}}\xspace)~{}^{+4.4}_{-1.4}~(\ensuremath{{\mathrm{syst.}}}\xspace)]\%$ and $[7.1 \pm 0.6~(\ensuremath{{\mathrm{stat.}}}\xspace)~{}^{+0.7}_{-0.5}~(\ensuremath{{\mathrm{syst.}}}\xspace)]\%$, respectively. The corresponding contributions to the proton-dissociative cross section are $[85.9 \pm 3.8~(\ensuremath{{\mathrm{stat.}}}\xspace)~{}^{+3.8}_{-3.2}~(\ensuremath{{\mathrm{syst.}}}\xspace)]\%$ and $[4.8 \pm 0.6~(\ensuremath{{\mathrm{stat.}}}\xspace)~{}^{+0.6}_{-0.5}~(\ensuremath{{\mathrm{syst.}}}\xspace)]\%$, respectively. Most noticeably, the relative non-resonant contribution is smaller for proton-dissociative than for elastic scattering resulting in a weaker skewing. In the analysed mass range, the non-resonant differential cross section exhibits little dependence on {\ensuremath{{m_{\pi\pi}}}}\xspace. For this reason the two fit parameters $\delta_{\ensuremath{{\mathrm{nr}}}\xspace}$ and $\Lambda_{\ensuremath{{\mathrm{nr}}}\xspace}$ are strongly correlated and cannot be constrained very well individually.
Qualitatively the result of the fit is similar to past HERA {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction analyses~\cite{Breitweg:1997ed}.
Quantitative comparisons of fit parameters are not possible because different parametrisations were used.
Variants of the fit were also explored but did not significantly improve the compatibility with data. When including a possible admixture of incoherent background, the respective contribution is found to be compatible with zero. This conclusion is largely independent of the assumed incoherent background shape. Models including barrier factors \cite{VonHippel:1972fg} also can be fitted well by the data. Obviously, all extracted particle masses and widths are model dependent. A detailed discussion of these effects is beyond the scope of this paper.
The rich structure of the interference term near 770~{\ensuremath{{\mathrm{MeV}}}}\xspace, visible in \figref{fig:results_dSigma_dm_fit}, originates from the $\omega$ resonance.
In the fit model, the $\omega$ resonance amplitude prevents the net interference contribution from vanishing at the {\ensuremath{{\rho^0}}}\xspace meson mass and shifts the position of the sign change of the interference term towards the $\omega$ meson mass.
In the {\ensuremath{{\pi^+\pi^-}}}\xspace production cross section this is visible as a characteristic step.
The impact of the $\omega$ meson on the measured differential cross sections is similar in size to observations made in $e^+e^- \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace$ production~\cite{ee_mrho}. In the present analysis, this effect is measured for the first time at HERA.
The masses of the ${\ensuremath{{\rho^0}}}\xspace$ and $\omega$ mesons and the width of the ${\ensuremath{{\rho^0}}}\xspace$ meson obtained from the fit are compared to the world average PDG values~\cite{PDG} in \tabref{tab:res_mgRhoComp}.
The PDG lists two sets of {\ensuremath{{\rho^0}}}\xspace parameters from {\ensuremath{{\rho^0}}}\xspace meson photoproduction and production in $e^+e^-$ scattering, respectively. The value measured for the {\ensuremath{{\rho^0}}}\xspace mass is compatible with previous photoproduction measurements~\cite{Breitweg:1997ed,Abramowicz:2011pk,Bartalucci:1977cp}, confirming a slightly lower {\ensuremath{{\rho^0}}}\xspace mass parameter as compared to typical $e^+e^-$ measurements~\cite{ee_mrho}. The mass uncertainty is dominated by the uncertainties in the magnetic field and the material in the detector affecting the track $p_T$ scale. The precision of the presented value is similar to the most precise previous photoproduction measurements. The {\ensuremath{{\rho^0}}}\xspace width is consistent with both PDG values and offers an uncertainty as good as previous photoproduction measurements. The $\omega$ mass is also compatible with the PDG value but has a sizeable uncertainty. The observed difference between the measured and the PDG $\omega$ mass is similar to that of the ${\ensuremath{{\rho^0}}}\xspace$ mass relative to the $e^+e^-$ measurement.
\begin{table}[!ttt] \centering
\small
\renewcommand{\arraystretch}{1.3}
\setlength{\tabcolsep}{1em}
\begin{tabular}{@{}l@{\hspace{3em}} c c c c@{}}
\toprule
Parameter & This measurement & PDG $\gamma p$ & PDG $e^+e^-$ & $\Delta$($e^+e^-$, H1)\\
\midrule
$ m_{{\ensuremath{{\rho^0}}}\xspace}[{\ensuremath{{\mathrm{MeV}}}}\xspace]$ & $770.8~{}^{+2.6}_{-2.7}$ & $769.0\pm 1.0$ & $775.26 \pm 0.25$ & $4.5 ^{+2.7}_{-2.6}$ \\
$ \Gamma_{{\ensuremath{{\rho^0}}}\xspace}~[{\ensuremath{{\mathrm{MeV}}}}\xspace]$ & $151.3~{}^{+2.7}_{-3.6}$ & $151.7\pm 2.6$ & $147.8\ \, \pm 0.9\ \,$ & \\
$ m_\omega~[{\ensuremath{{\mathrm{MeV}}}}\xspace]$ & $777.9~{}^{+4.8}_{-3.1}$ & & $782.65\pm 0.12$ & $4.8 ^{+3.1}_{-4.8}$ \\
\bottomrule
\end{tabular}
\caption{Comparison of measured ${\ensuremath{{\rho^0}}}\xspace$ and $\omega$ meson properties to the PDG values~\cite{PDG}. Only the total uncertainties are given. The last column gives the mass differences between the PDG $e^+e^-$ and the present measurement.}
\label{tab:res_mgRhoComp}
\end{table}
\subsubsection{Fiducial {\ensuremath{{\rho^0}}}\xspace and $\omega$ meson photoproduction cross sections}
The result of the fit is used to calculate the {\ensuremath{{\rho^0}}}\xspace meson contributions to the fiducial {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections.
This is achieved by integrating the {\ensuremath{{\rho^0}}}\xspace meson resonance amplitude in the range $2m_\pi < {\ensuremath{{m_{\pi\pi}}}}\xspace < 1.53~{\ensuremath{{\mathrm{GeV}}}}\xspace$ as discussed in \secref{sec:theo_xSecRho}. The integration yields
\begin{alignat}{3}
\ensuremath{{\sigma(\gamma p \rightarrow \rho^0 p)}}\xspace =&&\ 10.&97 &&\pm 0.18\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.72}_{-0.73}\ (\ensuremath{{\mathrm{syst.}}}\xspace)\ \ensuremath{{\mu\mathrm{b}}}\xspace \text{ and} \\
\ensuremath{{\sigma(\gamma p \rightarrow \rho^0 Y)}}\xspace =&&\ 4.&71 &&\pm 0.09\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.60}_{-0.59}\ (\ensuremath{{\mathrm{syst.}}}\xspace)\ \ensuremath{{\mu\mathrm{b}}}\xspace \text, \ \text{ for }\ m_p < {\ensuremath{{m_Y}}}\xspace < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace \text,
\end{alignat}
for the elastic and proton-dissociative ${\ensuremath{{\rho^0}}}\xspace$ meson photoproduction cross sections, respectively. The ratio of the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections in the measurement phase space is
\begin{equation}
\dfrac{\ensuremath{{\sigma(\gamma p \rightarrow \rho^0 Y)}}\xspace}{\ensuremath{{\sigma(\gamma p \rightarrow \rho^0 p)}}\xspace} = 0.429 \pm 0.009\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.056}_{-0.053}\ (\ensuremath{{\mathrm{syst.}}}\xspace) \text, \ \text{ for }\ m_p < {\ensuremath{{m_Y}}}\xspace < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace \text,
\end{equation}
taking into account correlations. The measured elastic {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross section is consistent with previous HERA measurements~\cite{Aid:1996bs,Derrick:1995vq,Breitweg:1997ed}, when accounting for {\ensuremath{{W_{\gamma p}}}}\xspace differences between the measurements. Suitable reference measurements for the proton-dissociative cross section are not available. However, the ratio of the proton-dissociative to the elastic {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross section is consistent with a previous measurement~\cite{Breitweg:1997ed}, assuming that phase space differences are covered by the large uncertainties.
With the fit model \eqnref{eqn:theo_rhoSoedMass}, also the indirect measurement of the fiducial $\omega$ meson photoproduction cross sections is possible by integrating the $\omega$-{\ensuremath{{\rho^0}}}\xspace mixing amplitude over the range $2m_\pi \leq {\ensuremath{{m_{\pi\pi}}}}\xspace \leq 0.82~{\ensuremath{{\mathrm{GeV}}}}\xspace$.
The result is corrected for the $\omega \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace$ branching fraction $\mathcal{BR}(\omega \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace) = 0.0153 \pm 0.0006\ (\ensuremath{{\mathrm{tot.}}}\xspace)$~\cite{PDG} to yield the cross sections
\begin{alignat}{3}
\sigma(\gamma p \rightarrow \omega p)
=&&\ 1.&06 &&\pm 0.26\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.13}_{-0.30}\ (\ensuremath{{\mathrm{syst.}}}\xspace)\ \ensuremath{{\mu\mathrm{b}}}\xspace \ \text{ and} \\
\sigma(\gamma p \rightarrow \omega Y)
=&&\ 0.&31 &&\pm 0.23\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.19}_{-0.16}\ (\ensuremath{{\mathrm{syst.}}}\xspace)\ \ensuremath{{\mu\mathrm{b}}}\xspace \text, \ \text{ for }\ m_p < {\ensuremath{{m_Y}}}\xspace < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace \text.
\end{alignat}
The uncertainty of the branching fraction is included in the systematic uncertainty. The ratios of the $\omega$ meson to the ${\ensuremath{{\rho^0}}}\xspace$ meson photoproduction cross sections are then determined to be
\begin{alignat}{2}
\dfrac{ \sigma(\gamma p \rightarrow \omega p)}{\ensuremath{{\sigma(\gamma p \rightarrow \rho^0 p)}}\xspace} =&\ 0.097 &&\pm 0.020\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.011}_{-0.026}\ (\ensuremath{{\mathrm{syst.}}}\xspace) \ \text{ and} \\
\dfrac{ \sigma(\gamma p \rightarrow \omega Y)}{\ensuremath{{\sigma(\gamma p \rightarrow \rho^0 Y)}}\xspace} =&\ 0.065 &&\pm 0.041\ (\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.042}_{-0.032}\ (\ensuremath{{\mathrm{syst.}}}\xspace) \text, \ \text{ for }\ m_p < {\ensuremath{{m_Y}}}\xspace < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace \text.
\end{alignat}
The elastic $\omega$ photoproduction cross section has been directly measured at HERA in the $\omega \rightarrow {\ensuremath{{\pi^+\pi^-}}}\xspace \pi^0$ channel by the ZEUS collaboration. In a similar phase space, a compatible value for the cross section was observed~\cite{Derrick:1996yt}. The present value is also consistent with the expectation that the {\ensuremath{{\rho^0}}}\xspace and $\omega$ meson photoproduction cross sections differ by a factor $c_{\omega}/c_{\rho} \simeq 1/9$ that originates from SU(2) flavour symmetry and the quark electric charges.
\subsection{Energy dependence of {\ensuremath{{\rho^0}}}\xspace meson photoproduction}
\label{sec:res_sigmaRhoOfW}
\noindent
The energy dependencies of the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections are obtained by unfolding two-dimensional distributions ${\ensuremath{{W_{\gamma p}^{\mathrm{rec}}}}}\xspace \otimes {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace$ and calculating the corresponding two-dimensional {\ensuremath{{\pi^+\pi^-}}}\xspace cross sections. Numerical results are presented in \apxref{apx:sec_pipiXSecTables} and the cross section distributions are displayed in \figref{fig:results_dSigma_dmW_wmFit_el} and \figref{fig:results_dSigma_dmW_wmFit_pd} for the elastic and proton-dissociative components, respectively.
To extract the {\ensuremath{{\rho^0}}}\xspace meson contributions, \eqnref{eqn:theo_rhoSoedMass} is fitted simultaneously to the {\ensuremath{{m_{\pi\pi}}}}\xspace distributions in every unfolded {\ensuremath{{W_{\gamma p}}}}\xspace bin with the parameter settings described in \tabref{tab:theo_SodingPars}.
In particular, the $\omega$ meson model parameters cannot be constrained well by the fit and are fixed to the values obtained in the fit to the one-dimensional mass distributions. The fit gives $\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace = 222.0/188$. The fitted curves are also shown figures \ref{fig:results_dSigma_dmW_wmFit_el} and \ref{fig:results_dSigma_dmW_wmFit_pd}.
The fit results are used to calculate the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections \ensuremath{{\sigma(\gamma p \rightarrow \rho^0 Y)}}\xspace as a function of {\ensuremath{{W_{\gamma p}}}}\xspace.
The results are given in \tabref{tab:tab_xSec_wRho} and displayed in \figref{fig:results_sigmaRho_w_fit}. The cross sections only have weak dependencies on {\ensuremath{{W_{\gamma p}}}}\xspace, as is expected for diffractive processes at such energies. Many sources of uncertainty considered here are observed to vary with {\ensuremath{{W_{\gamma p}}}}\xspace. For the elastic cross section, the full systematic uncertainty is about 6\% in the centre of the considered phase space and increases to 10\% and 8\% at the lowest and highest considered {\ensuremath{{W_{\gamma p}}}}\xspace, respectively. The corresponding proton-dissociative uncertainties are about twice as large.
\subsubsection{Fit of the energy dependence}
The energy dependencies of the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections in the range ${20 < {\ensuremath{{W_{\gamma p}}}}\xspace < 80~{\ensuremath{{\mathrm{GeV}}}}\xspace}$ are parametrised by a simple power law:
\begin{equation}
{\ensuremath{{\sigma_{\rho}}}}\xspace({\ensuremath{{W_{\gamma p}}}}\xspace) = {\ensuremath{{\sigma_{\rho}}}}\xspace(W_0)\ \left(\dfrac{{\ensuremath{{W_{\gamma p}}}}\xspace}{W_0} \right)^\delta\,.
\label{eqn:results_sigma_w_fit_fcn}
\end{equation}
The parametrisation is fitted simultaneously to the measured elastic and proton-disso\-cia\-tive distributions assuming independent elastic and proton-dissociative fit parameters. The function is evaluated at the geometric bin centres, and the reference energy is set to $W_0 = 40~{\ensuremath{{\mathrm{GeV}}}}\xspace$. The fit yields a very poor $\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace = 33.0 / 11 $, where however systematic uncertainties are not considered. The resulting fit parameters are presented in \tabref{tab:results_sigmaRho_w_pars}. The fitted curves are compared to the data in \figref{fig:results_sigmaRho_w_fit}
The fitted power parameter ${\delta_{\ensuremath{{\mathrm{el}}}\xspace}=0.171\pm0.009~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.039}_{-0.026}~(\ensuremath{{\mathrm{syst.}}}\xspace)}$ characterises the slow rise of the elastic cross section with increasing scattering energy. The measured value is compatible with previous measurements, e.g.,\ by the ZEUS collaboration~\cite{Breitweg:1997ed} or the CMS Collaboration ($pPb$)~\cite{Sirunyan:2019nog}, but is more precise. The fit results in a proton-dissociative parameter ${\delta_{\ensuremath{{\mathrm{pd}}}\xspace}=-0.156\pm0.026~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.081}_{-0.079}~(\ensuremath{{\mathrm{syst.}}}\xspace)}$ that is significantly different from the elastic parameter and suggests a decrease of the proton-dissociative cross section with increasing energy. However, the proton-dissociative cross section is expected to be strongly shaped by the restriction of the fiducial phase space; most notably by the requirement ${\ensuremath{{m_Y}}}\xspace < 10~{\ensuremath{{\mathrm{GeV}}}}\xspace$. Since energy is required to excite high masses {\ensuremath{{m_Y}}}\xspace, the cut suppresses the cross section more strongly for high than for low {\ensuremath{{W_{\gamma p}}}}\xspace. The phase space is not accounted for in \eqnref{eqn:results_sigma_w_fit_fcn} so that the proton-dissociative result $\delta_{\ensuremath{{\mathrm{pd}}}\xspace}$ cannot be interpreted directly in terms of a scattering amplitude.
In \figref{fig:results_sigmaRho_w_hist}, the measured elastic {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross section data are compared to selected measurements by fixed-target~\cite{Ballam:1971wq,Park:1971ts,Ballam:1972eq,Struczinski:1975ik,Egloff:1979mg,Aston:1982hr}, HERA~\cite{Aid:1996bs,Derrick:1995vq,Breitweg:1997ed}, and LHC~\cite{Sirunyan:2019nog} experiments\footnote{The phase space differences between the measurements with respect to the considered $t$ ranges are found to be negligible ($\lesssim 2\%$) relative to much larger overall normalisation uncertainties.}. The present elastic data are in agreement with other measurements and connect the fixed-target data with the high energy regime. The combination of data makes possible the study of the region of reggeon exchange at small scattering energies and of pomeron exchange at large energies simultaneously. For describing the energy range $2 < {\ensuremath{{W_{\gamma p}}}}\xspace < 200~{\ensuremath{{\mathrm{GeV}}}}\xspace$, the energy dependence of the elastic cross section is parametrised by the sum of two power-law functions:
\begin{equation}
{\ensuremath{{\sigma_{\rho}}}}\xspace({\ensuremath{{W_{\gamma p}}}}\xspace) = {\ensuremath{{\sigma_{\rho}}}}\xspace(W_0)\ \left( \left(\dfrac{{\ensuremath{{W_{\gamma p}}}}\xspace}{W_0} \right)^{\delta_{\ensuremath{{I\!\!P}}\xspace}} + f_{\ensuremath{{I\!\!R}}\xspace} \left(\dfrac{{\ensuremath{{W_{\gamma p}}}}\xspace}{W_0} \right)^{\delta_{\ensuremath{{I\!\!R}}\xspace}} \right).
\label{eqn:results_sigma_w_fit_fcn_DL}
\end{equation}
The two parameters $\delta_{\ensuremath{{I\!\!P}}\xspace}$ and $\delta_{\ensuremath{{I\!\!R}}\xspace}$ are associated with pomeron exchange at high and reggeon exchange at low {\ensuremath{{W_{\gamma p}}}}\xspace, respectively. The parametrisation is fitted to the HERA data~\cite{Aid:1996bs,Derrick:1995vq,Breitweg:1997ed} together with data from fixed-target experiments~\cite{Ballam:1971wq,Park:1971ts,Ballam:1972eq,Struczinski:1975ik,Egloff:1979mg,Aston:1982hr} as shown in \figref{fig:results_sigmaRho_w_hist}. In the fit, previous measurements enter with their respective uncorrelated and normalisation uncertainties. For the H1~\cite{Aid:1996bs} and ZEUS measurements~\cite{Breitweg:1997ed}, normalisation uncertainties of 8\% and 9\%, respectively, are extracted from the total uncertainties. Correlations between experiments are not considered. The uncorrelated uncertainties are accounted for in the data covariance, together with the statistical uncertainties of the present measurement. The normalisation uncertainties are included by offsetting the data and repeating the fit, as is also done for
systematic and normalisation uncertainties of the present data.
The fit yields $\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace = 84.3/43$ and results in $\delta_{\ensuremath{{I\!\!P}}\xspace} = 0.207 \pm 0.015~(\ensuremath{{\mathrm{stat.}}}\xspace) ^{+0.053}_{-0.033}~(\ensuremath{{\mathrm{syst.}}}\xspace)$ and $\delta_{\ensuremath{{I\!\!R}}\xspace} = -1.45 \pm 0.12~(\ensuremath{{\mathrm{stat.}}}\xspace) ^{+0.35}_{-0.21} ~(\ensuremath{{\mathrm{syst.}}}\xspace)$. All fit parameters are given in table \ref{tab:results_world_w_pars}.
The fitted curve is compared to the data in \figref{fig:results_sigmaRho_w_hist}. The extracted parameters are consistent with the broader picture of vector meson electroproduction, in particular with the observed dependence of $\delta_{\ensuremath{{I\!\!P}}\xspace}$ on the vector meson mass~\cite{Favart:2015umi}. The fit further suggests a small reggeon contribution of ${f_{\ensuremath{{I\!\!R}}\xspace} = [2.0 \pm 0.7~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+2.9}_{-1.3}~(\ensuremath{{\mathrm{syst.}}}\xspace)]\%}$ at $W_0=40~{\ensuremath{{\mathrm{GeV}}}}\xspace$. For this reason, the parameter $\delta_{\ensuremath{{\mathrm{el}}}\xspace}$ from the fit with a single power-law function (\tabref{tab:results_sigmaRho_w_pars}, ${\ensuremath{{m_Y}}}\xspace=m_p$) is interpreted as an effective parameter describing both pomeron and reggeon exchange contributions.
\subsection{$t$ dependence of {\ensuremath{{\rho^0}}}\xspace meson photoproduction}
\label{sec:res_dSigmaRhoOft}
\noindent
The $t$ dependencies of the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections are obtained by unfolding two-dimensional distributions $\trec \otimes {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace$ and calculating the corresponding two-dimensional {\ensuremath{{\pi^+\pi^-}}}\xspace cross sections. Numerical results are presented in \apxref{apx:sec_pipiXSecTables} and the cross section distributions are displayed in \figref{fig:results_dSigma_dmdt_mtFit_el} and \figref{fig:results_dSigma_dmdt_mtFit_pd} for the elastic and proton-dissociative component, respectively.
To extract the {\ensuremath{{\rho^0}}}\xspace meson contributions, \eqnref{eqn:theo_rhoSoedMass} is fitted simultaneously to the {\ensuremath{{m_{\pi\pi}}}}\xspace distributions in every unfolded $t$ bin with the parameter settings described in \tabref{tab:theo_SodingPars}.
The fit yields ${\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace = 353.4/247}$. The fitted curves are also shown figures \ref{fig:results_dSigma_dmdt_mtFit_el} and \ref{fig:results_dSigma_dmdt_mtFit_pd}.
The fit results are used to calculate the elastic and proton-dissociative differential {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections ${\ensuremath{{\dd \sigma(\gamma p \rightarrow \myrho Y)/\dd t}}}\xspace$ as a function of $t$.
The results are given in \tabref{tab:tab_xSec_tRho} and displayed in \figref{fig:results_dSigmaRho_dt_fit}. The cross sections fall of exponentially with $t$ as is expected for diffractive processes. The relative systematic uncertainties of roughly 6\% for the elastic distribution vary little with $t$ and increase moderately to 11\% in the highest $|t|$ bin. For the proton-dissociative distribution, the uncertainties increase from roughly 13\% at intermediate $|t|$ to 17\% in both the highest and lowest considered $|t|$ bins.
\subsubsection{Fit of the $t$ dependence}
The $t$ dependencies of the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections are parametrised by the function
\begin{align}
\dfrac{\ensuremath{{\mathrm{d}}} {\ensuremath{{\sigma_{\rho}}}}\xspace}{\ensuremath{{\mathrm{d}}} t}(t) = \dfrac{\ensuremath{{\mathrm{d}}} {\ensuremath{{\sigma_{\rho}}}}\xspace}{\ensuremath{{\mathrm{d}}} t}(t=0)\ \left(1 - \dfrac{ b\, t}{a} \right)^{-a} \text.
\label{eqn:results_dSigma_dt_fit_fcn}
\end{align}
It interpolates between an exponential $\ensuremath{{\mathrm{d}}} {\ensuremath{{\sigma_{\rho}}}}\xspace/\ensuremath{{\mathrm{d}}} t \propto \exp\left(b\, t \right) $ at low $|t|$ and a power-law $ \ensuremath{{\mathrm{d}}} {\ensuremath{{\sigma_{\rho}}}}\xspace / \ensuremath{{\mathrm{d}}} t \propto |t|^{-a}$ at large $|t|$. The function is fitted simultaneously to the elastic and proton-dissociative distributions assuming independent elastic and proton-dissociative model parameters. To match the cross section definition, the function is integrated over each bin and divided by the respective bin width in the fit. The proton-dissociative distribution is affected by phase space restrictions that are not accounted for in the parametrisation. The impact is particularly large at small $|t|$ because excitations with large ${\ensuremath{{m_Y}}}\xspace$ require a minimal\footnote{ $|t| \gtrsim \dfrac{{\ensuremath{{m_Y^2}}}\xspace}{W_{{\ensuremath{{\gamma p}}}\xspace}^4} \left(m_{\pi\pi}^2+Q^2\right)^2$~\cite{Abramowicz:1999eq}} $|t|$. For this reason, the lowest $|t|$ bin of the proton-dissociative distribution is not included in the fit. The fit yields ${\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace = 15.3 / 14 }$. The resulting fit parameters are presented in \tabref{tab:results_dSigmaRho_dt_pars}. The fitted curves are compared to data in \figref{fig:results_dSigmaRho_dt_fit}
The exponential $t$ dependencies of the elastic and proton-dissociative cross sections at small $|t|$ are quantified by the parameters ${b_{\ensuremath{{\mathrm{el}}}\xspace} = 9.61\pm0.15~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.20}_{-0.15}~(\ensuremath{{\mathrm{syst.}}}\xspace)~{\ensuremath{{\mathrm{GeV}^{-2}}}}\xspace}$ and ${b_{\ensuremath{{\mathrm{pd}}}\xspace} = 4.81\pm0.24~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.39}_{-0.37}~(\ensuremath{{\mathrm{syst.}}}\xspace)~{\ensuremath{{\mathrm{GeV}^{-2}}}}\xspace}$, respectively. With $b_{\ensuremath{{\mathrm{pd}}}\xspace} \sim 1/2\, b_{\ensuremath{{\mathrm{el}}}\xspace}$, the elastic spectrum falls off much more steeply than the proton-dissociative spectrum. The difference between the exponential slope parameters is $b_{\ensuremath{{\mathrm{el}}}\xspace} - b_{\ensuremath{{\mathrm{pd}}}\xspace} = 4.80 \pm 0.29~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.41}_{-0.40}~(\ensuremath{{\mathrm{syst.}}}\xspace)$, taking into account correlations. In the optical interpretation (or in an eikonal model approach~\cite{eikonalModel}), this difference reflects the difference in the size of the respective target, i.e.,\ proton in elastic vs parton in proton-dissociative interactions. The measured value is consistent with HERA measurements in {\ensuremath{{\rho^0}}}\xspace meson electroproduction~\cite{Aaron:2009xp}. There is evidence for a deviation from the purely exponential behaviour given by the finite values for $a_{\ensuremath{{\mathrm{el}}}\xspace} = 20.4\pm3.7~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+6.8}_{-5.1}~(\ensuremath{{\mathrm{syst.}}}\xspace)$ and $a_{\ensuremath{{\mathrm{pd}}}\xspace} = 8.5\pm1.7~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+2.7}_{-2.1}~(\ensuremath{{\mathrm{syst.}}}\xspace)$. The smaller $a_{\ensuremath{{\mathrm{pd}}}\xspace}$ indicates a stronger deviation for the proton-dissociative than for the elastic cross section. The present measurements of the $t$ dependencies of the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections can be compared with other HERA measurements~\cite{Aid:1996bs,Breitweg:1997ed}. The measured values for the elastic slope $b_{\ensuremath{{\mathrm{el}}}\xspace}$ are somewhat lower than observed previously but seem compatible within uncertainties when taking into account the differences in energy and the thus expected shrinkage. The present value is the most precise of these measurements. The observed deviation of the elastic slope from the exponential behaviour is compatible with the ZEUS measurement \cite{Breitweg:1997ed}. Similar to the elastic case, the slopes $b_{\ensuremath{{\mathrm{pd}}}\xspace}$ in proton-dissociative scattering are compatible with earlier measurements. The present measurement is more precise and does account for deviations from the exponential form. Differences in the definition of the proton-dissociative phase space were not considered in these comparisons.
\subsection{{\ensuremath{{\rho^0}}}\xspace meson photoproduction as a function of $t$ and {\ensuremath{{W_{\gamma p}}}}\xspace}
\label{sec:res_dSigmaRhoOfWt}
\noindent
The two-dimensional $t$ and {\ensuremath{{W_{\gamma p}}}}\xspace dependencies of the elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections are obtained by unfolding three-dimensional distributions $\trec \otimes {\ensuremath{{W_{\gamma p}^{\mathrm{rec}}}}}\xspace \otimes {\ensuremath{{m_{\pi\pi}^\mathrm{rec}}}}\xspace$ and calculating the corresponding {\ensuremath{{\pi^+\pi^-}}}\xspace cross sections\footnote{The underlying response matrix has 1243 bins on detector and 882 bins on truth level.}. Numerical results are presented in \apxref{apx:sec_pipiXSecTables} and the cross section distributions are displayed in \figref{fig:results_dSigma_dmdtW_twmFit_el} and \figref{fig:results_dSigma_dmdtW_twmFit_pd} for the elastic and proton-dissociative component, respectively.
To extract the {\ensuremath{{\rho^0}}}\xspace meson contributions, \eqnref{eqn:theo_rhoSoedMass} is fitted simultaneously to the {\ensuremath{{m_{\pi\pi}}}}\xspace distributions in every unfolded {\ensuremath{{W_{\gamma p}}}}\xspace and $t$ bin with the parameter settings described in \tabref{tab:theo_SodingPars}.
The fit yields $\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace = 804.0/607$. The fitted curves are also shown in figures \ref{fig:results_dSigma_dmdtW_twmFit_el} and \ref{fig:results_dSigma_dmdtW_twmFit_pd}.
The fit results are used to calculate the elastic and proton-dissociative differential {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections ${\ensuremath{{\dd \sigma(\gamma p \rightarrow \myrho Y)/\dd t}}}\xspace$ as a function of $t$ and {\ensuremath{{W_{\gamma p}}}}\xspace.
The results are given in \tabref{tab:tab_xSec_twRho_el} and \tabref{tab:tab_xSec_twRho_pd} for the elastic and proton-dissociative component, respectively, and are shown in \figref{fig:results_dSigmaRho_dtW_twfit}.
\subsubsection{Regge fit of the $t$ and ${\ensuremath{{W_{\gamma p}}}}\xspace$ dependence}
The elastic and proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections are parametrised by the function
\begin{equation}
\dfrac{\ensuremath{{\mathrm{d}}} {\ensuremath{{\sigma_{\rho}}}}\xspace}{\ensuremath{{\mathrm{d}}} t}(t;{\ensuremath{{W_{\gamma p}}}}\xspace) = \dfrac{\ensuremath{{\mathrm{d}}} {\ensuremath{{\sigma_{\rho}}}}\xspace}{\ensuremath{{\mathrm{d}}} t}(t;W_0) \left( \dfrac{{\ensuremath{{W_{\gamma p}}}}\xspace}{W_0} \right)^{4 (\alpha(t)-1) } \text,
\label{eqn:sigma_walpha_fcn}
\end{equation}
with an energy dependence predicted by Regge theory and with a single leading trajectory $\alpha(t)$. The $t$ dependence at the reference energy $W_0=40~{\ensuremath{{\mathrm{GeV}}}}\xspace$ is parametrised according to \eqnref{eqn:results_dSigma_dt_fit_fcn}. Regge trajectories are expected to continue linearly only in the region of small $|t|$. Since the analysis extends to $|t| \leq 1.5~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$, a modified parametrisation based on a Fermi function is used:
\begin{equation}
\alpha(t) = \alpha_0 + \beta\, \left( \left(\ensuremath{{\mathrm{e}}}^{ - \frac{4\alpha_1 t}{\beta} }+1\right)^{-1} -\dfrac{1}{2}\right)\text .
\label{eqn:sigma_alpha_fermi}
\end{equation}
It approximates a linear function $\alpha_0 + \alpha_1\, t$ for small $|t|$ and approaches a constant value ${\alpha_0 - \beta/2}$ for ${t \rightarrow -\infty}$. Such a lower bound is expected for large $|t|$ from QCD calculations~\cite{BoundReggeTraj} but deviations from a linear behaviour have not been observed yet in this reaction in the $t$ range considered here. The parametrisation in \eqnref{eqn:sigma_walpha_fcn} is fitted simultaneously to the elastic and proton-dissociative cross sections assuming independent elastic and proton-dissociative fit parameters. In the $\chi^2$ calculation, the function is integrated over each $t$ bin and evaluated at the geometric {\ensuremath{{W_{\gamma p}}}}\xspace bin centres to match the cross section definition. Following the argumentation discussed above, the respective lowest $|t|$ bins of the proton-dissociative cross sections are excluded from the fit. The fit yields $\ensuremath{{\chi_\mathrm{stat}^2/\mathrm{n}_\mathrm{dof}}}\xspace =31.7/32 $. The resulting fit parameters are presented \tabref{tab:results_fPar_wtRhoFit}. The fitted curves are compared to data in \figref{fig:results_dSigmaRho_dtW_twfit}.
\subsubsection{Shrinkage of the elastic differential cross section}
The fit parameters in \tabref{tab:results_fPar_wtRhoFit} that describe the $t$ dependencies of the differential {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections at $W_0=40~{\ensuremath{{\mathrm{GeV}}}}\xspace$ are in agreement with the parameters from the one-dimensional fit of the $t$ dependencies that is described above. For the elastic differential cross section, the finite slope of the measured trajectory $\alpha(t)$ (see below) results in a shrinkage of the forward peak with increasing {\ensuremath{{W_{\gamma p}}}}\xspace. At small $|t|$, where the differential cross section ${\ensuremath{{\dd \sigma(\gamma p \rightarrow \myrho p)/\dd t}}}\xspace$ falls exponentially with $t$ and the trajectory is linear, the shrinkage can be expressed in terms of a {\ensuremath{{W_{\gamma p}}}}\xspace dependence of the exponential slope parameter
\begin{equation}
b_{\ensuremath{{\mathrm{el}}}\xspace}({\ensuremath{{W_{\gamma p}}}}\xspace) = b_{\ensuremath{{\mathrm{el}}}\xspace}\left(W_0\right) + 4\alpha_{1} \log\left(\dfrac{{\ensuremath{{W_{\gamma p}}}}\xspace}{W_0} \right)\text.
\end{equation}
The function is plotted in \figref{fig:results_fPar_wtRho_b} using the parameters extracted for the elastic cross section in the two-dimensional fit (cf. \tabref{tab:results_fPar_wtRhoFit}, $m_Y{=}m_p$). For comparison, $b({\ensuremath{{W_{\gamma p}}}}\xspace)$ is measured in every {\ensuremath{{W_{\gamma p}}}}\xspace bin by fitting the parametrisation \eqnref{eqn:results_dSigma_dt_fit_fcn} with free fit parameters $b_t$ to the $t$ dependencies in all {\ensuremath{{W_{\gamma p}}}}\xspace bins. Deviations from the exponential form are accounted for by including a single parameter $a$ common to all for {\ensuremath{{W_{\gamma p}}}}\xspace bins.
The result is presented in \tabref{tab:tab_fPar_bRho} and also shown in \figref{fig:results_fPar_wtRho_b}. Further data from previous HERA~\cite{Aid:1996bs,Derrick:1995vq,Breitweg:1997ed} and fixed-target~\cite{Ballam:1971wq,Park:1971ts,Ballam:1972eq,Struczinski:1975ik,Egloff:1979mg,Aston:1982hr} measurements are included in the figure, as well. The present slope values are somewhat lower than those from previous HERA measurements but are much more precise. The present data alone clearly indicate that there is shrinkage of the elastic peak with increasing energy.
\subsubsection{Effective leading Regge trajectory}
For visualisation of the Regge trajectory parameters, $\alpha(t)$ is measured separately in each $t$ bin by fitting a simple power law $\propto {\ensuremath{{W_{\gamma p}}}}\xspace^{4(\alpha_t - 1)}$ with free fit parameters $\alpha_t$ to the {\ensuremath{{W_{\gamma p}}}}\xspace distributions in all $t$ bins. The resulting parameters $\alpha_t$ are presented in \tabref{tab:tab_fPar_alphaRho} and displayed as a function of the $t$ bin centres $t_{bc}$\footnote{
The bin centres $t_{bc}$ are defined to satisfy $\frac{\ensuremath{{\mathrm{d}}}\sigma}{\ensuremath{{\mathrm{d}}} t}\left(t_{bc}\right) = \frac{1}{\left(t_\mathrm{max}-t_\mathrm{min}\right)} \int_{t_\mathrm{min}}^{t_\mathrm{max}} \frac{\ensuremath{{\mathrm{d}}} \sigma}{\ensuremath{{\mathrm{d}}} t}\left(t\right)\, \ensuremath{{\mathrm{d}}} t$ and calculated using \eqnref{eqn:results_dSigma_dt_fit_fcn} and the fit parameters for the elastic and proton-dissociative cross-sections given in \tabref{tab:results_dSigmaRho_dt_pars}.
} in \figref{fig:results_fPar_wtRho_alpha}. The data are compared to the parametrisations $\alpha(t)$ obtained in the two-dimensional fit and to other curves as discussed below. A direct fit of a parametrisation $\alpha(t)$ to the so extracted $\alpha_t$ is ambiguous because of the definition of the bin centres $t_{bc}$.
For the elastic cross section, the measured trajectory is linear at small $|t|$. The intercept \linebreak and slope parameters are measured at ${\alpha_{0}=1.0654 \pm 0.0044~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.0088}_{-0.0050}~(\ensuremath{{\mathrm{syst.}}}\xspace)}$ and \linebreak ${\alpha_{1} = 0.233 \pm 0.064~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.020 }_{-0.038 }~(\ensuremath{{\mathrm{syst.}}}\xspace) ~{\ensuremath{{\mathrm{GeV}^{-2}}}}\xspace}$, respectively. Potential non-linearities start occurring for $t\lesssim -0.2~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$ and the trajectory is compatible with approaching a constant asymptotic value of approximately $0.98 \pm 0.04~(\ensuremath{{\mathrm{tot.}}}\xspace)$ for $t \rightarrow -\infty$. However, a linear trajectory over the full considered $t$ range is also able to describe the elastic cross section. The resulting intercept and slope parameters are ${\alpha_{0,\mathrm{lin}} = 1.0624 \pm 0.0033~(\ensuremath{{\mathrm{stat.}}}\xspace)~{}^{+0.0083}_{-0.0052}~(\ensuremath{{\mathrm{syst.}}}\xspace)}$ and $\alpha_{1,\mathrm{lin}} = 0.175 \pm 0.026~(\ensuremath{{\mathrm{stat.}}}\xspace)~{}^{+0.021}_{-0.027}~(\ensuremath{{\mathrm{syst.}}}\xspace)~{\ensuremath{{\mathrm{GeV}^{-2}}}}\xspace$, respectively. For comparison, the linear trajectory is also included in \figref{fig:results_fPar_wtRho_alpha}.
In contrast, the proton-dissociative cross section is not compatible with a linear trajectory $\alpha(t)$. The fitted parametrisation falls off steeply at low $|t|$ but is constant with a value of approximately ${0.93 \pm 0.03~(\ensuremath{{\mathrm{tot.}}}\xspace)}$ for $t \lesssim -0.2~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$. However, significant shaping effects are expected from the phase space restrictions, as discussed earlier. No attempt is made here to relate these fit parameters with the underlying amplitude.
In the high energy limit, the elastic cross section is expected to be governed by the soft pomeron trajectory. The canonical pomeron trajectory has been determined by Donnachie and Landshoff from $pp$ and $p\bar{p}$ scattering data as $\alpha_\mathrm{DL}(t) = 1.0808 + 0.25\,t/{\ensuremath{{\mathrm{GeV}^2}}}\xspace$~\cite{Donnachie:1983hf}. The trajectory parameters have been investigated in various vector meson photo- and electroproduction measurements. A recent analysis of HERA data with a two tensor pomeron model furthermore has lead to a precise measurement of the soft pomeron intercept using inclusive DIS cross sections at low Bjorken-$x$ and total photoproduction cross sections~\cite{Britzger:2019lvc}. In {\ensuremath{{\rho^0}}}\xspace meson photoproduction, the trajectory has previously been directly extracted~\cite{Breitweg:1999jy} in an analysis of the energy dependence of the differential cross section {\ensuremath{{\dd \sigma(\gamma p \rightarrow \myrho p)/\dd t}}}\xspace in the range $8.2 < {\ensuremath{{W_{\gamma p}}}}\xspace < 94~{\ensuremath{{\mathrm{GeV}}}}\xspace$ as measured by various experiments~\cite{Aston:1982hr,Aid:1996bs,Breitweg:1997ed,Derrick:1996vw,Breitweg:1999jy}. The present analysis has the advantage that it measures the leading trajectory from a single dataset. In \figref{fig:results_alphaPom_comp}, the measured intercept and slope at $t=0$ are compared to the canonical pomeron trajectory parameters and those measured by the cited works~\cite{Britzger:2019lvc,Breitweg:1999jy}. The measured intercept seems to be comparatively low. The analysis of the one-dimensional energy dependence (\secref{sec:res_sigmaRhoOfW}) suggests the presence of a subleading reggeon contribution that indeed may have lowered the measured effective trajectory by about $0.01$ units. For the measured slope, good agreement is observed with the canonical DL value. However, there is some tension with previous HERA measurements, such as the cited ZEUS measurement. A potential explanation could arise from possible deviations from the linear form of the trajectory, which are taken into account in the present analysis. For proton-dissociative {\ensuremath{{\rho^0}}}\xspace meson photoproduction, a constant trajectory has been previously measured at HERA for higher $|t|$ than are considered here~\cite{Chekanov:2002rm}.
\section{Summary}
\label{sec:Summary}
Elastic and proton-dissociative {\ensuremath{{\pi^+\pi^-}}}\xspace and {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections are measured as a function of the invariant {\ensuremath{{\pi^+\pi^-}}}\xspace mass {\ensuremath{{m_{\pi\pi}}}}\xspace, the scattering energy {\ensuremath{{W_{\gamma p}}}}\xspace, and the squared momentum transfer at the proton vertex $t$. The measurement covers a phase space of $0.5 < {\ensuremath{{m_{\pi\pi}}}}\xspace < 2.2~{\ensuremath{{\mathrm{GeV}}}}\xspace$, $20 < {\ensuremath{{W_{\gamma p}}}}\xspace < 80~{\ensuremath{{\mathrm{GeV}}}}\xspace$, and $|t|<1.5~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$. The cross sections are obtained by unfolding up to three-dimensional {\ensuremath{{\pi^+\pi^-}}}\xspace distributions. In the procedure, the elastic and proton-dissociative components are extracted simultaneously. The results are more precise than previous measurements, in particular the measurement of the proton-dissociative cross sections.
The {\ensuremath{{\pi^+\pi^-}}}\xspace mass spectra are analysed with a model in which the skewing of the {\ensuremath{{\rho^0}}}\xspace meson resonance peak is attributed to interference with non-resonant {\ensuremath{{\pi^+\pi^-}}}\xspace production, as originally proposed by S\"oding in 1966.
A fit of the model to the data yields ${m_\rho = 770.8 \pm 1.3~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+2.3}_{-2.4}~(\ensuremath{{\mathrm{syst.}}}\xspace) ~{\ensuremath{{\mathrm{MeV}}}}\xspace}$ and ${\Gamma_\rho = 151.3 \pm 2.2~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+1.6}_{-2.8}~(\ensuremath{{\mathrm{syst.}}}\xspace) ~{\ensuremath{{\mathrm{MeV}}}}\xspace}$ for the mass and width of the {\ensuremath{{\rho^0}}}\xspace meson, respectively. For the first time at HERA, the sensitivity of the data is sufficient to constrain the $\omega$ meson contribution to {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction and measure the $\omega$ meson mass at $m_\omega = 777.9 \pm 2.2~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+4.3}_{-2.2}~(\ensuremath{{\mathrm{syst.}}}\xspace) ~{\ensuremath{{\mathrm{MeV}}}}\xspace$.
The fit is used to extract the {\ensuremath{{\rho^0}}}\xspace meson contributions to the elastic and proton-dissociative {\ensuremath{{\pi^+\pi^-}}}\xspace photoproduction cross sections.
One- and two-dimensional {\ensuremath{{\rho^0}}}\xspace meson photoproduction cross sections as functions of {\ensuremath{{W_{\gamma p}}}}\xspace and $t$ are presented and subsequently interpreted with phenomenological models in the context of Regge theory. Precise parameters describing the $t$ and {\ensuremath{{W_{\gamma p}}}}\xspace dependencies are measured. The slope parameters describing the exponential drop of the cross sections with $t$ are measured at an average energy of $\langle {\ensuremath{{W_{\gamma p}}}}\xspace \rangle = 43~{\ensuremath{{\mathrm{GeV}}}}\xspace$ as ${b_{\ensuremath{{\mathrm{el}}}\xspace} = 9.61\pm0.15~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.20}_{-0.15}~(\ensuremath{{\mathrm{syst.}}}\xspace)~{\ensuremath{{\mathrm{GeV}^{-2}}}}\xspace}$ and
$b_{\ensuremath{{\mathrm{pd}}}\xspace} = 4.81\pm0.24~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.39}_{-0.37}~(\ensuremath{{\mathrm{syst.}}}\xspace)~{\ensuremath{{\mathrm{GeV}^{-2}}}}\xspace$ for the elastic and proton-dissociative components, respectively. From the analysis of the elastic cross section {\ensuremath{{\dd \sigma(\gamma p \rightarrow \myrho p)/\dd t}}}\xspace as a function of {\ensuremath{{W_{\gamma p}}}}\xspace and $t$ the effective leading Regge trajectory is extracted. Allowing for deviations from the linear form at large $|t|$, an intercept of $\alpha(t{=}0) = 1.0654 \pm 0.0044~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.0088}_{-0.0050}~(\ensuremath{{\mathrm{syst.}}}\xspace)$ and a slope of $\alpha^\prime(t{=}0) = 0.233 \pm 0.064~(\ensuremath{{\mathrm{stat.}}}\xspace)\ {}^{+0.020 }_{-0.038 }~(\ensuremath{{\mathrm{syst.}}}\xspace) ~{\ensuremath{{\mathrm{GeV}^{-2}}}}\xspace$ are measured. Since the probed energies are low, the intercept measurement may be obscured somewhat by sub-leading reggeon contributions that cannot be constrained by the present data alone. There are some indications that the trajectory deviates from a linear behaviour for $t < -0.2~{\ensuremath{{\mathrm{GeV}^2}}}\xspace$. Within uncertainties, the data are compatible with the trajectory becoming equal to unity for large $|t|$.
\section*{Acknowledgements}
We are grateful to the HERA machine group whose outstanding efforts have made this experiment possible. We thank the engineers and technicians for their work in constructing and maintaining the H1 detector, our funding agencies for financial support, the DESY technical staff for continual assistance and the DESY directorate for support and for the hospitality which they extend to the non DESY members of the collaboration. We would like to give credit to all partners contributing to the EGI computing infrastructure for their support for the H1 Collaboration.
We express our thanks to all those involved in securing not only the H1 data but also the software and working environment for long term use allowing the unique H1 dataset to continue to be explored in the coming years. The transfer from experiment specific to central resources with long term support, including both storage and batch systems has also been crucial to this enterprise. We therefore also acknowledge the role played by DESY-IT and all people involved during this transition and their future role in the years to come.
| 1ea1e1e4f4a2fcad0bac7205b744587c09d16342 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Conclusions}
\bibliographystyle{IEEEtran}
\subsection{Dynamic program to compute $\mathbf F(\cdot)$}
\label{app:F_compute}
We here describe how to compute $\mathbf F(p,v)$ in $O(mn)$ time and space complexity, where $p=(p_1,...,p_n)$ and $v=v_1...v_m$, via a dynamic programming approach.
We first define
\begin{align*}
&\mathbf G^{for}(k,j)\triangleq \mathbf F(p_{[1:k]},v_{[1:j]}).
\numberthis
\label{eq:G}
\end{align*}
Using Lemma~\ref{lemma:F_decomposition} with $i=n$, we get
\begin{align*}
\mathbf F(p,v) = \mathbf F(p_{[n-1]},v) + p_n^{v_m} (1-p_n)^{(1-v_m)} \mathbf F(p_{[n-1]},v_{[m-1]}).
\end{align*}
\noindent This translates to the following dynamic program for $\mathbf G^{for}$:
\begin{align*}
\mathbf G^{for}(k,j) = \mathbf G^{for}(k-1,j)+ p_{k}^{v_j}(1-p_{k})^{1-v_j}&\mathbf G^{for}(k-1,j-1),\numberthis \label{eq:approx_smap_dpfor}
\end{align*}
with the boundary conditions $\mathbf G^{for}(k,0)=1\ \forall\ k \geq 0$ and $\mathbf G^{for}(k,j)=0\ \forall\ k<j$.
The algorithm is now summarized as Alg.~\ref{alg:F_comp}.
\begin{algorithm}[h!]
\caption{Computing $\mathbf F(p,v)$}\label{alg:F_comp}
\begin{algorithmic}[1]
\State Inputs: $p \in [0,1]^n$, $v \in \{0,1\}^m$
\State Outputs: $\mathbf F(p_{[1:k]},v_{[1:j]})$ for all $k \in [n]$ and $j\in[m]$
\State Initialize $\mathbf G^{for}(k,0)=1\ \forall\ k$ and $\mathbf G^{for}(k,j)=0\ \forall\ k<j$
\For {$k = 1:n$ and $j = 1:m$}
\State Use \eqref{eq:approx_smap_dpfor} to update $\mathbf G^{for}(k,j)$
\EndFor
\State return $\mathbf G^{for}(k,j)\ \forall\ k,j$
\end{algorithmic}
\end{algorithm}
We note that a similar dynamic programming approach yields $\mathbf F(p_{[k:n]},v_{[j:m]})$ for all $k \in [n]$ and $j\in[m]$ in $O(mn)$ time and space complexity by defining
\begin{align*}
&\mathbf G^{rev}(k,j)\triangleq \mathbf F(p_{[k:n]},v_{[j:m]}).
\end{align*}
\subsection{An algebraic definition of the infiltration product.}
\label{app:infil_def}
We now give a more formal definition of the infiltration product (see \cite{lothaire1997combinatorics} for the equivalence of the two definitions and a more rigorous treatment). A \textit{formal series} with indeterminates (or variables) in a set $\mathcal A$ and coefficients in a commutative ring $\mathcal R$, is a mapping of $\mathcal A^*$ onto $\mathcal R$. Recall that a commutative ring is a set which forms an abelian group under an \textit{addition} operation, is a monoid under a \textit{multiplication} operation which commutes, and the multiplication operation distributes over addition. Here we consider $\mathbb Z$, the set of integers as the commutative ring $\mathcal{R}$. A formal series is called a \textit{polynomial} if only a finite number of sequences are mapped to non-zero values, the rest of the sequences map to zero. Consider two polynomials $\sigma,\tau: \mathcal{A}^*\rightarrow \mathbb Z$. The value taken by a sequence $w\in \mathcal A^*$ on $\sigma$ (or the coefficient of $w$ in $\sigma$) is denoted by $\langle \sigma,w\rangle \in \mathbb R$. We also define binary addition ($\oplus$) and multiplication operations ($\times$) on the set of polynomials as follows:
\begin{align}
\langle \sigma\oplus \tau,w\rangle \triangleq \langle \sigma,w\rangle + \langle \tau,w \rangle \quad \forall w\in \mathcal A^*,\label{eq:polynomial_add}\\
\langle \sigma\times \tau,w\rangle \triangleq \sum_{\substack{f,g\in \mathcal A^*:\\ f.g=w}}\langle \sigma,f\rangle \langle \tau,g \rangle \quad \forall w\in \mathcal A^*.\label{eq:polynomial_prod}
\end{align}
We will use the usual symbols $+$ and $.$ in place of $\oplus$ and $\times$ in this work for convenience. The meaning of the operation would be clear depending on the operands. With these operations the set of polynomials form a non-commutative ring, and is denoted by $\mathbb Z\langle\mathcal A \rangle$, also called the free $\mathbb Z$-algebra on $\mathcal A$ in ring theory. Note that the addition and multiplication operations defined in \eqref{eq:polynomial_add} and \eqref{eq:polynomial_prod} are similar to the operations defined on commutative polynomials, except that the multiplication operation under the summation in \eqref{eq:polynomial_prod} ($f.g=w$) is actually concatenation and is non-commutative. The multiplication inside the summation in \eqref{eq:polynomial_prod} is multiplication in the real field and hence commutative. It is also easy to prove that the multiplication defined in \eqref{eq:polynomial_prod} distributes over addition defined in \eqref{eq:polynomial_add}. Thus, a polynomial in $\mathbb Z\langle\mathcal A \rangle$ can be represented as a sum of monomials in $\mathcal A^*$ each with an associated coefficient in $\mathbb Z$, i.e., $\sigma=\sum\limits_{w\in \mathcal A^*} \langle\sigma,w \rangle w$. Define the \textit{degree} of a polynomial to be equal to the length of a longest sequence with a non-zero coefficient in the polynomial and the \textit{number of terms} of a polynomial as the number of sequences with non-zero coefficients in the polynomial. Note that a degree $d$ polynomial could have a number of terms upto $2^{d+1}-1$.
With this, the \textit{infiltration product} (in general, for two polynomials) is defined as follows:
\begin{align}
\forall f\in \mathcal{A}^*,& \quad f\uparrow e = e\uparrow f=f.\nonumber \\
\forall f,g\in \mathcal{A}^*&,\quad \forall a,b\in \mathcal{A}, \nonumber\\
fa\uparrow gb=(f\uparrow gb)a&+(fa\uparrow g)b+\mathbbm{1}_{a=b}(f\uparrow g)a.\nonumber \\
\forall \sigma,\tau \in \mathbb{Z}\langle\mathcal{A} \rangle, \quad &\sigma\uparrow \tau=\sum_{f,g\in \mathcal{A}^*} \langle \sigma,f \rangle \langle \tau,g \rangle (f\uparrow g). \label{def:infiltforseq}
\end{align}
\subsection{Proof of the channel equivalence in Theorem~\ref{thm:channel_equiv}.}
\label{app:channel_equiv}
We first state a lemma (introduced in \cite{Srini2019}) which is closely related to the channel equivalence. The proof of the Lemma is relegated to Appendix~\ref{app:bin_inf_lemma}.
\begin{restatable}{lemma}{bininfrelation}
\label{lemma:bin_inf_relation}
For $h,f_1,f_2,...,f_m \in \mathcal{A}^*$,\\
$${h \choose f_1} {h \choose f_2}...{h \choose f_m}=\sum_{w\in \mathcal{A}^*}\langle f_1\uparrow f_2\uparrow ...\uparrow f_m,w \rangle{h \choose w}.$$
\end{restatable
The channel equivalence can essentially be tied to this lemma as follows: consider the two channel models in Fig.~\ref{fig:channel_equiv}. The probability of observations given the input in both cases is proportional to the number of ways of obtaining the observations given the input.
\begin{itemize}
\item For the $t$-trace deletion channel model in Fig.~\ref{fig:channel_equiv} (a), the number of ways to obtain the traces given the input is equal to ${X \choose \tilde{Y}^{(1)}}{X \choose \tilde{Y}^{(2)}}...{X \choose \tilde{Y}^{(t)}}$.
\item For the cascade model in Fig.~\ref{fig:channel_equiv} (b), the number of ways to obtain the traces given the input is equal to $\sum_{z} {X \choose z} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle$ (this is shown in the proof of the theorem).
\end{itemize}
Clearly by Lemma~\ref{lemma:bin_inf_relation}, the above two are equal, thus proving the equivalence of the two channel models. On the contrary, this equivalence provides a nice physical interpretation of Lemma~\ref{lemma:bin_inf_relation}.
\channelequiv*
\begin{proof}
We show the probabilistic equivalence between the two channel models in Fig.~\ref{fig:channel_equiv}: 1) $t$ independent deletion channels 2) cascade of a deletion channel with parameter $\delta^t$ with the \textit{remnant} channel with parameter $\delta$ defined as follows: an input symbol to the remnant channel is included in $k>0$ given outputs and deleted in the rest with a probability $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
Given this definition, we now first compute the probability of a given set of output sequences given an input sequence for the remnant channel, namely $\Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z)$. First, note that there can be multiple deletion patterns corresponding to outputs $\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}$ resulting from a given input $Z$. The number of such patterns is equal to $\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle$, which essentially follows from the definition of the infiltration product. Consider one such valid deletion pattern, i.e., a deletion pattern $\mathcal{D}$ that is a mapping of the symbols in $Z$ onto the symbols in $\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}$: $\mathcal{D}=\{(1,S_1),(2,S_2),...,(|Z|,S_{|Z|})\}$. Here $(i,S_i)$ represents the fact that $Z_i$ is not deleted in the output set $\tilde{Y}^{(S_i)}$ and is deleted in the rest. Clearly, $|S_i|>0$ from the definition of the remnant channel. Also $\sum_{i=1}^{|Z|}|S_i|=\sum_{j=1}^t|\tilde{Y}^{(j)}|$ since every symbol of each output is associated with exactly one input symbol and hence corresponds to one particular $S_i$. Thus,
\begin{align*}
\Pr(&\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z) \\&= \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z,\mathcal{D})\\
&=\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \prod_{i=1}^{|Z|}\frac{(1-\delta)^{|S_i|}\delta^{t-|S_i|}}{1-\delta^t}\\
&=\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \frac{(1-\delta)^{\sum|S_i|}\delta^{|Z|t-\sum |S_i|}}{(1-\delta^t)^{|Z|}}\\
&=\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \frac{(1-\delta)^{\sum|\tilde{Y}^{(j)}|}\delta^{|Z|t-\sum |\tilde{Y}^{(j)}|}}{(1-\delta^t)^{|Z|}}.
\end{align*}
We can then compute the probability of the output given the input for the cascade channel as
\begin{align*}
\Pr(&\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|X)\\ &= \sum_{z} \Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)},Z=z|X)\\
&= \sum_{z} \Pr(Z=z|X)\Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z=z)\\
&= \sum_{z} \Bigg[{X \choose z} \delta^{t(|X|-|z|)}(1-\delta^t)^{|z|}\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle \frac{(1-\delta)^{\sum|\tilde{Y}^{(j)}|}\delta^{|z|t-\sum |\tilde{Y}^{(j)}|}}{(1-\delta^t)^{|z|}}\Bigg]\\
&= \left[\sum_{z} {X \choose z} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle\right] \delta^{t|X|-\sum |\tilde{Y}^{(j)}|} {(1-\delta)^{\sum|\tilde{Y}^{(j)}|}}\\
&={X \choose \tilde{Y}^{(1)}}{X \choose \tilde{Y}^{(2)}}...{X \choose \tilde{Y}^{(t)}} \delta^{t|X|-\sum |\tilde{Y}^{(j)}|} {(1-\delta)^{\sum|\tilde{Y}^{(j)}|}}\\
&=\prod_{j=1}^t {X \choose \tilde{Y}^{(j)}} \delta^{|X|-|\tilde{Y}^{(j)}|} {(1-\delta)^{|\tilde{Y}^{(j)}|}}\\
&=\Pr(Y^{(1)}, Y^{(2)},..., Y^{(t)}|X),
\end{align*}
proving the equivalence.
\end{proof}
\subsection{Proof of Lemma~\ref{lemma:F_decomposition}.}
\label{app:F_lemma_proof}
\deletionMLrellemma*
\begin{proof}
The proof of this lemma uses a similar approach as the proof of Thm.~\ref{thm:approx_smap_postprob}. First, in the expression for $\mathbf F(\cdot )$, we separate out the subsets that contain index $i$:
\begin{align*}
\mathbf F(p,Y) &= \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j}\\
&= \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m, \\ i \notin \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j} + \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j}\\
&= \mathbf F(p_{[n]\backslash \{i\}},Y) + \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j}. \numberthis
\label{eq:Flemma_proof1}
\end{align*}
Now the second term can be further split as,
\begin{align*}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j} &= \sum_{k=1}^m\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ \mathcal S_k = i}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j}.
\end{align*}
One could express the set $\mathcal S$ as the union $\mathcal S = \mathcal S' \cup \{i\} \cup \mathcal S''$ such that $\mathcal S' \subseteq [i-1]$ and $\mathcal S'' \subseteq [i+1:n]$ to get
\begin{align*}
&\sum_{k=1}^m\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ \mathcal S_k = i}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j}\\
&= \sum_{k=1}^m\sum_{\substack{\mathcal S'|\\ \mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\sum_{\substack{\mathcal S''|\\ \mathcal S'' \subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \left( \prod\limits_{j=1}^{k-1} p_{\mathcal S'_j}^{Y_j} (1-p_{\mathcal S'_j})^{1-Y_j} \right) \left( p_{i}^{Y_k} (1-p_{i})^{1-Y_k} \right) \left( \prod\limits_{j=1}^{m-k} p_{\mathcal S''_j}^{Y_{j+k}} (1-p_{\mathcal S''_j})^{1-Y_{j+k}} \right)\\
&= \sum_{k=1}^m p_{i}^{Y_k} (1-p_{i})^{1-Y_k}\left(\sum_{\substack{\mathcal S'|\\ \mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}} \prod\limits_{j=1}^{k-1} p_{\mathcal S'_j}^{Y_j} (1-p_{\mathcal S'_j})^{1-Y_j} \right) \left( \sum_{\substack{\mathcal S''|\\ \mathcal S'' \subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \prod\limits_{j=1}^{m-k} p_{\mathcal S''_j}^{Y_{j+k}} (1-p_{\mathcal S''_j})^{1-Y_{j+k}} \right)\\
&= \sum_{k=1}^m p_{i}^{Y_k} (1-p_{i})^{1-Y_k} \mathbf F(p_{[i-1]},Y_{[k-1]}) \mathbf F(p_{[i+1:n]},Y_{[k+1:m]}).
\end{align*}
The $\sum_{k=1}^m$ summation in the above expression could further be split into the two cases depending on whether $Y_k=0$ or $Y_k=1$, which simplifies the term $p_{i}^{Y_k} (1-p_{i})^{1-Y_k}$ to either $1-p_i$ or $p_i$ respectively. Thus,
\begin{align*}
&\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{Y_j} (1-p_{\mathcal S_j})^{1-Y_j}\\
&= (1-p_i)\sum_{k|Y_k=0} \mathbf F(p_{[i-1]},Y_{[k-1]}) \mathbf F(p_{[i+1:n]},Y_{[k+1:m]}) + p_i\sum_{k|Y_k=1} \mathbf F(p_{[i-1]},Y_{[k-1]}) \mathbf F(p_{[i+1:n]},Y_{[k+1:m]}).\numberthis
\label{eq:Flemma_proof2}
\end{align*}
\noindent Plugging \eqref{eq:Flemma_proof2} in \eqref{eq:Flemma_proof1} concludes the proof of the Lemma.
\end{proof}
\subsection{Symbolwise posterior probabilities for the remnant channel}
\label{app:remnant_postprob}
Consider the remnant channel shown below, and let $Z=Z_1Z_2...Z_n$. Also let $Z_i \sim \text{Ber}(0.5)$. We aim to compute $\Pr(Z_i=1|\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)})$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.4]{remnant_channel.pdf}
\caption{The remnant channel}
\end{figure}
As seen in appendix~\ref{app:channel_equiv}, the input-output relation for this channel is given by:
\begin{align*}
\Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z) =\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \frac{(1-\delta)^{\sum|\tilde{Y}^{(j)}|}\delta^{nt-\sum |\tilde{Y}^{(j)}|}}{(1-\delta^t)^{n}}.
\end{align*}
Now, one could write the symbolwise posterior probabilities for $Z$ as:
\begin{align*}
\Pr(Z_i=1&|\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}) = \sum_{\substack{z||z|=n,\\z_i=1}} \Pr(z|\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)})\\
&{=} \frac{1}{2^n \Pr(\tilde Y^{(1)}...,\tilde Y^{(t)})} \sum_{\substack{z||z|=n,\\z_i=1}} \Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|z)\\
&{=} \frac{{(1-\delta)^{\sum|\tilde{Y}^{(j)}|}\delta^{nt-\sum |\tilde{Y}^{(j)}|}}}{(1-\delta^t)^{n} 2^n \Pr(\tilde Y^{(1)},...,\tilde Y^{(t)})} \sum_{\substack{z||z|=n,\\z_i=1}} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle. \numberthis
\label{eq:remant_map_prob_1}
\end{align*}
A similar expression can be obtained for the case when $Z_i=0$ as
\begin{align*}
\Pr(Z_i=0&|\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)})\\
&{=} \frac{{(1-\delta)^{\sum|\tilde{Y}^{(j)}|}\delta^{nt-\sum |\tilde{Y}^{(j)}|}}}{(1-\delta^t)^{n} 2^n \Pr(\tilde Y^{(1)},...,\tilde Y^{(t)})} \sum_{\substack{z||z|=n,\\z_i=0}} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle. \numberthis
\label{eq:remant_map_prob_0}
\end{align*}
We could further simplify \eqref{eq:remant_map_prob_1} and \eqref{eq:remant_map_prob_0} using the fact that
\begin{align*}
\Pr(Z_i=1|\tilde Y^{(1)},...,\tilde Y^{(t)}) &= \frac{\Pr(Z_i=1|\tilde Y^{(1)},...,\tilde Y^{(t)})}{\Pr(Z_i=0|\tilde Y^{(1)},...,\tilde Y^{(t)})+\Pr(Z_i=1|\tilde Y^{(1)},...,\tilde Y^{(t)})} \\
&= \frac{\sum\limits_{\substack{z||z|=n,\\z_i=1}} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle}{\sum\limits_{\substack{z||z|=n}} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle}. \numberthis \label{eq:remnant_map_prob}
\end{align*}
We precisely describe the algorithm which computes the terms in \eqref{eq:remnant_map_prob} in section~\ref{sec:exactsmap}, by exploiting the edit graph interpretation of the infiltration product, but give a high level idea below. The complexity of such an algorithm is $O((2n)^t)$ which is equal to the number of edges in the edit graph. Note that for a fixed number of traces, this algorithm is polynomial in the blocklength as opposed to a naive approach of iterating through all the $n$-length sequences.
Recall that $\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle$ is the number of paths from origin to destination of the edit graph $\mathcal G(\tilde Y^{(1)}, ..., \tilde Y^{(t)})$ which correspond to $z$. Therefore,
$\sum_{\substack{z||z|=n}}\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle$ is equal to the number of $n$-length paths in $\mathcal G(\tilde Y^{(1)}, ..., \tilde Y^{(t)})$ from the origin to the destination. Note that the edit graph has no cycles, so this quantity can be efficiently computed via the following dynamic program -- the number of $n$ length paths from the origin to a vertex $v$ is equal to the sum of the number of $n-1$ length paths from the origin to the in-neighbors of $v$. Such a procedure iterates over the vertex set of $\mathcal G(\tilde Y^{(1)}, ..., \tilde Y^{(t)})$ exactly once.
The numerator term $\sum_{\substack{z||z|=n\\z_i=1}}\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle$ can be interpreted in a similar way: it is equal to the number of $n$-length paths in $\mathcal G(\tilde Y^{(1)}, ..., \tilde Y^{(t)})$ from the origin to the destination such that the $i^{th}$ edge of the path corresponds to a `1'. The algorithm for this, therefore, follows a similar principle but has an extra step. For each vertex $v$, we compute
\begin{itemize}
\item the number of paths from the origin to $v$ of length $0,1,...,n$,
\item the number of paths from $v$ to the destination of length $0,1,...,n$.
\end{itemize}
Next we iterate over all edges in $\mathcal G(\tilde Y^{(1)}, ..., \tilde Y^{(t)})$ corresponding to a `1' and accumulate the number of $n$ length paths which have this particular edge as its $i^{th}$ edge. Thus, this algorithm iterates over the vertex set twice and the edge set of $\mathcal G(\tilde Y^{(1)}, ..., \tilde Y^{(t)})$ once.
\subsection{Proof of Lemma~\ref{lemma:bin_inf_relation}}
\label{app:bin_inf_lemma}
\bininfrelation*
\begin{proof}
We use induction to prove the statement. The statement is trivially true when $m=1$ since, $\sum_{w}{h \choose w}\langle f_1,w \rangle={h \choose f_1}$ as $\langle f,w \rangle=\mathbbm{1}_{f=w}$. We refer the reader to \cite{lothaire1997combinatorics} for a proof of the proposition for the case when $m=2$. Assume that the statement is true for $m=k \in \mathbb{Z}, k\geq 2$. We next prove the validity when $m=k+1$. \\
Consider
\begin{align}
{h \choose f_1} {h \choose f_2}...{h \choose f_k}{h \choose f_{k+1}}\nonumber\\=\sum_{w}{h \choose w}\langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle {h \choose f_{k+1}}\nonumber\\
=\sum_{w}\left[{h \choose w} {h \choose f_{k+1}}\right]\langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle\nonumber\\
=\sum_{w}\left[\sum_v \langle w\uparrow f_{k+1},v \rangle {h \choose v}\right]\langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle\nonumber\\
=\sum_{v}{h \choose v}\left[\sum_w \langle w\uparrow f_{k+1},v \rangle \langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle\right]. \label{eq:prop2proof}
\end{align}
To evaluate the term in the square bracket, we use \eqref{def:infiltforseq}. For the case where $\tau \in \mathcal{A}^*,\sigma \in \mathbb{Z}\langle \mathcal{A} \rangle$ in \eqref{def:infiltforseq}, we have $$\sigma\uparrow \tau=\sum_{f\in \mathcal{A}^*} \langle \sigma,f \rangle (f\uparrow \tau),$$ and thus \begin{equation}
\langle \sigma \uparrow \tau,u \rangle=\sum_{f\in \mathcal{A}^*} \langle \sigma,f \rangle \langle f\uparrow \tau,u\rangle.
\label{eq:prop2proof2}
\end{equation} We use \eqref{eq:prop2proof2} to replace the term in the square bracket in \eqref{eq:prop2proof}, i.e.,
\begin{align}
{h \choose f_1} {h \choose f_2}...{h \choose f_k}{h \choose f_{k+1}}\nonumber\\
=\sum_{v}{h \choose v}\langle (f_1\uparrow f_2\uparrow ...\uparrow f_k) \uparrow f_{k+1},v \rangle,
\end{align}
and the proposition follows from the associativity property of the infiltration product.
\end{proof}
\subsection{Proof of Lemma~\ref{lemma:smapsum}}
\label{app:smapsum}
\begin{restatable}{lemma}{smapsum}
\label{lemma:smapsum}
\begin{align*}
\sum_{\substack{f||f|{=}n\\f_i{=}a}}{f \choose g} = 2^{n-|g|}\Bigg(\frac{1}{2}{n-1 \choose |g|} &+\sum_{j|g_j=a}{i-1 \choose j-1}{n-i \choose |g|-j}\Bigg),
\end{align*}
where $j\in\Big[\max\{1,|g|+i-n\}:\min\{i,|g|\}\Big]$.
\end{restatable}
\begin{proof}
First, observe that $${f \choose g} = \sum_{\substack{S\subseteq [n]:\\|S|=|g|}} \mathbbm 1_{f_S=g},$$ where the summation is over all ordered subsets of $[n]=\{1,2,...,n\}$ of size $|g|$ and $f_S$ corresponds to the subsequence of $f$ indexed by $S$. Thus,
\begin{align*}
\sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}}&{f \choose g} = \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \sum_{\substack{S\subseteq [n]|\\|S|=|g|}} \mathbbm 1_{f_S=g} = \sum_{\substack{S\subseteq [n]|\\|S|=|g|}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g}\\
&=\sum_{\substack{S\subseteq [n]|\\|S|=|g|\\i \notin S}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g} +
\sum_{\substack{S\subseteq [n]|\\|S|=|g|\\i \in S}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g}\\
&=\sum_{\substack{S\subseteq [n]|\\|S|=|g|\\i \notin S}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g} +
\sum_{j=1}^m \sum_{\substack{S\subseteq [n]|\\|S|=|g|\\S_j=i}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g}. \numberthis \label{eq:lemmasmapsum}
\end{align*}
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.25]{lemma1.pdf}
\vspace*{-2mm}
\caption{Figure illustrating proof of Lemma~\ref{lemma:smapsum}.}
\label{fig:lemmasmapsum}
\end{center}
\end{figure}
The two terms in \eqref{eq:lemmasmapsum} can be visualized as the number of ways to fill up the blank spaces (spaces without arrows pointing to it in $f$) in Fig.~\ref{fig:lemmasmapsum}(a) and (b) respectively. Solving this counting problem is easy and we get,
$$\sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}}{f \choose g}=2^{n-|g|}\left(\frac{1}{2}{n-1 \choose |g|}+\sum_{j|g_j=a}{i-1 \choose j-1}{n-i \choose |g|-j}\right).$$
\end{proof}
\section{Conclusions}
\label{sec:conclusions}
In this work we gave, to the best of our knowledge, the first results and techniques to compute posterior distributions over single and multiple deletion channels. We also provided a new perspective on the maximum-likelihood for the deletion channel by showing an equivalence between a discrete optimization problem and its relaxed version. In this process, we introduced a variety of tools (the relaxed binomial coefficient, edit graph and infiltration product) and demonstrated their use for analyzing deletion channels. We also presented numerical evaluations of our algorithms.
\section{Introduction}
\label{sec:intro}
Sequence reconstruction over deletion channels, both with and without a codebook, has received considerable attention in the information theory as well as in the theoretical computer science literature. From an information theory perspective, reconstruction over the deletion channel, or more specifically a maximum-likelihood (ML) argument for the deletion channel, would give further insight on the capacity of the deletion channel, a long-standing open problem (see \cite{mitzenmacher2009survey}). To quote \cite{mitzenmacher2009survey} -- ``at the very least, progress in this direction would likely surpass previous results on the capacity of the deletion channels''. Yet, there are no results on reconstruction over a deletion channel with statistical guarantees -- in this work, we take a step in this direction.
On the other hand, the problem of \textit{trace reconstruction}, as introduced in \cite{Batu2004}, has received renewed interest in the past few years (see \cite{Holenstein2008}, \cite{Peres2017}, \cite{De2017}, \cite{holden18}, \cite{Nazarov:2017}, \cite{holden2018lower}, \cite{chase2019new}). The problem of trace reconstruction can be stated simply as follows: consider a sequence $X$ which is simultaneously passed through $t$ independent deletion channels to yield $t$ deleted observations (also called \textit{traces}) of $X$ (see Fig.~\ref{fig:tdeletion}). How many such traces are needed to reconstruct $X$ perfectly? A variety of upper and lower bounds for this problem have been proposed, both for worst case and average case reconstruction. Our problem definition, stated in the following paragraph, is closely related to the average case trace reconstruction.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{tdeletion.pdf}
\caption{$t$-deletion channel model: sequence $X$ passed through $t$ independent deletion channels to yield $t$ \textit{traces}. We aim to estimate $X$ from the $Y^{(i)}$s.}
\label{fig:tdeletion}
\end{center}
\end{figure}
\noindent \textbf{Problem definition.} Given an input sequence of length $n$ (known apriori), the independently and identically distributed (i.i.d.) deletion channel deletes each input symbol indepedently with probability $\delta$, producing at its output a subsequence of the input sequence. Consider a sequence $X$ passed through $t$ ($t$ is fixed) such deletion
channels as shown in
Fig.~\ref{fig:tdeletion}. We call this the $t$-deletion channel model. We ask two questions:
\begin{enumerate}
\item For $t=1$ (see Fig.~\ref{fig:1deletion}, also called the single deletion channel model), and when $X_i \sim\ \text{ind. Ber}(p_i)$, compute the posterior distributions of $X_i$ given the trace $Y$.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.25]{1deletion.pdf}
\caption{The single deletion channel model. We assume $X_i \sim\ \text{ind. Ber}(p_i)$.}
\label{fig:1deletion}
\end{center}
\end{figure}
\item In the $t$ deletion channel model, for a fixed $t$ assume that $X_i \sim\ \text{i.i.d. Ber}(0.5)$ and compute the posterior distributions of $X_i$ given all traces $Y^{(1)}, Y^{(2)},...,Y^{(t)}$.
\end{enumerate}
Note that solving 1) above doesn't lead to a natural solution for 2). This is because for a memoryless channel, we have $Y^{(j)} - X_i - Y^{(k)}$ and hence $\Pr(X_i=\alpha|Y^{(j)}, Y^{(k)}) \propto \Pr(X_i=\alpha|Y^{(j)}) \Pr(X_i=\alpha|Y^{(k)})$; so one could independently combine the posterior probabilities from each noisy observation. This is not the case for deletion channels since the markov chain $Y^{(j)} - X_i - Y^{(k)}$ no longer holds. More intuitively, one needs to first ``align'' the traces for computing the likelihoods.
We point out that the problem considered in 2) asks a question complementary to the trace reconstruction: given a fixed (possibly a few) number of traces, what is our ``best'' guess of $X$? We provide algorithms which do this. Unlike trace reconstruction, we are not concerned with perfect reconstruction (since perfect reconstruction may not be possible with just a few traces), although it should also be obvious that performance guarantees for our algorithms (not a part of this work) would naturally lead to upper bounds for trace reconstruction.
Deletion channel by itself is known to be notoriously difficult to analyse. As stated earlier, the capacity of a single deletion channel is still unknown (\cite{diggavi2007capacity,diggavi2006information,diggavi2001transmission}); as are optimal coding schemes. Recent works have looked at the design of codes for deletion channels (\cite{ratzer2005marker,ratzer2000codes,thomas2017polar}); these works consider use of a codebook (we do not). As a result, statistical estimation over deletion channels is also a difficult problem due its highly combinatorial nature. To the best of our knowledge, as yet there are no efficient estimation algorithms over deletion channels with statistical guarantees; not even for ML over a single deletion channel.
\noindent \textbf{Biological motivation.} Trace reconstruction in itself was motivated, in part, by problems DNA sequence reconstruction. One such problem was to infer the DNA sequence of a common ancestor from the samples of its descendants. We argue that our problem definition fits more naturally in such a scenario since perfect reconstruction may not be feasible or even possible. Our motivation for considering this problem also comes from a recent DNA sequencing technology called \textit{nanopore sequencing}. The $t$-deletion channel model is a simplistic model to approximately capture the process of a DNA sequence passed through a nanopore sequencer\footnote{As seen in
\cite{Mao2017},\cite{MDK17} there are more complicated effects of
the nanopore reader not captured in this simple
representation.}.
Very recently, a variant of the trace reconstruction problem called \textit{coded trace reconstruction} has been proposed, motivated by portable DNA-based data storage systems using DNA nanopores (see \cite{abroshan2019coding}, \cite{cheraghchi2019coded}, \cite{brakensiek2019coded}) and we believe that the ideas in this work may prove useful in such a setting.
There are other works on sequence assembly (see for example, \cite{Li09fastand}, \cite{Shomorony2016}),
where multiple short reads (from different segments of a sequence) are used to reconstruct the bigger
sequence. This work differs from sequence assembly since we are interested to infer the entire length sequence and not just small segments of it (which are then ``stitched'' together in sequence assembly).
\noindent \textbf{Tools and techniques.} In terms of theoretical tools, the series of books by Lothaire (\cite{lothaire1997combinatorics,lothaire2002algebraic,lothaire2005applied}) extensively use algebraic tools for problems in the combinatorics of sequences (or \textit{words}), and our work is inspired by such techniques. We borrow many of their notations and results for our work.
\noindent \textbf{Contributions.} {\color{red}} Our main contribution is to provide tools to visualize and analyze error events (described in the next section) for the multiple deletion channel model in Fig.~\ref{fig:tdeletion}. We also provide algorithms to solve the problems stated in 1) and 2) earlier in the section.
\begin{itemize}[wide=5pt]
\item In section~\ref{sec:1deletion}, for the single deletion channel model, we provide an $O(n^2)$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i=1|Y)\ \forall\ i$ when $X_i \sim \text{ind. Ber}(p_i)$.
\item In Section~\ref{sec:exactsmap}, for the $t$-deletion channel model, we give an $O(2^t n^{t+2})$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i = 1|Y_1,...,Y_t)$ when $X_i \sim \text{ind. Ber}(0.5)$.
\end{itemize}
\section{Introduction}
\label{sec:intro}
Sequence reconstruction over deletion channels, both with and without a codebook, has received considerable attention in the information theory as well as in the theoretical computer science literature. From an information theory perspective, reconstruction over the deletion channel, or more specifically a maximum-likelihood (ML) argument for the deletion channel, would give further insight on the capacity of the deletion channel, a long-standing open problem (see \cite{mitzenmacher2009survey}). To quote \cite{mitzenmacher2009survey} -- ``at the very least, progress in this direction would likely surpass previous results on the capacity of the deletion channels''. Yet, there are no results on reconstruction over a deletion channel with statistical guarantees. In this work, we take steps in this direction.
In this space, the problem of \textit{trace reconstruction}, as introduced in \cite{Batu2004}, has also received renewed interest in the past few years (see \cite{Holenstein2008,Peres2017}, \cite{De2017}, \cite{holden18}, \cite{Nazarov:2017}, \cite{holden2018lower}, \cite{chase2019new}). The problem of trace reconstruction can be stated simply as follows: consider a sequence $X$ which is simultaneously passed through $t$ independent deletion channels to yield $t$ output subsequences (also called \textit{traces}) of $X$ (see Fig.~\ref{fig:tdeletion}). How many such traces are needed to reconstruct $X$ perfectly? A variety of upper and lower bounds for this problem have been proposed, both for worst case and average case reconstruction. Our problem formulation is complementary to this, as we discuss next.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{tdeletion.pdf}
\caption{The $t$-trace deletion channel model: the sequence $X$ is passed through $t$ independent deletion channels to yield $t$ \textit{traces}. We aim to estimate $X$ from the $Y^{(i)}$s.}
\label{fig:tdeletion}
\end{center}
\end{figure}
\noindent \textbf{Problem formulation.} Given an input sequence of length $n$ (known apriori), the independently and identically distributed (i.i.d.) deletion channel deletes each input symbol indepedently with probability $\delta$, producing at its output a subsequence of the input sequence. Consider a sequence $X$ passed through $t$ ($t$ is fixed) such deletion
channels as shown in
Fig.~\ref{fig:tdeletion}. We call this the $t$-trace deletion channel model. We ask three main questions:
\begin{enumerate}
\item For $t=1$ (also called \textit{single-trace deletion channel}, see Fig.~\ref{fig:1deletion}), what is the maximum-likelihood estimate of $X$ having observed $Y$, i.e., a solution to $\argmax\limits_{x\in \{0,1\}^n}\Pr(Y|X)$.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{1deletion.pdf}
\caption{The single-trace deletion channel model.}
\label{fig:1deletion}
\end{center}
\end{figure}
\item For $t=1$ and $X_i \sim\ \text{ind. Ber}(p_i)$ in Fig.~\ref{fig:1deletion}, what are the posterior distributions of $X_i$ given the trace $Y$, i.e., compute $\Pr(X_i=\alpha|Y)$.
\item For a fixed $t$, with $t>1$ and $X_i \sim\ \text{i.i.d. Ber}(0.5)$ in Fig.~\ref{fig:tdeletion}, what are the posterior distributions of $X_i$ given all traces $Y^{(1)}, Y^{(2)},...,Y^{(t)}$, i.e., compute $\Pr(X_i=\alpha|Y^{(1)}, Y^{(2)},...,Y^{(t)})$.
\end{enumerate}
We make two observations:\\
{\bf I.} An answer to 2) above doesn't lead to a natural solution for 3) since the deletion channel is not memoryless. In particular, for a memoryless channel, we have $Y^{(j)}_i - X_i - Y^{(k)}_i$ and hence $\Pr(X_i=\alpha|Y^{(j)}, Y^{(k)}) = \Pr(X_i=\alpha|Y^{(j)}_i, Y^{(k)}_i) \propto \Pr(X_i=\alpha|Y^{(j)}_i) \Pr(X_i=\alpha|Y^{(k)}_i)$; so one could first obtain the posterior probabilities from each independent observation and combine them after. However, this is not the case for deletion channels since the markov chain $Y^{(j)}_i - X_i - Y^{(k)}_i$ no longer holds. As a result, one first needs to ``align'' all the observations in order to compute the likelihoods. \\
{\bf II.} Solving 3) naturally leads to an algorithm for trace reconstruction, that selects the most likely value for each $X_i$. However, the problem formulation in 3) asks a question complementary to the trace reconstruction: given a fixed (possibly a few) number of traces, what is our ``best'' guess of $X$? Unlike trace reconstruction, we are not concerned with perfect reconstruction (since perfect reconstruction may not be possible with just a few traces). We also note that error rate guarantees for our algorithms (not a part of this work) would naturally lead to upper bounds for trace reconstruction.\\
\noindent \textbf{Contributions.}
Our main contributions are as follows.
\begin{itemize}[leftmargin=5mm]
\item We introduce mathematical tools and constructs to visualize and analyze single-trace and $t$-trace deletion error events (see Section~\ref{sec:notation}).
\item For the single-trace deletion channel, we establish an equivalence between finding the optimal ML decoder and a continuous optimization problem we introduce (see Section~\ref{sec:1deletion_ML}).
\item In Section~\ref{sec:1deletion}, we prove the following:
\begin{theorem}
For the single-trace deletion channel model with priors $X_i \sim \text{ind. Ber}(p_i)$ and observed trace $Y$, the symbolwise posterior probabilities $\Pr(X_i=1|Y)\ \forall\ i$ can be computed in $O(n^2)$ time complexity.
\end{theorem}
\item In Section~\ref{sec:exactsmap}, we prove the following:
\begin{theorem}
For the $t$-trace deletion channel model with priors $X_i \sim \text{i.i.d. Ber}(0.5)$ and observed traces $Y^{(1)},...,Y^{(t)}$, the symbolwise posterior probabilities $\Pr(X_i = 1|Y^{(1)},...,Y^{(t)})\ \forall\ i$ can be computed in $O(2^t n^{t+2})$ time complexity.\\
\end{theorem}
\end{itemize}
\noindent \textbf{Tools and techniques.} In terms of theoretical tools, the series of books by Lothaire (\cite{lothaire1997combinatorics,lothaire2002algebraic,lothaire2005applied}) extensively use algebraic tools for problems in the combinatorics of sequences (or \textit{words}), and our work is inspired by such techniques. We borrow some notation and leverage a few of their results in our work. \\
\noindent \textbf{Biological motivation.} Trace reconstruction in itself was motivated, in part, by problems in DNA sequence reconstruction. One such problem was to infer the DNA sequence of a common ancestor from the samples of its descendants. Our problem definition, that considers a fixed value of $t$, would fit naturally in a scenario with a fixed number of descendants where perfect reconstruction may not be possible.
Our motivation for considering this problem also comes from a recent DNA sequencing technology called \textit{nanopore sequencing}. The $t$-trace deletion channel model is a simplistic model to approximately capture the process of a DNA sequence passed through a nanopore sequencer\footnote{As seen in
\cite{Mao2017},\cite{MDK17} there are more complicated effects of
the nanopore reader not captured in this simple
representation.}. \\
\noindent \textbf{More related work.}
Our work falls under the general umbrella of sequence reconstruction over deletion channels, where we offer, to the best of our knowledge, the first non-trivial results on maximum likelihood and maximum aposteriori estimates for the single and multiple deletion channel. As mentioned earlier, the complementary problem of trace reconstruction falls closest to this work.
The deletion channel by itself is known to be notoriously difficult to analyse. As stated earlier, the capacity of a single deletion channel is still unknown (\cite{diggavi2007capacity,diggavi2006information,diggavi2001transmission}); as are optimal coding schemes. Prior works have looked at the design of codes for deletion channels (\cite{ratzer2005marker,ratzer2000codes,thomas2017polar}); these works consider use of a codebook (we do not). Statistical estimation over deletion channels is a difficult problem to analyze due its highly combinatorial nature. To the best of our knowledge, as yet there are no efficient estimation algorithms over deletion channels with statistical guarantees.
Very recently, a variant of the trace reconstruction problem called \textit{coded trace reconstruction} has been proposed, motivated by portable DNA-based data storage systems using DNA nanopores (see \cite{abroshan2019coding}, \cite{cheraghchi2019coded}, \cite{brakensiek2019coded}) and we believe that the ideas in this work may prove useful in such a setting.
There are other works on sequence assembly (see for example, \cite{Li09fastand}, \cite{Shomorony2016}),
where multiple short reads (from different segments of a sequence) are used to reconstruct the bigger
sequence. This work differs from sequence assembly since we are interested in inferring the entire length sequence and not just small segments of it (which are then ``stitched'' together in sequence assembly).\\
\noindent \textbf{Paper Organization.} Section~\ref{sec:notation} introduces our notation and visualization tools for the single and $t$-trace channel error events; Section~\ref{sec:1deletion_ML} provides a result concerning question 1) wherein we prove the equivalence of ML decoding to solving a continuous optimization problem; Section~\ref{sec:1deletion} answers question 2) for the single-trace channel;
Section~\ref{sec:exactsmap}) answers question 3) for the $t$-deletion channel; Section~\ref{sec:Numerics} gives numerical evaluations; and Section~\ref{sec:conclusions} concludes the paper.
\section{Symbolwise posterior probabilities for the $t$-trace deletion channel}
\label{sec:exactsmap}
In this section, we put to use the ideas and constructs introduced in section~\ref{sec:notation} to exactly compute the symbolwise posterior probabilities given $t$-traces, which in turn gives a symbolwise MAP estimate (with uniform priors) of the input sequence. In Appendix~\ref{app:remnant_postprob}, we also provide a method to compute the symbolwise posterior probabilities for the remnant channel -- we encourage the reader to use this appendix as a warm-up. For the $t$-trace deletion channel, similar expressions arise due to the channel equivalence result of Theorem~\ref{thm:channel_equiv}.
Let $\mathcal A = \{0,1\}$, and assume that $X\sim $ Uniform $\mathcal A^n$. Our goal is to compute the symbolwise posterior probabilities $\Pr(X_i{=}1|Y^{(1)},...,Y^{(t)})$, where $Y^{(j)}$ is the $j^{th}$ trace. Our proposed algorithm is provided in Alg.~\ref{alg:exact_smap} and estimates the symbolwise MAP (with uniform priors). We can directly leverage Alg.~\ref{alg:exact_smap} to reconstruct the input as follows: for each index $i$, compute $\Pr(X_i{=}1|Y^{(1)},...,Y^{(t)})$ and decide
\begin{align*}
\hat{X}_i = \begin{cases}
1,\quad &$if$\ \Pr(X_i{=}1|Y^{(1)},...,Y^{(t)})\geq 0.5 \\
0, \quad &$otherwise$.
\end{cases}
\end{align*}
Through the rest of this section, we show how to compute $\Pr(X_i{=}1|Y^{(1)},...,Y^{(t)})$\footnote{Symbolwise MAP with non-uniform priors is part of on-going work.} in two steps:
\begin{itemize}
\item We first give an expression for $\Pr(X_i{=}1|Y^{(1)},...,Y^{(t)})$ which sums over potentially an exponential number of terms.
\item We then show that this summation can be computed in polynomial time (polynomial in the blocklength $n$).
\end{itemize}
\noindent \textbf{Step 1: An expression for $\Pr(X_i{=}1|Y^{(1)},...,Y^{(t)})$.}
\begin{theorem}
\label{thm:exactSMAP_posteriorprob}
Assume $X\sim $ Uniform $\mathcal A^n$ or equivalently $X_i \sim\ \text{Ber}(0.5)$. The posterior probability of the $i^{th}$ bit given the $t$ traces can be expressed as
\begin{align*}
\Pr(X_i=1|Y^{(1)},&...,Y^{(t)}) &\\ =
& \Bigg[ \sum_{k=0}^n 2^{n-k-1}{n-1 \choose k} \sum_{w||w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \\
&+ \sum_{k=0}^n \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \Bigg] \Big/ \\
&\Bigg[ \sum_{k=0}^n 2^{n-k} {n \choose k} \sum_{\substack{w| |w|=k}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \Bigg]. \numberthis
\label{eq:posterior_prob}
\end{align*}
\end{theorem}
Note that the summation index, $w| |w|{=}k$ is over all sequences $w$ of length $k$; this is an alternate expression for $w|w{\in}\mathcal A^k$. We follow this convention throughout the rest of the paper.
\begin{proof}
\begin{align*}
\Pr(X_i=1&|Y^{(1)},...,Y^{(t)}) = \sum_{\substack{x||x|=n,\\x_i=1}} \Pr(x|Y^{(1)},...,Y^{(t)})\\
&\overset{(a)}{=} \frac{1}{2^n \Pr(Y^{(1)},...,Y^{(t)})} \sum_{\substack{x||x|=n,\\x_i=1}} \Pr(Y^{(1)},...,Y^{(t)}|x)\\
&\overset{(b)}{=} \frac{1}{2^n \Pr(Y^{(1)},...,Y^{(t)})} \sum_{\substack{x||x|=n,\\x_i=1}} \prod_{j=1}^t\Pr(Y^{(j)}|x),
\end{align*}
where $(a)$ uses Bayes' principle and $(b)$ is because each deletion channel acts independently. Recall that for a deletion channel with deletion probability $\delta$, $\Pr(Y|X=x)={x \choose Y}\delta^{|x|-|Y|}(1-\delta)^{|Y|}$. Also, using the fact that $\Pr(Y^{(1)},...,Y^{(t)})=\sum\limits_{\substack{x||x|=n}}\Pr(x) \Pr(Y^{(1)},...,Y^{(t)}|x)$ we have,
\begin{align*}
\Pr(X_i=1|Y^{(1)},...,Y^{(t)})= \frac{\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose Y^{(1)}}...{x \choose Y^{(t)}}}{\sum\limits_{\substack{x||x|=n}} {x \choose Y^{(1)}}...{x \choose Y^{(t)}}}. \numberthis \label{eq:thmexactsmap_proofterm0}
\end{align*}
We first simplify the numerator $\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose Y^{(1)}}...{x \choose Y^{(t)}}$; the denominator can be simplified using the same approach. Now,
\begin{align*}
\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose Y^{(1)}}...{x \choose Y^{(t)}} &\overset{(a)}{=} \sum_{\substack{x||x|=n,\\x_i=1}} \sum_{w\in \{0,1\}^*} {x \choose w} \langle Y^{(1)} \uparrow Y^{(2)} \uparrow ... \uparrow Y^{(t)},w \rangle \\
&=\sum_{w\in \mathcal A^*} \langle Y^{(1)} \uparrow Y^{(2)} \uparrow ... \uparrow Y^{(t)},w \rangle \sum_{\substack{x||x|=n,\\x_i=1}}{x \choose w}\\
&\overset{(b)}{=}\sum_{w\in \mathcal A^*} \langle Y^{(1)} \uparrow Y^{(2)} \uparrow ... \uparrow Y^{(t)},w \rangle \\
& \hspace{2cm} \times 2^{n-|w|}\left(\frac{1}{2}{n-1 \choose |w|}+\sum_{j|w_j=1}{i-1 \choose j-1}{n-i \choose |w|-j}\right)
\end{align*}
where $(a)$ is due to Lemma~\ref{lemma:bin_inf_relation} and $(b)$ due to Lemma~\ref{lemma:smapsum} (both introduced in \cite{Srini2018}); see Appendix~\ref{app:bin_inf_lemma} and Appendix~\ref{app:smapsum} for the statement and proof.
Note that Lemma~\ref{lemma:bin_inf_relation} is an alternate mathematical outlook of the channel equivalence in Theorem~\ref{thm:channel_equiv} (see Appendix~\ref{app:channel_equiv}).
\noindent Therefore we have,
\begin{align*}
\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose Y^{(1)}}...{x \choose Y^{(t)}}
&\overset{(a)}{=} \sum_{k=0}^{\infty} 2^{n-k-1}{n-1 \choose k} \sum_{w| |w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \\
&\hspace{1cm}+ \sum_{k=0}^{\infty} \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \\
&\overset{(b)}{=} \sum_{k=0}^n 2^{n-k-1}{n-1 \choose k} \sum_{w| |w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \\
&\hspace{1cm}+ \sum_{k=0}^n \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle, \numberthis
\label{eq:thmexactsmap_proofterm1}
\end{align*}
where in $(a)$ we first fix $|w|$ and then sum over all $w$ of the given length and $(b)$ holds because the combinatorial terms are $0$ when $k>n$.
A similar analysis gives \begin{align*}
\sum\limits_{x| |x|=n} &{x \choose Y^{(1)}}...{x \choose Y^{(t)}} = \sum_{k=0}^n 2^{n-k} {n \choose k} \sum_{\substack{w| |w|=k}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle.\numberthis
\label{eq:thmexactsmap_proofterm2}
\end{align*}
Plugging \eqref{eq:thmexactsmap_proofterm1} and \eqref{eq:thmexactsmap_proofterm2} in \eqref{eq:thmexactsmap_proofterm0}, we get the expression in Theorem~\ref{thm:exactSMAP_posteriorprob},
\begin{align*}
\Pr(X_i=1|Y^{(1)},&...,Y^{(t)}) &\\ =
& \Bigg[ \sum_{k=0}^n 2^{n-k-1}{n-1 \choose k} \sum_{w||w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \\
&+ \sum_{k=0}^n \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \Bigg] \Big/ \\
&\Bigg[ \sum_{k=0}^n 2^{n-k} {n \choose k} \sum_{\substack{w| |w|=k}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \Bigg].
\end{align*}
\end{proof}
\noindent \textbf{Step 2: Dynamic program to compute $\sum\limits_{w| |w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle$ and $\sum\limits_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle$.}
Note that the number of sequences $w| |w|=k$ is $O(2^k)$ so a naive evaluation is exponential in the blocklength $n$. We can, however, exploit the edit graph to come up with a dynamic program resulting in an algorithm which is polynomial in $n$.
Recall that in the edit graph, $\langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle$ is equal to the number of distinct paths from the origin $(0,...,0)$ to the destination $(|Y^{(1)}|,...,|Y^{(t)}|)$ and which correspond to $w$. Hence,
\begin{enumerate}[wide=0pt]
\item[(a)] $\sum\limits_{w| |w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle$ is the number of distinct paths of length $k$ from origin to destination and,
\item[(b)] $\sum\limits_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle$ is the number of such paths of length $k$ such that the $j^{th}$ edge of the path corresponds to a `1'.
\end{enumerate}
\noindent With this interpretation, the dynamic program for (a) follows naturally -- the number of $k$-length paths from the origin to any vertex is the sum of the number of $(k{-}1)$-length paths from the origin to all incoming neighbors of the vertex. To make this formal, associate a polynomial (in $\lambda$) for each vertex, such that the coefficient of $\lambda^k$
is equal to the number of paths of length $k$ from the origin to $v$: we call it the "forward-potential" polynomial $p^{for}_v(\lambda)$ for vertex $v$, the coefficient of $\lambda^k$ as earlier is denoted by $\langle p^{for}_v(\lambda),\lambda^k \rangle $. The dynamic program to compute $p^{for}_v(\lambda)$ for all $v$ can be expressed as:
\begin{equation}
p^{for}_v(\lambda) = \sum_{u|u\rightarrow v} \lambda p^{for}_u(\lambda).
\end{equation}
With this definition, we have $$\sum\limits_{w| |w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle =\langle p^{for}_{destination}(\lambda),\lambda^k \rangle.$$ In the example in Fig.~\ref{fig:editgraph_smap1}, one could do the following: order the vertices $(0,0)$ to $(3,3)$ lexicographically and then compute $p^{for}_v(\lambda)$ in the same order. Because of the directed grid nature of the edit graph, every vertex has incoming neighbors which are lexicographically ahead of itself. Also we initialize $p^{for}_{(0,0)}(\lambda)=1$. For the example in Fig.~\ref{fig:editgraph_smap1}, the forward-potentials are shown in Fig.~\ref{fig:editgraph_smap2}. The complexity of this dynamic program is $O(2^tn^{t+1})$ as it goes over $O(n^t)$ vertices and for each vertex it sums $O(2^t)$ polynomials, each of degree $O(n)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap2.pdf}
\caption{The forward-potential $p^{for}_v(\lambda)$ at each vertex.}
\label{fig:editgraph_smap2}
\end{figure}
We compute (b) as follows: pick an edge $(u{\rightarrow}v)$ which corresponds to `1', count the number of $(j{-}1)$-length paths from origin to $u$ and multiply it with the number of $(k{-}j)$-length paths from $v$ to the destination -- this is exactly the number of paths of length $k$ such that its $j^{th}$ edge is $(u{\rightarrow}v)$. Summing this term for all such edges which correspond to 1 gives us the term in (b). Note that we have already computed the number of $k$-length paths ($\forall k$) from origin to every vertex in $p^{for}_v(\lambda)$ . We can similarly compute the number of $k$-length paths ($\forall k$) from every vertex to the destination as $p^{rev}_v(\lambda)$ -- the "reverse potential" polynomial. The dynamic program for $p^{rev}_v(\lambda)$ is:
\begin{equation}
p^{rev}_v(\lambda) = \sum_{u|v\rightarrow u} \lambda p^{rev}_u(\lambda),
\end{equation}
with $p^{rev}_{destination}(\lambda)=1$. The reverse potentials for the example in Fig.~\ref{fig:editgraph_smap1} is shown in Fig.~\ref{fig:editgraph_smap3}. Like in the case of forward potential, we first order the vertices reverse lexicographically and then invoke the dynamic program above sequentially to compute the reverse potential polynomial at each vertex.
\begin{algorithm}[t!]
\caption{Computing the forward-potentials $p^{for}_u(\lambda)$ }\label{alg:forward_pot}
\begin{algorithmic}[1]
\item Input: Edit graph {$\mathcal G(Y^{(1)},...,Y^{(t)})$}\\ Outputs: $p^{for}_v(\lambda)\ \forall\ v$
\State Order the vertices from $(0,0,...,0)$ to $(|Y^{(1)}|,|Y^{(2)}|,...,|Y^{(t)}|)$ lexicogaphically; let the ordered list be $\mathcal V$
\State Initialise $p^{for}_{(0,...,0)}(\lambda)\gets 1$
\For{$v\ \in\ \mathcal V$}
\State \textbf{assign} $p^{for}_v(\lambda)\gets \sum_{u|u\rightarrow v} \lambda p^{for}_u(\lambda)$
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t!]
\caption{Computing the reverse-potentials $p^{rev}_u(\lambda)$ }\label{alg:reverse_pot}
\begin{algorithmic}[1]
\item Input: Edit graph {$\mathcal G(Y^{(1)},...,Y^{(t)})$}\\ Outputs: $p^{rev}_v(\lambda)\ \forall\ v$
\State Order the vertices from $(|Y^{(1)}|,|Y^{(2)}|,...,|Y^{(t)}|)$ to $(0,0,...,0)$ reverse lexicogaphically; let the ordered list be $\mathcal V$
\State Initialise $p^{rev}_{(|Y^{(1)}|,|Y^{(2)}|,...,|Y^{(t)}|)}(\lambda)\gets 1$
\For{$v\ \in\ \mathcal V$}
\State \textbf{assign} $p^{rev}_v(\lambda) \gets \sum_{u|v\rightarrow u} \lambda p^{rev}_u(\lambda)$
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap3.pdf}
\caption{The reverse-potential $p^{rev}_v(\lambda)$ at each vertex.}
\label{fig:editgraph_smap3}
\end{figure}
With this, the term in (b) can be expressed as:
\begin{align*}
\sum\limits_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle =\hspace{-2mm} \sum_{\substack{(u,v)|\\s(u\rightarrow v)=1}} \hspace{-2mm}\langle p^{for}_u(\lambda),\lambda^{j-1} \rangle \langle p^{rev}_v(\lambda),\lambda^{k-j} \rangle.
\end{align*}
Alg.~\ref{alg:exact_smap} now summarizes the computation of the posterior probabilities. This algorithm iterates over all the edges (we have $O((2n)^t)$ of these), and also $k,j$ ($O(n)$ each). The time complexity of Alg.~\ref{alg:exact_smap} hence is $O(2^tn^{t+2})$.
\begin{algorithm}[t!]
\caption{Symbolwise posterior probabilities with $t$ traces}
\label{alg:exact_smap}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{(1)},Y^{(2)},...,Y^{(t)}$, input length $n$}\\ Output: $\Pr(X_i=1|Y^{(1)},Y^{(2)},...,Y^{(t)})\ \forall\ i
\State Construct edit graph $\mathcal G(Y^{(1)},...,Y^{(t)})$
\State Use Alg.~\ref{alg:forward_pot} and Alg.~\ref{alg:reverse_pot} on $\mathcal G(Y^{(1)},...,Y^{(t)})$ to calculate $p^{for}_v(\lambda)$ and $p^{rev}_v(\lambda)$ $\forall\ v$
\For{$k \in\ [0:n]$}
\State \textbf{assign} $\sum\limits_{w| |w|=k} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \gets \langle p^{for}_{destination}(\lambda),\lambda^k \rangle.$
\For{each $j \in\ [1:n]$}
\State Initialize $temp \leftarrow $ 0
\For{each edge $u\rightarrow v\ \in\ \mathcal G$}
\If{$s(u{\rightarrow} v)=$ `1'}
\State $temp\ += \langle p^{for}_u(\lambda),\lambda^{j-1} \rangle \langle p^{rev}_v(\lambda),\lambda^{k-j} \rangle$
\EndIf
\EndFor
\State \textbf{assign} $\sum\limits_{\substack{w| |w|=k,\\w_j=1}} \langle Y^{(1)} \uparrow ... \uparrow Y^{(t)},w \rangle \gets temp$
\EndFor
\EndFor
\For{$i \in\ [1:n]$}
\State Use \eqref{eq:posterior_prob} to compute $\Pr(X_i=1|Y^{(1)},Y^{(2)},...,Y^{(t)})$
\EndFor
\end{algorithmic}
\end{algorithm}
\section{Notation and Tools}
\label{sec:notations}
\noindent \textbf{Basic notations:} In this work, we borrow a majority of the notation and tools from \cite{lothaire1997combinatorics} which deals with non-commutative algebra. We restate the definitions here for convenience. Calligraphic letters refer to sets, capitalized letters
correspond to random variables and bold letters are used for functions. Let $\mathcal{A}$ be the set of all
symbols. Throughout this work, we will focus on the case where $\mathcal{A}=\{0,1\}$,
though our methods extend to arbitrarily large sets of finite size. Define
$\mathcal{A}^n$ to be the set of all $n$-length sequences and $\mathcal{A}^*$ to be the set of all finite length sequences with
symbols in $\mathcal{A}$. For a sequence $f$, $|f|$ denotes the length
of $f$. For integers $i,j$, we define $[i:j] \triangleq \{i,i+1,...,j\}$ if $j\geq i$ and $[i:j] \triangleq \emptyset$ otherwise. Also define $[i] \triangleq [1:i]$.\\
\noindent \textbf{Binomial coefficient:} Given sequences $f$ and $g$ in $\mathcal{A}^*$, the number of
subsequence patterns of $f$ that are equal to $g$ is called the
\textit{binomial coefficient} of $g$ in $f$ and is denoted by $f
\choose g$. For example, ${'apple' \choose 'ape'} = 2$ since $'ape'$ can be obtained from two (overlapping) subsequences of $'apple'$.
When the alphabet $\mathcal{A}$ is of cardinality 1, ${f
\choose g} = {|f| \choose |g|}$, the classical binomial coefficient
with their respective lengths as the parameters. This definition hence
could be thought of as a generalization of the classical binomial
coefficients. We will denote by $e$ the sequence of length 0, and {define ${f
\choose e} \triangleq 1\ \forall\ f\ \in\ \mathcal{A}^*$.} We also define the classical binomial coefficient ${a \choose b}\triangleq 0,$ whenever $b>a$ or $b<0$ for ease of use.
The binomial coefficient is an integral aspect of this work and for analyzing error events in deletion channels because the input-output relation for a deletion channel (with deletion probability $\delta$, input $X$ and output $Y$) can be expressed as
\begin{equation}
\label{eq:in-out_relation}
\Pr(Y=y|X=x) = {x \choose y} \delta^{|x|-|y|} (1-\delta)^{|y|}.
\end{equation}
The proof is straight-forward -- the number of distinct error events that give rise to $y$ from $x$ is exactly the number of subsequences of $x$ which are equal to $y$. Each of these error events has a probability $\delta^{|x|-|y|} (1-\delta)^{|y|}$, wherein the exponent of $\delta$ corresponds to the deleted symbols and the exponent of $1-\delta$ to the undeleted symbols.
Given the definition of the binomial coefficient, the maximum-likelihood (ML) estimate over a deletion channel with observed output $Y$ can be cast in the following form:
\begin{align*}
\argmax_{x \in \mathcal C} {x \choose Y},\numberthis
\label{eq:ML_deletion}
\end{align*}
where $\mathcal C$ is the chosen codebook. In the case of multiple deletion channels with observed traces $Y^{(1)},...,Y^{(t)}$, the ML formulation is similar:
\begin{align*}
\argmax_{x \in \mathcal C} \prod_{j=1}^{t} {x \choose Y^{(j)}}.\numberthis
\label{eq:ML_multiple_deletion}
\end{align*}
As yet, there is no known efficient way to come up with a solution for the above two formulations, even for \eqref{eq:ML_deletion} with $\mathcal C = \{0,1\}^n$ (see \cite{mitzenmacher2009survey}). In this work, we attempt to take a step in this direction by showing that a continuous relaxation of \eqref{eq:ML_deletion} is equivalent to \eqref{eq:ML_deletion}. However, an efficient algorithm to solve the optimization \eqref{eq:ML_deletion} remains open. In the context of trace reconstruction, the ultimate pursuit would be an algorithm for \eqref{eq:ML_multiple_deletion} with $\mathcal C = \{0,1\}^n$ and error analysis thereof.
\noindent We now describe a function which can be thought of as a real-valued extension of the binomial coefficient. This function is used in sections~\ref{sec:1deletion_ML} and \ref{sec:1deletion}. Consider the function $\mathbf F(\cdot)$ defined as:
\begin{definition}
\label{def:f}
\begin{align*}
&\mathbf F: \mathbb{R}^k \times \{0,1\}^l \rightarrow \mathbb{R},\\
\mathbf F(q, v)\triangleq &\begin{cases}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [k],\\|\mathcal S|=l}} \quad \prod\limits_{i=1}^l q_{\mathcal S_i}^{v_i} (1-q_{\mathcal S_i})^{1-v_i} & 1 \leq l\leq k \\
1 & 0=l\leq k \\
0 & \text{else}.
\end{cases}
\end{align*}
\noindent An alternate definition is as follows: consider a random vector $Z\in\{0,1\}^k$ such that $Z_i\sim$ ind. Ber$(q_i)$, let $q$ be the vector of probabilities of length $k$, $\mathcal S\subseteq [k]$ a subset of the indices of size $l$, and ${v}$ a binary vector of length $l$. Then, $$\mathbf F(q,v)=\sum_{\substack{\mathcal S||\mathcal S|=l}} \Pr(Z_{\mathcal S}= v).$$ Note that if $q$ is a vector of $0$'s and $1$'s, then $\mathbf F( q, v)={q \choose v}$, the binomial coefficient of $v$ in $q$. Thus, $\mathbf F(\cdot)$ could be interpreted as an extension of the binomial coefficient where one of the parameters can take values in $[0,1]^n$ instead of $\{0,1\}^n$.
\end{definition}
Though at first sight, $\mathbf F(q,v)$ sums over an exponential number of subsets, a dynamic programming approach can be used to compute it in $O(|v|^2)$ time complexity. The dynamic program is described in section~\ref{sec:1deletion}.
\iffalse
\noindent \textbf{Edit distance:} The edit distance $d_e(f,g)$ measures similarity between two sequences of
possibly different lengths \cite{Navarro2001}. $d_e(f,g)$ is the minimum number of operations
needed to transform $f$ to $g$, where the permitted operations are
insertion, deletion or substitution of a symbol. In this work, we
quantify the performance of algorithms in Section \ref{sec:Numerics}
using the edit distance metric.
\fi
The following definitions and ideas are relevant only for \ref{sec:exactsmap} and can be omitted for the sections on single deletion channel. Before getting to the mathematical definitions, we first state a result of ours that aids in thinking about error events in multiple deletion channels.
\subsection{An alternate interpretation of multiple deletion channel model}
The events occurring in the $t$-deletion channel model can be categorized into two groups:
\begin{enumerate}
\item an input symbol is deleted in \textit{all} the $t$-traces,
\item an input symbol is reflected in at least one of the traces.
\end{enumerate}
The error events of the first kind are in some sense ``not correctable" or even ``detectable" since it is impossible to tell what and where the deleted symbol could have been (although the probabilities need not be uniform). The events of the second kind, however, can still be detected although they could likely be confused with a large set of similar events. This thought process gives rise to a natural decomposition of the $t$-deletion channel model into a cascade of two channels: the first one being a deletion channel which captures error events of the first kind and the second one is what we call the \textit{remnant channel} which captures events of the second kind (see Fig.~\ref{fig:channel_equiv}). More precisely, the remnant channel is defined as follows:
\begin{definition}
\textit{Remnant channel:} an input symbol to the remnant channel is reflected in $k>0$ given outputs and deleted in the rest with a probability $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
\end{definition}
It is easy to note that probability of the union of all possible events here is $\sum_{k=1}^t {t \choose k} \frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}=1$, validating our definition.
\begin{restatable}{theorem}{channelequiv}
\label{thm:channel_equiv}
The $t$-deletion channel model and the cascade of the deletion channel with remnant channel shown in Fig.~\ref{fig:channel_equiv} are probabilistically equivalent, i.e.,
$$\Pr({Y}^{(1)},{Y}^{(2)},...,{Y}^{(t)}|X = x) = \Pr(\tilde{Y}^{(1)},\tilde{Y}^{(2)},...,\tilde{Y}^{(t)}|X = x).$$
\end{restatable}
The formal proof of the theorem requires few more definitions and is relegated to the appendix.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.4]{channel_equiv.pdf}
\caption{A channel equivalence result: the $t$-deletion channel model in (a) is probabilistically equivalent to the the cascade of a deletion channel with the \textit{remnant channel} ($ \mathcal C_2$) in (b).}
\label{fig:channel_equiv}
\end{center}
\end{figure}
\noindent\textbf{Edit graph} (as defined in \cite{Gusfield1997}): We now define a graph construct which is closely related to the error events in the remnant channel. We start with a simple case and generalize subsequently. Define an \textit{edit graph} given two sequences $f$ and $g$, where every path connecting the ``origin'' to the ``destination'' on the edit graph yields a supersequence $h$ of $f,g$, where $h$ is ``covered'' by $f,g$ -- i.e., each symbol of $h$ comes from either $f$ or $g$ or
both. In other words, given that $f$ and $g$ are the outputs of the remnant channel (with two outputs), each path from the origin of the edit graph to the destination corresponds to a possible input $h$ to the remnant channel and to an error event which resulted in $f$ and $g$ with input $h$.
For $f$ and $g$ in $\mathcal
A^*$, we form a graph $\mathcal{G}(f,g)$ with $(|f|+1)(|g|+1)$
vertices each labelled with a distinct pair $(i,j), 0\leq i\leq
|f|,\ 0\leq j\leq |g|$. A directed edge
$(i_1,j_1)\rightarrow(i_2,j_2)$ exists iff at least one of the following
holds:
\begin{enumerate}
\item$i_2-i_1=1$ and $j_1=j_2$, or
\item $j_2-j_1=1$ and $i_1=i_2$, or
\item $i_2-i_1=1$, $j_2-j_1=1$ and $f_{i_2}=g_{j_2}$,
\end{enumerate}
where $f_i$ is the $i^{th}$ symbol of the sequence $f$. The origin is the vertex $(0,0)$ and the destination $(|f|,|g|)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap1.pdf}
\caption{ Edit graph for sequences $f=$ `001' and $g=$ `101'. An easy way to think about this is to write down $f$ vertically with each symbol aligned to a vertical set of edges and $g$ horizontally likewise. A diagonal edge in a small square exists if the corresponding $f$ and $g$ symbols are equal. The thick red edges form a path from the origin to the destination; this path corresponds to the sequence `0101' -- append the corresponding symbol at the left of an edge if it's vertical or diagonal, otherwise append the symbol at the top. It is also easy to verify that 0101 is a supersequence of both $f$ and $g$, and could be obtained as a covering of $f$ and $g$; the path itself gives one such covering. This covering also corresponds to an error event in the remnant channel which would result in outputs $f$ and $g$ with input $h=$ `0101' -- more precisely, the error event is one in which the first symbol of $h$ is deleted only in $g$, the second symbol is deleted only in $f$ and the third and fourth symbols not deleted in either $f$ or $g$.}
\label{fig:editgraph_smap1}
\end{figure}
Let $p=((i_1,j_1),(i_2,j_2),...,(i_m,j_m))$ be a path in $\mathcal{G}(f,g)$. We define $ s(p)$ to be the sequence corresponding to the path. Intuitively, $s(p)$ is formed by appending symbols in the following
way: append the corresponding $f$ symbol for a vertical edge, $g$ symbol for horizontal edge, and $f$ or $g$ symbol for diagonal edge (see example Fig.~\ref{fig:editgraph_smap1}). Any path from $(0,0)$ to $(|f|,|g|)$ corresponds to a supersequence of $f$ and $g$ and which is ``covered" by $f$ and $g$. More formally, define $ s(p)\triangleq x_1x_2...x_{m-1}$ where
$$x_k =
\begin{cases}
f_{i_{k+1}} \quad\text{if }j_{k}=j_{k+1},\\
g_{j_{k+1}} \quad\text{if }i_{k}=i_{k+1},\\
f_{i_{k+1}} \quad\text{else.}
\end{cases}
$$
The construct of edit graph can be extended to more than 2 sequences with the same idea. For sequences $f_1,f_2,...,f_t$, construct a $t$-dimensional grid with a number of vertices $(|f_1|+1)(|f_2|+1)...(|f_t|+1)$ labeled from $(0,0,...,0)$ to $(|f_1|,|f_2|,...,|f_t|)$. A vertex $u=(i_1,i_2,...,i_t)$ is connected to $v=(j_1,j_2,...,j_t)$ (we say $u \rightarrow v$) iff both of the following conditions are met:
\begin{itemize}
\item $j_l=i_l$ or $j_l=i_l+1$ $\forall\ l\in [t]$, i.e., $(i_1,...,i_t)$ and $(j_1,...,j_t)$ are vertices of a particular unit cube. Only these type of vertices can share an edge in the grid graph.
\item Let $\mathcal T \subseteq [t]$ be the collection of indices where $j_l=i_l+1$. Then $f_{i_l}$ is same $\forall\ l \in \mathcal T$. For instance, if $\mathcal T=\{1,3,5\}$, then $f_{i_1}=f_{i_3}=f_{i_5}$.
\end{itemize}
Define the vertex $(0,...,0)$ to be the origin of this graph and the vertex $(|f_1|,...,|f_t|)$ to be the destination. If $|f_j|=O(n)\ \forall\ j$, this graph has a number of vertices $O(n^t)$ and a maximum number of edges $O((2n)^t)$ since each vertex has at most $2^t$ outgoing edges.\\
\noindent\textbf{Infiltration product} (introduced in \cite{lothaire1997combinatorics}):
The infiltration product has been extensively used in \cite{lothaire1997combinatorics}, as a tool in non-commutative algebra. Here we give an edit-graph interpretation of this tool. We also give a formal definition later in this section. Using the edit graph we can construct the set of possible
supersequences $\mathcal{S}(f,g)$ of $f,g$ which are covered by
it. Clearly multiple paths could yield the same supersequence and we
can count the number of distinct ways $\mathbf N(h;f,g)$ one can construct
the same supersequence $h$ from $f,g$.
We can informally define the
\emph{infiltration product $f\uparrow g$} of $f$ and $g$, as a polynomial with monomials the supersequences $h$ in
$\mathcal{S}(f,g)$ and coefficients $\langle f\uparrow g,h\rangle$ equal to $\mathbf N(h;f,g)$. In Fig.~\ref{fig:editgraph_smap1}, it is easy to verify that there is exactly one path corresponding to `101001' and hence $\langle 001\uparrow 101,101001 \rangle=1$ and similarly $\langle 001\uparrow 101,01001 \rangle=2$. One could find these coefficients for all relevant sequences and form the polynomial as described. More examples: Let $\mathcal{A}=\{a,b\}$, then
\begin{itemize}[wide=2pt]
\item $ab\uparrow ab=ab+2aab+2abb+4aabb+2abab$,
\item $ab\uparrow ba=aba+bab+abab+2abba+2baab+baba.$
\end{itemize}
The infiltration operation is commutative and associative, and infiltration of two sequences $f\uparrow g$ is a polynomial with variables of length (or \textit{degree}) at most $|f|+|g|$; see \cite{lothaire1997combinatorics}.
The definition of infiltration extends to two polynomials via distributivity (precisely defined in later), and consequently to multiple sequences as well. For multiple sequences, infiltration has the same edit graph interpretation: $\langle f_1\uparrow f_2 \uparrow...\uparrow f_t, w \rangle$ is the number of distinct ways of constructing $w$ as a supersequence of $f_1, f_2, ... ,f_t$ so that the construction covers $w$, i.e., construct the $t$-dimensional edit graph of $f_1, f_2, ... ,f_t$ and count the number of paths corresponding to $w$.
\subsection{Formal definition of infiltration product}
We now give a more formal definition of the infiltration product (see \cite{lothaire1997combinatorics} for the equivalence of the two definitions and a more rigorous treatment). A \textit{formal series} with indeterminates (or variables) in a set $\mathcal A$ and coefficients in a commutative ring $\mathcal R$, is a mapping of $\mathcal A^*$ onto $\mathcal R$. Recall that a commutative ring is a set which forms an abelian group under an \textit{addition} operation, is a monoid under a \textit{multiplication} operation which commutes, and the multiplication operation distributes over addition. Here we consider $\mathbb Z$, the set of integers as the commutative ring $\mathcal{R}$. A formal series is called a \textit{polynomial} if only a finite number of sequences are mapped to non-zero values, the rest of the sequences map to zero. Consider two polynomials $\sigma,\tau: \mathcal{A}^*\rightarrow \mathbb Z$. The value taken by a sequence $w\in \mathcal A^*$ on $\sigma$ (or the coefficient of $w$ in $\sigma$) is denoted by $\langle \sigma,w\rangle \in \mathbb R$. We also define binary addition ($\oplus$) and multiplication operations ($\times$) on the set of polynomials as follows:
\begin{align}
\langle \sigma\oplus \tau,w\rangle \triangleq \langle \sigma,w\rangle + \langle \tau,w \rangle \quad \forall w\in \mathcal A^*,\label{eq:polynomial_add}\\
\langle \sigma\times \tau,w\rangle \triangleq \sum_{\substack{f,g\in \mathcal A^*:\\ f.g=w}}\langle \sigma,f\rangle \langle \tau,g \rangle \quad \forall w\in \mathcal A^*.\label{eq:polynomial_prod}
\end{align}
We will use the usual symbols $+$ and $.$ in place of $\oplus$ and $\times$ in this work for convenience. The meaning of the operation would be clear depending on the operands. With these operations the set of polynomials form a non-commutative ring, and is denoted by $\mathbb Z\langle\mathcal A \rangle$, also called the free $\mathbb Z$-algebra on $\mathcal A$ in ring theory. Note that the addition and multiplication operations defined in \eqref{eq:polynomial_add} and \eqref{eq:polynomial_prod} are similar to the operations defined on commutative polynomials, except that the multiplication operation under the summation in \eqref{eq:polynomial_prod} ($f.g=w$) is actually concatenation and is non-commutative. The multiplication inside the summation in \eqref{eq:polynomial_prod} is multiplication in the real field and hence commutative. It is also easy to prove that the multiplication defined in \eqref{eq:polynomial_prod} distributes over addition defined in \eqref{eq:polynomial_add}. Thus, a polynomial in $\mathbb Z\langle\mathcal A \rangle$ can be represented as a sum of monomials in $\mathcal A^*$ each with an associated coefficient in $\mathbb Z$, i.e., $\sigma=\sum\limits_{w\in \mathcal A^*} \langle\sigma,w \rangle w$. Define the \textit{degree} of a polynomial to be equal to the length of a longest sequence with a non-zero coefficient in the polynomial and the \textit{number of terms} of a polynomial as the number of sequences with non-zero coefficients in the polynomial. Note that a degree $d$ polynomial could have a number of terms upto $2^{d+1}-1$.
With this, the \textit{infiltration product} (in general, for two polynomials) is defined as follows:
\begin{align}
\forall f\in \mathcal{A}^*,& \quad f\uparrow e = e\uparrow f=f.\nonumber \\
\forall f,g\in \mathcal{A}^*&,\quad \forall a,b\in \mathcal{A}, \nonumber\\
fa\uparrow gb=(f\uparrow gb)a&+(fa\uparrow g)b+\mathbbm{1}_{a=b}(f\uparrow g)a.\nonumber \\
\forall \sigma,\tau \in \mathbb{Z}\langle\mathcal{A} \rangle, \quad &\sigma\uparrow \tau=\sum_{f,g\in \mathcal{A}^*} \langle \sigma,f \rangle \langle \tau,g \rangle (f\uparrow g). \label{def:infiltforseq}
\end{align}
\textbf{Summary of definitions and ideas introduced in this section:}
\begin{itemize}
\item The binomial coefficient captures the likelihood of observations for deletion channels.
\item An extension of the binomial coefficient function where one of the parameters can take real values has been introduced; this function is pivotal for our results on the single-trace deletion channel.
\item For multiple deletion channels, the error events can be categorized into two groups -- one where an input symbol is deleted in all the traces, and second, the complement of this event. This categorization results in a natural decomposition of multiple deletion channel model into a cascade model involving the remnant channel.
\item The remnant channel disregards the error events where a symbol is deleted in all the traces.
\item The edit graph provides a way to visualize all the possible error events and input sequences to a remnant channel given its outputs.
\item The infiltration product serves the same purpose as the edit graph, but has an algebraic flavor and provides rigor for proofs and analyses. The edit graph, on the other hand, is more helpful in designing reconstruction algorithms over deletion channels.
\end{itemize}
\section{Notation and Tools}
\label{sec:notation}
\noindent \textbf{Basic notation:} We borrow some notation from \cite{lothaire1997combinatorics} which deals with non-commutative algebra; we restate them here for convenience. Calligraphic letters refer to sets, capitalized letters
correspond to random variables and bold letters are used for functions. Let $\mathcal{A}$ be the set of all
symbols. Throughout this work, we will focus on the case where $\mathcal{A}=\{0,1\}$,
though our methods extend to arbitrarily large sets of finite size. Define
$\mathcal{A}^n$ to be the set of all $n$-length sequences and $\mathcal{A}^*$ to be the set of all finite length sequences with
symbols in $\mathcal{A}$. For a sequence $f$, $|f|$ denotes the length
of $f$.
For integers $i,j$, we define $[i:j] \triangleq \{i,i+1,...,j\}$ if $j\geq i$ and $[i:j] \triangleq \emptyset$ otherwise. We also define $[i] \triangleq [1:i]$.
For a vector or sequence $x=(x_1,x_2,...,x_{i-1},x_i,x_{i+1},...,x_n)$, define $$x^{(i\rightarrow s)}\triangleq (x_1,x_2,...,x_{i-1},s,x_{i+1},...,x_n),$$ where the $i^{th}$ coordinate of $x$ is replaced by symbol $s$. \\
\noindent \textbf{Binomial coefficient \cite{lothaire1997combinatorics}:} Given sequences $f$ and $g$ in $\mathcal{A}^*$, the number of
subsequence patterns of $f$ that are equal to $g$ is called the
\textit{binomial coefficient} of $g$ in $f$ and is denoted by $f\choose g$. For example, ${'apple' \choose 'ape'} = 2$ since $'ape'$ can be obtained from two (overlapping) subsequences of $'apple'$.
When the alphabet $\mathcal{A}$ is of cardinality 1, ${f \choose g} = {|f| \choose |g|}$, the classical binomial coefficient with their respective lengths as the parameters. This definition hence
could be thought of as a generalization of the classical binomial
coefficients. We will denote by $e$ the sequence of length 0, and {define ${f
\choose e} \triangleq 1\ \forall\ f\ \in\ \mathcal{A}^*$.} We also define the classical binomial coefficient ${a \choose b}\triangleq 0,$ whenever $b>a$ or $b<0$ for ease of use.
The binomial coefficient forms the backbone for the probabilistic analysis of deletion channels since the input-output relation for a deletion channel (with deletion probability $\delta$, input $X$ and output $Y$) can be expressed as
\begin{equation}
\label{eq:in-out_relation}
\Pr(Y=y|X=x) = {x \choose y} \delta^{|x|-|y|} (1-\delta)^{|y|}.
\end{equation}
The proof is straightforward -- the number of distinct error events that give rise to $y$ from $x$ is exactly the number of subsequences of $x$ which are equal to $y$. Each of these error events has a probability $\delta^{|x|-|y|} (1-\delta)^{|y|}$, wherein the exponent of $\delta$ corresponds to the deleted symbols and the exponent of $1-\delta$ to the undeleted symbols. \\
\noindent \textbf{Maximum Likelihood (ML) estimate:}
Given the definition of the binomial coefficient, the maximum-likelihood (ML) estimate over a deletion channel with observed output $Y$ can be cast in the following form:
\begin{align*}
\argmax_{x \in \{0,1\}^n} {x \choose Y}.\numberthis
\label{eq:ML_deletion}
\end{align*}
In the case of multiple deletion channels with observed traces $Y^{(1)},...,Y^{(t)}$, the ML formulation is similar:
\begin{align*}
\argmax_{x \in \{0,1\}^n} \prod_{j=1}^{t} {x \choose Y^{(j)}}.\numberthis
\label{eq:ML_multiple_deletion}
\end{align*}
As yet, there is no known efficient way to come up with a solution for either of the above two formulations (see \cite{mitzenmacher2009survey}).\\
\noindent \textbf{Relaxed binomial coefficient.}
We now introduce the function $\mathbf F(\cdot)$ which can be thought of as a real-valued relaxation of the binomial coefficient. This function is used in sections~\ref{sec:1deletion_ML} and~\ref{sec:1deletion}.
An intuitive definition is as follows: Consider a random vector $Z\in\{0,1\}^n$ such that $Z_i\sim$ ind. Ber$(p_i)$, let $p$ be the vector of probabilities of length $n$, $\mathcal S\subseteq [n]$ a subset of the indices of size $m$, and ${v}$ a binary vector of length $m$. Then, $$\mathbf F(p,v)=\sum_{\substack{\mathcal S||\mathcal S|=m}} \Pr(Z_{\mathcal S}= v).$$
It is easy to see that $\mathbf F(p,v)=\mathbb E_{Z\sim p}\ {Z \choose v}$; in words, suppose $Z$ is sampled from a Bernoulli distribution parametrized by $p$, $\mathbf F(p,v)$ is the expected number of times $v$ appears as a subsequence of $Z$. Clearly if $p \in \{0,1\}^n$, then $Z=p$ with probability 1 and $\mathbf F(p,v) = {p \choose v}$.
More precisely, $\mathbf F(\cdot)$ is defined as:
\begin{definition}
\label{def:f}
\begin{align*}
&\mathbf F: [0,1]^n \times \{0,1\}^m \rightarrow \mathbb{R},\\
\mathbf F(p, v)\triangleq &\begin{cases}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m}} \quad \prod\limits_{i=1}^m p_{\mathcal S_i}^{v_i} (1-p_{\mathcal S_i})^{1-v_i} & 1 \leq m\leq n \\
1 & 0=m\leq n \\
0 & \text{else}.
\end{cases}
\end{align*}
\end{definition}
Though at first sight $\mathbf F(p,v)$ sums over an exponential number of subsets, a dynamic programming approach can be used to compute it in $O(nm)$ time complexity (see Appendix~\ref{app:F_compute}).\\
\noindent \textbf{Decomposition of the $t$-trace deletion channel:}
The following definitions and ideas are relevant only for Section \ref{sec:exactsmap} and can be omitted for the sections on single-trace deletion channel. We first state a result of ours that aids in thinking about error events in multiple deletion channels.
The events occurring in the $t$-deletion channel model can be categorized into two groups:
\begin{enumerate}
\item an input symbol is deleted in \textit{all} the $t$-traces,
\item an input symbol is reflected in at least one of the traces.
\end{enumerate}
The error events of the first kind are in some sense ``not correctable'' or even ``detectable'' in any situation since it is impossible to tell with absolute certainty what and where the deleted symbol could have been (although the probabilities need not be uniform). The events of the second kind, however, can be detected and corrected in some situations. This thought process gives rise to a natural decomposition of the $t$-deletion channel model into a cascade of two channels: the first one being a deletion channel which captures error events of the first kind and the second one is what we call the \textit{remnant channel} which captures events of the second kind (see Fig.~\ref{fig:channel_equiv}). More precisely, we define the remnant channel as follows:
\begin{definition}
\textit{Remnant channel:} an input symbol to the remnant channel is reflected in $k>0$ given outputs and deleted in the rest with a probability $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
\end{definition}
It is easy to note that probability of the union of all possible events here is $\sum_{k=1}^t {t \choose k} \frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}=1$, validating our definition.
\begin{restatable}{theorem}{channelequiv}
\label{thm:channel_equiv}
The $t$-deletion channel model and the cascade of the deletion channel with remnant channel shown in Fig.~\ref{fig:channel_equiv} are probabilistically equivalent, i.e.,
$$\Pr({Y}^{(1)},{Y}^{(2)},...,{Y}^{(t)}|X = x) = \Pr(\tilde{Y}^{(1)},\tilde{Y}^{(2)},...,\tilde{Y}^{(t)}|X = x).$$
\end{restatable}
An algebraic proof of the theorem requires a few more definitions and is relegated to Appendix~\ref{app:channel_equiv}.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.4]{channel_equiv.pdf}
\caption{A channel equivalence result: the $t$-trace deletion channel model in (a) is probabilistically equivalent to the the cascade of a deletion channel with the \textit{remnant channel} ($ \mathcal C_2$) in (b).}
\label{fig:channel_equiv}
\end{center}
\end{figure}
\noindent\textbf{Edit graph} (\cite{Gusfield1997}): Similar graph constructs have been defined in related problems on common supersequences and subsequences (see \cite{Nicosia2001} for example). This graph is closely related to the error events in the remnant channel. We start with a simple case and generalize subsequently. Define an \textit{edit graph} given two sequences $f$ and $g$, where every path connecting the ``origin'' to the ``destination'' on the edit graph yields a supersequence $h$ of $f,g$, where $h$ is ``covered'' by $f,g$ -- i.e., each symbol of $h$ comes from either $f$ or $g$ or
both. In other words, given that $f$ and $g$ are the outputs of the remnant channel (with two outputs), each path from the origin of the edit graph to the destination corresponds to a possible input $h$ to the remnant channel and to an error event which resulted in outputs $f,g$ with input $h$.
For $f$ and $g$ in $\mathcal
A^*$, we form a graph $\mathcal{G}(f,g)$ with $(|f|+1)(|g|+1)$
vertices each labelled with a distinct pair $(i,j), 0\leq i\leq
|f|,\ 0\leq j\leq |g|$. A directed edge
$(i_1,j_1)\rightarrow(i_2,j_2)$ exists iff at least one of the following
holds:
\begin{enumerate}
\item$i_2-i_1=1$ and $j_1=j_2$, or
\item $j_2-j_1=1$ and $i_1=i_2$, or
\item $i_2-i_1=1$, $j_2-j_1=1$ and $f_{i_2}=g_{j_2}$,
\end{enumerate}
where $f_i$ is the $i^{th}$ symbol of the sequence $f$. The origin is the vertex $(0,0)$ and the destination $(|f|,|g|)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap1.pdf}
\caption{ Edit graph for sequences $f=$ `001' and $g=$ `101'. Make a grid so the vertical edges are aligned with a symbol in $f$ and horizontal edges with $g$ as shown. A diagonal edge $(i{-}1,j{-}1)\rightarrow (i,j)$ exists if $f_i = g_j$. The thick red edges form a path from the origin to the destination; this path corresponds to $h=$`0101' -- sequentially append the corresponding symbol to which each edge is aligned. It is also easy to verify that $h$ is a supersequence of both $f$ and $g$, and could be obtained as a covering of $f$ and $g$; the path itself gives one such covering. This covering also corresponds to an error event (or a deletion pattern) in the remnant channel which would result in outputs $f$ and $g$ with input $h=$ `0101' -- the deletion pattern is shown in the figure.}
\label{fig:editgraph_smap1}
\end{figure}
Let $p=((i_1,j_1),(i_2,j_2),...,(i_m,j_m))$ be a path in $\mathcal{G}(f,g)$. We define $ s(p)$ to be the sequence corresponding to the path. Intuitively, $s(p)$ is formed by appending symbols in the following
way: append the corresponding $f$ symbol for a vertical edge, $g$ symbol for horizontal edge, and $f$ or $g$ symbol for diagonal edge (see example Fig.~\ref{fig:editgraph_smap1}). Any path from $(0,0)$ to $(|f|,|g|)$ corresponds to a supersequence of $f$ and $g$ and which is covered by $f$ and $g$. More formally, define $ s(p)\triangleq x_1x_2...x_{m-1}$ where
$$x_k =
\begin{cases}
f_{i_{k+1}} \quad\text{if }j_{k}=j_{k+1},\\
g_{j_{k+1}} \quad\text{if }i_{k}=i_{k+1},\\
f_{i_{k+1}} \quad\text{else.}
\end{cases}
$$
The construct of edit graph can be extended to more than 2 sequences with the same idea. For sequences $f_1,f_2,...,f_t$, construct a $t$-dimensional grid with a number of vertices $(|f_1|+1)(|f_2|+1)...(|f_t|+1)$ labeled from $(0,0,...,0)$ to $(|f_1|,|f_2|,...,|f_t|)$. A vertex $u=(i_1,i_2,...,i_t)$ is connected to $v=(j_1,j_2,...,j_t)$ (we say $u \rightarrow v$) iff both of the following conditions are met:
\begin{itemize}
\item $j_l=i_l$ or $j_l=i_l+1$ $\forall\ l\in [t]$, i.e., $(i_1,...,i_t)$ and $(j_1,...,j_t)$ are vertices of a particular unit cube. Only these type of vertices can share an edge in the grid graph.
\item Let $\mathcal T \subseteq [t]$ be the collection of indices where $j_l=i_l+1$. Then ${f_l}_{j_l}$ is equal $\forall\ l \in \mathcal T$. For example in 4 dimensional grid, consider the two vertices $(10,5,8,2)$ and $(10,6,9,2)$. In this case $\mathcal T = \{2,3\}$ since the second and third coordinates differ by 1. Therefore $(10,5,8,2)\rightarrow (10,6,9,2)$ iff ${f_2}_{5}={f_3}_{9}$. Note that if only one coordinate differs by 1 in the two vertices, a directed edge always exists (in other words all non-diagonal edges exist).
\end{itemize}
Define the vertex $(0,...,0)$ to be the origin of this graph and the vertex $(|f_1|,...,|f_t|)$ to be the destination. If $|f_j|=O(n)\ \forall\ j$, this graph has a number of vertices $O(n^t)$ and a maximum number of edges $O((2n)^t)$ since each vertex has at most $2^t-1$ outgoing edges.\\
\noindent\textbf{Infiltration product} (introduced in \cite{lothaire1997combinatorics}):
The infiltration product has been extensively used in \cite{lothaire1997combinatorics}, as a tool in non-commutative algebra. Here, we give an edit-graph interpretation of this tool. A formal algebraic definition of the infiltration product is in Appendix~\ref{app:infil_def}. Using the edit graph we can construct the set of possible
supersequences $\mathcal{S}(f,g)$ of $f$, $g$ that are covered by the symbols in $f$ and $g$.
Clearly multiple paths could yield the same supersequence and we
can count the number of distinct ways $\mathbf N(h;f,g)$ one can construct
the same supersequence $h$ from $f$, $g$.
We can informally define the
\emph{infiltration product $f\uparrow g$} of $f$ and $g$, as a polynomial with monomials the supersequences $h$ in
$\mathcal{S}(f,g)$ and coefficients $\langle f\uparrow g,h\rangle$ equal to $\mathbf N(h;f,g)$. For the example in Fig.~\ref{fig:editgraph_smap1}, it is easy to verify that there is exactly one path corresponding to `101001' and hence $\langle 001\uparrow 101,101001 \rangle=1$ and similarly $\langle 001\uparrow 101,01001 \rangle=2$. One could find these coefficients for all relevant sequences and form the polynomial as described. As additional examples, let $\mathcal{A}=\{a,b\}$, then
\begin{itemize}[wide=2pt]
\item $ab\uparrow ab=ab+2aab+2abb+4aabb+2abab$,
\item $ab\uparrow ba=aba+bab+abab+2abba+2baab+baba.$
\end{itemize}
The infiltration operation is commutative and associative, and infiltration of two sequences $f\uparrow g$ is a polynomial with variables of length (or \textit{degree}) at most $|f|+|g|$; see \cite{lothaire1997combinatorics}.
The definition of infiltration extends to two polynomials via distributivity (precisely defined in Appendix~\ref{app:infil_def}), and consequently to multiple sequences as well. For multiple sequences, infiltration has the same edit graph interpretation: $\langle f_1\uparrow f_2 \uparrow...\uparrow f_t, w \rangle$ is the number of distinct ways of constructing $w$ as a supersequence of $f_1, f_2, ... ,f_t$ so that the construction covers $w$, i.e., construct the $t$-dimensional edit graph of $f_1, f_2, ... ,f_t$ and count the number of paths corresponding to $w$.\\
\noindent \textbf{Summary of definitions and ideas introduced in this section:}
\begin{itemize}
\item The binomial coefficient captures the likelihood of observations for deletion channels.
\item A real-valued extension of the binomial coefficient function is introduced; this function is pivotal to our results on the single-trace deletion channel.
\item For multiple deletion channels, the error events can be categorized into two groups -- one where an input symbol is deleted in all the traces, and second, the complement of this event. This categorization results in a natural decomposition of multiple deletion channel model into a cascade model involving the remnant channel.
\item The remnant channel disregards the error events where a symbol is deleted in all the traces.
\item The edit graph provides a way to visualize and construct all the possible error events and input sequences to a remnant channel given its outputs.
\item The infiltration product serves the same purpose as the edit graph, but has an algebraic flavor and provides rigor for proofs and analysis. The edit graph, on the other hand, is more helpful in designing algorithms.
\end{itemize}
\section{Maximum likelihood for the single-trace deletion channel}
\label{sec:1deletion_ML}
Consider the ML formulation for the single-trace deletion channel given non-empty trace $Y$ and all possible $n$-length inputs allowed, restated here for convenience:
\begin{equation}
\argmax_{x\in \{0,1\}^n} {x \choose Y}.
\label{eq:1deletion_ML}
\end{equation}
To the best of our knowledge, the only known method to solve \eqref{eq:1deletion_ML} involves iterating over all possible choices of $x$ and computing the objective value for each of the choices. We here show that it is sufficient to solve a continuous relaxation of above problem to obtain a solution to \eqref{eq:1deletion_ML}. Note that there could be multiple solutions to \eqref{eq:1deletion_ML}. Before going in the main result, we first state a useful lemma which factors out a given coordinate $p_i$ out of $\mathbf F(p,Y)$. The proof of the lemma is relegated to the appendix.
\begin{restatable}{lemma}{deletionMLrellemma}
For $p=[p_1,p_2,..,p_i,...,p_n]$ and $Y=Y_1Y_2...Y_m$ with $n \geq m > 0$, we have
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
\label{lemma:F_decomposition}
\end{restatable}
\begin{theorem} An alternative optimization to characterize ML for the single-trace deletion channel.
\begin{equation}
\max_{x\in \{0,1\}^n} {x \choose Y} = \max_{p\in [0,1]^n} \mathbf F(p,Y).
\label{eq:ml_opt_equiv}
\end{equation}
Furthermore, given any non-integral $p^* \in [0,1]^n$ that maximizes $\mathbf F(p,Y)$, one could construct a corresponding integral solution $x^* \in \{0,1\}^n$ that maximizes $\mathbf F(p,Y)$ and consequently is also a solution to $\max_{x\in \{0,1\}^n} {x \choose Y}$.
\end{theorem}
\begin{proof}
As noted earlier, we have ${x \choose Y} = \mathbf F(x,Y)$. Therefore, we are interested in proving the following:
\begin{align*}
\max_{x\in \{0,1\}^n} \mathbf F(x,Y) \equiv \max_{p\in [0,1]^n} \mathbf F(p,Y).\numberthis
\label{eq:ml_opt_equiv_proof1}
\end{align*}
To show this and also the second statement, we prove that given $p=(p_1,p_2,...,p_i,...,p_n)$, at least one of the following holds true:
\begin{itemize}
\item$\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},0,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 0.
\item $\mathbf F(p^{(i\rightarrow 1)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},1,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 1.
\end{itemize}
Thus if $p^*$ is an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ with $p_i\in (0,1)$, then at least one of $p^{(i\rightarrow 0)}$ or $p^{(i\rightarrow 1)}$ is also an optimal solution. Sequentially applying this argument for each coordinate of $p$ shows that there exists a point in $\{0,1\}^n$ which is also an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ and consequently to $\max_{x\in \{0,1\}^n} \mathbf F(x,Y)$.
Now to prove our claim, we use Lemma~\ref{lemma:F_decomposition} to factor out $p_i$ terms in $\mathbf F(p,Y)$:
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
There are 3 cases
\begin{enumerate}
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) = \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) = \mathbf F(p,Y) = \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) > \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \leq \mathbf F(p,Y) \leq \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) < \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y) \geq \mathbf F(p^{(i\rightarrow 1)},Y).$
\end{enumerate}
Thus in each of the 3 cases, we see that at least one of $\mathbf F(p^{(i\rightarrow 0)},Y)$ or $\mathbf F(p^{(i\rightarrow 1)},Y)$ is at least as large as $\mathbf F(p,Y)$ thus proving the theorem. Note that the proof gives a natural way to find an optimal lattice point $p_{lattice}^*$ given a non-lattice point $p^*$ by iterating through each coordinate of $p^*$ and switching them to $0$ or $1$ by comparing $\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})$ with $\sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$
\end{proof}
\section{Numerical results}
\label{sec:Numerics}
In this section we show numerics supporting our theoretical results. The metric we use to measure the performance of different reconstruction algorithms is \textit{Hamming error rate}, which is defined as the average Hamming distance (the Hamming distance between the actual and estimated input sequence divided by the input sequence length).\\
\noindent \textbf{Single-trace deletion channel.}
\begin{itemize}[leftmargin=3mm]
\item Fig.~\ref{fig:numerics_single1} compares the error rate for reconstruction via ML heuristic (Alg.~\ref{alg:cood_switch}) when the initialization is a random lattice point and when the initialization is a random interior point, and also the symbolwise MAP via Alg.~\ref{alg:apprx_smap_dp} for a small blocklength $n=10$. For Alg.~\ref{alg:cood_switch}, we start with 5 random initializations and choose the one that gives the maximum objective value. As can be seen, initialization with an interior point far outperforms initialization with lattice point, thus validating the usefulness of our relaxation.
\end{itemize}
\begin{figure}[t!]
\centering
\includegraphics[scale=0.48]{ml_int_vs_ver_error.eps}
\caption{Hamming error rate for reconstruction via coordinate switch ML heuristic (Alg.~\ref{alg:cood_switch}) when the initialization is a random lattice point and when the initialization is a random interior point, and reconstruction via symbolwise MAP (Alg.~\ref{alg:apprx_smap_dp}) for a blocklength $n=10$.}
\label{fig:numerics_single1}
\end{figure}
\vspace{10pt} \noindent \textbf{$t$-trace deletion channel model.}
Fig.~\ref{fig:numerics1} and Fig.~\ref{fig:numerics2} show results comparing reconstruction via symbolwise MAP for $t$-trace deletion channel model (Alg.~\ref{alg:exact_smap}), reconstruction via sequentially updating posterior probabilities over the traces (Alg.~\ref{alg:apprx_smap}), reconstruction via independently combining posteriors from each of the traces (described as Alg.~\ref{alg:ind_comb}) and a trace reconstruction algorithm from \cite{Batu2004} called the \textit{Bitwise Majority Alignment (BMA)} reproduced here as Alg.~\ref{alg:bitwise_majority}. We next make two observations, related to two of these algorithms.
\noindent \textbf{Independent posterior combination:} As pointed in the introduction, computing the posterior probabilities for each deletion channel and combining them as if they came from independent observations does not provide a natural solution for computing the posterior probabilities for the $t$-trace deletion channel. One could, however, check how such a naive combination of posteriors compares with our reconstruction algorithms for $t$-traces.
\noindent \textbf{Bitwise Majority Alignment:} BMA reconstructs the input sequence by first ``aligning'' the traces using a pointer for each trace, and then taking the majority of the pointed symbols. We note that the state-of-the-art trace reconstruction algorithms in the literature are applicable in the asymptotic regime where the blocklength $n$ and the number of traces $t$ approach $\infty$; it is not clear how to adapt such algorithms for a finite blocklength and a small number of traces.
\begin{algorithm}[t!]
\caption{Trace reconstruction via independent posterior combination}\label{alg:ind_comb}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{(1)},...,Y^{(t)}$}, input length $n$ \\ Outputs: Estimate of the input $\hat X$
\State Initialize priors $p^{old} \gets (0.5,0.5,...,0.5)$
\For {$l=1:t$}
\State Use Alg.~\ref{alg:apprx_smap_dp} with $p^{old}$ and $Y^{(l)}$ to compute posteriors $p^{(l),new}$
\EndFor
\For {$i=1:n$}
\If {$\prod_{l=1}^t p^{(l),new}_i \geq \prod_{l=1}^t (1-p^{(l),new}_i)$}
$\ \hat X_i \gets 1$
\Else
$\ \hat X_i \gets 0$
\EndIf
\EndFor
\State \textbf{return} $\hat X_1 \hat X_2 ... \hat X_n$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[!t]
\caption{Bitwise Majority Alignment}
\label{alg:bitwise_majority}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{(1)},Y^{(2)},...,Y^{(t)}$, input length $n$}\\ Output: estimate of input $\hat X = \hat X_1 \hat X_2...\hat X_n$.
\State Initialize $c_j=1$ for $j\in [t]$.
\State Initialize $\hat X_i = 1$ for $i \in [n]$.
\For{$i \in\ [0:n]$}
\State Let $b$ be the majority over all $t$ of $Y^{(j)}_{c_j}$
\State $\hat X_i \gets b$
\State Increment $c_j$ for each $j$ such that $Y^{(j)}_{c_j} = b$
\EndFor
\State \textbf{return} $\hat X_1 \hat X_2 ... \hat X_n$
\end{algorithmic}
\end{algorithm}
Our numerical evaluation results are as follows.
\begin{itemize}[leftmargin=3mm]
\item Fig.~\ref{fig:numerics1} compares symbolwise MAP for the $t$-trace deletion channel model (Alg.~\ref{alg:exact_smap}), trace reconstruction heuristic via sequentially updating posterior probabilities (Alg.~\ref{alg:apprx_smap}) with BMA (Alg.~\ref{alg:bitwise_majority}) for a small blocklength $n=10$. As seen, both of our algorithms significantly outperform BMA. Moreover, the heuristic Alg.~\ref{alg:apprx_smap} is not much worse than Alg.~\ref{alg:exact_smap}, at least for this blocklength. Note that Alg.~\ref{alg:apprx_smap} is polynomial in $n$ and $t$ ($O(tn^2)$) while Alg.~\ref{alg:exact_smap} is exponential in $t$ ($O(2^tn^{t+2})$).
\item Fig.~\ref{fig:numerics2} compares the reconstruction heuristic via sequentially updating posterior probabilities (Alg.~\ref{alg:apprx_smap}) with reconstruction via independently combining posteriors from each of the traces (Alg.~\ref{alg:ind_comb}) and BMA (Alg.~\ref{alg:bitwise_majority}), over a larger blocklength of $n=100$. Clearly, Alg.~\ref{alg:apprx_smap} outperforms both the others, indicating that it is an interesting heuristic for trace reconstruction.
\begin{figure}[t!]
\centering
\includegraphics[width=16cm]{errors_BL10.eps}
\caption{Hamming error rate comparison for symbolwise MAP for the $t$-trace deletion channel model (Alg.~\ref{alg:exact_smap}), trace reconstruction heuristic via sequentially updating posterior probabilities (Alg.~\ref{alg:apprx_smap}) with BMA (Alg.~\ref{alg:bitwise_majority}) for a blocklength $n=10$.}
\label{fig:numerics1}
\end{figure}
\item We note that with the number of traces, though the performance of all algorithms improves, the returns tend to diminish. This intuitively makes sense because not all the information in each additional trace is ``new''; some of it is already conveyed by existing traces.
\item Even when the deletion probability is $\delta = 0.5$, the error rate is still less than $0.5$. This is not unexpected since the observed deleted sequence still carries some information about each of the input bits.
\end{itemize}
\begin{figure}[t!]
\centering
\includegraphics[width=16cm]{errors_BL100.eps}
\caption{Hamming error rate comparison for reconstruction heuristic via sequentially updating posterior probabilities (Alg.~\ref{alg:apprx_smap}) with reconstruction via independently combining posteriors from each of the traces (Alg.~\ref{alg:ind_comb}) and BMA (Alg.~\ref{alg:bitwise_majority}) for a larger blocklength of $n=100$.}
\label{fig:numerics2}
\end{figure}
\section{Introduction}
\label{sec:intro}
Sequence reconstruction over deletion channels, both with and without a codebook, has received considerable attention in the information theory as well as in the theoretical computer science literature. From an information theory perspective, reconstruction over the deletion channel, or more specifically a maximum-likelihood (ML) argument for the deletion channel, would give further insight on the capacity of the deletion channel, a long-standing open problem (see \cite{mitzenmacher2009survey}). To quote \cite{mitzenmacher2009survey} -- ``at the very least, progress in this direction would likely surpass previous results on the capacity of the deletion channels''. Yet, there are no results on reconstruction over a deletion channel with statistical guarantees -- in this work, we take a step in this direction.
On the other hand, the problem of \textit{trace reconstruction}, as introduced in \cite{Batu2004}, has received renewed interest in the past few years (see \cite{Holenstein2008}, \cite{Peres2017}, \cite{De2017}, \cite{holden18}, \cite{Nazarov:2017}, \cite{holden2018lower}, \cite{chase2019new}). The problem of trace reconstruction can be stated simply as follows: consider a sequence $X$ which is simultaneously passed through $t$ independent deletion channels to yield $t$ deleted observations (also called \textit{traces}) of $X$ (see Fig.~\ref{fig:tdeletion}). How many such traces are needed to reconstruct $X$ perfectly? A variety of upper and lower bounds for this problem have been proposed, both for worst case and average case reconstruction. Our problem definition, stated in the following paragraph, is closely related to the average case trace reconstruction.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{tdeletion.pdf}
\caption{$t$-deletion channel model: sequence $X$ passed through $t$ independent deletion channels to yield $t$ \textit{traces}. We aim to estimate $X$ from the $Y^{(i)}$s.}
\label{fig:tdeletion}
\end{center}
\end{figure}
\noindent \textbf{Problem definition.} Given an input sequence of length $n$ (known apriori), the independently and identically distributed (i.i.d.) deletion channel deletes each input symbol indepedently with probability $\delta$, producing at its output a subsequence of the input sequence. Consider a sequence $X$ passed through $t$ ($t$ is fixed) such deletion
channels as shown in
Fig.~\ref{fig:tdeletion}. We call this the $t$-deletion channel model. We ask two questions:
\begin{enumerate}
\item For $t=1$ (see Fig.~\ref{fig:1deletion}, also called the single deletion channel model), and when $X_i \sim\ \text{ind. Ber}(p_i)$, compute the posterior distributions of $X_i$ given the trace $Y$.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.25]{1deletion.pdf}
\caption{The single deletion channel model. We assume $X_i \sim\ \text{ind. Ber}(p_i)$.}
\label{fig:1deletion}
\end{center}
\end{figure}
\item In the $t$ deletion channel model, for a fixed $t$ assume that $X_i \sim\ \text{i.i.d. Ber}(0.5)$ and compute the posterior distributions of $X_i$ given all traces $Y^{(1)}, Y^{(2)},...,Y^{(t)}$.
\end{enumerate}
Note that solving 1) above doesn't lead to a natural solution for 2). This is because for a memoryless channel, we have $Y^{(j)} - X_i - Y^{(k)}$ and hence $\Pr(X_i=\alpha|Y^{(j)}, Y^{(k)}) \propto \Pr(X_i=\alpha|Y^{(j)}) \Pr(X_i=\alpha|Y^{(k)})$; so one could independently combine the posterior probabilities from each noisy observation. This is not the case for deletion channels since the markov chain $Y^{(j)} - X_i - Y^{(k)}$ no longer holds. More intuitively, one needs to first ``align'' the traces for computing the likelihoods.
We point out that the problem considered in 2) asks a question complementary to the trace reconstruction: given a fixed (possibly a few) number of traces, what is our ``best'' guess of $X$? We provide algorithms which do this. Unlike trace reconstruction, we are not concerned with perfect reconstruction (since perfect reconstruction may not be possible with just a few traces), although it should also be obvious that performance guarantees for our algorithms (not a part of this work) would naturally lead to upper bounds for trace reconstruction.
Deletion channel by itself is known to be notoriously difficult to analyse. As stated earlier, the capacity of a single deletion channel is still unknown (\cite{diggavi2007capacity,diggavi2006information,diggavi2001transmission}); as are optimal coding schemes. Recent works have looked at the design of codes for deletion channels (\cite{ratzer2005marker,ratzer2000codes,thomas2017polar}); these works consider use of a codebook (we do not). As a result, statistical estimation over deletion channels is also a difficult problem due its highly combinatorial nature. To the best of our knowledge, as yet there are no efficient estimation algorithms over deletion channels with statistical guarantees; not even for ML over a single deletion channel.
\noindent \textbf{Biological motivation.} Trace reconstruction in itself was motivated, in part, by problems DNA sequence reconstruction. One such problem was to infer the DNA sequence of a common ancestor from the samples of its descendants. We argue that our problem definition fits more naturally in such a scenario since perfect reconstruction may not be feasible or even possible. Our motivation for considering this problem also comes from a recent DNA sequencing technology called \textit{nanopore sequencing}. The $t$-deletion channel model is a simplistic model to approximately capture the process of a DNA sequence passed through a nanopore sequencer\footnote{As seen in
\cite{Mao2017},\cite{MDK17} there are more complicated effects of
the nanopore reader not captured in this simple
representation.}.
Very recently, a variant of the trace reconstruction problem called \textit{coded trace reconstruction} has been proposed, motivated by portable DNA-based data storage systems using DNA nanopores (see \cite{abroshan2019coding}, \cite{cheraghchi2019coded}, \cite{brakensiek2019coded}) and we believe that the ideas in this work may prove useful in such a setting.
There are other works on sequence assembly (see for example, \cite{Li09fastand}, \cite{Shomorony2016}),
where multiple short reads (from different segments of a sequence) are used to reconstruct the bigger
sequence. This work differs from sequence assembly since we are interested to infer the entire length sequence and not just small segments of it (which are then ``stitched'' together in sequence assembly).
\noindent \textbf{Tools and techniques.} In terms of theoretical tools, the series of books by Lothaire (\cite{lothaire1997combinatorics,lothaire2002algebraic,lothaire2005applied}) extensively use algebraic tools for problems in the combinatorics of sequences (or \textit{words}), and our work is inspired by such techniques. We borrow many of their notations and results for our work.
\noindent \textbf{Contributions.} {\color{red}} Our main contribution is to provide tools to visualize and analyze error events (described in the next section) for the multiple deletion channel model in Fig.~\ref{fig:tdeletion}. We also provide algorithms to solve the problems stated in 1) and 2) earlier in the section.
\begin{itemize}[wide=5pt]
\item In section~\ref{sec:1deletion}, for the single deletion channel model, we provide an $O(n^2)$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i=1|Y)\ \forall\ i$ when $X_i \sim \text{ind. Ber}(p_i)$.
\item In Section~\ref{sec:exactsmap}, for the $t$-deletion channel model, we give an $O(2^t n^{t+2})$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i = 1|Y_1,...,Y_t)$ when $X_i \sim \text{ind. Ber}(0.5)$.
\end{itemize}
\section{Notation and Tools}
\label{sec:notations}
\noindent \textbf{Basic notations:} In this work, we borrow a majority of the notation and tools from \cite{lothaire1997combinatorics} which deals with non-commutative algebra. We restate the definitions here for convenience. Calligraphic letters refer to sets, capitalized letters
correspond to random variables and bold letters are used for functions. Let $\mathcal{A}$ be the set of all
symbols. Throughout this work, we will focus on the case where $\mathcal{A}=\{0,1\}$,
though our methods extend to arbitrarily large sets of finite size. Define
$\mathcal{A}^n$ to be the set of all $n$-length sequences and $\mathcal{A}^*$ to be the set of all finite length sequences with
symbols in $\mathcal{A}$. For a sequence $f$, $|f|$ denotes the length
of $f$. For integers $i,j$, we define $[i:j] \triangleq \{i,i+1,...,j\}$ if $j\geq i$ and $[i:j] \triangleq \emptyset$ otherwise. Also define $[i] \triangleq [1:i]$.\\
\noindent \textbf{Binomial coefficient:} Given sequences $f$ and $g$ in $\mathcal{A}^*$, the number of
subsequence patterns of $f$ that are equal to $g$ is called the
\textit{binomial coefficient} of $g$ in $f$ and is denoted by $f
\choose g$. For example, ${'apple' \choose 'ape'} = 2$ since $'ape'$ can be obtained from two (overlapping) subsequences of $'apple'$.
When the alphabet $\mathcal{A}$ is of cardinality 1, ${f
\choose g} = {|f| \choose |g|}$, the classical binomial coefficient
with their respective lengths as the parameters. This definition hence
could be thought of as a generalization of the classical binomial
coefficients. We will denote by $e$ the sequence of length 0, and {define ${f
\choose e} \triangleq 1\ \forall\ f\ \in\ \mathcal{A}^*$.} We also define the classical binomial coefficient ${a \choose b}\triangleq 0,$ whenever $b>a$ or $b<0$ for ease of use.
The binomial coefficient is an integral aspect of this work and for analyzing error events in deletion channels because the input-output relation for a deletion channel (with deletion probability $\delta$, input $X$ and output $Y$) can be expressed as
\begin{equation}
\label{eq:in-out_relation}
\Pr(Y=y|X=x) = {x \choose y} \delta^{|x|-|y|} (1-\delta)^{|y|}.
\end{equation}
The proof is straight-forward -- the number of distinct error events that give rise to $y$ from $x$ is exactly the number of subsequences of $x$ which are equal to $y$. Each of these error events has a probability $\delta^{|x|-|y|} (1-\delta)^{|y|}$, wherein the exponent of $\delta$ corresponds to the deleted symbols and the exponent of $1-\delta$ to the undeleted symbols.
Given the definition of the binomial coefficient, the maximum-likelihood (ML) estimate over a deletion channel with observed output $Y$ can be cast in the following form:
\begin{align*}
\argmax_{x \in \mathcal C} {x \choose Y},\numberthis
\label{eq:ML_deletion}
\end{align*}
where $\mathcal C$ is the chosen codebook. In the case of multiple deletion channels with observed traces $Y^{(1)},...,Y^{(t)}$, the ML formulation is similar:
\begin{align*}
\argmax_{x \in \mathcal C} \prod_{j=1}^{t} {x \choose Y^{(j)}}.\numberthis
\label{eq:ML_multiple_deletion}
\end{align*}
As yet, there is no known efficient way to come up with a solution for the above two formulations, even for \eqref{eq:ML_deletion} with $\mathcal C = \{0,1\}^n$ (see \cite{mitzenmacher2009survey}). In this work, we attempt to take a step in this direction by showing that a continuous relaxation of \eqref{eq:ML_deletion} is equivalent to \eqref{eq:ML_deletion}. However, an efficient algorithm to solve the optimization \eqref{eq:ML_deletion} remains open. In the context of trace reconstruction, the ultimate pursuit would be an algorithm for \eqref{eq:ML_multiple_deletion} with $\mathcal C = \{0,1\}^n$ and error analysis thereof.
\noindent We now describe a function which can be thought of as a real-valued extension of the binomial coefficient. This function is used in sections~\ref{sec:1deletion_ML} and \ref{sec:1deletion}. Consider the function $\mathbf F(\cdot)$ defined as:
\begin{definition}
\label{def:f}
\begin{align*}
&\mathbf F: \mathbb{R}^k \times \{0,1\}^l \rightarrow \mathbb{R},\\
\mathbf F(q, v)\triangleq &\begin{cases}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [k],\\|\mathcal S|=l}} \quad \prod\limits_{i=1}^l q_{\mathcal S_i}^{v_i} (1-q_{\mathcal S_i})^{1-v_i} & 1 \leq l\leq k \\
1 & 0=l\leq k \\
0 & \text{else}.
\end{cases}
\end{align*}
\noindent An alternate definition is as follows: consider a random vector $Z\in\{0,1\}^k$ such that $Z_i\sim$ ind. Ber$(q_i)$, let $q$ be the vector of probabilities of length $k$, $\mathcal S\subseteq [k]$ a subset of the indices of size $l$, and ${v}$ a binary vector of length $l$. Then, $$\mathbf F(q,v)=\sum_{\substack{\mathcal S||\mathcal S|=l}} \Pr(Z_{\mathcal S}= v).$$ Note that if $q$ is a vector of $0$'s and $1$'s, then $\mathbf F( q, v)={q \choose v}$, the binomial coefficient of $v$ in $q$. Thus, $\mathbf F(\cdot)$ could be interpreted as an extension of the binomial coefficient where one of the parameters can take values in $[0,1]^n$ instead of $\{0,1\}^n$.
\end{definition}
Though at first sight, $\mathbf F(q,v)$ sums over an exponential number of subsets, a dynamic programming approach can be used to compute it in $O(|v|^2)$ time complexity. The dynamic program is described in section~\ref{sec:1deletion}.
\iffalse
\noindent \textbf{Edit distance:} The edit distance $d_e(f,g)$ measures similarity between two sequences of
possibly different lengths \cite{Navarro2001}. $d_e(f,g)$ is the minimum number of operations
needed to transform $f$ to $g$, where the permitted operations are
insertion, deletion or substitution of a symbol. In this work, we
quantify the performance of algorithms in Section \ref{sec:Numerics}
using the edit distance metric.
\fi
The following definitions and ideas are relevant only for \ref{sec:exactsmap} and can be omitted for the sections on single deletion channel. Before getting to the mathematical definitions, we first state a result of ours that aids in thinking about error events in multiple deletion channels.
\subsection{An alternate interpretation of multiple deletion channel model}
The events occurring in the $t$-deletion channel model can be categorized into two groups:
\begin{enumerate}
\item an input symbol is deleted in \textit{all} the $t$-traces,
\item an input symbol is reflected in at least one of the traces.
\end{enumerate}
The error events of the first kind are in some sense ``not correctable" or even ``detectable" since it is impossible to tell what and where the deleted symbol could have been (although the probabilities need not be uniform). The events of the second kind, however, can still be detected although they could likely be confused with a large set of similar events. This thought process gives rise to a natural decomposition of the $t$-deletion channel model into a cascade of two channels: the first one being a deletion channel which captures error events of the first kind and the second one is what we call the \textit{remnant channel} which captures events of the second kind (see Fig.~\ref{fig:channel_equiv}). More precisely, the remnant channel is defined as follows:
\begin{definition}
\textit{Remnant channel:} an input symbol to the remnant channel is reflected in $k>0$ given outputs and deleted in the rest with a probability $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
\end{definition}
It is easy to note that probability of the union of all possible events here is $\sum_{k=1}^t {t \choose k} \frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}=1$, validating our definition.
\begin{restatable}{theorem}{channelequiv}
\label{thm:channel_equiv}
The $t$-deletion channel model and the cascade of the deletion channel with remnant channel shown in Fig.~\ref{fig:channel_equiv} are probabilistically equivalent, i.e.,
$$\Pr({Y}^{(1)},{Y}^{(2)},...,{Y}^{(t)}|X = x) = \Pr(\tilde{Y}^{(1)},\tilde{Y}^{(2)},...,\tilde{Y}^{(t)}|X = x).$$
\end{restatable}
The formal proof of the theorem requires few more definitions and is relegated to the appendix.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.4]{channel_equiv.pdf}
\caption{A channel equivalence result: the $t$-deletion channel model in (a) is probabilistically equivalent to the the cascade of a deletion channel with the \textit{remnant channel} ($ \mathcal C_2$) in (b).}
\label{fig:channel_equiv}
\end{center}
\end{figure}
\noindent\textbf{Edit graph} (as defined in \cite{Gusfield1997}): We now define a graph construct which is closely related to the error events in the remnant channel. We start with a simple case and generalize subsequently. Define an \textit{edit graph} given two sequences $f$ and $g$, where every path connecting the ``origin'' to the ``destination'' on the edit graph yields a supersequence $h$ of $f,g$, where $h$ is ``covered'' by $f,g$ -- i.e., each symbol of $h$ comes from either $f$ or $g$ or
both. In other words, given that $f$ and $g$ are the outputs of the remnant channel (with two outputs), each path from the origin of the edit graph to the destination corresponds to a possible input $h$ to the remnant channel and to an error event which resulted in $f$ and $g$ with input $h$.
For $f$ and $g$ in $\mathcal
A^*$, we form a graph $\mathcal{G}(f,g)$ with $(|f|+1)(|g|+1)$
vertices each labelled with a distinct pair $(i,j), 0\leq i\leq
|f|,\ 0\leq j\leq |g|$. A directed edge
$(i_1,j_1)\rightarrow(i_2,j_2)$ exists iff at least one of the following
holds:
\begin{enumerate}
\item$i_2-i_1=1$ and $j_1=j_2$, or
\item $j_2-j_1=1$ and $i_1=i_2$, or
\item $i_2-i_1=1$, $j_2-j_1=1$ and $f_{i_2}=g_{j_2}$,
\end{enumerate}
where $f_i$ is the $i^{th}$ symbol of the sequence $f$. The origin is the vertex $(0,0)$ and the destination $(|f|,|g|)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap1.pdf}
\caption{ Edit graph for sequences $f=$ `001' and $g=$ `101'. An easy way to think about this is to write down $f$ vertically with each symbol aligned to a vertical set of edges and $g$ horizontally likewise. A diagonal edge in a small square exists if the corresponding $f$ and $g$ symbols are equal. The thick red edges form a path from the origin to the destination; this path corresponds to the sequence `0101' -- append the corresponding symbol at the left of an edge if it's vertical or diagonal, otherwise append the symbol at the top. It is also easy to verify that 0101 is a supersequence of both $f$ and $g$, and could be obtained as a covering of $f$ and $g$; the path itself gives one such covering. This covering also corresponds to an error event in the remnant channel which would result in outputs $f$ and $g$ with input $h=$ `0101' -- more precisely, the error event is one in which the first symbol of $h$ is deleted only in $g$, the second symbol is deleted only in $f$ and the third and fourth symbols not deleted in either $f$ or $g$.}
\label{fig:editgraph_smap1}
\end{figure}
Let $p=((i_1,j_1),(i_2,j_2),...,(i_m,j_m))$ be a path in $\mathcal{G}(f,g)$. We define $ s(p)$ to be the sequence corresponding to the path. Intuitively, $s(p)$ is formed by appending symbols in the following
way: append the corresponding $f$ symbol for a vertical edge, $g$ symbol for horizontal edge, and $f$ or $g$ symbol for diagonal edge (see example Fig.~\ref{fig:editgraph_smap1}). Any path from $(0,0)$ to $(|f|,|g|)$ corresponds to a supersequence of $f$ and $g$ and which is ``covered" by $f$ and $g$. More formally, define $ s(p)\triangleq x_1x_2...x_{m-1}$ where
$$x_k =
\begin{cases}
f_{i_{k+1}} \quad\text{if }j_{k}=j_{k+1},\\
g_{j_{k+1}} \quad\text{if }i_{k}=i_{k+1},\\
f_{i_{k+1}} \quad\text{else.}
\end{cases}
$$
The construct of edit graph can be extended to more than 2 sequences with the same idea. For sequences $f_1,f_2,...,f_t$, construct a $t$-dimensional grid with a number of vertices $(|f_1|+1)(|f_2|+1)...(|f_t|+1)$ labeled from $(0,0,...,0)$ to $(|f_1|,|f_2|,...,|f_t|)$. A vertex $u=(i_1,i_2,...,i_t)$ is connected to $v=(j_1,j_2,...,j_t)$ (we say $u \rightarrow v$) iff both of the following conditions are met:
\begin{itemize}
\item $j_l=i_l$ or $j_l=i_l+1$ $\forall\ l\in [t]$, i.e., $(i_1,...,i_t)$ and $(j_1,...,j_t)$ are vertices of a particular unit cube. Only these type of vertices can share an edge in the grid graph.
\item Let $\mathcal T \subseteq [t]$ be the collection of indices where $j_l=i_l+1$. Then $f_{i_l}$ is same $\forall\ l \in \mathcal T$. For instance, if $\mathcal T=\{1,3,5\}$, then $f_{i_1}=f_{i_3}=f_{i_5}$.
\end{itemize}
Define the vertex $(0,...,0)$ to be the origin of this graph and the vertex $(|f_1|,...,|f_t|)$ to be the destination. If $|f_j|=O(n)\ \forall\ j$, this graph has a number of vertices $O(n^t)$ and a maximum number of edges $O((2n)^t)$ since each vertex has at most $2^t$ outgoing edges.\\
\noindent\textbf{Infiltration product} (introduced in \cite{lothaire1997combinatorics}):
The infiltration product has been extensively used in \cite{lothaire1997combinatorics}, as a tool in non-commutative algebra. Here we give an edit-graph interpretation of this tool. We also give a formal definition later in this section. Using the edit graph we can construct the set of possible
supersequences $\mathcal{S}(f,g)$ of $f,g$ which are covered by
it. Clearly multiple paths could yield the same supersequence and we
can count the number of distinct ways $\mathbf N(h;f,g)$ one can construct
the same supersequence $h$ from $f,g$.
We can informally define the
\emph{infiltration product $f\uparrow g$} of $f$ and $g$, as a polynomial with monomials the supersequences $h$ in
$\mathcal{S}(f,g)$ and coefficients $\langle f\uparrow g,h\rangle$ equal to $\mathbf N(h;f,g)$. In Fig.~\ref{fig:editgraph_smap1}, it is easy to verify that there is exactly one path corresponding to `101001' and hence $\langle 001\uparrow 101,101001 \rangle=1$ and similarly $\langle 001\uparrow 101,01001 \rangle=2$. One could find these coefficients for all relevant sequences and form the polynomial as described. More examples: Let $\mathcal{A}=\{a,b\}$, then
\begin{itemize}[wide=2pt]
\item $ab\uparrow ab=ab+2aab+2abb+4aabb+2abab$,
\item $ab\uparrow ba=aba+bab+abab+2abba+2baab+baba.$
\end{itemize}
The infiltration operation is commutative and associative, and infiltration of two sequences $f\uparrow g$ is a polynomial with variables of length (or \textit{degree}) at most $|f|+|g|$; see \cite{lothaire1997combinatorics}.
The definition of infiltration extends to two polynomials via distributivity (precisely defined in later), and consequently to multiple sequences as well. For multiple sequences, infiltration has the same edit graph interpretation: $\langle f_1\uparrow f_2 \uparrow...\uparrow f_t, w \rangle$ is the number of distinct ways of constructing $w$ as a supersequence of $f_1, f_2, ... ,f_t$ so that the construction covers $w$, i.e., construct the $t$-dimensional edit graph of $f_1, f_2, ... ,f_t$ and count the number of paths corresponding to $w$.
\subsection{Formal definition of infiltration product}
We now give a more formal definition of the infiltration product (see \cite{lothaire1997combinatorics} for the equivalence of the two definitions and a more rigorous treatment). A \textit{formal series} with indeterminates (or variables) in a set $\mathcal A$ and coefficients in a commutative ring $\mathcal R$, is a mapping of $\mathcal A^*$ onto $\mathcal R$. Recall that a commutative ring is a set which forms an abelian group under an \textit{addition} operation, is a monoid under a \textit{multiplication} operation which commutes, and the multiplication operation distributes over addition. Here we consider $\mathbb Z$, the set of integers as the commutative ring $\mathcal{R}$. A formal series is called a \textit{polynomial} if only a finite number of sequences are mapped to non-zero values, the rest of the sequences map to zero. Consider two polynomials $\sigma,\tau: \mathcal{A}^*\rightarrow \mathbb Z$. The value taken by a sequence $w\in \mathcal A^*$ on $\sigma$ (or the coefficient of $w$ in $\sigma$) is denoted by $\langle \sigma,w\rangle \in \mathbb R$. We also define binary addition ($\oplus$) and multiplication operations ($\times$) on the set of polynomials as follows:
\begin{align}
\langle \sigma\oplus \tau,w\rangle \triangleq \langle \sigma,w\rangle + \langle \tau,w \rangle \quad \forall w\in \mathcal A^*,\label{eq:polynomial_add}\\
\langle \sigma\times \tau,w\rangle \triangleq \sum_{\substack{f,g\in \mathcal A^*:\\ f.g=w}}\langle \sigma,f\rangle \langle \tau,g \rangle \quad \forall w\in \mathcal A^*.\label{eq:polynomial_prod}
\end{align}
We will use the usual symbols $+$ and $.$ in place of $\oplus$ and $\times$ in this work for convenience. The meaning of the operation would be clear depending on the operands. With these operations the set of polynomials form a non-commutative ring, and is denoted by $\mathbb Z\langle\mathcal A \rangle$, also called the free $\mathbb Z$-algebra on $\mathcal A$ in ring theory. Note that the addition and multiplication operations defined in \eqref{eq:polynomial_add} and \eqref{eq:polynomial_prod} are similar to the operations defined on commutative polynomials, except that the multiplication operation under the summation in \eqref{eq:polynomial_prod} ($f.g=w$) is actually concatenation and is non-commutative. The multiplication inside the summation in \eqref{eq:polynomial_prod} is multiplication in the real field and hence commutative. It is also easy to prove that the multiplication defined in \eqref{eq:polynomial_prod} distributes over addition defined in \eqref{eq:polynomial_add}. Thus, a polynomial in $\mathbb Z\langle\mathcal A \rangle$ can be represented as a sum of monomials in $\mathcal A^*$ each with an associated coefficient in $\mathbb Z$, i.e., $\sigma=\sum\limits_{w\in \mathcal A^*} \langle\sigma,w \rangle w$. Define the \textit{degree} of a polynomial to be equal to the length of a longest sequence with a non-zero coefficient in the polynomial and the \textit{number of terms} of a polynomial as the number of sequences with non-zero coefficients in the polynomial. Note that a degree $d$ polynomial could have a number of terms upto $2^{d+1}-1$.
With this, the \textit{infiltration product} (in general, for two polynomials) is defined as follows:
\begin{align}
\forall f\in \mathcal{A}^*,& \quad f\uparrow e = e\uparrow f=f.\nonumber \\
\forall f,g\in \mathcal{A}^*&,\quad \forall a,b\in \mathcal{A}, \nonumber\\
fa\uparrow gb=(f\uparrow gb)a&+(fa\uparrow g)b+\mathbbm{1}_{a=b}(f\uparrow g)a.\nonumber \\
\forall \sigma,\tau \in \mathbb{Z}\langle\mathcal{A} \rangle, \quad &\sigma\uparrow \tau=\sum_{f,g\in \mathcal{A}^*} \langle \sigma,f \rangle \langle \tau,g \rangle (f\uparrow g). \label{def:infiltforseq}
\end{align}
\textbf{Summary of definitions and ideas introduced in this section:}
\begin{itemize}
\item The binomial coefficient captures the likelihood of observations for deletion channels.
\item An extension of the binomial coefficient function where one of the parameters can take real values has been introduced; this function is pivotal for our results on the single-trace deletion channel.
\item For multiple deletion channels, the error events can be categorized into two groups -- one where an input symbol is deleted in all the traces, and second, the complement of this event. This categorization results in a natural decomposition of multiple deletion channel model into a cascade model involving the remnant channel.
\item The remnant channel disregards the error events where a symbol is deleted in all the traces.
\item The edit graph provides a way to visualize all the possible error events and input sequences to a remnant channel given its outputs.
\item The infiltration product serves the same purpose as the edit graph, but has an algebraic flavor and provides rigor for proofs and analyses. The edit graph, on the other hand, is more helpful in designing reconstruction algorithms over deletion channels.
\end{itemize}
\section{Maximum likelihood for the single-trace deletion channel}
\label{sec:1deletion_ML}
Consider the ML formulation for the single-trace deletion channel given non-empty trace $Y$ and all possible $n$-length inputs allowed, restated here for convenience:
\begin{equation}
\argmax_{x\in \{0,1\}^n} {x \choose Y}.
\label{eq:1deletion_ML}
\end{equation}
To the best of our knowledge, the only known method to solve \eqref{eq:1deletion_ML} involves iterating over all possible choices of $x$ and computing the objective value for each of the choices. We here show that it is sufficient to solve a continuous relaxation of above problem to obtain a solution to \eqref{eq:1deletion_ML}. Note that there could be multiple solutions to \eqref{eq:1deletion_ML}. Before going in the main result, we first state a useful lemma which factors out a given coordinate $p_i$ out of $\mathbf F(p,Y)$. The proof of the lemma is relegated to the appendix.
\begin{restatable}{lemma}{deletionMLrellemma}
For $p=[p_1,p_2,..,p_i,...,p_n]$ and $Y=Y_1Y_2...Y_m$ with $n \geq m > 0$, we have
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
\label{lemma:F_decomposition}
\end{restatable}
\begin{theorem} An alternative optimization to characterize ML for the single-trace deletion channel.
\begin{equation}
\max_{x\in \{0,1\}^n} {x \choose Y} = \max_{p\in [0,1]^n} \mathbf F(p,Y).
\label{eq:ml_opt_equiv}
\end{equation}
Furthermore, given any non-integral $p^* \in [0,1]^n$ that maximizes $\mathbf F(p,Y)$, one could construct a corresponding integral solution $x^* \in \{0,1\}^n$ that maximizes $\mathbf F(p,Y)$ and consequently is also a solution to $\max_{x\in \{0,1\}^n} {x \choose Y}$.
\end{theorem}
\begin{proof}
As noted earlier, we have ${x \choose Y} = \mathbf F(x,Y)$. Therefore, we are interested in proving the following:
\begin{align*}
\max_{x\in \{0,1\}^n} \mathbf F(x,Y) \equiv \max_{p\in [0,1]^n} \mathbf F(p,Y).\numberthis
\label{eq:ml_opt_equiv_proof1}
\end{align*}
To show this and also the second statement, we prove that given $p=(p_1,p_2,...,p_i,...,p_n)$, at least one of the following holds true:
\begin{itemize}
\item$\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},0,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 0.
\item $\mathbf F(p^{(i\rightarrow 1)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},1,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 1.
\end{itemize}
Thus if $p^*$ is an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ with $p_i\in (0,1)$, then at least one of $p^{(i\rightarrow 0)}$ or $p^{(i\rightarrow 1)}$ is also an optimal solution. Sequentially applying this argument for each coordinate of $p$ shows that there exists a point in $\{0,1\}^n$ which is also an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ and consequently to $\max_{x\in \{0,1\}^n} \mathbf F(x,Y)$.
Now to prove our claim, we use Lemma~\ref{lemma:F_decomposition} to factor out $p_i$ terms in $\mathbf F(p,Y)$:
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
There are 3 cases
\begin{enumerate}
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) = \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) = \mathbf F(p,Y) = \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) > \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \leq \mathbf F(p,Y) \leq \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) < \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y) \geq \mathbf F(p^{(i\rightarrow 1)},Y).$
\end{enumerate}
Thus in each of the 3 cases, we see that at least one of $\mathbf F(p^{(i\rightarrow 0)},Y)$ or $\mathbf F(p^{(i\rightarrow 1)},Y)$ is at least as large as $\mathbf F(p,Y)$ thus proving the theorem. Note that the proof gives a natural way to find an optimal lattice point $p_{lattice}^*$ given a non-lattice point $p^*$ by iterating through each coordinate of $p^*$ and switching them to $0$ or $1$ by comparing $\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})$ with $\sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$
\end{proof}
\section{Symbolwise posterior probabilities for the single-trace deletion channel}
\label{sec:1deletion}
We here develop an algorithm to compute the symbolwise posterior probabilities for the single-trace deletion channel when the input symbols are independently generated with arbitrary priors. Consider the single deletion channel model in Fig.~\ref{fig:1deletion}, where $X=X_1...X_n$, each input symbol is generated $X_i \sim \text{ind. Ber}\ (p_i)$, and we observe at the output the trace $Y = Y_1Y_2...Y_m$ with $m\leq n$. Define the vector of priors as $p\triangleq(p_1,p_2,...,p_n)$. We first give an $O(n^2)$ algorithm to calculate the posterior probabilities $\Pr(X_i=1|Y)$, which in turn provides the symbolwise MAP estimate for the considered model. We then show how this algorithm can be used for trace reconstruction. We take three steps to present the algorithm.
\noindent \textbf{An expression for $\Pr(X_i=1|Y)$.}
Let $\Pr(X_i=1)=p_i$. As a first step, we have
\vspace{6pt}
\begin{align*}
\Pr(X_i=1|{Y}) &= \frac{\Pr(X_i=1,Y)}{\Pr(Y)}
= \frac{ \sum\limits_{\substack{ x| x_i=1}} \Pr({X=x}) \Pr(Y|X=x)}{ \sum_{\substack{x}} \Pr({X=x}) \Pr(Y|X=x)} \\
&\overset{(a)}{=} \frac{ \sum\limits_{\substack{ x| x_i=1}} \Pr({X=x}) { x \choose Y}}{ \sum_{\substack{x}} \Pr({X=x}) { x \choose Y}}, \numberthis
\label{eq:approx_smap_1}
\end{align*}
where $(a)$ is because for a deletion channel $\Pr(Y|X=x)={x \choose Y} \delta^{|x|-|Y|}(1-\delta)^{|Y|}$.
To proceed, we need to evaluate the summation in the numerator and the denominator.
Theorem~\ref{thm:approx_smap_postprob} expresses \eqref{eq:approx_smap_1} in terms of relaxed binomial coefficient terms $\mathbf F(\cdot)$. Recall one way to define $\mathbf F(p,Y) \triangleq \mathbb E_{X\sim p} {X \choose Y}$, which is the denominator term in \eqref{eq:approx_smap_1}.
\begin{theorem}
\label{thm:approx_smap_postprob}
Let $X=X_1...X_n$ where $X_i \sim \text{ind. Ber}\ (p_i)$, and let $Y$ be the observed trace when $X$ is passed through a deletion channel. Then,
\begin{align*}
\Pr(X_i&=1|{Y}) = \frac{p_i}{\mathbf F( p, Y)} \left( \mathbf F( p_{[n]\backslash \{i\}}, Y) + \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) \right) \numberthis
\label{eq:smap_postprob_thm}
\end{align*}
where $\mathbf F(\cdot)$ is as given in Definition~\ref{def:f}.
\end{theorem}
\begin{proof}
The proof of this theorem employs the same trick used in the proof of Lemma~\ref{lemma:F_decomposition}. Let $|Y|=m$. From \eqref{eq:approx_smap_1}
\begin{align*}
\Pr(X_i = 1 | Y) = \frac{ \sum\limits_{\substack{ x| x_i=1}} \Pr({X=x}) { x \choose Y}}{\mathbf F(p,Y)}.
\end{align*}
Now,
\begin{align*}
\sum_{\substack{ x| x_i=1}} & \Pr({X=x}) { x \choose Y}
=\sum_{\substack{ x|x_i=1}} \Pr({X=x}) \sum_{\substack{\mathcal S\subseteq [n]\\ |\mathcal S|=m}} \mathbbm{1}\{ x_{\mathcal S}= Y\}\\
&=\sum_{\substack{\mathcal S\subseteq [n]\\ |\mathcal S|=m}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=Y}} \Pr({X=x}).\numberthis \label{eq:smapiter1}
\end{align*}
We first separate the outer summation into two cases: (a) $\mathcal S|i \notin \mathcal S$ and (b) $\mathcal S|i\in \mathcal S$.
We can express the first case as
\begin{align*}
&\hspace{-1cm}\sum_{\substack{\mathcal S\subseteq [n] \\ |\mathcal S|=m,i\notin \mathcal S}}\sum_{\substack{ x|x_i=1\\x_{\mathcal S}=Y}} \Pr({X=x})
=\sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=Y}} \Pr({X=x})\\
&=\sum_{\substack{S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} \sum_{\substack{ x|x_i=1\\ x_{\mathcal S}= Y}} \Big(\Pr(X_i=1)\Pr( X_{\mathcal S}= Y) \Pr( X_{[n]\backslash \mathcal S\cup\{i\}}= x_{[n]\backslash \mathcal S\cup\{i\}}) \Big)\\
&=\sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} p_i \Pr( X_{\mathcal S}= Y) \left(\sum_{\substack{ x|x_i=1\\ x_{\mathcal S}= Y}} \Pr( X_{[n]\backslash \mathcal S\cup\{i\}}= x_{[n]\backslash \mathcal S\cup\{i\}})\right)\\
&=\sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} p_i \Pr( X_{\mathcal S}= Y) \left(\sum_{(x_j|j\in [n]\backslash \mathcal S\cup \{i\})} \Pr( X_{[n]\backslash \mathcal S\cup\{i\}}= x_{[n]\backslash \mathcal S\cup\{i\}})\right)\\
&=p_i \sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} \Pr( X_{\mathcal S}= Y) = p_i \mathbf F( p_{[n]\backslash \{i\}}, Y).\numberthis \label{eq:lemma3proof1}
\end{align*}
For the second term, we express the set $\mathcal S$ as a union $\mathcal S = \mathcal S' \cup \{i\} \cup \mathcal S''$ such that $\mathcal S' \subseteq [i-1]$ and $\mathcal S'' \subseteq [i+1:n]$ to get:
\begin{align*}
&\sum_{\substack{\mathcal S\subseteq [n]\\ |\mathcal S|=m,\\i\in \mathcal S}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=Y}} \Pr({X=x})= \sum_{k=1}^m\sum\limits_{\substack{\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ \mathcal S_k = i}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=Y}} \Pr({X=x})\\
&=\sum_{k=1}^m\sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=Y}}\mathbbm{1}_{\{Y_k=1\}} \Pr({X=x}) \\
&=\sum_{k:Y_k=1}\sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \sum_{\substack{ x|x_i=1\\ x_{\mathcal S'}= Y_{[1:k-1]}\\ x_{\mathcal S''}= Y_{[k+1:m]}}} \Bigg ( \Pr(X_i=1) \Pr( X_{\mathcal S'}= Y_{[1:k-1]})
\Pr( X_{\mathcal S''}= Y_{[k+1:m]}) \\&\hspace{7cm} \Pr( X_{[n]\backslash \mathcal S'\cup \mathcal S''\cup \{i\}}= x_{[n]\backslash \mathcal S'\cup \mathcal S'' \cup\{i\}})\Bigg )\\
&=p_i\sum_{k:Y_k=1}\Bigg ( \Big( \sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\Pr( X_{\mathcal S'}= Y_{[1:k-1]})\Big) \Big(\sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}}\Pr( X_{\mathcal S''}= Y_{[k+1:m]})\Big ) \\
& \hspace{5cm} \Big( \sum_{\substack{ x|x_i=1\\ x_{\mathcal S'}= Y_{[1:k-1]}\\ x_{\mathcal S''}= Y_{[k+1:m]}}} \Pr( X_{[n]\backslash \mathcal S'\cup \mathcal S''\cup \{i\}}= x_{[n]\backslash \mathcal S'\cup \mathcal S'' \cup\{i\}})\Big) \Bigg )\\
&=p_i\sum_{k|Y_k=1}\Bigg(\Big( \sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\Pr( X_{\mathcal S'}= Y_{[1:k-1]})\Big) \Big( \sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}}\Pr( X_{\mathcal S''}= Y_{[k+1:m]}) \Big)\Bigg)\\
&=p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}). \numberthis \label{eq:lemma3proof2}
\end{align*}
Plugging in \eqref{eq:lemma3proof1} and \eqref{eq:lemma3proof2} in \eqref{eq:approx_smap_1} proves the theorem.
\end{proof}
Alg.~\ref{alg:apprx_smap_dp} summarizes the computation of $\Pr(X_i=1|Y)$. The complexity of the algorithm is $O(n^2)$ which is also the complexity of computing $\mathbf F(p,Y)$; note that $m=O(n)$ since $Y$ is a deleted version of the input.
\begin{algorithm}[t!]
\caption{Symbolwise posterior probabilities with one trace}\label{alg:apprx_smap_dp}
\begin{algorithmic}[1]
\item Input: Trace {$Y$}, priors $p$\\ Outputs: Posteriors $\Pr(X_i=1|Y)\ \forall\ i$
\State Compute $\mathbf F(p_{[1:k]},Y_{[1:j]})\ \forall\ k,j$ and $\mathbf F(p_{[k:n]},Y_{[j:m]})\ \forall\ k,j$ via Alg.~\ref{alg:F_comp}
\For {$i=1:n$}
\State Use \eqref{eq:smap_postprob_thm} to compute $\Pr(X_i=1|Y)$
\EndFor
\end{algorithmic}
\end{algorithm}
\noindent\textbf{A trace reconstruction heuristic with $t$ traces.}
The posterior probability computation in Alg.~\ref{alg:apprx_smap_dp} naturally gives rise to a trace reconstruction heuristic that updates the symbolwise statistics sequentially on the traces, where we use Alg.~\ref{alg:apprx_smap_dp} with one trace at a time to continually update $\Pr(X_i=1|Y)$. The overall heuristic is described in Alg.~\ref{alg:apprx_smap}. The complexity of the algorithm is $O(tn^2)$ since Alg.~\ref{alg:apprx_smap} amounts to just $t$ uses of Alg.~\ref{alg:apprx_smap_dp}.
\begin{algorithm}[t!]
\caption{Trace reconstruction via iterative single-trace posterior probabilities}\label{alg:apprx_smap}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{(1)},...,Y^{(t)}$}, input length $n$ \\ Outputs: Estimate of the input $\hat X$
\State Initialize priors $p^{old}=p^{new} \gets (0.5,0.5,...,0.5)$
\For {$l=1:t$}
\State Use Alg.~\ref{alg:apprx_smap_dp} with $p^{old}$ and $Y^{(l)}$ to update $p^{new}$
\State $p^{old}\gets p^{new}$
\EndFor
\For {$i=1:n$}
\If {$p^{new}_i\geq 0.5$}
$\ \hat X_i \gets 1$
\Else
$\ \hat X_i \gets 0$
\EndIf
\EndFor
\State \textbf{return} $\hat X_1 \hat X_2 ... \hat X_n$
\end{algorithmic}
\end{algorithm}
\section{On Maximum Likelihood Decoding For The Single-Trace Deletion Channel}
\label{sec:1deletion_ML}
We here consider the single-trace ML decoding in (\ref{eq:ML_deletion}),
assuming that the output sequence $Y$ is non-empty.
To the best of our knowledge, the only known method to solve \eqref{eq:ML_deletion} involves solving a combinatorial optimization, essentially iterating over all possible choices of $x$ and computing the objective value for each of the choices. We here show that one could equivalently solve the continuous relaxation of \eqref{eq:ML_deletion} to obtain a solution of \eqref{eq:ML_deletion}.
Before presenting the main result, we first state a useful lemma which factors a given coordinate $p_i$ out of the relaxed binomial coefficient $\mathbf F(p,Y)$ we introduced in Definition~\ref{def:f}.
\begin{restatable}{lemma}{deletionMLrellemma}
For $p=(p_1,p_2,..,p_i,...,p_n)$ and $Y=Y_1Y_2...Y_m$ with $n \geq m > 0$, we have
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{
[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
\label{lemma:F_decomposition}
\end{restatable}
Recall that $\mathbf F(p,Y)$ sums over all $m$-length subsets $\mathcal S$ and associates $p_{\mathcal S}$ with $Y.$ Intuitively, this recursive relationship considers separately the cases where
\begin{itemize}
\item $i \notin \mathcal S$,
\item $i \in \mathcal S$ and is associated with a particular $Y_k$ where $Y_k = 1$,
\item $i \in \mathcal S$ and is associated with a particular $Y_k$ where $Y_k = 0$.
\end{itemize}
The detailed proof can be found in Appendix~\ref{app:F_lemma_proof}. It is clear from Lemma~\ref{lemma:F_decomposition} that $\mathbf F(p,Y)$ is affine when projected onto each coordinate $p_i$. Thus, the extrema of $\mathbf F(p,Y)$ must occur at the boundary of the support set of $p_i$; i.e., at either $p_i = 0$ or $p_i = 1$. Combining this with the fact that $\mathbf F(\cdot)$ is a relaxed version of the binomial coefficient, we observe that the maximization problem in \eqref{eq:ML_deletion} is equivalent to its real-valued relaxation. The following result makes this precise.
\begin{theorem} The ML decoding problem for the single-trace deletion channel
\begin{equation}
\max_{x\in \{0,1\}^n} {x \choose Y}
\label{eq:ml_opt_equiv1}
\end{equation}
is equivalent to the problem
\begin{equation}
\max_{p\in [0,1]^n} \mathbf F(p,Y).
\label{eq:ml_opt_equiv2}
\end{equation}
Furthermore, given any non-integral $p^* \in [0,1]^n$ that maximizes $\mathbf F(p,Y)$, we can construct a corresponding integral solution $x^* \in \{0,1\}^n$ that maximizes $\mathbf F(x,Y)$ and consequently also maximizes ${x \choose Y}$.
\label{thm:ML_relaxation}
\end{theorem}
\begin{proof}
As noted earlier, we have ${x \choose Y} = \mathbf F(x,Y)$. Therefore, we are interested in proving the following:
\begin{align*}
\max_{x\in \{0,1\}^n} \mathbf F(x,Y) \equiv \max_{p\in [0,1]^n} \mathbf F(p,Y),\numberthis
\label{eq:ml_opt_equiv_proof1}
\end{align*}
where $\equiv$ refers to that the two problems are equivalent (have the same optimal objective value).
We prove this by applying the following claim.\\
\textbf{Claim:} Given any feasible $p=(p_1,p_2,...,p_i,...,p_n)$, at least one of the following holds true:
\begin{itemize}
\item$\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y)$. Recall from notation that $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},0,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by $0$.
\item $\mathbf F(p^{(i\rightarrow 1)},Y) \geq \mathbf F(p,Y)$.
\end{itemize}
Thus if $p^*$ is an optimal solution to \eqref{eq:ml_opt_equiv2} with $p_i\in (0,1)$, then at least one of $p^{(i\rightarrow 0)}$ or $p^{(i\rightarrow 1)}$ is also an optimal solution. Sequentially applying this argument for each coordinate of $p$ shows that there exists a point in $\{0,1\}^n$ which is an optimal solution to \eqref{eq:ml_opt_equiv2} and consequently to \eqref{eq:ml_opt_equiv1}.
It remains to prove our claim. We use Lemma~\ref{lemma:F_decomposition} to factor out $p_i$ terms in $\mathbf F(p,Y)$:
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
Now we express $\mathbf F(p^{(i\rightarrow 0)},Y)$ and $\mathbf F(p^{(i\rightarrow 1)},Y)$ as
$$\mathbf F(p^{(i\rightarrow 0)},Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}),$$
$$\mathbf F(p^{(i\rightarrow 1)},Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$
\noindent Because $0\leq p_i\leq 1$ it directly follows that $$\min \left\{\mathbf F(p^{(i\rightarrow 0)},Y),\mathbf F(p^{(i\rightarrow 1)},Y)\right\} \leq \mathbf F(p,Y) \leq \max \left\{\mathbf F(p^{(i\rightarrow 0)},Y),\mathbf F(p^{(i\rightarrow 1)},Y)\right\},$$
thus proving our claim.
Note that the proof gives a natural way to find an optimal lattice point $p_{lattice}^*$ given an optimal interior point $p^*$ by iterating through each coordinate of $p^*$ and switching them to $0$ or $1$ by comparing $\mathbf F(p^{(i\rightarrow 0)},Y)$ and $\mathbf F(p^{(i\rightarrow 1)},Y)$.
\end{proof}
The real-valued optimization problem in \eqref{eq:ml_opt_equiv2} falls under the umbrella of signomial optimization which is, in general, NP-hard (see for example, \cite{xu2014signomial}, \cite{chand2016signomial}). With a slight change of variables, \eqref{eq:ml_opt_equiv2} could also be expressed as a maximization of a convex function in a convex set. However, it is unclear if \eqref{eq:ml_opt_equiv2} is solvable in polynomial time or inherently hard to solve.
\begin{algorithm}[t!]
\caption{Coordinate switch ML heuristic}\label{alg:cood_switch}
\begin{algorithmic}[1]
\item Input: Blocklength $n$, Trace {$Y$}, Initial point $p = (p_1,p_2,...,p_n)$ \\ Outputs: Estimated sequence $\hat X$
\State Initialize visited set $\mathcal V = \emptyset$
\While {True}
\State Compute $\mathcal F_i = |\mathbf F(p^{(i\rightarrow 1)},Y)- \mathbf F(p^{(i\rightarrow 0)},Y)|\ \forall\ i$ and let $\mathcal F = (\mathcal F_1,\mathcal F_2,...,\mathcal F_n)$.
\State Define the ordered list $\mathcal S =$ \texttt{argsort}$(\mathcal F)$ where \texttt{argsort}$(\mathcal F)$ returns the index set $[n]$ sorted by descending order of $\mathcal F$, i.e., $\mathcal F_{\mathcal S_1}\geq \mathcal F_{\mathcal S_2}\geq ... \geq \mathcal F_{\mathcal S_n}$.
\For {$i \in \mathcal S$ (ordered traversal)}
\If {$\mathbf F(p^{(i\rightarrow 1)},Y)- \mathbf F(p^{(i\rightarrow 0)},Y) \geq 0$}
\State update $p \leftarrow p^{(i\rightarrow 1)}$
\Else
\State update $p \leftarrow p^{(i\rightarrow 0)}$
\EndIf
\EndFor
\If {$p \in \mathcal{V}$} break
\EndIf
\State $\mathcal V = \mathcal V \cup \{p\}$
\EndWhile
\State \textbf{return} $\hat X = p$
\end{algorithmic}
\end{algorithm}
The proof of Theorem~\ref{thm:ML_relaxation} inspires a heuristic for sequence reconstruction (see Alg.~\ref{alg:cood_switch}):
\begin{itemize}
\item Start from a given point $p = (p_1,...,p_n) \in [0,1]^n$.
\item One round of iteration is defined as follows: fix a traversal order for the indices $\{1,2,...,n\}$. Traverse through the indices $i$ in order and make $p_i$ either 0 or 1 depending on whether $\mathbf F(p^{(i\rightarrow 0)},Y)$ or $\mathbf F(p^{(i\rightarrow 1)},Y)$ is larger. This ensures that $\mathbf F(p,Y)$ never decreases.
\item At the end of the round, check if the resultant $p$ was already obtained at the end of a previous round: if so, end the algorithm (to prevent it from going into an endless cycle). Otherwise, start a new round from the resultant $p$.
\end{itemize}
Clearly the resultant $p$ at the end of a round is a lattice point since we make each $p_i$ to be 0 or 1. Therefore, the algorithm will end after a finite number of steps; in the worst case it will iterate through all $2^n$ sequences, although in practice we observe that it ends in 4-5 rounds (tested upto a blocklength of 100). We also note that the complexity of each round is $O(n^3)$ since it iterates through $n$ coordinates and for each coordinate computes $\mathbf F(\cdot)$, which is $O(n^2)$.
A natural question is whether it makes a difference if Alg.~\ref{alg:cood_switch} starts from an interior point ($p = (p_1,...,p_n) \in [0,1]^n$ where $\exists\ p_i \in (0,1)$) as compared to starting from a lattice point ($p = (p_1,...,p_n) \in \{0,1\}^n$). In Section~\ref{sec:Numerics} we show that starting from an interior point results in better estimation accuracy, thus supporting the usefulness of the ML relaxation.
\subsection{Proofs.}
\subsubsection{Proof of Theorem~\ref{thm:channel_equiv}}
\label{app:channel_equiv}
\iffalse
We first state a lemma (introduced in \cite{Srini2019}) which is closely related to the channel equivalence. The proof of the Lemma is relegated to Appendix~\ref{app:bin_inf_lemma}.
\bininfrelation*
The channel equivalence can essentially be tied to this lemma as follows: consider the two channel models in Fig.~\ref{fig:channel_equiv}. The probability of observations given the input in both cases is proportional to the number of ways of obtaining the observations given the input.
\begin{itemize}
\item For the $t$-trace deletion channel model in Fig.~\ref{fig:channel_equiv} (a), the number of ways to obtain the traces given the input is equal to ${X \choose \tilde{Y}^{(1)}}{X \choose \tilde{Y}^{(2)}}...{X \choose \tilde{Y}^{(t)}}$.
\item For the cascade model in Fig.~\ref{fig:channel_equiv} (b), the number of ways to obtain the traces given the input is equal to $\sum_{z} {X \choose z} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle$ (this is shown in the proof of the theorem).
\end{itemize}
Clearly by Lemma~\ref{lemma:bin_inf_relation}, the above two are equal, thus proving the equivalence of the two channel models. On the contrary, this equivalence provides a nice physical interpretation of Lemma~\ref{lemma:bin_inf_relation}.
\channelequiv*
\begin{proof}
We show the probabilistic equivalence between the two channel models in Fig.~\ref{fig:channel_equiv}: 1) $t$ independent deletion channels 2) cascade of a deletion channel with parameter $\delta^t$ with the \textit{remnant} channel with parameter $\delta$ defined as follows: an input symbol to the remnant channel is included in $k>0$ given outputs and deleted in the rest with a probability $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
Given this definition, we now first compute the probability of a given set of output sequences given an input sequence for the remnant channel, namely $\Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z)$. First, note that there can be multiple deletion patterns corresponding to outputs $\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}$ resulting from a given input $Z$. The number of such patterns is equal to $\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle$, which essentially follows from the definition of the infiltration product. Consider one such valid deletion pattern, i.e., a deletion pattern $\mathcal{D}$ that is a mapping of the symbols in $Z$ onto the symbols in $\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}$: $\mathcal{D}=\{(1,S_1),(2,S_2),...,(|Z|,S_{|Z|})\}$. Here $(i,S_i)$ represents the fact that $Z_i$ is not deleted in the output set $\tilde{Y}^{(S_i)}$ and is deleted in the rest. Clearly, $|S_i|>0$ from the definition of the remnant channel. Also $\sum_{i=1}^{|Z|}|S_i|=\sum_{j=1}^t|\tilde{Y}^{(j)}|$ since every symbol of each output is associated with exactly one input symbol and hence corresponds to one particular $S_i$. Thus,
\begin{align*}
\Pr(&\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z) \\&= \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z,\mathcal{D})\\
&=\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \prod_{i=1}^{|Z|}\frac{(1-\delta)^{|S_i|}\delta^{t-|S_i|}}{1-\delta^t}\\
&=\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \frac{(1-\delta)^{\sum|S_i|}\delta^{|Z|t-\sum |S_i|}}{(1-\delta^t)^{|Z|}}\\
&=\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},Z \rangle \frac{(1-\delta)^{\sum|\tilde{Y}^{(j)}|}\delta^{|Z|t-\sum |\tilde{Y}^{(j)}|}}{(1-\delta^t)^{|Z|}}.
\end{align*}
We can then compute the probability of the output given the input for the cascade channel as
\begin{align*}
\Pr(&\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|X)\\ &= \sum_{z} \Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)},Z=z|X)\\
&= \sum_{z} \Pr(Z=z|X)\Pr(\tilde Y^{(1)},\tilde Y^{(2)},...,\tilde Y^{(t)}|Z=z)\\
&= \sum_{z} \Bigg[{X \choose z} \delta^{t(|X|-|z|)}(1-\delta^t)^{|z|}\langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle \frac{(1-\delta)^{\sum|\tilde{Y}^{(j)}|}\delta^{|z|t-\sum |\tilde{Y}^{(j)}|}}{(1-\delta^t)^{|z|}}\Bigg]\\
&= \left[\sum_{z} {X \choose z} \langle \tilde Y^{(1)}\uparrow \tilde Y^{(2)} \uparrow...\uparrow \tilde Y^{(t)},z \rangle\right] \delta^{t|X|-\sum |\tilde{Y}^{(j)}|} {(1-\delta)^{\sum|\tilde{Y}^{(j)}|}}\\
&={X \choose \tilde{Y}^{(1)}}{X \choose \tilde{Y}^{(2)}}...{X \choose \tilde{Y}^{(t)}} \delta^{t|X|-\sum |\tilde{Y}^{(j)}|} {(1-\delta)^{\sum|\tilde{Y}^{(j)}|}}\\
&=\prod_{j=1}^t {X \choose \tilde{Y}^{(j)}} \delta^{|X|-|\tilde{Y}^{(j)}|} {(1-\delta)^{|\tilde{Y}^{(j)}|}}\\
&=\Pr(Y^{(1)}, Y^{(2)},..., Y^{(t)}|X),
\end{align*}
proving the equivalence.
\end{proof}
\fi
The intuition behind the theorem is that the cascade model splits the error events in the $t$-trace deletion channel into 2 parts:\\
- When an input symbol is deleted in all the traces, which is captured by the deletion channel with parameter $\delta^t$.\\
- When an input symbol is not deleted in at least one of the traces, captured by the remnant channel.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.4]{channel_equiv_proof.pdf}
\caption{The deletion error events occurring in the two channel models. Here `$-$' corresponds to a symbol being deleted and `$+$' corresponds to a transmission. The deletion pattern $D_i$ corresponds to the input symbol $X_i$.}
\end{figure}
In order to prove the theorem, we need to prove that the deletion patterns arising in the $t$-trace channel model and in the cascade model have the same distribution, i.e.,
\begin{align*}
\Pr(D_1=d_1,D_2=d_2,...,D_n=d_n) = \Pr(\widetilde D_1=d_1,\widetilde D_2=d_2,...,\widetilde D_n=d_n),
\end{align*}
where $d_i \in \{-,+\}^t$, where a $-$ corresponds to a deletion and a $+$ corresponds to a transmission.
Also from the definition of our channel models, the deletions act independently on each input symbol i.e., $D_i\ \raisebox{0.05em}{\rotatebox[origin=c]{90}{$\models$}}\ D_j$ for $i\neq j$. So it is sufficient to prove that the distributions of each $D_i$ and $\widetilde D_i$ are the same.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.5]{channel_equiv_proof_2.pdf}
\caption{The error events of the cascade model, expressed in terms of the error events of its components.}
\end{figure}
Consider $\widetilde D_i$ -- this is influenced by $\breve {D^0_i}$ which is the deletion in channel $\mathcal C_1$ and by $\breve D_i$ which are the deletion in the remnant channel $\mathcal C_2$. To prove the equivalence, we consider 2 cases:
\begin{itemize}
\item $d_i = (-,-,-,...,-)$, the error event where a symbol is deleted in all the observations. It can be seen that $\Pr(D_i = d_i)$ for this case is $\delta^t$. On the other hand, to compute $\Pr(\widetilde D_i = d_i)$, we note that this event is possible if and only if $\breve {D^0_i} = -$, since by definition, the remnant channel cannot delete the input symbol in all the $t$ observations. Therefore, $\Pr(\widetilde D_i = d_i)=\Pr(\breve D^0_i = -) = \delta^t$.
\item $d_i \neq (-,-,-,...,-)$, i.e., the input symbol is not deleted in at least one trace. Also let us define $k$ to be the count of $-$ in $d_i$. In this case, $\Pr(D_i = d_i)=\delta^{\text{Count(-) in } d_i}(1-\delta)^{\text{Count(+) in } d_i} = \delta^k (1-\delta)^{t-k}$. For the cascade model, this event requires that
$\breve {D^0_i} = +$ and $\breve {D_i}=d_i$. Thus, $$\Pr(\tilde D_i=d_i)=\Pr(\breve {D^0_i} = +)\cdot \Pr(\breve {D_i}=d_i) = (1-\delta^t) \frac{\delta^k (1-\delta)^{t-k}}{1-\delta^t}=\delta^k (1-\delta)^{t-k}.$$
\end{itemize}
In both cases, the distributions of $D_i$ and $\widetilde D_i$ are the same, proving the equivalence.
\subsubsection{Proof of Lemma~\ref{lemma:F_decomposition}}
\label{app:F_lemma_proof}
\deletionMLrellemma*
\begin{proof}
The proof of this lemma uses a similar approach as the proof of Thm.~\ref{thm:approx_smap_postprob}. First, in the expression for $\mathbf F(\cdot )$, we separate out the subsets that contain index $i$:
\begin{align*}
\mathbf F(p,y) &= \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j}\\
&= \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m, \\ i \notin \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j} + \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j}\\
&= \mathbf F(p_{[n]\backslash \{i\}},y) + \sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j}. \numberthis
\label{eq:Flemma_proof1}
\end{align*}
Now the second term can be further split as,
\begin{align*}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j} &= \sum_{k=1}^m\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ \mathcal S_k = i}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j}.
\end{align*}
One could express the set $\mathcal S$ as the union $\mathcal S = \mathcal S' \cup \{i\} \cup \mathcal S''$ such that $\mathcal S' \subseteq [i-1]$ and $\mathcal S'' \subseteq [i+1:n]$ to get
\begin{align*}
&\sum_{k=1}^m\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ \mathcal S_k = i}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j}\\
&= \sum_{k=1}^m\sum_{\substack{\mathcal S'|\\ \mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\sum_{\substack{\mathcal S''|\\ \mathcal S'' \subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \left( \prod\limits_{j=1}^{k-1} p_{\mathcal S'_j}^{y_j} (1-p_{\mathcal S'_j})^{1-y_j} \right) \left( p_{i}^{y_k} (1-p_{i})^{1-y_k} \right) \left( \prod\limits_{j=1}^{m-k} p_{\mathcal S''_j}^{y_{j+k}} (1-p_{\mathcal S''_j})^{1-y_{j+k}} \right)\\
&= \sum_{k=1}^m p_{i}^{y_k} (1-p_{i})^{1-y_k}\left(\sum_{\substack{\mathcal S'|\\ \mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}} \prod\limits_{j=1}^{k-1} p_{\mathcal S'_j}^{y_j} (1-p_{\mathcal S'_j})^{1-y_j} \right) \left( \sum_{\substack{\mathcal S''|\\ \mathcal S'' \subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \prod\limits_{j=1}^{m-k} p_{\mathcal S''_j}^{y_{j+k}} (1-p_{\mathcal S''_j})^{1-y_{j+k}} \right)\\
&= \sum_{k=1}^m p_{i}^{y_k} (1-p_{i})^{1-y_k} \mathbf F(p_{[i-1]},y_{[k-1]}) \mathbf F(p_{[i+1:n]},y_{[k+1:m]}).
\end{align*}
The $\sum_{k=1}^m$ summation in the above expression could further be split into the two cases depending on whether $y_k=0$ or $y_k=1$, which simplifies the term $p_{i}^{y_k} (1-p_{i})^{1-y_k}$ to either $1-p_i$ or $p_i$ respectively. Thus,
\begin{align*}
&\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ i \in \mathcal S}} \quad \prod\limits_{j=1}^m p_{\mathcal S_j}^{y_j} (1-p_{\mathcal S_j})^{1-y_j}\\
&= (1-p_i)\sum_{k|y_k=0} \mathbf F(p_{[i-1]},y_{[k-1]}) \mathbf F(p_{[i+1:n]},y_{[k+1:m]}) + p_i\sum_{k|y_k=1} \mathbf F(p_{[i-1]},y_{[k-1]}) \mathbf F(p_{[i+1:n]},y_{[k+1:m]}).\numberthis
\label{eq:Flemma_proof2}
\end{align*}
\noindent Plugging \eqref{eq:Flemma_proof2} in \eqref{eq:Flemma_proof1} concludes the proof of the Lemma.
\end{proof}
\subsubsection{Proof of Lemma~\ref{lemma:bin_inf_relation}}
The following Lemma forms the backbone of the analyses for multiple traces. This lemma is also closely related to the channel equivalence in Theorem~\ref{thm:channel_equiv}.
\label{app:bin_inf_lemma}
\bininfrelation*
\begin{proof}
The channel equivalence can essentially be tied to this lemma as follows: consider the two channel models in Fig.~\ref{fig:channel_equiv}. The probability of observations given the input in both cases is proportional to the number of ways of obtaining the observations given the input.
\begin{itemize}
\item For the $t$-trace deletion channel model in Fig.~\ref{fig:channel_equiv} (a), the number of ways to obtain the traces given the input is equal to ${X \choose {Y}^{1}}{X \choose {Y}^{2}}...{X \choose {Y}^{t}}$.
\item For the cascade model in Fig.~\ref{fig:channel_equiv} (b), the number of ways to obtain the traces given the input is equal to $\sum_{z} {X \choose z} \langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},z \rangle$, which we show below.
\end{itemize}
The above two are expression must be equal since the two channel models are equivalent.
We now first compute the probability of a given set of output sequences given an input sequence for the remnant channel, namely $\Pr(\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t}|Z)$. First, note that there can be multiple deletion patterns corresponding to outputs $\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t}$ resulting from a given input $Z$. The number of such patterns is equal to $\langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},Z \rangle$, which essentially follows from the definition of the infiltration product. Consider one such valid deletion pattern, i.e., a deletion pattern $\mathcal{D}$ that is a mapping of the symbols in $Z$ onto the symbols in $\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t}$: $\mathcal{D}=\{(1,S_1),(2,S_2),...,(|Z|,S_{|Z|})\}$. Here $(i,S_i)$ represents the fact that $Z_i$ is not deleted in the output set $\tilde{Y}^{S_i}$ and is deleted in the rest. From the definition of the remnant channel, we have $|S_i|>0$ . Also $\sum_{i=1}^{|Z|}|S_i|=\sum_{j=1}^t|\tilde{Y}^{j}|$ since every symbol of each output is associated with exactly one input symbol and hence corresponds to one particular $S_i$. Thus,
\begin{align*}
\Pr(\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t}|Z) &= \langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},Z \rangle \Pr(\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t}|Z,\mathcal{D})\\
&=\langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},Z \rangle \prod_{i=1}^{|Z|}\frac{(1-\delta)^{|S_i|}\delta^{t-|S_i|}}{1-\delta^t}\\
&=\langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},Z \rangle \frac{(1-\delta)^{\sum|S_i|}\delta^{|Z|t-\sum |S_i|}}{(1-\delta^t)^{|Z|}}\\
&=\langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},Z \rangle \frac{(1-\delta)^{\sum|\tilde{Y}^{j}|}\delta^{|Z|t-\sum |\tilde{Y}^{j}|}}{(1-\delta^t)^{|Z|}}.
\end{align*}
We can then compute the probability of the output given the input for the cascade channel as
\begin{align*}
\Pr(&\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t}|X)\\ &= \sum_{z} \Pr(\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t},Z=z|X)\\
&= \sum_{z} \Pr(Z=z|X)\Pr(\tilde Y^{1},\tilde Y^{2},...,\tilde Y^{t}|Z=z)\\
&= \sum_{z} \Bigg[{X \choose z} \delta^{t(|X|-|z|)}(1-\delta^t)^{|z|}\langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},z \rangle \frac{(1-\delta)^{\sum|\tilde{Y}^{j}|}\delta^{|z|t-\sum |\tilde{Y}^{j}|}}{(1-\delta^t)^{|z|}}\Bigg]\\
&= \left[\sum_{z} {X \choose z} \langle \tilde Y^{1}\uparrow \tilde Y^{2} \uparrow...\uparrow \tilde Y^{t},z \rangle\right] \delta^{t|X|-\sum |\tilde{Y}^{j}|} {(1-\delta)^{\sum|\tilde{Y}^{j}|}}. \numberthis
\label{eq:lemma2_proof_1}
\end{align*}
For the $t$-trace deletion channel model, we have:
\begin{align*}
\Pr(Y^1,Y^2,...,Y^t|X) &= \prod_{j=1}^t {X \choose Y^j} \delta^{|X|-|Y^j|}(1-\delta)^{|Y_j|} \\
&= {X \choose {Y}^{1}}{X \choose {Y}^{2}}...{X \choose {Y}^{t}} \delta^{t|X|-\sum |{Y}^{j}|} {(1-\delta)^{\sum|{Y}^{j}|}}. \numberthis
\label{eq:lemma2_proof_2}
\end{align*}
Equating \eqref{eq:lemma2_proof_1} and \eqref{eq:lemma2_proof_2} with $X = h$ and traces as $Y^j = \tilde Y^j = f_j$ proves the Lemma.
Alternatively, we use also induction to prove the statement as we do below. The statement is trivially true when $m=1$ since, $\sum_{w}{h \choose w}\langle f_1,w \rangle={h \choose f_1}$ as $\langle f,w \rangle=\mathbbm{1}_{f=w}$. We refer the reader to equation 6.3.25 in \cite{lothaire1997combinatorics} for the proof of the lemma for the case $m=2$. Assume that the statement is true for $m=k \in \mathbb{Z}, k\geq 2$. We next prove the validity when $m=k+1$. \\
Consider
\begin{align}
{h \choose f_1} {h \choose f_2}...{h \choose f_k}{h \choose f_{k+1}}&=\sum_{w}{h \choose w}\langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle {h \choose f_{k+1}}\nonumber\\
&=\sum_{w}\left[{h \choose w} {h \choose f_{k+1}}\right]\langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle\nonumber\\
&=\sum_{w}\left[\sum_v \langle w\uparrow f_{k+1},v \rangle {h \choose v}\right]\langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle\nonumber\\
&=\sum_{v}{h \choose v}\left[\sum_w \langle w\uparrow f_{k+1},v \rangle \langle f_1\uparrow f_2\uparrow ...\uparrow f_k,w \rangle\right]. \label{eq:prop2proof}
\end{align}
To evaluate the term in the square bracket, we use \eqref{def:infiltforseq}. For the case where $\tau \in \mathcal{A}^*,\sigma \in \mathbb{Z}\langle \mathcal{A} \rangle$ in \eqref{def:infiltforseq}, we have $$\sigma\uparrow \tau=\sum_{f\in \mathcal{A}^*} \langle \sigma,f \rangle (f\uparrow \tau),$$ and thus \begin{equation}
\langle \sigma \uparrow \tau,u \rangle=\sum_{f\in \mathcal{A}^*} \langle \sigma,f \rangle \langle f\uparrow \tau,u\rangle.
\label{eq:prop2proof2}
\end{equation} We use \eqref{eq:prop2proof2} to replace the term in the square bracket in \eqref{eq:prop2proof}, i.e.,
\begin{align}
{h \choose f_1} {h \choose f_2}...{h \choose f_k}{h \choose f_{k+1}}\nonumber\\
=\sum_{v}{h \choose v}\langle (f_1\uparrow f_2\uparrow ...\uparrow f_k) \uparrow f_{k+1},v \rangle,
\end{align}
and the lemma follows from the associativity property of the infiltration product.
\end{proof}
\subsubsection{Proof of Lemma~\ref{lemma:smapsum}}
\label{app:smapsum}
\begin{restatable}{lemma}{smapsum}
\label{lemma:smapsum}
\begin{align*}
\sum_{\substack{f||f|{=}n\\f_i{=}a}}{f \choose g} = 2^{n-|g|}\Bigg(\frac{1}{2}{n-1 \choose |g|} &+\sum_{j|g_j=a}{i-1 \choose j-1}{n-i \choose |g|-j}\Bigg),
\end{align*}
where $j\in\Big[\max\{1,|g|+i-n\}:\min\{i,|g|\}\Big]$.
\end{restatable}
\begin{proof}
First, observe that $${f \choose g} = \sum_{\substack{S\subseteq [n]:\\|S|=|g|}} \mathbbm 1_{f_S=g},$$ where the summation is over all ordered subsets of $[n]=\{1,2,...,n\}$ of size $|g|$ and $f_S$ corresponds to the subsequence of $f$ indexed by $S$. Thus,
\begin{align*}
\sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}}&{f \choose g} = \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \sum_{\substack{S\subseteq [n]|\\|S|=|g|}} \mathbbm 1_{f_S=g} = \sum_{\substack{S\subseteq [n]|\\|S|=|g|}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g}\\
&=\sum_{\substack{S\subseteq [n]|\\|S|=|g|\\i \notin S}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g} +
\sum_{\substack{S\subseteq [n]|\\|S|=|g|\\i \in S}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g}\\
&=\sum_{\substack{S\subseteq [n]|\\|S|=|g|\\i \notin S}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g} +
\sum_{j=1}^m \sum_{\substack{S\subseteq [n]|\\|S|=|g|\\S_j=i}} \sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}} \mathbbm 1_{f_S=g}. \numberthis \label{eq:lemmasmapsum}
\end{align*}
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.25]{lemma1.pdf}
\vspace*{-2mm}
\caption{Figure illustrating proof of Lemma~\ref{lemma:smapsum}.}
\label{fig:lemmasmapsum}
\end{center}
\end{figure}
The two terms in \eqref{eq:lemmasmapsum} can be visualized as the number of ways to fill up the blank spaces (spaces without arrows pointing to it in $f$) in Fig.~\ref{fig:lemmasmapsum}(a) and (b) respectively. Solving this counting problem, we get
$$\sum_{\substack{f\in\mathcal{A}^n|\\f_i=a}}{f \choose g}=2^{n-|g|}\left(\frac{1}{2}{n-1 \choose |g|}+\sum_{j|g_j=a}{i-1 \choose j-1}{n-i \choose |g|-j}\right).$$
\end{proof}
\subsection{Dynamic program to compute $\mathbf F(\cdot)$ and $\nabla \mathbf F(\cdot)$}
\subsubsection{Computation of $\mathbf F(p,v)$}
\label{app:F_compute}
We here describe how to compute $\mathbf F(p,v)$ in $O(mn)$ time and space complexity, where $p=(p_1,...,p_n)$ and $v=v_1...v_m$, via a dynamic programming approach. Note that $m \leq n$ otherwise $\mathbf F(p,v)=0$.
We first define
\begin{align*}
&\mathbf G^{for}(k,j)\triangleq \mathbf F(p_{[1:k]},v_{[1:j]}).
\numberthis
\label{eq:G}
\end{align*}
Using Lemma~\ref{lemma:F_decomposition} with $i=n$, we get
\begin{align*}
\mathbf F(p,v) = \mathbf F(p_{[n-1]},v) + p_n^{v_m} (1-p_n)^{(1-v_m)} \mathbf F(p_{[n-1]},v_{[m-1]}).
\end{align*}
\noindent This translates to the following dynamic program for $\mathbf G^{for}$:
\begin{align*}
\mathbf G^{for}(k,j) = \mathbf G^{for}(k-1,j)+ p_{k}^{v_j}(1-p_{k})^{1-v_j}&\mathbf G^{for}(k-1,j-1),\numberthis \label{eq:approx_smap_dpfor}
\end{align*}
with the boundary conditions $\mathbf G^{for}(k,0)=1\ \forall\ k \geq 0$ and $\mathbf G^{for}(k,j)=0\ \forall\ k<j$.
The algorithm is now summarized as Alg.~\ref{alg:F_comp}.
\begin{algorithm}[h!]
\caption{Computing $\mathbf F(p,v)$}\label{alg:F_comp}
\begin{algorithmic}[1]
\State Inputs: $p \in [0,1]^n$, $v \in \{0,1\}^m$
\State Outputs: $\mathbf F(p_{[1:k]},v_{[1:j]})$ for all $k \in [n]$ and $j\in[m]$
\State Initialize $\mathbf G^{for}(k,0)=1\ \forall\ k$ and $\mathbf G^{for}(k,j)=0\ \forall\ k<j$
\For {$k = 1:n$ and $j = 1:m$}
\State Use \eqref{eq:approx_smap_dpfor} to update $\mathbf G^{for}(k,j)$
\EndFor
\State return $\mathbf G^{for}(k,j)\ \forall\ k,j$
\end{algorithmic}
\end{algorithm}
We note that a similar dynamic programming approach yields $\mathbf F(p_{[k+1:n]},v_{[j+1:m]})$ for all $k \in [n]$ and $j\in[m]$ in $O(mn)$ time and space complexity by defining
\begin{align*}
&\mathbf G^{rev}(k,j)\triangleq \mathbf F(p_{[k+1:n]},v_{[j+1:m]}).
\end{align*}
\noindent The following dynamic program can be used for $\mathbf G^{rev}$:
\begin{align*}
\mathbf G^{rev}(k,j) = \mathbf G^{rev}(k+1,j)+ p_{k+1}^{v_{j+1}}(1-p_{k+1})^{1-v_{j+1}}&\mathbf G^{rev}(k+1,j+1),\numberthis \label{eq:approx_smap_dprev}
\end{align*}
with the boundary conditions $\mathbf G^{rev}(k,m)=1\ \forall\ k \geq 0$ and $\mathbf G^{rev}(k,j)=0\ \forall\ k,j: n-k<m-j$. \\
\subsubsection{Computation of $\nabla_p \mathbf F(p,v)$}
\label{app:F_grad_comp}
First, from Lemma~\ref{lemma:F_decomposition}, we have
\begin{align*}
\mathbf F(p,v) = \mathbf F(p_{[n]\backslash \{i\}},v) +& (1-p_i)\sum_{k|v_k=0} \mathbf F(p_{[i-1]},v_{[k-1]}) \mathbf F(p_{[i+1:n]},v_{[k+1:m]}) \\&+ p_i\sum_{k|v_k=1} \mathbf F(p_{[i-1]},v_{[k-1]}) \mathbf F(p_{[i+1:n]},v_{[k+1:m]}).
\end{align*}
Differentiating with respect to $p_i$, we get
\begin{align*}
\frac{\partial \mathbf F(p,v)}{\partial p_i} &= \sum_{k|v_k=1} \mathbf F(p_{[i-1]},v_{[k-1]}) \mathbf F(p_{[i+1:n]},v_{[k+1:m]}) - \sum_{k|v_k=0} \mathbf F(p_{[i-1]},v_{[k-1]}) \mathbf F(p_{[i+1:n]},v_{[k+1:m]}) \\
&= \sum_{k|v_k=1} \mathbf G^{for}(i{-}1,k{-}1)\mathbf G^{rev}(i,k) - \sum_{k|v_k=0}\mathbf G^{for}(i{-}1,k{-}1)\mathbf G^{rev}(i,k). \numberthis
\label{eq:F_grad_comp}
\end{align*}
Thus, computing the $\mathbf G^{for}$ and $\mathbf G^{rev}$ terms is sufficient to compute the gradient. As discussed above, this computation requires $O(nm)$ operations. Given $\mathbf G^{for}$ and $\mathbf G^{rev}$, the computation of each partial derivative $\frac{\partial \mathbf F(p,v)}{\partial p_i}$ requires $O(m)$ operations, and we need to compute $n$ such partial derivatives. Thus, the complexity of computing $\nabla_p \mathbf F(p,v)$ can be done in $O(nm)$ time and space complexity.
\begin{algorithm}[h!]
\caption{Computing $\nabla_p \mathbf F(p,v)$}\label{alg:F_grad_comp}
\begin{algorithmic}[1]
\State Inputs: $p \in [0,1]^n$, $v \in \{0,1\}^m$
\State Outputs: $\nabla_p \mathbf F(p,v)$
\State Initialize $\mathbf G^{for}(k,0)=1\ \forall\ k$ and $\mathbf G^{for}(k,j)=0\ \forall\ k<j$
\State Initialize $\mathbf G^{rev}(k,m)=1\ \forall\ k$ and $\mathbf G^{rev}(k,j)=0\ \forall\ k,j: n-k<m-j$
\For {$k = 1:n$ and $j = 1:m$}
\State Use \eqref{eq:approx_smap_dpfor} and \eqref{eq:approx_smap_dprev} to compute $\mathbf G^{for}(k,j)$ and $\mathbf G^{rev}(k,j)$
\EndFor
\For {$i = 1:n$}
\State Use \eqref{eq:F_grad_comp} to compute $\frac{\partial \mathbf F(p,v)}{\partial p_i} $
\EndFor
\State return $\nabla_p \mathbf F(p,v)$
\end{algorithmic}
\end{algorithm}
\subsection{An algebraic definition of the infiltration product.}
\label{app:infil_def}
For completeness, we reproduce the formal definition of the infiltration product from Section 6.3 of \cite{lothaire1997combinatorics} (also see there for the equivalence of the two definitions). A \textit{formal series} with indeterminates (or variables) in a set $\mathcal A$ and coefficients in a commutative ring $\mathcal R$, is a mapping of $\mathcal A^*$ onto $\mathcal R$. Recall that a commutative ring is a set which forms an abelian group under an \textit{addition} operation, is a monoid under a \textit{multiplication} operation which commutes, and the multiplication operation distributes over addition. Here we consider $\mathbb Z$, the set of integers as the commutative ring $\mathcal{R}$. A formal series is called a \textit{polynomial} if only a finite number of sequences are mapped to non-zero values, the rest of the sequences map to zero. Consider two polynomials $\sigma,\tau: \mathcal{A}^*\rightarrow \mathbb Z$. The value taken by a sequence $w\in \mathcal A^*$ on $\sigma$ (or the coefficient of $w$ in $\sigma$) is denoted by $\langle \sigma,w\rangle \in \mathbb R$. We also define binary addition ($\oplus$) and multiplication operations ($\times$) on the set of polynomials as follows:
\begin{align}
\langle \sigma\oplus \tau,w\rangle \triangleq \langle \sigma,w\rangle + \langle \tau,w \rangle \quad \forall w\in \mathcal A^*,\label{eq:polynomial_add}\\
\langle \sigma\times \tau,w\rangle \triangleq \sum_{\substack{f,g\in \mathcal A^*:\\ f.g=w}}\langle \sigma,f\rangle \langle \tau,g \rangle \quad \forall w\in \mathcal A^*.\label{eq:polynomial_prod}
\end{align}
We will use the usual symbols $+$ and $.$ in place of $\oplus$ and $\times$ in this work for convenience. The meaning of the operation would be clear depending on the operands. With these operations the set of polynomials form a non-commutative ring, and is denoted by $\mathbb Z\langle\mathcal A \rangle$, also called the free $\mathbb Z$-algebra on $\mathcal A$ in ring theory. Note that the addition and multiplication operations defined in \eqref{eq:polynomial_add} and \eqref{eq:polynomial_prod} are similar to the operations defined on commutative polynomials, except that the multiplication operation under the summation in \eqref{eq:polynomial_prod} ($f.g=w$) is actually concatenation and is non-commutative. The multiplication inside the summation in \eqref{eq:polynomial_prod} is multiplication in the real field and hence commutative. The multiplication defined in \eqref{eq:polynomial_prod} distributes over addition defined in \eqref{eq:polynomial_add}. Thus, a polynomial in $\mathbb Z\langle\mathcal A \rangle$ can be represented as a sum of monomials in $\mathcal A^*$ each with an associated coefficient in $\mathbb Z$, i.e., $\sigma=\sum\limits_{w\in \mathcal A^*} \langle\sigma,w \rangle w$. Define the \textit{degree} of a polynomial to be equal to the length of a longest sequence with a non-zero coefficient in the polynomial and the \textit{number of terms} of a polynomial as the number of sequences with non-zero coefficients in the polynomial. Note that a degree $d$ polynomial could have a number of terms upto $2^{d+1}-1$.
With this, the \textit{infiltration product} (in general, for two polynomials) is defined as follows:
\begin{align}
\forall f\in \mathcal{A}^*,& \quad f\uparrow e = e\uparrow f=f.\nonumber \\
\forall f,g\in \mathcal{A}^*&,\quad \forall a,b\in \mathcal{A}, \nonumber\\
fa\uparrow gb=(f\uparrow gb)a&+(fa\uparrow g)b+\mathbbm{1}_{a=b}(f\uparrow g)a.\nonumber \\
\forall \sigma,\tau \in \mathbb{Z}\langle\mathcal{A} \rangle, \quad &\sigma\uparrow \tau=\sum_{f,g\in \mathcal{A}^*} \langle \sigma,f \rangle \langle \tau,g \rangle (f\uparrow g). \label{def:infiltforseq}
\end{align}
\subsection{Symbolwise posterior probabilities for the remnant channel}
\label{app:remnant_postprob}
Consider the remnant channel shown below, and let $Z=Z_1Z_2...Z_n$. Also let $Z_i \sim \text{Ber}(0.5)$. We aim to compute $\Pr(Z_i=1|\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.4]{remnant_channel.pdf}
\caption{The remnant channel}
\end{figure}
From the definition of the infiltration product, the input-output relation for this channel can be derived to be:
\begin{align*}
\Pr(\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t|Z) =\langle y^1\uparrow y^2 \uparrow...\uparrow y^t,Z \rangle \frac{(1-\delta)^{\sum|y^j|}\delta^{nt-\sum |y^j|}}{(1-\delta^t)^{n}}.
\end{align*}
Now, one could write the symbolwise posterior probabilities for $Z$ as:
\begin{align*}
\Pr(Z_i=1&|\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t) = \sum_{\substack{z||z|=n,\\z_i=1}} \Pr(z|\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t)\\
&{=} \frac{1}{2^n \Pr(\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t)} \sum_{\substack{z||z|=n,\\z_i=1}} \Pr(\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t|z)\\
&{=} \frac{{(1-\delta)^{\sum|y^{j}|}\delta^{nt-\sum |y^{j}|}}}{(1-\delta^t)^{n} 2^n \Pr(\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t)} \sum_{\substack{z||z|=n,\\z_i=1}} \langle y^1 \uparrow y^2 \uparrow ... \uparrow y^t,z \rangle. \numberthis
\label{eq:remant_map_prob_1}
\end{align*}
A similar expression can be obtained for the case when $Z_i=0$ as
\begin{align*}
\Pr(Z_i=0&|\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t)\\
&{=} \frac{{(1-\delta)^{\sum|y^{j}|}\delta^{nt-\sum |y^{j}|}}}{(1-\delta^t)^{n} 2^n \Pr(\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t)} \sum_{\substack{z||z|=n,\\z_i=0}} \langle y^1 \uparrow y^2 \uparrow ... \uparrow y^t,z \rangle. \numberthis
\label{eq:remant_map_prob_0}
\end{align*}
We could further simplify \eqref{eq:remant_map_prob_1} and \eqref{eq:remant_map_prob_0} using the fact that the expressions in \eqref{eq:remant_map_prob_1} and \eqref{eq:remant_map_prob_0} must sum to 1, leading us to
\begin{align*}
\Pr(Z_i=1|\tilde Y^{1}=y^1,\tilde Y^{2}=y^2,...,\tilde Y^{t}=y^t) = \frac{ \sum\limits_{\substack{z||z|=n,\\z_i=1}} \langle y^1 \uparrow y^2 \uparrow ... \uparrow y^t,z \rangle}{\sum\limits_{\substack{z||z|=n}} \langle y^1 \uparrow y^2 \uparrow ... \uparrow y^t,z \rangle}. \numberthis \label{eq:remnant_map_prob}
\end{align*}
We precisely describe the algorithm which computes the terms in \eqref{eq:remnant_map_prob} in section~\ref{sec:exactsmap}, by exploiting the edit graph interpretation of the infiltration product, but give a high level idea below. The complexity of such an algorithm is $O((2n)^t)$ which is equal to the number of edges in the edit graph. Note that for a fixed number of traces, this algorithm is polynomial in the blocklength as opposed to a naive approach of iterating through all the $n$-length sequences.
Recall that $\langle y^1 \uparrow y^2 \uparrow ... \uparrow y^t,z \rangle$ is the number of paths from origin to destination of the edit graph $\mathcal G(y^1,y^2,...,y^t)$ which correspond to $z$. Therefore,
$\sum_{\substack{z||z|=n}} \langle y^1 \uparrow y^2 \uparrow ... \uparrow y^t,z \rangle$ is equal to the number of $n$-length paths in $\mathcal G(y^1,y^2,...,y^t)$ from the origin to the destination. Note that the edit graph has no cycles, so this quantity can be efficiently computed via the following dynamic program -- the number of $n$ length paths from the origin to a vertex $v$ is equal to the sum of the number of $n-1$ length paths from the origin to the in-neighbors of $v$. Such a procedure iterates over the vertex set of $\mathcal G(y^1,y^2,...,y^t)$ exactly once.
The numerator term $\sum_{\substack{z||z|=n\\z_i=1}}\langle y^1 \uparrow y^2 \uparrow ... \uparrow y^t,z \rangle$ can be interpreted in a similar way: it is equal to the number of $n$-length paths in $\mathcal G(y^1,y^2,...,y^t)$ from the origin to the destination such that the $i^{th}$ edge of the path corresponds to a `1'. The algorithm for this, therefore, follows a similar principle but has an extra step. For each vertex $v$, we compute
\begin{itemize}
\item the number of paths from the origin to $v$ of length $0,1,...,n$,
\item the number of paths from $v$ to the destination of length $0,1,...,n$.
\end{itemize}
Next we iterate over all edges in $\mathcal G(y^1,y^2,...,y^t)$ corresponding to a `1' and accumulate the number of $n$ length paths which have this particular edge as its $i^{th}$ edge. Thus, this algorithm iterates over the vertex set twice and the edge set of $\mathcal G(y^1,y^2,...,y^t)$ once.
\subsection{A heuristic for ML optimization with a single trace.}
{The proof of Theorem~\ref{thm:ML_relaxation} inspires a heuristic for sequence reconstruction (see Alg.~\ref{alg:cood_switch}):
\begin{itemize}
\item Start from a given point $p = (p_1,...,p_n) \in [0,1]^n$.
\item One round of iteration is defined as follows: fix a traversal order for the indices $\{1,2,...,n\}$. Traverse through the indices $i$ in order and make $p_i$ either 0 or 1 depending on whether $\mathbf F(p^{(i\rightarrow 0)},y)$ or $\mathbf F(p^{(i\rightarrow 1)},y)$ is larger. This ensures that $\mathbf F(p,y)$ never decreases.
\item At the end of the round, check if the resultant $p$ was already obtained at the end of a previous round: if so, end the algorithm (to prevent it from going into an endless cycle). Otherwise, start a new round from the resultant $p$.
\end{itemize}
The resultant $p$ at the end of a round is a lattice point since we make each $p_i$ to be 0 or 1. Therefore, the algorithm will end after a finite number of steps; in the worst case it will iterate through all $2^n$ sequences, although in practice we observe that it ends in 4-5 rounds (tested up to a blocklength of 100). We also note that the complexity of each round is $O(n^3)$ since it iterates through $n$ coordinates and for each coordinate computes $\mathbf F(\cdot)$, which is $O(n^2)$.}
\begin{algorithm}[t!]
\caption{Coordinate switch ML heuristic}\label{alg:cood_switch}
\begin{algorithmic}[1]
\item Input: Blocklength $n$, Trace {$Y=y$}, Initial point $p = (p_1,p_2,...,p_n)$ \\ Outputs: Estimated sequence $\hat X$
\State Initialize visited set $\mathcal V = \emptyset$
\While {True}
\State Compute $\mathcal F_i = |\mathbf F(p^{(i\rightarrow 1)},y)- \mathbf F(p^{(i\rightarrow 0)},y)|\ \forall\ i$ and let $\mathcal F = (\mathcal F_1,\mathcal F_2,...,\mathcal F_n)$.
\State Define the ordered list $\mathcal S =$ \texttt{argsort}$(\mathcal F)$ where \texttt{argsort}$(\mathcal F)$ returns the index set $[n]$ sorted by descending order of $\mathcal F$, i.e., $\mathcal F_{\mathcal S_1}\geq \mathcal F_{\mathcal S_2}\geq ... \geq \mathcal F_{\mathcal S_n}$.
\For {$i \in \mathcal S$ (ordered traversal)}
\If {$\mathbf F(p^{(i\rightarrow 1)},y)- \mathbf F(p^{(i\rightarrow 0)},y) \geq 0$}
\State update $p \leftarrow p^{(i\rightarrow 1)}$
\Else
\State update $p \leftarrow p^{(i\rightarrow 0)}$
\EndIf
\EndFor
\If {$p \in \mathcal{V}$} break
\EndIf
\State $\mathcal V = \mathcal V \cup \{p\}$
\EndWhile
\State \textbf{return} $\hat X = p$
\end{algorithmic}
\end{algorithm}
A natural question is whether it makes a difference if Alg.~\ref{alg:cood_switch} starts from an interior point ($p = (p_1,...,p_n) \in [0,1]^n$ where $\exists\ p_i \in (0,1)$) as compared to starting from a lattice point (for instance, we could start from $p = (y,0,...,0) \in \{0,1\}^n$) which is the $n$-length sequence obtained via appending $y$ with zeros. It turns out that starting from an interior point results in better accuracy on both Hamming and edit error rate metrics, thus supporting the usefulness of our ML relaxation result.
In Fig.~\ref{fig:singletrace}, we compare the performance of Coordinate switch heuristic with the other trace reconstruction heuristics in Section~\ref{sec:Numerics}. We see that the coordinate switch with interior point initialization performs very similar to the true ML sequence (obtained via exhaustive search), in terms of both the Hamming error rate as well as the edit error rate. This intuitively supports the idea that this is a good heuristic for the ML optimization problem. However, at this point the heuristic is applicable for reconstruction using just a single trace and it is unclear on how to extend it to multiple traces.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.5]{Numerics/single_trace.pdf}
\caption{Numerics for reconstruction from a single trace for a blocklength $n=20$. This plot compares the performance of coordinate switch heuristic (abbreviated ``Coodsw. interior init.'' and ``Coodsw. lattice init.'') with other trace reconstruction algorithms from Section~\ref{sec:Numerics}. ``ML'' refers to the true ML sequence obtained via an exhaustive search on all 20 length binary sequences. The interior point initialization initializes $p=(0.5,0.5,...,0.5)$ while the lattice point initialization appends the trace $y$ with zeros to obtain an $n$-length vector $p=(y,0,...,0)$.}
\label{fig:singletrace}
\end{figure}
\subsection{Symbolwise MAP as the minimizer of Hamming error rate}
\label{app:smap_hamming}
Symbolwise MAP is an optimal estimator for minimizing the Hamming error rate for any channel, regardless of whether it is memoryless or not. This fact can be seen from the following argument: Consider a fixed observation $y$ (note that $y$ here can also be a collection of multiple observations, our arguments which follow remain unchanged) and that we aim to estimate a binary input sequence $X$; let the estimate of the input be $\hat X(y)$. Note that the estimate is a function of observation $y$ alone. Now the Hamming error rate of any estimator given $y$ is the expectation (over all inputs) of number of symbol mismatches divided by the blocklength, i.e.,
\begin{align*}
\frac{1}{n}\E \left[ \sum_{i=1}^n \mathbbm{1} \{X_i \neq \hat X_i(y)\} \Big| Y=y \right ]
&= \frac{1}{n} \sum_{i=1}^n \E \left [ \mathbbm{1}\{X_i \neq \hat X_i(y)\} \Big| Y=y \right ]\\
& = \frac{1}{n} \sum_{i=1}^n \Pr\left ( X_i \neq \hat X_i(y)\Big| Y=y \right ) \\
& = \frac{1}{n} \sum_{i=1}^n \Bigg( \Pr(X_i = 0|Y=y)\Pr(\hat X_i(y) = 1|X_i = 0,Y=y) \\ & \hspace{1cm}+
\Pr(X_i = 1|Y=y)\Pr(\hat X_i(y) = 0|X_i = 1,Y=y) \Bigg ).
\end{align*}
But, $\hat X_i$ is a function of only $y$ and hence is conditionally independent of $X_i$ given $y$, which implies the following:
\begin{align*}
\frac{1}{n}\E \left[ \sum_{i=1}^n \mathbbm{1} \{X_i \neq \hat X_i(y)\} \Big| Y=y \right ]
& = \frac{1}{n} \sum_{i=1}^n \Bigg( \Pr(X_i = 0|Y=y)\Pr(\hat X_i(y) = 1|Y=y) \\ & \hspace{2cm}+
\Pr(X_i = 1|Y=y)\Pr(\hat X_i(y) = 0|Y=y)\Bigg ).
\end{align*}
To simplify notation, let the posterior probabilities be $q_i(y) \triangleq \Pr(X_i = 1|Y=y)$ and let $\alpha_i(y) \triangleq \Pr(\hat X_i(y) = 1|Y=y)$. Note that $q_i(y)$ is a property of the channel and is fixed given $y$, while $\alpha_i(y)$ depends on the design of our estimator. With this, the above expression can be re-written as
$$\frac{1}{n}\E \left[ \sum_{i=1}^n \mathbbm{1} \{X_i \neq \hat X_i(y)\} \Big| Y=y \right ] = \frac{1}{n} \sum_{i=1}^n \Bigg( (1-q_i(y)) \alpha_i(y) + q_i(y) (1-\alpha_i(y))\Bigg ).$$ The optimal assignment of $\alpha_i(y)$ to minimize this expression is $\alpha_i(y) = 1$ if $q_i(y) \geq 0.5$ and $\alpha_i(y) = 0$ otherwise, which coincides with the symbolwise MAP estimate. This proves the optimality of symbolwise MAP for minimizing the Hamming error rate given any observation $y$, for any channel.
\section{Conclusions}
\label{sec:conclusions}
In this work we gave, to the best of our knowledge, the first results and techniques to compute posterior distributions over single and multiple deletion channels. We also provided a new perspective on the maximum-likelihood for the deletion channel by showing an equivalence between a discrete optimization problem and its relaxed version. In this process, we introduced a variety of tools (the relaxed binomial coefficient, edit graph and infiltration product) and demonstrated their use for analyzing deletion channels. We also presented numerical evaluations of our algorithms and showed performance improvements over existing trace reconstruction algorithms.
\section{Introduction}
\label{sec:intro}
Sequence reconstruction over deletion channels, both with and without a codebook, has received considerable attention in the information theory as well as in the theoretical computer science literature. From an information theory perspective, reconstruction over the deletion channel, or more specifically a maximum-likelihood (ML) argument for the deletion channel, would give further insight on the capacity of the deletion channel, a long-standing open problem (see \cite{mitzenmacher2009survey}). To quote \cite{mitzenmacher2009survey} -- ``at the very least, progress in this direction would likely surpass previous results on the capacity of the deletion channels''. Yet, there are no results on reconstruction over a deletion channel with statistical guarantees -- in this work, we take a step in this direction.
On the other hand, the problem of \textit{trace reconstruction}, as introduced in \cite{Batu2004}, has received renewed interest in the past few years (see \cite{Holenstein2008}, \cite{Peres2017}, \cite{De2017}, \cite{holden18}, \cite{Nazarov:2017}, \cite{holden2018lower}, \cite{chase2019new}). The problem of trace reconstruction can be stated simply as follows: consider a sequence $X$ which is simultaneously passed through $t$ independent deletion channels to yield $t$ deleted observations (also called \textit{traces}) of $X$ (see Fig.~\ref{fig:tdeletion}). How many such traces are needed to reconstruct $X$ perfectly? A variety of upper and lower bounds for this problem have been proposed, both for worst case and average case reconstruction. Our problem definition, stated in the following paragraph, is closely related to the average case trace reconstruction.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{tdeletion.pdf}
\caption{$t$-deletion channel model: sequence $X$ passed through $t$ independent deletion channels to yield $t$ \textit{traces}. We aim to estimate $X$ from the $Y^{(i)}$s.}
\label{fig:tdeletion}
\end{center}
\end{figure}
\noindent \textbf{Problem definition.} Given an input sequence of length $n$ (known apriori), the independently and identically distributed (i.i.d.) deletion channel deletes each input symbol indepedently with probability $\delta$, producing at its output a subsequence of the input sequence. Consider a sequence $X$ passed through $t$ ($t$ is fixed) such deletion
channels as shown in
Fig.~\ref{fig:tdeletion}. We call this the $t$-deletion channel model. We ask two questions:
\begin{enumerate}
\item For $t=1$ (see Fig.~\ref{fig:1deletion}, also called the single deletion channel model), and when $X_i \sim\ \text{ind. Ber}(p_i)$, compute the posterior distributions of $X_i$ given the trace $Y$.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.25]{1deletion.pdf}
\caption{The single deletion channel model. We assume $X_i \sim\ \text{ind. Ber}(p_i)$.}
\label{fig:1deletion}
\end{center}
\end{figure}
\item In the $t$ deletion channel model, for a fixed $t$ assume that $X_i \sim\ \text{i.i.d. Ber}(0.5)$ and compute the posterior distributions of $X_i$ given all traces $Y^{(1)}, Y^{(2)},...,Y^{(t)}$.
\end{enumerate}
Note that solving 1) above doesn't lead to a natural solution for 2). This is because for a memoryless channel, we have $Y^{(j)} - X_i - Y^{(k)}$ and hence $\Pr(X_i=\alpha|Y^{(j)}, Y^{(k)}) \propto \Pr(X_i=\alpha|Y^{(j)}) \Pr(X_i=\alpha|Y^{(k)})$; so one could independently combine the posterior probabilities from each noisy observation. This is not the case for deletion channels since the markov chain $Y^{(j)} - X_i - Y^{(k)}$ no longer holds. More intuitively, one needs to first ``align'' the traces for computing the likelihoods.
We point out that the problem considered in 2) asks a question complementary to the trace reconstruction: given a fixed (possibly a few) number of traces, what is our ``best'' guess of $X$? We provide algorithms which do this. Unlike trace reconstruction, we are not concerned with perfect reconstruction (since perfect reconstruction may not be possible with just a few traces), although it should also be obvious that performance guarantees for our algorithms (not a part of this work) would naturally lead to upper bounds for trace reconstruction.
Deletion channel by itself is known to be notoriously difficult to analyse. As stated earlier, the capacity of a single deletion channel is still unknown (\cite{diggavi2007capacity,diggavi2006information,diggavi2001transmission}); as are optimal coding schemes. Recent works have looked at the design of codes for deletion channels (\cite{ratzer2005marker,ratzer2000codes,thomas2017polar}); these works consider use of a codebook (we do not). As a result, statistical estimation over deletion channels is also a difficult problem due its highly combinatorial nature. To the best of our knowledge, as yet there are no efficient estimation algorithms over deletion channels with statistical guarantees; not even for ML over a single deletion channel.
\noindent \textbf{Biological motivation.} Trace reconstruction in itself was motivated, in part, by problems DNA sequence reconstruction. One such problem was to infer the DNA sequence of a common ancestor from the samples of its descendants. We argue that our problem definition fits more naturally in such a scenario since perfect reconstruction may not be feasible or even possible. Our motivation for considering this problem also comes from a recent DNA sequencing technology called \textit{nanopore sequencing}. The $t$-deletion channel model is a simplistic model to approximately capture the process of a DNA sequence passed through a nanopore sequencer\footnote{As seen in
\cite{Mao2017},\cite{MDK17} there are more complicated effects of
the nanopore reader not captured in this simple
representation.}.
Very recently, a variant of the trace reconstruction problem called \textit{coded trace reconstruction} has been proposed, motivated by portable DNA-based data storage systems using DNA nanopores (see \cite{abroshan2019coding}, \cite{cheraghchi2019coded}, \cite{brakensiek2019coded}) and we believe that the ideas in this work may prove useful in such a setting.
There are other works on sequence assembly (see for example, \cite{Li09fastand}, \cite{Shomorony2016}),
where multiple short reads (from different segments of a sequence) are used to reconstruct the bigger
sequence. This work differs from sequence assembly since we are interested to infer the entire length sequence and not just small segments of it (which are then ``stitched'' together in sequence assembly).
\noindent \textbf{Tools and techniques.} In terms of theoretical tools, the series of books by Lothaire (\cite{lothaire1997combinatorics,lothaire2002algebraic,lothaire2005applied}) extensively use algebraic tools for problems in the combinatorics of sequences (or \textit{words}), and our work is inspired by such techniques. We borrow many of their notations and results for our work.
\noindent \textbf{Contributions.} {\color{red}} Our main contribution is to provide tools to visualize and analyze error events (described in the next section) for the multiple deletion channel model in Fig.~\ref{fig:tdeletion}. We also provide algorithms to solve the problems stated in 1) and 2) earlier in the section.
\begin{itemize}[wide=5pt]
\item In section~\ref{sec:1deletion}, for the single deletion channel model, we provide an $O(n^2)$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i=1|Y)\ \forall\ i$ when $X_i \sim \text{ind. Ber}(p_i)$.
\item In Section~\ref{sec:exactsmap}, for the $t$-deletion channel model, we give an $O(2^t n^{t+2})$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i = 1|Y_1,...,Y_t)$ when $X_i \sim \text{ind. Ber}(0.5)$.
\end{itemize}
\section{Introduction}
\label{sec:intro}
Sequence reconstruction over deletion channels, both with and without a codebook, has received considerable attention in the information theory as well as in the theoretical computer science literature. From an information theory perspective, reconstruction over the deletion channel, or more specifically a maximum-likelihood (ML) argument for the deletion channel, would give further insight on the capacity of the deletion channel, a long-standing open problem (see \cite{mitzenmacher2009survey}). To quote \cite{mitzenmacher2009survey} -- ``at the very least, progress in this direction would likely surpass previous results on the capacity of the deletion channels''. Yet, there are no results on reconstruction over a deletion channel with statistical guarantees. In this work, we take steps in this direction.
In this space, the problem of \textit{trace reconstruction}, as introduced in \cite{Batu2004}, has also received renewed interest in the past few years (see \cite{Holenstein2008,Peres2017}, \cite{De2017}, \cite{holden18}, \cite{Nazarov:2017}, \cite{holden2018lower}, \cite{chase2019new}). The problem of trace reconstruction can be stated simply as follows: consider a sequence $X$ which is simultaneously passed through $t$ independent deletion channels to yield $t$ output subsequences (also called \textit{traces}) of $X$ (see Fig.~\ref{fig:tdeletion}). How many such traces are needed to reconstruct $X$ perfectly? A variety of upper and lower bounds for this problem have been proposed, both for worst case and average case reconstruction. Our problem formulation is complementary to this, as we discuss next.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{tdeletion.pdf}
\caption{The $t$-trace deletion channel model: the sequence $X$ is passed through $t$ independent deletion channels to yield $t$ \textit{traces}. We aim to estimate $X$ from the $Y^{i}$s.}
\label{fig:tdeletion}
\end{center}
\end{figure}
\noindent \textbf{Problem formulation.} Given an input sequence of length $n$ (known apriori), the independently and identically distributed (i.i.d.) deletion channel deletes each input symbol indepedently with probability $\delta$, producing at its output a subsequence of the input sequence. Consider a sequence $X$ passed through $t$ ($t$ is fixed) such deletion
channels as shown in
Fig.~\ref{fig:tdeletion}. We call this the $t$-trace deletion channel model. We ask four main questions:
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{1deletion.pdf}
\caption{The single-trace deletion channel model.}
\label{fig:1deletion}
\end{center}
\end{figure}
\begin{enumerate}[leftmargin = *]
\item \textbf{Sequencewise maximum-likelihood with one trace:} For $t=1$ (also called \textit{single-trace deletion channel}, see Fig.~\ref{fig:1deletion}), what is the maximum-likelihood estimate of $X$ having observed $Y=y$, i.e., a solution to $\argmax\limits_{x\in \{0,1\}^n}\ \Pr(Y=y|X=x)$.
\vspace{8pt}
\item \textbf{Sequencewise maximum-likelihood with multiple traces:} For a fixed $t$, with $t>1$, what is the maximum-likelihood estimate of $X$ having observed $Y^{1}=y^1,Y^{2}=y^2,...,Y^{t}=y^t $, i.e., $$\argmax\limits_{x\in \{0,1\}^n}\ \Pr(Y^{1}=y^1,Y^{2}=y^2,...,Y^{t}=y^t|X=x).$$
\vspace{-12pt}
\item \textbf{Symbolwise MAP with one trace:} For $t=1$ and $X_i \sim\ \text{ind. Ber}(p_i)$ in Fig.~\ref{fig:1deletion}, what are the posterior distributions of $X_i$ given the trace $Y=y$, i.e., compute $\Pr(X_i=\alpha|Y=y)$.
\item \textbf{Symbolwise MAP with multiple traces:} For a fixed $t$, with $t>1$ and $X_i \sim\ \text{i.i.d. Ber}(0.5)$ in Fig.~\ref{fig:tdeletion}, what are the posterior distributions of $X_i$ given all traces $Y^{1}=y^1, Y^{2}=y^2,...,Y^{t}=y^t$, i.e., compute $\Pr(X_i=\alpha|Y^{1}=y^1, Y^{2}=y^2,...,Y^{t}=y^t)$.
\end{enumerate}
\vspace{5mm}
\noindent We make a few notes.
\begin{itemize}[leftmargin = *]
\item For a channel with memory such as the deletion channel, the symbolwise MAP/ML estimate and sequencewise MAP/ML estimate are not equivalent. For example, consider $t=1$, $n =6$ in Fig.~\ref{fig:1deletion} and say we observe the trace $Y = 1010$. The symbolwise MAP estimate with uniform priors for this case can be computed to be $\hat X_{smap} = 100110$ whereas the sequencewise ML estimate is $\hat X_{ml} = 101010$.
\item An answer to 3) above doesn't lead to a natural solution for 4) which is also due to deletion channels possessing memory. In particular, for a memoryless channel, we have $Y^{j}_i - X_i - Y^{k}_i$ and hence $\Pr(X_i=\alpha|Y^{j}, Y^{k}) \propto \Pr(Y^{j}_i, Y^{k}_i|X_i=\alpha) = \Pr(Y^{j}_i|X_i=\alpha) \Pr(Y^{k}_i|X_i=\alpha)\propto \Pr(X_i=\alpha| Y^{j}) \Pr(X_i=\alpha | Y^{k})$; so one could first obtain the posterior probabilities from each independent observation and combine them after. However, this is not the case for deletion channels since the markov chain $Y^{j}_i - X_i - Y^{k}_i$ no longer holds. As a result, one first needs to ``align'' all the observations in order to compute the likelihoods.
\item Solving 2) and 4) naturally leads to two different algorithms for average-case trace reconstruction -- one that selects the most likely sequence $X$ and the other that selects the most likely value for each symbol $X_i$. However, the problem formulations in 3) and 4) ask a question complementary to that of trace reconstruction: given a fixed (possibly a few) number of traces, what is our ``best'' guess of $X$? The two problems 2) and 4) have different quantification of the word ``best''. Unlike trace reconstruction, we are not concerned with perfect reconstruction (since perfect reconstruction may not be possible with just a few traces). We also note that error rate guarantees for our algorithms (not a part of this work) would naturally lead to upper bounds for trace reconstruction.
\item The challenges associated with solving 1) and 2) and solving 3) and 4) are very different. On the one hand, solving 1) and 2) amounts to discovering alternate, equivalent or approximate formulations for the seemingly difficult discrete optimization problems. On the other hand, the challenge with 3) and 4) involves the design of efficient algorithms that are capable of exactly computing/approximating the symbolwise posterior probabilities, for which ``closed form'' expressions can be derived.
\end{itemize}
\vspace{5mm}
\noindent \textbf{Contributions.}
Our main contributions are as follows.
\begin{itemize}[leftmargin=*]
\item We introduce mathematical tools and constructs to visualize and analyze single-trace and $t$-trace deletion error events (see Section~\ref{sec:notation}).
\item For the single-trace deletion channel, we establish an equivalence between finding the optimal ML decoder and a continuous optimization problem we introduce (see Section~\ref{sec:1deletion_ML}). This equivalence allows for the use of existing techniques for continuous optimization to be employed for a seemingly difficult discrete optimization problem. This continuous optimization problem also turns out to be a signomial optimization. Furthermore we also provide a polynomial time trace reconstruction heuristic with multiple traces that exploits this formulation.
\item In Section~\ref{sec:1deletion}, we prove the following:
\begin{theorem}
For the single-trace deletion channel model with priors $X_i \sim \text{ind. Ber}(p_i)$ and observed trace $Y=y$, the symbolwise posterior probabilities $\Pr(X_i=1|Y=y)\ \forall\ i$ can be computed in $O(n^2)$ time complexity.
\end{theorem}
\item In Section~\ref{sec:exactsmap}, we prove the following:
\begin{theorem}
For the $t$-trace deletion channel model with priors $X_i \sim \text{i.i.d. Ber}(0.5)$ and observed traces $Y^{1}=y^1,...,Y^{t}=y^t$, the symbolwise posterior probabilities $\Pr(X_i = 1|Y^{1}=y^1,...,Y^{t}=y^t)\ \forall\ i$ can be computed in $O(2^t n^{t+2})$ time complexity.\\
\end{theorem}
\end{itemize}
\noindent \textbf{Tools and techniques.} In terms of theoretical tools, the series of books by Lothaire (\cite{lothaire1997combinatorics,lothaire2002algebraic,lothaire2005applied}) extensively use algebraic tools for problems in the combinatorics of sequences (or \textit{words}), and our work is inspired by such techniques. We borrow some notation and leverage a few of their results in our work. \\
\noindent \textbf{Biological motivation.} Trace reconstruction in itself was motivated, in part, by problems in DNA sequence reconstruction. One such problem was to infer the DNA sequence of a common ancestor from the samples of its descendants. Our problem definition, that considers a fixed value of $t$, would fit naturally in a scenario with a fixed number of descendants where perfect reconstruction may not be possible.
Our motivation for considering this problem also comes from a recent DNA sequencing technology called \textit{nanopore sequencing}. The $t$-trace deletion channel model is a simplistic model to approximately capture the process of a DNA sequence passed through a nanopore sequencer\footnote{As seen in
\cite{Mao2017},\cite{MDK17} there are more complicated effects of
the nanopore reader not captured in this simple
representation.}. \\
\noindent \textbf{More related work.}
Our work falls under the general umbrella of sequence reconstruction over deletion channels (also see Levenshtein's work \cite{levenshtein2001efficient}), where we offer, to the best of our knowledge, the first non-trivial results on maximum likelihood and maximum aposteriori estimates for the single and multiple deletion channel. As mentioned earlier, the complementary problem of trace reconstruction falls closest to this work.
The deletion channel by itself is known to be notoriously difficult to analyse. As stated earlier, the capacity of a single deletion channel is still unknown (\cite{diggavi2007capacity,diggavi2006information,diggavi2001transmission}); as are optimal coding schemes. Prior works have looked at the design of codes for deletion channels (\cite{ratzer2005marker,ratzer2000codes,thomas2017polar}); these works consider use of a codebook (we do not). Statistical estimation over deletion channels is a difficult problem to analyze due its highly combinatorial nature. To the best of our knowledge, as yet there are no efficient estimation algorithms over deletion channels with statistical guarantees.
Very recently, a variant of the trace reconstruction problem called \textit{coded trace reconstruction} has been proposed, motivated by portable DNA-based data storage systems using DNA nanopores (see \cite{abroshan2019coding}, \cite{cheraghchi2019coded}, \cite{brakensiek2019coded}) and we believe that the ideas in this work may prove useful in such a setting.
There are other works on sequence assembly (see for example, \cite{Li09fastand}, \cite{Shomorony2016}),
where multiple short reads (from different segments of a sequence) are used to reconstruct the bigger
sequence. This work differs from sequence assembly since we are interested in inferring the entire length sequence and not just small segments of it (which are then ``stitched'' together in sequence assembly).\\
\noindent \textbf{Paper Organization.} Section~\ref{sec:notation} introduces our notation and visualization tools for the single and $t$-trace channel error events; Section~\ref{sec:1deletion_ML} provides a result concerning questions 1) and 2) wherein we prove the equivalence of ML decoding in question 1) to solving a continuous optimization problem; Section~\ref{sec:1deletion} answers question 3) for the single-trace channel;
Section~\ref{sec:exactsmap}) answers question 4) for the $t$-deletion channel; Section~\ref{sec:Numerics} gives numerical evaluations; and Section~\ref{sec:conclusions} concludes the paper.
\section{Symbolwise MAP for the $t$-trace deletion channel}
\label{sec:exactsmap}
In this section, we put to use the ideas and constructs introduced in section~\ref{sec:notation} to exactly compute the symbolwise posterior probabilities given $t$-traces, which in turn gives a symbolwise MAP estimate with uniform input priors (motivated by average case trace reconstruction). With this formulation the symbolwise MAP with uniform priors can be seen as a minimizer of the symbol error rate in the context of average case trace reconstruction. In Appendix~\ref{app:remnant_postprob}, we also provide a method to compute the symbolwise posterior probabilities for the remnant channel -- we encourage the reader to use this appendix as a warm-up. For the $t$-trace deletion channel, similar expressions arise due to the channel equivalence result of Theorem~\ref{thm:channel_equiv}.
Let $\mathcal A = \{0,1\}$, and assume that $X\sim $ Uniform $\mathcal A^n$. Our goal is to compute the symbolwise posterior probabilities $\Pr(X_i=1|Y^1=y^1,...,Y^t=y^t)$, where $Y^j$ is the $j^{th}$ trace. Our proposed algorithm is provided in Alg.~\ref{alg:exact_smap} and estimates the symbolwise MAP (with uniform priors). We can directly leverage Alg.~\ref{alg:exact_smap} to reconstruct the input as follows: for each index $i$, compute $\Pr(X_i =1|Y^1=y^1,...,Y^t=y^t)$ and decide
\begin{align*}
\hat{X}_i = \begin{cases}
1,\quad &$if$\ \Pr(X_i =1|Y^1=y^1,...,Y^t=y^t)\geq 0.5 \\
0, \quad &$otherwise$.
\end{cases}
\end{align*}
Through the rest of this section, we show how to compute $\Pr(X_i =1|Y^1=y^1,...,Y^t=y^t)$\footnote{Symbolwise MAP with non-uniform priors is part of on-going work.} in two steps:
\begin{itemize}
\item We first give an expression for $\Pr(X_i =1|Y^1=y^1,...,Y^t=y^t)$ which sums over potentially an exponential number of terms.
\item We then show that this summation can be computed in polynomial time (polynomial in the blocklength $n$).
\end{itemize}
\noindent \textbf{Step 1: An expression for $\Pr(X_i =1|Y^1=y^1,...,Y^t=y^t)$.}
\begin{theorem}
\label{thm:exactSMAP_posteriorprob}
Assume $X\sim $ Uniform $\mathcal A^n$ or equivalently $X_i \sim\ \text{Ber}(0.5)$. The posterior probability of the $i^{th}$ bit given the $t$ traces can be expressed as
\begin{align*}
\Pr(X_i =1|Y^1=y^1,...,Y^t&=y^t) \\ =
& \Bigg[ \sum_{k=0}^n 2^{n-k-1}{n-1 \choose k} \sum_{w||w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \\
&+ \sum_{k=0}^n \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \Bigg] \Big/ \\
&\Bigg[ \sum_{k=0}^n 2^{n-k} {n \choose k} \sum_{\substack{w| |w|=k}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \Bigg]. \numberthis
\label{eq:posterior_prob}
\end{align*}
\end{theorem}
Note that the summation index, $w| |w|{=}k$ is over all sequences $w$ of length $k$; this is an alternate expression for $w|w{\in}\mathcal A^k$. We follow this convention throughout the rest of the paper.
\begin{proof}
\begin{align*}
\Pr(X_i=1&|Y^1=y^1,...,Y^t=y^t) = \sum_{\substack{x||x|=n,\\x_i=1}} \Pr(X=x|Y^1=y^1,...,Y^t=y^t)\\
&\overset{(a)}{=} \frac{1}{2^n \Pr(Y^1=y^1,...,Y^t=y^t)} \sum_{\substack{x||x|=n,\\x_i=1}} \Pr(Y^1=y^1,...,Y^t=y^t|X=x)\\
&\overset{(b)}{=} \frac{1}{2^n \Pr(Y^1=y^1,...,Y^t=y^t)} \sum_{\substack{x||x|=n,\\x_i=1}} \prod_{j=1}^t\Pr(Y^{j}=y^j|X=x),
\end{align*}
where $(a)$ uses Bayes' principle and $(b)$ is because each deletion channel acts independently. Recall that for a deletion channel with deletion probability $\delta$, $\Pr(Y=y|X=x)={x \choose y}\delta^{|x|-|y|}(1-\delta)^{|y|}$. Also, using the fact that $\Pr(Y^1=y^1,...,Y^t=y^t)=\sum\limits_{\substack{x||x|=n}}\Pr(x) \Pr(Y^1=y^1,...,Y^t=y^t|X=x)$ we have,
\begin{align*}
\Pr(X_i=1|Y^1=y^1,...,Y^t=y^t)= \frac{\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose y^1}...{x \choose y^t}}{\sum\limits_{\substack{x||x|=n}} {x \choose y^1}...{x \choose y^t}}. \numberthis \label{eq:thmexactsmap_proofterm0}
\end{align*}
We first simplify the numerator $\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose y^1}...{x \choose y^t}$; the denominator can be simplified using the same approach. Now,
\begin{align*}
\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose y^1}...{x \choose y^t} &\overset{(a)}{=} \sum_{\substack{x||x|=n,\\x_i=1}} \sum_{w\in \{0,1\}^*} {x \choose w} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \\
&=\sum_{w\in \mathcal A^*} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \sum_{\substack{x||x|=n,\\x_i=1}}{x \choose w}\\
&\overset{(b)}{=}\sum_{w\in \mathcal A^*} 2^{n-|w|} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \left(\frac{1}{2}{n-1 \choose |w|}+\sum_{j|w_j=1}{i-1 \choose j-1}{n-i \choose |w|-j}\right)
\end{align*}
where $(a)$ is due to Lemma~\ref{lemma:bin_inf_relation} and $(b)$ due to Lemma~\ref{lemma:smapsum} (both introduced in \cite{Srini2018}); see Appendix~\ref{app:bin_inf_lemma} and Appendix~\ref{app:smapsum} for the statement and proof.
\noindent Therefore we have,
\begin{align*}
\sum\limits_{\substack{x||x|=n,\\x_i=1}} {x \choose y^1}...{x \choose y^t}
&\overset{(a)}{=} \sum_{k=0}^{\infty} 2^{n-k-1}{n-1 \choose k} \sum_{w| |w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \\
&\hspace{1cm}+ \sum_{k=0}^{\infty} \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \\
&\overset{(b)}{=} \sum_{k=0}^n 2^{n-k-1}{n-1 \choose k} \sum_{w| |w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \\
&\hspace{1cm}+ \sum_{k=0}^n \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle, \numberthis
\label{eq:thmexactsmap_proofterm1}
\end{align*}
where in $(a)$ we first fix $|w|$ and then sum over all $w$ of the given length and $(b)$ holds because the combinatorial terms are $0$ when $k>n$.
A similar analysis gives \begin{align*}
\sum\limits_{x| |x|=n} &{x \choose y^1}...{x \choose y^t} = \sum_{k=0}^n 2^{n-k} {n \choose k} \sum_{\substack{w| |w|=k}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle.\numberthis
\label{eq:thmexactsmap_proofterm2}
\end{align*}
Plugging \eqref{eq:thmexactsmap_proofterm1} and \eqref{eq:thmexactsmap_proofterm2} in \eqref{eq:thmexactsmap_proofterm0}, we get the expression in Theorem~\ref{thm:exactSMAP_posteriorprob},
\begin{align*}
\Pr(X_i =1|Y^1=y^1,...,&Y^t=y^t) \\ =
& \Bigg[ \sum_{k=0}^n 2^{n-k-1}{n-1 \choose k} \sum_{w||w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \\
&+ \sum_{k=0}^n \sum_{j=1}^k 2^{n-k}{i-1 \choose j-1}{n-i \choose k-j} \sum_{\substack{w| |w|=k,\\w_j=1}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \Bigg] \Big/ \\
&\Bigg[ \sum_{k=0}^n 2^{n-k} {n \choose k} \sum_{\substack{w| |w|=k}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \Bigg].
\end{align*}
\end{proof}
\noindent \textbf{Step 2: Dynamic program to compute $\sum\limits_{w| |w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle$ and $\sum\limits_{\substack{w| |w|=k,\\w_j=1}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle$.}
Note that the number of sequences $w$ such that $|w|=k$ is $O(2^k)$ so a naive evaluation is exponential in the blocklength $n$. We can, however, exploit the edit graph to come up with a dynamic program resulting in an algorithm which is polynomial in $n$.
Recall that in the edit graph, $\langle y^1 \uparrow ... \uparrow y^t,w \rangle$ is equal to the number of distinct paths from the origin $(0,...,0)$ to the destination $(|y^1|,...,|y^t|)$ and which correspond to $w$. Hence,
\begin{enumerate}[wide=0pt]
\item[(a)] $\sum\limits_{w| |w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle$ is the number of distinct paths of length $k$ from origin to destination and,
\item[(b)] $\sum\limits_{\substack{w| |w|=k,\\w_j=1}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle$ is the number of such paths of length $k$ such that the $j^{th}$ edge of the path corresponds to a `1'.
\end{enumerate}
\noindent With this interpretation, the dynamic program for (a) follows naturally -- the number of $k$-length paths from the origin to any vertex is the sum of the number of $(k{-}1)$-length paths from the origin to all incoming neighbors of the vertex. To make this formal, associate a polynomial (in $\lambda$) for each vertex, such that the coefficient of $\lambda^k$
is equal to the number of paths of length $k$ from the origin to $v$: we call it the "forward-potential" polynomial $p^{for}_v(\lambda)$ for vertex $v$, the coefficient of $\lambda^k$ as earlier is denoted by $\langle p^{for}_v(\lambda),\lambda^k \rangle $. The dynamic program to compute $p^{for}_v(\lambda)$ for all $v$ can be expressed as:
\begin{equation}
p^{for}_v(\lambda) = \sum_{u|u\rightarrow v} \lambda p^{for}_u(\lambda).
\end{equation}
With this definition, we have $$\sum\limits_{w| |w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle =\langle p^{for}_{destination}(\lambda),\lambda^k \rangle.$$ In the example in Fig.~\ref{fig:editgraph_smap1}, one could do the following: order the vertices $(0,0)$ to $(3,3)$ lexicographically and then compute $p^{for}_v(\lambda)$ in the same order. Because of the directed grid nature of the edit graph, every vertex has incoming neighbors which are lexicographically ahead of itself. Also we initialize $p^{for}_{(0,0)}(\lambda)=1$. For the example in Fig.~\ref{fig:editgraph_smap1}, the forward-potentials are shown in Fig.~\ref{fig:editgraph_smap2}. The complexity of this dynamic program is $O(2^tn^{t+1})$ as it goes over $O(n^t)$ vertices and for each vertex it sums $O(2^t)$ polynomials, each of degree $O(n)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap2.pdf}
\caption{The forward-potential $p^{for}_v(\lambda)$ at each vertex.}
\label{fig:editgraph_smap2}
\end{figure}
We compute (b) as follows: pick an edge $(u{\rightarrow}v)$ which corresponds to `1', count the number of $(j{-}1)$-length paths from origin to $u$ and multiply it with the number of $(k{-}j)$-length paths from $v$ to the destination -- this is exactly the number of paths of length $k$ such that its $j^{th}$ edge is $(u{\rightarrow}v)$. Summing this term for all such edges which correspond to 1 gives us the term in (b). Note that we have already computed the number of $k$-length paths ($\forall k$) from origin to every vertex in $p^{for}_v(\lambda)$ . We can similarly compute the number of $k$-length paths ($\forall k$) from every vertex to the destination as $p^{rev}_v(\lambda)$ -- the "reverse potential" polynomial. The dynamic program for $p^{rev}_v(\lambda)$ is:
\begin{equation}
p^{rev}_v(\lambda) = \sum_{u|v\rightarrow u} \lambda p^{rev}_u(\lambda),
\end{equation}
with $p^{rev}_{destination}(\lambda)=1$. The reverse potentials for the example in Fig.~\ref{fig:editgraph_smap1} is shown in Fig.~\ref{fig:editgraph_smap3}. Like in the case of forward potential, we first order the vertices reverse lexicographically and then invoke the dynamic program above sequentially to compute the reverse potential polynomial at each vertex.
\begin{algorithm}[t!]
\caption{Computing the forward-potentials $p^{for}_u(\lambda)$ }\label{alg:forward_pot}
\begin{algorithmic}[1]
\item Input: Edit graph {$\mathcal G(y^1,...,y^t)$}\\ Outputs: $p^{for}_v(\lambda)\ \forall\ v$
\State Order the vertices from $(0,0,...,0)$ to $(|y^1|,|y^2|,...,|y^t|)$ lexicogaphically; let the ordered list be $\mathcal V$
\State Initialise $p^{for}_{(0,...,0)}(\lambda)\gets 1$
\For{$v\ \in\ \mathcal V$}
\State \textbf{assign} $p^{for}_v(\lambda)\gets \sum_{u|u\rightarrow v} \lambda p^{for}_u(\lambda)$
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t!]
\caption{Computing the reverse-potentials $p^{rev}_u(\lambda)$ }\label{alg:reverse_pot}
\begin{algorithmic}[1]
\item Input: Edit graph {$\mathcal G(y^1,...,y^t)$}\\ Outputs: $p^{rev}_v(\lambda)\ \forall\ v$
\State Order the vertices from $(|y^1|,|y^2|,...,|y^t|)$ to $(0,0,...,0)$ reverse lexicogaphically; let the ordered list be $\mathcal V$
\State Initialise $p^{rev}_{(|y^1|,|y^2|,...,|y^t|)}(\lambda)\gets 1$
\For{$v\ \in\ \mathcal V$}
\State \textbf{assign} $p^{rev}_v(\lambda) \gets \sum_{u|v\rightarrow u} \lambda p^{rev}_u(\lambda)$
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap3.pdf}
\caption{The reverse-potential $p^{rev}_v(\lambda)$ at each vertex.}
\label{fig:editgraph_smap3}
\end{figure}
With this, the term in (b) can be expressed as:
\begin{align*}
\sum\limits_{\substack{w| |w|=k,\\w_j=1}}\langle y^1 \uparrow ... \uparrow y^t,w \rangle =\hspace{-2mm} \sum_{\substack{(u,v)|\\s(u\rightarrow v)=1}} \hspace{-2mm}\langle p^{for}_u(\lambda),\lambda^{j-1} \rangle \langle p^{rev}_v(\lambda),\lambda^{k-j} \rangle.
\end{align*}
Alg.~\ref{alg:exact_smap} now summarizes the computation of the posterior probabilities. This algorithm iterates over all the edges (we have $O((2n)^t)$ of these), and also $k,j$ ($O(n)$ each). The time complexity of Alg.~\ref{alg:exact_smap} hence is $O(2^tn^{t+2})$.
\begin{algorithm}[t!]
\caption{Symbolwise MAP with $t$ traces}
\label{alg:exact_smap}
\begin{algorithmic}[1]
\item Input: Traces {$Y^1=y^1,...,Y^t=y^t$, input length $n$}\\ Output: $\hat X = \hat X_1\hat X_2...\hat X_n
\State Construct edit graph $\mathcal G(y^1,...,y^t)$
\State Use Alg.~\ref{alg:forward_pot} and Alg.~\ref{alg:reverse_pot} on $\mathcal G(y^1,...,y^t)$ to calculate $p^{for}_v(\lambda)$ and $p^{rev}_v(\lambda)$ $\forall\ v$
\For{$k \in\ [0:n]$}
\State \textbf{assign} $\sum\limits_{w| |w|=k} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \gets \langle p^{for}_{destination}(\lambda),\lambda^k \rangle.$
\For{each $j \in\ [1:n]$}
\State Initialize $temp \leftarrow $ 0
\For{each edge $u\rightarrow v\ \in\ \mathcal G$}
\If{$s(u{\rightarrow} v)=$ `1'}
\State $temp\ += \langle p^{for}_u(\lambda),\lambda^{j-1} \rangle \langle p^{rev}_v(\lambda),\lambda^{k-j} \rangle$
\EndIf
\EndFor
\State \textbf{assign} $\sum\limits_{\substack{w| |w|=k,\\w_j=1}} \langle y^1 \uparrow ... \uparrow y^t,w \rangle \gets temp$
\EndFor
\EndFor
\For{$i \in\ [1:n]$}
\State Use \eqref{eq:posterior_prob} to compute $\Pr(X_i=1|Y^1=y^1,...,Y^t=y^t)$
\State $\hat X_i \leftarrow 1$ if $\Pr(X_i=1|Y^1=y^1,...,Y^t=y^t) > 0.5$ and $\hat X_i \leftarrow 0$ otherwise
\EndFor
\State \textbf{return} $\hat X_1\hat X_2...\hat X_n$
\end{algorithmic}
\end{algorithm}
\section{Notation and Tools}
\label{sec:notations}
\noindent \textbf{Basic notations:} In this work, we borrow a majority of the notation and tools from \cite{lothaire1997combinatorics} which deals with non-commutative algebra. We restate the definitions here for convenience. Calligraphic letters refer to sets, capitalized letters
correspond to random variables and bold letters are used for functions. Let $\mathcal{A}$ be the set of all
symbols. Throughout this work, we will focus on the case where $\mathcal{A}=\{0,1\}$,
though our methods extend to arbitrarily large sets of finite size. Define
$\mathcal{A}^n$ to be the set of all $n$-length sequences and $\mathcal{A}^*$ to be the set of all finite length sequences with
symbols in $\mathcal{A}$. For a sequence $f$, $|f|$ denotes the length
of $f$. For integers $i,j$, we define $[i:j] \triangleq \{i,i+1,...,j\}$ if $j\geq i$ and $[i:j] \triangleq \emptyset$ otherwise. Also define $[i] \triangleq [1:i]$.\\
\noindent \textbf{Binomial coefficient:} Given sequences $f$ and $g$ in $\mathcal{A}^*$, the number of
subsequence patterns of $f$ that are equal to $g$ is called the
\textit{binomial coefficient} of $g$ in $f$ and is denoted by $f
\choose g$. For example, ${'apple' \choose 'ape'} = 2$ since $'ape'$ can be obtained from two (overlapping) subsequences of $'apple'$.
When the alphabet $\mathcal{A}$ is of cardinality 1, ${f
\choose g} = {|f| \choose |g|}$, the classical binomial coefficient
with their respective lengths as the parameters. This definition hence
could be thought of as a generalization of the classical binomial
coefficients. We will denote by $e$ the sequence of length 0, and {define ${f
\choose e} \triangleq 1\ \forall\ f\ \in\ \mathcal{A}^*$.} We also define the classical binomial coefficient ${a \choose b}\triangleq 0,$ whenever $b>a$ or $b<0$ for ease of use.
The binomial coefficient is an integral aspect of this work and for analyzing error events in deletion channels because the input-output relation for a deletion channel (with deletion probability $\delta$, input $X$ and output $Y$) can be expressed as
\begin{equation}
\label{eq:in-out_relation}
\Pr(Y=y|X=x) = {x \choose y} \delta^{|x|-|y|} (1-\delta)^{|y|}.
\end{equation}
The proof is straight-forward -- the number of distinct error events that give rise to $y$ from $x$ is exactly the number of subsequences of $x$ which are equal to $y$. Each of these error events has a probability $\delta^{|x|-|y|} (1-\delta)^{|y|}$, wherein the exponent of $\delta$ corresponds to the deleted symbols and the exponent of $1-\delta$ to the undeleted symbols.
Given the definition of the binomial coefficient, the maximum-likelihood (ML) estimate over a deletion channel with observed output $Y$ can be cast in the following form:
\begin{align*}
\argmax_{x \in \mathcal C} {x \choose Y},\numberthis
\label{eq:ML_deletion}
\end{align*}
where $\mathcal C$ is the chosen codebook. In the case of multiple deletion channels with observed traces $Y^{(1)},...,Y^{(t)}$, the ML formulation is similar:
\begin{align*}
\argmax_{x \in \mathcal C} \prod_{j=1}^{t} {x \choose Y^{(j)}}.\numberthis
\label{eq:ML_multiple_deletion}
\end{align*}
As yet, there is no known efficient way to come up with a solution for the above two formulations, even for \eqref{eq:ML_deletion} with $\mathcal C = \{0,1\}^n$ (see \cite{mitzenmacher2009survey}). In this work, we attempt to take a step in this direction by showing that a continuous relaxation of \eqref{eq:ML_deletion} is equivalent to \eqref{eq:ML_deletion}. However, an efficient algorithm to solve the optimization \eqref{eq:ML_deletion} remains open. In the context of trace reconstruction, the ultimate pursuit would be an algorithm for \eqref{eq:ML_multiple_deletion} with $\mathcal C = \{0,1\}^n$ and error analysis thereof.
\noindent We now describe a function which can be thought of as a real-valued extension of the binomial coefficient. This function is used in sections~\ref{sec:1deletion_ML} and \ref{sec:1deletion}. Consider the function $\mathbf F(\cdot)$ defined as:
\begin{definition}
\label{def:f}
\begin{align*}
&\mathbf F: \mathbb{R}^k \times \{0,1\}^l \rightarrow \mathbb{R},\\
\mathbf F(q, v)\triangleq &\begin{cases}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [k],\\|\mathcal S|=l}} \quad \prod\limits_{i=1}^l q_{\mathcal S_i}^{v_i} (1-q_{\mathcal S_i})^{1-v_i} & 1 \leq l\leq k \\
1 & 0=l\leq k \\
0 & \text{else}.
\end{cases}
\end{align*}
\noindent An alternate definition is as follows: consider a random vector $Z\in\{0,1\}^k$ such that $Z_i\sim$ ind. Ber$(q_i)$, let $q$ be the vector of probabilities of length $k$, $\mathcal S\subseteq [k]$ a subset of the indices of size $l$, and ${v}$ a binary vector of length $l$. Then, $$\mathbf F(q,v)=\sum_{\substack{\mathcal S||\mathcal S|=l}} \Pr(Z_{\mathcal S}= v).$$ Note that if $q$ is a vector of $0$'s and $1$'s, then $\mathbf F( q, v)={q \choose v}$, the binomial coefficient of $v$ in $q$. Thus, $\mathbf F(\cdot)$ could be interpreted as an extension of the binomial coefficient where one of the parameters can take values in $[0,1]^n$ instead of $\{0,1\}^n$.
\end{definition}
Though at first sight, $\mathbf F(q,v)$ sums over an exponential number of subsets, a dynamic programming approach can be used to compute it in $O(|v|^2)$ time complexity. The dynamic program is described in section~\ref{sec:1deletion}.
\iffalse
\noindent \textbf{Edit distance:} The edit distance $d_e(f,g)$ measures similarity between two sequences of
possibly different lengths \cite{Navarro2001}. $d_e(f,g)$ is the minimum number of operations
needed to transform $f$ to $g$, where the permitted operations are
insertion, deletion or substitution of a symbol. In this work, we
quantify the performance of algorithms in Section \ref{sec:Numerics}
using the edit distance metric.
\fi
The following definitions and ideas are relevant only for \ref{sec:exactsmap} and can be omitted for the sections on single deletion channel. Before getting to the mathematical definitions, we first state a result of ours that aids in thinking about error events in multiple deletion channels.
\subsection{An alternate interpretation of multiple deletion channel model}
The events occurring in the $t$-deletion channel model can be categorized into two groups:
\begin{enumerate}
\item an input symbol is deleted in \textit{all} the $t$-traces,
\item an input symbol is reflected in at least one of the traces.
\end{enumerate}
The error events of the first kind are in some sense ``not correctable" or even ``detectable" since it is impossible to tell what and where the deleted symbol could have been (although the probabilities need not be uniform). The events of the second kind, however, can still be detected although they could likely be confused with a large set of similar events. This thought process gives rise to a natural decomposition of the $t$-deletion channel model into a cascade of two channels: the first one being a deletion channel which captures error events of the first kind and the second one is what we call the \textit{remnant channel} which captures events of the second kind (see Fig.~\ref{fig:channel_equiv}). More precisely, the remnant channel is defined as follows:
\begin{definition}
\textit{Remnant channel:} an input symbol to the remnant channel is reflected in $k>0$ given outputs and deleted in the rest with a probability $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
\end{definition}
It is easy to note that probability of the union of all possible events here is $\sum_{k=1}^t {t \choose k} \frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}=1$, validating our definition.
\begin{restatable}{theorem}{channelequiv}
\label{thm:channel_equiv}
The $t$-deletion channel model and the cascade of the deletion channel with remnant channel shown in Fig.~\ref{fig:channel_equiv} are probabilistically equivalent, i.e.,
$$\Pr({Y}^{(1)},{Y}^{(2)},...,{Y}^{(t)}|X = x) = \Pr(\tilde{Y}^{(1)},\tilde{Y}^{(2)},...,\tilde{Y}^{(t)}|X = x).$$
\end{restatable}
The formal proof of the theorem requires few more definitions and is relegated to the appendix.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.4]{channel_equiv.pdf}
\caption{A channel equivalence result: the $t$-deletion channel model in (a) is probabilistically equivalent to the the cascade of a deletion channel with the \textit{remnant channel} ($ \mathcal C_2$) in (b).}
\label{fig:channel_equiv}
\end{center}
\end{figure}
\noindent\textbf{Edit graph} (as defined in \cite{Gusfield1997}): We now define a graph construct which is closely related to the error events in the remnant channel. We start with a simple case and generalize subsequently. Define an \textit{edit graph} given two sequences $f$ and $g$, where every path connecting the ``origin'' to the ``destination'' on the edit graph yields a supersequence $h$ of $f,g$, where $h$ is ``covered'' by $f,g$ -- i.e., each symbol of $h$ comes from either $f$ or $g$ or
both. In other words, given that $f$ and $g$ are the outputs of the remnant channel (with two outputs), each path from the origin of the edit graph to the destination corresponds to a possible input $h$ to the remnant channel and to an error event which resulted in $f$ and $g$ with input $h$.
For $f$ and $g$ in $\mathcal
A^*$, we form a graph $\mathcal{G}(f,g)$ with $(|f|+1)(|g|+1)$
vertices each labelled with a distinct pair $(i,j), 0\leq i\leq
|f|,\ 0\leq j\leq |g|$. A directed edge
$(i_1,j_1)\rightarrow(i_2,j_2)$ exists iff at least one of the following
holds:
\begin{enumerate}
\item$i_2-i_1=1$ and $j_1=j_2$, or
\item $j_2-j_1=1$ and $i_1=i_2$, or
\item $i_2-i_1=1$, $j_2-j_1=1$ and $f_{i_2}=g_{j_2}$,
\end{enumerate}
where $f_i$ is the $i^{th}$ symbol of the sequence $f$. The origin is the vertex $(0,0)$ and the destination $(|f|,|g|)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap1.pdf}
\caption{ Edit graph for sequences $f=$ `001' and $g=$ `101'. An easy way to think about this is to write down $f$ vertically with each symbol aligned to a vertical set of edges and $g$ horizontally likewise. A diagonal edge in a small square exists if the corresponding $f$ and $g$ symbols are equal. The thick red edges form a path from the origin to the destination; this path corresponds to the sequence `0101' -- append the corresponding symbol at the left of an edge if it's vertical or diagonal, otherwise append the symbol at the top. It is also easy to verify that 0101 is a supersequence of both $f$ and $g$, and could be obtained as a covering of $f$ and $g$; the path itself gives one such covering. This covering also corresponds to an error event in the remnant channel which would result in outputs $f$ and $g$ with input $h=$ `0101' -- more precisely, the error event is one in which the first symbol of $h$ is deleted only in $g$, the second symbol is deleted only in $f$ and the third and fourth symbols not deleted in either $f$ or $g$.}
\label{fig:editgraph_smap1}
\end{figure}
Let $p=((i_1,j_1),(i_2,j_2),...,(i_m,j_m))$ be a path in $\mathcal{G}(f,g)$. We define $ s(p)$ to be the sequence corresponding to the path. Intuitively, $s(p)$ is formed by appending symbols in the following
way: append the corresponding $f$ symbol for a vertical edge, $g$ symbol for horizontal edge, and $f$ or $g$ symbol for diagonal edge (see example Fig.~\ref{fig:editgraph_smap1}). Any path from $(0,0)$ to $(|f|,|g|)$ corresponds to a supersequence of $f$ and $g$ and which is ``covered" by $f$ and $g$. More formally, define $ s(p)\triangleq x_1x_2...x_{m-1}$ where
$$x_k =
\begin{cases}
f_{i_{k+1}} \quad\text{if }j_{k}=j_{k+1},\\
g_{j_{k+1}} \quad\text{if }i_{k}=i_{k+1},\\
f_{i_{k+1}} \quad\text{else.}
\end{cases}
$$
The construct of edit graph can be extended to more than 2 sequences with the same idea. For sequences $f_1,f_2,...,f_t$, construct a $t$-dimensional grid with a number of vertices $(|f_1|+1)(|f_2|+1)...(|f_t|+1)$ labeled from $(0,0,...,0)$ to $(|f_1|,|f_2|,...,|f_t|)$. A vertex $u=(i_1,i_2,...,i_t)$ is connected to $v=(j_1,j_2,...,j_t)$ (we say $u \rightarrow v$) iff both of the following conditions are met:
\begin{itemize}
\item $j_l=i_l$ or $j_l=i_l+1$ $\forall\ l\in [t]$, i.e., $(i_1,...,i_t)$ and $(j_1,...,j_t)$ are vertices of a particular unit cube. Only these type of vertices can share an edge in the grid graph.
\item Let $\mathcal T \subseteq [t]$ be the collection of indices where $j_l=i_l+1$. Then $f_{i_l}$ is same $\forall\ l \in \mathcal T$. For instance, if $\mathcal T=\{1,3,5\}$, then $f_{i_1}=f_{i_3}=f_{i_5}$.
\end{itemize}
Define the vertex $(0,...,0)$ to be the origin of this graph and the vertex $(|f_1|,...,|f_t|)$ to be the destination. If $|f_j|=O(n)\ \forall\ j$, this graph has a number of vertices $O(n^t)$ and a maximum number of edges $O((2n)^t)$ since each vertex has at most $2^t$ outgoing edges.\\
\noindent\textbf{Infiltration product} (introduced in \cite{lothaire1997combinatorics}):
The infiltration product has been extensively used in \cite{lothaire1997combinatorics}, as a tool in non-commutative algebra. Here we give an edit-graph interpretation of this tool. We also give a formal definition later in this section. Using the edit graph we can construct the set of possible
supersequences $\mathcal{S}(f,g)$ of $f,g$ which are covered by
it. Clearly multiple paths could yield the same supersequence and we
can count the number of distinct ways $\mathbf N(h;f,g)$ one can construct
the same supersequence $h$ from $f,g$.
We can informally define the
\emph{infiltration product $f\uparrow g$} of $f$ and $g$, as a polynomial with monomials the supersequences $h$ in
$\mathcal{S}(f,g)$ and coefficients $\langle f\uparrow g,h\rangle$ equal to $\mathbf N(h;f,g)$. In Fig.~\ref{fig:editgraph_smap1}, it is easy to verify that there is exactly one path corresponding to `101001' and hence $\langle 001\uparrow 101,101001 \rangle=1$ and similarly $\langle 001\uparrow 101,01001 \rangle=2$. One could find these coefficients for all relevant sequences and form the polynomial as described. More examples: Let $\mathcal{A}=\{a,b\}$, then
\begin{itemize}[wide=2pt]
\item $ab\uparrow ab=ab+2aab+2abb+4aabb+2abab$,
\item $ab\uparrow ba=aba+bab+abab+2abba+2baab+baba.$
\end{itemize}
The infiltration operation is commutative and associative, and infiltration of two sequences $f\uparrow g$ is a polynomial with variables of length (or \textit{degree}) at most $|f|+|g|$; see \cite{lothaire1997combinatorics}.
The definition of infiltration extends to two polynomials via distributivity (precisely defined in later), and consequently to multiple sequences as well. For multiple sequences, infiltration has the same edit graph interpretation: $\langle f_1\uparrow f_2 \uparrow...\uparrow f_t, w \rangle$ is the number of distinct ways of constructing $w$ as a supersequence of $f_1, f_2, ... ,f_t$ so that the construction covers $w$, i.e., construct the $t$-dimensional edit graph of $f_1, f_2, ... ,f_t$ and count the number of paths corresponding to $w$.
\subsection{Formal definition of infiltration product}
We now give a more formal definition of the infiltration product (see \cite{lothaire1997combinatorics} for the equivalence of the two definitions and a more rigorous treatment). A \textit{formal series} with indeterminates (or variables) in a set $\mathcal A$ and coefficients in a commutative ring $\mathcal R$, is a mapping of $\mathcal A^*$ onto $\mathcal R$. Recall that a commutative ring is a set which forms an abelian group under an \textit{addition} operation, is a monoid under a \textit{multiplication} operation which commutes, and the multiplication operation distributes over addition. Here we consider $\mathbb Z$, the set of integers as the commutative ring $\mathcal{R}$. A formal series is called a \textit{polynomial} if only a finite number of sequences are mapped to non-zero values, the rest of the sequences map to zero. Consider two polynomials $\sigma,\tau: \mathcal{A}^*\rightarrow \mathbb Z$. The value taken by a sequence $w\in \mathcal A^*$ on $\sigma$ (or the coefficient of $w$ in $\sigma$) is denoted by $\langle \sigma,w\rangle \in \mathbb R$. We also define binary addition ($\oplus$) and multiplication operations ($\times$) on the set of polynomials as follows:
\begin{align}
\langle \sigma\oplus \tau,w\rangle \triangleq \langle \sigma,w\rangle + \langle \tau,w \rangle \quad \forall w\in \mathcal A^*,\label{eq:polynomial_add}\\
\langle \sigma\times \tau,w\rangle \triangleq \sum_{\substack{f,g\in \mathcal A^*:\\ f.g=w}}\langle \sigma,f\rangle \langle \tau,g \rangle \quad \forall w\in \mathcal A^*.\label{eq:polynomial_prod}
\end{align}
We will use the usual symbols $+$ and $.$ in place of $\oplus$ and $\times$ in this work for convenience. The meaning of the operation would be clear depending on the operands. With these operations the set of polynomials form a non-commutative ring, and is denoted by $\mathbb Z\langle\mathcal A \rangle$, also called the free $\mathbb Z$-algebra on $\mathcal A$ in ring theory. Note that the addition and multiplication operations defined in \eqref{eq:polynomial_add} and \eqref{eq:polynomial_prod} are similar to the operations defined on commutative polynomials, except that the multiplication operation under the summation in \eqref{eq:polynomial_prod} ($f.g=w$) is actually concatenation and is non-commutative. The multiplication inside the summation in \eqref{eq:polynomial_prod} is multiplication in the real field and hence commutative. It is also easy to prove that the multiplication defined in \eqref{eq:polynomial_prod} distributes over addition defined in \eqref{eq:polynomial_add}. Thus, a polynomial in $\mathbb Z\langle\mathcal A \rangle$ can be represented as a sum of monomials in $\mathcal A^*$ each with an associated coefficient in $\mathbb Z$, i.e., $\sigma=\sum\limits_{w\in \mathcal A^*} \langle\sigma,w \rangle w$. Define the \textit{degree} of a polynomial to be equal to the length of a longest sequence with a non-zero coefficient in the polynomial and the \textit{number of terms} of a polynomial as the number of sequences with non-zero coefficients in the polynomial. Note that a degree $d$ polynomial could have a number of terms upto $2^{d+1}-1$.
With this, the \textit{infiltration product} (in general, for two polynomials) is defined as follows:
\begin{align}
\forall f\in \mathcal{A}^*,& \quad f\uparrow e = e\uparrow f=f.\nonumber \\
\forall f,g\in \mathcal{A}^*&,\quad \forall a,b\in \mathcal{A}, \nonumber\\
fa\uparrow gb=(f\uparrow gb)a&+(fa\uparrow g)b+\mathbbm{1}_{a=b}(f\uparrow g)a.\nonumber \\
\forall \sigma,\tau \in \mathbb{Z}\langle\mathcal{A} \rangle, \quad &\sigma\uparrow \tau=\sum_{f,g\in \mathcal{A}^*} \langle \sigma,f \rangle \langle \tau,g \rangle (f\uparrow g). \label{def:infiltforseq}
\end{align}
\textbf{Summary of definitions and ideas introduced in this section:}
\begin{itemize}
\item The binomial coefficient captures the likelihood of observations for deletion channels.
\item An extension of the binomial coefficient function where one of the parameters can take real values has been introduced; this function is pivotal for our results on the single-trace deletion channel.
\item For multiple deletion channels, the error events can be categorized into two groups -- one where an input symbol is deleted in all the traces, and second, the complement of this event. This categorization results in a natural decomposition of multiple deletion channel model into a cascade model involving the remnant channel.
\item The remnant channel disregards the error events where a symbol is deleted in all the traces.
\item The edit graph provides a way to visualize all the possible error events and input sequences to a remnant channel given its outputs.
\item The infiltration product serves the same purpose as the edit graph, but has an algebraic flavor and provides rigor for proofs and analyses. The edit graph, on the other hand, is more helpful in designing reconstruction algorithms over deletion channels.
\end{itemize}
\section{Notation and Tools}
\label{sec:notation}
\noindent \textbf{Basic notation:} We borrow some notation from \cite{lothaire1997combinatorics} which deals with non-commutative algebra; we restate them here for convenience. Calligraphic letters refer to sets, capitalized letters
correspond to random variables and bold letters are used for functions. Let $\mathcal{A}$ be the set of all
symbols. Throughout this work, we will focus on the case where $\mathcal{A}=\{0,1\}$,
though our methods extend to arbitrarily large sets of finite size. Define
$\mathcal{A}^n$ to be the set of all $n$-length sequences and $\mathcal{A}^*$ to be the set of all finite length sequences with
symbols in $\mathcal{A}$. For a sequence $f$, $|f|$ denotes the length
of $f$.
For integers $i,j$, we define $[i:j] \triangleq \{i,i+1,...,j\}$ if $j\geq i$ and $[i:j] \triangleq \emptyset$ otherwise. We also define $[i] \triangleq [1:i]$.
For a vector or sequence $x=(x_1,x_2,...,x_{i-1},x_i,x_{i+1},...,x_n)$, define $$x^{(i\rightarrow s)}\triangleq (x_1,x_2,...,x_{i-1},s,x_{i+1},...,x_n),$$ where the $i^{th}$ coordinate of $x$ is replaced by symbol $s$. \\
\noindent \textbf{Binomial coefficient (section 6.3 in \cite{lothaire1997combinatorics}):} Given sequences $f$ and $g$ in $\mathcal{A}^*$, the number of
subsequence patterns of $f$ that are equal to $g$ is called the
\textit{binomial coefficient} of $g$ in $f$ and is denoted by $f\choose g$. For example, ${'apple' \choose 'ape'} = 2$ since $'ape'$ can be obtained from two (overlapping) subsequences of $'apple'$. This quantity has also been referred to as the \textit{embedding number} by another line of work \cite{elzinga2008algorithms}. For two sequences of lengths $n$ and $m$, the binomial coefficient can be computed using a dynamic programming approach in $O(nm)$ (see \cite{elzinga2008algorithms} or Proposition 6.3.2 in \cite{lothaire1997combinatorics}).
When the alphabet $\mathcal{A}$ is of cardinality 1, ${f \choose g} = {|f| \choose |g|}$, the classical binomial coefficient with their respective lengths as the parameters. This definition hence
could be thought of as a generalization of the classical binomial
coefficients. We will denote by $e$ the sequence of length 0, and {define ${f
\choose e} \triangleq 1\ \forall\ f\ \in\ \mathcal{A}^*$.} We also define the classical binomial coefficient ${a \choose b}\triangleq 0,$ whenever $b>a$ or $b<0$ for ease of use.
The binomial coefficient forms the backbone for the probabilistic analysis of deletion channels since the input-output relation for a deletion channel (with deletion probability $\delta$, input $X$ and output $Y$) can be expressed as
\begin{equation}
\label{eq:in-out_relation}
\Pr(Y=y|X=x) = {x \choose y} \delta^{|x|-|y|} (1-\delta)^{|y|}.
\end{equation}
The proof is straightforward -- the number of distinct error events that give rise to $y$ from $x$ is exactly the number of subsequences of $x$ which are equal to $y$. Each of these error events has a probability $\delta^{|x|-|y|} (1-\delta)^{|y|}$, wherein the exponent of $\delta$ corresponds to the deleted symbols and the exponent of $1-\delta$ to the undeleted symbols. \\
\noindent \textbf{Maximum Likelihood (ML) estimate:}
Given the definition of the binomial coefficient, the maximum-likelihood (ML) estimate over a deletion channel with observed output $Y=y$ can be cast in the following form:
\begin{align*}
\argmax_{x \in \{0,1\}^n} {x \choose y}.\numberthis
\label{eq:ML_deletion}
\end{align*}
In the case of multiple deletion channels with observed traces $Y^{1}=y^1,...,Y^{t}=y^t$, the ML formulation is similar:
\begin{align*}
\argmax_{x \in \{0,1\}^n} \prod_{j=1}^{t} {x \choose y^{j}}.\numberthis
\label{eq:ML_multiple_deletion}
\end{align*}
As yet, there is no known efficient way to come up with a solution for either of the above two formulations (see \cite{mitzenmacher2009survey}).\\
\noindent \textbf{Relaxed binomial coefficient.}
We now introduce the function $\mathbf F(\cdot)$ which can be thought of as a real-valued relaxation of the binomial coefficient. This function is used in sections~\ref{sec:1deletion_ML} and~\ref{sec:1deletion}.
An intuitive definition is as follows: Consider a random vector $Z\in\{0,1\}^n$ such that $Z_i\sim$ ind. Ber$(p_i)$, and let $p$ be the vector of probabilities of length $n$. Then $\mathbf F(p,v)=\mathbb E_{Z\sim p}\ {Z \choose v}$, i.e., $\mathbf F(p,v)$ is the expected number of times $v$ appears as a subsequence of $Z$. If $p \in \{0,1\}^n$, then $Z=p$ with probability 1 and $\mathbf F(p,v) = {p \choose v}$.
More precisely, $\mathbf F(\cdot)$ is defined as:
\begin{definition}
\label{def:f}
\begin{align*}
&\mathbf F: [0,1]^n \times \{0,1\}^m \rightarrow \mathbb{R},\\
\mathbf F(p, v)\triangleq &\begin{cases}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [n],\\|\mathcal S|=m}} \quad \prod\limits_{i=1}^m p_{\mathcal S_i}^{v_i} (1-p_{\mathcal S_i})^{1-v_i} & 1 \leq m\leq n \\
1 & 0=m\leq n \\
0 & \text{else}.
\end{cases}
\end{align*}
\end{definition}
Though at first sight $\mathbf F(p,v)$ sums over an exponential number of subsets, a dynamic programming approach can be used to compute it in $O(nm)$ time complexity (see Appendix~\ref{app:F_compute}). Note that this is the same complexity as computing the binomial coefficient.\\
\noindent \textbf{Decomposition of the $t$-trace deletion channel:}
The following definitions and ideas are relevant to the results pertaining to multiple traces. We first state a result that aids in thinking about error events in multiple deletion channels.
The events occurring in the $t$-deletion channel model can be categorized into two groups:
\begin{enumerate}
\item an input symbol is deleted in \textit{all} the $t$-traces,
\item an input symbol is reflected in at least one of the traces.
\end{enumerate}
The error events of the first kind are in some sense ``not correctable'' or even ``detectable'' in any situation since it is impossible to tell with absolute certainty what and where the deleted symbol could have been (although the probabilities need not be uniform). The events of the second kind, however, can be detected and corrected in some situations. This thought process gives rise to a natural decomposition of the $t$-deletion channel model into a cascade of two channels: the first one being a deletion channel which captures error events of the first kind and the second one is what we call the \textit{remnant channel} which captures events of the second kind (see Fig.~\ref{fig:channel_equiv}). More precisely, we define the remnant channel as follows:
\begin{definition}
\textit{Remnant channel:} an input symbol to the remnant channel is reflected in any $k>0$ uniformly random traces and deleted in the rest with a probability ${t \choose k} \frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$. Thus, the probability of an input symbol reflected in a \textit{fixed} set of $k>0$ traces is equal to $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
\end{definition}
Note that probability of the union of all possible events here is $\sum_{k=1}^t {t \choose k} \frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}=1$, validating our definition.
\vspace{-5mm}
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.4]{channel_equiv.pdf}
\caption{A channel equivalence result: the $t$-trace deletion channel model in (a) is probabilistically equivalent to the the cascade of a deletion channel with the \textit{remnant channel} ($ \mathcal C_2$) in (b).}
\label{fig:channel_equiv}
\end{center}
\end{figure}
\vspace{-10mm}
\begin{restatable}{theorem}{channelequiv}
\label{thm:channel_equiv}
The $t$-deletion channel model and the cascade of the deletion channel with remnant channel shown in Fig.~\ref{fig:channel_equiv} are probabilistically equivalent, i.e.,
$$\Pr({Y}^{1}=y^1,{Y}^{2}=y^2,...,{Y}^{t}=y^t|X = x) = \Pr(\tilde{Y}^{1}=y^1,\tilde{Y}^{2}=y^2,...,\tilde{Y}^{t}=y^t|X = x).$$
\end{restatable}
A rigorous proof of this theorem for arbitrary length sequences can be found in Appendix~\ref{app:channel_equiv}. A similar, though not equivalent, decomposition has been exploited in \cite{haeupler2014repeated} albeit for the purpose of characterizing the capacity of multiple deletion channels -- there the authors consider deletion patterns which are ``undetectable''; for example, a deletion in the deletion channel $\mathcal C_1$ in the cascade model is undetectable since none of the traces will reflect that input symbol. However, our channel decomposition result does not appear in \cite{haeupler2014repeated}.\\
\noindent\textbf{Edit graph} (\cite{Gusfield1997}): Similar graph constructs have been defined in related problems on common supersequences and subsequences (see \cite{Nicosia2001} for example). This graph is closely related to the error events in the remnant channel. We start with a simple case and generalize subsequently. Define a directed graph called \textit{edit graph} given two sequences $f$ and $g$, where every path connecting the ``origin'' to the ``destination'' on the edit graph yields a supersequence $h$ of $f,g$, where $h$ is ``covered'' by $f,g$ -- i.e., each symbol of $h$ comes from either $f$ or $g$ or
both. In other words, given that $f$ and $g$ are the outputs of the remnant channel (with two outputs), each path from the origin of the edit graph to the destination corresponds to a possible input $h$ to the remnant channel and to an error event which resulted in outputs $f,g$ with input $h$.
For $f$ and $g$ in $\mathcal
A^*$, we form a directed graph $\mathcal{G}(f,g)$ with $(|f|+1)(|g|+1)$
vertices each labelled with a distinct pair $(i,j), 0\leq i\leq
|f|,\ 0\leq j\leq |g|$. A directed edge
$(i_1,j_1)\rightarrow(i_2,j_2)$ exists iff at least one of the following
holds:
\begin{enumerate}
\item$i_2-i_1=1$ and $j_1=j_2$, or
\item $j_2-j_1=1$ and $i_1=i_2$, or
\item $i_2-i_1=1$, $j_2-j_1=1$ and $f_{i_2}=g_{j_2}$,
\end{enumerate}
where $f_i$ is the $i^{th}$ symbol of the sequence $f$. The origin is the vertex $(0,0)$ and the destination $(|f|,|g|)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap1.pdf}
\caption{ Edit graph for sequences $f=$ `001' and $g=$ `101'. Make a grid so the vertical edges are aligned with a symbol in $f$ and horizontal edges with $g$ as shown. A diagonal edge $(i{-}1,j{-}1)\rightarrow (i,j)$ exists if $f_i = g_j$. The thick red edges form a path from the origin to the destination; this path corresponds to $h=$`0101' -- sequentially append the corresponding symbol to which each edge is aligned. It can also be verified that $h$ is a supersequence of both $f$ and $g$, and could be obtained as a covering of $f$ and $g$; the path itself gives one such covering. This covering also corresponds to an error event (or a deletion pattern) in the remnant channel which would result in outputs $f$ and $g$ with input $h=$ `0101' -- the deletion pattern is shown in the figure.}
\label{fig:editgraph_smap1}
\end{figure}
Let $p=((i_1,j_1),(i_2,j_2),...,(i_m,j_m))$ be a path in $\mathcal{G}(f,g)$. We define $ s(p)$ to be the sequence corresponding to the path. Intuitively, $s(p)$ is formed by appending symbols in the following
way: append the corresponding $f$ symbol for a vertical edge, $g$ symbol for horizontal edge, and $f$ or $g$ symbol for diagonal edge (see example Fig.~\ref{fig:editgraph_smap1}). Any path from $(0,0)$ to $(|f|,|g|)$ corresponds to a supersequence of $f$ and $g$ and which is covered by $f$ and $g$. More formally, define $ s(p)\triangleq x_1x_2...x_{m-1}$ where
$$x_k =
\begin{cases}
f_{i_{k+1}} \quad\text{if }j_{k}=j_{k+1},\\
g_{j_{k+1}} \quad\text{if }i_{k}=i_{k+1},\\
f_{i_{k+1}} \quad\text{else.}
\end{cases}
$$
The construct of edit graph can be extended to more than 2 sequences with the same idea. For sequences $f_1,f_2,...,f_t$, construct a $t$-dimensional grid with a number of vertices $(|f_1|+1)(|f_2|+1)...(|f_t|+1)$ labeled from $(0,0,...,0)$ to $(|f_1|,|f_2|,...,|f_t|)$. A vertex $u=(i_1,i_2,...,i_t)$ is connected to $v=(j_1,j_2,...,j_t)$ (we say $u \rightarrow v$) iff both of the following conditions are met:
\begin{itemize}
\item $j_l=i_l$ or $j_l=i_l+1$ $\forall\ l\in [t]$, i.e., $(i_1,...,i_t)$ and $(j_1,...,j_t)$ are vertices of a particular unit cube. Only these type of vertices can share an edge in the grid graph.
\item Let $\mathcal T \subseteq [t]$ be the collection of indices where $j_l=i_l+1$. Then ${f_l}_{j_l}$ is equal $\forall\ l \in \mathcal T$. For example in 4 dimensional grid, consider the two vertices $(10,5,8,2)$ and $(10,6,9,2)$. In this case $\mathcal T = \{2,3\}$ since the second and third coordinates differ by 1. Therefore $(10,5,8,2)\rightarrow (10,6,9,2)$ iff ${f_2}_{5}={f_3}_{9}$. Note that if only one coordinate differs by 1 in the two vertices, a directed edge always exists (in other words all non-diagonal edges exist).
\end{itemize}
Define the vertex $(0,...,0)$ to be the origin of this graph and the vertex $(|f_1|,...,|f_t|)$ to be the destination. If $|f_j|=O(n)\ \forall\ j$, this graph has a number of vertices $O(n^t)$ and a maximum number of edges $O((2n)^t)$ since each vertex has at most $2^t-1$ outgoing edges.\\
\noindent\textbf{Infiltration product} (introduced in section 6.3 of \cite{lothaire1997combinatorics}):
The infiltration product has been extensively used in \cite{lothaire1997combinatorics}, as a tool in non-commutative algebra. Here, we give an edit-graph interpretation of this tool. A formal algebraic definition of the infiltration product is in Appendix~\ref{app:infil_def}. Using the edit graph we can construct the set of possible
supersequences $\mathcal{S}(f,g)$ of $f$, $g$ that are covered by the symbols in $f$ and $g$.
Indeed, multiple paths could yield the same supersequence and we
can count the number of distinct ways $\mathbf N(h;f,g)$ one can construct
the same supersequence $h$ from $f$, $g$.
We can informally define the
\emph{infiltration product $f\uparrow g$} of $f$ and $g$, as a polynomial with monomials the supersequences $h$ in
$\mathcal{S}(f,g)$ and coefficients $\langle f\uparrow g,h\rangle$ equal to $\mathbf N(h;f,g)$. For the example in Fig.~\ref{fig:editgraph_smap1}, there is exactly one path corresponding to `101001' and hence $\langle 001\uparrow 101,101001 \rangle=1$ and similarly $\langle 001\uparrow 101,01001 \rangle=2$. One could find these coefficients for all relevant sequences and form the polynomial as described. We now give additional examples (see 6.3.14 in \cite{lothaire1997combinatorics}). Let $\mathcal{A}=\{a,b\}$, then
\begin{itemize}[wide=2pt]
\item $ab\uparrow ab=ab+2aab+2abb+4aabb+2abab$,
\item $ab\uparrow ba=aba+bab+abab+2abba+2baab+baba.$
\end{itemize}
The infiltration operation is commutative and associative, and infiltration of two sequences $f\uparrow g$ is a polynomial with variables of length (or \textit{degree}) at most $|f|+|g|$; see \cite{lothaire1997combinatorics}.
The definition of infiltration extends to two polynomials via distributivity (precisely defined in Appendix~\ref{app:infil_def}), and consequently to multiple sequences as well. For multiple sequences, infiltration has the same edit graph interpretation: $\langle f_1\uparrow f_2 \uparrow...\uparrow f_t, w \rangle$ is the number of distinct ways of constructing $w$ as a supersequence of $f_1, f_2, ... ,f_t$ so that the construction covers $w$, i.e., construct the $t$-dimensional edit graph of $f_1, f_2, ... ,f_t$ and count the number of paths corresponding to $w$.
\\
{\small
\begin{center}
\begin{tabular}{ |P{3cm}|P{10cm}| }
\hline
\multicolumn{2}{|c|}{Table of notation} \\
\hline
$\mathcal A$ & A set \\
\hline
$X$ & A random variable or a random vector \\
\hline
$x$ & A scalar or a vector variable\\
\hline
$|x|$ & Length of the sequence $x$\\
\hline
$[i:j]$ & $\{i,i+1,...,j\}$\\
\hline
$x^{(i\rightarrow s)}$ & $(x_1,x_2,...,x_{i-1},s,x_{i+1},...,x_n)$ \\
\hline
${f \choose g}$& Binomial coefficient: number of subsequence patters of $f$ equal to $g$ \\
\hline
$\mathbf F(p,v)$ & Relaxed binomial coefficient: $\mathbb E_{Z\sim p} {Z \choose v}$ \\
\hline
$\langle f \uparrow g,h \rangle$ & Infiltration product: number of ways of obtaining sequence $h$ as a ``covered'' supersequence of $f$ and $g$ \\
\hline
\end{tabular}
\end{center}} \vspace{2mm}
\section{Maximum likelihood for the single-trace deletion channel}
\label{sec:1deletion_ML}
Consider the ML formulation for the single-trace deletion channel given non-empty trace $Y$ and all possible $n$-length inputs allowed, restated here for convenience:
\begin{equation}
\argmax_{x\in \{0,1\}^n} {x \choose Y}.
\label{eq:1deletion_ML}
\end{equation}
To the best of our knowledge, the only known method to solve \eqref{eq:1deletion_ML} involves iterating over all possible choices of $x$ and computing the objective value for each of the choices. We here show that it is sufficient to solve a continuous relaxation of above problem to obtain a solution to \eqref{eq:1deletion_ML}. Note that there could be multiple solutions to \eqref{eq:1deletion_ML}. Before going in the main result, we first state a useful lemma which factors out a given coordinate $p_i$ out of $\mathbf F(p,Y)$. The proof of the lemma is relegated to the appendix.
\begin{restatable}{lemma}{deletionMLrellemma}
For $p=[p_1,p_2,..,p_i,...,p_n]$ and $Y=Y_1Y_2...Y_m$ with $n \geq m > 0$, we have
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
\label{lemma:F_decomposition}
\end{restatable}
\begin{theorem} An alternative optimization to characterize ML for the single-trace deletion channel.
\begin{equation}
\max_{x\in \{0,1\}^n} {x \choose Y} = \max_{p\in [0,1]^n} \mathbf F(p,Y).
\label{eq:ml_opt_equiv}
\end{equation}
Furthermore, given any non-integral $p^* \in [0,1]^n$ that maximizes $\mathbf F(p,Y)$, one could construct a corresponding integral solution $x^* \in \{0,1\}^n$ that maximizes $\mathbf F(p,Y)$ and consequently is also a solution to $\max_{x\in \{0,1\}^n} {x \choose Y}$.
\end{theorem}
\begin{proof}
As noted earlier, we have ${x \choose Y} = \mathbf F(x,Y)$. Therefore, we are interested in proving the following:
\begin{align*}
\max_{x\in \{0,1\}^n} \mathbf F(x,Y) \equiv \max_{p\in [0,1]^n} \mathbf F(p,Y).\numberthis
\label{eq:ml_opt_equiv_proof1}
\end{align*}
To show this and also the second statement, we prove that given $p=(p_1,p_2,...,p_i,...,p_n)$, at least one of the following holds true:
\begin{itemize}
\item$\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},0,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 0.
\item $\mathbf F(p^{(i\rightarrow 1)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},1,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 1.
\end{itemize}
Thus if $p^*$ is an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ with $p_i\in (0,1)$, then at least one of $p^{(i\rightarrow 0)}$ or $p^{(i\rightarrow 1)}$ is also an optimal solution. Sequentially applying this argument for each coordinate of $p$ shows that there exists a point in $\{0,1\}^n$ which is also an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ and consequently to $\max_{x\in \{0,1\}^n} \mathbf F(x,Y)$.
Now to prove our claim, we use Lemma~\ref{lemma:F_decomposition} to factor out $p_i$ terms in $\mathbf F(p,Y)$:
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
There are 3 cases
\begin{enumerate}
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) = \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) = \mathbf F(p,Y) = \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) > \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \leq \mathbf F(p,Y) \leq \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) < \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y) \geq \mathbf F(p^{(i\rightarrow 1)},Y).$
\end{enumerate}
Thus in each of the 3 cases, we see that at least one of $\mathbf F(p^{(i\rightarrow 0)},Y)$ or $\mathbf F(p^{(i\rightarrow 1)},Y)$ is at least as large as $\mathbf F(p,Y)$ thus proving the theorem. Note that the proof gives a natural way to find an optimal lattice point $p_{lattice}^*$ given a non-lattice point $p^*$ by iterating through each coordinate of $p^*$ and switching them to $0$ or $1$ by comparing $\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})$ with $\sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$
\end{proof}
\section{Numerical results}
\label{sec:Numerics}
In this section we show numerics supporting our theoretical results. In all of our experiments, we generate the input sequence uniformly at random (motivated by average case trace reconstruction), and obtain the $t$ traces by passing the input through a deletion channel (with a deletion probability $\delta$) $t$ times. We then reconstruct the input from the obtained traces and measure how \textit{close} the reconstructed sequence is, to the actual input sequence.
We use two metrics to measure the performance of the reconstruction algorithms: 1. \textit{Hamming error rate}, which is defined as the average Hamming distance between the actual input and the estimated sequence divided by the length of the input sequence and 2. \textit{Edit error rate}, which is defined as the average edit distance between the actual input and the estimated sequence divided by the length of the input sequence. The reason for using Hamming error rate is that our goal is to reconstruct a \textit{known-length} sequence, which has been the problem formulation throughout this work. Moreover, the Hamming error rate is also of special interest to us since the symbolwise MAP is an optimal estimator for minimizing the Hamming error rate (see Appendix~\ref{app:smap_hamming} for a proof). We also use edit error rate as it is a typical metric used in the context of insertion/deletion channels.
{\small
\begin{center}
\begin{tabular}{ |P{3cm}|P{9.5cm}|P{2.5cm}| }
\hline
\multicolumn{3}{|c|}{List of trace reconstruction algorithms compared in this work.} \\
\hline
\bf Abbreviation & \bf Description & \bf Complexity\\
\hline
Ind. post. comb. & Independent posterior combination (Alg.~\ref{alg:ind_comb}) & $O(n^2t)$\\
\hline
BMA & Bitwise majority alignment of \cite{Batu2004} (Alg.~\ref{alg:bitwise_majority}) & $O(nt)$ \\
\hline
Trace stats. & Algorithm based on trace symbolwise statistics from \cite{Holenstein2008} (Alg.~\ref{alg:trace_statistics})
& $O(n^{3.37}+nt)$\\
\hline
Grad asc. & Projected gradient ascent (Alg.~\ref{alg:grad_asc_traces}) & $O(n^2t)$\\
\hline
SMAP seq. & Sequential symbolwise MAP heuristic (Alg.~\ref{alg:apprx_smap}) & $O(n^2t)$ \\
\hline
SMAP exact & Exact symbolwise MAP (Alg.~\ref{alg:exact_smap}) & $O(n^{t+2}2^t)$\\
\hline
\end{tabular}
\end{center}}
\vspace{5mm}
\noindent \textbf{Baseline algorithms:}
\begin{enumerate}[leftmargin = *]
\item \textbf{Independent posterior combination:} As pointed in the introduction, computing the posterior probabilities for each deletion channel and combining them as if they came from independent observations does not provide a natural solution for computing the posterior probabilities for the $t$-trace deletion channel. One could, however, check how such a naive combination of posteriors compares with our reconstruction algorithms for $t$-traces. This is detailed as Alg.~\ref{alg:ind_comb}. The complexity of this algorithm is $O(n^2t)$ since computing the posteriors takes $O(n^2)$ and we compute posteriors for $t$ traces.
\item \textbf{Bitwise Majority Alignment (introduced in \cite{Batu2004}):} BMA reconstructs the input sequence by first ``aligning'' the traces using a pointer for each trace, and then taking the majority of the pointed symbols. BMA is detailed as Alg.~\ref{alg:bitwise_majority}. From an efficiency standpoint, BMA is the most efficient of all the algorithms since it is linear in the blocklength as well as the number of traces ($O(nt)$).
\item \textbf{Trace statistics algorithm:} An algorithm based on trace symbol statistics (also called mean-based algorithms and summary statistics algorithms) has been extensively studied for worst-case trace reconstruction (see \cite{Holenstein2008}, \cite{De2017}, \cite{Nazarov:2017}). In essence, the algorithm first estimates the ``trace symbol statistics'' -- $\Pr(Y_i=1)\ \forall\ i$ -- from the obtained traces and uses only these estimates to reconstruct $X$. However, it uses a new set of traces for every position $i$, thus requiring at least $n$ traces (see (3.6) and the paragraph below (3.8) in \cite{Holenstein2008}). Here we modify the algorithm to adapt them for an arbitrary number of traces; in particular, we reuse the traces while estimating $\Pr(Y_i=1)\ \forall\ i$. The algorithm is detailed in Alg.~\ref{alg:trace_statistics}.
The complexity analysis for this gets tricky since it depends on the algorithm used to solve the set of $2n$ linear programs. The state-of-the-art algorithm for solving a linear program in $n$ variables takes approximately $O(n^{2.37})$ (see \cite{cohen2019solving}); thus the complexity of Trace statistics algorithm is $O(n^{3.37}+nt)$, where the $nt$ term corresponds to the complexity of computing $\hat p_j$. However, in our implementation we use the solver from the "SciPy" Python library which uses primal-dual interior point methods for solving linear programs. The complexity of such methods is typically $O(n^3)$ making our implementation $O(n^4+nt)$. Also note that these are iterative methods and have many hidden constants (such as the number of iterations for convergence).
\end{enumerate}
We note that the state-of-the-art average-case trace reconstruction algorithms in the literature are applicable in the asymptotic regime where the blocklength $n$ and the number of traces $t$ approach $\infty$; it is not clear how to adapt such algorithms for a finite blocklength and a small number of traces. It is for this reason that we chose to compare against BMA and Trace statistics algorithm, which can be easily adapted for the finite blocklength regime and for a small number of traces. It should also be noted that the performance of the above two algorithms may not be reliable with a small number of traces (as they are not designed for this regime), yet we include them owing to the lack of better baselines.\\
\begin{algorithm}[t!]
\caption{Trace reconstruction via independent posterior combination}\label{alg:ind_comb}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{1}=y^1,...,Y^{t}=y^t$}, input length $n$ \\ Outputs: Estimate of the input $\hat X$
\State Initialize priors $p^{old} \gets (0.5,0.5,...,0.5)$
\For {$l=1:t$}
\State Use Alg.~\ref{alg:apprx_smap_dp} with $p^{old}$ and $y^l$ to compute posteriors $p^{l,new}$
\EndFor
\For {$i=1:n$}
\If {$\prod_{l=1}^t p^{l,new}_i \geq \prod_{l=1}^t (1-p^{l,new}_i)$}
$\ \hat X_i \gets 1$
\Else
$\ \hat X_i \gets 0$
\EndIf
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[!t]
\caption{Bitwise Majority Alignment}
\label{alg:bitwise_majority}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{1}=y^1,...,Y^{t}=y^t$, input length $n$}\\ Output: estimate of input $\hat X = \hat X_1 \hat X_2...\hat X_n$.
\State Initialize $c_j=1$ for $j\in [t]$.
\State Initialize $\hat X_i = 1$ for $i \in [n]$.
\For{$i \in\ [1:n]$}
\State Let $b$ be the majority over all $t$ of $y^j_{c_j}$
\State $\hat X_i \gets b$
\State Increment $c_j$ for each $j$ such that $y^j_{c_j} = b$
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[!t]
\caption{Trace statistics heuristic}
\label{alg:trace_statistics}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{1}=y^1,...,Y^{t}=y^t$, input length $n$}\\ Output: estimate of input $\hat X = \hat X_1 \hat X_2...\hat X_n$.
\State Append each trace $y^j$ with zeros until each of them is of length $n$.
\State Assign $\hat p_j \leftarrow \frac{|\{y^l:y^l_j=1\}|}{t}$.
\For{$i \in\ [1:n]$}
\State Solve the 2 linear programs (3.6) in \cite{Holenstein2008} by fixing $x_i=0$ and $x_i=1$: let the optimum value in the two cases be $m_0$ and $m_1$ respectively.
\State If $m_0<m_1$, assign $\hat X_i = x_i \leftarrow 0$. Else fix $\hat X_i = x_i \leftarrow 1$.
\EndFor
\end{algorithmic}
\end{algorithm}
\newpage
\noindent \textbf{Algorithms introduced in this paper:}
\begin{enumerate}[leftmargin = *]
\item \textbf{Projected gradient ascent:} Alg.~\ref{alg:grad_asc_traces} used as described, with max iterations $M=100$ and convergence criteria $C$ set as follows: the percentage difference in $\sum_j \mathbf F(p,y^j)$ over two consecutive iterations is less than 0.1\%.
\item \textbf{Symbolwise MAP sequentially used one trace at a time:} Alg.~\ref{alg:apprx_smap} used as described.
\item \textbf{Exact symbolwise MAP:} Alg.~\ref{alg:exact_smap} used as described.
\end{enumerate}
\vspace{7mm}
\textbf{Observations:} In Fig.~\ref{fig:numerics_hamming} and Fig.~\ref{fig:numerics_edit}, we compare the Hamming and edit error rates for the different algorithms described above.
\begin{itemize}[leftmargin = *]
\item The 3 algorithms introduced in this work outperform the 3 baselines in most cases. The Hamming error rate of Grad asc. with 2 and 3 traces is a notable exception as it does worse than Ind. post. comb. However, it improves rapidly as we increase the number of traces as seen in Fig.~\ref{fig:numerics_hamming}.
\item Both Ind. post. comb. as well as our SMAP seq. struggle with the problem of \textit{diminishing returns} for Hamming error rate as they do not improve much with the number of traces. This could indicate that considering traces one at a time could fail to accumulate extrinsic information (for instance, it completely neglects the possible alignments given multiple traces); one needs to simultaneously consider multiple traces in order to accomplish this. SMAP seq. however, improves with the number of traces with respect to edit error rate.
\item The Grad asc. is the ``champion'' amongst the algorithms we compare here, when it comes to the edit error rate as illustrated by Fig.~\ref{fig:numerics_edit}. The Grad asc. was constructed with the aim of maximizing the likelihood of the observed traces, and this in turn seems to have some correlation with minimizing the edit distance -- it is not clear why this is the case.
\item As seen in Fig.~\ref{fig:numerics_hamming} (a) and (b), SMAP exact has the minimum Hamming error rate. This supports the fact that symbolwise MAP is the minimizer of the Hamming error rate. However, note that this does not necessarily minimize the edit error rate, as seen from Fig.~\ref{fig:numerics_edit} (a) and (b).
\end{itemize}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.85]{Numerics/hamming_error_rates.pdf}
\caption{Comparison of Hamming error rates for a blocklength $n=100$ illustrated with 2,3,5 and 10 observed traces. Note that we do not run SMAP exact. for 5 and 10 traces since its complexity grows exponentially with the number of traces. All the subplots are plotted on the same scale to aid comparability across subplots. Few of the subplots which contain algorithms with similar error rates also contain a zoomed-in inset view.}
\label{fig:numerics_hamming}
\end{figure}
\clearpage
\begin{figure}[!t]
\centering
\includegraphics[scale=0.85]{Numerics/edit_error_rates.pdf}
\caption{Comparison of edit error rates for a blocklength $n=100$ illustrated with 2,3,5 and 10 observed traces. Note that we do not run SMAP exact. for 5 and 10 traces since its complexity grows exponentially with the number of traces. All the subplots are plotted on the same scale to aid comparability across subplots. Few of the subplots which contain algorithms with similar error rates also contain a zoomed-in inset view.}
\label{fig:numerics_edit}
\end{figure}
\clearpage
\section{Introduction}
\label{sec:intro}
Sequence reconstruction over deletion channels, both with and without a codebook, has received considerable attention in the information theory as well as in the theoretical computer science literature. From an information theory perspective, reconstruction over the deletion channel, or more specifically a maximum-likelihood (ML) argument for the deletion channel, would give further insight on the capacity of the deletion channel, a long-standing open problem (see \cite{mitzenmacher2009survey}). To quote \cite{mitzenmacher2009survey} -- ``at the very least, progress in this direction would likely surpass previous results on the capacity of the deletion channels''. Yet, there are no results on reconstruction over a deletion channel with statistical guarantees -- in this work, we take a step in this direction.
On the other hand, the problem of \textit{trace reconstruction}, as introduced in \cite{Batu2004}, has received renewed interest in the past few years (see \cite{Holenstein2008}, \cite{Peres2017}, \cite{De2017}, \cite{holden18}, \cite{Nazarov:2017}, \cite{holden2018lower}, \cite{chase2019new}). The problem of trace reconstruction can be stated simply as follows: consider a sequence $X$ which is simultaneously passed through $t$ independent deletion channels to yield $t$ deleted observations (also called \textit{traces}) of $X$ (see Fig.~\ref{fig:tdeletion}). How many such traces are needed to reconstruct $X$ perfectly? A variety of upper and lower bounds for this problem have been proposed, both for worst case and average case reconstruction. Our problem definition, stated in the following paragraph, is closely related to the average case trace reconstruction.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.2]{tdeletion.pdf}
\caption{$t$-deletion channel model: sequence $X$ passed through $t$ independent deletion channels to yield $t$ \textit{traces}. We aim to estimate $X$ from the $Y^{(i)}$s.}
\label{fig:tdeletion}
\end{center}
\end{figure}
\noindent \textbf{Problem definition.} Given an input sequence of length $n$ (known apriori), the independently and identically distributed (i.i.d.) deletion channel deletes each input symbol indepedently with probability $\delta$, producing at its output a subsequence of the input sequence. Consider a sequence $X$ passed through $t$ ($t$ is fixed) such deletion
channels as shown in
Fig.~\ref{fig:tdeletion}. We call this the $t$-deletion channel model. We ask two questions:
\begin{enumerate}
\item For $t=1$ (see Fig.~\ref{fig:1deletion}, also called the single deletion channel model), and when $X_i \sim\ \text{ind. Ber}(p_i)$, compute the posterior distributions of $X_i$ given the trace $Y$.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.25]{1deletion.pdf}
\caption{The single deletion channel model. We assume $X_i \sim\ \text{ind. Ber}(p_i)$.}
\label{fig:1deletion}
\end{center}
\end{figure}
\item In the $t$ deletion channel model, for a fixed $t$ assume that $X_i \sim\ \text{i.i.d. Ber}(0.5)$ and compute the posterior distributions of $X_i$ given all traces $Y^{(1)}, Y^{(2)},...,Y^{(t)}$.
\end{enumerate}
Note that solving 1) above doesn't lead to a natural solution for 2). This is because for a memoryless channel, we have $Y^{(j)} - X_i - Y^{(k)}$ and hence $\Pr(X_i=\alpha|Y^{(j)}, Y^{(k)}) \propto \Pr(X_i=\alpha|Y^{(j)}) \Pr(X_i=\alpha|Y^{(k)})$; so one could independently combine the posterior probabilities from each noisy observation. This is not the case for deletion channels since the markov chain $Y^{(j)} - X_i - Y^{(k)}$ no longer holds. More intuitively, one needs to first ``align'' the traces for computing the likelihoods.
We point out that the problem considered in 2) asks a question complementary to the trace reconstruction: given a fixed (possibly a few) number of traces, what is our ``best'' guess of $X$? We provide algorithms which do this. Unlike trace reconstruction, we are not concerned with perfect reconstruction (since perfect reconstruction may not be possible with just a few traces), although it should also be obvious that performance guarantees for our algorithms (not a part of this work) would naturally lead to upper bounds for trace reconstruction.
Deletion channel by itself is known to be notoriously difficult to analyse. As stated earlier, the capacity of a single deletion channel is still unknown (\cite{diggavi2007capacity,diggavi2006information,diggavi2001transmission}); as are optimal coding schemes. Recent works have looked at the design of codes for deletion channels (\cite{ratzer2005marker,ratzer2000codes,thomas2017polar}); these works consider use of a codebook (we do not). As a result, statistical estimation over deletion channels is also a difficult problem due its highly combinatorial nature. To the best of our knowledge, as yet there are no efficient estimation algorithms over deletion channels with statistical guarantees; not even for ML over a single deletion channel.
\noindent \textbf{Biological motivation.} Trace reconstruction in itself was motivated, in part, by problems DNA sequence reconstruction. One such problem was to infer the DNA sequence of a common ancestor from the samples of its descendants. We argue that our problem definition fits more naturally in such a scenario since perfect reconstruction may not be feasible or even possible. Our motivation for considering this problem also comes from a recent DNA sequencing technology called \textit{nanopore sequencing}. The $t$-deletion channel model is a simplistic model to approximately capture the process of a DNA sequence passed through a nanopore sequencer\footnote{As seen in
\cite{Mao2017},\cite{MDK17} there are more complicated effects of
the nanopore reader not captured in this simple
representation.}.
Very recently, a variant of the trace reconstruction problem called \textit{coded trace reconstruction} has been proposed, motivated by portable DNA-based data storage systems using DNA nanopores (see \cite{abroshan2019coding}, \cite{cheraghchi2019coded}, \cite{brakensiek2019coded}) and we believe that the ideas in this work may prove useful in such a setting.
There are other works on sequence assembly (see for example, \cite{Li09fastand}, \cite{Shomorony2016}),
where multiple short reads (from different segments of a sequence) are used to reconstruct the bigger
sequence. This work differs from sequence assembly since we are interested to infer the entire length sequence and not just small segments of it (which are then ``stitched'' together in sequence assembly).
\noindent \textbf{Tools and techniques.} In terms of theoretical tools, the series of books by Lothaire (\cite{lothaire1997combinatorics,lothaire2002algebraic,lothaire2005applied}) extensively use algebraic tools for problems in the combinatorics of sequences (or \textit{words}), and our work is inspired by such techniques. We borrow many of their notations and results for our work.
\noindent \textbf{Contributions.} {\color{red}} Our main contribution is to provide tools to visualize and analyze error events (described in the next section) for the multiple deletion channel model in Fig.~\ref{fig:tdeletion}. We also provide algorithms to solve the problems stated in 1) and 2) earlier in the section.
\begin{itemize}[wide=5pt]
\item In section~\ref{sec:1deletion}, for the single deletion channel model, we provide an $O(n^2)$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i=1|Y)\ \forall\ i$ when $X_i \sim \text{ind. Ber}(p_i)$.
\item In Section~\ref{sec:exactsmap}, for the $t$-deletion channel model, we give an $O(2^t n^{t+2})$ algorithm to calculate the symbolwise posterior probabilities $\Pr(X_i = 1|Y_1,...,Y_t)$ when $X_i \sim \text{ind. Ber}(0.5)$.
\end{itemize}
\section{Notation and Tools}
\label{sec:notations}
\noindent \textbf{Basic notations:} In this work, we borrow a majority of the notation and tools from \cite{lothaire1997combinatorics} which deals with non-commutative algebra. We restate the definitions here for convenience. Calligraphic letters refer to sets, capitalized letters
correspond to random variables and bold letters are used for functions. Let $\mathcal{A}$ be the set of all
symbols. Throughout this work, we will focus on the case where $\mathcal{A}=\{0,1\}$,
though our methods extend to arbitrarily large sets of finite size. Define
$\mathcal{A}^n$ to be the set of all $n$-length sequences and $\mathcal{A}^*$ to be the set of all finite length sequences with
symbols in $\mathcal{A}$. For a sequence $f$, $|f|$ denotes the length
of $f$. For integers $i,j$, we define $[i:j] \triangleq \{i,i+1,...,j\}$ if $j\geq i$ and $[i:j] \triangleq \emptyset$ otherwise. Also define $[i] \triangleq [1:i]$.\\
\noindent \textbf{Binomial coefficient:} Given sequences $f$ and $g$ in $\mathcal{A}^*$, the number of
subsequence patterns of $f$ that are equal to $g$ is called the
\textit{binomial coefficient} of $g$ in $f$ and is denoted by $f
\choose g$. For example, ${'apple' \choose 'ape'} = 2$ since $'ape'$ can be obtained from two (overlapping) subsequences of $'apple'$.
When the alphabet $\mathcal{A}$ is of cardinality 1, ${f
\choose g} = {|f| \choose |g|}$, the classical binomial coefficient
with their respective lengths as the parameters. This definition hence
could be thought of as a generalization of the classical binomial
coefficients. We will denote by $e$ the sequence of length 0, and {define ${f
\choose e} \triangleq 1\ \forall\ f\ \in\ \mathcal{A}^*$.} We also define the classical binomial coefficient ${a \choose b}\triangleq 0,$ whenever $b>a$ or $b<0$ for ease of use.
The binomial coefficient is an integral aspect of this work and for analyzing error events in deletion channels because the input-output relation for a deletion channel (with deletion probability $\delta$, input $X$ and output $Y$) can be expressed as
\begin{equation}
\label{eq:in-out_relation}
\Pr(Y=y|X=x) = {x \choose y} \delta^{|x|-|y|} (1-\delta)^{|y|}.
\end{equation}
The proof is straight-forward -- the number of distinct error events that give rise to $y$ from $x$ is exactly the number of subsequences of $x$ which are equal to $y$. Each of these error events has a probability $\delta^{|x|-|y|} (1-\delta)^{|y|}$, wherein the exponent of $\delta$ corresponds to the deleted symbols and the exponent of $1-\delta$ to the undeleted symbols.
Given the definition of the binomial coefficient, the maximum-likelihood (ML) estimate over a deletion channel with observed output $Y$ can be cast in the following form:
\begin{align*}
\argmax_{x \in \mathcal C} {x \choose Y},\numberthis
\label{eq:ML_deletion}
\end{align*}
where $\mathcal C$ is the chosen codebook. In the case of multiple deletion channels with observed traces $Y^{(1)},...,Y^{(t)}$, the ML formulation is similar:
\begin{align*}
\argmax_{x \in \mathcal C} \prod_{j=1}^{t} {x \choose Y^{(j)}}.\numberthis
\label{eq:ML_multiple_deletion}
\end{align*}
As yet, there is no known efficient way to come up with a solution for the above two formulations, even for \eqref{eq:ML_deletion} with $\mathcal C = \{0,1\}^n$ (see \cite{mitzenmacher2009survey}). In this work, we attempt to take a step in this direction by showing that a continuous relaxation of \eqref{eq:ML_deletion} is equivalent to \eqref{eq:ML_deletion}. However, an efficient algorithm to solve the optimization \eqref{eq:ML_deletion} remains open. In the context of trace reconstruction, the ultimate pursuit would be an algorithm for \eqref{eq:ML_multiple_deletion} with $\mathcal C = \{0,1\}^n$ and error analysis thereof.
\noindent We now describe a function which can be thought of as a real-valued extension of the binomial coefficient. This function is used in sections~\ref{sec:1deletion_ML} and \ref{sec:1deletion}. Consider the function $\mathbf F(\cdot)$ defined as:
\begin{definition}
\label{def:f}
\begin{align*}
&\mathbf F: \mathbb{R}^k \times \{0,1\}^l \rightarrow \mathbb{R},\\
\mathbf F(q, v)\triangleq &\begin{cases}
\sum\limits_{\substack{\mathcal S|\mathcal S\subseteq [k],\\|\mathcal S|=l}} \quad \prod\limits_{i=1}^l q_{\mathcal S_i}^{v_i} (1-q_{\mathcal S_i})^{1-v_i} & 1 \leq l\leq k \\
1 & 0=l\leq k \\
0 & \text{else}.
\end{cases}
\end{align*}
\noindent An alternate definition is as follows: consider a random vector $Z\in\{0,1\}^k$ such that $Z_i\sim$ ind. Ber$(q_i)$, let $q$ be the vector of probabilities of length $k$, $\mathcal S\subseteq [k]$ a subset of the indices of size $l$, and ${v}$ a binary vector of length $l$. Then, $$\mathbf F(q,v)=\sum_{\substack{\mathcal S||\mathcal S|=l}} \Pr(Z_{\mathcal S}= v).$$ Note that if $q$ is a vector of $0$'s and $1$'s, then $\mathbf F( q, v)={q \choose v}$, the binomial coefficient of $v$ in $q$. Thus, $\mathbf F(\cdot)$ could be interpreted as an extension of the binomial coefficient where one of the parameters can take values in $[0,1]^n$ instead of $\{0,1\}^n$.
\end{definition}
Though at first sight, $\mathbf F(q,v)$ sums over an exponential number of subsets, a dynamic programming approach can be used to compute it in $O(|v|^2)$ time complexity. The dynamic program is described in section~\ref{sec:1deletion}.
\iffalse
\noindent \textbf{Edit distance:} The edit distance $d_e(f,g)$ measures similarity between two sequences of
possibly different lengths \cite{Navarro2001}. $d_e(f,g)$ is the minimum number of operations
needed to transform $f$ to $g$, where the permitted operations are
insertion, deletion or substitution of a symbol. In this work, we
quantify the performance of algorithms in Section \ref{sec:Numerics}
using the edit distance metric.
\fi
The following definitions and ideas are relevant only for \ref{sec:exactsmap} and can be omitted for the sections on single deletion channel. Before getting to the mathematical definitions, we first state a result of ours that aids in thinking about error events in multiple deletion channels.
\subsection{An alternate interpretation of multiple deletion channel model}
The events occurring in the $t$-deletion channel model can be categorized into two groups:
\begin{enumerate}
\item an input symbol is deleted in \textit{all} the $t$-traces,
\item an input symbol is reflected in at least one of the traces.
\end{enumerate}
The error events of the first kind are in some sense ``not correctable" or even ``detectable" since it is impossible to tell what and where the deleted symbol could have been (although the probabilities need not be uniform). The events of the second kind, however, can still be detected although they could likely be confused with a large set of similar events. This thought process gives rise to a natural decomposition of the $t$-deletion channel model into a cascade of two channels: the first one being a deletion channel which captures error events of the first kind and the second one is what we call the \textit{remnant channel} which captures events of the second kind (see Fig.~\ref{fig:channel_equiv}). More precisely, the remnant channel is defined as follows:
\begin{definition}
\textit{Remnant channel:} an input symbol to the remnant channel is reflected in $k>0$ given outputs and deleted in the rest with a probability $\frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}$.
\end{definition}
It is easy to note that probability of the union of all possible events here is $\sum_{k=1}^t {t \choose k} \frac{\delta^{t-k}(1-\delta)^k}{1-\delta^t}=1$, validating our definition.
\begin{restatable}{theorem}{channelequiv}
\label{thm:channel_equiv}
The $t$-deletion channel model and the cascade of the deletion channel with remnant channel shown in Fig.~\ref{fig:channel_equiv} are probabilistically equivalent, i.e.,
$$\Pr({Y}^{(1)},{Y}^{(2)},...,{Y}^{(t)}|X = x) = \Pr(\tilde{Y}^{(1)},\tilde{Y}^{(2)},...,\tilde{Y}^{(t)}|X = x).$$
\end{restatable}
The formal proof of the theorem requires few more definitions and is relegated to the appendix.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=0.4]{channel_equiv.pdf}
\caption{A channel equivalence result: the $t$-deletion channel model in (a) is probabilistically equivalent to the the cascade of a deletion channel with the \textit{remnant channel} ($ \mathcal C_2$) in (b).}
\label{fig:channel_equiv}
\end{center}
\end{figure}
\noindent\textbf{Edit graph} (as defined in \cite{Gusfield1997}): We now define a graph construct which is closely related to the error events in the remnant channel. We start with a simple case and generalize subsequently. Define an \textit{edit graph} given two sequences $f$ and $g$, where every path connecting the ``origin'' to the ``destination'' on the edit graph yields a supersequence $h$ of $f,g$, where $h$ is ``covered'' by $f,g$ -- i.e., each symbol of $h$ comes from either $f$ or $g$ or
both. In other words, given that $f$ and $g$ are the outputs of the remnant channel (with two outputs), each path from the origin of the edit graph to the destination corresponds to a possible input $h$ to the remnant channel and to an error event which resulted in $f$ and $g$ with input $h$.
For $f$ and $g$ in $\mathcal
A^*$, we form a graph $\mathcal{G}(f,g)$ with $(|f|+1)(|g|+1)$
vertices each labelled with a distinct pair $(i,j), 0\leq i\leq
|f|,\ 0\leq j\leq |g|$. A directed edge
$(i_1,j_1)\rightarrow(i_2,j_2)$ exists iff at least one of the following
holds:
\begin{enumerate}
\item$i_2-i_1=1$ and $j_1=j_2$, or
\item $j_2-j_1=1$ and $i_1=i_2$, or
\item $i_2-i_1=1$, $j_2-j_1=1$ and $f_{i_2}=g_{j_2}$,
\end{enumerate}
where $f_i$ is the $i^{th}$ symbol of the sequence $f$. The origin is the vertex $(0,0)$ and the destination $(|f|,|g|)$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.25]{editgraph_smap1.pdf}
\caption{ Edit graph for sequences $f=$ `001' and $g=$ `101'. An easy way to think about this is to write down $f$ vertically with each symbol aligned to a vertical set of edges and $g$ horizontally likewise. A diagonal edge in a small square exists if the corresponding $f$ and $g$ symbols are equal. The thick red edges form a path from the origin to the destination; this path corresponds to the sequence `0101' -- append the corresponding symbol at the left of an edge if it's vertical or diagonal, otherwise append the symbol at the top. It is also easy to verify that 0101 is a supersequence of both $f$ and $g$, and could be obtained as a covering of $f$ and $g$; the path itself gives one such covering. This covering also corresponds to an error event in the remnant channel which would result in outputs $f$ and $g$ with input $h=$ `0101' -- more precisely, the error event is one in which the first symbol of $h$ is deleted only in $g$, the second symbol is deleted only in $f$ and the third and fourth symbols not deleted in either $f$ or $g$.}
\label{fig:editgraph_smap1}
\end{figure}
Let $p=((i_1,j_1),(i_2,j_2),...,(i_m,j_m))$ be a path in $\mathcal{G}(f,g)$. We define $ s(p)$ to be the sequence corresponding to the path. Intuitively, $s(p)$ is formed by appending symbols in the following
way: append the corresponding $f$ symbol for a vertical edge, $g$ symbol for horizontal edge, and $f$ or $g$ symbol for diagonal edge (see example Fig.~\ref{fig:editgraph_smap1}). Any path from $(0,0)$ to $(|f|,|g|)$ corresponds to a supersequence of $f$ and $g$ and which is ``covered" by $f$ and $g$. More formally, define $ s(p)\triangleq x_1x_2...x_{m-1}$ where
$$x_k =
\begin{cases}
f_{i_{k+1}} \quad\text{if }j_{k}=j_{k+1},\\
g_{j_{k+1}} \quad\text{if }i_{k}=i_{k+1},\\
f_{i_{k+1}} \quad\text{else.}
\end{cases}
$$
The construct of edit graph can be extended to more than 2 sequences with the same idea. For sequences $f_1,f_2,...,f_t$, construct a $t$-dimensional grid with a number of vertices $(|f_1|+1)(|f_2|+1)...(|f_t|+1)$ labeled from $(0,0,...,0)$ to $(|f_1|,|f_2|,...,|f_t|)$. A vertex $u=(i_1,i_2,...,i_t)$ is connected to $v=(j_1,j_2,...,j_t)$ (we say $u \rightarrow v$) iff both of the following conditions are met:
\begin{itemize}
\item $j_l=i_l$ or $j_l=i_l+1$ $\forall\ l\in [t]$, i.e., $(i_1,...,i_t)$ and $(j_1,...,j_t)$ are vertices of a particular unit cube. Only these type of vertices can share an edge in the grid graph.
\item Let $\mathcal T \subseteq [t]$ be the collection of indices where $j_l=i_l+1$. Then $f_{i_l}$ is same $\forall\ l \in \mathcal T$. For instance, if $\mathcal T=\{1,3,5\}$, then $f_{i_1}=f_{i_3}=f_{i_5}$.
\end{itemize}
Define the vertex $(0,...,0)$ to be the origin of this graph and the vertex $(|f_1|,...,|f_t|)$ to be the destination. If $|f_j|=O(n)\ \forall\ j$, this graph has a number of vertices $O(n^t)$ and a maximum number of edges $O((2n)^t)$ since each vertex has at most $2^t$ outgoing edges.\\
\noindent\textbf{Infiltration product} (introduced in \cite{lothaire1997combinatorics}):
The infiltration product has been extensively used in \cite{lothaire1997combinatorics}, as a tool in non-commutative algebra. Here we give an edit-graph interpretation of this tool. We also give a formal definition later in this section. Using the edit graph we can construct the set of possible
supersequences $\mathcal{S}(f,g)$ of $f,g$ which are covered by
it. Clearly multiple paths could yield the same supersequence and we
can count the number of distinct ways $\mathbf N(h;f,g)$ one can construct
the same supersequence $h$ from $f,g$.
We can informally define the
\emph{infiltration product $f\uparrow g$} of $f$ and $g$, as a polynomial with monomials the supersequences $h$ in
$\mathcal{S}(f,g)$ and coefficients $\langle f\uparrow g,h\rangle$ equal to $\mathbf N(h;f,g)$. In Fig.~\ref{fig:editgraph_smap1}, it is easy to verify that there is exactly one path corresponding to `101001' and hence $\langle 001\uparrow 101,101001 \rangle=1$ and similarly $\langle 001\uparrow 101,01001 \rangle=2$. One could find these coefficients for all relevant sequences and form the polynomial as described. More examples: Let $\mathcal{A}=\{a,b\}$, then
\begin{itemize}[wide=2pt]
\item $ab\uparrow ab=ab+2aab+2abb+4aabb+2abab$,
\item $ab\uparrow ba=aba+bab+abab+2abba+2baab+baba.$
\end{itemize}
The infiltration operation is commutative and associative, and infiltration of two sequences $f\uparrow g$ is a polynomial with variables of length (or \textit{degree}) at most $|f|+|g|$; see \cite{lothaire1997combinatorics}.
The definition of infiltration extends to two polynomials via distributivity (precisely defined in later), and consequently to multiple sequences as well. For multiple sequences, infiltration has the same edit graph interpretation: $\langle f_1\uparrow f_2 \uparrow...\uparrow f_t, w \rangle$ is the number of distinct ways of constructing $w$ as a supersequence of $f_1, f_2, ... ,f_t$ so that the construction covers $w$, i.e., construct the $t$-dimensional edit graph of $f_1, f_2, ... ,f_t$ and count the number of paths corresponding to $w$.
\subsection{Formal definition of infiltration product}
We now give a more formal definition of the infiltration product (see \cite{lothaire1997combinatorics} for the equivalence of the two definitions and a more rigorous treatment). A \textit{formal series} with indeterminates (or variables) in a set $\mathcal A$ and coefficients in a commutative ring $\mathcal R$, is a mapping of $\mathcal A^*$ onto $\mathcal R$. Recall that a commutative ring is a set which forms an abelian group under an \textit{addition} operation, is a monoid under a \textit{multiplication} operation which commutes, and the multiplication operation distributes over addition. Here we consider $\mathbb Z$, the set of integers as the commutative ring $\mathcal{R}$. A formal series is called a \textit{polynomial} if only a finite number of sequences are mapped to non-zero values, the rest of the sequences map to zero. Consider two polynomials $\sigma,\tau: \mathcal{A}^*\rightarrow \mathbb Z$. The value taken by a sequence $w\in \mathcal A^*$ on $\sigma$ (or the coefficient of $w$ in $\sigma$) is denoted by $\langle \sigma,w\rangle \in \mathbb R$. We also define binary addition ($\oplus$) and multiplication operations ($\times$) on the set of polynomials as follows:
\begin{align}
\langle \sigma\oplus \tau,w\rangle \triangleq \langle \sigma,w\rangle + \langle \tau,w \rangle \quad \forall w\in \mathcal A^*,\label{eq:polynomial_add}\\
\langle \sigma\times \tau,w\rangle \triangleq \sum_{\substack{f,g\in \mathcal A^*:\\ f.g=w}}\langle \sigma,f\rangle \langle \tau,g \rangle \quad \forall w\in \mathcal A^*.\label{eq:polynomial_prod}
\end{align}
We will use the usual symbols $+$ and $.$ in place of $\oplus$ and $\times$ in this work for convenience. The meaning of the operation would be clear depending on the operands. With these operations the set of polynomials form a non-commutative ring, and is denoted by $\mathbb Z\langle\mathcal A \rangle$, also called the free $\mathbb Z$-algebra on $\mathcal A$ in ring theory. Note that the addition and multiplication operations defined in \eqref{eq:polynomial_add} and \eqref{eq:polynomial_prod} are similar to the operations defined on commutative polynomials, except that the multiplication operation under the summation in \eqref{eq:polynomial_prod} ($f.g=w$) is actually concatenation and is non-commutative. The multiplication inside the summation in \eqref{eq:polynomial_prod} is multiplication in the real field and hence commutative. It is also easy to prove that the multiplication defined in \eqref{eq:polynomial_prod} distributes over addition defined in \eqref{eq:polynomial_add}. Thus, a polynomial in $\mathbb Z\langle\mathcal A \rangle$ can be represented as a sum of monomials in $\mathcal A^*$ each with an associated coefficient in $\mathbb Z$, i.e., $\sigma=\sum\limits_{w\in \mathcal A^*} \langle\sigma,w \rangle w$. Define the \textit{degree} of a polynomial to be equal to the length of a longest sequence with a non-zero coefficient in the polynomial and the \textit{number of terms} of a polynomial as the number of sequences with non-zero coefficients in the polynomial. Note that a degree $d$ polynomial could have a number of terms upto $2^{d+1}-1$.
With this, the \textit{infiltration product} (in general, for two polynomials) is defined as follows:
\begin{align}
\forall f\in \mathcal{A}^*,& \quad f\uparrow e = e\uparrow f=f.\nonumber \\
\forall f,g\in \mathcal{A}^*&,\quad \forall a,b\in \mathcal{A}, \nonumber\\
fa\uparrow gb=(f\uparrow gb)a&+(fa\uparrow g)b+\mathbbm{1}_{a=b}(f\uparrow g)a.\nonumber \\
\forall \sigma,\tau \in \mathbb{Z}\langle\mathcal{A} \rangle, \quad &\sigma\uparrow \tau=\sum_{f,g\in \mathcal{A}^*} \langle \sigma,f \rangle \langle \tau,g \rangle (f\uparrow g). \label{def:infiltforseq}
\end{align}
\textbf{Summary of definitions and ideas introduced in this section:}
\begin{itemize}
\item The binomial coefficient captures the likelihood of observations for deletion channels.
\item An extension of the binomial coefficient function where one of the parameters can take real values has been introduced; this function is pivotal for our results on the single-trace deletion channel.
\item For multiple deletion channels, the error events can be categorized into two groups -- one where an input symbol is deleted in all the traces, and second, the complement of this event. This categorization results in a natural decomposition of multiple deletion channel model into a cascade model involving the remnant channel.
\item The remnant channel disregards the error events where a symbol is deleted in all the traces.
\item The edit graph provides a way to visualize all the possible error events and input sequences to a remnant channel given its outputs.
\item The infiltration product serves the same purpose as the edit graph, but has an algebraic flavor and provides rigor for proofs and analyses. The edit graph, on the other hand, is more helpful in designing reconstruction algorithms over deletion channels.
\end{itemize}
\section{Maximum likelihood for the single-trace deletion channel}
\label{sec:1deletion_ML}
Consider the ML formulation for the single-trace deletion channel given non-empty trace $Y$ and all possible $n$-length inputs allowed, restated here for convenience:
\begin{equation}
\argmax_{x\in \{0,1\}^n} {x \choose Y}.
\label{eq:1deletion_ML}
\end{equation}
To the best of our knowledge, the only known method to solve \eqref{eq:1deletion_ML} involves iterating over all possible choices of $x$ and computing the objective value for each of the choices. We here show that it is sufficient to solve a continuous relaxation of above problem to obtain a solution to \eqref{eq:1deletion_ML}. Note that there could be multiple solutions to \eqref{eq:1deletion_ML}. Before going in the main result, we first state a useful lemma which factors out a given coordinate $p_i$ out of $\mathbf F(p,Y)$. The proof of the lemma is relegated to the appendix.
\begin{restatable}{lemma}{deletionMLrellemma}
For $p=[p_1,p_2,..,p_i,...,p_n]$ and $Y=Y_1Y_2...Y_m$ with $n \geq m > 0$, we have
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
\label{lemma:F_decomposition}
\end{restatable}
\begin{theorem} An alternative optimization to characterize ML for the single-trace deletion channel.
\begin{equation}
\max_{x\in \{0,1\}^n} {x \choose Y} = \max_{p\in [0,1]^n} \mathbf F(p,Y).
\label{eq:ml_opt_equiv}
\end{equation}
Furthermore, given any non-integral $p^* \in [0,1]^n$ that maximizes $\mathbf F(p,Y)$, one could construct a corresponding integral solution $x^* \in \{0,1\}^n$ that maximizes $\mathbf F(p,Y)$ and consequently is also a solution to $\max_{x\in \{0,1\}^n} {x \choose Y}$.
\end{theorem}
\begin{proof}
As noted earlier, we have ${x \choose Y} = \mathbf F(x,Y)$. Therefore, we are interested in proving the following:
\begin{align*}
\max_{x\in \{0,1\}^n} \mathbf F(x,Y) \equiv \max_{p\in [0,1]^n} \mathbf F(p,Y).\numberthis
\label{eq:ml_opt_equiv_proof1}
\end{align*}
To show this and also the second statement, we prove that given $p=(p_1,p_2,...,p_i,...,p_n)$, at least one of the following holds true:
\begin{itemize}
\item$\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},0,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 0.
\item $\mathbf F(p^{(i\rightarrow 1)},Y) \geq \mathbf F(p,Y)$, where $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},1,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by 1.
\end{itemize}
Thus if $p^*$ is an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ with $p_i\in (0,1)$, then at least one of $p^{(i\rightarrow 0)}$ or $p^{(i\rightarrow 1)}$ is also an optimal solution. Sequentially applying this argument for each coordinate of $p$ shows that there exists a point in $\{0,1\}^n$ which is also an optimal solution to $\max_{p\in [0,1]^n} \mathbf F(p,Y)$ and consequently to $\max_{x\in \{0,1\}^n} \mathbf F(x,Y)$.
Now to prove our claim, we use Lemma~\ref{lemma:F_decomposition} to factor out $p_i$ terms in $\mathbf F(p,Y)$:
\begin{align*}
\mathbf F(p,Y) = \mathbf F( p_{[n]\backslash \{i\}},Y) + p_i \sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).
\end{align*}
There are 3 cases
\begin{enumerate}
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) = \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) = \mathbf F(p,Y) = \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) > \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \leq \mathbf F(p,Y) \leq \mathbf F(p^{(i\rightarrow 1)},Y).$
\item $$\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}) < \sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$$ \\In this case it is easy to verify that $\mathbf F(p^{(i\rightarrow 0)},Y) \geq \mathbf F(p,Y) \geq \mathbf F(p^{(i\rightarrow 1)},Y).$
\end{enumerate}
Thus in each of the 3 cases, we see that at least one of $\mathbf F(p^{(i\rightarrow 0)},Y)$ or $\mathbf F(p^{(i\rightarrow 1)},Y)$ is at least as large as $\mathbf F(p,Y)$ thus proving the theorem. Note that the proof gives a natural way to find an optimal lattice point $p_{lattice}^*$ given a non-lattice point $p^*$ by iterating through each coordinate of $p^*$ and switching them to $0$ or $1$ by comparing $\sum\limits_{k|Y_k=1}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]})$ with $\sum\limits_{k|Y_k=0}\mathbf F( p_{[1:i-1]}, Y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, Y_{[k+1,m]}).$
\end{proof}
\section{Symbolwise MAP for the single-trace deletion channel}
\label{sec:1deletion}
We here develop an algorithm to compute the symbolwise posterior probabilities for the single-trace deletion channel when the input symbols are independently generated with arbitrary priors. Consider the single deletion channel model in Fig.~\ref{fig:1deletion}, where $X=X_1...X_n$, each input symbol is generated $X_i \sim \text{ind. Ber}\ (p_i)$, and we observe the trace $Y = y = y_1y_2...y_m$ with $m\leq n$. Define the vector of priors as $p\triangleq(p_1,p_2,...,p_n)$. We first give an $O(n^2)$ algorithm to calculate the posterior probabilities $\Pr(X_i=1|Y=y)$, which in turn provides the symbolwise MAP estimate for the considered model. We then show how this algorithm can be used for trace reconstruction. We take three steps to present the algorithm.
\noindent \textbf{An expression for $\Pr(X_i=1|Y=y)$.}
Let $\Pr(X_i=1)=p_i$. As a first step, we have
\vspace{6pt}
\begin{align*}
\Pr(X_i=1|{Y=y}) &= \frac{\Pr(X_i=1,Y=y)}{\Pr(Y=y)}
= \frac{ \sum\limits_{\substack{ x| x_i=1}} \Pr({X=x}) \Pr(Y=y|X=x)}{ \sum_{\substack{x}} \Pr({X=x}) \Pr(Y=y|X=x)} \\
&\overset{(a)}{=} \frac{ \sum\limits_{\substack{ x| x_i=1}} \Pr({X=x}) { x \choose y}}{ \sum_{\substack{x}} \Pr({X=x}) { x \choose y}}, \numberthis
\label{eq:approx_smap_1}
\end{align*}
where $(a)$ is because for a deletion channel $\Pr(Y=y|X=x)={x \choose y} \delta^{|x|-|y|}(1-\delta)^{|y|}$.
To proceed, we need to evaluate the summation in the numerator and the denominator.
Theorem~\ref{thm:approx_smap_postprob} expresses \eqref{eq:approx_smap_1} in terms of relaxed binomial coefficient terms $\mathbf F(\cdot)$. Recall that $\mathbf F(p,y) \triangleq \mathbb E_{X\sim p} {X \choose y}$, which is the denominator term in \eqref{eq:approx_smap_1}.
\begin{theorem}
\label{thm:approx_smap_postprob}
Let $X=X_1...X_n$ where $X_i \sim \text{ind. Ber}\ (p_i)$, and let $Y=y$ be the observed trace when $X$ is passed through a deletion channel. Then,
\begin{align*}
\Pr(X_i&=1|Y=y) = \frac{p_i}{\mathbf F( p, y)} \left( \mathbf F( p_{[n]\backslash \{i\}}, y) + \sum\limits_{k|y_k=1}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]}) \right). \numberthis
\label{eq:smap_postprob_thm}
\end{align*}
\end{theorem}
\begin{proof}
The proof of this theorem employs the same trick used in the proof of Lemma~\ref{lemma:F_decomposition}. From \eqref{eq:approx_smap_1}, we have
\begin{align*}
\Pr(X_i = 1 | Y=y) = \frac{ \sum\limits_{\substack{ x| x_i=1}} \Pr({X=x}) { x \choose y}}{\mathbf F(p,y)}.
\end{align*}
Now,
\begin{align*}
\sum_{\substack{ x| x_i=1}} & \Pr({X=x}) { x \choose y}
=\sum_{\substack{ x|x_i=1}} \Pr({X=x}) \sum_{\substack{\mathcal S\subseteq [n]\\ |\mathcal S|=m}} \mathbbm{1}\{ x_{\mathcal S}= y\}\\
&=\sum_{\substack{\mathcal S\subseteq [n]\\ |\mathcal S|=m}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=y}} \Pr({X=x}).\numberthis \label{eq:smapiter1}
\end{align*}
We first separate the outer summation into two cases: (a) $\mathcal S|i \notin \mathcal S$ and (b) $\mathcal S|i\in \mathcal S$.
We can express the first case as
\begin{align*}
&\hspace{-1cm}\sum_{\substack{\mathcal S\subseteq [n] \\ |\mathcal S|=m,i\notin \mathcal S}}\sum_{\substack{ x|x_i=1\\x_{\mathcal S}=y}} \Pr({X=x})
=\sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=y}} \Pr({X=x})\\
&=\sum_{\substack{S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} \sum_{\substack{ x|x_i=1\\ x_{\mathcal S}= y}} \Big(\Pr(X_i=1)\Pr( X_{\mathcal S}= y) \Pr( X_{[n]\backslash \mathcal S\cup\{i\}}= x_{[n]\backslash \mathcal S\cup\{i\}}) \Big)\\
&=\sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} p_i \Pr( X_{\mathcal S}= y) \left(\sum_{\substack{ x|x_i=1\\ x_{\mathcal S}= y}} \Pr( X_{[n]\backslash \mathcal S\cup\{i\}}= x_{[n]\backslash \mathcal S\cup\{i\}})\right)\\
&=\sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} p_i \Pr( X_{\mathcal S}= y) \left(\sum_{(x_j|j\in [n]\backslash \mathcal S\cup \{i\})} \Pr( X_{[n]\backslash \mathcal S\cup\{i\}}= x_{[n]\backslash \mathcal S\cup\{i\}})\right)\\
&=p_i \sum_{\substack{\mathcal S\subseteq [n]\backslash \{i\}\\ |\mathcal S|=m}} \Pr( X_{\mathcal S}= y) = p_i \mathbf F( p_{[n]\backslash \{i\}}, y).\numberthis \label{eq:lemma3proof1}
\end{align*}
For the second term, we express the set $\mathcal S$ as a union $\mathcal S = \mathcal S' \cup \{i\} \cup \mathcal S''$ such that $\mathcal S' \subseteq [i-1]$ and $\mathcal S'' \subseteq [i+1:n]$ to get:
\begin{align*}
&\sum_{\substack{\mathcal S\subseteq [n]\\ |\mathcal S|=m,\\i\in \mathcal S}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=y}} \Pr({X=x})= \sum_{k=1}^m\sum\limits_{\substack{\mathcal S\subseteq [n],\\|\mathcal S|=m,\\ \mathcal S_k = i}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=y}} \Pr({X=x})\\
&=\sum_{k=1}^m\sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \sum_{\substack{ x|x_i=1\\x_{\mathcal S}=y}}\mathbbm{1}_{\{y_k=1\}} \Pr({X=x}) \\
&=\sum_{k:y_k=1}\sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}} \sum_{\substack{ x|x_i=1\\ x_{\mathcal S'}= y_{[1:k-1]}\\ x_{\mathcal S''}= y_{[k+1:m]}}} \Bigg ( \Pr(X_i=1) \Pr( X_{\mathcal S'}= y_{[1:k-1]})
\Pr( X_{\mathcal S''}= y_{[k+1:m]}) \\&\hspace{7cm} \Pr( X_{[n]\backslash \mathcal S'\cup \mathcal S''\cup \{i\}}= x_{[n]\backslash \mathcal S'\cup \mathcal S'' \cup\{i\}})\Bigg )\\
&=p_i\sum_{k:y_k=1}\Bigg ( \Big( \sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\Pr( X_{\mathcal S'}= y_{[1:k-1]})\Big) \Big(\sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}}\Pr( X_{\mathcal S''}= y_{[k+1:m]})\Big ) \\
& \hspace{5cm} \Big( \sum_{\substack{ x|x_i=1\\ x_{\mathcal S'}= y_{[1:k-1]}\\ x_{\mathcal S''}= y_{[k+1:m]}}} \Pr( X_{[n]\backslash \mathcal S'\cup \mathcal S''\cup \{i\}}= x_{[n]\backslash \mathcal S'\cup \mathcal S'' \cup\{i\}})\Big) \Bigg )\\
&=p_i\sum_{k|y_k=1}\Bigg(\Big( \sum_{\substack{\mathcal S'\subseteq [i-1]\\ |\mathcal S'|=k-1}}\Pr( X_{\mathcal S'}= y_{[1:k-1]})\Big) \Big( \sum_{\substack{\mathcal S''\subseteq [i+1:n]\\ |\mathcal S''|=m-k}}\Pr( X_{\mathcal S''}=y_{[k+1:m]}) \Big)\Bigg)\\
&=p_i \sum\limits_{k|y_k=1}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]}). \numberthis \label{eq:lemma3proof2}
\end{align*}
Plugging in \eqref{eq:lemma3proof1} and \eqref{eq:lemma3proof2} in \eqref{eq:approx_smap_1} proves the theorem.
\end{proof}
Alg.~\ref{alg:apprx_smap_dp} summarizes the computation of $\Pr(X_i=1|Y=y)$.
\begin{algorithm}[t!]
\caption{Symbolwise posterior probabilities with one trace}\label{alg:apprx_smap_dp}
\begin{algorithmic}[1]
\item Input: Trace {$Y=y$}, priors $p$\\ Outputs: Posteriors $\Pr(X_i=1|Y=y)\ \forall\ i$
\State Compute $\mathbf F(p_{[1:k]},y_{[1:j]})\ \forall\ k,j$ and $\mathbf F(p_{[k:n]},y_{[j:m]})\ \forall\ k,j$ via Alg.~\ref{alg:F_comp}
\For {$i=1:n$}
\State Use \eqref{eq:smap_postprob_thm} to compute $\Pr(X_i=1|Y=y)$
\EndFor
\end{algorithmic}
\end{algorithm}
\noindent\textbf{A trace reconstruction heuristic with $t$ traces.}
The posterior probability computation in Alg.~\ref{alg:apprx_smap_dp} naturally gives rise to a trace reconstruction heuristic that updates the symbolwise statistics sequentially on the traces, where we use Alg.~\ref{alg:apprx_smap_dp} with one trace at a time to continually update $\Pr(X_i=1|Y=y)$. The overall heuristic is described in Alg.~\ref{alg:apprx_smap}. Note that the algorithm first needs to compute $\mathbf F(p_{[1:k]},y_{[1:j]})\ \forall\ k,j$ and $\mathbf F(p_{[k:n]},y_{[j:m]})\ \forall\ k,j$ which requires $O(n^2)$ operations, as described in Appendix~\ref{app:F_compute}. Given this, the algorithm iterates over the $n$ indices and computes the posteriors in $O(n)$ for each of the index. Thus, the complexity of the algorithm is $O(n^2)$; note that $m=O(n)$ since $y$ is a deleted version of the input.
\begin{algorithm}[t!]
\caption{Trace reconstruction via iterative single-trace posterior probabilities}\label{alg:apprx_smap}
\begin{algorithmic}[1]
\item Input: Traces {$Y^{1}=y^1,...,Y^{t}=y^t$}, input length $n$ \\ Outputs: Estimate of the input $\hat X$
\State Initialize priors $p^{old}=p^{new} \gets (0.5,0.5,...,0.5)$
\For {$l=1:t$}
\State Use Alg.~\ref{alg:apprx_smap_dp} with $p^{old}$ and $y^{l}$ to update $p^{new}$
\State $p^{old}\gets p^{new}$
\EndFor
\For {$i=1:n$}
\If {$p^{new}_i\geq 0.5$}
$\ \hat X_i \gets 1$
\Else
$\ \hat X_i \gets 0$
\EndIf
\EndFor
\State \textbf{return} $\hat X_1 \hat X_2 ... \hat X_n$
\end{algorithmic}
\end{algorithm}
\section{Sequencewise ML for the deletion channel}
\label{sec:1deletion_ML}
\subsection{A continuous optimization formulation for the single trace ML}
We here consider the single-trace ML decoding in (\ref{eq:ML_deletion}),
assuming that the output sequence $Y=y$ is non-empty.
To the best of our knowledge, the only known method to solve \eqref{eq:ML_deletion} involves solving a combinatorial optimization, essentially iterating over all possible choices of $x$ and computing the objective value for each of the choices.
The reason is that there seems to be no discernible pattern exhibited by the true ML sequence; as we see in the table below, the true ML sequence at times extends a few runs, and at times even introduces new runs! Here, we list a few examples of the trace and the corresponding 10-length ML sequences.
\begin{center}
\begin{tabular}{ | c | c| }
\hline
$y$ & The set of all $x_{ml}$ sequences \\
\hline
10111 & 1100111111 \\
\hline
1010 & 1101010100 \\
\hline
000100 & 0000001000, \quad 0000010000, \quad 0000011000 \\
\hline
111101 & 1111111001,\quad 1111111011 \\
\hline
\end{tabular}
\end{center}
\vspace{5mm}
In this section, we show that one could equivalently solve the continuous relaxation of \eqref{eq:ML_deletion} to obtain a solution for \eqref{eq:ML_deletion}.
Before presenting the main result, we first state a useful lemma which factors a given coordinate $p_i$ out of the relaxed binomial coefficient $\mathbf F(p,y)$ we introduced in Definition~\ref{def:f}.
\begin{restatable}{lemma}{deletionMLrellemma}
For $p=(p_1,p_2,..,p_i,...,p_n)$ and $Y=y=y_1...y_m$ with $n \geq m > 0$, we have
\begin{align*}
\mathbf F(p,y) = \mathbf F( p_{
[n]\backslash \{i\}},y) + p_i \sum\limits_{k|y_k=1}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|y_k=0}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]}).
\end{align*}
\label{lemma:F_decomposition}
\end{restatable}
Recall that $\mathbf F(p,y)$ sums over all $m$-length subsets $\mathcal S$ and associates $p_{\mathcal S}$ with $y.$ Intuitively, this recursive relationship considers separately the cases where
\begin{itemize}
\item $i \notin \mathcal S$,
\item $i \in \mathcal S$ and is associated with a particular $y_k$ where $y_k = 1$,
\item $i \in \mathcal S$ and is associated with a particular $y_k$ where $y_k = 0$.
\end{itemize}
The detailed proof can be found in Appendix~\ref{app:F_lemma_proof}. It is clear from Lemma~\ref{lemma:F_decomposition} that $\mathbf F(p,y)$ is affine when projected onto each coordinate $p_i$. Thus, the extrema of $\mathbf F(p,y)$ must occur at the boundary of the support set of $p_i$; i.e., at either $p_i = 0$ or $p_i = 1$. Combining this with the fact that $\mathbf F(\cdot)$ is a relaxed version of the binomial coefficient, we observe that the maximization problem in \eqref{eq:ML_deletion} is equivalent to its real-valued relaxation. The following result makes this precise.
\begin{theorem} The ML decoding problem for the single-trace deletion channel
\begin{equation}
\max_{x\in \{0,1\}^n} {x \choose y}
\label{eq:ml_opt_equiv1}
\end{equation}
is equivalent to the problem
\begin{equation}
\max_{p\in [0,1]^n} \mathbf F(p,y).
\label{eq:ml_opt_equiv2}
\end{equation}
Furthermore, given any non-integral $p^* \in [0,1]^n$ that maximizes $\mathbf F(p,y)$, we can construct a corresponding integral solution $x^* \in \{0,1\}^n$ that maximizes $\mathbf F(x,y)$ and consequently also maximizes ${x \choose y}$.
\label{thm:ML_relaxation}
\end{theorem}
\begin{proof}
As noted earlier, we have ${x \choose y} = \mathbf F(x,y)$. Therefore, we are interested in proving the following:
\begin{align*}
\max_{x\in \{0,1\}^n} \mathbf F(x,y) \equiv \max_{p\in [0,1]^n} \mathbf F(p,y),\numberthis
\label{eq:ml_opt_equiv_proof1}
\end{align*}
where $\equiv$ refers to that the two problems are equivalent (have the same optimal objective value).
We prove this by applying the following claim.\\
\textbf{Claim:} Given any feasible $p=(p_1,p_2,...,p_i,...,p_n)$, at least one of the following holds true:
\begin{itemize}
\item$\mathbf F(p^{(i\rightarrow 0)},y) \geq \mathbf F(p,y)$. Recall from notation that $p^{(i\rightarrow 0)}=(p_1,p_2,...,p_{i-1},0,p_{i+1}...,p_n)$ is the vector where the $i^{th}$ coordinate is replaced by $0$.
\item $\mathbf F(p^{(i\rightarrow 1)},y) \geq \mathbf F(p,y)$.
\end{itemize}
Thus if $p^*$ is an optimal solution to \eqref{eq:ml_opt_equiv2} with $p_i\in (0,1)$, then at least one of $p^{(i\rightarrow 0)}$ or $p^{(i\rightarrow 1)}$ is also an optimal solution. Sequentially applying this argument for each coordinate of $p$ shows that there exists a point in $\{0,1\}^n$ which is an optimal solution to \eqref{eq:ml_opt_equiv2} and consequently to \eqref{eq:ml_opt_equiv1}.
It remains to prove our claim. We use Lemma~\ref{lemma:F_decomposition} to factor out $p_i$ terms in $\mathbf F(p,Y)$:
\begin{align*}
\mathbf F(p,y) = \mathbf F( p_{[n]\backslash \{i\}},y) + p_i \sum\limits_{k|y_k=1}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]})\\ + (1-p_i) \sum\limits_{k|y_k=0}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]}).
\end{align*}
Now we express $\mathbf F(p^{(i\rightarrow 0)},y)$ and $\mathbf F(p^{(i\rightarrow 1)},y)$ as
$$\mathbf F(p^{(i\rightarrow 0)},y) = \mathbf F( p_{[n]\backslash \{i\}},y) + \sum\limits_{k|y_k=0}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]}),$$
$$\mathbf F(p^{(i\rightarrow 1)},y) = \mathbf F( p_{[n]\backslash \{i\}},y) + \sum\limits_{k|y_k=1}\mathbf F( p_{[1:i-1]}, y_{[1:k-1]})\mathbf F( p_{[i+1:n]}, y_{[k+1,m]}).$$
\noindent Because $0\leq p_i\leq 1$ it directly follows that $$\min \left\{\mathbf F(p^{(i\rightarrow 0)},y),\mathbf F(p^{(i\rightarrow 1)},y)\right\} \leq \mathbf F(p,y) \leq \max \left\{\mathbf F(p^{(i\rightarrow 0)},y),\mathbf F(p^{(i\rightarrow 1)},y)\right\},$$
thus proving our claim.
\end{proof}
The real-valued optimization problem in \eqref{eq:ml_opt_equiv2} falls under the umbrella of signomial optimization which is, in general, NP-hard (see for example, \cite{xu2014signomial}, \cite{chand2016signomial}). A standard technique for signomial optimization uses convexification strategies to approximate the optimal value. In particular,
as stated in \cite{chand2016signomial}, the main observation underlying their methods is that certifying the nonnegativity of a signomial with at most \textit{one negative coefficient} can be accomplished efficiently. However, there are two problems with this approach in relation to our work -- 1. when expressed as a signomial optimization problem, \textit{all} the coefficients are negative in the ML optimization objective function, and 2. the objective function has an exponential number of signomial terms as can be seen from Definition~\ref{def:f}. As a result, such strategies turn out to not be useful for the ML optimization problem. For instance, the techniques in \cite{chand2016signomial} resulted in the bound $\mathbf F(p,Y) \leq {|p| \choose |Y|}$ for most instances of $p$ and $Y$, where $|\cdot|$ denotes the length of the vector/sequence. This is a trivial bound that uses no information about $p$ and $Y$ other than their lengths. Moreover, with a slight change of variables, \eqref{eq:ml_opt_equiv2} could also be expressed as a maximization of a convex function in a convex set. With that being said, it is still unclear if \eqref{eq:ml_opt_equiv2} is solvable in polynomial time or not.
\subsection{ML via gradient ascent}
Given the continuous variable formulation of the ML problem in \eqref{eq:ml_opt_equiv2}, a natural heuristic to find an estimate of the ML sequence is to employ \textit{projected gradient ascent} to solve \eqref{eq:ml_opt_equiv2}. The algorithm, in short, can be described as follows (the exact algorithm is detailed as Alg.~\ref{alg:grad_asc}):\\
Step I: Start from a randomly chosen interior point (in our case, we start from $p=(0.5,0.5,...,0.5)$, the point corresponding to the uniform distribution).\\
Step II: Take a small step in the direction of the gradient $\nabla_p\ \mathbf F(p,y)$.\\
Step III: If the gradient step results in $p$ moving out of $[0,1]^n$, project it back onto $[0,1]^n$. Repeat Steps II and III until convergence.\\
Step IV: From the final $p$, determine the closest binary sequence to be the reconstructed sequence.
Moreover in Appendix~\ref{app:F_grad_comp}, we show using Lemma~\ref{lemma:F_decomposition} that $\nabla_p\ \mathbf F(p,y)$ can be computed in $O(n^2)$ as a ``by-product'' of computing $\mathbf F(p,y)$.
\begin{algorithm}[t!]
\caption{Single trace projected gradient ascent for ML}\label{alg:grad_asc}
\begin{algorithmic}[1]
\item Input: Blocklength $n$, Trace {$Y=y$}, Initial point $p = (p_1,p_2,...,p_n)$, step-size $\epsilon$, Max iterations $M$, Convergence criteria $C$ \\ Outputs: Estimated sequence $\hat X$
\State Iteration count $j=0$
\While {$C$ is FALSE and $j<M$}
\State $p \leftarrow p + \epsilon \frac{\nabla_p \mathbf F(p,y)}{\mathbf F(p,y)}$
\State Replace $p_i \leftarrow 1$ for all $i:p_i > 1$
\State Replace $p_i \leftarrow 0$ for all $i:p_i < 0$
\State $j\leftarrow j+1$
\EndWhile
\State For each $i$, set $\hat X_i = \mathbbm 1 \{p_i>0.5\}$.
\State \textbf{return} $\hat X = \hat X_1 \hat X_2 ... \hat X_n$
\end{algorithmic}
\end{algorithm}
\newpage
\subsection{A heuristic for multiple traces}
The continuous variable ML formulation in \eqref{eq:ml_opt_equiv2} optimizes over the distributions $p$, instead of sequences $x$. In particular, we proved the following:
\begin{equation*}
\max_{x\in \{0,1\}^n} {x \choose y} \equiv \max_{p\in [0,1]^n} \mathbf F(p,y) \equiv\ \max_{p\in [0,1]^n}\ \mathbb E_{Z \sim p} {Z \choose y}.
\end{equation*}
At this point, one could ask how this formulation extends to multiple traces $Y^1=y^1,Y^2=y^2,...,Y^t=y^t$. The following theorem gives such a continuous optimization formulation with multiple traces.
\begin{theorem} The ML decoding with multiple traces
\begin{equation}
\max_{x\in \{0,1\}^n} {x \choose y^1}{x \choose y^2}...{x \choose y^t}
\label{eq:ml_opt_traces1}
\end{equation}
is equivalent to
\begin{equation}
\max_{p\in [0,1]^n} \mathbb E_{Z\sim p} \left[{Z \choose y^1}{Z \choose y^2}...{Z \choose y^t} \right].
\label{eq:ml_opt_traces2}
\end{equation}
Furthermore, given any non-integral $p^* \in [0,1]^n$ that maximizes $\mathbb E_{Z\sim p} \left[{Z \choose y^1}{Z \choose y^2}...{Z \choose y^t} \right]$, we can construct a corresponding integral solution $x^* \in \{0,1\}^n$ that also maximizes ${x \choose y^1}{x \choose y^2}...{x \choose y^t} $.
\label{thm:ML_multi_traces}
\end{theorem}
\begin{proof}
This theorem can be proved in the same way as Theorem~\ref{thm:ML_relaxation}, by showing that \\ $\mathbb E_{Z\sim p} \left[{Z \choose y^1}{Z \choose y^2}...{Z \choose y^t} \right]$ is an affine function of each $p_i$; here we only prove this fact and the rest of the arguments follow exactly as in the proof of Theorem~\ref{thm:ML_relaxation}.
To show this we use Lemma~\ref{lemma:bin_inf_relation} stated below; this Lemma is also closely related to the channel equivalence of Theorem~\ref{thm:channel_equiv} (see Appendix~\ref{app:bin_inf_lemma}).
\begin{restatable}{lemma}{bininfrelation}
\label{lemma:bin_inf_relation}
For $h,f_1,f_2,...,f_m \in \mathcal{A}^*$,\\
$${h \choose f_1} {h \choose f_2}...{h \choose f_m}=\sum_{w\in \mathcal{A}^*}\langle f_1\uparrow f_2\uparrow ...\uparrow f_m,w \rangle{h \choose w}.$$
\end{restatable
Using Lemma~\ref{lemma:bin_inf_relation}, we now have
\begin{align*}
\mathbb E_{Z\sim p} \left[{Z \choose y^1}{Z \choose y^2}...{Z \choose y^t}\right] &= \mathbb E_{Z\sim p} \sum_{w\in \mathcal{A}^*}\langle y^1\uparrow y^2\uparrow ...\uparrow y^t,w \rangle{Z \choose w}\\
&= \sum_{w\in \mathcal{A}^*}\langle y^1\uparrow y^2\uparrow ...\uparrow y^t,w \rangle \mathbb E_{Z\sim p} {Z \choose w}\\
& = \sum_{w\in \mathcal{A}^*}\langle y^1\uparrow y^2\uparrow ...\uparrow y^t,w \rangle \mathbf F(p,w).
\end{align*}
Note that $\mathbf F(p,w)$ is affine in each $p_i$. Thus $\mathbb E_{Z\sim p} \left[{Z \choose y^1}{Z \choose y^2}...{Z \choose y^t}\right]$ is a linear combination of affine functions of each $p_i$, and hence is also affine in each $p_i$.
\end{proof}
The formulation of \eqref{eq:ml_opt_traces2}, by itself, is not very useful as it is unclear on how to efficiently compute $\mathbb E_{Z\sim p} \left[{Z \choose y^1}{Z \choose y^2}...{Z \choose y^t}\right]$. Indeed, if ${Z \choose y^i}\ \raisebox{0.05em}{\rotatebox[origin=c]{90}{$\models$}}\ {Z \choose y^j}$, the expectation of products would decompose into the product $\prod_{j} \mathbb E_{Z\sim p} {Z \choose y^j} = \prod_{j} \mathbf F(p,y^j)$, and each of the terms in the product can be computed in $O(n^2)$ as detailed in Appendix~\ref{app:F_compute} -- this is however not the case as ${Z \choose y^i}$ and ${Z \choose y^j}$ are not independent.
Having said that, we can now solve the maximization problem $\argmax_{p\in [0,1]^n} \prod_{j=1}^t \mathbf F(p,y^j)$ and hope that the resultant solution is also a good solution for $\argmax_{p\in [0,1]^n} \mathbb E_{Z\sim p} \left[{Z \choose y^1}...{Z \choose y^t}\right]$; Algorithm~\ref{alg:grad_asc_traces} makes this idea precise. Moreover, instead of maximizing $\prod_{j=1}^t \mathbf F(p,y^j)$, we can further simplify the gradient computations by taking the log of the objective function, i.e., we solve $\argmax_{p\in [0,1]^n} \sum_{j=1}^t \log \mathbf F(p,y^j)$. This heuristic turns out to perform well in a variety of situations, as illustrated in Section~\ref{sec:Numerics}. As for the complexity, note that Alg.~\ref{alg:grad_asc_traces} involves the computation of $t$ gradients (each of which takes $O(n^2)$) at each gradient iteration. For a fixed number of max iterations $M$, the complexity of the algorithm is $O(n^2t)$.
\begin{algorithm}[t!]
\caption{Trace reconstruction heuristic via projected gradient ascent}\label{alg:grad_asc_traces}
\begin{algorithmic}[1]
\item Input: Blocklength $n$, Traces {$Y^1=y^1,Y^2=y^2,...,Y^t=y^t$}, Initial point $p = (p_1,p_2,...,p_n)$, step-size $\epsilon$, Max iterations $M$, Convergence criteria $C$ \\ Outputs: Estimated sequence $\hat X$
\State Iteration count $j=0$
\While {$C$ is FALSE and $j<M$}
\State $p \leftarrow p + \epsilon \sum_{j=1}^t \frac{\nabla_p \mathbf F(p,y^j)}{\mathbf F(p,y^j)}$
\State Replace $p_i \leftarrow 1$ for all $i:p_i > 1$
\State Replace $p_i \leftarrow 0$ for all $i:p_i < 0$
\State $j\leftarrow j+1$
\EndWhile
\State For each $i$, set $\hat X_i = \mathbbm 1 \{p_i>0.5\}$.
\State \textbf{return} $\hat X = \hat X_1 \hat X_2 ... \hat X_n$
\end{algorithmic}
\end{algorithm}
| b49aee9a211c76414de4737c4e404431a9f7bb86 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Computer-aided thoracic disease diagnosis based on chest X-ray has been
significantly advanced by deep convolutional neural networks (DCNNs) in recent
years~\cite{wang2017chestx,tang2018attention,sedai2018deep,wang2018chestnet}.
Most of these approaches are formulated as a multi-task binary classification
problem, where a CNN is trained to predict the risk probabilities of different
thoracic diseases. In clinical practices, visual localization of the
lesions on chest X-ray, such as heatmaps or segmentation masks,
is also preferred to provide interpretable supports for the classification results.
Precise lesion localization often requires training CNNs with strong supervision,
such as bounding boxes~\cite{wang2017chestx}, beyond merely image-level labels.
However, accurately annotating lesion locations is difficult, time-consuming, and
infeasible to practice in large-scale. For example, one of the largest
publicly available chest X-ray datasets ChestX-ray14~\cite{wang2017chestx} contains
more than one hundred thousands images with image-level labels, among which
only less than one thousand images are further annotated with bounding boxes for
benchmarking. Therefore, weakly supervised lesion localization on chest X-ray based
on image-level labels remains a challenge but vital problem for computer-aided
thoracic disease diagnosis.
The recent work of Class Activation Map (CAM)~\cite{zhou2016learning} demonstrates
the excellent localization ability of CNNs trained on nature images with only image-level
supervision. On chest X-ray images, CAM and its variations
have also been used for lesions localization~\cite{wang2017chestx,tang2018attention,sedai2018deep,wang2018chestnet}.
However, most of these approaches utilize CAM as a post-processing technique to
first generate lesion localization heatmaps, then threshold the heatmap scores
and generate lesion bounding boxes. We argue that it may be beneficial to leverage
CAM even during training given its excellent localization ability.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{fig1.png}
\end{center}
\caption{The framework of Probabilistic-CAM (PCAM) pooling.}
\label{fig1}
\end{figure}
In this work, we propose a novel and simple extension to the CAM-based framework for lesion
localization on chest X-ray with image-level supervision. Specifically, we propose
a new global pooling operation that explicitly leverages CAM for localization during
training in a probabilistic fashion, namely Probabilistic-CAM (PCAM) pooling. Figure~\ref{fig1}
shows the framework of PCAM pooling. A fully convolutional backbone network first
processes the input chest
X-ray image and generates a feature map. Then, for a particular label of thoracic
disease, e.g. ``Pneumonia'', each feature embedding within the feature map goes through a fully
connected (fc) layer implemented as a $1 \times 1$ convolutional layer and generates
the class activation score that monotonically measure the disease likelihood
of each embedding. Unlike the standard practice that directly uses the class activation score for localization, we
further bound it with the sigmoid function and interpret the output as
the disease probability of each embedding. Finally, the output probability
map is normalized to attention weights of each embedding, following the
multiple-instance learning (MIL) framework~\cite{ilse2018attention,yao2018weakly},
which are used to pool the original feature map by weighted average pooling.
The pooled embedding goes through
the same fc layer introduced above and generates the image-level disease probability
for training. During inference time, we directly use the probability map for lesion
localization, and apply a simple probability thresholding to obtain disease regions
and bounding boxes.
PCAM pooling does not introduce any additional training parameters and is easy
to implement. We experiment the efficacy of PCAM pooling for lesion localization
with image-level supervision on the ChestX-ray14~\cite{wang2017chestx} dataset. A
ResNet-34~\cite{he2016deep} model trained with PCAM pooling significantly outperforms the ChestX-ray14
baseline\cite{wang2017chestx} in both the classification task and the localization
task. Qualitative visual examination shows the probability maps generated by PCAM
pooling tend to have clear and sharp boundaries around lesion regions compared
to the typical class activation map.
\section{Related work}
\subsection{Weakly supervised lesion localization}
Various methods have been proposed to localize lesions on chest X-ray through
image-level supervision~\cite{wang2017chestx,tang2018attention,sedai2018deep,yao2018weakly,wang2018chestnet}.
Wang et al.~\cite{wang2017chestx} introduce the ChestX-ray14 dataset, together
with a baseline for evaluating weakly supervised lesion
localization. Instead of the typical global average pooling, Wang et al. use
the Log-Sum-Exp (LSE) pooling~\cite{pinheiro2015image} to encourage model training
focusing more on the discriminative regions of the chest X-ray. Sedai et al.~\cite{sedai2018deep}
utilize the intermediate feature maps from CNN layers at different scales together
with learned layer reference weights to improve the localization performance of
small lesions. Tang et al.~\cite{tang2018attention}
combine CNN with attention-guided curriculum learning to gradually learn distinctive
convolutional features for localization. Most of these
approaches utilize the standard CAM technique to localize lesions, where our
proposed PCAM pooling serves as an extension to the standard CAM to improve lesion localization
with image-level supervision.
\subsection{Global pooling}\label{sec2.2}
Global average pooling is arguably the most widely used global pooling operation
for image classification. While it is less prone to overfitting as the network
tries to identify all discriminative regions of an object in natural images~\cite{zhou2016learning},
it may also fail to highlight the most discriminative regions within a chest X-ray,
given most chest X-ray images share the same anatomic structures and may only differ
in fine-grained lesion regions. Therefore, different global pooling operations
have also been used to analyze chest X-rays. For example, Wang et al.~\cite{wang2017chestx} uses LSE pooling,
which can be viewed as an adjustable operation between max pooling and average
pooling by controlling the hyper-parameter $\gamma$.
We summarize the mathematical differences and correlations between different
global pooling operations in Table~\ref{tab1}. Particularly, $X$ with shape
$(C, H, W)$ denotes the feature map from the last convolutional layer of a CNN.
$x$ denotes the feature embedding of length $C$ after global pooling. $C$ denotes
the channel dimension, $H$ and $W$ denote the height and width of the feature map.
\begin{table}
\centering
\caption{Different types of global pooling.}\label{tab1}
\begin{tabular}{|l|l|}
\hline
Pooling type & Formulation\\
\hline
Average & $x_c = \sum_{i,j}^{H,W} w_{c,i,j} X_{c,i,j} ,\ w_{c,i,j} = \frac{1}{H*W}$ \\
Linear~\cite{wang2019comparison} & $x_c = \sum_{i,j}^{H,W} w_{c,i,j} X_{c,i,j} ,\ w_{c,i,j} = \frac{X_{c,i,j}}{\sum_{i,j}^{H,W} X_{c,i,j}}$ \\
Exponential~\cite{wang2019comparison} & $x_c = \sum_{i,j}^{H,W} w_{c,i,j} X_{c,i,j} ,\ w_{c,i,j} = \frac{\exp(X_{c,i,j})}{\sum_{i,j}^{H,W} \exp(X_{c,i,j})}$ \\
LSE~\cite{pinheiro2015image} & $x_c = \frac{1}{\gamma} \log\left[ \frac{1}{H*W} \sum_{i,j}^{H,W} \exp(\gamma X_{c,i,j}) \right]$ \\
LSE-LBA~\cite{yao2018weakly} & $s = \frac{1}{\gamma_0 + \exp(\beta)} \log\left[ \frac{1}{H*W} \sum_{i,j}^{H,W} \exp\left[(\gamma_0 + \exp(\beta)) S_{i,j}\right] \right]$ \\
Attention~\cite{ilse2018attention} & $x = \sum_{i,j}^{H,W} w_{i,j} X_{i,j} ,\ w_{i,j} = \frac{\exp\left[\mathbf{w}^\intercal\tanh(\mathbf{V} X_{i,j})\right]}{\sum_{i,j}^{H,W} \exp\left[\mathbf{w}^\intercal\tanh(\mathbf{V} X_{i,j})\right]} $ \\
\hline
\end{tabular}
\end{table}
We can see that, global pooling operations differ mainly in the ways to compute the
weights for feature map averaging. Note that, most of the pooling is performed for each
channel $X_{c,i,j}$ independently, except for Attention pooling~\cite{ilse2018attention} that
computes the attention weight at embedding level $X_{i,j}$ with extra trainable parameters,
i.e. $\mathbf{V}, \mathbf{w}$ in Table~\ref{tab1}. All the channels within the
same embedding share the same weight. The basic assumption of Attention
pooling follows the multiple-instance learning (MIL) framework, that treats each
embedding as an instance, and the chest X-ray as a bag of instances is positive,
e.g. certain thoracic disease, as long as one of the instance is positive.
PCAM pooling follows the same MIL framework, but uses a different
method to compute the attention weight for each embedding based on
the localization ability of CAM without introducing extra trainable parameters.
LSE-LBA~\cite{yao2018weakly} also falls within the MIL framework, but it performs
the pooling operation on the saliency map $S$ of shape $(H, W)$ instead of the
feature map $X$ to obtain a saliency score $s$ for training.
\section{Probabilistic-CAM (PCAM) Pooling}
The main idea of PCAM pooling is to explicitly leverage the localization ability
of CAM~\cite{zhou2016learning} through the global pooling operation during training.
Using the same notation from Section~\ref{sec2.2}, given a fully convolutional
network trained for multi-task binary classification, the class activation map of
a particular thoracic disease is given by $\{s_{i,j} = \mathbf{w}^\intercal X_{i,j} + b|i,j \in H,W\}$.
$X_{i,j}$ is the feature embedding of length $C$ at the position $(i, j)$
of a feature map $X$ with shape $(C, H, W)$ from the last convolutional layer.
$\mathbf{w}, b$ are the weights and bias of the last fc layer for binary
classification. In other words, $s_{i,j}$ is the logit before sigmoid function
under the binary classification setting. $s_{i,j}$ monotonically measures the
disease likelihood of $X_{i,j}$, and is used to generate the localization heatmap
after the model is trained in the standard CAM framework.
PCAM pooling utilizes $s_{i,j}$ to guide lesion localization during training
through the global pooling operation under the MIL framework~\cite{ilse2018attention}.
The MIL framework assumes the chest X-ray as a bag of embeddings is positive, as
long as one of the embedding is positive. To measure each embedding's contribution
to the whole bag, the MIL framework assigns normalized attention weights to each embedding for
weighted global average pooling~\cite{ilse2018attention}. Because the numerical
range of $s_{i,j}$ is unbounded, i.e. $s_{i,j} \in (-\infty, +\infty)$ in theory,
it's neither interpretable nor directly applicable to compute the attention weights.
Therefore, we further bound $s_{i,j}$ with the sigmoid function, $p_{i,j} = \text{sigmoid}(s_{i,j})$,
and normalize $p_{i,j}$ as the attention weights. In summary, PCAM
pooling can be formulated as
\begin{eqnarray}\label{eq1}
x = \sum_{i,j}^{H,W} w_{i,j} X_{i,j} ,\ w_{i,j} = \frac{\text{sigmoid}(\mathbf{w}^\intercal X_{i,j} + b)}{\sum_{i,j}^{H,W} \text{sigmoid}(\mathbf{w}^\intercal X_{i,j} + b)}
\end{eqnarray}
where $w_{i,j}$ is the attention weight for $X_{i,j}$ and $x$ is the pooled
feature embedding which goes through the same fc layer for final image level classification.
During the inference time, we interpret the sigmoid-bounded $p_{i,j}$ as the
probability of embedding $X_{i,j}$ being positive, thus named Probabilistic-CAM,
and we use the probability map $\{p_{i,j}|i,j \in H,W\}$ as the localization heatmap.
We use a simple probability thresholding on the probability map to obtain regions
of interest. In comparison, because $s_{i,j}$ is unbounded with different numerical
ranges in different chest X-ray images, it is usually normalized into $[0, 255]$
within each image and thresholded with some ad-hoc ranges, e.g. $[60, 180]$
in~\cite{wang2017chestx}, to generate regions of interest. We show in Section~\ref{sec4.3}
section that PCAM pooling generates localization heatmaps with better visual quality
around lesion boundary regions compared to the standard CAM.
\section{Experiments}
\subsection{Dataset and experiments setup}\label{sec4.1}
We evaluate lesion localization with image-level supervision on the ChestX-ray14~\cite{wang2017chestx}
dataset, which contains 112,120 frontal-view chest X-ray images with 14 thoracic
disease labels. 8 out of the 14 diseases are further annotated with 984 bounding
boxes. We randomly split the official train\_valid set into $75\%$ for
training and $25\%$ for validation. On the official test set, we evaluate the
classification task on the 14 diseases and the localization task on the 8 diseases
that have bounding boxes. Note that the 984 bounding boxes are not used for training.
We use ResNet-34~\cite{he2016deep} as the backbone network and process
the input images on the original $1024 \times 1024$ scale following~\cite{wang2017chestx}.
The network is trained with a batch size of 36 and a learning rate of $1e^{-4}$
for 10 epoches. We balance the binary cross entropy loss of positive and negative
samples within each batch following~\cite{wang2017chestx}. For the localization
task, we first apply a probability of 0.9 to threshold the probability maps
from PCAM pooling, then generate bounding boxes that cover the isolated regions
in the binary masks. To compare the visual quality of localization heatmaps from
PCAM pooling with previous methods, we also train a ResNet-34 with LSE pooling
following~\cite{wang2017chestx}. We normalize the class activation maps from LSE
pooling into $[0, 255]$ for each image individually, and then apply a threshold
of 180 following~\cite{wang2017chestx}. Note that the performances of LSE pooling
reported in Table~\ref{tab2} and Table~\ref{tab3} are from Wang et al.~\cite{wang2017chestx}.
\subsection{Classification task}\label{sec4.2}
\begin{table}
\centering
\caption{Classification AUCs of the 14 diseases on the ChestX-ray14 test set.}\label{tab2}
\begin{tabular}{|l|c|c|c|c|c|}
\hline
Method & LSE~\cite{wang2017chestx} & LSE-LBA~\cite{yao2018weakly} & AGCL~\cite{tang2018attention} & ChestNet~\cite{wang2018chestnet} & PCAM pooling \\
\hline
Atelectasis & 0.700 & 0.733 & 0.756 & 0.743 & \textbf{0.772} \\
Cardiomegaly & 0.810 & 0.856 & \textbf{0.887} & 0.875 & 0.864 \\
Effusion & 0.759 & 0.806 & 0.819 & 0.811 & \textbf{0.825} \\
Infiltration & 0.661 & 0.673 & 0.689 & 0.677 & \textbf{0.694} \\
Mass & 0.693 & 0.777 & \textbf{0.814} & 0.783 & 0.813 \\
Nodule & 0.669 & 0.718 & 0.755 & 0.698 & \textbf{0.783} \\
Pneumonia & 0.658 & 0.684 & \textbf{0.729} & 0.696 & 0.721 \\
Pneumothorax & 0.799 & 0.805 & 0.850 & 0.810 & \textbf{0.868} \\
Consolidation & 0.703 & 0.711 & 0.728 & 0.726 & \textbf{0.732} \\
Edema & 0.805 & 0.806 & \textbf{0.848} & 0.833 & 0.833 \\
Emphysema & 0.833 & 0.842 & 0.908 & 0.822 & \textbf{0.931} \\
Fibrosis & 0.786 & 0.743 & 0.818 & 0.804 & \textbf{0.819} \\
Pleural Thickening & 0.684 & 0.724 & 0.765 & 0.751 & \textbf{0.788} \\
Hernia & 0.872 & 0.775 & 0.875 & \textbf{0.900} & 0.784 \\
\hline
\end{tabular}
\end{table}
Table~\ref{tab2} shows the Area Under the receiver operating characteristic Curves
(AUCs) of the classification task on the 14 thoracic diseases from the ChestX-ray14
official test set. PCAM pooling
outperforms most of the other state-of-the-art methods including the baseline
reported in ChestX-ray14~\cite{wang2017chestx}. We suspect explicitly utilizing CAM for localization
during training may also benefit the classification task. Note the results from
AGCL~\cite{tang2018attention} are obtained by training with additional severity-level
information.
\subsection{Localization task}\label{sec4.3}
\begin{table}
\centering
\caption{Localization accuracies and average false positives of the 8 diseases that have bounding boxes on the ChestX-ray14 test set.}\label{tab3}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|}
\hline
Method & AT & CM & EF & IF & MS & ND & PMN & PMT \\
\hline
\multicolumn{9}{|c|}{Localization accuracy, IoBB $>$ 0.5} \\
\hline
LSE-baseline~\cite{wang2017chestx} & 0.2833 & 0.8767 & 0.3333 & 0.4227 & 0.1411 & 0.0126 & 0.3833 & 0.1836 \\
PCAM pooling & \textbf{0.3500} & \textbf{0.9657} & \textbf{0.5359} & \textbf{0.7642} & \textbf{0.4118} & \textbf{0.0759} & \textbf{0.7667} & \textbf{0.1939} \\
\hline
\multicolumn{9}{|c|}{Average false positive} \\
\hline
LSE-baseline~\cite{wang2017chestx} & 1.020 & 0.563 & \textbf{0.927} & \textbf{0.659} & \textbf{0.694} & \textbf{0.619} & \textbf{1.013} & \textbf{0.529} \\
PCAM pooling & \textbf{0.867} & \textbf{0.021} & 1.137 & 1.805 & 1.000 & 1.228 & 1.200 & 1.684 \\
\hline
\end{tabular}
\begin{tablenotes}
\small
\item AT: Atelectasis, CM: Cardiomegaly, EF: Effusion, IF: Infiltration,
MS: Mass, ND: Nodule, PMN: Pneumonia, PMT: Pneumothorax.
\end{tablenotes}
\end{table}
Table~\ref{tab3} shows the localization accuracies and the average false positives
of the localization task on the 8 thoracic diseases that have bounding boxes. We use Intersection over
the predicted B-Box area ratio (IoBB) to measure the overlap between predicted
bounding boxes and ground truth bounding boxes annotated by radiologists
following~\cite{wang2017chestx}. A correct localization is defined as at least
one predicted bounding box is overlapped with the ground truth bounding box with
IoBB $>$ 0.5~\cite{wang2017chestx}. PCAM pooling outperforms the baseline
localization accuracy~\cite{wang2017chestx} by a significant margin on all of
the diseases, demonstrating its efficacy in weakly supervised lesion localization.
Figure~\ref{fig2} shows a few selected examples of the probability maps generated
by PCAM pooling and the class activation maps generated by LSE pooling together with
the predicted bounding boxes. Compared to the class activation maps, the probability
maps are visually more clear with sharp boundaries around lesion regions.
We attribute the improved visual quality to the probabilistic interpretation of
the sigmoid-bounded class activation map and explicitly using it for training
with global pooling.
We notice the probability maps generated by PCAM pooling tend to enlarge regions
of interest in general than class activation maps from LSE pooling, especially when
the ground truth regions are small, such as ``Nodule'' in Figure~\ref{fig2}.
This may explain the fact that PCAM pooling has relatively larger average false
positives than CAM with LSE pooling.
\begin{figure}
\centering
\begin{subfigure}[b]{0.48\linewidth}
\includegraphics[width=\linewidth]{Atelectasis.jpg}
\caption{Atelectasis}
\label{fig2:a}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\linewidth}
\includegraphics[width=\linewidth]{Cardiomegaly.jpg}
\caption{Cardiomegaly}
\label{fig2:b}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\linewidth}
\includegraphics[width=\linewidth]{Effusion.jpg}
\caption{Effusion}
\label{fig2:c}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\linewidth}
\includegraphics[width=\linewidth]{Pneumonia.jpg}
\caption{Pneumonia}
\label{fig2:d}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\linewidth}
\includegraphics[width=\linewidth]{Mass.jpg}
\caption{Mass}
\label{fig2:e}
\end{subfigure}
~
\begin{subfigure}[b]{0.48\linewidth}
\includegraphics[width=\linewidth]{Nodule.jpg}
\caption{Nodule}
\label{fig2:f}
\end{subfigure}
~
\caption{Selected samples of localization heatmaps and their bounding boxes
generated by LSE pooling and PCAM pooling on the test set of
ChestX-ray14~\cite{wang2017chestx}. In each subfigure, the left panel is the original
chest X-ray with the ground truth bounding boxes (green), the middle panel
is the class activation map and predicted bounding boxes (blue) by LSE pooling,
the right panel is the probability map and predicted bounding boxes (blue) by
PCAM pooling.}
\label{fig2}
\end{figure}
\section{Conclusion}
In this work, we present Probabilistic-CAM (PCAM) pooling, a new global pooling
operation to explicitly leverage the localization ability of CAM for training.
PCAM pooling is easy to implement and does not introduce any additional training
parameters. Experiments of weakly supervised lesion localization on the ChestX-ray14~\cite{wang2017chestx}
dataset demonstrate its efficacy in improving both the classification task and
the localization task compared to several state-of-the-art baselines. Visual
inspection of the probability maps generated by PCAM pooling shows clear and sharp
boundaries around lesion regions compared to the standard class activation maps.
Currently, PCAM pooling tends to generate localization maps that enlarge regions
of interest, which may increase false positives especially for small lesions. We
are working on reducing this effect as our future direction.
\small
\bibliographystyle{abbrv}
\input{document.bbl}
\end{document}
| 5310ca72d4adc1b51ed8e2461d38c3ee9f53a113 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
Subsets and Splits